Tobias Klauser | d533f67 | 2005-09-10 00:26:46 -0700 | [diff] [blame] | 1 | Below is the original README file from the descore.shar package. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2 | ------------------------------------------------------------------------------ |
| 3 | |
| 4 | des - fast & portable DES encryption & decryption. |
| 5 | Copyright (C) 1992 Dana L. How |
| 6 | |
| 7 | This program is free software; you can redistribute it and/or modify |
| 8 | it under the terms of the GNU Library General Public License as published by |
| 9 | the Free Software Foundation; either version 2 of the License, or |
| 10 | (at your option) any later version. |
| 11 | |
| 12 | This program is distributed in the hope that it will be useful, |
| 13 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 15 | GNU Library General Public License for more details. |
| 16 | |
| 17 | You should have received a copy of the GNU Library General Public License |
| 18 | along with this program; if not, write to the Free Software |
| 19 | Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
| 20 | |
| 21 | Author's address: how@isl.stanford.edu |
| 22 | |
| 23 | $Id: README,v 1.15 1992/05/20 00:25:32 how E $ |
| 24 | |
| 25 | |
| 26 | ==>> To compile after untarring/unsharring, just `make' <<== |
| 27 | |
| 28 | |
| 29 | This package was designed with the following goals: |
| 30 | 1. Highest possible encryption/decryption PERFORMANCE. |
| 31 | 2. PORTABILITY to any byte-addressable host with a 32bit unsigned C type |
| 32 | 3. Plug-compatible replacement for KERBEROS's low-level routines. |
| 33 | |
| 34 | This second release includes a number of performance enhancements for |
| 35 | register-starved machines. My discussions with Richard Outerbridge, |
| 36 | 71755.204@compuserve.com, sparked a number of these enhancements. |
| 37 | |
| 38 | To more rapidly understand the code in this package, inspect desSmallFips.i |
| 39 | (created by typing `make') BEFORE you tackle desCode.h. The latter is set |
| 40 | up in a parameterized fashion so it can easily be modified by speed-daemon |
| 41 | hackers in pursuit of that last microsecond. You will find it more |
| 42 | illuminating to inspect one specific implementation, |
| 43 | and then move on to the common abstract skeleton with this one in mind. |
| 44 | |
| 45 | |
| 46 | performance comparison to other available des code which i could |
| 47 | compile on a SPARCStation 1 (cc -O4, gcc -O2): |
| 48 | |
| 49 | this code (byte-order independent): |
| 50 | 30us per encryption (options: 64k tables, no IP/FP) |
| 51 | 33us per encryption (options: 64k tables, FIPS standard bit ordering) |
| 52 | 45us per encryption (options: 2k tables, no IP/FP) |
| 53 | 48us per encryption (options: 2k tables, FIPS standard bit ordering) |
| 54 | 275us to set a new key (uses 1k of key tables) |
| 55 | this has the quickest encryption/decryption routines i've seen. |
| 56 | since i was interested in fast des filters rather than crypt(3) |
| 57 | and password cracking, i haven't really bothered yet to speed up |
| 58 | the key setting routine. also, i have no interest in re-implementing |
| 59 | all the other junk in the mit kerberos des library, so i've just |
| 60 | provided my routines with little stub interfaces so they can be |
| 61 | used as drop-in replacements with mit's code or any of the mit- |
| 62 | compatible packages below. (note that the first two timings above |
| 63 | are highly variable because of cache effects). |
| 64 | |
| 65 | kerberos des replacement from australia (version 1.95): |
| 66 | 53us per encryption (uses 2k of tables) |
| 67 | 96us to set a new key (uses 2.25k of key tables) |
| 68 | so despite the author's inclusion of some of the performance |
| 69 | improvements i had suggested to him, this package's |
| 70 | encryption/decryption is still slower on the sparc and 68000. |
| 71 | more specifically, 19-40% slower on the 68020 and 11-35% slower |
| 72 | on the sparc, depending on the compiler; |
| 73 | in full gory detail (ALT_ECB is a libdes variant): |
| 74 | compiler machine desCore libdes ALT_ECB slower by |
| 75 | gcc 2.1 -O2 Sun 3/110 304 uS 369.5uS 461.8uS 22% |
| 76 | cc -O1 Sun 3/110 336 uS 436.6uS 399.3uS 19% |
| 77 | cc -O2 Sun 3/110 360 uS 532.4uS 505.1uS 40% |
| 78 | cc -O4 Sun 3/110 365 uS 532.3uS 505.3uS 38% |
| 79 | gcc 2.1 -O2 Sun 4/50 48 uS 53.4uS 57.5uS 11% |
| 80 | cc -O2 Sun 4/50 48 uS 64.6uS 64.7uS 35% |
| 81 | cc -O4 Sun 4/50 48 uS 64.7uS 64.9uS 35% |
| 82 | (my time measurements are not as accurate as his). |
| 83 | the comments in my first release of desCore on version 1.92: |
| 84 | 68us per encryption (uses 2k of tables) |
| 85 | 96us to set a new key (uses 2.25k of key tables) |
| 86 | this is a very nice package which implements the most important |
| 87 | of the optimizations which i did in my encryption routines. |
| 88 | it's a bit weak on common low-level optimizations which is why |
| 89 | it's 39%-106% slower. because he was interested in fast crypt(3) and |
| 90 | password-cracking applications, he also used the same ideas to |
