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 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 �-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		52[54]		83[95]		23
107# AMD K8	46[41]		66[70]		18
108# PIII		41[50]		60[77]		24
109# Core 2	31[36]		45[64]		18.5
110# Atom		76[100]		96[138]		60
111# Pentium	115		150		77
112#
113# Version 4.1 switches to compact S-box even in key schedule setup.
114#
115# Version 4.2 prefetches compact S-box in every SSE round or in other
116# words every cache-line is *guaranteed* to be accessed within ~50
117# cycles window. Why just SSE? Because it's needed on hyper-threading
118# CPU! Which is also why it's prefetched with 64 byte stride. Best
119# part is that it has no negative effect on performance:-)
120#
121# Version 4.3 implements switch between compact and non-compact block
122# functions in AES_cbc_encrypt depending on how much data was asked
123# to be processed in one stroke.
124#
125######################################################################
126# Timing attacks are classified in two classes: synchronous when
127# attacker consciously initiates cryptographic operation and collects
128# timing data of various character afterwards, and asynchronous when
129# malicious code is executed on same CPU simultaneously with AES,
130# instruments itself and performs statistical analysis of this data.
131#
132# As far as synchronous attacks go the root to the AES timing
133# vulnerability is twofold. Firstly, of 256 S-box elements at most 160
134# are referred to in single 128-bit block operation. Well, in C
135# implementation with 4 distinct tables it's actually as little as 40
136# references per 256 elements table, but anyway... Secondly, even
137# though S-box elements are clustered into smaller amount of cache-
138# lines, smaller than 160 and even 40, it turned out that for certain
139# plain-text pattern[s] or simply put chosen plain-text and given key
140# few cache-lines remain unaccessed during block operation. Now, if
141# attacker can figure out this access pattern, he can deduct the key
142# [or at least part of it]. The natural way to mitigate this kind of
143# attacks is to minimize the amount of cache-lines in S-box and/or
144# prefetch them to ensure that every one is accessed for more uniform
145# timing. But note that *if* plain-text was concealed in such way that
146# input to block function is distributed *uniformly*, then attack
147# wouldn't apply. Now note that some encryption modes, most notably
148# CBC, do mask the plain-text in this exact way [secure cipher output
149# is distributed uniformly]. Yes, one still might find input that
150# would reveal the information about given key, but if amount of
151# candidate inputs to be tried is larger than amount of possible key
152# combinations then attack becomes infeasible. This is why revised
153# AES_cbc_encrypt "dares" to switch to larger S-box when larger chunk
154# of data is to be processed in one stroke. The current size limit of
155# 512 bytes is chosen to provide same [diminishigly low] probability
156# for cache-line to remain untouched in large chunk operation with
157# large S-box as for single block operation with compact S-box and
158# surely needs more careful consideration...
159#
160# As for asynchronous attacks. There are two flavours: attacker code
161# being interleaved with AES on hyper-threading CPU at *instruction*
162# level, and two processes time sharing single core. As for latter.
163# Two vectors. 1. Given that attacker process has higher priority,
164# yield execution to process performing AES just before timer fires
165# off the scheduler, immediately regain control of CPU and analyze the
166# cache state. For this attack to be efficient attacker would have to
167# effectively slow down the operation by several *orders* of magnitute,
168# by ratio of time slice to duration of handful of AES rounds, which
169# unlikely to remain unnoticed. Not to mention that this also means
170# that he would spend correspondigly more time to collect enough
171# statistical data to mount the attack. It's probably appropriate to
172# say that if adeversary reckons that this attack is beneficial and
173# risks to be noticed, you probably have larger problems having him
174# mere opportunity. In other words suggested code design expects you
175# to preclude/mitigate this attack by overall system security design.
176# 2. Attacker manages to make his code interrupt driven. In order for
177# this kind of attack to be feasible, interrupt rate has to be high
178# enough, again comparable to duration of handful of AES rounds. But
179# is there interrupt source of such rate? Hardly, not even 1Gbps NIC
180# generates interrupts at such raging rate...
181#
182# And now back to the former, hyper-threading CPU or more specifically
183# Intel P4. Recall that asynchronous attack implies that malicious
184# code instruments itself. And naturally instrumentation granularity
185# has be noticeably lower than duration of codepath accessing S-box.
186# Given that all cache-lines are accessed during that time that is.
187# Current implementation accesses *all* cache-lines within ~50 cycles
188# window, which is actually *less* than RDTSC latency on Intel P4!
189
190$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
191push(@INC,"${dir}","${dir}../../perlasm");
192require "x86asm.pl";
193
194&asm_init($ARGV[0],"aes-586.pl",$x86only = $ARGV[$#ARGV] eq "386");
195&static_label("AES_Te");
196&static_label("AES_Td");
197
198$s0="eax";
199$s1="ebx";
200$s2="ecx";
201$s3="edx";
202$key="edi";
203$acc="esi";
204$tbl="ebp";
205
206# stack frame layout in _[x86|sse]_AES_* routines, frame is allocated
207# by caller
208$__ra=&DWP(0,"esp");	# return address
209$__s0=&DWP(4,"esp");	# s0 backing store
210$__s1=&DWP(8,"esp");	# s1 backing store
211$__s2=&DWP(12,"esp");	# s2 backing store
212$__s3=&DWP(16,"esp");	# s3 backing store
213$__key=&DWP(20,"esp");	# pointer to key schedule
214$__end=&DWP(24,"esp");	# pointer to end of key schedule
215$__tbl=&DWP(28,"esp");	# %ebp backing store
216
217# stack frame layout in AES_[en|crypt] routines, which differs from
218# above by 4 and overlaps by %ebp backing store
219$_tbl=&DWP(24,"esp");
220$_esp=&DWP(28,"esp");
221
222sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
223
224$speed_limit=512;	# chunks smaller than $speed_limit are
225			# processed with compact routine in CBC mode
226$small_footprint=1;	# $small_footprint=1 code is ~5% slower [on
227			# recent �-archs], but ~5 times smaller!
228			# I favor compact code to minimize cache
229			# contention and in hope to "collect" 5% back
230			# in real-life applications...
231
232$vertical_spin=0;	# shift "verticaly" defaults to 0, because of
233			# its proof-of-concept status...
234# Note that there is no decvert(), as well as last encryption round is
235# performed with "horizontal" shifts. This is because this "vertical"
236# implementation [one which groups shifts on a given $s[i] to form a
237# "column," unlike "horizontal" one, which groups shifts on different
238# $s[i] to form a "row"] is work in progress. It was observed to run
239# few percents faster on Intel cores, but not AMD. On AMD K8 core it's
240# whole 12% slower:-( So we face a trade-off... Shall it be resolved
241# some day? Till then the code is considered experimental and by
242# default remains dormant...
243
244sub encvert()
245{ my ($te,@s) = @_;
246  my ($v0,$v1) = ($acc,$key);
247
248	&mov	($v0,$s[3]);				# copy s3
249	&mov	(&DWP(4,"esp"),$s[2]);			# save s2
250	&mov	($v1,$s[0]);				# copy s0
251	&mov	(&DWP(8,"esp"),$s[1]);			# save s1
252
253	&movz	($s[2],&HB($s[0]));
254	&and	($s[0],0xFF);
255	&mov	($s[0],&DWP(0,$te,$s[0],8));		# s0>>0
256	&shr	($v1,16);
257	&mov	($s[3],&DWP(3,$te,$s[2],8));		# s0>>8
258	&movz	($s[1],&HB($v1));
259	&and	($v1,0xFF);
260	&mov	($s[2],&DWP(2,$te,$v1,8));		# s0>>16
261	 &mov	($v1,$v0);
262	&mov	($s[1],&DWP(1,$te,$s[1],8));		# s0>>24
263
264	&and	($v0,0xFF);
265	&xor	($s[3],&DWP(0,$te,$v0,8));		# s3>>0
266	&movz	($v0,&HB($v1));
267	&shr	($v1,16);
268	&xor	($s[2],&DWP(3,$te,$v0,8));		# s3>>8
269	&movz	($v0,&HB($v1));
270	&and	($v1,0xFF);
271	&xor	($s[1],&DWP(2,$te,$v1,8));		# s3>>16
272	 &mov	($v1,&DWP(4,"esp"));			# restore s2
273	&xor	($s[0],&DWP(1,$te,$v0,8));		# s3>>24
274
275	&mov	($v0,$v1);
276	&and	($v1,0xFF);
277	&xor	($s[2],&DWP(0,$te,$v1,8));		# s2>>0
278	&movz	($v1,&HB($v0));
279	&shr	($v0,16);
280	&xor	($s[1],&DWP(3,$te,$v1,8));		# s2>>8
281	&movz	($v1,&HB($v0));
282	&and	($v0,0xFF);
283	&xor	($s[0],&DWP(2,$te,$v0,8));		# s2>>16
284	 &mov	($v0,&DWP(8,"esp"));			# restore s1
285	&xor	($s[3],&DWP(1,$te,$v1,8));		# s2>>24
286
287	&mov	($v1,$v0);
288	&and	($v0,0xFF);
289	&xor	($s[1],&DWP(0,$te,$v0,8));		# s1>>0
290	&movz	($v0,&HB($v1));
291	&shr	($v1,16);
292	&xor	($s[0],&DWP(3,$te,$v0,8));		# s1>>8
293	&movz	($v0,&HB($v1));
294	&and	($v1,0xFF);
295	&xor	($s[3],&DWP(2,$te,$v1,8));		# s1>>16
296	 &mov	($key,$__key);				# reincarnate v1 as key
297	&xor	($s[2],&DWP(1,$te,$v0,8));		# s1>>24
298}
299
300# Another experimental routine, which features "horizontal spin," but
301# eliminates one reference to stack. Strangely enough runs slower...
302sub enchoriz()
303{ my ($v0,$v1) = ($key,$acc);
304
305	&movz	($v0,&LB($s0));			#  3, 2, 1, 0*
306	&rotr	($s2,8);			#  8,11,10, 9
307	&mov	($v1,&DWP(0,$te,$v0,8));	#  0
308	&movz	($v0,&HB($s1));			#  7, 6, 5*, 4
309	&rotr	($s3,16);			# 13,12,15,14
310	&xor	($v1,&DWP(3,$te,$v0,8));	#  5
311	&movz	($v0,&HB($s2));			#  8,11,10*, 9
312	&rotr	($s0,16);			#  1, 0, 3, 2
313	&xor	($v1,&DWP(2,$te,$v0,8));	# 10
314	&movz	($v0,&HB($s3));			# 13,12,15*,14
315	&xor	($v1,&DWP(1,$te,$v0,8));	# 15, t[0] collected
316	&mov	($__s0,$v1);			# t[0] saved
317
318	&movz	($v0,&LB($s1));			#  7, 6, 5, 4*
319	&shr	($s1,16);			#  -, -, 7, 6
320	&mov	($v1,&DWP(0,$te,$v0,8));	#  4
321	&movz	($v0,&LB($s3));			# 13,12,15,14*
322	&xor	($v1,&DWP(2,$te,$v0,8));	# 14
323	&movz	($v0,&HB($s0));			#  1, 0, 3*, 2
324	&and	($s3,0xffff0000);		# 13,12, -, -
325	&xor	($v1,&DWP(1,$te,$v0,8));	#  3
326	&movz	($v0,&LB($s2));			#  8,11,10, 9*
327	&or	($s3,$s1);			# 13,12, 7, 6
328	&xor	($v1,&DWP(3,$te,$v0,8));	#  9, t[1] collected
329	&mov	($s1,$v1);			#  s[1]=t[1]
330
331	&movz	($v0,&LB($s0));			#  1, 0, 3, 2*
332	&shr	($s2,16);			#  -, -, 8,11
333	&mov	($v1,&DWP(2,$te,$v0,8));	#  2
334	&movz	($v0,&HB($s3));			# 13,12, 7*, 6
335	&xor	($v1,&DWP(1,$te,$v0,8));	#  7
336	&movz	($v0,&HB($s2));			#  -, -, 8*,11
337	&xor	($v1,&DWP(0,$te,$v0,8));	#  8
338	&mov	($v0,$s3);
339	&shr	($v0,24);			# 13
340	&xor	($v1,&DWP(3,$te,$v0,8));	# 13, t[2] collected
341
342	&movz	($v0,&LB($s2));			#  -, -, 8,11*
343	&shr	($s0,24);			#  1*
344	&mov	($s2,&DWP(1,$te,$v0,8));	# 11
345	&xor	($s2,&DWP(3,$te,$s0,8));	#  1
346	&mov	($s0,$__s0);			# s[0]=t[0]
347	&movz	($v0,&LB($s3));			# 13,12, 7, 6*
348	&shr	($s3,16);			#   ,  ,13,12
349	&xor	($s2,&DWP(2,$te,$v0,8));	#  6
350	&mov	($key,$__key);			# reincarnate v0 as key
351	&and	($s3,0xff);			#   ,  ,13,12*
352	&mov	($s3,&DWP(0,$te,$s3,8));	# 12
353	&xor	($s3,$s2);			# s[2]=t[3] collected
354	&mov	($s2,$v1);			# s[2]=t[2]
355}
356
357# More experimental code... SSE one... Even though this one eliminates
358# *all* references to stack, it's not faster...
