Improve 386 portability of aes-586.pl.
[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. Rights for redistribution and usage in source and binary
6 # forms are granted according to the OpenSSL license.
7 # ====================================================================
8 #
9 # Version 4.3.
10 #
11 # You might fail to appreciate this module performance from the first
12 # try. If compared to "vanilla" linux-ia32-icc target, i.e. considered
13 # to be *the* best Intel C compiler without -KPIC, performance appears
14 # to be virtually identical... But try to re-configure with shared
15 # library support... Aha! Intel compiler "suddenly" lags behind by 30%
16 # [on P4, more on others]:-) And if compared to position-independent
17 # code generated by GNU C, this code performs *more* than *twice* as
18 # fast! Yes, all this buzz about PIC means that unlike other hand-
19 # coded implementations, this one was explicitly designed to be safe
20 # to use even in shared library context... This also means that this
21 # code isn't necessarily absolutely fastest "ever," because in order
22 # to achieve position independence an extra register has to be
23 # off-loaded to stack, which affects the benchmark result.
24 #
25 # Special note about instruction choice. Do you recall RC4_INT code
26 # performing poorly on P4? It might be the time to figure out why.
27 # RC4_INT code implies effective address calculations in base+offset*4
28 # form. Trouble is that it seems that offset scaling turned to be
29 # critical path... At least eliminating scaling resulted in 2.8x RC4
30 # performance improvement [as you might recall]. As AES code is hungry
31 # for scaling too, I [try to] avoid the latter by favoring off-by-2
32 # shifts and masking the result with 0xFF<<2 instead of "boring" 0xFF.
33 #
34 # As was shown by Dean Gaudet <dean@arctic.org>, the above note turned
35 # void. Performance improvement with off-by-2 shifts was observed on
36 # intermediate implementation, which was spilling yet another register
37 # to stack... Final offset*4 code below runs just a tad faster on P4,
38 # but exhibits up to 10% improvement on other cores.
39 #
40 # Second version is "monolithic" replacement for aes_core.c, which in
41 # addition to AES_[de|en]crypt implements AES_set_[de|en]cryption_key.
42 # This made it possible to implement little-endian variant of the
43 # algorithm without modifying the base C code. Motivating factor for
44 # the undertaken effort was that it appeared that in tight IA-32
45 # register window little-endian flavor could achieve slightly higher
46 # Instruction Level Parallelism, and it indeed resulted in up to 15%
47 # better performance on most recent µ-archs...
48 #
49 # Third version adds AES_cbc_encrypt implementation, which resulted in
50 # up to 40% performance imrovement of CBC benchmark results. 40% was
51 # observed on P4 core, where "overall" imrovement coefficient, i.e. if
52 # compared to PIC generated by GCC and in CBC mode, was observed to be
53 # as large as 4x:-) CBC performance is virtually identical to ECB now
54 # and on some platforms even better, e.g. 17.6 "small" cycles/byte on
55 # Opteron, because certain function prologues and epilogues are
56 # effectively taken out of the loop...
57 #
58 # Version 3.2 implements compressed tables and prefetch of these tables
59 # in CBC[!] mode. Former means that 3/4 of table references are now
60 # misaligned, which unfortunately has negative impact on elder IA-32
61 # implementations, Pentium suffered 30% penalty, PIII - 10%.
62 #
63 # Version 3.3 avoids L1 cache aliasing between stack frame and
64 # S-boxes, and 3.4 - L1 cache aliasing even between key schedule. The
65 # latter is achieved by copying the key schedule to controlled place in
66 # stack. This unfortunately has rather strong impact on small block CBC
67 # performance, ~2x deterioration on 16-byte block if compared to 3.3.
68 #
69 # Version 3.5 checks if there is L1 cache aliasing between user-supplied
70 # key schedule and S-boxes and abstains from copying the former if
71 # there is no. This allows end-user to consciously retain small block
72 # performance by aligning key schedule in specific manner.
73 #
74 # Version 3.6 compresses Td4 to 256 bytes and prefetches it in ECB.
75 #
76 # Current ECB performance numbers for 128-bit key in CPU cycles per
77 # processed byte [measure commonly used by AES benchmarkers] are:
78 #
79 #               small footprint         fully unrolled
80 # P4            24                      22
81 # AMD K8        20                      19
82 # PIII          25                      23
83 # Pentium       81                      78
84 #
85 # Version 3.7 reimplements outer rounds as "compact." Meaning that
86 # first and last rounds reference compact 256 bytes S-box. This means
87 # that first round consumes a lot more CPU cycles and that encrypt
88 # and decrypt performance becomes asymmetric. Encrypt performance
89 # drops by 10-12%, while decrypt - by 20-25%:-( 256 bytes S-box is
90 # aggressively pre-fetched.
91 #
92 # Version 4.0 effectively rolls back to 3.6 and instead implements
93 # additional set of functions, _[x86|sse]_AES_[en|de]crypt_compact,
94 # which use exclusively 256 byte S-box. These functions are to be
95 # called in modes not concealing plain text, such as ECB, or when
96 # we're asked to process smaller amount of data [or unconditionally
97 # on hyper-threading CPU]. Currently it's called unconditionally from
98 # AES_[en|de]crypt, which affects all modes, but CBC. CBC routine
99 # still needs to be modified to switch between slower and faster
100 # mode when appropriate... But in either case benchmark landscape
101 # changes dramatically and below numbers are CPU cycles per processed
102 # byte for 128-bit key.
103 #
104 #               ECB encrypt     ECB decrypt     CBC large chunk
105 # P4            56[60]          84[100]         23
106 # AMD K8        48[44]          70[79]          18
107 # PIII          41[50]          61[91]          24
108 # Pentium       120             160             77
109 #
110 # Version 4.1 switches to compact S-box even in key schedule setup.
111 #
112 # Version 4.2 prefetches compact S-box in every SSE round or in other
113 # words every cache-line is *guaranteed* to be accessed within ~50
114 # cycles window. Why just SSE? Because it's needed on hyper-threading
115 # CPU! Which is also why it's prefetched with 64 byte stride. Best
116 # part is that it has no negative effect on performance:-)  
117 #
118 # Version 4.3 implements switch between compact and non-compact block
119 # functions in AES_cbc_encrypt depending on how much data was asked
120 # to be processed in one stroke.
121 #
122 ######################################################################
123 # Timing attacks are classified in two classes: synchronous when
124 # attacker consciously initiates cryptographic operation and collects
125 # timing data of various character afterwards, and asynchronous when
126 # malicious code is executed on same CPU simultaneously with AES,
127 # instruments itself and performs statistical analysis of this data.
128 #
129 # As far as synchronous attacks go the root to the AES timing
130 # vulnerability is twofold. Firstly, of 256 S-box elements at most 160
131 # are referred to in single 128-bit block operation. Well, in C
132 # implementation with 4 distinct tables it's actually as little as 40
133 # references per 256 elements table, but anyway... Secondly, even
134 # though S-box elements are clustered into smaller amount of cache-
135 # lines, smaller than 160 and even 40, it turned out that for certain
136 # plain-text pattern[s] or simply put chosen plain-text and given key
137 # few cache-lines remain unaccessed during block operation. Now, if
138 # attacker can figure out this access pattern, he can deduct the key
139 # [or at least part of it]. The natural way to mitigate this kind of
140 # attacks is to minimize the amount of cache-lines in S-box and/or
141 # prefetch them to ensure that every one is accessed for more uniform
142 # timing. But note that *if* plain-text was concealed in such way that
143 # input to block function is distributed *uniformly*, then attack
144 # wouldn't apply. Now note that some encryption modes, most notably
145 # CBC, do mask the plain-text in this exact way [secure cipher output
146 # is distributed uniformly]. Yes, one still might find input that
147 # would reveal the information about given key, but if amount of
148 # candidate inputs to be tried is larger than amount of possible key
149 # combinations then attack becomes infeasible. This is why revised
150 # AES_cbc_encrypt "dares" to switch to larger S-box when larger chunk
151 # of data is to be processed in one stroke. The current size limit of
152 # 512 bytes is chosen to provide same [diminishigly low] probability
153 # for cache-line to remain untouched in large chunk operation with
154 # large S-box as for single block operation with compact S-box and
155 # surely needs more careful consideration...
156 #
157 # As for asynchronous attacks. There are two flavours: attacker code
158 # being interleaved with AES on hyper-threading CPU at *instruction*
159 # level, and two processes time sharing single core. As for latter.
160 # Two vectors. 1. Given that attacker process has higher priority,
161 # yield execution to process performing AES just before timer fires
162 # off the scheduler, immediately regain control of CPU and analyze the
163 # cache state. For this attack to be efficient attacker would have to
164 # effectively slow down the operation by several *orders* of magnitute,
165 # by ratio of time slice to duration of handful of AES rounds, which
166 # unlikely to remain unnoticed. Not to mention that this also means
167 # that he would spend correspondigly more time to collect enough
168 # statistical data to mount the attack. It's probably appropriate to
169 # say that if adeversary reckons that this attack is beneficial and
170 # risks to be noticed, you probably have larger problems having him
171 # mere opportunity. In other words suggested code design expects you
172 # to preclude/mitigate this attack by overall system security design.
173 # 2. Attacker manages to make his code interrupt driven. In order for
174 # this kind of attack to be feasible, interrupt rate has to be high
175 # enough, again comparable to duration of handful of AES rounds. But
176 # is there interrupt source of such rate? Hardly, not even 1Gbps NIC
177 # generates interrupts at such raging rate...
178 #
179 # And now back to the former, hyper-threading CPU or more specifically
180 # Intel P4. Recall that asynchronous attack implies that malicious
181 # code instruments itself. And naturally instrumentation granularity
182 # has be noticeably lower than duration of codepath accessing S-box.
183 # Given that all cache-lines are accessed during that time that is.
184 # Current implementation accesses *all* cache-lines within ~50 cycles
185 # window, which is actually *less* than RDTSC latency on Intel P4!
186
187 push(@INC,"perlasm","../../perlasm");
188 require "x86asm.pl";
189
190 &asm_init($ARGV[0],"aes-586.pl",$x86only = $ARGV[$#ARGV] eq "386");
191
192 $s0="eax";
193 $s1="ebx";
194 $s2="ecx";
195 $s3="edx";
196 $key="edi";
197 $acc="esi";
198 $tbl="ebp";
199
200 # stack frame layout in _[x86|sse]_AES_* routines, frame is allocated
201 # by caller
202 $__ra=&DWP(0,"esp");    # return address
203 $__s0=&DWP(4,"esp");    # s0 backing store
204 $__s1=&DWP(8,"esp");    # s1 backing store
205 $__s2=&DWP(12,"esp");   # s2 backing store
206 $__s3=&DWP(16,"esp");   # s3 backing store
207 $__key=&DWP(20,"esp");  # pointer to key schedule
208 $__end=&DWP(24,"esp");  # pointer to end of key schedule
209 $__tbl=&DWP(28,"esp");  # %ebp backing store
210
211 # stack frame layout in AES_[en|crypt] routines, which differs from
212 # above by 4 and overlaps by %ebp backing store
213 $_tbl=&DWP(24,"esp");
214 $_esp=&DWP(28,"esp");
215
216 sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
217
218 $speed_limit=512;       # chunks smaller than $speed_limit are
219                         # processed with compact routine in CBC mode
220 $small_footprint=1;     # $small_footprint=1 code is ~5% slower [on
221                         # recent µ-archs], but ~5 times smaller!
222                         # I favor compact code to minimize cache
223                         # contention and in hope to "collect" 5% back
224                         # in real-life applications...
225
226 $vertical_spin=0;       # shift "verticaly" defaults to 0, because of
227                         # its proof-of-concept status...
228 # Note that there is no decvert(), as well as last encryption round is
229 # performed with "horizontal" shifts. This is because this "vertical"
230 # implementation [one which groups shifts on a given $s[i] to form a
231 # "column," unlike "horizontal" one, which groups shifts on different
232 # $s[i] to form a "row"] is work in progress. It was observed to run
233 # few percents faster on Intel cores, but not AMD. On AMD K8 core it's
234 # whole 12% slower:-( So we face a trade-off... Shall it be resolved
235 # some day? Till then the code is considered experimental and by
236 # default remains dormant...
237
238 sub encvert()
239 { my ($te,@s) = @_;
240   my $v0 = $acc, $v1 = $key;
241
242         &mov    ($v0,$s[3]);                            # copy s3
243         &mov    (&DWP(4,"esp"),$s[2]);                  # save s2
244         &mov    ($v1,$s[0]);                            # copy s0
245         &mov    (&DWP(8,"esp"),$s[1]);                  # save s1
246
247         &movz   ($s[2],&HB($s[0]));
248         &and    ($s[0],0xFF);
249         &mov    ($s[0],&DWP(0,$te,$s[0],8));            # s0>>0
250         &shr    ($v1,16);
251         &mov    ($s[3],&DWP(3,$te,$s[2],8));            # s0>>8
252         &movz   ($s[1],&HB($v1));
253         &and    ($v1,0xFF);
254         &mov    ($s[2],&DWP(2,$te,$v1,8));              # s0>>16
255          &mov   ($v1,$v0);
256         &mov    ($s[1],&DWP(1,$te,$s[1],8));            # s0>>24
257
258         &and    ($v0,0xFF);
259         &xor    ($s[3],&DWP(0,$te,$v0,8));              # s3>>0
260         &movz   ($v0,&HB($v1));
261         &shr    ($v1,16);
262         &xor    ($s[2],&DWP(3,$te,$v0,8));              # s3>>8
263         &movz   ($v0,&HB($v1));
264         &and    ($v1,0xFF);
265         &xor    ($s[1],&DWP(2,$te,$v1,8));              # s3>>16
266          &mov   ($v1,&DWP(4,"esp"));                    # restore s2
267         &xor    ($s[0],&DWP(1,$te,$v0,8));              # s3>>24
268
269         &mov    ($v0,$v1);
270         &and    ($v1,0xFF);
271         &xor    ($s[2],&DWP(0,$te,$v1,8));              # s2>>0
272         &movz   ($v1,&HB($v0));
273         &shr    ($v0,16);
274         &xor    ($s[1],&DWP(3,$te,$v1,8));              # s2>>8
275         &movz   ($v1,&HB($v0));
276         &and    ($v0,0xFF);
277         &xor    ($s[0],&DWP(2,$te,$v0,8));              # s2>>16
278          &mov   ($v0,&DWP(8,"esp"));                    # restore s1
279         &xor    ($s[3],&DWP(1,$te,$v1,8));              # s2>>24
280
281         &mov    ($v1,$v0);
282         &and    ($v0,0xFF);
283         &xor    ($s[1],&DWP(0,$te,$v0,8));              # s1>>0
284         &movz   ($v0,&HB($v1));
285         &shr    ($v1,16);
286         &xor    ($s[0],&DWP(3,$te,$v0,8));              # s1>>8
287         &movz   ($v0,&HB($v1));
288         &and    ($v1,0xFF);
289         &xor    ($s[3],&DWP(2,$te,$v1,8));              # s1>>16
290          &mov   ($key,$__key);                          # reincarnate v1 as key
291         &xor    ($s[2],&DWP(1,$te,$v0,8));              # s1>>24
292 }
293
294 # Another experimental routine, which features "horizontal spin," but
295 # eliminates one reference to stack. Strangely enough runs slower...
