1.. SPDX-License-Identifier: GPL-2.0
2.. include:: <isonum.txt>
3
4===========================================
5Fast & Portable DES encryption & decryption
6===========================================
7
8.. note::
9
10   Below is the original README file from the descore.shar package,
11   converted to ReST format.
12
13------------------------------------------------------------------------------
14
15des - fast & portable DES encryption & decryption.
16
17Copyright |copy| 1992  Dana L. How
18
19This program is free software; you can redistribute it and/or modify
20it under the terms of the GNU Library General Public License as published by
21the Free Software Foundation; either version 2 of the License, or
22(at your option) any later version.
23
24This program is distributed in the hope that it will be useful,
25but WITHOUT ANY WARRANTY; without even the implied warranty of
26MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
27GNU Library General Public License for more details.
28
29You should have received a copy of the GNU Library General Public License
30along with this program; if not, write to the Free Software
31Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32
33Author's address: how@isl.stanford.edu
34
35.. README,v 1.15 1992/05/20 00:25:32 how E
36
37==>> To compile after untarring/unsharring, just ``make`` <<==
38
39This package was designed with the following goals:
40
411.	Highest possible encryption/decryption PERFORMANCE.
422.	PORTABILITY to any byte-addressable host with a 32bit unsigned C type
433.	Plug-compatible replacement for KERBEROS's low-level routines.
44
45This second release includes a number of performance enhancements for
46register-starved machines.  My discussions with Richard Outerbridge,
4771755.204@compuserve.com, sparked a number of these enhancements.
48
49To more rapidly understand the code in this package, inspect desSmallFips.i
50(created by typing ``make``) BEFORE you tackle desCode.h.  The latter is set
51up in a parameterized fashion so it can easily be modified by speed-daemon
52hackers in pursuit of that last microsecond.  You will find it more
53illuminating to inspect one specific implementation,
54and then move on to the common abstract skeleton with this one in mind.
55
56
57performance comparison to other available des code which i could
58compile on a SPARCStation 1 (cc -O4, gcc -O2):
59
60this code (byte-order independent):
61
62  - 30us per encryption (options: 64k tables, no IP/FP)
63  - 33us per encryption (options: 64k tables, FIPS standard bit ordering)
64  - 45us per encryption (options:  2k tables, no IP/FP)
65  - 48us per encryption (options:  2k tables, FIPS standard bit ordering)
66  - 275us to set a new key (uses 1k of key tables)
67
68	this has the quickest encryption/decryption routines i've seen.
69	since i was interested in fast des filters rather than crypt(3)
70	and password cracking, i haven't really bothered yet to speed up
71	the key setting routine. also, i have no interest in re-implementing
72	all the other junk in the mit kerberos des library, so i've just
73	provided my routines with little stub interfaces so they can be
74	used as drop-in replacements with mit's code or any of the mit-
75	compatible packages below. (note that the first two timings above
76	are highly variable because of cache effects).
77
78kerberos des replacement from australia (version 1.95):
79
80  - 53us per encryption (uses 2k of tables)
81  - 96us to set a new key (uses 2.25k of key tables)
82
83	so despite the author's inclusion of some of the performance
84	improvements i had suggested to him, this package's
85	encryption/decryption is still slower on the sparc and 68000.
86	more specifically, 19-40% slower on the 68020 and 11-35% slower
87	on the sparc,  depending on the compiler;
88	in full gory detail (ALT_ECB is a libdes variant):
89
90	===============	==============	===============	=================
91	compiler   	machine		desCore	libdes	ALT_ECB	slower by
92	===============	==============	===============	=================
93	gcc 2.1 -O2	Sun 3/110	304  uS	369.5uS	461.8uS	 22%
94	cc      -O1	Sun 3/110	336  uS	436.6uS	399.3uS	 19%
95	cc      -O2	Sun 3/110	360  uS	532.4uS	505.1uS	 40%
96	cc      -O4	Sun 3/110	365  uS	532.3uS	505.3uS	 38%
97	gcc 2.1 -O2	Sun 4/50	 48  uS	 53.4uS	 57.5uS	 11%
98	cc      -O2	Sun 4/50	 48  uS	 64.6uS	 64.7uS	 35%
99	cc      -O4	Sun 4/50	 48  uS	 64.7uS	 64.9uS	 35%
100	===============	==============	===============	=================
101
102	(my time measurements are not as accurate as his).