359sub sse_encbody()
360{
361	&movz	($acc,&LB("eax"));		#  0
362	&mov	("ecx",&DWP(0,$tbl,$acc,8));	#  0
363	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
364	&movz	("edx",&HB("eax"));		#  1
365	&mov	("edx",&DWP(3,$tbl,"edx",8));	#  1
366	&shr	("eax",16);			#  5, 4
367
368	&movz	($acc,&LB("ebx"));		# 10
369	&xor	("ecx",&DWP(2,$tbl,$acc,8));	# 10
370	&pshufw	("mm6","mm4",0x08);		# 13,12, 9, 8
371	&movz	($acc,&HB("ebx"));		# 11
372	&xor	("edx",&DWP(1,$tbl,$acc,8));	# 11
373	&shr	("ebx",16);			# 15,14
374
375	&movz	($acc,&HB("eax"));		#  5
376	&xor	("ecx",&DWP(3,$tbl,$acc,8));	#  5
377	&movq	("mm3",QWP(16,$key));
378	&movz	($acc,&HB("ebx"));		# 15
379	&xor	("ecx",&DWP(1,$tbl,$acc,8));	# 15
380	&movd	("mm0","ecx");			# t[0] collected
381
382	&movz	($acc,&LB("eax"));		#  4
383	&mov	("ecx",&DWP(0,$tbl,$acc,8));	#  4
384	&movd	("eax","mm2");			#  7, 6, 3, 2
385	&movz	($acc,&LB("ebx"));		# 14
386	&xor	("ecx",&DWP(2,$tbl,$acc,8));	# 14
387	&movd	("ebx","mm6");			# 13,12, 9, 8
388
389	&movz	($acc,&HB("eax"));		#  3
390	&xor	("ecx",&DWP(1,$tbl,$acc,8));	#  3
391	&movz	($acc,&HB("ebx"));		#  9
392	&xor	("ecx",&DWP(3,$tbl,$acc,8));	#  9
393	&movd	("mm1","ecx");			# t[1] collected
394
395	&movz	($acc,&LB("eax"));		#  2
396	&mov	("ecx",&DWP(2,$tbl,$acc,8));	#  2
397	&shr	("eax",16);			#  7, 6
398	&punpckldq	("mm0","mm1");		# t[0,1] collected
399	&movz	($acc,&LB("ebx"));		#  8
400	&xor	("ecx",&DWP(0,$tbl,$acc,8));	#  8
401	&shr	("ebx",16);			# 13,12
402
403	&movz	($acc,&HB("eax"));		#  7
404	&xor	("ecx",&DWP(1,$tbl,$acc,8));	#  7
405	&pxor	("mm0","mm3");
406	&movz	("eax",&LB("eax"));		#  6
407	&xor	("edx",&DWP(2,$tbl,"eax",8));	#  6
408	&pshufw	("mm1","mm0",0x08);		#  5, 4, 1, 0
409	&movz	($acc,&HB("ebx"));		# 13
410	&xor	("ecx",&DWP(3,$tbl,$acc,8));	# 13
411	&xor	("ecx",&DWP(24,$key));		# t[2]
412	&movd	("mm4","ecx");			# t[2] collected
413	&movz	("ebx",&LB("ebx"));		# 12
414	&xor	("edx",&DWP(0,$tbl,"ebx",8));	# 12
415	&shr	("ecx",16);
416	&movd	("eax","mm1");			#  5, 4, 1, 0
417	&mov	("ebx",&DWP(28,$key));		# t[3]
418	&xor	("ebx","edx");
419	&movd	("mm5","ebx");			# t[3] collected
420	&and	("ebx",0xffff0000);
421	&or	("ebx","ecx");
422
423	&punpckldq	("mm4","mm5");		# t[2,3] collected
424}
425
426######################################################################
427# "Compact" block function
428######################################################################
429
430sub enccompact()
431{ my $Fn = \&mov;
432  while ($#_>5) { pop(@_); $Fn=sub{}; }
433  my ($i,$te,@s)=@_;
434  my $tmp = $key;
435  my $out = $i==3?$s[0]:$acc;
436
437	# $Fn is used in first compact round and its purpose is to
438	# void restoration of some values from stack, so that after
439	# 4xenccompact with extra argument $key value is left there...
440	if ($i==3)  {	&$Fn	($key,$__key);			}##%edx
441	else        {	&mov	($out,$s[0]);			}
442			&and	($out,0xFF);
443	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
444	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
445			&movz	($out,&BP(-128,$te,$out,1));
446
447	if ($i==3)  {	$tmp=$s[1];				}##%eax
448			&movz	($tmp,&HB($s[1]));
449			&movz	($tmp,&BP(-128,$te,$tmp,1));
450			&shl	($tmp,8);
451			&xor	($out,$tmp);
452
453	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
454	else        {	&mov	($tmp,$s[2]);
455			&shr	($tmp,16);			}
456	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
457			&and	($tmp,0xFF);
458			&movz	($tmp,&BP(-128,$te,$tmp,1));
459			&shl	($tmp,16);
460			&xor	($out,$tmp);
461
462	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
463	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
464	else        {	&mov	($tmp,$s[3]);
465			&shr	($tmp,24);			}
466			&movz	($tmp,&BP(-128,$te,$tmp,1));
467			&shl	($tmp,24);
468			&xor	($out,$tmp);
469	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
470	if ($i==3)  {	&mov	($s[3],$acc);			}
471	&comment();
472}
473
474sub enctransform()
475{ my @s = ($s0,$s1,$s2,$s3);
476  my $i = shift;
477  my $tmp = $tbl;
478  my $r2  = $key ;
479
480	&and	($tmp,$s[$i]);
481	&lea	($r2,&DWP(0,$s[$i],$s[$i]));
482	&mov	($acc,$tmp);
483	&shr	($tmp,7);
484	&and	($r2,0xfefefefe);
485	&sub	($acc,$tmp);
486	&mov	($tmp,$s[$i]);
487	&and	($acc,0x1b1b1b1b);
488	&rotr	($tmp,16);
489	&xor	($acc,$r2);	# r2
490	&mov	($r2,$s[$i]);
491
492	&xor	($s[$i],$acc);	# r0 ^ r2
493	&rotr	($r2,16+8);
494	&xor	($acc,$tmp);
495	&rotl	($s[$i],24);
496	&xor	($acc,$r2);
497	&mov	($tmp,0x80808080)	if ($i!=1);
498	&xor	($s[$i],$acc);	# ROTATE(r2^r0,24) ^ r2
499}
500
501&function_begin_B("_x86_AES_encrypt_compact");
502	# note that caller is expected to allocate stack frame for me!
503	&mov	($__key,$key);			# save key
504
505	&xor	($s0,&DWP(0,$key));		# xor with key
506	&xor	($s1,&DWP(4,$key));
507	&xor	($s2,&DWP(8,$key));
508	&xor	($s3,&DWP(12,$key));
509
510	&mov	($acc,&DWP(240,$key));		# load key->rounds
511	&lea	($acc,&DWP(-2,$acc,$acc));
512	&lea	($acc,&DWP(0,$key,$acc,8));
513	&mov	($__end,$acc);			# end of key schedule
514
515	# prefetch Te4
516	&mov	($key,&DWP(0-128,$tbl));
517	&mov	($acc,&DWP(32-128,$tbl));
518	&mov	($key,&DWP(64-128,$tbl));
519	&mov	($acc,&DWP(96-128,$tbl));
520	&mov	($key,&DWP(128-128,$tbl));
521	&mov	($acc,&DWP(160-128,$tbl));
522	&mov	($key,&DWP(192-128,$tbl));
523	&mov	($acc,&DWP(224-128,$tbl));
524
525	&set_label("loop",16);
526
527		&enccompact(0,$tbl,$s0,$s1,$s2,$s3,1);
528		&enccompact(1,$tbl,$s1,$s2,$s3,$s0,1);
529		&enccompact(2,$tbl,$s2,$s3,$s0,$s1,1);
530		&enccompact(3,$tbl,$s3,$s0,$s1,$s2,1);
531		&mov	($tbl,0x80808080);
532		&enctransform(2);
533		&enctransform(3);
534		&enctransform(0);
535		&enctransform(1);
536		&mov 	($key,$__key);
537		&mov	($tbl,$__tbl);
538		&add	($key,16);		# advance rd_key
539		&xor	($s0,&DWP(0,$key));
540		&xor	($s1,&DWP(4,$key));
541		&xor	($s2,&DWP(8,$key));
542		&xor	($s3,&DWP(12,$key));
543
544	&cmp	($key,$__end);
545	&mov	($__key,$key);
546	&jb	(&label("loop"));
547
548	&enccompact(0,$tbl,$s0,$s1,$s2,$s3);
549	&enccompact(1,$tbl,$s1,$s2,$s3,$s0);
550	&enccompact(2,$tbl,$s2,$s3,$s0,$s1);
551	&enccompact(3,$tbl,$s3,$s0,$s1,$s2);
552
553	&xor	($s0,&DWP(16,$key));
554	&xor	($s1,&DWP(20,$key));
555	&xor	($s2,&DWP(24,$key));
556	&xor	($s3,&DWP(28,$key));
557
558	&ret	();
559&function_end_B("_x86_AES_encrypt_compact");
560
561######################################################################
562# "Compact" SSE block function.
563######################################################################
564#
565# Performance is not actually extraordinary in comparison to pure
566# x86 code. In particular encrypt performance is virtually the same.
567# Decrypt performance on the other hand is 15-20% better on newer
568# �-archs [but we're thankful for *any* improvement here], and ~50%
569# better on PIII:-) And additionally on the pros side this code
570# eliminates redundant references to stack and thus relieves/
571# minimizes the pressure on the memory bus.
572#
573# MMX register layout                           lsb
574# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
575# |          mm4          |          mm0          |
576# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
577# |     s3    |     s2    |     s1    |     s0    |
578# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
579# |15|14|13|12|11|10| 9| 8| 7| 6| 5| 4| 3| 2| 1| 0|
580# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
581#
582# Indexes translate as s[N/4]>>(8*(N%4)), e.g. 5 means s1>>8.
583# In this terms encryption and decryption "compact" permutation
584# matrices can be depicted as following:
585#
586# encryption              lsb	# decryption              lsb
587# +----++----+----+----+----+	# +----++----+----+----+----+
588# | t0 || 15 | 10 |  5 |  0 |	# | t0 ||  7 | 10 | 13 |  0 |
589# +----++----+----+----+----+	# +----++----+----+----+----+
590# | t1 ||  3 | 14 |  9 |  4 |	# | t1 || 11 | 14 |  1 |  4 |
591# +----++----+----+----+----+	# +----++----+----+----+----+
592# | t2 ||  7 |  2 | 13 |  8 |	# | t2 || 15 |  2 |  5 |  8 |
593# +----++----+----+----+----+	# +----++----+----+----+----+
594# | t3 || 11 |  6 |  1 | 12 |	# | t3 ||  3 |  6 |  9 | 12 |
595# +----++----+----+----+----+	# +----++----+----+----+----+
596#
597######################################################################
598# Why not xmm registers? Short answer. It was actually tested and
599# was not any faster, but *contrary*, most notably on Intel CPUs.
600# Longer answer. Main advantage of using mm registers is that movd
601# latency is lower, especially on Intel P4. While arithmetic
602# instructions are twice as many, they can be scheduled every cycle
603# and not every second one when they are operating on xmm register,
604# so that "arithmetic throughput" remains virtually the same. And
605# finally the code can be executed even on elder SSE-only CPUs:-)
606
607sub sse_enccompact()
608{
609	&pshufw	("mm1","mm0",0x08);		#  5, 4, 1, 0
610	&pshufw	("mm5","mm4",0x0d);		# 15,14,11,10
611	&movd	("eax","mm1");			#  5, 4, 1, 0
612	&movd	("ebx","mm5");			# 15,14,11,10
613	&mov	($__key,$key);
614
615	&movz	($acc,&LB("eax"));		#  0
616	&movz	("edx",&HB("eax"));		#  1
617	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
618	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
619	&movz	($key,&LB("ebx"));		# 10
620	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
621	&shr	("eax",16);			#  5, 4
622	&shl	("edx",8);			#  1
623
624	&movz	($acc,&BP(-128,$tbl,$key,1));	# 10
625	&movz	($key,&HB("ebx"));		# 11
626	&shl	($acc,16);			# 10
627	&pshufw	("mm6","mm4",0x08);		# 13,12, 9, 8
628	&or	("ecx",$acc);			# 10
629	&movz	($acc,&BP(-128,$tbl,$key,1));	# 11
630	&movz	($key,&HB("eax"));		#  5
631	&shl	($acc,24);			# 11
632	&shr	("ebx",16);			# 15,14
633	&or	("edx",$acc);			# 11
634
635	&movz	($acc,&BP(-128,$tbl,$key,1));	#  5
636	&movz	($key,&HB("ebx"));		# 15
637	&shl	($acc,8);			#  5
638	&or	("ecx",$acc);			#  5
639	&movz	($acc,&BP(-128,$tbl,$key,1));	# 15
640	&movz	($key,&LB("eax"));		#  4
641	&shl	($acc,24);			# 15
642	&or	("ecx",$acc);			# 15
643
644	&movz	($acc,&BP(-128,$tbl,$key,1));	#  4
645	&movz	($key,&LB("ebx"));		# 14
646	&movd	("eax","mm2");			#  7, 6, 3, 2
647	&movd	("mm0","ecx");			# t[0] collected
648	&movz	("ecx",&BP(-128,$tbl,$key,1));	# 14
649	&movz	($key,&HB("eax"));		#  3
650	&shl	("ecx",16);			# 14
651	&movd	("ebx","mm6");			# 13,12, 9, 8
652	&or	("ecx",$acc);			# 14
653
654	&movz	($acc,&BP(-128,$tbl,$key,1));	#  3
655	&movz	($key,&HB("ebx"));		#  9
656	&shl	($acc,24);			#  3
657	&or	("ecx",$acc);			#  3
658	&movz	($acc,&BP(-128,$tbl,$key,1));	#  9
659	&movz	($key,&LB("ebx"));		#  8
660	&shl	($acc,8);			#  9
661	&shr	("ebx",16);			# 13,12
662	&or	("ecx",$acc);			#  9
663
664	&movz	($acc,&BP(-128,$tbl,$key,1));	#  8
665	&movz	($key,&LB("eax"));		#  2
666	&shr	("eax",16);			#  7, 6
667	&movd	("mm1","ecx");			# t[1] collected
668	&movz	("ecx",&BP(-128,$tbl,$key,1));	#  2
669	&movz	($key,&HB("eax"));		#  7
670	&shl	("ecx",16);			#  2
671	&and	("eax",0xff);			#  6
672	&or	("ecx",$acc);			#  2
673
674	&punpckldq	("mm0","mm1");		# t[0,1] collected
675
676	&movz	($acc,&BP(-128,$tbl,$key,1));	#  7
677	&movz	($key,&HB("ebx"));		# 13
678	&shl	($acc,24);			#  7
679	&and	("ebx",0xff);			# 12
680	&movz	("eax",&BP(-128,$tbl,"eax",1));	#  6
681	&or	("ecx",$acc);			#  7
682	&shl	("eax",16);			#  6
683	&movz	($acc,&BP(-128,$tbl,$key,1));	# 13
684	&or	("edx","eax");			#  6
685	&shl	($acc,8);			# 13
686	&movz	("ebx",&BP(-128,$tbl,"ebx",1));	# 12
687	&or	("ecx",$acc);			# 13
688	&or	("edx","ebx");			# 12
689	&mov	($key,$__key);
690	&movd	("mm4","ecx");			# t[2] collected
691	&movd	("mm5","edx");			# t[3] collected
692
693	&punpckldq	("mm4","mm5");		# t[2,3] collected
694}
695
696					if (!$x86only) {
697&function_begin_B("_sse_AES_encrypt_compact");
698	&pxor	("mm0",&QWP(0,$key));	#  7, 6, 5, 4, 3, 2, 1, 0
699	&pxor	("mm4",&QWP(8,$key));	# 15,14,13,12,11,10, 9, 8
700
701	# note that caller is expected to allocate stack frame for me!