296 sub enchoriz()
297 { my $v0 = $key, $v1 = $acc;
298
299         &movz   ($v0,&LB($s0));                 #  3, 2, 1, 0*
300         &rotr   ($s2,8);                        #  8,11,10, 9
301         &mov    ($v1,&DWP(0,$te,$v0,8));        #  0
302         &movz   ($v0,&HB($s1));                 #  7, 6, 5*, 4
303         &rotr   ($s3,16);                       # 13,12,15,14
304         &xor    ($v1,&DWP(3,$te,$v0,8));        #  5
305         &movz   ($v0,&HB($s2));                 #  8,11,10*, 9
306         &rotr   ($s0,16);                       #  1, 0, 3, 2
307         &xor    ($v1,&DWP(2,$te,$v0,8));        # 10
308         &movz   ($v0,&HB($s3));                 # 13,12,15*,14
309         &xor    ($v1,&DWP(1,$te,$v0,8));        # 15, t[0] collected
310         &mov    ($__s0,$v1);                    # t[0] saved
311
312         &movz   ($v0,&LB($s1));                 #  7, 6, 5, 4*
313         &shr    ($s1,16);                       #  -, -, 7, 6
314         &mov    ($v1,&DWP(0,$te,$v0,8));        #  4
315         &movz   ($v0,&LB($s3));                 # 13,12,15,14*
316         &xor    ($v1,&DWP(2,$te,$v0,8));        # 14
317         &movz   ($v0,&HB($s0));                 #  1, 0, 3*, 2
318         &and    ($s3,0xffff0000);               # 13,12, -, -
319         &xor    ($v1,&DWP(1,$te,$v0,8));        #  3
320         &movz   ($v0,&LB($s2));                 #  8,11,10, 9*
321         &or     ($s3,$s1);                      # 13,12, 7, 6
322         &xor    ($v1,&DWP(3,$te,$v0,8));        #  9, t[1] collected
323         &mov    ($s1,$v1);                      #  s[1]=t[1]
324
325         &movz   ($v0,&LB($s0));                 #  1, 0, 3, 2*
326         &shr    ($s2,16);                       #  -, -, 8,11
327         &mov    ($v1,&DWP(2,$te,$v0,8));        #  2
328         &movz   ($v0,&HB($s3));                 # 13,12, 7*, 6
329         &xor    ($v1,&DWP(1,$te,$v0,8));        #  7
330         &movz   ($v0,&HB($s2));                 #  -, -, 8*,11
331         &xor    ($v1,&DWP(0,$te,$v0,8));        #  8
332         &mov    ($v0,$s3);
333         &shr    ($v0,24);                       # 13
334         &xor    ($v1,&DWP(3,$te,$v0,8));        # 13, t[2] collected
335
336         &movz   ($v0,&LB($s2));                 #  -, -, 8,11*
337         &shr    ($s0,24);                       #  1*
338         &mov    ($s2,&DWP(1,$te,$v0,8));        # 11
339         &xor    ($s2,&DWP(3,$te,$s0,8));        #  1
340         &mov    ($s0,$__s0);                    # s[0]=t[0]
341         &movz   ($v0,&LB($s3));                 # 13,12, 7, 6*
342         &shr    ($s3,16);                       #   ,  ,13,12
343         &xor    ($s2,&DWP(2,$te,$v0,8));        #  6
344         &mov    ($key,$__key);                  # reincarnate v0 as key
345         &and    ($s3,0xff);                     #   ,  ,13,12*
346         &mov    ($s3,&DWP(0,$te,$s3,8));        # 12
347         &xor    ($s3,$s2);                      # s[2]=t[3] collected
348         &mov    ($s2,$v1);                      # s[2]=t[2]
349 }
350
351 # More experimental code... SSE one... Even though this one eliminates
352 # *all* references to stack, it's not faster...
353 sub sse_encbody()
354 {
355         &movz   ($acc,&LB("eax"));              #  0
356         &mov    ("ecx",&DWP(0,$tbl,$acc,8));    #  0
357         &pshufw ("mm2","mm0",0x0d);             #  7, 6, 3, 2
358         &movz   ("edx",&HB("eax"));             #  1
359         &mov    ("edx",&DWP(3,$tbl,"edx",8));   #  1
360         &shr    ("eax",16);                     #  5, 4
361
362         &movz   ($acc,&LB("ebx"));              # 10
363         &xor    ("ecx",&DWP(2,$tbl,$acc,8));    # 10
364         &pshufw ("mm6","mm4",0x08);             # 13,12, 9, 8
365         &movz   ($acc,&HB("ebx"));              # 11
366         &xor    ("edx",&DWP(1,$tbl,$acc,8));    # 11
367         &shr    ("ebx",16);                     # 15,14
368
369         &movz   ($acc,&HB("eax"));              #  5
370         &xor    ("ecx",&DWP(3,$tbl,$acc,8));    #  5
371         &movq   ("mm3",QWP(16,$key));
372         &movz   ($acc,&HB("ebx"));              # 15
373         &xor    ("ecx",&DWP(1,$tbl,$acc,8));    # 15
374         &movd   ("mm0","ecx");                  # t[0] collected
375
376         &movz   ($acc,&LB("eax"));              #  4
377         &mov    ("ecx",&DWP(0,$tbl,$acc,8));    #  4
378         &movd   ("eax","mm2");                  #  7, 6, 3, 2
379         &movz   ($acc,&LB("ebx"));              # 14
380         &xor    ("ecx",&DWP(2,$tbl,$acc,8));    # 14
381         &movd   ("ebx","mm6");                  # 13,12, 9, 8
382
383         &movz   ($acc,&HB("eax"));              #  3
384         &xor    ("ecx",&DWP(1,$tbl,$acc,8));    #  3
385         &movz   ($acc,&HB("ebx"));              #  9
386         &xor    ("ecx",&DWP(3,$tbl,$acc,8));    #  9
387         &movd   ("mm1","ecx");                  # t[1] collected
388
389         &movz   ($acc,&LB("eax"));              #  2
390         &mov    ("ecx",&DWP(2,$tbl,$acc,8));    #  2
391         &shr    ("eax",16);                     #  7, 6
392         &punpckldq      ("mm0","mm1");          # t[0,1] collected
393         &movz   ($acc,&LB("ebx"));              #  8
394         &xor    ("ecx",&DWP(0,$tbl,$acc,8));    #  8
395         &shr    ("ebx",16);                     # 13,12
396
397         &movz   ($acc,&HB("eax"));              #  7
398         &xor    ("ecx",&DWP(1,$tbl,$acc,8));    #  7
399         &pxor   ("mm0","mm3");
400         &movz   ("eax",&LB("eax"));             #  6
401         &xor    ("edx",&DWP(2,$tbl,"eax",8));   #  6
402         &pshufw ("mm1","mm0",0x08);             #  5, 4, 1, 0
403         &movz   ($acc,&HB("ebx"));              # 13
404         &xor    ("ecx",&DWP(3,$tbl,$acc,8));    # 13
405         &xor    ("ecx",&DWP(24,$key));          # t[2]
406         &movd   ("mm4","ecx");                  # t[2] collected
407         &movz   ("ebx",&LB("ebx"));             # 12
408         &xor    ("edx",&DWP(0,$tbl,"ebx",8));   # 12
409         &shr    ("ecx",16);
410         &movd   ("eax","mm1");                  #  5, 4, 1, 0
411         &mov    ("ebx",&DWP(28,$key));          # t[3]
412         &xor    ("ebx","edx");
413         &movd   ("mm5","ebx");                  # t[3] collected
414         &and    ("ebx",0xffff0000);
415         &or     ("ebx","ecx");
416
417         &punpckldq      ("mm4","mm5");          # t[2,3] collected
418 }
419
420 ######################################################################
421 # "Compact" block function
422 ######################################################################
423
424 sub enccompact()
425 { my $Fn = mov;
426   while ($#_>5) { pop(@_); $Fn=sub{}; }
427   my ($i,$te,@s)=@_;
428   my $tmp = $key;
429   my $out = $i==3?$s[0]:$acc;
430
431         # $Fn is used in first compact round and its purpose is to
432         # void restoration of some values from stack, so that after
433         # 4xenccompact with extra argument $key value is left there...
434         if ($i==3)  {   &$Fn    ($key,$__key);                  }##%edx
435         else        {   &mov    ($out,$s[0]);                   }
436                         &and    ($out,0xFF);
437         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
438         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
439                         &movz   ($out,&BP(-128,$te,$out,1));
440
441         if ($i==3)  {   $tmp=$s[1];                             }##%eax
442                         &movz   ($tmp,&HB($s[1]));
443                         &movz   ($tmp,&BP(-128,$te,$tmp,1));
444                         &shl    ($tmp,8);
445                         &xor    ($out,$tmp);
446
447         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$__s0);         }##%ebx
448         else        {   &mov    ($tmp,$s[2]);
449                         &shr    ($tmp,16);                      }
450         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
451                         &and    ($tmp,0xFF);
452                         &movz   ($tmp,&BP(-128,$te,$tmp,1));
453                         &shl    ($tmp,16);
454                         &xor    ($out,$tmp);
455
456         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],$__s1);         }##%ecx
457         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
458         else        {   &mov    ($tmp,$s[3]);
459                         &shr    ($tmp,24);                      }
460                         &movz   ($tmp,&BP(-128,$te,$tmp,1));
461                         &shl    ($tmp,24);
462                         &xor    ($out,$tmp);
463         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
464         if ($i==3)  {   &mov    ($s[3],$acc);                   }
465         &comment();
466 }
467
468 sub enctransform()
469 { my @s = ($s0,$s1,$s2,$s3);
470   my $i = shift;
471   my $tmp = $tbl;
472   my $r2  = $key ;
473
474         &mov    ($acc,$s[$i]);
475         &and    ($acc,0x80808080);
476         &mov    ($tmp,$acc);
477         &mov    ($r2,$s[$i]);
478         &shr    ($tmp,7);
479         &and    ($r2,0x7f7f7f7f);
480         &sub    ($acc,$tmp);
481         &lea    ($r2,&DWP(0,$r2,$r2));
482         &and    ($acc,0x1b1b1b1b);
483         &mov    ($tmp,$s[$i]);
484         &xor    ($acc,$r2);     # r2
485
486         &xor    ($s[$i],$acc);  # r0 ^ r2
487         &rotl   ($s[$i],24);
488         &xor    ($s[$i],$acc)   # ROTATE(r2^r0,24) ^ r2
489         &rotr   ($tmp,16);
490         &xor    ($s[$i],$tmp);
491         &rotr   ($tmp,8);
492         &xor    ($s[$i],$tmp);
493 }
494
495 &public_label("AES_Te");
496 &function_begin_B("_x86_AES_encrypt_compact");
497         # note that caller is expected to allocate stack frame for me!
498         &mov    ($__key,$key);                  # save key
499
500         &xor    ($s0,&DWP(0,$key));             # xor with key
501         &xor    ($s1,&DWP(4,$key));
502         &xor    ($s2,&DWP(8,$key));
503         &xor    ($s3,&DWP(12,$key));
504
505         &mov    ($acc,&DWP(240,$key));          # load key->rounds
506         &lea    ($acc,&DWP(-2,$acc,$acc));
507         &lea    ($acc,&DWP(0,$key,$acc,8));
508         &mov    ($__end,$acc);                  # end of key schedule
509
510         # prefetch Te4
511         &mov    ($key,&DWP(0-128,$tbl));
512         &mov    ($acc,&DWP(32-128,$tbl));
513         &mov    ($key,&DWP(64-128,$tbl));
514         &mov    ($acc,&DWP(96-128,$tbl));
515         &mov    ($key,&DWP(128-128,$tbl));
516         &mov    ($acc,&DWP(160-128,$tbl));
517         &mov    ($key,&DWP(192-128,$tbl));
518         &mov    ($acc,&DWP(224-128,$tbl));
519
520         &set_label("loop",16);
521
522                 &enccompact(0,$tbl,$s0,$s1,$s2,$s3,1);
523                 &enccompact(1,$tbl,$s1,$s2,$s3,$s0,1);
524                 &enccompact(2,$tbl,$s2,$s3,$s0,$s1,1);
525                 &enccompact(3,$tbl,$s3,$s0,$s1,$s2,1);
526                 &enctransform(2);
527                 &enctransform(3);
528                 &enctransform(0);
529                 &enctransform(1);
530                 &mov    ($key,$__key);
531                 &mov    ($tbl,$__tbl);
532                 &add    ($key,16);              # advance rd_key
533                 &xor    ($s0,&DWP(0,$key));
534                 &xor    ($s1,&DWP(4,$key));
535                 &xor    ($s2,&DWP(8,$key));
536                 &xor    ($s3,&DWP(12,$key));
537
538         &cmp    ($key,$__end);
539         &mov    ($__key,$key);
540         &jb     (&label("loop"));
541
542         &enccompact(0,$tbl,$s0,$s1,$s2,$s3);
543         &enccompact(1,$tbl,$s1,$s2,$s3,$s0);
544         &enccompact(2,$tbl,$s2,$s3,$s0,$s1);
545         &enccompact(3,$tbl,$s3,$s0,$s1,$s2);
546
547         &xor    ($s0,&DWP(16,$key));
548         &xor    ($s1,&DWP(20,$key));
549         &xor    ($s2,&DWP(24,$key));
550         &xor    ($s3,&DWP(28,$key));
551
552         &ret    ();
553 &function_end_B("_x86_AES_encrypt_compact");
554
555 ######################################################################
556 # "Compact" SSE block function.
557 ######################################################################
558 #
559 # Performance is not actually extraordinary in comparison to pure
560 # x86 code. In particular encrypt performance is virtually the same.
561 # Decrypt performance on the other hand is 15-20% better on newer
562 # µ-archs [but we're thankful for *any* improvement here], and ~50%
563 # better on PIII:-) And additionally on the pros side this code
564 # eliminates redundant references to stack and thus relieves/
565 # minimizes the pressure on the memory bus.
566 #
567 # MMX register layout                           lsb
568 # +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
569 # |          mm4          |          mm0          |
570 # +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
571 # |     s3    |     s2    |     s1    |     s0    |    
572 # +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
573 # |15|14|13|12|11|10| 9| 8| 7| 6| 5| 4| 3| 2| 1| 0|
574 # +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
575 #
576 # Indexes translate as s[N/4]>>(8*(N%4)), e.g. 5 means s1>>8.
577 # In this terms encryption and decryption "compact" permutation
578 # matrices can be depicted as following:
579 #
580 # encryption              lsb   # decryption              lsb
581 # +----++----+----+----+----+   # +----++----+----+----+----+
582 # | t0 || 15 | 10 |  5 |  0 |   # | t0 ||  7 | 10 | 13 |  0 |
583 # +----++----+----+----+----+   # +----++----+----+----+----+
584 # | t1 ||  3 | 14 |  9 |  4 |   # | t1 || 11 | 14 |  1 |  4 |
585 # +----++----+----+----+----+   # +----++----+----+----+----+
586 # | t2 ||  7 |  2 | 13 |  8 |   # | t2 || 15 |  2 |  5 |  8 |
587 # +----++----+----+----+----+   # +----++----+----+----+----+
588 # | t3 || 11 |  6 |  1 | 12 |   # | t3 ||  3 |  6 |  9 | 12 |
589 # +----++----+----+----+----+   # +----++----+----+----+----+
590 #
591 ######################################################################
592 # Why not xmm registers? Short answer. It was actually tested and
593 # was not any faster, but *contrary*, most notably on Intel CPUs.
594 # Longer answer. Main advantage of using mm registers is that movd
595 # latency is lower, especially on Intel P4. While arithmetic
596 # instructions are twice as many, they can be scheduled every cycle
597 # and not every second one when they are operating on xmm register,
598 # so that "arithmetic throughput" remains virtually the same. And
599 # finally the code can be executed even on elder SSE-only CPUs:-)
600
601 sub sse_enccompact()
602 {
603         &pshufw ("mm1","mm0",0x08);             #  5, 4, 1, 0
604         &pshufw ("mm5","mm4",0x0d);             # 15,14,11,10
605         &movd   ("eax","mm1");                  #  5, 4, 1, 0
606         &movd   ("ebx","mm5");                  # 15,14,11,10
607
608         &movz   ($acc,&LB("eax"));              #  0
609         &movz   ("ecx",&BP(-128,$tbl,$acc,1));  #  0
610         &pshufw ("mm2","mm0",0x0d);             #  7, 6, 3, 2
611         &movz   ("edx",&HB("eax"));             #  1
612         &movz   ("edx",&BP(-128,$tbl,"edx",1)); #  1
613         &shl    ("edx",8);                      #  1
614         &shr    ("eax",16);                     #  5, 4
615
616         &movz   ($acc,&LB("ebx"));              # 10
617         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 10
618         &shl    ($acc,16);                      # 10
619         &or     ("ecx",$acc);                   # 10
620         &pshufw ("mm6","mm4",0x08);             # 13,12, 9, 8
621         &movz   ($acc,&HB("ebx"));              # 11
622         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 11
623         &shl    ($acc,24);                      # 11
624         &or     ("edx",$acc);                   # 11
625         &shr    ("ebx",16);                     # 15,14
626
627         &movz   ($acc,&HB("eax"));              #  5
628         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  5
629         &shl    ($acc,8);                       #  5
630         &or     ("ecx",$acc);                   #  5
631         &movz   ($acc,&HB("ebx"));              # 15
632         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 15
633         &shl    ($acc,24);                      # 15
634         &or     ("ecx",$acc);                   # 15
635         &movd   ("mm0","ecx");                  # t[0] collected
636
637         &movz   ($acc,&LB("eax"));              #  4
638         &movz   ("ecx",&BP(-128,$tbl,$acc,1));  #  4
639         &movd   ("eax","mm2");                  #  7, 6, 3, 2
640         &movz   ($acc,&LB("ebx"));              # 14
641         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 14
642         &shl    ($acc,16);                      # 14
643         &or     ("ecx",$acc);                   # 14
644
645         &movd   ("ebx","mm6");                  # 13,12, 9, 8
646         &movz   ($acc,&HB("eax"));              #  3
647         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  3
648         &shl    ($acc,24);                      #  3
649         &or     ("ecx",$acc);                   #  3
650         &movz   ($acc,&HB("ebx"));              #  9
651         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  9
652         &shl    ($acc,8);                       #  9
653         &or     ("ecx",$acc);                   #  9
654         &movd   ("mm1","ecx");                  # t[1] collected
655
656         &movz   ($acc,&LB("ebx"));              #  8
657         &movz   ("ecx",&BP(-128,$tbl,$acc,1));  #  8
658         &shr    ("ebx",16);                     # 13,12
659         &movz   ($acc,&LB("eax"));              #  2
660         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  2
661         &shl    ($acc,16);                      #  2
662         &or     ("ecx",$acc);                   #  2
663         &shr    ("eax",16);                     #  7, 6
664
665         &punpckldq      ("mm0","mm1");          # t[0,1] collected
666
667         &movz   ($acc,&HB("eax"));              #  7
668         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  7
669         &shl    ($acc,24);                      #  7
670         &or     ("ecx",$acc);                   #  7
671         &and    ("eax",0xff);                   #  6
672         &movz   ("eax",&BP(-128,$tbl,"eax",1)); #  6
673         &shl    ("eax",16);                     #  6
674         &or     ("edx","eax");                  #  6
675         &movz   ($acc,&HB("ebx"));              # 13
676         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 13
677         &shl    ($acc,8);                       # 13
678         &or     ("ecx",$acc);                   # 13
679         &movd   ("mm4","ecx");                  # t[2] collected
680         &and    ("ebx",0xff);                   # 12
681         &movz   ("ebx",&BP(-128,$tbl,"ebx",1)); # 12
682         &or     ("edx","ebx");                  # 12
683         &movd   ("mm5","edx");                  # t[3] collected
684
685         &punpckldq      ("mm4","mm5");          # t[2,3] collected
686 }
687
688                                         if (!$x86only) {
689 &public_label("AES_Te");
690 &function_begin_B("_sse_AES_encrypt_compact");
691         &pxor   ("mm0",&QWP(0,$key));   #  7, 6, 5, 4, 3, 2, 1, 0
692         &pxor   ("mm4",&QWP(8,$key));   # 15,14,13,12,11,10, 9, 8
693
694         # note that caller is expected to allocate stack frame for me!