103
104   the comments in my first release of desCore on version 1.92:
105
106   - 68us per encryption (uses 2k of tables)
107   - 96us to set a new key (uses 2.25k of key tables)
108
109	this is a very nice package which implements the most important
110	of the optimizations which i did in my encryption routines.
111	it's a bit weak on common low-level optimizations which is why
112	it's 39%-106% slower.  because he was interested in fast crypt(3) and
113	password-cracking applications,  he also used the same ideas to
114	speed up the key-setting routines with impressive results.
115	(at some point i may do the same in my package).  he also implements
116	the rest of the mit des library.
117
118	(code from eay@psych.psy.uq.oz.au via comp.sources.misc)
119
120fast crypt(3) package from denmark:
121
122	the des routine here is buried inside a loop to do the
123	crypt function and i didn't feel like ripping it out and measuring
124	performance. his code takes 26 sparc instructions to compute one
125	des iteration; above, Quick (64k) takes 21 and Small (2k) takes 37.
126	he claims to use 280k of tables but the iteration calculation seems
127	to use only 128k.  his tables and code are machine independent.
128
129	(code from glad@daimi.aau.dk via alt.sources or comp.sources.misc)
130
131swedish reimplementation of Kerberos des library
132
133  - 108us per encryption (uses 34k worth of tables)
134  - 134us to set a new key (uses 32k of key tables to get this speed!)
135
136	the tables used seem to be machine-independent;
137	he seems to have included a lot of special case code
138	so that, e.g., ``long`` loads can be used instead of 4 ``char`` loads
139	when the machine's architecture allows it.
140
141	(code obtained from chalmers.se:pub/des)
142
143crack 3.3c package from england:
144
145	as in crypt above, the des routine is buried in a loop. it's
146	also very modified for crypt.  his iteration code uses 16k
147	of tables and appears to be slow.
148
149	(code obtained from aem@aber.ac.uk via alt.sources or comp.sources.misc)
150
151``highly optimized`` and tweaked Kerberos/Athena code (byte-order dependent):
152
153  - 165us per encryption (uses 6k worth of tables)
154  - 478us to set a new key (uses <1k of key tables)
155
156	so despite the comments in this code, it was possible to get
157	faster code AND smaller tables, as well as making the tables
158	machine-independent.
159	(code obtained from prep.ai.mit.edu)
160
161UC Berkeley code (depends on machine-endedness):
162  -  226us per encryption
163  - 10848us to set a new key
164
165	table sizes are unclear, but they don't look very small
166	(code obtained from wuarchive.wustl.edu)
167
168
169motivation and history
170======================
171
172a while ago i wanted some des routines and the routines documented on sun's
173man pages either didn't exist or dumped core.  i had heard of kerberos,
174and knew that it used des,  so i figured i'd use its routines.  but once
175i got it and looked at the code,  it really set off a lot of pet peeves -
176it was too convoluted, the code had been written without taking
177advantage of the regular structure of operations such as IP, E, and FP
178(i.e. the author didn't sit down and think before coding),
179it was excessively slow,  the author had attempted to clarify the code
180by adding MORE statements to make the data movement more ``consistent``
181instead of simplifying his implementation and cutting down on all data
182movement (in particular, his use of L1, R1, L2, R2), and it was full of
183idiotic ``tweaks`` for particular machines which failed to deliver significant
184speedups but which did obfuscate everything.  so i took the test data
185from his verification program and rewrote everything else.
186
187a while later i ran across the great crypt(3) package mentioned above.
188the fact that this guy was computing 2 sboxes per table lookup rather
189than one (and using a MUCH larger table in the process) emboldened me to
190do the same - it was a trivial change from which i had been scared away
191by the larger table size.  in his case he didn't realize you don't need to keep
192the working data in TWO forms, one for easy use of half the sboxes in
193indexing, the other for easy use of the other half; instead you can keep
194it in the form for the first half and use a simple rotate to get the other
195half.  this means i have (almost) half the data manipulation and half
196the table size.  in fairness though he might be encoding something particular
197to crypt(3) in his tables - i didn't check.
198
199i'm glad that i implemented it the way i did, because this C version is
200portable (the ifdef's are performance enhancements) and it is faster
201than versions hand-written in assembly for the sparc!
