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