695         &mov    ($acc,&DWP(240,$key));          # load key->rounds
696         &lea    ($acc,&DWP(-2,$acc,$acc));
697         &lea    ($acc,&DWP(0,$key,$acc,8));
698         &mov    ($__end,$acc);                  # end of key schedule
699
700         &mov    ($s0,0x1b1b1b1b);               # magic constant
701         &mov    (&DWP(8,"esp"),$s0);
702         &mov    (&DWP(12,"esp"),$s0);
703
704         # prefetch Te4
705         &mov    ($s0,&DWP(0-128,$tbl));
706         &mov    ($s1,&DWP(32-128,$tbl));
707         &mov    ($s2,&DWP(64-128,$tbl));
708         &mov    ($s3,&DWP(96-128,$tbl));
709         &mov    ($s0,&DWP(128-128,$tbl));
710         &mov    ($s1,&DWP(160-128,$tbl));
711         &mov    ($s2,&DWP(192-128,$tbl));
712         &mov    ($s3,&DWP(224-128,$tbl));
713
714         &set_label("loop",16);
715                 &sse_enccompact();
716                 &add    ($key,16);
717                 &cmp    ($key,$__end);
718                 &ja     (&label("out"));
719
720                 &movq   ("mm2",&QWP(8,"esp"));
721                 &pxor   ("mm3","mm3");          &pxor   ("mm7","mm7");
722                 &movq   ("mm1","mm0");          &movq   ("mm5","mm4");  # r0
723                 &pcmpgtb("mm3","mm0");          &pcmpgtb("mm7","mm4");
724                 &pand   ("mm3","mm2");          &pand   ("mm7","mm2");
725                 &pshufw ("mm2","mm0",0xb1);     &pshufw ("mm6","mm4",0xb1);# ROTATE(r0,16)
726                 &paddb  ("mm0","mm0");          &paddb  ("mm4","mm4");
727                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # = r2
728                 &pshufw ("mm3","mm2",0xb1);     &pshufw ("mm7","mm6",0xb1);# r0
729                 &pxor   ("mm1","mm0");          &pxor   ("mm5","mm4");  # r0^r2
730                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= ROTATE(r0,16)
731
732                 &movq   ("mm2","mm3");          &movq   ("mm6","mm7");
733                 &pslld  ("mm3",8);              &pslld  ("mm7",8);
734                 &psrld  ("mm2",24);             &psrld  ("mm6",24);
735                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= r0<<8
736                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= r0>>24
737
738                 &movq   ("mm3","mm1");          &movq   ("mm7","mm5");
739                 &movq   ("mm2",&QWP(0,$key));   &movq   ("mm6",&QWP(8,$key));
740                 &psrld  ("mm1",8);              &psrld  ("mm5",8);
741                 &mov    ($s0,&DWP(0-128,$tbl));
742                 &pslld  ("mm3",24);             &pslld  ("mm7",24);
743                 &mov    ($s1,&DWP(64-128,$tbl));
744                 &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= (r2^r0)<<8
745                 &mov    ($s2,&DWP(128-128,$tbl));
746                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= (r2^r0)>>24
747                 &mov    ($s3,&DWP(192-128,$tbl));
748
749                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");
750         &jmp    (&label("loop"));
751
752         &set_label("out",16);
753         &pxor   ("mm0",&QWP(0,$key));
754         &pxor   ("mm4",&QWP(8,$key));
755
756         &ret    ();
757 &function_end_B("_sse_AES_encrypt_compact");
758                                         }
759
760 ######################################################################
761 # Vanilla block function.
762 ######################################################################
763
764 sub encstep()
765 { my ($i,$te,@s) = @_;
766   my $tmp = $key;
767   my $out = $i==3?$s[0]:$acc;
768
769         # lines marked with #%e?x[i] denote "reordered" instructions...
770         if ($i==3)  {   &mov    ($key,$__key);                  }##%edx
771         else        {   &mov    ($out,$s[0]);
772                         &and    ($out,0xFF);                    }
773         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
774         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
775                         &mov    ($out,&DWP(0,$te,$out,8));
776
777         if ($i==3)  {   $tmp=$s[1];                             }##%eax
778                         &movz   ($tmp,&HB($s[1]));
779                         &xor    ($out,&DWP(3,$te,$tmp,8));
780
781         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$__s0);         }##%ebx
782         else        {   &mov    ($tmp,$s[2]);
783                         &shr    ($tmp,16);                      }
784         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
785                         &and    ($tmp,0xFF);
786                         &xor    ($out,&DWP(2,$te,$tmp,8));
787
788         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],$__s1);         }##%ecx
789         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
790         else        {   &mov    ($tmp,$s[3]); 
791                         &shr    ($tmp,24)                       }
792                         &xor    ($out,&DWP(1,$te,$tmp,8));
793         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
794         if ($i==3)  {   &mov    ($s[3],$acc);                   }
795                         &comment();
796 }
797
798 sub enclast()
799 { my ($i,$te,@s)=@_;
800   my $tmp = $key;
801   my $out = $i==3?$s[0]:$acc;
802
803         if ($i==3)  {   &mov    ($key,$__key);                  }##%edx
804         else        {   &mov    ($out,$s[0]);                   }
805                         &and    ($out,0xFF);
806         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
807         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
808                         &mov    ($out,&DWP(2,$te,$out,8));
809                         &and    ($out,0x000000ff);
810
811         if ($i==3)  {   $tmp=$s[1];                             }##%eax
812                         &movz   ($tmp,&HB($s[1]));
813                         &mov    ($tmp,&DWP(0,$te,$tmp,8));
814                         &and    ($tmp,0x0000ff00);
815                         &xor    ($out,$tmp);
816
817         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$__s0);         }##%ebx
818         else        {   &mov    ($tmp,$s[2]);
819                         &shr    ($tmp,16);                      }
820         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
821                         &and    ($tmp,0xFF);
822                         &mov    ($tmp,&DWP(0,$te,$tmp,8));
823                         &and    ($tmp,0x00ff0000);
824                         &xor    ($out,$tmp);
825
826         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],$__s1);         }##%ecx
827         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
828         else        {   &mov    ($tmp,$s[3]);
829                         &shr    ($tmp,24);                      }
830                         &mov    ($tmp,&DWP(2,$te,$tmp,8));
831                         &and    ($tmp,0xff000000);
832                         &xor    ($out,$tmp);
833         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
834         if ($i==3)  {   &mov    ($s[3],$acc);                   }
835 }
836
837 &public_label("AES_Te");
838 &function_begin_B("_x86_AES_encrypt");
839         if ($vertical_spin) {
840                 # I need high parts of volatile registers to be accessible...
841                 &exch   ($s1="edi",$key="ebx");
842                 &mov    ($s2="esi",$acc="ecx");
843         }
844
845         # note that caller is expected to allocate stack frame for me!
846         &mov    ($__key,$key);                  # save key
847
848         &xor    ($s0,&DWP(0,$key));             # xor with key
849         &xor    ($s1,&DWP(4,$key));
850         &xor    ($s2,&DWP(8,$key));
851         &xor    ($s3,&DWP(12,$key));
852
853         &mov    ($acc,&DWP(240,$key));          # load key->rounds
854
855         if ($small_footprint) {
856             &lea        ($acc,&DWP(-2,$acc,$acc));
857             &lea        ($acc,&DWP(0,$key,$acc,8));
858             &mov        ($__end,$acc);          # end of key schedule
859
860             &set_label("loop",16);
861                 if ($vertical_spin) {
862                     &encvert($tbl,$s0,$s1,$s2,$s3);
863                 } else {
864                     &encstep(0,$tbl,$s0,$s1,$s2,$s3);
865                     &encstep(1,$tbl,$s1,$s2,$s3,$s0);
866                     &encstep(2,$tbl,$s2,$s3,$s0,$s1);
867                     &encstep(3,$tbl,$s3,$s0,$s1,$s2);
868                 }
869                 &add    ($key,16);              # advance rd_key
870                 &xor    ($s0,&DWP(0,$key));
871                 &xor    ($s1,&DWP(4,$key));
872                 &xor    ($s2,&DWP(8,$key));
873                 &xor    ($s3,&DWP(12,$key));
874             &cmp        ($key,$__end);
875             &mov        ($__key,$key);
876             &jb         (&label("loop"));
877         }
878         else {
879             &cmp        ($acc,10);
880             &jle        (&label("10rounds"));
881             &cmp        ($acc,12);
882             &jle        (&label("12rounds"));
883
884         &set_label("14rounds",4);
885             for ($i=1;$i<3;$i++) {
886                 if ($vertical_spin) {
887                     &encvert($tbl,$s0,$s1,$s2,$s3);
888                 } else {
889                     &encstep(0,$tbl,$s0,$s1,$s2,$s3);
890                     &encstep(1,$tbl,$s1,$s2,$s3,$s0);
891                     &encstep(2,$tbl,$s2,$s3,$s0,$s1);
892                     &encstep(3,$tbl,$s3,$s0,$s1,$s2);
893                 }
894                 &xor    ($s0,&DWP(16*$i+0,$key));
895                 &xor    ($s1,&DWP(16*$i+4,$key));
896                 &xor    ($s2,&DWP(16*$i+8,$key));
897                 &xor    ($s3,&DWP(16*$i+12,$key));
898             }
899             &add        ($key,32);
900             &mov        ($__key,$key);          # advance rd_key
901         &set_label("12rounds",4);
902             for ($i=1;$i<3;$i++) {
903                 if ($vertical_spin) {
904                     &encvert($tbl,$s0,$s1,$s2,$s3);
905                 } else {
906                     &encstep(0,$tbl,$s0,$s1,$s2,$s3);
907                     &encstep(1,$tbl,$s1,$s2,$s3,$s0);
908                     &encstep(2,$tbl,$s2,$s3,$s0,$s1);
909                     &encstep(3,$tbl,$s3,$s0,$s1,$s2);
910                 }
911                 &xor    ($s0,&DWP(16*$i+0,$key));
912                 &xor    ($s1,&DWP(16*$i+4,$key));
913                 &xor    ($s2,&DWP(16*$i+8,$key));
914                 &xor    ($s3,&DWP(16*$i+12,$key));
915             }
916             &add        ($key,32);
917             &mov        ($__key,$key);          # advance rd_key
918         &set_label("10rounds",4);
919             for ($i=1;$i<10;$i++) {
920                 if ($vertical_spin) {
921                     &encvert($tbl,$s0,$s1,$s2,$s3);
922                 } else {
923                     &encstep(0,$tbl,$s0,$s1,$s2,$s3);
924                     &encstep(1,$tbl,$s1,$s2,$s3,$s0);
925                     &encstep(2,$tbl,$s2,$s3,$s0,$s1);
926                     &encstep(3,$tbl,$s3,$s0,$s1,$s2);
927                 }
928                 &xor    ($s0,&DWP(16*$i+0,$key));
929                 &xor    ($s1,&DWP(16*$i+4,$key));
930                 &xor    ($s2,&DWP(16*$i+8,$key));
931                 &xor    ($s3,&DWP(16*$i+12,$key));
932             }
933         }
934
935         if ($vertical_spin) {
936             # "reincarnate" some registers for "horizontal" spin...
937             &mov        ($s1="ebx",$key="edi");
938             &mov        ($s2="ecx",$acc="esi");
939         }
940         &enclast(0,$tbl,$s0,$s1,$s2,$s3);
941         &enclast(1,$tbl,$s1,$s2,$s3,$s0);
942         &enclast(2,$tbl,$s2,$s3,$s0,$s1);
943         &enclast(3,$tbl,$s3,$s0,$s1,$s2);
944
945         &add    ($key,$small_footprint?16:160);
946         &xor    ($s0,&DWP(0,$key));
947         &xor    ($s1,&DWP(4,$key));
948         &xor    ($s2,&DWP(8,$key));
949         &xor    ($s3,&DWP(12,$key));
950
951         &ret    ();
952
953 &set_label("AES_Te",64);        # Yes! I keep it in the code segment!
954         &_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
955         &_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
956         &_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
957         &_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
958         &_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
959         &_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
960         &_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
961         &_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
962         &_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
963         &_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
964         &_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
965         &_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
966         &_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
967         &_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
968         &_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
969         &_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
970         &_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
971         &_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
972         &_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
973         &_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
974         &_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
975         &_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
976         &_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
977         &_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
978         &_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
979         &_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
980         &_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
981         &_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
982         &_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
983         &_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
984         &_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
985         &_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
986         &_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
987         &_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
988         &_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
989         &_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
990         &_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
991         &_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
992         &_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
993         &_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
994         &_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
995         &_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
996         &_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
997         &_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
998         &_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
999         &_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
1000         &_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
1001         &_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);
1002         &_data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
1003         &_data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
1004         &_data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
1005         &_data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
1006         &_data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
1007         &_data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
1008         &_data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
1009         &_data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
1010         &_data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
1011         &_data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
1012         &_data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
1013         &_data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
1014         &_data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
1015         &_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
1016         &_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
1017         &_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
1018
1019 #Te4    # four copies of Te4 to choose from to avoid L1 aliasing
1020         &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1021         &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1022         &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1023         &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1024         &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1025         &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1026         &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1027         &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1028         &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1029         &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1030         &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1031         &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1032         &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1033         &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1034         &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1035         &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1036         &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1037         &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1038         &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1039         &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1040         &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1041         &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1042         &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1043         &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1044         &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1045         &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1046         &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1047         &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1048         &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1049         &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1050         &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1051         &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1052
1053         &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1054         &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1055         &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1056         &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1057         &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1058         &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1059         &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1060         &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1061         &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1062         &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1063         &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1064         &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1065         &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1066         &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1067         &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1068         &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1069         &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1070         &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1071         &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1072         &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1073         &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1074         &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1075         &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1076         &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1077         &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1078         &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1079         &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1080         &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1081         &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1082         &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1083         &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1084         &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1085
1086         &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1087         &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1088         &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1089         &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1090         &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1091         &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1092         &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1093         &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1094         &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1095         &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1096         &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1097         &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1098         &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1099         &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1100         &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1101         &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1102         &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1103         &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1104         &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1105         &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1106         &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1107         &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1108         &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1109         &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1110         &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1111         &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1112         &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1113         &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1114         &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1115         &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1116         &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1117         &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1118
1119         &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
1120         &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
1121         &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
1122         &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
1123         &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
1124         &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
1125         &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
1126         &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
1127         &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
1128         &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
1129         &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
1130         &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
1131         &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
1132         &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
1133         &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
1134         &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
1135         &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
1136         &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
1137         &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
1138         &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
1139         &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
1140         &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
1141         &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
1142         &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
1143         &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
1144         &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
1145         &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
1146         &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
1147         &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
1148         &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
1149         &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
1150         &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
1151 #rcon:
1152         &data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008);
1153         &data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080);
1154         &data_word(0x0000001b, 0x00000036, 0x00000000, 0x00000000);
1155         &data_word(0x00000000, 0x00000000, 0x00000000, 0x00000000);
1156 &function_end_B("_x86_AES_encrypt");
1157
1158 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
1159 &public_label("AES_Te");
1160 &function_begin("AES_encrypt");
1161         &mov    ($acc,&wparam(0));              # load inp
1162         &mov    ($key,&wparam(2));              # load key
1163
1164         &mov    ($s0,"esp");
1165         &sub    ("esp",36);
1166         &and    ("esp",-64);                    # align to cache-line
1167
1168         # place stack frame just "above" the key schedule
1169         &lea    ($s1,&DWP(-64-63,$key));
1170         &sub    ($s1,"esp");
1171         &neg    ($s1);
1172         &and    ($s1,0x3C0);    # modulo 1024, but aligned to cache-line
1173         &sub    ("esp",$s1);
1174         &add    ("esp",4);      # 4 is reserved for caller's return address
1175         &mov    ($_esp,$s0);                    # save stack pointer
1176
1177         &call   (&label("pic_point"));          # make it PIC!