202
203
204porting notes
205=============
206
207one thing i did not want to do was write an enormous mess
208which depended on endedness and other machine quirks,
209and which necessarily produced different code and different lookup tables
210for different machines.  see the kerberos code for an example
211of what i didn't want to do; all their endedness-specific ``optimizations``
212obfuscate the code and in the end were slower than a simpler machine
213independent approach.  however, there are always some portability
214considerations of some kind, and i have included some options
215for varying numbers of register variables.
216perhaps some will still regard the result as a mess!
217
2181) i assume everything is byte addressable, although i don't actually
219   depend on the byte order, and that bytes are 8 bits.
220   i assume word pointers can be freely cast to and from char pointers.
221   note that 99% of C programs make these assumptions.
222   i always use unsigned char's if the high bit could be set.
2232) the typedef ``word`` means a 32 bit unsigned integral type.
224   if ``unsigned long`` is not 32 bits, change the typedef in desCore.h.
225   i assume sizeof(word) == 4 EVERYWHERE.
226
227the (worst-case) cost of my NOT doing endedness-specific optimizations
228in the data loading and storing code surrounding the key iterations
229is less than 12%.  also, there is the added benefit that
230the input and output work areas do not need to be word-aligned.
231
232
233OPTIONAL performance optimizations
234==================================
235
2361) you should define one of ``i386,`` ``vax,`` ``mc68000,`` or ``sparc,``
237   whichever one is closest to the capabilities of your machine.
238   see the start of desCode.h to see exactly what this selection implies.
239   note that if you select the wrong one, the des code will still work;
240   these are just performance tweaks.
2412) for those with functional ``asm`` keywords: you should change the
242   ROR and ROL macros to use machine rotate instructions if you have them.
243   this will save 2 instructions and a temporary per use,
244   or about 32 to 40 instructions per en/decryption.
245
246   note that gcc is smart enough to translate the ROL/R macros into
247   machine rotates!
248
249these optimizations are all rather persnickety, yet with them you should
250be able to get performance equal to assembly-coding, except that:
251
2521) with the lack of a bit rotate operator in C, rotates have to be synthesized
253   from shifts.  so access to ``asm`` will speed things up if your machine
254   has rotates, as explained above in (3) (not necessary if you use gcc).
2552) if your machine has less than 12 32-bit registers i doubt your compiler will
256   generate good code.
257
258   ``i386`` tries to configure the code for a 386 by only declaring 3 registers
259   (it appears that gcc can use ebx, esi and edi to hold register variables).
260   however, if you like assembly coding, the 386 does have 7 32-bit registers,
261   and if you use ALL of them, use ``scaled by 8`` address modes with displacement
262   and other tricks, you can get reasonable routines for DesQuickCore... with
263   about 250 instructions apiece.  For DesSmall... it will help to rearrange
264   des_keymap, i.e., now the sbox # is the high part of the index and
265   the 6 bits of data is the low part; it helps to exchange these.
266
267   since i have no way to conveniently test it i have not provided my
268   shoehorned 386 version.  note that with this release of desCore, gcc is able
269   to put everything in registers(!), and generate about 370 instructions apiece
270   for the DesQuickCore... routines!
271
272coding notes
273============
274
275the en/decryption routines each use 6 necessary register variables,
276with 4 being actively used at once during the inner iterations.
277if you don't have 4 register variables get a new machine.
278up to 8 more registers are used to hold constants in some configurations.
279
280i assume that the use of a constant is more expensive than using a register:
281
282a) additionally, i have tried to put the larger constants in registers.
283   registering priority was by the following:
284
285	- anything more than 12 bits (bad for RISC and CISC)
286	- greater than 127 in value (can't use movq or byte immediate on CISC)
287	- 9-127 (may not be able to use CISC shift immediate or add/sub quick),
288	- 1-8 were never registered, being the cheapest constants.
289
290b) the compiler may be too stupid to realize table and table+256 should
291   be assigned to different constant registers and instead repetitively
292   do the arithmetic, so i assign these to explicit ``m`` register variables
293   when possible and helpful.
294
295i assume that indexing is cheaper or equivalent to auto increment/decrement,
296where the index is 7 bits unsigned or smaller.
297this assumption is reversed for 68k and vax.
298
299i assume that addresses can be cheaply formed from two registers,
300or from a register and a small constant.