702	&mov	($acc,&DWP(240,$key));		# load key->rounds
703	&lea	($acc,&DWP(-2,$acc,$acc));
704	&lea	($acc,&DWP(0,$key,$acc,8));
705	&mov	($__end,$acc);			# end of key schedule
706
707	&mov	($s0,0x1b1b1b1b);		# magic constant
708	&mov	(&DWP(8,"esp"),$s0);
709	&mov	(&DWP(12,"esp"),$s0);
710
711	# prefetch Te4
712	&mov	($s0,&DWP(0-128,$tbl));
713	&mov	($s1,&DWP(32-128,$tbl));
714	&mov	($s2,&DWP(64-128,$tbl));
715	&mov	($s3,&DWP(96-128,$tbl));
716	&mov	($s0,&DWP(128-128,$tbl));
717	&mov	($s1,&DWP(160-128,$tbl));
718	&mov	($s2,&DWP(192-128,$tbl));
719	&mov	($s3,&DWP(224-128,$tbl));
720
721	&set_label("loop",16);
722		&sse_enccompact();
723		&add	($key,16);
724		&cmp	($key,$__end);
725		&ja	(&label("out"));
726
727		&movq	("mm2",&QWP(8,"esp"));
728		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
729		&movq	("mm1","mm0");		&movq	("mm5","mm4");	# r0
730		&pcmpgtb("mm3","mm0");		&pcmpgtb("mm7","mm4");
731		&pand	("mm3","mm2");		&pand	("mm7","mm2");
732		&pshufw	("mm2","mm0",0xb1);	&pshufw	("mm6","mm4",0xb1);# ROTATE(r0,16)
733		&paddb	("mm0","mm0");		&paddb	("mm4","mm4");
734		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# = r2
735		&pshufw	("mm3","mm2",0xb1);	&pshufw	("mm7","mm6",0xb1);# r0
736		&pxor	("mm1","mm0");		&pxor	("mm5","mm4");	# r0^r2
737		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= ROTATE(r0,16)
738
739		&movq	("mm2","mm3");		&movq	("mm6","mm7");
740		&pslld	("mm3",8);		&pslld	("mm7",8);
741		&psrld	("mm2",24);		&psrld	("mm6",24);
742		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= r0<<8
743		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= r0>>24
744
745		&movq	("mm3","mm1");		&movq	("mm7","mm5");
746		&movq	("mm2",&QWP(0,$key));	&movq	("mm6",&QWP(8,$key));
747		&psrld	("mm1",8);		&psrld	("mm5",8);
748		&mov	($s0,&DWP(0-128,$tbl));
749		&pslld	("mm3",24);		&pslld	("mm7",24);
750		&mov	($s1,&DWP(64-128,$tbl));
751		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= (r2^r0)<<8
752		&mov	($s2,&DWP(128-128,$tbl));
753		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= (r2^r0)>>24
754		&mov	($s3,&DWP(192-128,$tbl));
755
756		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");
757	&jmp	(&label("loop"));
758
759	&set_label("out",16);
760	&pxor	("mm0",&QWP(0,$key));
761	&pxor	("mm4",&QWP(8,$key));
762
763	&ret	();
764&function_end_B("_sse_AES_encrypt_compact");
765					}
766
767######################################################################
768# Vanilla block function.
769######################################################################
770
771sub encstep()
772{ my ($i,$te,@s) = @_;
773  my $tmp = $key;
774  my $out = $i==3?$s[0]:$acc;
775
776	# lines marked with #%e?x[i] denote "reordered" instructions...
777	if ($i==3)  {	&mov	($key,$__key);			}##%edx
778	else        {	&mov	($out,$s[0]);
779			&and	($out,0xFF);			}
780	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
781	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
782			&mov	($out,&DWP(0,$te,$out,8));
783
784	if ($i==3)  {	$tmp=$s[1];				}##%eax
785			&movz	($tmp,&HB($s[1]));
786			&xor	($out,&DWP(3,$te,$tmp,8));
787
788	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
789	else        {	&mov	($tmp,$s[2]);
790			&shr	($tmp,16);			}
791	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
792			&and	($tmp,0xFF);
793			&xor	($out,&DWP(2,$te,$tmp,8));
794
795	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
796	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
797	else        {	&mov	($tmp,$s[3]);
798			&shr	($tmp,24)			}
799			&xor	($out,&DWP(1,$te,$tmp,8));
800	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
801	if ($i==3)  {	&mov	($s[3],$acc);			}
802			&comment();
803}
804
805sub enclast()
806{ my ($i,$te,@s)=@_;
807  my $tmp = $key;
808  my $out = $i==3?$s[0]:$acc;
809
810	if ($i==3)  {	&mov	($key,$__key);			}##%edx
811	else        {	&mov	($out,$s[0]);			}
812			&and	($out,0xFF);
813	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
814	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
815			&mov	($out,&DWP(2,$te,$out,8));
816			&and	($out,0x000000ff);
817
818	if ($i==3)  {	$tmp=$s[1];				}##%eax
819			&movz	($tmp,&HB($s[1]));
820			&mov	($tmp,&DWP(0,$te,$tmp,8));
821			&and	($tmp,0x0000ff00);
822			&xor	($out,$tmp);
823
824	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
825	else        {	&mov	($tmp,$s[2]);
826			&shr	($tmp,16);			}
827	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
828			&and	($tmp,0xFF);
829			&mov	($tmp,&DWP(0,$te,$tmp,8));
830			&and	($tmp,0x00ff0000);
831			&xor	($out,$tmp);
832
833	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
834	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
835	else        {	&mov	($tmp,$s[3]);
836			&shr	($tmp,24);			}
837			&mov	($tmp,&DWP(2,$te,$tmp,8));
838			&and	($tmp,0xff000000);
839			&xor	($out,$tmp);
840	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
841	if ($i==3)  {	&mov	($s[3],$acc);			}
842}
843
844&function_begin_B("_x86_AES_encrypt");
845	if ($vertical_spin) {
846		# I need high parts of volatile registers to be accessible...
847		&exch	($s1="edi",$key="ebx");
848		&mov	($s2="esi",$acc="ecx");
849	}
850
851	# note that caller is expected to allocate stack frame for me!
852	&mov	($__key,$key);			# save key
853
854	&xor	($s0,&DWP(0,$key));		# xor with key
855	&xor	($s1,&DWP(4,$key));
856	&xor	($s2,&DWP(8,$key));
857	&xor	($s3,&DWP(12,$key));
858
859	&mov	($acc,&DWP(240,$key));		# load key->rounds
860
861	if ($small_footprint) {
862	    &lea	($acc,&DWP(-2,$acc,$acc));
863	    &lea	($acc,&DWP(0,$key,$acc,8));
864	    &mov	($__end,$acc);		# end of key schedule
865
866	    &set_label("loop",16);
867		if ($vertical_spin) {
868		    &encvert($tbl,$s0,$s1,$s2,$s3);
869		} else {
870		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
871		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
872		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
873		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
874		}
875		&add	($key,16);		# advance rd_key
876		&xor	($s0,&DWP(0,$key));
877		&xor	($s1,&DWP(4,$key));
878		&xor	($s2,&DWP(8,$key));
879		&xor	($s3,&DWP(12,$key));
880	    &cmp	($key,$__end);
881	    &mov	($__key,$key);
882	    &jb		(&label("loop"));
883	}
884	else {
885	    &cmp	($acc,10);
886	    &jle	(&label("10rounds"));
887	    &cmp	($acc,12);
888	    &jle	(&label("12rounds"));
889
890	&set_label("14rounds",4);
891	    for ($i=1;$i<3;$i++) {
892		if ($vertical_spin) {
893		    &encvert($tbl,$s0,$s1,$s2,$s3);
894		} else {
895		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
896		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
897		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
898		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
899		}
900		&xor	($s0,&DWP(16*$i+0,$key));
901		&xor	($s1,&DWP(16*$i+4,$key));
902		&xor	($s2,&DWP(16*$i+8,$key));
903		&xor	($s3,&DWP(16*$i+12,$key));
904	    }
905	    &add	($key,32);
906	    &mov	($__key,$key);		# advance rd_key
907	&set_label("12rounds",4);
908	    for ($i=1;$i<3;$i++) {
909		if ($vertical_spin) {
910		    &encvert($tbl,$s0,$s1,$s2,$s3);
911		} else {
912		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
913		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
914		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
915		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
916		}
917		&xor	($s0,&DWP(16*$i+0,$key));
918		&xor	($s1,&DWP(16*$i+4,$key));
919		&xor	($s2,&DWP(16*$i+8,$key));
920		&xor	($s3,&DWP(16*$i+12,$key));
921	    }
922	    &add	($key,32);
923	    &mov	($__key,$key);		# advance rd_key
924	&set_label("10rounds",4);
925	    for ($i=1;$i<10;$i++) {
926		if ($vertical_spin) {
927		    &encvert($tbl,$s0,$s1,$s2,$s3);
928		} else {
929		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
930		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
931		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
932		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
933		}
934		&xor	($s0,&DWP(16*$i+0,$key));
935		&xor	($s1,&DWP(16*$i+4,$key));
936		&xor	($s2,&DWP(16*$i+8,$key));
937		&xor	($s3,&DWP(16*$i+12,$key));
938	    }
939	}
940
941	if ($vertical_spin) {
942	    # "reincarnate" some registers for "horizontal" spin...
943	    &mov	($s1="ebx",$key="edi");
944	    &mov	($s2="ecx",$acc="esi");
945	}
946	&enclast(0,$tbl,$s0,$s1,$s2,$s3);
947	&enclast(1,$tbl,$s1,$s2,$s3,$s0);
948	&enclast(2,$tbl,$s2,$s3,$s0,$s1);
949	&enclast(3,$tbl,$s3,$s0,$s1,$s2);
950
951	&add	($key,$small_footprint?16:160);
952	&xor	($s0,&DWP(0,$key));
953	&xor	($s1,&DWP(4,$key));
954	&xor	($s2,&DWP(8,$key));
955	&xor	($s3,&DWP(12,$key));
956
957	&ret	();
958
959&set_label("AES_Te",64);	# Yes! I keep it in the code segment!
960	&_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
961	&_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
962	&_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
963	&_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
964	&_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
965	&_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
966	&_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
967	&_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
968	&_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
969	&_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
970	&_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
971	&_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
972	&_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
973	&_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
974	&_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
975	&_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
976	&_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
977	&_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
978	&_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
979	&_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
980	&_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
981	&_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
982	&_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
983	&_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
984	&_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
985	&_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
986	&_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
987	&_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
988	&_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
989	&_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
990	&_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
991	&_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
992	&_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
993	&_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
994	&_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
995	&_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
996	&_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
997	&_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
998	&_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
999	&_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
1000	&_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
1001	&_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
1002	&_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
1003	&_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
1004	&_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
1005	&_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
1006	&_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
1007	&_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);
1008	&_data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
1009	&_data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
1010	&_data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
1011	&_data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
1012	&_data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
1013	&_data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
1014	&_data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
1015	&_data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
1016	&_data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
1017	&_data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
1018	&_data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
1019	&_data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
1020	&_data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
1021	&_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
1022	&_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
1023	&_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
1024
1025#Te4	# four copies of Te4 to choose from to avoid L1 aliasing
1026	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1027	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1028	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1029	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1030	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1031	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1032	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1033	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1034	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1035	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1036	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1037	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1038	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1039	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1040	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1041	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1042	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1043	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1044	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1045	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1046	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1047	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1048	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1049	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1050	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1051	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1052	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1053	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1054	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1055	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1056	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1057	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1058
1059	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1060	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1061	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1062	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1063	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1064	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1065	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1066	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1067	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1068	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1069	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1070	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1071	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1072	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1073	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1074	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1075	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1076	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1077	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1078	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1079	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1080	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1081	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1082	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1083	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1084	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1085	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1086	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1087	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1088	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1089	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1090	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1091
1092	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1093	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1094	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1095	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1096	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1097	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1098	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1099	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1100	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1101	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1102	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1103	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1104	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1105	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1106	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1107	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1108	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1109	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1110	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1111	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1112	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1113	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1114	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1115	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1116	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1117	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1118	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1119	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1120	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1121	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1122	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1123	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1124
1125	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1126	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1127	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1128	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1129	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1130	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1131	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1132	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1133	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1134	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1135	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1136	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1137	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1138	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1139	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1140	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1141	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1142	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1143	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1144	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1145	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1146	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1147	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1148	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1149	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1150	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1151	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1152	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1153	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1154	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1155	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1156	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1157#rcon:
1158	&data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008);
1159	&data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080);
1160	&data_word(0x0000001b, 0x00000036, 0x00000000, 0x00000000);
1161	&data_word(0x00000000, 0x00000000, 0x00000000, 0x00000000);
1162&function_end_B("_x86_AES_encrypt");
1163
1164# void asm_AES_encrypt (const void *inp,void *out,const AES_KEY *key);
1165&function_begin("asm_AES_encrypt");
1166	&mov	($acc,&wparam(0));		# load inp
1167	&mov	($key,&wparam(2));		# load key
1168
1169	&mov	($s0,"esp");
1170	&sub	("esp",36);
1171	&and	("esp",-64);			# align to cache-line
1172
1173	# place stack frame just "above" the key schedule
1174	&lea	($s1,&DWP(-64-63,$key));
1175	&sub	($s1,"esp");
1176	&neg	($s1);
1177	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
1178	&sub	("esp",$s1);
1179	&add	("esp",4);	# 4 is reserved for caller's return address
1180	&mov	($_esp,$s0);			# save stack pointer
1181
1182	&call   (&label("pic_point"));          # make it PIC!
1183	&set_label("pic_point");
1184	&blindpop($tbl);
1185	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if (!$x86only);
1186	&lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
1187
1188	# pick Te4 copy which can't "overlap" with stack frame or key schedule
1189	&lea	($s1,&DWP(768-4,"esp"));
1190	&sub	($s1,$tbl);
1191	&and	($s1,0x300);
1192	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
1193
1194					if (!$x86only) {
1195	&bt	(&DWP(0,$s0),25);	# check for SSE bit
1196	&jnc	(&label("x86"));
1197
1198	&movq	("mm0",&QWP(0,$acc));
1199	&movq	("mm4",&QWP(8,$acc));
1200	&call	("_sse_AES_encrypt_compact");
1201	&mov	("esp",$_esp);			# restore stack pointer
1202	&mov	($acc,&wparam(1));		# load out
1203	&movq	(&QWP(0,$acc),"mm0");		# write output data
1204	&movq	(&QWP(8,$acc),"mm4");
1205	&emms	();
1206	&function_end_A();
1207					}
1208	&set_label("x86",16);
1209	&mov	($_tbl,$tbl);
1210	&mov	($s0,&DWP(0,$acc));		# load input data
1211	&mov	($s1,&DWP(4,$acc));
1212	&mov	($s2,&DWP(8,$acc));
1213	&mov	($s3,&DWP(12,$acc));
1214	&call	("_x86_AES_encrypt_compact");
1215	&mov	("esp",$_esp);			# restore stack pointer
1216	&mov	($acc,&wparam(1));		# load out
1217	&mov	(&DWP(0,$acc),$s0);		# write output data
1218	&mov	(&DWP(4,$acc),$s1);
1219	&mov	(&DWP(8,$acc),$s2);
1220	&mov	(&DWP(12,$acc),$s3);
1221&function_end("asm_AES_encrypt");
1222
1223#--------------------------------------------------------------------#
1224
1225######################################################################
1226# "Compact" block function
1227######################################################################
1228
1229sub deccompact()
1230{ my $Fn = \&mov;
1231  while ($#_>5) { pop(@_); $Fn=sub{}; }
1232  my ($i,$td,@s)=@_;
1233  my $tmp = $key;
1234  my $out = $i==3?$s[0]:$acc;
1235
1236	# $Fn is used in first compact round and its purpose is to
1237	# void restoration of some values from stack, so that after
1238	# 4xdeccompact with extra argument $key, $s0 and $s1 values
1239	# are left there...