1178         &set_label("pic_point");
1179         &blindpop($tbl);
1180         &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if (!$x86only);
1181         &lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
1182
1183         # pick Te4 copy which can't "overlap" with stack frame or key schedule
1184         &lea    ($s1,&DWP(768-4,"esp"));
1185         &sub    ($s1,$tbl);
1186         &and    ($s1,0x300);
1187         &lea    ($tbl,&DWP(2048+128,$tbl,$s1));
1188
1189                                         if (!$x86only) {
1190         &bt     (&DWP(0,$s0),25);       # check for SSE bit
1191         &jnc    (&label("x86"));
1192
1193         &movq   ("mm0",&QWP(0,$acc));
1194         &movq   ("mm4",&QWP(8,$acc));
1195         &call   ("_sse_AES_encrypt_compact");
1196         &mov    ("esp",$_esp);                  # restore stack pointer
1197         &mov    ($acc,&wparam(1));              # load out
1198         &movq   (&QWP(0,$acc),"mm0");           # write output data
1199         &movq   (&QWP(8,$acc),"mm4");
1200         &emms   ();
1201         &function_end_A();
1202                                         }
1203         &set_label("x86",16);
1204         &mov    ($_tbl,$tbl);
1205         &mov    ($s0,&DWP(0,$acc));             # load input data
1206         &mov    ($s1,&DWP(4,$acc));
1207         &mov    ($s2,&DWP(8,$acc));
1208         &mov    ($s3,&DWP(12,$acc));
1209         &call   ("_x86_AES_encrypt_compact");
1210         &mov    ("esp",$_esp);                  # restore stack pointer
1211         &mov    ($acc,&wparam(1));              # load out
1212         &mov    (&DWP(0,$acc),$s0);             # write output data
1213         &mov    (&DWP(4,$acc),$s1);
1214         &mov    (&DWP(8,$acc),$s2);
1215         &mov    (&DWP(12,$acc),$s3);
1216 &function_end("AES_encrypt");
1217
1218 #--------------------------------------------------------------------#
1219
1220 ######################################################################
1221 # "Compact" block function
1222 ######################################################################
1223
1224 sub deccompact()
1225 { my $Fn = mov;
1226   while ($#_>5) { pop(@_); $Fn=sub{}; }
1227   my ($i,$td,@s)=@_;
1228   my $tmp = $key;
1229   my $out = $i==3?$s[0]:$acc;
1230
1231         # $Fn is used in first compact round and its purpose is to
1232         # void restoration of some values from stack, so that after
1233         # 4xdeccompact with extra argument $key, $s0 and $s1 values
1234         # are left there...
1235         if($i==3)   {   &$Fn    ($key,$__key);                  }
1236         else        {   &mov    ($out,$s[0]);                   }
1237                         &and    ($out,0xFF);
1238                         &movz   ($out,&BP(-128,$td,$out,1));
1239
1240         if ($i==3)  {   $tmp=$s[1];                             }
1241                         &movz   ($tmp,&HB($s[1]));
1242                         &movz   ($tmp,&BP(-128,$td,$tmp,1));
1243                         &shl    ($tmp,8);
1244                         &xor    ($out,$tmp);
1245
1246         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
1247         else        {   mov     ($tmp,$s[2]);                   }
1248                         &shr    ($tmp,16);
1249                         &and    ($tmp,0xFF);
1250                         &movz   ($tmp,&BP(-128,$td,$tmp,1));
1251                         &shl    ($tmp,16);
1252                         &xor    ($out,$tmp);
1253
1254         if ($i==3)  {   $tmp=$s[3]; &$Fn ($s[2],$__s1);         }
1255         else        {   &mov    ($tmp,$s[3]);                   }
1256                         &shr    ($tmp,24);
1257                         &movz   ($tmp,&BP(-128,$td,$tmp,1));
1258                         &shl    ($tmp,24);
1259                         &xor    ($out,$tmp);
1260         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
1261         if ($i==3)  {   &$Fn    ($s[3],$__s0);                  }
1262 }
1263
1264 # must be called with 2,3,0,1 as argument sequence!!!
1265 sub dectransform()
1266 { my @s = ($s0,$s1,$s2,$s3);
1267   my $i = shift;
1268   my $tmp = $key;
1269   my $tp2 = @s[($i+2)%4]; $tp2 = @s[2] if ($i==1);
1270   my $tp4 = @s[($i+3)%4]; $tp4 = @s[3] if ($i==1);
1271   my $tp8 = $tbl;
1272
1273         &mov    ($acc,$s[$i]);
1274         &and    ($acc,0x80808080);
1275         &mov    ($tmp,$acc);
1276         &mov    ($tp2,$s[$i]);
1277         &shr    ($tmp,7);
1278         &and    ($tp2,0x7f7f7f7f);
1279         &sub    ($acc,$tmp);
1280         &add    ($tp2,$tp2);
1281         &and    ($acc,0x1b1b1b1b);
1282         &xor    ($acc,$tp2);
1283         &mov    ($tp2,$acc);
1284
1285         &and    ($acc,0x80808080);
1286         &mov    ($tmp,$acc);
1287         &mov    ($tp4,$tp2);
1288          &xor   ($tp2,$s[$i]);  # tp2^tp1
1289         &shr    ($tmp,7);
1290         &and    ($tp4,0x7f7f7f7f);
1291         &sub    ($acc,$tmp);
1292         &add    ($tp4,$tp4);
1293         &and    ($acc,0x1b1b1b1b);
1294         &xor    ($acc,$tp4);
1295         &mov    ($tp4,$acc);
1296
1297         &and    ($acc,0x80808080);
1298         &mov    ($tmp,$acc);
1299         &mov    ($tp8,$tp4);
1300          &xor   ($tp4,$s[$i]);  # tp4^tp1
1301         &shr    ($tmp,7);
1302         &and    ($tp8,0x7f7f7f7f);
1303         &sub    ($acc,$tmp);
1304         &add    ($tp8,$tp8);
1305         &and    ($acc,0x1b1b1b1b);
1306          &rotl  ($s[$i],8);     # = ROTATE(tp1,8)
1307         &xor    ($tp8,$acc);
1308
1309         &xor    ($s[$i],$tp2);
1310         &xor    ($tp2,$tp8);
1311         &xor    ($s[$i],$tp4);
1312         &rotl   ($tp2,24);
1313         &xor    ($tp4,$tp8);
1314         &xor    ($s[$i],$tp8);  # ^= tp8^(tp4^tp1)^(tp2^tp1)
1315         &rotl   ($tp4,16);
1316         &xor    ($s[$i],$tp2);  # ^= ROTATE(tp8^tp2^tp1,24)
1317         &rotl   ($tp8,8);
1318         &xor    ($s[$i],$tp4);  # ^= ROTATE(tp8^tp4^tp1,16)
1319         &xor    ($s[$i],$tp8);  # ^= ROTATE(tp8,8)
1320
1321         &mov    ($s[0],$__s0)                   if($i==2); #prefetch $s0
1322         &mov    ($s[1],$__s1)                   if($i==3); #prefetch $s1
1323         &mov    ($s[2],$__s2)                   if($i==1);
1324         &mov    ($s[3],$__s3)                   if($i==1);
1325         &mov    (&DWP(4+4*$i,"esp"),$s[$i])     if($i>=2);
1326 }
1327
1328 &public_label("AES_Td");
1329 &function_begin_B("_x86_AES_decrypt_compact");
1330         # note that caller is expected to allocate stack frame for me!
1331         &mov    ($__key,$key);                  # save key
1332
1333         &xor    ($s0,&DWP(0,$key));             # xor with key
1334         &xor    ($s1,&DWP(4,$key));
1335         &xor    ($s2,&DWP(8,$key));
1336         &xor    ($s3,&DWP(12,$key));
1337
1338         &mov    ($acc,&DWP(240,$key));          # load key->rounds
1339
1340         &lea    ($acc,&DWP(-2,$acc,$acc));
1341         &lea    ($acc,&DWP(0,$key,$acc,8));
1342         &mov    ($__end,$acc);                  # end of key schedule
1343
1344         # prefetch Td4
1345         &mov    ($key,&DWP(0-128,$tbl));
1346         &mov    ($acc,&DWP(32-128,$tbl));
1347         &mov    ($key,&DWP(64-128,$tbl));
1348         &mov    ($acc,&DWP(96-128,$tbl));
1349         &mov    ($key,&DWP(128-128,$tbl));
1350         &mov    ($acc,&DWP(160-128,$tbl));
1351         &mov    ($key,&DWP(192-128,$tbl));
1352         &mov    ($acc,&DWP(224-128,$tbl));
1353
1354         &set_label("loop",16);
1355
1356                 &deccompact(0,$tbl,$s0,$s3,$s2,$s1,1);
1357                 &deccompact(1,$tbl,$s1,$s0,$s3,$s2,1);
1358                 &deccompact(2,$tbl,$s2,$s1,$s0,$s3,1);
1359                 &deccompact(3,$tbl,$s3,$s2,$s1,$s0,1);
1360                 &dectransform(2);
1361                 &dectransform(3);
1362                 &dectransform(0);
1363                 &dectransform(1);
1364                 &mov    ($key,$__key);
1365                 &mov    ($tbl,$__tbl);
1366                 &add    ($key,16);              # advance rd_key
1367                 &xor    ($s0,&DWP(0,$key));
1368                 &xor    ($s1,&DWP(4,$key));
1369                 &xor    ($s2,&DWP(8,$key));
1370                 &xor    ($s3,&DWP(12,$key));
1371
1372         &cmp    ($key,$__end);
1373         &mov    ($__key,$key);
1374         &jb     (&label("loop"));
1375
1376         &deccompact(0,$tbl,$s0,$s3,$s2,$s1);
1377         &deccompact(1,$tbl,$s1,$s0,$s3,$s2);
1378         &deccompact(2,$tbl,$s2,$s1,$s0,$s3);
1379         &deccompact(3,$tbl,$s3,$s2,$s1,$s0);
1380
1381         &xor    ($s0,&DWP(16,$key));
1382         &xor    ($s1,&DWP(20,$key));
1383         &xor    ($s2,&DWP(24,$key));
1384         &xor    ($s3,&DWP(28,$key));
1385
1386         &ret    ();
1387 &function_end_B("_x86_AES_decrypt_compact");
1388
1389 ######################################################################
1390 # "Compact" SSE block function.
1391 ######################################################################
1392
1393 sub sse_deccompact()
1394 {
1395         &pshufw ("mm1","mm0",0x0c);             #  7, 6, 1, 0
1396         &movd   ("eax","mm1");                  #  7, 6, 1, 0
1397
1398         &pshufw ("mm5","mm4",0x09);             # 13,12,11,10
1399         &movz   ($acc,&LB("eax"));              #  0
1400         &movz   ("ecx",&BP(-128,$tbl,$acc,1));  #  0
1401         &movd   ("ebx","mm5");                  # 13,12,11,10
1402         &movz   ("edx",&HB("eax"));             #  1
1403         &movz   ("edx",&BP(-128,$tbl,"edx",1)); #  1
1404         &shl    ("edx",8);                      #  1
1405
1406         &pshufw ("mm2","mm0",0x06);             #  3, 2, 5, 4
1407         &movz   ($acc,&LB("ebx"));              # 10
1408         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 10
1409         &shl    ($acc,16);                      # 10
1410         &or     ("ecx",$acc);                   # 10
1411         &shr    ("eax",16);                     #  7, 6
1412         &movz   ($acc,&HB("ebx"));              # 11
1413         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 11
1414         &shl    ($acc,24);                      # 11
1415         &or     ("edx",$acc);                   # 11
1416         &shr    ("ebx",16);                     # 13,12
1417
1418         &pshufw ("mm6","mm4",0x03);             # 9, 8,15,14
1419         &movz   ($acc,&HB("eax"));              #  7
1420         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  7
1421         &shl    ($acc,24);                      #  7
1422         &or     ("ecx",$acc);                   #  7
1423         &movz   ($acc,&HB("ebx"));              # 13
1424         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 13
1425         &shl    ($acc,8);                       # 13
1426         &or     ("ecx",$acc);                   # 13
1427         &movd   ("mm0","ecx");                  # t[0] collected
1428
1429         &movz   ($acc,&LB("eax"));              #  6
1430         &movd   ("eax","mm2");                  #  3, 2, 5, 4
1431         &movz   ("ecx",&BP(-128,$tbl,$acc,1));  #  6
1432         &shl    ("ecx",16);                     #  6
1433         &movz   ($acc,&LB("ebx"));              # 12
1434         &movd   ("ebx","mm6");                  #  9, 8,15,14
1435         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 12
1436         &or     ("ecx",$acc);                   # 12
1437
1438         &movz   ($acc,&LB("eax"));              #  4
1439         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  4
1440         &or     ("edx",$acc);                   #  4
1441         &movz   ($acc,&LB("ebx"));              # 14
1442         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 14
1443         &shl    ($acc,16);                      # 14
1444         &or     ("edx",$acc);                   # 14
1445         &movd   ("mm1","edx");                  # t[1] collected
1446
1447         &movz   ($acc,&HB("eax"));              #  5
1448         &movz   ("edx",&BP(-128,$tbl,$acc,1));  #  5
1449         &shl    ("edx",8);                      #  5
1450         &movz   ($acc,&HB("ebx"));              # 15
1451         &shr    ("eax",16);                     #  3, 2
1452         &movz   ($acc,&BP(-128,$tbl,$acc,1));   # 15
1453         &shl    ($acc,24);                      # 15
1454         &or     ("edx",$acc);                   # 15
1455         &shr    ("ebx",16);                     #  9, 8
1456
1457         &punpckldq      ("mm0","mm1");          # t[0,1] collected
1458
1459         &movz   ($acc,&HB("ebx"));              #  9
1460         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  9
1461         &shl    ($acc,8);                       #  9
1462         &or     ("ecx",$acc);                   #  9
1463         &and    ("ebx",0xff);                   #  8
1464         &movz   ("ebx",&BP(-128,$tbl,"ebx",1)); #  8
1465         &or     ("edx","ebx");                  #  8
1466         &movz   ($acc,&LB("eax"));              #  2
1467         &movz   ($acc,&BP(-128,$tbl,$acc,1));   #  2
1468         &shl    ($acc,16);                      #  2
1469         &or     ("edx",$acc);                   #  2
1470         &movd   ("mm4","edx");                  # t[2] collected
1471         &movz   ("eax",&HB("eax"));             #  3
1472         &movz   ("eax",&BP(-128,$tbl,"eax",1)); #  3
1473         &shl    ("eax",24);                     #  3
1474         &or     ("ecx","eax");                  #  3
1475         &movd   ("mm5","ecx");                  # t[3] collected
1476
1477         &punpckldq      ("mm4","mm5");          # t[2,3] collected
1478 }
1479
1480                                         if (!$x86only) {
1481 &public_label("AES_Td");
1482 &function_begin_B("_sse_AES_decrypt_compact");
1483         &pxor   ("mm0",&QWP(0,$key));   #  7, 6, 5, 4, 3, 2, 1, 0
1484         &pxor   ("mm4",&QWP(8,$key));   # 15,14,13,12,11,10, 9, 8
1485
1486         # note that caller is expected to allocate stack frame for me!