301for the 68000, the ``two registers and small offset`` form is used sparingly.
302all index scaling is done explicitly - no hidden shifts by log2(sizeof).
303
304the code is written so that even a dumb compiler
305should never need more than one hidden temporary,
306increasing the chance that everything will fit in the registers.
307KEEP THIS MORE SUBTLE POINT IN MIND IF YOU REWRITE ANYTHING.
308
309(actually, there are some code fragments now which do require two temps,
310but fixing it would either break the structure of the macros or
311require declaring another temporary).
312
313
314special efficient data format
315==============================
316
317bits are manipulated in this arrangement most of the time (S7 S5 S3 S1)::
318
319	003130292827xxxx242322212019xxxx161514131211xxxx080706050403xxxx
320
321(the x bits are still there, i'm just emphasizing where the S boxes are).
322bits are rotated left 4 when computing S6 S4 S2 S0::
323
324	282726252423xxxx201918171615xxxx121110090807xxxx040302010031xxxx
325
326the rightmost two bits are usually cleared so the lower byte can be used
327as an index into an sbox mapping table. the next two x'd bits are set
328to various values to access different parts of the tables.
329
330
331how to use the routines
332
333datatypes:
334	pointer to 8 byte area of type DesData
335	used to hold keys and input/output blocks to des.
336
337	pointer to 128 byte area of type DesKeys
338	used to hold full 768-bit key.
339	must be long-aligned.
340
341DesQuickInit()
342	call this before using any other routine with ``Quick`` in its name.
343	it generates the special 64k table these routines need.
344DesQuickDone()
345	frees this table
346
347DesMethod(m, k)
348	m points to a 128byte block, k points to an 8 byte des key
349	which must have odd parity (or -1 is returned) and which must
350	not be a (semi-)weak key (or -2 is returned).
351	normally DesMethod() returns 0.
352
353	m is filled in from k so that when one of the routines below
354	is called with m, the routine will act like standard des
355	en/decryption with the key k. if you use DesMethod,
356	you supply a standard 56bit key; however, if you fill in
357	m yourself, you will get a 768bit key - but then it won't
358	be standard.  it's 768bits not 1024 because the least significant
359	two bits of each byte are not used.  note that these two bits
360	will be set to magic constants which speed up the encryption/decryption
361	on some machines.  and yes, each byte controls
362	a specific sbox during a specific iteration.
363
364	you really shouldn't use the 768bit format directly;  i should
365	provide a routine that converts 128 6-bit bytes (specified in
366	S-box mapping order or something) into the right format for you.
367	this would entail some byte concatenation and rotation.
368
369Des{Small|Quick}{Fips|Core}{Encrypt|Decrypt}(d, m, s)
370	performs des on the 8 bytes at s into the 8 bytes at
371	``d. (d,s: char *)``.
372
373	uses m as a 768bit key as explained above.
374
375	the Encrypt|Decrypt choice is obvious.
376
377	Fips|Core determines whether a completely standard FIPS initial
378	and final permutation is done; if not, then the data is loaded
379	and stored in a nonstandard bit order (FIPS w/o IP/FP).
380
381	Fips slows down Quick by 10%, Small by 9%.
382
383	Small|Quick determines whether you use the normal routine
384	or the crazy quick one which gobbles up 64k more of memory.
385	Small is 50% slower then Quick, but Quick needs 32 times as much
386	memory.  Quick is included for programs that do nothing but DES,
387	e.g., encryption filters, etc.
388
389
390Getting it to compile on your machine
391=====================================
392
393there are no machine-dependencies in the code (see porting),
394except perhaps the ``now()`` macro in desTest.c.
395ALL generated tables are machine independent.
396you should edit the Makefile with the appropriate optimization flags
397for your compiler (MAX optimization).
398
399
400Speeding up kerberos (and/or its des library)
401=============================================
402
403note that i have included a kerberos-compatible interface in desUtil.c
404through the functions des_key_sched() and des_ecb_encrypt().
405to use these with kerberos or kerberos-compatible code put desCore.a
406ahead of the kerberos-compatible library on your linker's command line.
407you should not need to #include desCore.h;  just include the header
408file provided with the kerberos library.
409
410Other uses
411==========
412
413the macros in desCode.h would be very useful for putting inline des
414functions in more complicated encryption routines.
415