1240	if($i==3)   {	&$Fn	($key,$__key);			}
1241	else        {	&mov	($out,$s[0]);			}
1242			&and	($out,0xFF);
1243			&movz	($out,&BP(-128,$td,$out,1));
1244
1245	if ($i==3)  {	$tmp=$s[1];				}
1246			&movz	($tmp,&HB($s[1]));
1247			&movz	($tmp,&BP(-128,$td,$tmp,1));
1248			&shl	($tmp,8);
1249			&xor	($out,$tmp);
1250
1251	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$acc);		}
1252	else        {	mov	($tmp,$s[2]);			}
1253			&shr	($tmp,16);
1254			&and	($tmp,0xFF);
1255			&movz	($tmp,&BP(-128,$td,$tmp,1));
1256			&shl	($tmp,16);
1257			&xor	($out,$tmp);
1258
1259	if ($i==3)  {	$tmp=$s[3]; &$Fn ($s[2],$__s1);		}
1260	else        {	&mov	($tmp,$s[3]);			}
1261			&shr	($tmp,24);
1262			&movz	($tmp,&BP(-128,$td,$tmp,1));
1263			&shl	($tmp,24);
1264			&xor	($out,$tmp);
1265	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
1266	if ($i==3)  {	&$Fn	($s[3],$__s0);			}
1267}
1268
1269# must be called with 2,3,0,1 as argument sequence!!!
1270sub dectransform()
1271{ my @s = ($s0,$s1,$s2,$s3);
1272  my $i = shift;
1273  my $tmp = $key;
1274  my $tp2 = @s[($i+2)%4]; $tp2 = @s[2] if ($i==1);
1275  my $tp4 = @s[($i+3)%4]; $tp4 = @s[3] if ($i==1);
1276  my $tp8 = $tbl;
1277
1278	&mov	($tmp,0x80808080);
1279	&and	($tmp,$s[$i]);
1280	&mov	($acc,$tmp);
1281	&shr	($tmp,7);
1282	&lea	($tp2,&DWP(0,$s[$i],$s[$i]));
1283	&sub	($acc,$tmp);
1284	&and	($tp2,0xfefefefe);
1285	&and	($acc,0x1b1b1b1b);
1286	&xor	($tp2,$acc);
1287	&mov	($tmp,0x80808080);
1288
1289	&and	($tmp,$tp2);
1290	&mov	($acc,$tmp);
1291	&shr	($tmp,7);
1292	&lea	($tp4,&DWP(0,$tp2,$tp2));
1293	&sub	($acc,$tmp);
1294	&and	($tp4,0xfefefefe);
1295	&and	($acc,0x1b1b1b1b);
1296	 &xor	($tp2,$s[$i]);	# tp2^tp1
1297	&xor	($tp4,$acc);
1298	&mov	($tmp,0x80808080);
1299
1300	&and	($tmp,$tp4);
1301	&mov	($acc,$tmp);
1302	&shr	($tmp,7);
1303	&lea	($tp8,&DWP(0,$tp4,$tp4));
1304	&sub	($acc,$tmp);
1305	&and	($tp8,0xfefefefe);
1306	&and	($acc,0x1b1b1b1b);
1307	 &xor	($tp4,$s[$i]);	# tp4^tp1
1308	 &rotl	($s[$i],8);	# = ROTATE(tp1,8)
1309	&xor	($tp8,$acc);
1310
1311	&xor	($s[$i],$tp2);
1312	&xor	($tp2,$tp8);
1313	&xor	($s[$i],$tp4);
1314	&xor	($tp4,$tp8);
1315	&rotl	($tp2,24);
1316	&xor	($s[$i],$tp8);	# ^= tp8^(tp4^tp1)^(tp2^tp1)
1317	&rotl	($tp4,16);
1318	&xor	($s[$i],$tp2);	# ^= ROTATE(tp8^tp2^tp1,24)
1319	&rotl	($tp8,8);
1320	&xor	($s[$i],$tp4);	# ^= ROTATE(tp8^tp4^tp1,16)
1321	 &mov	($s[0],$__s0)			if($i==2); #prefetch $s0
1322	 &mov	($s[1],$__s1)			if($i==3); #prefetch $s1
1323	 &mov	($s[2],$__s2)			if($i==1);
1324	&xor	($s[$i],$tp8);	# ^= ROTATE(tp8,8)
1325
1326	&mov	($s[3],$__s3)			if($i==1);
1327	&mov	(&DWP(4+4*$i,"esp"),$s[$i])	if($i>=2);
1328}
1329
1330&function_begin_B("_x86_AES_decrypt_compact");
1331	# note that caller is expected to allocate stack frame for me!
1332	&mov	($__key,$key);			# save key
1333
1334	&xor	($s0,&DWP(0,$key));		# xor with key
1335	&xor	($s1,&DWP(4,$key));
1336	&xor	($s2,&DWP(8,$key));
1337	&xor	($s3,&DWP(12,$key));
1338
1339	&mov	($acc,&DWP(240,$key));		# load key->rounds
1340
1341	&lea	($acc,&DWP(-2,$acc,$acc));
1342	&lea	($acc,&DWP(0,$key,$acc,8));
1343	&mov	($__end,$acc);			# end of key schedule
1344
1345	# prefetch Td4
1346	&mov	($key,&DWP(0-128,$tbl));
1347	&mov	($acc,&DWP(32-128,$tbl));
1348	&mov	($key,&DWP(64-128,$tbl));
1349	&mov	($acc,&DWP(96-128,$tbl));
1350	&mov	($key,&DWP(128-128,$tbl));
1351	&mov	($acc,&DWP(160-128,$tbl));
1352	&mov	($key,&DWP(192-128,$tbl));
1353	&mov	($acc,&DWP(224-128,$tbl));
1354
1355	&set_label("loop",16);
1356
1357		&deccompact(0,$tbl,$s0,$s3,$s2,$s1,1);
1358		&deccompact(1,$tbl,$s1,$s0,$s3,$s2,1);
1359		&deccompact(2,$tbl,$s2,$s1,$s0,$s3,1);
1360		&deccompact(3,$tbl,$s3,$s2,$s1,$s0,1);
1361		&dectransform(2);
1362		&dectransform(3);
1363		&dectransform(0);
1364		&dectransform(1);
1365		&mov 	($key,$__key);
1366		&mov	($tbl,$__tbl);
1367		&add	($key,16);		# advance rd_key
1368		&xor	($s0,&DWP(0,$key));
1369		&xor	($s1,&DWP(4,$key));
1370		&xor	($s2,&DWP(8,$key));
1371		&xor	($s3,&DWP(12,$key));
1372
1373	&cmp	($key,$__end);
1374	&mov	($__key,$key);
1375	&jb	(&label("loop"));
1376
1377	&deccompact(0,$tbl,$s0,$s3,$s2,$s1);
1378	&deccompact(1,$tbl,$s1,$s0,$s3,$s2);
1379	&deccompact(2,$tbl,$s2,$s1,$s0,$s3);
1380	&deccompact(3,$tbl,$s3,$s2,$s1,$s0);
1381
1382	&xor	($s0,&DWP(16,$key));
1383	&xor	($s1,&DWP(20,$key));
1384	&xor	($s2,&DWP(24,$key));
1385	&xor	($s3,&DWP(28,$key));
1386
1387	&ret	();
1388&function_end_B("_x86_AES_decrypt_compact");
1389
1390######################################################################
1391# "Compact" SSE block function.
1392######################################################################
1393
1394sub sse_deccompact()
1395{
1396	&pshufw	("mm1","mm0",0x0c);		#  7, 6, 1, 0
1397	&pshufw	("mm5","mm4",0x09);		# 13,12,11,10
1398	&movd	("eax","mm1");			#  7, 6, 1, 0
1399	&movd	("ebx","mm5");			# 13,12,11,10
1400	&mov	($__key,$key);
1401
1402	&movz	($acc,&LB("eax"));		#  0
1403	&movz	("edx",&HB("eax"));		#  1
1404	&pshufw	("mm2","mm0",0x06);		#  3, 2, 5, 4
1405	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
1406	&movz	($key,&LB("ebx"));		# 10
1407	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
1408	&shr	("eax",16);			#  7, 6
1409	&shl	("edx",8);			#  1
1410
1411	&movz	($acc,&BP(-128,$tbl,$key,1));	# 10
1412	&movz	($key,&HB("ebx"));		# 11
1413	&shl	($acc,16);			# 10
1414	&pshufw	("mm6","mm4",0x03);		# 9, 8,15,14
1415	&or	("ecx",$acc);			# 10
1416	&movz	($acc,&BP(-128,$tbl,$key,1));	# 11
1417	&movz	($key,&HB("eax"));		#  7
1418	&shl	($acc,24);			# 11
1419	&shr	("ebx",16);			# 13,12
1420	&or	("edx",$acc);			# 11
1421
1422	&movz	($acc,&BP(-128,$tbl,$key,1));	#  7
1423	&movz	($key,&HB("ebx"));		# 13
1424	&shl	($acc,24);			#  7
1425	&or	("ecx",$acc);			#  7
1426	&movz	($acc,&BP(-128,$tbl,$key,1));	# 13
1427	&movz	($key,&LB("eax"));		#  6
1428	&shl	($acc,8);			# 13
1429	&movd	("eax","mm2");			#  3, 2, 5, 4
1430	&or	("ecx",$acc);			# 13
1431
1432	&movz	($acc,&BP(-128,$tbl,$key,1));	#  6
1433	&movz	($key,&LB("ebx"));		# 12
1434	&shl	($acc,16);			#  6
1435	&movd	("ebx","mm6");			#  9, 8,15,14
1436	&movd	("mm0","ecx");			# t[0] collected
1437	&movz	("ecx",&BP(-128,$tbl,$key,1));	# 12
1438	&movz	($key,&LB("eax"));		#  4
1439	&or	("ecx",$acc);			# 12
1440
1441	&movz	($acc,&BP(-128,$tbl,$key,1));	#  4
1442	&movz	($key,&LB("ebx"));		# 14
1443	&or	("edx",$acc);			#  4
1444	&movz	($acc,&BP(-128,$tbl,$key,1));	# 14
1445	&movz	($key,&HB("eax"));		#  5
1446	&shl	($acc,16);			# 14
1447	&shr	("eax",16);			#  3, 2
1448	&or	("edx",$acc);			# 14
1449
1450	&movz	($acc,&BP(-128,$tbl,$key,1));	#  5
1451	&movz	($key,&HB("ebx"));		# 15
1452	&shr	("ebx",16);			#  9, 8
1453	&shl	($acc,8);			#  5
1454	&movd	("mm1","edx");			# t[1] collected
1455	&movz	("edx",&BP(-128,$tbl,$key,1));	# 15
1456	&movz	($key,&HB("ebx"));		#  9
1457	&shl	("edx",24);			# 15
1458	&and	("ebx",0xff);			#  8
1459	&or	("edx",$acc);			# 15
1460
1461	&punpckldq	("mm0","mm1");		# t[0,1] collected
1462
1463	&movz	($acc,&BP(-128,$tbl,$key,1));	#  9
1464	&movz	($key,&LB("eax"));		#  2
1465	&shl	($acc,8);			#  9
1466	&movz	("eax",&HB("eax"));		#  3
1467	&movz	("ebx",&BP(-128,$tbl,"ebx",1));	#  8
1468	&or	("ecx",$acc);			#  9
1469	&movz	($acc,&BP(-128,$tbl,$key,1));	#  2
1470	&or	("edx","ebx");			#  8
1471	&shl	($acc,16);			#  2
1472	&movz	("eax",&BP(-128,$tbl,"eax",1));	#  3
1473	&or	("edx",$acc);			#  2
1474	&shl	("eax",24);			#  3
1475	&or	("ecx","eax");			#  3
1476	&mov	($key,$__key);
1477	&movd	("mm4","edx");			# t[2] collected
1478	&movd	("mm5","ecx");			# t[3] collected
1479
1480	&punpckldq	("mm4","mm5");		# t[2,3] collected
1481}
1482
1483					if (!$x86only) {
1484&function_begin_B("_sse_AES_decrypt_compact");
1485	&pxor	("mm0",&QWP(0,$key));	#  7, 6, 5, 4, 3, 2, 1, 0
1486	&pxor	("mm4",&QWP(8,$key));	# 15,14,13,12,11,10, 9, 8
1487
1488	# note that caller is expected to allocate stack frame for me!