1487         &mov    ($acc,&DWP(240,$key));          # load key->rounds
1488         &lea    ($acc,&DWP(-2,$acc,$acc));
1489         &lea    ($acc,&DWP(0,$key,$acc,8));
1490         &mov    ($__end,$acc);                  # end of key schedule
1491
1492         &mov    ($s0,0x1b1b1b1b);               # magic constant
1493         &mov    (&DWP(8,"esp"),$s0);
1494         &mov    (&DWP(12,"esp"),$s0);
1495
1496         # prefetch Td4
1497         &mov    ($s0,&DWP(0-128,$tbl));
1498         &mov    ($s1,&DWP(32-128,$tbl));
1499         &mov    ($s2,&DWP(64-128,$tbl));
1500         &mov    ($s3,&DWP(96-128,$tbl));
1501         &mov    ($s0,&DWP(128-128,$tbl));
1502         &mov    ($s1,&DWP(160-128,$tbl));
1503         &mov    ($s2,&DWP(192-128,$tbl));
1504         &mov    ($s3,&DWP(224-128,$tbl));
1505
1506         &set_label("loop",16);
1507                 &sse_deccompact();
1508                 &add    ($key,16);
1509                 &cmp    ($key,$__end);
1510                 &ja     (&label("out"));
1511
1512                 # ROTATE(x^y,N) == ROTATE(x,N)^ROTATE(y,N)
1513                 &movq   ("mm3","mm0");          &movq   ("mm7","mm4");
1514                 &movq   ("mm2","mm0",1);        &movq   ("mm6","mm4",1);
1515                 &movq   ("mm1","mm0");          &movq   ("mm5","mm4");
1516                 &pshufw ("mm0","mm0",0xb1);     &pshufw ("mm4","mm4",0xb1);# = ROTATE(tp0,16)
1517                 &pslld  ("mm2",8);              &pslld  ("mm6",8);
1518                 &psrld  ("mm3",8);              &psrld  ("mm7",8);
1519                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= tp0<<8
1520                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp0>>8
1521                 &pslld  ("mm2",16);             &pslld  ("mm6",16);
1522                 &psrld  ("mm3",16);             &psrld  ("mm7",16);
1523                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= tp0<<24
1524                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp0>>24
1525
1526                 &movq   ("mm3",&QWP(8,"esp"));
1527                 &pxor   ("mm2","mm2");          &pxor   ("mm6","mm6");
1528                 &pcmpgtb("mm2","mm1");          &pcmpgtb("mm6","mm5");
1529                 &pand   ("mm2","mm3");          &pand   ("mm6","mm3");
1530                 &paddb  ("mm1","mm1");          &paddb  ("mm5","mm5");
1531                 &pxor   ("mm1","mm2");          &pxor   ("mm5","mm6");  # tp2
1532                 &movq   ("mm3","mm1");          &movq   ("mm7","mm5");
1533                 &movq   ("mm2","mm1");          &movq   ("mm6","mm5");
1534                 &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp2
1535                 &pslld  ("mm3",24);             &pslld  ("mm7",24);
1536                 &psrld  ("mm2",8);              &psrld  ("mm6",8);
1537                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp2<<24
1538                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= tp2>>8
1539
1540                 &movq   ("mm2",&QWP(8,"esp"));
1541                 &pxor   ("mm3","mm3");          &pxor   ("mm7","mm7");
1542                 &pcmpgtb("mm3","mm1");          &pcmpgtb("mm7","mm5");
1543                 &pand   ("mm3","mm2");          &pand   ("mm7","mm2");
1544                 &paddb  ("mm1","mm1");          &paddb  ("mm5","mm5");
1545                 &pxor   ("mm1","mm3");          &pxor   ("mm5","mm7");  # tp4
1546                 &pshufw ("mm3","mm1",0xb1);     &pshufw ("mm7","mm5",0xb1);
1547                 &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp4
1548                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= ROTATE(tp4,16)     
1549
1550                 &pxor   ("mm3","mm3");          &pxor   ("mm7","mm7");
1551                 &pcmpgtb("mm3","mm1");          &pcmpgtb("mm7","mm5");
1552                 &pand   ("mm3","mm2");          &pand   ("mm7","mm2");
1553                 &paddb  ("mm1","mm1");          &paddb  ("mm5","mm5");
1554                 &pxor   ("mm1","mm3");          &pxor   ("mm5","mm7");  # tp8
1555                 &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp8
1556                 &movq   ("mm3","mm1");          &movq   ("mm7","mm5");
1557                 &pshufw ("mm2","mm1",0xb1);     &pshufw ("mm6","mm5",0xb1);
1558                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= ROTATE(tp8,16)
1559                 &pslld  ("mm1",8);              &pslld  ("mm5",8);
1560                 &psrld  ("mm3",8);              &psrld  ("mm7",8);
1561                 &movq   ("mm2",&QWP(0,$key));   &movq   ("mm6",&QWP(8,$key));
1562                 &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp8<<8
1563                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp8>>8
1564                 &mov    ($s0,&DWP(0-128,$tbl));
1565                 &pslld  ("mm1",16);             &pslld  ("mm5",16);
1566                 &mov    ($s1,&DWP(64-128,$tbl));
1567                 &psrld  ("mm3",16);             &psrld  ("mm7",16);
1568                 &mov    ($s2,&DWP(128-128,$tbl));
1569                 &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp8<<24
1570                 &mov    ($s3,&DWP(192-128,$tbl));
1571                 &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp8>>24
1572
1573                 &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");
1574         &jmp    (&label("loop"));
1575
1576         &set_label("out",16);
1577         &pxor   ("mm0",&QWP(0,$key));
1578         &pxor   ("mm4",&QWP(8,$key));
1579
1580         &ret    ();
1581 &function_end_B("_sse_AES_decrypt_compact");
1582                                         }
1583
1584 ######################################################################
1585 # Vanilla block function.
1586 ######################################################################
1587
1588 sub decstep()
1589 { my ($i,$td,@s) = @_;
1590   my $tmp = $key;
1591   my $out = $i==3?$s[0]:$acc;
1592
1593         # no instructions are reordered, as performance appears
1594         # optimal... or rather that all attempts to reorder didn't
1595         # result in better performance [which by the way is not a
1596         # bit lower than ecryption].
1597         if($i==3)   {   &mov    ($key,$__key);                  }
1598         else        {   &mov    ($out,$s[0]);                   }
1599                         &and    ($out,0xFF);
1600                         &mov    ($out,&DWP(0,$td,$out,8));
1601
1602         if ($i==3)  {   $tmp=$s[1];                             }
1603                         &movz   ($tmp,&HB($s[1]));
1604                         &xor    ($out,&DWP(3,$td,$tmp,8));
1605
1606         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
1607         else        {   &mov    ($tmp,$s[2]);                   }
1608                         &shr    ($tmp,16);
1609                         &and    ($tmp,0xFF);
1610                         &xor    ($out,&DWP(2,$td,$tmp,8));
1611
1612         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],$__s1);         }
1613         else        {   &mov    ($tmp,$s[3]);                   }
1614                         &shr    ($tmp,24);
1615                         &xor    ($out,&DWP(1,$td,$tmp,8));
1616         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
1617         if ($i==3)  {   &mov    ($s[3],$__s0);                  }
1618                         &comment();
1619 }
1620
1621 sub declast()
1622 { my ($i,$td,@s)=@_;
1623   my $tmp = $key;
1624   my $out = $i==3?$s[0]:$acc;
1625
1626         if($i==0)   {   &lea    ($td,&DWP(2048+128,$td));
1627                         &mov    ($tmp,&DWP(0-128,$td));
1628                         &mov    ($acc,&DWP(32-128,$td));
1629                         &mov    ($tmp,&DWP(64-128,$td));
1630                         &mov    ($acc,&DWP(96-128,$td));
1631                         &mov    ($tmp,&DWP(128-128,$td));
1632                         &mov    ($acc,&DWP(160-128,$td));
1633                         &mov    ($tmp,&DWP(192-128,$td));
1634                         &mov    ($acc,&DWP(224-128,$td));
1635                         &lea    ($td,&DWP(-128,$td));           }
1636         if($i==3)   {   &mov    ($key,$__key);                  }
1637         else        {   &mov    ($out,$s[0]);                   }
1638                         &and    ($out,0xFF);
1639                         &movz   ($out,&BP(0,$td,$out,1));
1640
1641         if ($i==3)  {   $tmp=$s[1];                             }
1642                         &movz   ($tmp,&HB($s[1]));
1643                         &movz   ($tmp,&BP(0,$td,$tmp,1));
1644                         &shl    ($tmp,8);
1645                         &xor    ($out,$tmp);
1646
1647         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
1648         else        {   mov     ($tmp,$s[2]);                   }
1649                         &shr    ($tmp,16);
1650                         &and    ($tmp,0xFF);
1651                         &movz   ($tmp,&BP(0,$td,$tmp,1));
1652                         &shl    ($tmp,16);
1653                         &xor    ($out,$tmp);
1654
1655         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],$__s1);         }
1656         else        {   &mov    ($tmp,$s[3]);                   }
1657                         &shr    ($tmp,24);
1658                         &movz   ($tmp,&BP(0,$td,$tmp,1));
1659                         &shl    ($tmp,24);
1660                         &xor    ($out,$tmp);
1661         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
1662         if ($i==3)  {   &mov    ($s[3],$__s0);
1663                         &lea    ($td,&DWP(-2048,$td));          }
1664 }
1665
1666 &public_label("AES_Td");
1667 &function_begin_B("_x86_AES_decrypt");
1668         # note that caller is expected to allocate stack frame for me!
1669         &mov    ($__key,$key);                  # save key
1670
1671         &xor    ($s0,&DWP(0,$key));             # xor with key
1672         &xor    ($s1,&DWP(4,$key));
1673         &xor    ($s2,&DWP(8,$key));
1674         &xor    ($s3,&DWP(12,$key));
1675
1676         &mov    ($acc,&DWP(240,$key));          # load key->rounds
1677
1678         if ($small_footprint) {
1679             &lea        ($acc,&DWP(-2,$acc,$acc));
1680             &lea        ($acc,&DWP(0,$key,$acc,8));
1681             &mov        ($__end,$acc);          # end of key schedule
1682             &set_label("loop",16);
1683                 &decstep(0,$tbl,$s0,$s3,$s2,$s1);
1684                 &decstep(1,$tbl,$s1,$s0,$s3,$s2);
1685                 &decstep(2,$tbl,$s2,$s1,$s0,$s3);
1686                 &decstep(3,$tbl,$s3,$s2,$s1,$s0);
1687                 &add    ($key,16);              # advance rd_key
1688                 &xor    ($s0,&DWP(0,$key));
1689                 &xor    ($s1,&DWP(4,$key));
1690                 &xor    ($s2,&DWP(8,$key));
1691                 &xor    ($s3,&DWP(12,$key));
1692             &cmp        ($key,$__end);
1693             &mov        ($__key,$key);
1694             &jb         (&label("loop"));
1695         }
1696         else {
1697             &cmp        ($acc,10);
1698             &jle        (&label("10rounds"));
1699             &cmp        ($acc,12);
1700             &jle        (&label("12rounds"));
1701
1702         &set_label("14rounds",4);
1703             for ($i=1;$i<3;$i++) {
1704                 &decstep(0,$tbl,$s0,$s3,$s2,$s1);
1705                 &decstep(1,$tbl,$s1,$s0,$s3,$s2);
1706                 &decstep(2,$tbl,$s2,$s1,$s0,$s3);
1707                 &decstep(3,$tbl,$s3,$s2,$s1,$s0);
1708                 &xor    ($s0,&DWP(16*$i+0,$key));
1709                 &xor    ($s1,&DWP(16*$i+4,$key));
1710                 &xor    ($s2,&DWP(16*$i+8,$key));
1711                 &xor    ($s3,&DWP(16*$i+12,$key));
1712             }
1713             &add        ($key,32);
1714             &mov        ($__key,$key);          # advance rd_key
1715         &set_label("12rounds",4);
1716             for ($i=1;$i<3;$i++) {
1717                 &decstep(0,$tbl,$s0,$s3,$s2,$s1);
1718                 &decstep(1,$tbl,$s1,$s0,$s3,$s2);
1719                 &decstep(2,$tbl,$s2,$s1,$s0,$s3);
1720                 &decstep(3,$tbl,$s3,$s2,$s1,$s0);
1721                 &xor    ($s0,&DWP(16*$i+0,$key));
1722                 &xor    ($s1,&DWP(16*$i+4,$key));
1723                 &xor    ($s2,&DWP(16*$i+8,$key));
1724                 &xor    ($s3,&DWP(16*$i+12,$key));
1725             }
1726             &add        ($key,32);
1727             &mov        ($__key,$key);          # advance rd_key
1728         &set_label("10rounds",4);
1729             for ($i=1;$i<10;$i++) {
1730                 &decstep(0,$tbl,$s0,$s3,$s2,$s1);
1731                 &decstep(1,$tbl,$s1,$s0,$s3,$s2);
1732                 &decstep(2,$tbl,$s2,$s1,$s0,$s3);
1733                 &decstep(3,$tbl,$s3,$s2,$s1,$s0);
1734                 &xor    ($s0,&DWP(16*$i+0,$key));
1735                 &xor    ($s1,&DWP(16*$i+4,$key));
1736                 &xor    ($s2,&DWP(16*$i+8,$key));
1737                 &xor    ($s3,&DWP(16*$i+12,$key));
1738             }
1739         }
1740
1741         &declast(0,$tbl,$s0,$s3,$s2,$s1);
1742         &declast(1,$tbl,$s1,$s0,$s3,$s2);
1743         &declast(2,$tbl,$s2,$s1,$s0,$s3);
1744         &declast(3,$tbl,$s3,$s2,$s1,$s0);
1745
1746         &add    ($key,$small_footprint?16:160);
1747         &xor    ($s0,&DWP(0,$key));
1748         &xor    ($s1,&DWP(4,$key));
1749         &xor    ($s2,&DWP(8,$key));
1750         &xor    ($s3,&DWP(12,$key));
1751
1752         &ret    ();
1753
1754 &set_label("AES_Td",64);        # Yes! I keep it in the code segment!
1755         &_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
1756         &_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
1757         &_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
1758         &_data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5);
1759         &_data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d);
1760         &_data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b);
1761         &_data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295);
1762         &_data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e);
1763         &_data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927);
1764         &_data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d);
1765         &_data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362);
1766         &_data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9);
1767         &_data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52);
1768         &_data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566);
1769         &_data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3);
1770         &_data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed);
1771         &_data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e);
1772         &_data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4);
1773         &_data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4);
1774         &_data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd);
1775         &_data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d);
1776         &_data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060);
1777         &_data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967);
1778         &_data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879);
1779         &_data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000);
1780         &_data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c);
1781         &_data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36);
1782         &_data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624);
1783         &_data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b);
1784         &_data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c);
1785         &_data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12);
1786         &_data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14);
1787         &_data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3);
1788         &_data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b);
1789         &_data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8);
1790         &_data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684);
1791         &_data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7);
1792         &_data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177);
1793         &_data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947);
1794         &_data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322);
1795         &_data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498);
1796         &_data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f);
1797         &_data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54);
1798         &_data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382);
1799         &_data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf);
1800         &_data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb);
1801         &_data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83);
1802         &_data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef);
1803         &_data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029);
1804         &_data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235);
1805         &_data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733);
1806         &_data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117);
1807         &_data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4);
1808         &_data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546);
1809         &_data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb);
1810         &_data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d);
1811         &_data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb);
1812         &_data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a);
1813         &_data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773);
1814         &_data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478);
1815         &_data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2);
1816         &_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
1817         &_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
1818         &_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
1819
1820 #Td4:   # four copies of Td4 to choose from to avoid L1 aliasing
1821         &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1822         &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1823         &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1824         &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1825         &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1826         &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1827         &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1828         &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1829         &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1830         &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1831         &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1832         &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1833         &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1834         &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1835         &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1836         &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1837         &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1838         &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1839         &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1840         &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1841         &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1842         &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1843         &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1844         &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1845         &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1846         &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1847         &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1848         &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1849         &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1850         &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1851         &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1852         &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1853
1854         &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1855         &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1856         &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1857         &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1858         &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1859         &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1860         &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1861         &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1862         &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1863         &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1864         &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1865         &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1866         &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1867         &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1868         &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1869         &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1870         &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1871         &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1872         &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1873         &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1874         &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1875         &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1876         &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1877         &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1878         &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1879         &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1880         &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1881         &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1882         &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1883         &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1884         &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1885         &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1886
1887         &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1888         &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1889         &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1890         &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1891         &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1892         &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1893         &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1894         &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1895         &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1896         &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1897         &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1898         &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1899         &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1900         &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1901         &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1902         &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1903         &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1904         &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1905         &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1906         &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1907         &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1908         &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1909         &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1910         &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1911         &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1912         &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1913         &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1914         &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1915         &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1916         &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1917         &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1918         &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1919
1920         &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
1921         &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
1922         &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
1923         &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
1924         &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
1925         &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
1926         &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
1927         &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
1928         &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
1929         &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
1930         &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
1931         &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
1932         &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
1933         &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
1934         &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
1935         &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
1936         &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
1937         &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
1938         &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
1939         &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
1940         &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
1941         &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
1942         &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
1943         &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
1944         &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
1945         &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
1946         &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
1947         &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
1948         &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
1949         &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
1950         &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
1951         &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
1952 &function_end_B("_x86_AES_decrypt");
1953
1954 # void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
1955 &public_label("AES_Td");
1956 &function_begin("AES_decrypt");
1957         &mov    ($acc,&wparam(0));              # load inp
1958         &mov    ($key,&wparam(2));              # load key
1959
1960         &mov    ($s0,"esp");
1961         &sub    ("esp",36);
1962         &and    ("esp",-64);                    # align to cache-line
1963
1964         # place stack frame just "above" the key schedule
1965         &lea    ($s1,&DWP(-64-63,$key));
1966         &sub    ($s1,"esp");
1967         &neg    ($s1);
1968         &and    ($s1,0x3C0);    # modulo 1024, but aligned to cache-line
1969         &sub    ("esp",$s1);
1970         &add    ("esp",4);      # 4 is reserved for caller's return address
1971         &mov    ($_esp,$s0);    # save stack pointer
1972
1973         &call   (&label("pic_point"));          # make it PIC!
1974         &set_label("pic_point");
1975         &blindpop($tbl);
1976         &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
1977         &lea    ($tbl,&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl));
1978
1979         # pick Td4 copy which can't "overlap" with stack frame or key schedule
1980         &lea    ($s1,&DWP(768-4,"esp"));
1981         &sub    ($s1,$tbl);
1982         &and    ($s1,0x300);
1983         &lea    ($tbl,&DWP(2048+128,$tbl,$s1));
1984
1985                                         if (!$x86only) {
1986         &bt     (&DWP(0,$s0),25);       # check for SSE bit
1987         &jnc    (&label("x86"));
1988
1989         &movq   ("mm0",&QWP(0,$acc));
1990         &movq   ("mm4",&QWP(8,$acc));
1991         &call   ("_sse_AES_decrypt_compact");
1992         &mov    ("esp",$_esp);                  # restore stack pointer
1993         &mov    ($acc,&wparam(1));              # load out
1994         &movq   (&QWP(0,$acc),"mm0");           # write output data
1995         &movq   (&QWP(8,$acc),"mm4");
1996         &emms   ();
1997         &function_end_A();
1998                                         }
1999         &set_label("x86",16);
2000         &mov    ($_tbl,$tbl);
2001         &mov    ($s0,&DWP(0,$acc));             # load input data
2002         &mov    ($s1,&DWP(4,$acc));
2003         &mov    ($s2,&DWP(8,$acc));
2004         &mov    ($s3,&DWP(12,$acc));
2005         &call   ("_x86_AES_decrypt_compact");
2006         &mov    ("esp",$_esp);                  # restore stack pointer
2007         &mov    ($acc,&wparam(1));              # load out
2008         &mov    (&DWP(0,$acc),$s0);             # write output data
2009         &mov    (&DWP(4,$acc),$s1);
2010         &mov    (&DWP(8,$acc),$s2);
2011         &mov    (&DWP(12,$acc),$s3);
2012 &function_end("AES_decrypt");
2013
2014 # void AES_cbc_encrypt (const void char *inp, unsigned char *out,
2015 #                       size_t length, const AES_KEY *key,
2016 #                       unsigned char *ivp,const int enc);
2017 {
2018 # stack frame layout
2019 #             -4(%esp)          # return address         0(%esp)
2020 #              0(%esp)          # s0 backing store       4(%esp)        
2021 #              4(%esp)          # s1 backing store       8(%esp)
2022 #              8(%esp)          # s2 backing store      12(%esp)
2023 #             12(%esp)          # s3 backing store      16(%esp)
2024 #             16(%esp)          # key backup            20(%esp)
2025 #             20(%esp)          # end of key schedule   24(%esp)
2026 #             24(%esp)          # %ebp backup           28(%esp)
2027 #             28(%esp)          # %esp backup
2028 my $_inp=&DWP(32,"esp");        # copy of wparam(0)
2029 my $_out=&DWP(36,"esp");        # copy of wparam(1)
2030 my $_len=&DWP(40,"esp");        # copy of wparam(2)
2031 my $_key=&DWP(44,"esp");        # copy of wparam(3)
2032 my $_ivp=&DWP(48,"esp");        # copy of wparam(4)
2033 my $_tmp=&DWP(52,"esp");        # volatile variable
2034 #
2035 my $ivec=&DWP(60,"esp");        # ivec[16]
2036 my $aes_key=&DWP(76,"esp");     # copy of aes_key
2037 my $mark=&DWP(76+240,"esp");    # copy of aes_key->rounds
2038
2039 &public_label("AES_Te");
2040 &public_label("AES_Td");
2041 &function_begin("AES_cbc_encrypt");
2042         &mov    ($s2 eq "ecx"? $s2 : "",&wparam(2));    # load len
2043         &cmp    ($s2,0);
2044         &je     (&label("drop_out"));
2045
2046         &call   (&label("pic_point"));          # make it PIC!
2047         &set_label("pic_point");
2048         &blindpop($tbl);
2049         &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
2050
2051         &cmp    (&wparam(5),0);
2052         &lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
2053         &jne    (&label("picked_te"));
2054         &lea    ($tbl,&DWP(&label("AES_Td")."-".&label("AES_Te"),$tbl));
2055         &set_label("picked_te");
2056
2057         # one can argue if this is required
2058         &pushf  ();
2059         &cld    ();
2060
2061         &cmp    ($s2,$speed_limit);
2062         &jb     (&label("slow_way"));
2063         &test   ($s2,15);
2064         &jnz    (&label("slow_way"));
2065                                         if (!$x86only) {
2066         &bt     (&DWP(0,$s0),28);       # check for hyper-threading bit
2067         &jc     (&label("slow_way"));
2068                                         }
2069         # pre-allocate aligned stack frame...
2070         &lea    ($acc,&DWP(-80-244,"esp"));
2071         &and    ($acc,-64);
2072
2073         # ... and make sure it doesn't alias with $tbl modulo 4096
2074         &mov    ($s0,$tbl);
2075         &lea    ($s1,&DWP(2048+256,$tbl));
2076         &mov    ($s3,$acc);
2077         &and    ($s0,0xfff);            # s = %ebp&0xfff
2078         &and    ($s1,0xfff);            # e = (%ebp+2048+256)&0xfff
2079         &and    ($s3,0xfff);            # p = %esp&0xfff
2080
2081         &cmp    ($s3,$s1);              # if (p>=e) %esp =- (p-e);
2082         &jb     (&label("tbl_break_out"));
2083         &sub    ($s3,$s1);
2084         &sub    ($acc,$s3);
2085         &jmp    (&label("tbl_ok"));
2086         &set_label("tbl_break_out",4);  # else %esp -= (p-s)&0xfff + framesz;
2087         &sub    ($s3,$s0);
2088         &and    ($s3,0xfff);
2089         &add    ($s3,384);
2090         &sub    ($acc,$s3);
2091         &set_label("tbl_ok",4);
2092
2093         &lea    ($s3,&wparam(0));       # obtain pointer to parameter block
2094         &exch   ("esp",$acc);           # allocate stack frame
2095         &add    ("esp",4);              # reserve for return address!
2096         &mov    ($_tbl,$tbl);           # save %ebp
2097         &mov    ($_esp,$acc);           # save %esp
2098
2099         &mov    ($s0,&DWP(0,$s3));      # load inp
2100         &mov    ($s1,&DWP(4,$s3));      # load out
2101         #&mov   ($s2,&DWP(8,$s3));      # load len
2102         &mov    ($key,&DWP(12,$s3));    # load key
2103         &mov    ($acc,&DWP(16,$s3));    # load ivp
2104         &mov    ($s3,&DWP(20,$s3));     # load enc flag
2105
2106         &mov    ($_inp,$s0);            # save copy of inp
2107         &mov    ($_out,$s1);            # save copy of out
2108         &mov    ($_len,$s2);            # save copy of len
2109         &mov    ($_key,$key);           # save copy of key
2110         &mov    ($_ivp,$acc);           # save copy of ivp
2111
2112         &mov    ($mark,0);              # copy of aes_key->rounds = 0;
2113         # do we copy key schedule to stack?
2114         &mov    ($s1 eq "ebx" ? $s1 : "",$key);
2115         &mov    ($s2 eq "ecx" ? $s2 : "",244/4);
2116         &sub    ($s1,$tbl);
2117         &mov    ("esi",$key);
2118         &and    ($s1,0xfff);
2119         &lea    ("edi",$aes_key);
2120         &cmp    ($s1,2048+256);
2121         &jb     (&label("do_copy"));
2122         &cmp    ($s1,4096-244);
2123         &jb     (&label("skip_copy"));
2124         &set_label("do_copy",4);
2125                 &mov    ($_key,"edi");
2126                 &data_word(0xA5F3F689); # rep movsd
2127         &set_label("skip_copy");
2128
2129         &mov    ($key,16);
2130         &set_label("prefetch_tbl",4);
2131                 &mov    ($s0,&DWP(0,$tbl));
2132                 &mov    ($s1,&DWP(32,$tbl));
2133                 &mov    ($s2,&DWP(64,$tbl));
2134                 &mov    ($acc,&DWP(96,$tbl));
2135                 &lea    ($tbl,&DWP(128,$tbl));
2136                 &sub    ($key,1);
2137         &jnz    (&label("prefetch_tbl"));
2138         &sub    ($tbl,2048);
2139
2140         &mov    ($acc,$_inp);
2141         &mov    ($key,$_ivp);
2142
2143         &cmp    ($s3,0);
2144         &je     (&label("fast_decrypt"));
2145
2146 #----------------------------- ENCRYPT -----------------------------#
2147         &mov    ($s0,&DWP(0,$key));             # load iv
2148         &mov    ($s1,&DWP(4,$key));
2149
2150         &set_label("fast_enc_loop",16);
2151                 &mov    ($s2,&DWP(8,$key));
2152                 &mov    ($s3,&DWP(12,$key));
2153
2154                 &xor    ($s0,&DWP(0,$acc));     # xor input data
2155                 &xor    ($s1,&DWP(4,$acc));
2156                 &xor    ($s2,&DWP(8,$acc));
2157                 &xor    ($s3,&DWP(12,$acc));
2158
2159                 &mov    ($key,$_key);           # load key
2160                 &call   ("_x86_AES_encrypt");
2161
2162                 &mov    ($acc,$_inp);           # load inp
2163                 &mov    ($key,$_out);           # load out
2164
2165                 &mov    (&DWP(0,$key),$s0);     # save output data
2166                 &mov    (&DWP(4,$key),$s1);
2167                 &mov    (&DWP(8,$key),$s2);
2168                 &mov    (&DWP(12,$key),$s3);
2169
2170                 &lea    ($acc,&DWP(16,$acc));   # advance inp
2171                 &mov    ($s2,$_len);            # load len
2172                 &mov    ($_inp,$acc);           # save inp
2173                 &lea    ($s3,&DWP(16,$key));    # advance out
2174                 &mov    ($_out,$s3);            # save out
2175                 &sub    ($s2,16);               # decrease len
2176                 &mov    ($_len,$s2);            # save len
2177         &jnz    (&label("fast_enc_loop"));
2178         &mov    ($acc,$_ivp);           # load ivp
2179         &mov    ($s2,&DWP(8,$key));     # restore last 2 dwords
2180         &mov    ($s3,&DWP(12,$key));
2181         &mov    (&DWP(0,$acc),$s0);     # save ivec
2182         &mov    (&DWP(4,$acc),$s1);
2183         &mov    (&DWP(8,$acc),$s2);
2184         &mov    (&DWP(12,$acc),$s3);
2185
2186         &cmp    ($mark,0);              # was the key schedule copied?
2187         &mov    ("edi",$_key);
2188         &je     (&label("skip_ezero"));
2189         # zero copy of key schedule
2190         &mov    ("ecx",240/4);
2191         &xor    ("eax","eax");
2192         &align  (4);
2193         &data_word(0xABF3F689); # rep stosd
2194         &set_label("skip_ezero")
2195         &mov    ("esp",$_esp);
2196         &popf   ();
2197     &set_label("drop_out");
2198         &function_end_A();
2199         &pushf  ();                     # kludge, never executed
2200
2201 #----------------------------- DECRYPT -----------------------------#
2202 &set_label("fast_decrypt",16);
2203
2204         &cmp    ($acc,$_out);
2205         &je     (&label("fast_dec_in_place"));  # in-place processing...
2206
2207         &mov    ($_tmp,$key);
2208
2209         &align  (4);
2210         &set_label("fast_dec_loop",16);
2211                 &mov    ($s0,&DWP(0,$acc));     # read input
2212                 &mov    ($s1,&DWP(4,$acc));
2213                 &mov    ($s2,&DWP(8,$acc));
2214                 &mov    ($s3,&DWP(12,$acc));
2215
2216                 &mov    ($key,$_key);           # load key
2217                 &call   ("_x86_AES_decrypt");
2218
2219                 &mov    ($key,$_tmp);           # load ivp
2220                 &mov    ($acc,$_len);           # load len
2221                 &xor    ($s0,&DWP(0,$key));     # xor iv
2222                 &xor    ($s1,&DWP(4,$key));
2223                 &xor    ($s2,&DWP(8,$key));
2224                 &xor    ($s3,&DWP(12,$key));
2225
2226                 &mov    ($key,$_out);           # load out
2227                 &mov    ($acc,$_inp);           # load inp
2228
2229                 &mov    (&DWP(0,$key),$s0);     # write output
2230                 &mov    (&DWP(4,$key),$s1);
2231                 &mov    (&DWP(8,$key),$s2);
2232                 &mov    (&DWP(12,$key),$s3);
2233
2234                 &mov    ($s2,$_len);            # load len
2235                 &mov    ($_tmp,$acc);           # save ivp
2236                 &lea    ($acc,&DWP(16,$acc));   # advance inp
2237                 &mov    ($_inp,$acc);           # save inp
2238                 &lea    ($key,&DWP(16,$key));   # advance out
2239                 &mov    ($_out,$key);           # save out
2240                 &sub    ($s2,16);               # decrease len
2241                 &mov    ($_len,$s2);            # save len
2242         &jnz    (&label("fast_dec_loop"));
2243         &mov    ($key,$_tmp);           # load temp ivp
2244         &mov    ($acc,$_ivp);           # load user ivp
2245         &mov    ($s0,&DWP(0,$key));     # load iv
2246         &mov    ($s1,&DWP(4,$key));
2247         &mov    ($s2,&DWP(8,$key));
2248         &mov    ($s3,&DWP(12,$key));
2249         &mov    (&DWP(0,$acc),$s0);     # copy back to user
2250         &mov    (&DWP(4,$acc),$s1);
2251         &mov    (&DWP(8,$acc),$s2);
2252         &mov    (&DWP(12,$acc),$s3);
2253         &jmp    (&label("fast_dec_out"));
2254
2255     &set_label("fast_dec_in_place",16);
2256         &set_label("fast_dec_in_place_loop");
2257                 &mov    ($s0,&DWP(0,$acc));     # read input
2258                 &mov    ($s1,&DWP(4,$acc));
2259                 &mov    ($s2,&DWP(8,$acc));
2260                 &mov    ($s3,&DWP(12,$acc));
2261
2262                 &lea    ($key,$ivec);
2263                 &mov    (&DWP(0,$key),$s0);     # copy to temp
2264                 &mov    (&DWP(4,$key),$s1);
2265                 &mov    (&DWP(8,$key),$s2);
2266                 &mov    (&DWP(12,$key),$s3);
2267
2268                 &mov    ($key,$_key);           # load key
2269                 &call   ("_x86_AES_decrypt");
2270
2271                 &mov    ($key,$_ivp);           # load ivp
2272                 &mov    ($acc,$_out);           # load out
2273                 &xor    ($s0,&DWP(0,$key));     # xor iv
2274                 &xor    ($s1,&DWP(4,$key));
2275                 &xor    ($s2,&DWP(8,$key));
2276                 &xor    ($s3,&DWP(12,$key));
2277
2278                 &mov    (&DWP(0,$acc),$s0);     # write output
2279                 &mov    (&DWP(4,$acc),$s1);
2280                 &mov    (&DWP(8,$acc),$s2);
2281                 &mov    (&DWP(12,$acc),$s3);
2282
2283                 &lea    ($acc,&DWP(16,$acc));   # advance out
2284                 &mov    ($_out,$acc);           # save out
2285
2286                 &lea    ($acc,$ivec);
2287                 &mov    ($s0,&DWP(0,$acc));     # read temp
2288                 &mov    ($s1,&DWP(4,$acc));
2289                 &mov    ($s2,&DWP(8,$acc));
2290                 &mov    ($s3,&DWP(12,$acc));
2291
2292                 &mov    (&DWP(0,$key),$s0);     # copy iv
2293                 &mov    (&DWP(4,$key),$s1);
2294                 &mov    (&DWP(8,$key),$s2);
2295                 &mov    (&DWP(12,$key),$s3);
2296
2297                 &mov    ($acc,$_inp);           # load inp
2298                 &mov    ($s2,$_len);            # load len
2299                 &lea    ($acc,&DWP(16,$acc));   # advance inp
2300                 &mov    ($_inp,$acc);           # save inp
2301                 &sub    ($s2,16);               # decrease len
2302                 &mov    ($_len,$s2);            # save len
2303         &jnz    (&label("fast_dec_in_place_loop"));
2304
2305     &set_label("fast_dec_out",4);
2306         &cmp    ($mark,0);              # was the key schedule copied?