1489	&mov	($acc,&DWP(240,$key));		# load key->rounds
1490	&lea	($acc,&DWP(-2,$acc,$acc));
1491	&lea	($acc,&DWP(0,$key,$acc,8));
1492	&mov	($__end,$acc);			# end of key schedule
1493
1494	&mov	($s0,0x1b1b1b1b);		# magic constant
1495	&mov	(&DWP(8,"esp"),$s0);
1496	&mov	(&DWP(12,"esp"),$s0);
1497
1498	# prefetch Td4
1499	&mov	($s0,&DWP(0-128,$tbl));
1500	&mov	($s1,&DWP(32-128,$tbl));
1501	&mov	($s2,&DWP(64-128,$tbl));
1502	&mov	($s3,&DWP(96-128,$tbl));
1503	&mov	($s0,&DWP(128-128,$tbl));
1504	&mov	($s1,&DWP(160-128,$tbl));
1505	&mov	($s2,&DWP(192-128,$tbl));
1506	&mov	($s3,&DWP(224-128,$tbl));
1507
1508	&set_label("loop",16);
1509		&sse_deccompact();
1510		&add	($key,16);
1511		&cmp	($key,$__end);
1512		&ja	(&label("out"));
1513
1514		# ROTATE(x^y,N) == ROTATE(x,N)^ROTATE(y,N)
1515		&movq	("mm3","mm0");		&movq	("mm7","mm4");
1516		&movq	("mm2","mm0",1);	&movq	("mm6","mm4",1);
1517		&movq	("mm1","mm0");		&movq	("mm5","mm4");
1518		&pshufw	("mm0","mm0",0xb1);	&pshufw	("mm4","mm4",0xb1);# = ROTATE(tp0,16)
1519		&pslld	("mm2",8);		&pslld	("mm6",8);
1520		&psrld	("mm3",8);		&psrld	("mm7",8);
1521		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp0<<8
1522		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp0>>8
1523		&pslld	("mm2",16);		&pslld	("mm6",16);
1524		&psrld	("mm3",16);		&psrld	("mm7",16);
1525		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp0<<24
1526		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp0>>24
1527
1528		&movq	("mm3",&QWP(8,"esp"));
1529		&pxor	("mm2","mm2");		&pxor	("mm6","mm6");
1530		&pcmpgtb("mm2","mm1");		&pcmpgtb("mm6","mm5");
1531		&pand	("mm2","mm3");		&pand	("mm6","mm3");
1532		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
1533		&pxor	("mm1","mm2");		&pxor	("mm5","mm6");	# tp2
1534		&movq	("mm3","mm1");		&movq	("mm7","mm5");
1535		&movq	("mm2","mm1");		&movq	("mm6","mm5");
1536		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp2
1537		&pslld	("mm3",24);		&pslld	("mm7",24);
1538		&psrld	("mm2",8);		&psrld	("mm6",8);
1539		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp2<<24
1540		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp2>>8
1541
1542		&movq	("mm2",&QWP(8,"esp"));
1543		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
1544		&pcmpgtb("mm3","mm1");		&pcmpgtb("mm7","mm5");
1545		&pand	("mm3","mm2");		&pand	("mm7","mm2");
1546		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
1547		&pxor	("mm1","mm3");		&pxor	("mm5","mm7");	# tp4
1548		&pshufw	("mm3","mm1",0xb1);	&pshufw	("mm7","mm5",0xb1);
1549		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp4
1550		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= ROTATE(tp4,16)
1551
1552		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
1553		&pcmpgtb("mm3","mm1");		&pcmpgtb("mm7","mm5");
1554		&pand	("mm3","mm2");		&pand	("mm7","mm2");
1555		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
1556		&pxor	("mm1","mm3");		&pxor	("mm5","mm7");	# tp8
1557		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8
1558		&movq	("mm3","mm1");		&movq	("mm7","mm5");
1559		&pshufw	("mm2","mm1",0xb1);	&pshufw	("mm6","mm5",0xb1);
1560		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= ROTATE(tp8,16)
1561		&pslld	("mm1",8);		&pslld	("mm5",8);
1562		&psrld	("mm3",8);		&psrld	("mm7",8);
1563		&movq	("mm2",&QWP(0,$key));	&movq	("mm6",&QWP(8,$key));
1564		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8<<8
1565		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp8>>8
1566		&mov	($s0,&DWP(0-128,$tbl));
1567		&pslld	("mm1",16);		&pslld	("mm5",16);
1568		&mov	($s1,&DWP(64-128,$tbl));
1569		&psrld	("mm3",16);		&psrld	("mm7",16);
1570		&mov	($s2,&DWP(128-128,$tbl));
1571		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8<<24
1572		&mov	($s3,&DWP(192-128,$tbl));
1573		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp8>>24
1574
1575		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");
1576	&jmp	(&label("loop"));
1577
1578	&set_label("out",16);
1579	&pxor	("mm0",&QWP(0,$key));
1580	&pxor	("mm4",&QWP(8,$key));
1581
1582	&ret	();
1583&function_end_B("_sse_AES_decrypt_compact");
1584					}
1585
1586######################################################################
1587# Vanilla block function.
1588######################################################################
1589
1590sub decstep()
1591{ my ($i,$td,@s) = @_;
1592  my $tmp = $key;
1593  my $out = $i==3?$s[0]:$acc;
1594
1595	# no instructions are reordered, as performance appears
1596	# optimal... or rather that all attempts to reorder didn't
1597	# result in better performance [which by the way is not a
1598	# bit lower than ecryption].
1599	if($i==3)   {	&mov	($key,$__key);			}
1600	else        {	&mov	($out,$s[0]);			}
1601			&and	($out,0xFF);
1602			&mov	($out,&DWP(0,$td,$out,8));
1603
1604	if ($i==3)  {	$tmp=$s[1];				}
1605			&movz	($tmp,&HB($s[1]));
1606			&xor	($out,&DWP(3,$td,$tmp,8));
1607
1608	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$acc);		}
1609	else        {	&mov	($tmp,$s[2]);			}
1610			&shr	($tmp,16);
1611			&and	($tmp,0xFF);
1612			&xor	($out,&DWP(2,$td,$tmp,8));
1613
1614	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}
1615	else        {	&mov	($tmp,$s[3]);			}
1616			&shr	($tmp,24);
1617			&xor	($out,&DWP(1,$td,$tmp,8));
1618	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
1619	if ($i==3)  {	&mov	($s[3],$__s0);			}
1620			&comment();
1621}
1622
1623sub declast()
1624{ my ($i,$td,@s)=@_;
1625  my $tmp = $key;
1626  my $out = $i==3?$s[0]:$acc;
1627
1628	if($i==0)   {	&lea	($td,&DWP(2048+128,$td));
1629			&mov	($tmp,&DWP(0-128,$td));
1630			&mov	($acc,&DWP(32-128,$td));
1631			&mov	($tmp,&DWP(64-128,$td));
1632			&mov	($acc,&DWP(96-128,$td));
1633			&mov	($tmp,&DWP(128-128,$td));
1634			&mov	($acc,&DWP(160-128,$td));
1635			&mov	($tmp,&DWP(192-128,$td));
1636			&mov	($acc,&DWP(224-128,$td));
1637			&lea	($td,&DWP(-128,$td));		}
1638	if($i==3)   {	&mov	($key,$__key);			}
1639	else        {	&mov	($out,$s[0]);			}
1640			&and	($out,0xFF);
1641			&movz	($out,&BP(0,$td,$out,1));
1642
1643	if ($i==3)  {	$tmp=$s[1];				}
1644			&movz	($tmp,&HB($s[1]));
1645			&movz	($tmp,&BP(0,$td,$tmp,1));
1646			&shl	($tmp,8);
1647			&xor	($out,$tmp);
1648
1649	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$acc);		}
1650	else        {	mov	($tmp,$s[2]);			}
1651			&shr	($tmp,16);
1652			&and	($tmp,0xFF);
1653			&movz	($tmp,&BP(0,$td,$tmp,1));
1654			&shl	($tmp,16);
1655			&xor	($out,$tmp);
1656
1657	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}
1658	else        {	&mov	($tmp,$s[3]);			}
1659			&shr	($tmp,24);
1660			&movz	($tmp,&BP(0,$td,$tmp,1));
1661			&shl	($tmp,24);
1662			&xor	($out,$tmp);
1663	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
1664	if ($i==3)  {	&mov	($s[3],$__s0);
1665			&lea	($td,&DWP(-2048,$td));		}
1666}
1667
1668&function_begin_B("_x86_AES_decrypt");
1669	# note that caller is expected to allocate stack frame for me!
1670	&mov	($__key,$key);			# save key
1671
1672	&xor	($s0,&DWP(0,$key));		# xor with key
1673	&xor	($s1,&DWP(4,$key));
1674	&xor	($s2,&DWP(8,$key));
1675	&xor	($s3,&DWP(12,$key));
1676
1677	&mov	($acc,&DWP(240,$key));		# load key->rounds
1678
1679	if ($small_footprint) {
1680	    &lea	($acc,&DWP(-2,$acc,$acc));
1681	    &lea	($acc,&DWP(0,$key,$acc,8));
1682	    &mov	($__end,$acc);		# end of key schedule
1683	    &set_label("loop",16);
1684		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1685		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1686		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1687		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1688		&add	($key,16);		# advance rd_key
1689		&xor	($s0,&DWP(0,$key));
1690		&xor	($s1,&DWP(4,$key));
1691		&xor	($s2,&DWP(8,$key));
1692		&xor	($s3,&DWP(12,$key));
1693	    &cmp	($key,$__end);
1694	    &mov	($__key,$key);
1695	    &jb		(&label("loop"));
1696	}
1697	else {
1698	    &cmp	($acc,10);
1699	    &jle	(&label("10rounds"));
1700	    &cmp	($acc,12);
1701	    &jle	(&label("12rounds"));
1702
1703	&set_label("14rounds",4);
1704	    for ($i=1;$i<3;$i++) {
1705		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1706		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1707		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1708		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1709		&xor	($s0,&DWP(16*$i+0,$key));
1710		&xor	($s1,&DWP(16*$i+4,$key));
1711		&xor	($s2,&DWP(16*$i+8,$key));
1712		&xor	($s3,&DWP(16*$i+12,$key));
1713	    }
1714	    &add	($key,32);
1715	    &mov	($__key,$key);		# advance rd_key
1716	&set_label("12rounds",4);
1717	    for ($i=1;$i<3;$i++) {
1718		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1719		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1720		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1721		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1722		&xor	($s0,&DWP(16*$i+0,$key));
1723		&xor	($s1,&DWP(16*$i+4,$key));
1724		&xor	($s2,&DWP(16*$i+8,$key));
1725		&xor	($s3,&DWP(16*$i+12,$key));
1726	    }
1727	    &add	($key,32);
1728	    &mov	($__key,$key);		# advance rd_key
1729	&set_label("10rounds",4);
1730	    for ($i=1;$i<10;$i++) {
1731		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
1732		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
1733		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
1734		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
1735		&xor	($s0,&DWP(16*$i+0,$key));
1736		&xor	($s1,&DWP(16*$i+4,$key));
1737		&xor	($s2,&DWP(16*$i+8,$key));
1738		&xor	($s3,&DWP(16*$i+12,$key));
1739	    }
1740	}
1741
1742	&declast(0,$tbl,$s0,$s3,$s2,$s1);
1743	&declast(1,$tbl,$s1,$s0,$s3,$s2);
1744	&declast(2,$tbl,$s2,$s1,$s0,$s3);
1745	&declast(3,$tbl,$s3,$s2,$s1,$s0);
1746
1747	&add	($key,$small_footprint?16:160);
1748	&xor	($s0,&DWP(0,$key));
1749	&xor	($s1,&DWP(4,$key));
1750	&xor	($s2,&DWP(8,$key));
1751	&xor	($s3,&DWP(12,$key));
1752
1753	&ret	();
1754
1755&set_label("AES_Td",64);	# Yes! I keep it in the code segment!
1756	&_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
1757	&_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
1758	&_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
1759	&_data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5);
1760	&_data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d);
1761	&_data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b);
1762	&_data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295);
1763	&_data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e);
1764	&_data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927);
1765	&_data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d);
1766	&_data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362);
1767	&_data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9);
1768	&_data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52);
1769	&_data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566);
1770	&_data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3);
1771	&_data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed);
1772	&_data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e);
1773	&_data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4);
1774	&_data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4);
1775	&_data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd);
1776	&_data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d);
1777	&_data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060);
1778	&_data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967);
1779	&_data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879);
1780	&_data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000);
1781	&_data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c);
1782	&_data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36);
1783	&_data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624);
1784	&_data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b);
1785	&_data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c);
1786	&_data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12);
1787	&_data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14);
1788	&_data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3);
1789	&_data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b);
1790	&_data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8);
1791	&_data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684);
1792	&_data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7);
1793	&_data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177);
1794	&_data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947);
1795	&_data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322);
1796	&_data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498);
1797	&_data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f);
1798	&_data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54);
1799	&_data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382);
1800	&_data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf);
1801	&_data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb);
1802	&_data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83);
1803	&_data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef);
1804	&_data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029);
1805	&_data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235);
1806	&_data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733);
1807	&_data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117);
1808	&_data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4);
1809	&_data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546);
1810	&_data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb);
1811	&_data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d);
1812	&_data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb);
1813	&_data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a);
1814	&_data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773);
1815	&_data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478);
1816	&_data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2);
1817	&_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
1818	&_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
1819	&_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
1820
1821#Td4:	# four copies of Td4 to choose from to avoid L1 aliasing
1822	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1823	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1824	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1825	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1826	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1827	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1828	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1829	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1830	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1831	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1832	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1833	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1834	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1835	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1836	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1837	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1838	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1839	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1840	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1841	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1842	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1843	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1844	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1845	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1846	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1847	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1848	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1849	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1850	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1851	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1852	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1853	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1854
1855	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1856	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1857	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1858	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1859	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1860	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1861	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1862	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1863	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1864	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1865	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1866	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1867	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1868	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1869	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1870	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1871	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1872	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1873	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1874	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1875	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1876	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1877	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1878	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1879	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1880	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1881	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1882	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1883	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1884	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1885	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1886	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1887
1888	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1889	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1890	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1891	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1892	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1893	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1894	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1895	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1896	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1897	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1898	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1899	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1900	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1901	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1902	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1903	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1904	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1905	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1906	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1907	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1908	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1909	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1910	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1911	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1912	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1913	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1914	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1915	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1916	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1917	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1918	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1919	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1920
1921	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1922	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1923	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1924	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1925	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1926	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1927	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1928	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1929	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1930	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1931	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1932	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1933	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1934	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1935	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1936	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1937	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1938	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1939	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1940	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1941	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1942	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1943	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1944	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1945	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1946	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1947	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1948	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1949	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1950	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1951	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1952	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1953&function_end_B("_x86_AES_decrypt");
1954
1955# void asm_AES_decrypt (const void *inp,void *out,const AES_KEY *key);
1956&function_begin("asm_AES_decrypt");
1957	&mov	($acc,&wparam(0));		# load inp
1958	&mov	($key,&wparam(2));		# load key
1959
1960	&mov	($s0,"esp");
1961	&sub	("esp",36);
1962	&and	("esp",-64);			# align to cache-line
1963
1964	# place stack frame just "above" the key schedule
1965	&lea	($s1,&DWP(-64-63,$key));
1966	&sub	($s1,"esp");
1967	&neg	($s1);
1968	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
1969	&sub	("esp",$s1);
1970	&add	("esp",4);	# 4 is reserved for caller's return address
1971	&mov	($_esp,$s0);	# save stack pointer
1972
1973	&call   (&label("pic_point"));          # make it PIC!