2307         &mov    ("edi",$_key);
2308         &je     (&label("skip_dzero"));
2309         # zero copy of key schedule
2310         &mov    ("ecx",240/4);
2311         &xor    ("eax","eax");
2312         &align  (4);
2313         &data_word(0xABF3F689); # rep stosd
2314         &set_label("skip_dzero")
2315         &mov    ("esp",$_esp);
2316         &popf   ();
2317         &function_end_A();
2318         &pushf  ();                     # kludge, never executed
2319
2320 #--------------------------- SLOW ROUTINE ---------------------------#
2321 &set_label("slow_way",16);
2322
2323         &mov    ($s0,&DWP(0,$s0)) if (!$x86only);# load OPENSSL_ia32cap
2324         &mov    ($key,&wparam(3));      # load key
2325
2326         # pre-allocate aligned stack frame...
2327         &lea    ($acc,&DWP(-80,"esp"));
2328         &and    ($acc,-64);
2329
2330         # ... and make sure it doesn't alias with $key modulo 1024
2331         &lea    ($s1,&DWP(-80-63,$key));
2332         &sub    ($s1,$acc);
2333         &neg    ($s1);
2334         &and    ($s1,0x3C0);    # modulo 1024, but aligned to cache-line
2335         &sub    ($acc,$s1);
2336
2337         # pick S-box copy which can't overlap with stack frame or $key
2338         &lea    ($s1,&DWP(768,$acc));
2339         &sub    ($s1,$tbl);
2340         &and    ($s1,0x300);
2341         &lea    ($tbl,&DWP(2048+128,$tbl,$s1));
2342
2343         &lea    ($s3,&wparam(0));       # pointer to parameter block
2344
2345         &exch   ("esp",$acc);
2346         &add    ("esp",4);              # reserve for return address!
2347         &mov    ($_tbl,$tbl);           # save %ebp
2348         &mov    ($_esp,$acc);           # save %esp
2349         &mov    ($_tmp,$s0);            # save OPENSSL_ia32cap
2350
2351         &mov    ($s0,&DWP(0,$s3));      # load inp
2352         &mov    ($s1,&DWP(4,$s3));      # load out
2353         #&mov   ($s2,&DWP(8,$s3));      # load len
2354         #&mov   ($key,&DWP(12,$s3));    # load key
2355         &mov    ($acc,&DWP(16,$s3));    # load ivp
2356         &mov    ($s3,&DWP(20,$s3));     # load enc flag
2357
2358         &mov    ($_inp,$s0);            # save copy of inp
2359         &mov    ($_out,$s1);            # save copy of out
2360         &mov    ($_len,$s2);            # save copy of len
2361         &mov    ($_key,$key);           # save copy of key
2362         &mov    ($_ivp,$acc);           # save copy of ivp
2363
2364         &mov    ($key,$acc);
2365         &mov    ($acc,$s0);
2366
2367         &cmp    ($s3,0);
2368         &je     (&label("slow_decrypt"));
2369
2370 #--------------------------- SLOW ENCRYPT ---------------------------#
2371         &cmp    ($s2,16);
2372         &jb     (&label("slow_enc_tail"));
2373
2374                                         if (!$x86only) {
2375         &bt     ($_tmp,25);             # check for SSE bit
2376         &jnc    (&label("slow_enc_x86"));
2377
2378         &movq   ("mm0",&QWP(0,$key));   # load iv
2379         &movq   ("mm4",&QWP(8,$key));
2380
2381         &set_label("slow_enc_loop_sse",16);
2382                 &pxor   ("mm0",&QWP(0,$acc));   # xor input data
2383                 &pxor   ("mm4",&QWP(8,$acc));
2384
2385                 &mov    ($key,$_key);
2386                 &call   ("_sse_AES_encrypt_compact");
2387
2388                 &mov    ($acc,$_inp);           # load inp
2389                 &mov    ($key,$_out);           # load out
2390                 &mov    ($s2,$_len);            # load len
2391
2392                 &movq   (&QWP(0,$key),"mm0");   # save output data
2393                 &movq   (&QWP(8,$key),"mm4");
2394
2395                 &lea    ($acc,&DWP(16,$acc));   # advance inp
2396                 &mov    ($_inp,$acc);           # save inp
2397                 &lea    ($s3,&DWP(16,$key));    # advance out
2398                 &mov    ($_out,$s3);            # save out
2399                 &sub    ($s2,16);               # decrease len
2400                 &cmp    ($s2,16);
2401                 &mov    ($_len,$s2);            # save len
2402         &jae    (&label("slow_enc_loop_sse"));
2403         &test   ($s2,15);
2404         &jnz    (&label("slow_enc_tail"));
2405         &mov    ($acc,$_ivp);           # load ivp
2406         &movq   (&QWP(0,$acc),"mm0");   # save ivec
2407         &movq   (&QWP(8,$acc),"mm4");
2408         &emms   ();
2409         &mov    ("esp",$_esp);
2410         &popf   ();
2411         &function_end_A();
2412         &pushf  ();                     # kludge, never executed
2413                                         }
2414     &set_label("slow_enc_x86",16);
2415         &mov    ($s0,&DWP(0,$key));     # load iv
2416         &mov    ($s1,&DWP(4,$key));
2417
2418         &set_label("slow_enc_loop_x86",4);
2419                 &mov    ($s2,&DWP(8,$key));
2420                 &mov    ($s3,&DWP(12,$key));
2421
2422                 &xor    ($s0,&DWP(0,$acc));     # xor input data
2423                 &xor    ($s1,&DWP(4,$acc));
2424                 &xor    ($s2,&DWP(8,$acc));
2425                 &xor    ($s3,&DWP(12,$acc));
2426
2427                 &mov    ($key,$_key);           # load key
2428                 &call   ("_x86_AES_encrypt_compact");
2429
2430                 &mov    ($acc,$_inp);           # load inp
2431                 &mov    ($key,$_out);           # load out
2432
2433                 &mov    (&DWP(0,$key),$s0);     # save output data
2434                 &mov    (&DWP(4,$key),$s1);
2435                 &mov    (&DWP(8,$key),$s2);
2436                 &mov    (&DWP(12,$key),$s3);
2437
2438                 &mov    ($s2,$_len);            # load len
2439                 &lea    ($acc,&DWP(16,$acc));   # advance inp
2440                 &mov    ($_inp,$acc);           # save inp
2441                 &lea    ($s3,&DWP(16,$key));    # advance out
2442                 &mov    ($_out,$s3);            # save out
2443                 &sub    ($s2,16);               # decrease len
2444                 &cmp    ($s2,16);
2445                 &mov    ($_len,$s2);            # save len
2446         &jae    (&label("slow_enc_loop_x86"));
2447         &test   ($s2,15);
2448         &jnz    (&label("slow_enc_tail"));
2449         &mov    ($acc,$_ivp);           # load ivp
2450         &mov    ($s2,&DWP(8,$key));     # restore last dwords
2451         &mov    ($s3,&DWP(12,$key));
2452         &mov    (&DWP(0,$acc),$s0);     # save ivec
2453         &mov    (&DWP(4,$acc),$s1);
2454         &mov    (&DWP(8,$acc),$s2);
2455         &mov    (&DWP(12,$acc),$s3);
2456
2457         &mov    ("esp",$_esp);
2458         &popf   ();
2459         &function_end_A();
2460         &pushf  ();                     # kludge, never executed
2461
2462     &set_label("slow_enc_tail",16);
2463         &emms   ()      if (!$x86only);
2464         &mov    ($key eq "edi"? $key:"",$s3);   # load out to edi
2465         &mov    ($s1,16);
2466         &sub    ($s1,$s2);
2467         &cmp    ($key,$acc eq "esi"? $acc:"");  # compare with inp
2468         &je     (&label("enc_in_place"));
2469         &align  (4);
2470         &data_word(0xA4F3F689); # rep movsb     # copy input
2471         &jmp    (&label("enc_skip_in_place"));
2472     &set_label("enc_in_place");
2473         &lea    ($key,&DWP(0,$key,$s2));
2474     &set_label("enc_skip_in_place");
2475         &mov    ($s2,$s1);
2476         &xor    ($s0,$s0);
2477         &align  (4);
2478         &data_word(0xAAF3F689); # rep stosb     # zero tail
2479
2480         &lea    ($key,&DWP(-16,$s3));           # restore ivp
2481         &mov    ($acc,$s3);                     # output as input
2482         &mov    ($s0,&DWP(0,$key));
2483         &mov    ($s1,&DWP(4,$key));
2484         &mov    ($_len,16);                     # len=16
2485         &jmp    (&label("slow_enc_loop_x86"));  # one more spin...
2486
2487 #--------------------------- SLOW DECRYPT ---------------------------#
2488 &set_label("slow_decrypt",16);
2489                                         if (!$x86only) {
2490         &bt     ($_tmp,25);             # check for SSE bit
2491         &jnc    (&label("slow_dec_loop_x86"));
2492
2493         &set_label("slow_dec_loop_sse",4);
2494                 &movq   ("mm0",&QWP(0,$acc));   # read input
2495                 &movq   ("mm4",&QWP(8,$acc));
2496
2497                 &mov    ($key,$_key);
2498                 &call   ("_sse_AES_decrypt_compact");
2499
2500                 &mov    ($acc,$_inp);           # load inp
2501                 &lea    ($s0,$ivec);
2502                 &mov    ($s1,$_out);            # load out
2503                 &mov    ($s2,$_len);            # load len
2504                 &mov    ($key,$_ivp);           # load ivp
2505
2506                 &movq   ("mm1",&QWP(0,$acc));   # re-read input
2507                 &movq   ("mm5",&QWP(8,$acc));
2508
2509                 &pxor   ("mm0",&QWP(0,$key));   # xor iv
2510                 &pxor   ("mm4",&QWP(8,$key));
2511
2512                 &movq   (&QWP(0,$key),"mm1");   # copy input to iv
2513                 &movq   (&QWP(8,$key),"mm5");
2514
2515                 &sub    ($s2,16);               # decrease len
2516                 &jc     (&label("slow_dec_partial_sse"));
2517
2518                 &movq   (&QWP(0,$s1),"mm0");    # write output
2519                 &movq   (&QWP(8,$s1),"mm4");
2520
2521                 &lea    ($s1,&DWP(16,$s1));     # advance out
2522                 &mov    ($_out,$s1);            # save out
2523                 &lea    ($acc,&DWP(16,$acc));   # advance inp
2524                 &mov    ($_inp,$acc);           # save inp
2525                 &mov    ($_len,$s2);            # save len
2526         &jnz    (&label("slow_dec_loop_sse"));
2527         &emms   ();
2528         &mov    ("esp",$_esp);
2529         &popf   ();
2530         &function_end_A();
2531         &pushf  ();                     # kludge, never executed
2532
2533     &set_label("slow_dec_partial_sse",16);
2534         &movq   (&QWP(0,$s0),"mm0");    # save output to temp
2535         &movq   (&QWP(8,$s0),"mm4");
2536         &emms   ();
2537
2538         &add    ($s2 eq "ecx" ? "ecx":"",16);
2539         &mov    ("edi",$s1);            # out
2540         &mov    ("esi",$s0);            # temp
2541         &align  (4);
2542         &data_word(0xA4F3F689);         # rep movsb # copy partial output
2543
2544         &mov    ("esp",$_esp);
2545         &popf   ();
2546         &function_end_A();
2547         &pushf  ();                     # kludge, never executed
2548                                         }
2549         &set_label("slow_dec_loop_x86",16);
2550                 &mov    ($s0,&DWP(0,$acc));     # read input
2551                 &mov    ($s1,&DWP(4,$acc));
2552                 &mov    ($s2,&DWP(8,$acc));
2553                 &mov    ($s3,&DWP(12,$acc));
2554
2555                 &lea    ($key,$ivec);
2556                 &mov    (&DWP(0,$key),$s0);     # copy to temp
2557                 &mov    (&DWP(4,$key),$s1);
2558                 &mov    (&DWP(8,$key),$s2);
2559                 &mov    (&DWP(12,$key),$s3);
2560
2561                 &mov    ($key,$_key);           # load key
2562                 &call   ("_x86_AES_decrypt_compact");
2563
2564                 &mov    ($key,$_ivp);           # load ivp
2565                 &mov    ($acc,$_len);           # load len
2566                 &xor    ($s0,&DWP(0,$key));     # xor iv
2567                 &xor    ($s1,&DWP(4,$key));
2568                 &xor    ($s2,&DWP(8,$key));
2569                 &xor    ($s3,&DWP(12,$key));
2570
2571                 &sub    ($acc,16);
2572                 &jc     (&label("slow_dec_partial_x86"));
2573
2574                 &mov    ($_len,$acc);           # save len
2575                 &mov    ($acc,$_out);           # load out
2576
2577                 &mov    (&DWP(0,$acc),$s0);     # write output
2578                 &mov    (&DWP(4,$acc),$s1);
2579                 &mov    (&DWP(8,$acc),$s2);
2580                 &mov    (&DWP(12,$acc),$s3);
2581
2582                 &lea    ($acc,&DWP(16,$acc));   # advance out
2583                 &mov    ($_out,$acc);           # save out
2584
2585                 &lea    ($acc,$ivec);
2586                 &mov    ($s0,&DWP(0,$acc));     # read temp
2587                 &mov    ($s1,&DWP(4,$acc));
2588                 &mov    ($s2,&DWP(8,$acc));
2589                 &mov    ($s3,&DWP(12,$acc));
2590
2591                 &mov    (&DWP(0,$key),$s0);     # copy it to iv
2592                 &mov    (&DWP(4,$key),$s1);
2593                 &mov    (&DWP(8,$key),$s2);
2594                 &mov    (&DWP(12,$key),$s3);
2595
2596                 &mov    ($acc,$_inp);           # load inp
2597                 &lea    ($acc,&DWP(16,$acc));   # advance inp
2598                 &mov    ($_inp,$acc);           # save inp
2599         &jnz    (&label("slow_dec_loop_x86"));
2600         &mov    ("esp",$_esp);
2601         &popf   ();
2602         &function_end_A();
2603         &pushf  ();                     # kludge, never executed
2604
2605     &set_label("slow_dec_partial_x86",16);
2606         &lea    ($acc,$ivec);
2607         &mov    (&DWP(0,$acc),$s0);     # save output to temp
2608         &mov    (&DWP(4,$acc),$s1);
2609         &mov    (&DWP(8,$acc),$s2);
2610         &mov    (&DWP(12,$acc),$s3);
2611
2612         &mov    ($acc,$_inp);
2613         &mov    ($s0,&DWP(0,$acc));     # re-read input
2614         &mov    ($s1,&DWP(4,$acc));
2615         &mov    ($s2,&DWP(8,$acc));
2616         &mov    ($s3,&DWP(12,$acc));
2617
2618         &mov    (&DWP(0,$key),$s0);     # copy it to iv
2619         &mov    (&DWP(4,$key),$s1);
2620         &mov    (&DWP(8,$key),$s2);
2621         &mov    (&DWP(12,$key),$s3);
2622
2623         &mov    ("ecx",$_len);
2624         &mov    ("edi",$_out);
2625         &lea    ("esi",$ivec);
2626         &align  (4);
2627         &data_word(0xA4F3F689);         # rep movsb # copy partial output
2628
2629         &mov    ("esp",$_esp);
2630         &popf   ();
2631 &function_end("AES_cbc_encrypt");
2632 }
2633
2634 #------------------------------------------------------------------#
2635
2636 sub enckey()
2637 {
2638         &movz   ("esi",&LB("edx"));             # rk[i]>>0
2639         &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2640         &movz   ("esi",&HB("edx"));             # rk[i]>>8
2641         &shl    ("ebx",24);
2642         &xor    ("eax","ebx");
2643
2644         &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2645         &shr    ("edx",16);