1974	&set_label("pic_point");
1975	&blindpop($tbl);
1976	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
1977	&lea    ($tbl,&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl));
1978
1979	# pick Td4 copy which can't "overlap" with stack frame or key schedule
1980	&lea	($s1,&DWP(768-4,"esp"));
1981	&sub	($s1,$tbl);
1982	&and	($s1,0x300);
1983	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
1984
1985					if (!$x86only) {
1986	&bt	(&DWP(0,$s0),25);	# check for SSE bit
1987	&jnc	(&label("x86"));
1988
1989	&movq	("mm0",&QWP(0,$acc));
1990	&movq	("mm4",&QWP(8,$acc));
1991	&call	("_sse_AES_decrypt_compact");
1992	&mov	("esp",$_esp);			# restore stack pointer
1993	&mov	($acc,&wparam(1));		# load out
1994	&movq	(&QWP(0,$acc),"mm0");		# write output data
1995	&movq	(&QWP(8,$acc),"mm4");
1996	&emms	();
1997	&function_end_A();
1998					}
1999	&set_label("x86",16);
2000	&mov	($_tbl,$tbl);
2001	&mov	($s0,&DWP(0,$acc));		# load input data
2002	&mov	($s1,&DWP(4,$acc));
2003	&mov	($s2,&DWP(8,$acc));
2004	&mov	($s3,&DWP(12,$acc));
2005	&call	("_x86_AES_decrypt_compact");
2006	&mov	("esp",$_esp);			# restore stack pointer
2007	&mov	($acc,&wparam(1));		# load out
2008	&mov	(&DWP(0,$acc),$s0);		# write output data
2009	&mov	(&DWP(4,$acc),$s1);
2010	&mov	(&DWP(8,$acc),$s2);
2011	&mov	(&DWP(12,$acc),$s3);
2012&function_end("asm_AES_decrypt");
2013
2014# void asm_AES_cbc_encrypt (const void char *inp, unsigned char *out,
2015#			    size_t length, const AES_KEY *key,
2016#			    unsigned char *ivp,const int enc);
2017{
2018# stack frame layout
2019#             -4(%esp)		# return address	 0(%esp)
2020#              0(%esp)		# s0 backing store	 4(%esp)
2021#              4(%esp)		# s1 backing store	 8(%esp)
2022#              8(%esp)		# s2 backing store	12(%esp)
2023#             12(%esp)		# s3 backing store	16(%esp)
2024#             16(%esp)		# key backup		20(%esp)
2025#             20(%esp)		# end of key schedule	24(%esp)
2026#             24(%esp)		# %ebp backup		28(%esp)
2027#             28(%esp)		# %esp backup
2028my $_inp=&DWP(32,"esp");	# copy of wparam(0)
2029my $_out=&DWP(36,"esp");	# copy of wparam(1)
2030my $_len=&DWP(40,"esp");	# copy of wparam(2)
2031my $_key=&DWP(44,"esp");	# copy of wparam(3)
2032my $_ivp=&DWP(48,"esp");	# copy of wparam(4)
2033my $_tmp=&DWP(52,"esp");	# volatile variable
2034#
2035my $ivec=&DWP(60,"esp");	# ivec[16]
2036my $aes_key=&DWP(76,"esp");	# copy of aes_key
2037my $mark=&DWP(76+240,"esp");	# copy of aes_key->rounds
2038
2039&function_begin("asm_AES_cbc_encrypt");
2040	&mov	($s2 eq "ecx"? $s2 : "",&wparam(2));	# load len
2041	&cmp	($s2,0);
2042	&je	(&label("drop_out"));
2043
2044	&call   (&label("pic_point"));		# make it PIC!
2045	&set_label("pic_point");
2046	&blindpop($tbl);
2047	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
2048
2049	&cmp	(&wparam(5),0);
2050	&lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
2051	&jne	(&label("picked_te"));
2052	&lea	($tbl,&DWP(&label("AES_Td")."-".&label("AES_Te"),$tbl));
2053	&set_label("picked_te");
2054
2055	# one can argue if this is required
2056	&pushf	();
2057	&cld	();
2058
2059	&cmp	($s2,$speed_limit);
2060	&jb	(&label("slow_way"));
2061	&test	($s2,15);
2062	&jnz	(&label("slow_way"));
2063					if (!$x86only) {
2064	&bt	(&DWP(0,$s0),28);	# check for hyper-threading bit
2065	&jc	(&label("slow_way"));
2066					}
2067	# pre-allocate aligned stack frame...
2068	&lea	($acc,&DWP(-80-244,"esp"));
2069	&and	($acc,-64);
2070
2071	# ... and make sure it doesn't alias with $tbl modulo 4096
2072	&mov	($s0,$tbl);
2073	&lea	($s1,&DWP(2048+256,$tbl));
2074	&mov	($s3,$acc);
2075	&and	($s0,0xfff);		# s = %ebp&0xfff
2076	&and	($s1,0xfff);		# e = (%ebp+2048+256)&0xfff
2077	&and	($s3,0xfff);		# p = %esp&0xfff
2078
2079	&cmp	($s3,$s1);		# if (p>=e) %esp =- (p-e);
2080	&jb	(&label("tbl_break_out"));
2081	&sub	($s3,$s1);
2082	&sub	($acc,$s3);
2083	&jmp	(&label("tbl_ok"));
2084	&set_label("tbl_break_out",4);	# else %esp -= (p-s)&0xfff + framesz;
2085	&sub	($s3,$s0);
2086	&and	($s3,0xfff);
2087	&add	($s3,384);
2088	&sub	($acc,$s3);
2089	&set_label("tbl_ok",4);
2090
2091	&lea	($s3,&wparam(0));	# obtain pointer to parameter block
2092	&exch	("esp",$acc);		# allocate stack frame
2093	&add	("esp",4);		# reserve for return address!
2094	&mov	($_tbl,$tbl);		# save %ebp
2095	&mov	($_esp,$acc);		# save %esp
2096
2097	&mov	($s0,&DWP(0,$s3));	# load inp
2098	&mov	($s1,&DWP(4,$s3));	# load out
2099	#&mov	($s2,&DWP(8,$s3));	# load len
2100	&mov	($key,&DWP(12,$s3));	# load key
2101	&mov	($acc,&DWP(16,$s3));	# load ivp
2102	&mov	($s3,&DWP(20,$s3));	# load enc flag
2103
2104	&mov	($_inp,$s0);		# save copy of inp
2105	&mov	($_out,$s1);		# save copy of out
2106	&mov	($_len,$s2);		# save copy of len
2107	&mov	($_key,$key);		# save copy of key
2108	&mov	($_ivp,$acc);		# save copy of ivp
2109
2110	&mov	($mark,0);		# copy of aes_key->rounds = 0;
2111	# do we copy key schedule to stack?
2112	&mov	($s1 eq "ebx" ? $s1 : "",$key);
2113	&mov	($s2 eq "ecx" ? $s2 : "",244/4);
2114	&sub	($s1,$tbl);
2115	&mov	("esi",$key);
2116	&and	($s1,0xfff);
2117	&lea	("edi",$aes_key);
2118	&cmp	($s1,2048+256);
2119	&jb	(&label("do_copy"));
2120	&cmp	($s1,4096-244);
2121	&jb	(&label("skip_copy"));
2122	&set_label("do_copy",4);
2123		&mov	($_key,"edi");
2124		&data_word(0xA5F3F689);	# rep movsd
2125	&set_label("skip_copy");
2126
2127	&mov	($key,16);
2128	&set_label("prefetch_tbl",4);
2129		&mov	($s0,&DWP(0,$tbl));
2130		&mov	($s1,&DWP(32,$tbl));
2131		&mov	($s2,&DWP(64,$tbl));
2132		&mov	($acc,&DWP(96,$tbl));
2133		&lea	($tbl,&DWP(128,$tbl));
2134		&sub	($key,1);
2135	&jnz	(&label("prefetch_tbl"));
2136	&sub	($tbl,2048);
2137
2138	&mov	($acc,$_inp);
2139	&mov	($key,$_ivp);
2140
2141	&cmp	($s3,0);
2142	&je	(&label("fast_decrypt"));
2143
2144#----------------------------- ENCRYPT -----------------------------#
2145	&mov	($s0,&DWP(0,$key));		# load iv
2146	&mov	($s1,&DWP(4,$key));
2147
2148	&set_label("fast_enc_loop",16);
2149		&mov	($s2,&DWP(8,$key));
2150		&mov	($s3,&DWP(12,$key));
2151
2152		&xor	($s0,&DWP(0,$acc));	# xor input data
2153		&xor	($s1,&DWP(4,$acc));
2154		&xor	($s2,&DWP(8,$acc));
2155		&xor	($s3,&DWP(12,$acc));
2156
2157		&mov	($key,$_key);		# load key
2158		&call	("_x86_AES_encrypt");
2159
2160		&mov	($acc,$_inp);		# load inp
2161		&mov	($key,$_out);		# load out
2162
2163		&mov	(&DWP(0,$key),$s0);	# save output data
2164		&mov	(&DWP(4,$key),$s1);
2165		&mov	(&DWP(8,$key),$s2);
2166		&mov	(&DWP(12,$key),$s3);
2167
2168		&lea	($acc,&DWP(16,$acc));	# advance inp
2169		&mov	($s2,$_len);		# load len
2170		&mov	($_inp,$acc);		# save inp
2171		&lea	($s3,&DWP(16,$key));	# advance out
2172		&mov	($_out,$s3);		# save out
2173		&sub	($s2,16);		# decrease len
2174		&mov	($_len,$s2);		# save len
2175	&jnz	(&label("fast_enc_loop"));
2176	&mov	($acc,$_ivp);		# load ivp
2177	&mov	($s2,&DWP(8,$key));	# restore last 2 dwords
2178	&mov	($s3,&DWP(12,$key));
2179	&mov	(&DWP(0,$acc),$s0);	# save ivec
2180	&mov	(&DWP(4,$acc),$s1);
2181	&mov	(&DWP(8,$acc),$s2);
2182	&mov	(&DWP(12,$acc),$s3);
2183
2184	&cmp	($mark,0);		# was the key schedule copied?
2185	&mov	("edi",$_key);
2186	&je	(&label("skip_ezero"));
2187	# zero copy of key schedule
2188	&mov	("ecx",240/4);
2189	&xor	("eax","eax");
2190	&align	(4);
2191	&data_word(0xABF3F689);		# rep stosd
2192	&set_label("skip_ezero");
2193	&mov	("esp",$_esp);
2194	&popf	();
2195    &set_label("drop_out");
2196	&function_end_A();
2197	&pushf	();			# kludge, never executed
2198
2199#----------------------------- DECRYPT -----------------------------#
2200&set_label("fast_decrypt",16);
2201
2202	&cmp	($acc,$_out);
2203	&je	(&label("fast_dec_in_place"));	# in-place processing...
2204
2205	&mov	($_tmp,$key);
2206
2207	&align	(4);
2208	&set_label("fast_dec_loop",16);
2209		&mov	($s0,&DWP(0,$acc));	# read input
2210		&mov	($s1,&DWP(4,$acc));
2211		&mov	($s2,&DWP(8,$acc));
2212		&mov	($s3,&DWP(12,$acc));
2213
2214		&mov	($key,$_key);		# load key
2215		&call	("_x86_AES_decrypt");
2216
2217		&mov	($key,$_tmp);		# load ivp
2218		&mov	($acc,$_len);		# load len
2219		&xor	($s0,&DWP(0,$key));	# xor iv
2220		&xor	($s1,&DWP(4,$key));
2221		&xor	($s2,&DWP(8,$key));
2222		&xor	($s3,&DWP(12,$key));
2223
2224		&mov	($key,$_out);		# load out
2225		&mov	($acc,$_inp);		# load inp
2226
2227		&mov	(&DWP(0,$key),$s0);	# write output
2228		&mov	(&DWP(4,$key),$s1);
2229		&mov	(&DWP(8,$key),$s2);
2230		&mov	(&DWP(12,$key),$s3);
2231
2232		&mov	($s2,$_len);		# load len
2233		&mov	($_tmp,$acc);		# save ivp
2234		&lea	($acc,&DWP(16,$acc));	# advance inp
2235		&mov	($_inp,$acc);		# save inp
2236		&lea	($key,&DWP(16,$key));	# advance out
2237		&mov	($_out,$key);		# save out
2238		&sub	($s2,16);		# decrease len
2239		&mov	($_len,$s2);		# save len
2240	&jnz	(&label("fast_dec_loop"));
2241	&mov	($key,$_tmp);		# load temp ivp
2242	&mov	($acc,$_ivp);		# load user ivp
2243	&mov	($s0,&DWP(0,$key));	# load iv
2244	&mov	($s1,&DWP(4,$key));
2245	&mov	($s2,&DWP(8,$key));
2246	&mov	($s3,&DWP(12,$key));
2247	&mov	(&DWP(0,$acc),$s0);	# copy back to user
2248	&mov	(&DWP(4,$acc),$s1);
2249	&mov	(&DWP(8,$acc),$s2);
2250	&mov	(&DWP(12,$acc),$s3);
2251	&jmp	(&label("fast_dec_out"));
2252
2253    &set_label("fast_dec_in_place",16);
2254	&set_label("fast_dec_in_place_loop");
2255		&mov	($s0,&DWP(0,$acc));	# read input
2256		&mov	($s1,&DWP(4,$acc));
2257		&mov	($s2,&DWP(8,$acc));
2258		&mov	($s3,&DWP(12,$acc));
2259
2260		&lea	($key,$ivec);
2261		&mov	(&DWP(0,$key),$s0);	# copy to temp
2262		&mov	(&DWP(4,$key),$s1);
2263		&mov	(&DWP(8,$key),$s2);
2264		&mov	(&DWP(12,$key),$s3);
2265
2266		&mov	($key,$_key);		# load key
2267		&call	("_x86_AES_decrypt");
2268
2269		&mov	($key,$_ivp);		# load ivp
2270		&mov	($acc,$_out);		# load out
2271		&xor	($s0,&DWP(0,$key));	# xor iv
2272		&xor	($s1,&DWP(4,$key));
2273		&xor	($s2,&DWP(8,$key));
2274		&xor	($s3,&DWP(12,$key));
2275
2276		&mov	(&DWP(0,$acc),$s0);	# write output
2277		&mov	(&DWP(4,$acc),$s1);
2278		&mov	(&DWP(8,$acc),$s2);
2279		&mov	(&DWP(12,$acc),$s3);
2280
2281		&lea	($acc,&DWP(16,$acc));	# advance out
2282		&mov	($_out,$acc);		# save out
2283
2284		&lea	($acc,$ivec);
2285		&mov	($s0,&DWP(0,$acc));	# read temp
2286		&mov	($s1,&DWP(4,$acc));
2287		&mov	($s2,&DWP(8,$acc));
2288		&mov	($s3,&DWP(12,$acc));
2289
2290		&mov	(&DWP(0,$key),$s0);	# copy iv
2291		&mov	(&DWP(4,$key),$s1);
2292		&mov	(&DWP(8,$key),$s2);
2293		&mov	(&DWP(12,$key),$s3);
2294
2295		&mov	($acc,$_inp);		# load inp
2296		&mov	($s2,$_len);		# load len
2297		&lea	($acc,&DWP(16,$acc));	# advance inp
2298		&mov	($_inp,$acc);		# save inp
2299		&sub	($s2,16);		# decrease len
2300		&mov	($_len,$s2);		# save len
2301	&jnz	(&label("fast_dec_in_place_loop"));
2302
2303    &set_label("fast_dec_out",4);
2304	&cmp	($mark,0);		# was the key schedule copied?
2305	&mov	("edi",$_key);
2306	&je	(&label("skip_dzero"));
2307	# zero copy of key schedule
2308	&mov	("ecx",240/4);
2309	&xor	("eax","eax");
2310	&align	(4);
2311	&data_word(0xABF3F689);		# rep stosd
2312	&set_label("skip_dzero");
2313	&mov	("esp",$_esp);
2314	&popf	();
2315	&function_end_A();
2316	&pushf	();			# kludge, never executed
2317
2318#--------------------------- SLOW ROUTINE ---------------------------#
2319&set_label("slow_way",16);
2320
2321	&mov	($s0,&DWP(0,$s0)) if (!$x86only);# load OPENSSL_ia32cap
2322	&mov	($key,&wparam(3));	# load key
2323
2324	# pre-allocate aligned stack frame...
2325	&lea	($acc,&DWP(-80,"esp"));
2326	&and	($acc,-64);
2327
2328	# ... and make sure it doesn't alias with $key modulo 1024
2329	&lea	($s1,&DWP(-80-63,$key));
2330	&sub	($s1,$acc);
2331	&neg	($s1);
2332	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
2333	&sub	($acc,$s1);
2334
2335	# pick S-box copy which can't overlap with stack frame or $key
2336	&lea	($s1,&DWP(768,$acc));
2337	&sub	($s1,$tbl);
2338	&and	($s1,0x300);
2339	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
2340
2341	&lea	($s3,&wparam(0));	# pointer to parameter block
2342
2343	&exch	("esp",$acc);
2344	&add	("esp",4);		# reserve for return address!
2345	&mov	($_tbl,$tbl);		# save %ebp
2346	&mov	($_esp,$acc);		# save %esp
2347	&mov	($_tmp,$s0);		# save OPENSSL_ia32cap
2348
2349	&mov	($s0,&DWP(0,$s3));	# load inp
2350	&mov	($s1,&DWP(4,$s3));	# load out
2351	#&mov	($s2,&DWP(8,$s3));	# load len
2352	#&mov	($key,&DWP(12,$s3));	# load key
2353	&mov	($acc,&DWP(16,$s3));	# load ivp
2354	&mov	($s3,&DWP(20,$s3));	# load enc flag
2355
2356	&mov	($_inp,$s0);		# save copy of inp
2357	&mov	($_out,$s1);		# save copy of out
2358	&mov	($_len,$s2);		# save copy of len
2359	&mov	($_key,$key);		# save copy of key
2360	&mov	($_ivp,$acc);		# save copy of ivp
2361
2362	&mov	($key,$acc);
2363	&mov	($acc,$s0);
2364
2365	&cmp	($s3,0);
2366	&je	(&label("slow_decrypt"));
2367
2368#--------------------------- SLOW ENCRYPT ---------------------------#
2369	&cmp	($s2,16);
2370	&mov	($s3,$s1);
2371	&jb	(&label("slow_enc_tail"));
2372
2373					if (!$x86only) {
2374	&bt	($_tmp,25);		# check for SSE bit
2375	&jnc	(&label("slow_enc_x86"));
2376
2377	&movq	("mm0",&QWP(0,$key));	# load iv
2378	&movq	("mm4",&QWP(8,$key));
2379
2380	&set_label("slow_enc_loop_sse",16);
2381		&pxor	("mm0",&QWP(0,$acc));	# xor input data
2382		&pxor	("mm4",&QWP(8,$acc));
2383
2384		&mov	($key,$_key);
2385		&call	("_sse_AES_encrypt_compact");
2386
2387		&mov	($acc,$_inp);		# load inp
2388		&mov	($key,$_out);		# load out
2389		&mov	($s2,$_len);		# load len
2390
2391		&movq	(&QWP(0,$key),"mm0");	# save output data
2392		&movq	(&QWP(8,$key),"mm4");
2393
2394		&lea	($acc,&DWP(16,$acc));	# advance inp
2395		&mov	($_inp,$acc);		# save inp
2396		&lea	($s3,&DWP(16,$key));	# advance out
2397		&mov	($_out,$s3);		# save out
2398		&sub	($s2,16);		# decrease len
2399		&cmp	($s2,16);
2400		&mov	($_len,$s2);		# save len
2401	&jae	(&label("slow_enc_loop_sse"));
2402	&test	($s2,15);
2403	&jnz	(&label("slow_enc_tail"));
2404	&mov	($acc,$_ivp);		# load ivp
2405	&movq	(&QWP(0,$acc),"mm0");	# save ivec
2406	&movq	(&QWP(8,$acc),"mm4");
2407	&emms	();
2408	&mov	("esp",$_esp);
2409	&popf	();
2410	&function_end_A();
2411	&pushf	();			# kludge, never executed
2412					}
2413    &set_label("slow_enc_x86",16);
2414	&mov	($s0,&DWP(0,$key));	# load iv
2415	&mov	($s1,&DWP(4,$key));
2416
2417	&set_label("slow_enc_loop_x86",4);
2418		&mov	($s2,&DWP(8,$key));
2419		&mov	($s3,&DWP(12,$key));
2420
2421		&xor	($s0,&DWP(0,$acc));	# xor input data
2422		&xor	($s1,&DWP(4,$acc));
2423		&xor	($s2,&DWP(8,$acc));
2424		&xor	($s3,&DWP(12,$acc));
2425
2426		&mov	($key,$_key);		# load key
2427		&call	("_x86_AES_encrypt_compact");
2428
2429		&mov	($acc,$_inp);		# load inp
2430		&mov	($key,$_out);		# load out
2431
2432		&mov	(&DWP(0,$key),$s0);	# save output data
2433		&mov	(&DWP(4,$key),$s1);
2434		&mov	(&DWP(8,$key),$s2);
2435		&mov	(&DWP(12,$key),$s3);
2436
2437		&mov	($s2,$_len);		# load len
2438		&lea	($acc,&DWP(16,$acc));	# advance inp
2439		&mov	($_inp,$acc);		# save inp
2440		&lea	($s3,&DWP(16,$key));	# advance out
2441		&mov	($_out,$s3);		# save out
2442		&sub	($s2,16);		# decrease len
2443		&cmp	($s2,16);
2444		&mov	($_len,$s2);		# save len
2445	&jae	(&label("slow_enc_loop_x86"));
2446	&test	($s2,15);
2447	&jnz	(&label("slow_enc_tail"));
2448	&mov	($acc,$_ivp);		# load ivp
2449	&mov	($s2,&DWP(8,$key));	# restore last dwords
2450	&mov	($s3,&DWP(12,$key));
2451	&mov	(&DWP(0,$acc),$s0);	# save ivec
2452	&mov	(&DWP(4,$acc),$s1);
2453	&mov	(&DWP(8,$acc),$s2);
2454	&mov	(&DWP(12,$acc),$s3);
2455
2456	&mov	("esp",$_esp);
2457	&popf	();
2458	&function_end_A();
2459	&pushf	();			# kludge, never executed
2460
2461    &set_label("slow_enc_tail",16);
2462	&emms	()	if (!$x86only);
2463	&mov	($key eq "edi"? $key:"",$s3);	# load out to edi
2464	&mov	($s1,16);
2465	&sub	($s1,$s2);
2466	&cmp	($key,$acc eq "esi"? $acc:"");	# compare with inp
2467	&je	(&label("enc_in_place"));
2468	&align	(4);
2469	&data_word(0xA4F3F689);	# rep movsb	# copy input
2470	&jmp	(&label("enc_skip_in_place"));
2471    &set_label("enc_in_place");
2472	&lea	($key,&DWP(0,$key,$s2));
2473    &set_label("enc_skip_in_place");
2474	&mov	($s2,$s1);
2475	&xor	($s0,$s0);
2476	&align	(4);
2477	&data_word(0xAAF3F689);	# rep stosb	# zero tail
2478
2479	&mov	($key,$_ivp);			# restore ivp
2480	&mov	($acc,$s3);			# output as input
2481	&mov	($s0,&DWP(0,$key));
2482	&mov	($s1,&DWP(4,$key));
2483	&mov	($_len,16);			# len=16
2484	&jmp	(&label("slow_enc_loop_x86"));	# one more spin...
2485
2486#--------------------------- SLOW DECRYPT ---------------------------#
2487&set_label("slow_decrypt",16);
2488					if (!$x86only) {
2489	&bt	($_tmp,25);		# check for SSE bit
2490	&jnc	(&label("slow_dec_loop_x86"));
2491
2492	&set_label("slow_dec_loop_sse",4);
2493		&movq	("mm0",&QWP(0,$acc));	# read input
2494		&movq	("mm4",&QWP(8,$acc));
2495
2496		&mov	($key,$_key);
2497		&call	("_sse_AES_decrypt_compact");
2498
2499		&mov	($acc,$_inp);		# load inp
2500		&lea	($s0,$ivec);
2501		&mov	($s1,$_out);		# load out
2502		&mov	($s2,$_len);		# load len
2503		&mov	($key,$_ivp);		# load ivp
2504
2505		&movq	("mm1",&QWP(0,$acc));	# re-read input
2506		&movq	("mm5",&QWP(8,$acc));
2507
2508		&pxor	("mm0",&QWP(0,$key));	# xor iv
2509		&pxor	("mm4",&QWP(8,$key));
2510
2511		&movq	(&QWP(0,$key),"mm1");	# copy input to iv
2512		&movq	(&QWP(8,$key),"mm5");
2513
2514		&sub	($s2,16);		# decrease len
2515		&jc	(&label("slow_dec_partial_sse"));
2516
2517		&movq	(&QWP(0,$s1),"mm0");	# write output
2518		&movq	(&QWP(8,$s1),"mm4");
2519
2520		&lea	($s1,&DWP(16,$s1));	# advance out
2521		&mov	($_out,$s1);		# save out
2522		&lea	($acc,&DWP(16,$acc));	# advance inp
2523		&mov	($_inp,$acc);		# save inp
2524		&mov	($_len,$s2);		# save len
2525	&jnz	(&label("slow_dec_loop_sse"));
2526	&emms	();
2527	&mov	("esp",$_esp);
2528	&popf	();
2529	&function_end_A();
2530	&pushf	();			# kludge, never executed
2531
2532    &set_label("slow_dec_partial_sse",16);
2533	&movq	(&QWP(0,$s0),"mm0");	# save output to temp
2534	&movq	(&QWP(8,$s0),"mm4");
2535	&emms	();
2536
2537	&add	($s2 eq "ecx" ? "ecx":"",16);
2538	&mov	("edi",$s1);		# out
2539	&mov	("esi",$s0);		# temp
2540	&align	(4);
2541	&data_word(0xA4F3F689);		# rep movsb # copy partial output
2542
2543	&mov	("esp",$_esp);
2544	&popf	();
2545	&function_end_A();
2546	&pushf	();			# kludge, never executed
2547					}
2548	&set_label("slow_dec_loop_x86",16);
2549		&mov	($s0,&DWP(0,$acc));	# read input
2550		&mov	($s1,&DWP(4,$acc));
2551		&mov	($s2,&DWP(8,$acc));
2552		&mov	($s3,&DWP(12,$acc));
2553
2554		&lea	($key,$ivec);
2555		&mov	(&DWP(0,$key),$s0);	# copy to temp
2556		&mov	(&DWP(4,$key),$s1);
2557		&mov	(&DWP(8,$key),$s2);
2558		&mov	(&DWP(12,$key),$s3);
2559
2560		&mov	($key,$_key);		# load key
2561		&call	("_x86_AES_decrypt_compact");
2562
2563		&mov	($key,$_ivp);		# load ivp
2564		&mov	($acc,$_len);		# load len
2565		&xor	($s0,&DWP(0,$key));	# xor iv
2566		&xor	($s1,&DWP(4,$key));
2567		&xor	($s2,&DWP(8,$key));
2568		&xor	($s3,&DWP(12,$key));
2569
2570		&sub	($acc,16);
2571		&jc	(&label("slow_dec_partial_x86"));
2572
2573		&mov	($_len,$acc);		# save len
2574		&mov	($acc,$_out);		# load out
2575
2576		&mov	(&DWP(0,$acc),$s0);	# write output
2577		&mov	(&DWP(4,$acc),$s1);
2578		&mov	(&DWP(8,$acc),$s2);
2579		&mov	(&DWP(12,$acc),$s3);
2580
2581		&lea	($acc,&DWP(16,$acc));	# advance out
2582		&mov	($_out,$acc);		# save out
2583
2584		&lea	($acc,$ivec);
2585		&mov	($s0,&DWP(0,$acc));	# read temp
2586		&mov	($s1,&DWP(4,$acc));
2587		&mov	($s2,&DWP(8,$acc));
2588		&mov	($s3,&DWP(12,$acc));
2589
2590		&mov	(&DWP(0,$key),$s0);	# copy it to iv
2591		&mov	(&DWP(4,$key),$s1);
2592		&mov	(&DWP(8,$key),$s2);
2593		&mov	(&DWP(12,$key),$s3);
2594
2595		&mov	($acc,$_inp);		# load inp
2596		&lea	($acc,&DWP(16,$acc));	# advance inp
2597		&mov	($_inp,$acc);		# save inp
2598	&jnz	(&label("slow_dec_loop_x86"));
2599	&mov	("esp",$_esp);
2600	&popf	();
2601	&function_end_A();
2602	&pushf	();			# kludge, never executed
2603
2604    &set_label("slow_dec_partial_x86",16);
2605	&lea	($acc,$ivec);
2606	&mov	(&DWP(0,$acc),$s0);	# save output to temp
2607	&mov	(&DWP(4,$acc),$s1);
2608	&mov	(&DWP(8,$acc),$s2);
2609	&mov	(&DWP(12,$acc),$s3);
2610
2611	&mov	($acc,$_inp);
2612	&mov	($s0,&DWP(0,$acc));	# re-read input
2613	&mov	($s1,&DWP(4,$acc));
2614	&mov	($s2,&DWP(8,$acc));
2615	&mov	($s3,&DWP(12,$acc));
2616
2617	&mov	(&DWP(0,$key),$s0);	# copy it to iv
2618	&mov	(&DWP(4,$key),$s1);
2619	&mov	(&DWP(8,$key),$s2);
2620	&mov	(&DWP(12,$key),$s3);
2621
2622	&mov	("ecx",$_len);
2623	&mov	("edi",$_out);
2624	&lea	("esi",$ivec);
2625	&align	(4);
2626	&data_word(0xA4F3F689);		# rep movsb # copy partial output
2627
2628	&mov	("esp",$_esp);
2629	&popf	();
2630&function_end("asm_AES_cbc_encrypt");
2631}
2632
2633#------------------------------------------------------------------#
2634
2635sub enckey()
2636{
2637	&movz	("esi",&LB("edx"));		# rk[i]>>0
2638	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2639	&movz	("esi",&HB("edx"));		# rk[i]>>8
2640	&shl	("ebx",24);
2641	&xor	("eax","ebx");
2642
2643	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2644	&shr	("edx",16);
2645	&movz	("esi",&LB("edx"));		# rk[i]>>16
2646	&xor	("eax","ebx");
2647
2648	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2649	&movz	("esi",&HB("edx"));		# rk[i]>>24
2650	&shl	("ebx",8);
2651	&xor	("eax","ebx");
2652
2653	&movz	("ebx",&BP(-128,$tbl,"esi",1));
2654	&shl	("ebx",16);
2655	&xor	("eax","ebx");