2646         &movz   ("esi",&LB("edx"));             # rk[i]>>16
2647         &xor    ("eax","ebx");
2648
2649         &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2650         &movz   ("esi",&HB("edx"));             # rk[i]>>24
2651         &shl    ("ebx",8);
2652         &xor    ("eax","ebx");
2653
2654         &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2655         &shl    ("ebx",16);
2656         &xor    ("eax","ebx");
2657
2658         &xor    ("eax",&DWP(1024-128,$tbl,"ecx",4));    # rcon
2659 }
2660
2661 # int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
2662 #                        AES_KEY *key)
2663 &public_label("AES_Te");
2664 &function_begin("AES_set_encrypt_key");
2665         &mov    ("esi",&wparam(0));             # user supplied key
2666         &mov    ("edi",&wparam(2));             # private key schedule
2667
2668         &test   ("esi",-1);
2669         &jz     (&label("badpointer"));
2670         &test   ("edi",-1);
2671         &jz     (&label("badpointer"));
2672
2673         &call   (&label("pic_point"));
2674         &set_label("pic_point");
2675         &blindpop($tbl);
2676         &lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
2677         &lea    ($tbl,&DWP(2048+128,$tbl));
2678
2679         # prefetch Te4
2680         &mov    ("eax",&DWP(0-128,$tbl));
2681         &mov    ("ebx",&DWP(32-128,$tbl));
2682         &mov    ("ecx",&DWP(64-128,$tbl));
2683         &mov    ("edx",&DWP(96-128,$tbl));
2684         &mov    ("eax",&DWP(128-128,$tbl));
2685         &mov    ("ebx",&DWP(160-128,$tbl));
2686         &mov    ("ecx",&DWP(192-128,$tbl));
2687         &mov    ("edx",&DWP(224-128,$tbl));
2688
2689         &mov    ("ecx",&wparam(1));             # number of bits in key
2690         &cmp    ("ecx",128);
2691         &je     (&label("10rounds"));
2692         &cmp    ("ecx",192);
2693         &je     (&label("12rounds"));
2694         &cmp    ("ecx",256);
2695         &je     (&label("14rounds"));
2696         &mov    ("eax",-2);                     # invalid number of bits
2697         &jmp    (&label("exit"));
2698
2699     &set_label("10rounds");
2700         &mov    ("eax",&DWP(0,"esi"));          # copy first 4 dwords
2701         &mov    ("ebx",&DWP(4,"esi"));
2702         &mov    ("ecx",&DWP(8,"esi"));
2703         &mov    ("edx",&DWP(12,"esi"));
2704         &mov    (&DWP(0,"edi"),"eax");
2705         &mov    (&DWP(4,"edi"),"ebx");
2706         &mov    (&DWP(8,"edi"),"ecx");
2707         &mov    (&DWP(12,"edi"),"edx");
2708
2709         &xor    ("ecx","ecx");
2710         &jmp    (&label("10shortcut"));
2711
2712         &align  (4);
2713         &set_label("10loop");
2714                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
2715                 &mov    ("edx",&DWP(12,"edi"));         # rk[3]
2716         &set_label("10shortcut");
2717                 &enckey ();
2718
2719                 &mov    (&DWP(16,"edi"),"eax");         # rk[4]
2720                 &xor    ("eax",&DWP(4,"edi"));
2721                 &mov    (&DWP(20,"edi"),"eax");         # rk[5]
2722                 &xor    ("eax",&DWP(8,"edi"));
2723                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
2724                 &xor    ("eax",&DWP(12,"edi"));
2725                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
2726                 &inc    ("ecx");
2727                 &add    ("edi",16);
2728                 &cmp    ("ecx",10);
2729         &jl     (&label("10loop"));
2730
2731         &mov    (&DWP(80,"edi"),10);            # setup number of rounds
2732         &xor    ("eax","eax");
2733         &jmp    (&label("exit"));
2734                 
2735     &set_label("12rounds");
2736         &mov    ("eax",&DWP(0,"esi"));          # copy first 6 dwords
2737         &mov    ("ebx",&DWP(4,"esi"));
2738         &mov    ("ecx",&DWP(8,"esi"));
2739         &mov    ("edx",&DWP(12,"esi"));
2740         &mov    (&DWP(0,"edi"),"eax");
2741         &mov    (&DWP(4,"edi"),"ebx");
2742         &mov    (&DWP(8,"edi"),"ecx");
2743         &mov    (&DWP(12,"edi"),"edx");
2744         &mov    ("ecx",&DWP(16,"esi"));
2745         &mov    ("edx",&DWP(20,"esi"));
2746         &mov    (&DWP(16,"edi"),"ecx");
2747         &mov    (&DWP(20,"edi"),"edx");
2748
2749         &xor    ("ecx","ecx");
2750         &jmp    (&label("12shortcut"));
2751
2752         &align  (4);
2753         &set_label("12loop");
2754                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
2755                 &mov    ("edx",&DWP(20,"edi"));         # rk[5]
2756         &set_label("12shortcut");
2757                 &enckey ();
2758
2759                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
2760                 &xor    ("eax",&DWP(4,"edi"));
2761                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
2762                 &xor    ("eax",&DWP(8,"edi"));
2763                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
2764                 &xor    ("eax",&DWP(12,"edi"));
2765                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
2766
2767                 &cmp    ("ecx",7);
2768                 &je     (&label("12break"));
2769                 &inc    ("ecx");
2770
2771                 &xor    ("eax",&DWP(16,"edi"));
2772                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
2773                 &xor    ("eax",&DWP(20,"edi"));
2774                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
2775
2776                 &add    ("edi",24);
2777         &jmp    (&label("12loop"));
2778
2779         &set_label("12break");
2780         &mov    (&DWP(72,"edi"),12);            # setup number of rounds
2781         &xor    ("eax","eax");
2782         &jmp    (&label("exit"));
2783
2784     &set_label("14rounds");
2785         &mov    ("eax",&DWP(0,"esi"));          # copy first 8 dwords
2786         &mov    ("ebx",&DWP(4,"esi"));
2787         &mov    ("ecx",&DWP(8,"esi"));
2788         &mov    ("edx",&DWP(12,"esi"));
2789         &mov    (&DWP(0,"edi"),"eax");
2790         &mov    (&DWP(4,"edi"),"ebx");
2791         &mov    (&DWP(8,"edi"),"ecx");
2792         &mov    (&DWP(12,"edi"),"edx");
2793         &mov    ("eax",&DWP(16,"esi"));
2794         &mov    ("ebx",&DWP(20,"esi"));
2795         &mov    ("ecx",&DWP(24,"esi"));
2796         &mov    ("edx",&DWP(28,"esi"));
2797         &mov    (&DWP(16,"edi"),"eax");
2798         &mov    (&DWP(20,"edi"),"ebx");
2799         &mov    (&DWP(24,"edi"),"ecx");
2800         &mov    (&DWP(28,"edi"),"edx");
2801
2802         &xor    ("ecx","ecx");
2803         &jmp    (&label("14shortcut"));
2804
2805         &align  (4);
2806         &set_label("14loop");
2807                 &mov    ("edx",&DWP(28,"edi"));         # rk[7]
2808         &set_label("14shortcut");
2809                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
2810
2811                 &enckey ();
2812
2813                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
2814                 &xor    ("eax",&DWP(4,"edi"));
2815                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
2816                 &xor    ("eax",&DWP(8,"edi"));
2817                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
2818                 &xor    ("eax",&DWP(12,"edi"));
2819                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
2820
2821                 &cmp    ("ecx",6);
2822                 &je     (&label("14break"));
2823                 &inc    ("ecx");
2824
2825                 &mov    ("edx","eax");
2826                 &mov    ("eax",&DWP(16,"edi"));         # rk[4]
2827                 &movz   ("esi",&LB("edx"));             # rk[11]>>0
2828                 &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2829                 &movz   ("esi",&HB("edx"));             # rk[11]>>8
2830                 &xor    ("eax","ebx");
2831
2832                 &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2833                 &shr    ("edx",16);
2834                 &shl    ("ebx",8);
2835                 &movz   ("esi",&LB("edx"));             # rk[11]>>16
2836                 &xor    ("eax","ebx");
2837
2838                 &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2839                 &movz   ("esi",&HB("edx"));             # rk[11]>>24
2840                 &shl    ("ebx",16);
2841                 &xor    ("eax","ebx");
2842
2843                 &movz   ("ebx",&BP(-128,$tbl,"esi",1));
2844                 &shl    ("ebx",24);
2845                 &xor    ("eax","ebx");
2846
2847                 &mov    (&DWP(48,"edi"),"eax");         # rk[12]
2848                 &xor    ("eax",&DWP(20,"edi"));
2849                 &mov    (&DWP(52,"edi"),"eax");         # rk[13]
2850                 &xor    ("eax",&DWP(24,"edi"));
2851                 &mov    (&DWP(56,"edi"),"eax");         # rk[14]
2852                 &xor    ("eax",&DWP(28,"edi"));
2853                 &mov    (&DWP(60,"edi"),"eax");         # rk[15]
2854
2855                 &add    ("edi",32);
2856         &jmp    (&label("14loop"));
2857
2858         &set_label("14break");
2859         &mov    (&DWP(48,"edi"),14);            # setup number of rounds
2860         &xor    ("eax","eax");
2861         &jmp    (&label("exit"));
2862
2863     &set_label("badpointer");
2864         &mov    ("eax",-1);
2865     &set_label("exit");
2866 &function_end("AES_set_encrypt_key");
2867
2868 sub deckey()
2869 { my ($i,$key,$tp1,$tp2,$tp4,$tp8) = @_;
2870   my $tmp = $tbl;
2871
2872         &mov    ($acc,$tp1);
2873         &and    ($acc,0x80808080);
2874         &mov    ($tmp,$acc);
2875         &mov    ($tp2,$tp1);
2876         &shr    ($tmp,7);
2877         &and    ($tp2,0x7f7f7f7f);
2878         &sub    ($acc,$tmp);
2879         &add    ($tp2,$tp2);
2880         &and    ($acc,0x1b1b1b1b);
2881         &xor    ($acc,$tp2);
2882         &mov    ($tp2,$acc);
2883
2884         &and    ($acc,0x80808080);
2885         &mov    ($tmp,$acc);
2886         &mov    ($tp4,$tp2);
2887          &xor   ($tp2,$tp1);    # tp2^tp1
2888         &shr    ($tmp,7);
2889         &and    ($tp4,0x7f7f7f7f);
2890         &sub    ($acc,$tmp);
2891         &add    ($tp4,$tp4);
2892         &and    ($acc,0x1b1b1b1b);
2893         &xor    ($acc,$tp4);
2894         &mov    ($tp4,$acc);
2895
2896         &and    ($acc,0x80808080);
2897         &mov    ($tmp,$acc);
2898         &mov    ($tp8,$tp4);
2899          &xor   ($tp4,$tp1);    # tp4^tp1
2900         &shr    ($tmp,7);
2901         &and    ($tp8,0x7f7f7f7f);
2902         &sub    ($acc,$tmp);
2903         &add    ($tp8,$tp8);
2904         &and    ($acc,0x1b1b1b1b);
2905          &rotl  ($tp1,8);       # = ROTATE(tp1,8)
2906         &xor    ($tp8,$acc);
2907
2908         &mov    ($tmp,&DWP(4*($i+1),$key));     # modulo-scheduled load
2909
2910         &xor    ($tp1,$tp2);
2911         &xor    ($tp2,$tp8);
2912         &xor    ($tp1,$tp4);
2913         &rotl   ($tp2,24);
2914         &xor    ($tp4,$tp8);
2915         &xor    ($tp1,$tp8);    # ^= tp8^(tp4^tp1)^(tp2^tp1)
2916         &rotl   ($tp4,16);
2917         &xor    ($tp1,$tp2);    # ^= ROTATE(tp8^tp2^tp1,24)
2918         &rotl   ($tp8,8);
2919         &xor    ($tp1,$tp4);    # ^= ROTATE(tp8^tp4^tp1,16)
2920         &mov    ($tp2,$tmp);
2921         &xor    ($tp1,$tp8);    # ^= ROTATE(tp8,8)
2922
2923         &mov    (&DWP(4*$i,$key),$tp1);
2924 }
2925
2926 # int AES_set_decrypt_key(const unsigned char *userKey, const int bits,
2927 #                        AES_KEY *key)
2928 &public_label("AES_Td");
2929 &public_label("AES_Te");
2930 &function_begin_B("AES_set_decrypt_key");
2931         &mov    ("eax",&wparam(0));
2932         &mov    ("ecx",&wparam(1));
2933         &mov    ("edx",&wparam(2));
2934         &sub    ("esp",12);
2935         &mov    (&DWP(0,"esp"),"eax");
2936         &mov    (&DWP(4,"esp"),"ecx");
2937         &mov    (&DWP(8,"esp"),"edx");
2938         &call   ("AES_set_encrypt_key");
2939         &add    ("esp",12);
2940         &cmp    ("eax",0);
2941         &je     (&label("proceed"));
2942         &ret    ();
2943
2944     &set_label("proceed");
2945         &push   ("ebp");
2946         &push   ("ebx");
2947         &push   ("esi");
2948         &push   ("edi");
2949
2950         &mov    ("esi",&wparam(2));
2951         &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
2952         &lea    ("ecx",&DWP(0,"","ecx",4));
2953         &lea    ("edi",&DWP(0,"esi","ecx",4));  # pointer to last chunk
2954
2955         &set_label("invert",4);                 # invert order of chunks
2956                 &mov    ("eax",&DWP(0,"esi"));
2957                 &mov    ("ebx",&DWP(4,"esi"));
2958                 &mov    ("ecx",&DWP(0,"edi"));
2959                 &mov    ("edx",&DWP(4,"edi"));
2960                 &mov    (&DWP(0,"edi"),"eax");
2961                 &mov    (&DWP(4,"edi"),"ebx");
2962                 &mov    (&DWP(0,"esi"),"ecx");
2963                 &mov    (&DWP(4,"esi"),"edx");
2964                 &mov    ("eax",&DWP(8,"esi"));
2965                 &mov    ("ebx",&DWP(12,"esi"));
2966                 &mov    ("ecx",&DWP(8,"edi"));
2967                 &mov    ("edx",&DWP(12,"edi"));
2968                 &mov    (&DWP(8,"edi"),"eax");
2969                 &mov    (&DWP(12,"edi"),"ebx");
2970                 &mov    (&DWP(8,"esi"),"ecx");
2971                 &mov    (&DWP(12,"esi"),"edx");
2972                 &add    ("esi",16);
2973                 &sub    ("edi",16);
2974                 &cmp    ("esi","edi");
2975         &jne    (&label("invert"));
2976
2977         &mov    ($key,&wparam(2));
2978         &mov    ($acc,&DWP(240,$key));          # pull number of rounds
2979         &lea    ($acc,&DWP(-2,$acc,$acc));
2980         &lea    ($acc,&DWP(0,$key,$acc,8));
2981         &mov    (&wparam(2),$acc);
2982
2983         &mov    ($s0,&DWP(16,$key));            # modulo-scheduled load
2984         &set_label("permute",4);                # permute the key schedule
2985                 &add    ($key,16);
2986                 &deckey (0,$key,$s0,$s1,$s2,$s3);
2987                 &deckey (1,$key,$s1,$s2,$s3,$s0);
2988                 &deckey (2,$key,$s2,$s3,$s0,$s1);
2989                 &deckey (3,$key,$s3,$s0,$s1,$s2);
2990                 &cmp    ($key,&wparam(2));
2991         &jb     (&label("permute"));
2992
2993         &xor    ("eax","eax");                  # return success
2994 &function_end("AES_set_decrypt_key");
2995
2996 &asm_finish();