2656
2657	&xor	("eax",&DWP(1024-128,$tbl,"ecx",4));	# rcon
2658}
2659
2660&function_begin("_x86_AES_set_encrypt_key");
2661	&mov	("esi",&wparam(1));		# user supplied key
2662	&mov	("edi",&wparam(3));		# private key schedule
2663
2664	&test	("esi",-1);
2665	&jz	(&label("badpointer"));
2666	&test	("edi",-1);
2667	&jz	(&label("badpointer"));
2668
2669	&call	(&label("pic_point"));
2670	&set_label("pic_point");
2671	&blindpop($tbl);
2672	&lea	($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
2673	&lea	($tbl,&DWP(2048+128,$tbl));
2674
2675	# prefetch Te4
2676	&mov	("eax",&DWP(0-128,$tbl));
2677	&mov	("ebx",&DWP(32-128,$tbl));
2678	&mov	("ecx",&DWP(64-128,$tbl));
2679	&mov	("edx",&DWP(96-128,$tbl));
2680	&mov	("eax",&DWP(128-128,$tbl));
2681	&mov	("ebx",&DWP(160-128,$tbl));
2682	&mov	("ecx",&DWP(192-128,$tbl));
2683	&mov	("edx",&DWP(224-128,$tbl));
2684
2685	&mov	("ecx",&wparam(2));		# number of bits in key
2686	&cmp	("ecx",128);
2687	&je	(&label("10rounds"));
2688	&cmp	("ecx",192);
2689	&je	(&label("12rounds"));
2690	&cmp	("ecx",256);
2691	&je	(&label("14rounds"));
2692	&mov	("eax",-2);			# invalid number of bits
2693	&jmp	(&label("exit"));
2694
2695    &set_label("10rounds");
2696	&mov	("eax",&DWP(0,"esi"));		# copy first 4 dwords
2697	&mov	("ebx",&DWP(4,"esi"));
2698	&mov	("ecx",&DWP(8,"esi"));
2699	&mov	("edx",&DWP(12,"esi"));
2700	&mov	(&DWP(0,"edi"),"eax");
2701	&mov	(&DWP(4,"edi"),"ebx");
2702	&mov	(&DWP(8,"edi"),"ecx");
2703	&mov	(&DWP(12,"edi"),"edx");
2704
2705	&xor	("ecx","ecx");
2706	&jmp	(&label("10shortcut"));
2707
2708	&align	(4);
2709	&set_label("10loop");
2710		&mov	("eax",&DWP(0,"edi"));		# rk[0]
2711		&mov	("edx",&DWP(12,"edi"));		# rk[3]
2712	&set_label("10shortcut");
2713		&enckey	();
2714
2715		&mov	(&DWP(16,"edi"),"eax");		# rk[4]
2716		&xor	("eax",&DWP(4,"edi"));
2717		&mov	(&DWP(20,"edi"),"eax");		# rk[5]
2718		&xor	("eax",&DWP(8,"edi"));
2719		&mov	(&DWP(24,"edi"),"eax");		# rk[6]
2720		&xor	("eax",&DWP(12,"edi"));
2721		&mov	(&DWP(28,"edi"),"eax");		# rk[7]
2722		&inc	("ecx");
2723		&add	("edi",16);
2724		&cmp	("ecx",10);
2725	&jl	(&label("10loop"));
2726
2727	&mov	(&DWP(80,"edi"),10);		# setup number of rounds
2728	&xor	("eax","eax");
2729	&jmp	(&label("exit"));
2730
2731    &set_label("12rounds");
2732	&mov	("eax",&DWP(0,"esi"));		# copy first 6 dwords
2733	&mov	("ebx",&DWP(4,"esi"));
2734	&mov	("ecx",&DWP(8,"esi"));
2735	&mov	("edx",&DWP(12,"esi"));
2736	&mov	(&DWP(0,"edi"),"eax");
2737	&mov	(&DWP(4,"edi"),"ebx");
2738	&mov	(&DWP(8,"edi"),"ecx");
2739	&mov	(&DWP(12,"edi"),"edx");
2740	&mov	("ecx",&DWP(16,"esi"));
2741	&mov	("edx",&DWP(20,"esi"));
2742	&mov	(&DWP(16,"edi"),"ecx");
2743	&mov	(&DWP(20,"edi"),"edx");
2744
2745	&xor	("ecx","ecx");
2746	&jmp	(&label("12shortcut"));
2747
2748	&align	(4);
2749	&set_label("12loop");
2750		&mov	("eax",&DWP(0,"edi"));		# rk[0]
2751		&mov	("edx",&DWP(20,"edi"));		# rk[5]
2752	&set_label("12shortcut");
2753		&enckey	();
2754
2755		&mov	(&DWP(24,"edi"),"eax");		# rk[6]
2756		&xor	("eax",&DWP(4,"edi"));
2757		&mov	(&DWP(28,"edi"),"eax");		# rk[7]
2758		&xor	("eax",&DWP(8,"edi"));
2759		&mov	(&DWP(32,"edi"),"eax");		# rk[8]
2760		&xor	("eax",&DWP(12,"edi"));
2761		&mov	(&DWP(36,"edi"),"eax");		# rk[9]
2762
2763		&cmp	("ecx",7);
2764		&je	(&label("12break"));
2765		&inc	("ecx");
2766
2767		&xor	("eax",&DWP(16,"edi"));
2768		&mov	(&DWP(40,"edi"),"eax");		# rk[10]
2769		&xor	("eax",&DWP(20,"edi"));
2770		&mov	(&DWP(44,"edi"),"eax");		# rk[11]
2771
2772		&add	("edi",24);
2773	&jmp	(&label("12loop"));
2774
2775	&set_label("12break");
2776	&mov	(&DWP(72,"edi"),12);		# setup number of rounds
2777	&xor	("eax","eax");
2778	&jmp	(&label("exit"));
2779
2780    &set_label("14rounds");
2781	&mov	("eax",&DWP(0,"esi"));		# copy first 8 dwords
2782	&mov	("ebx",&DWP(4,"esi"));
2783	&mov	("ecx",&DWP(8,"esi"));
2784	&mov	("edx",&DWP(12,"esi"));
2785	&mov	(&DWP(0,"edi"),"eax");
2786	&mov	(&DWP(4,"edi"),"ebx");
2787	&mov	(&DWP(8,"edi"),"ecx");
2788	&mov	(&DWP(12,"edi"),"edx");
2789	&mov	("eax",&DWP(16,"esi"));
2790	&mov	("ebx",&DWP(20,"esi"));
2791	&mov	("ecx",&DWP(24,"esi"));
2792	&mov	("edx",&DWP(28,"esi"));
2793	&mov	(&DWP(16,"edi"),"eax");
2794	&mov	(&DWP(20,"edi"),"ebx");
2795	&mov	(&DWP(24,"edi"),"ecx");
2796	&mov	(&DWP(28,"edi"),"edx");
2797
2798	&xor	("ecx","ecx");
2799	&jmp	(&label("14shortcut"));
2800
2801	&align	(4);
2802	&set_label("14loop");
2803		&mov	("edx",&DWP(28,"edi"));		# rk[7]
2804	&set_label("14shortcut");
2805		&mov	("eax",&DWP(0,"edi"));		# rk[0]
2806
2807		&enckey	();
2808
2809		&mov	(&DWP(32,"edi"),"eax");		# rk[8]
2810		&xor	("eax",&DWP(4,"edi"));
2811		&mov	(&DWP(36,"edi"),"eax");		# rk[9]
2812		&xor	("eax",&DWP(8,"edi"));
2813		&mov	(&DWP(40,"edi"),"eax");		# rk[10]
2814		&xor	("eax",&DWP(12,"edi"));
2815		&mov	(&DWP(44,"edi"),"eax");		# rk[11]
2816
2817		&cmp	("ecx",6);
2818		&je	(&label("14break"));
2819		&inc	("ecx");
2820
2821		&mov	("edx","eax");
2822		&mov	("eax",&DWP(16,"edi"));		# rk[4]
2823		&movz	("esi",&LB("edx"));		# rk[11]>>0
2824		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2825		&movz	("esi",&HB("edx"));		# rk[11]>>8
2826		&xor	("eax","ebx");
2827
2828		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2829		&shr	("edx",16);
2830		&shl	("ebx",8);
2831		&movz	("esi",&LB("edx"));		# rk[11]>>16
2832		&xor	("eax","ebx");
2833
2834		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2835		&movz	("esi",&HB("edx"));		# rk[11]>>24
2836		&shl	("ebx",16);
2837		&xor	("eax","ebx");
2838
2839		&movz	("ebx",&BP(-128,$tbl,"esi",1));
2840		&shl	("ebx",24);
2841		&xor	("eax","ebx");
2842
2843		&mov	(&DWP(48,"edi"),"eax");		# rk[12]
2844		&xor	("eax",&DWP(20,"edi"));
2845		&mov	(&DWP(52,"edi"),"eax");		# rk[13]
2846		&xor	("eax",&DWP(24,"edi"));
2847		&mov	(&DWP(56,"edi"),"eax");		# rk[14]
2848		&xor	("eax",&DWP(28,"edi"));
2849		&mov	(&DWP(60,"edi"),"eax");		# rk[15]
2850
2851		&add	("edi",32);
2852	&jmp	(&label("14loop"));
2853
2854	&set_label("14break");
2855	&mov	(&DWP(48,"edi"),14);		# setup number of rounds
2856	&xor	("eax","eax");
2857	&jmp	(&label("exit"));
2858
2859    &set_label("badpointer");
2860	&mov	("eax",-1);
2861    &set_label("exit");
2862&function_end("_x86_AES_set_encrypt_key");
2863
2864# int asm_AES_set_encrypt_key(const unsigned char *userKey, const int bits,
2865#                             AES_KEY *key)
2866&function_begin_B("asm_AES_set_encrypt_key");
2867	&call	("_x86_AES_set_encrypt_key");
2868	&ret	();
2869&function_end_B("asm_AES_set_encrypt_key");
2870
2871sub deckey()
2872{ my ($i,$key,$tp1,$tp2,$tp4,$tp8) = @_;
2873  my $tmp = $tbl;
2874
2875	&mov	($tmp,0x80808080);
2876	&and	($tmp,$tp1);
2877	&lea	($tp2,&DWP(0,$tp1,$tp1));
2878	&mov	($acc,$tmp);
2879	&shr	($tmp,7);
2880	&sub	($acc,$tmp);
2881	&and	($tp2,0xfefefefe);
2882	&and	($acc,0x1b1b1b1b);
2883	&xor	($tp2,$acc);
2884	&mov	($tmp,0x80808080);
2885
2886	&and	($tmp,$tp2);
2887	&lea	($tp4,&DWP(0,$tp2,$tp2));
2888	&mov	($acc,$tmp);
2889	&shr	($tmp,7);
2890	&sub	($acc,$tmp);
2891	&and	($tp4,0xfefefefe);
2892	&and	($acc,0x1b1b1b1b);
2893	 &xor	($tp2,$tp1);	# tp2^tp1
2894	&xor	($tp4,$acc);
2895	&mov	($tmp,0x80808080);
2896
2897	&and	($tmp,$tp4);
2898	&lea	($tp8,&DWP(0,$tp4,$tp4));
2899	&mov	($acc,$tmp);
2900	&shr	($tmp,7);
2901	 &xor	($tp4,$tp1);	# tp4^tp1
2902	&sub	($acc,$tmp);
2903	&and	($tp8,0xfefefefe);
2904	&and	($acc,0x1b1b1b1b);
2905	 &rotl	($tp1,8);	# = ROTATE(tp1,8)
2906	&xor	($tp8,$acc);
2907
2908	&mov	($tmp,&DWP(4*($i+1),$key));	# modulo-scheduled load
2909
2910	&xor	($tp1,$tp2);
2911	&xor	($tp2,$tp8);
2912	&xor	($tp1,$tp4);
2913	&rotl	($tp2,24);
2914	&xor	($tp4,$tp8);
2915	&xor	($tp1,$tp8);	# ^= tp8^(tp4^tp1)^(tp2^tp1)
2916	&rotl	($tp4,16);
2917	&xor	($tp1,$tp2);	# ^= ROTATE(tp8^tp2^tp1,24)
2918	&rotl	($tp8,8);
2919	&xor	($tp1,$tp4);	# ^= ROTATE(tp8^tp4^tp1,16)
2920	&mov	($tp2,$tmp);
2921	&xor	($tp1,$tp8);	# ^= ROTATE(tp8,8)
2922
2923	&mov	(&DWP(4*$i,$key),$tp1);
2924}
2925
2926# int asm_AES_set_decrypt_key(const unsigned char *userKey, const int bits,
2927#                             AES_KEY *key)
2928&function_begin_B("asm_AES_set_decrypt_key");
2929	&call	("_x86_AES_set_encrypt_key");
2930	&cmp	("eax",0);
2931	&je	(&label("proceed"));
2932	&ret	();
2933
2934    &set_label("proceed");
2935	&push	("ebp");
2936	&push	("ebx");
2937	&push	("esi");
2938	&push	("edi");
2939
2940	&mov	("esi",&wparam(2));
2941	&mov	("ecx",&DWP(240,"esi"));	# pull number of rounds
2942	&lea	("ecx",&DWP(0,"","ecx",4));
2943	&lea	("edi",&DWP(0,"esi","ecx",4));	# pointer to last chunk
2944
2945	&set_label("invert",4);			# invert order of chunks
2946		&mov	("eax",&DWP(0,"esi"));
2947		&mov	("ebx",&DWP(4,"esi"));
2948		&mov	("ecx",&DWP(0,"edi"));
2949		&mov	("edx",&DWP(4,"edi"));
2950		&mov	(&DWP(0,"edi"),"eax");
2951		&mov	(&DWP(4,"edi"),"ebx");
2952		&mov	(&DWP(0,"esi"),"ecx");
2953		&mov	(&DWP(4,"esi"),"edx");
2954		&mov	("eax",&DWP(8,"esi"));
2955		&mov	("ebx",&DWP(12,"esi"));
2956		&mov	("ecx",&DWP(8,"edi"));
2957		&mov	("edx",&DWP(12,"edi"));
2958		&mov	(&DWP(8,"edi"),"eax");
2959		&mov	(&DWP(12,"edi"),"ebx");
2960		&mov	(&DWP(8,"esi"),"ecx");
2961		&mov	(&DWP(12,"esi"),"edx");
2962		&add	("esi",16);
2963		&sub	("edi",16);
2964		&cmp	("esi","edi");
2965	&jne	(&label("invert"));
2966
2967	&mov	($key,&wparam(2));
2968	&mov	($acc,&DWP(240,$key));		# pull number of rounds
2969	&lea	($acc,&DWP(-2,$acc,$acc));
2970	&lea	($acc,&DWP(0,$key,$acc,8));
2971	&mov	(&wparam(2),$acc);
2972
2973	&mov	($s0,&DWP(16,$key));		# modulo-scheduled load
2974	&set_label("permute",4);		# permute the key schedule
2975		&add	($key,16);
2976		&deckey	(0,$key,$s0,$s1,$s2,$s3);
2977		&deckey	(1,$key,$s1,$s2,$s3,$s0);
2978		&deckey	(2,$key,$s2,$s3,$s0,$s1);
2979		&deckey	(3,$key,$s3,$s0,$s1,$s2);
2980		&cmp	($key,&wparam(2));
2981	&jb	(&label("permute"));
2982
2983	&xor	("eax","eax");			# return success
2984&function_end("asm_AES_set_decrypt_key");
2985&asciz("AES for x86, CRYPTOGAMS by <appro\@openssl.org>");
2986
2987&asm_finish();
2988