| 91 | speed up the key-setting routines with impressive results. |
| 92 | (at some point i may do the same in my package). he also implements |
| 93 | the rest of the mit des library. |
| 94 | (code from eay@psych.psy.uq.oz.au via comp.sources.misc) |
| 95 | |
| 96 | fast crypt(3) package from denmark: |
| 97 | the des routine here is buried inside a loop to do the |
| 98 | crypt function and i didn't feel like ripping it out and measuring |
| 99 | performance. his code takes 26 sparc instructions to compute one |
| 100 | des iteration; above, Quick (64k) takes 21 and Small (2k) takes 37. |
| 101 | he claims to use 280k of tables but the iteration calculation seems |
| 102 | to use only 128k. his tables and code are machine independent. |
| 103 | (code from glad@daimi.aau.dk via alt.sources or comp.sources.misc) |
| 104 | |
| 105 | swedish reimplementation of Kerberos des library |
| 106 | 108us per encryption (uses 34k worth of tables) |
| 107 | 134us to set a new key (uses 32k of key tables to get this speed!) |
| 108 | the tables used seem to be machine-independent; |
| 109 | he seems to have included a lot of special case code |
| 110 | so that, e.g., `long' loads can be used instead of 4 `char' loads |
| 111 | when the machine's architecture allows it. |
| 112 | (code obtained from chalmers.se:pub/des) |
| 113 | |
| 114 | crack 3.3c package from england: |
| 115 | as in crypt above, the des routine is buried in a loop. it's |
| 116 | also very modified for crypt. his iteration code uses 16k |
| 117 | of tables and appears to be slow. |
| 118 | (code obtained from aem@aber.ac.uk via alt.sources or comp.sources.misc) |
| 119 | |
| 120 | ``highly optimized'' and tweaked Kerberos/Athena code (byte-order dependent): |
| 121 | 165us per encryption (uses 6k worth of tables) |
| 122 | 478us to set a new key (uses <1k of key tables) |
| 123 | so despite the comments in this code, it was possible to get |
| 124 | faster code AND smaller tables, as well as making the tables |
| 125 | machine-independent. |
| 126 | (code obtained from prep.ai.mit.edu) |
| 127 | |
| 128 | UC Berkeley code (depends on machine-endedness): |
| 129 | 226us per encryption |
| 130 | 10848us to set a new key |
| 131 | table sizes are unclear, but they don't look very small |
| 132 | (code obtained from wuarchive.wustl.edu) |
| 133 | |
| 134 | |
| 135 | motivation and history |
| 136 | |
| 137 | a while ago i wanted some des routines and the routines documented on sun's |
| 138 | man pages either didn't exist or dumped core. i had heard of kerberos, |
| 139 | and knew that it used des, so i figured i'd use its routines. but once |
| 140 | i got it and looked at the code, it really set off a lot of pet peeves - |
| 141 | it was too convoluted, the code had been written without taking |
| 142 | advantage of the regular structure of operations such as IP, E, and FP |
| 143 | (i.e. the author didn't sit down and think before coding), |
| 144 | it was excessively slow, the author had attempted to clarify the code |
| 145 | by adding MORE statements to make the data movement more `consistent' |
| 146 | instead of simplifying his implementation and cutting down on all data |
| 147 | movement (in particular, his use of L1, R1, L2, R2), and it was full of |
| 148 | idiotic `tweaks' for particular machines which failed to deliver significant |
| 149 | speedups but which did obfuscate everything. so i took the test data |
| 150 | from his verification program and rewrote everything else. |
| 151 | |
| 152 | a while later i ran across the great crypt(3) package mentioned above. |
| 153 | the fact that this guy was computing 2 sboxes per table lookup rather |
| 154 | than one (and using a MUCH larger table in the process) emboldened me to |
| 155 | do the same - it was a trivial change from which i had been scared away |
| 156 | by the larger table size. in his case he didn't realize you don't need to keep |
| 157 | the working data in TWO forms, one for easy use of half the sboxes in |
| 158 | indexing, the other for easy use of the other half; instead you can keep |
| 159 | it in the form for the first half and use a simple rotate to get the other |
| 160 | half. this means i have (almost) half the data manipulation and half |
| 161 | the table size. in fairness though he might be encoding something particular |
| 162 | to crypt(3) in his tables - i didn't check. |
| 163 | |
| 164 | i'm glad that i implemented it the way i did, because this C version is |
| 165 | portable (the ifdef's are performance enhancements) and it is faster |
| 166 | than versions hand-written in assembly for the sparc! |
| 167 | |
| 168 | |
| 169 | porting notes |
| 170 | |
| 171 | one thing i did not want to do was write an enormous mess |
| 172 | which depended on endedness and other machine quirks, |
| 173 | and which necessarily produced different code and different lookup tables |
| 174 | for different machines. see the kerberos code for an example |
| 175 | of what i didn't want to do; all their endedness-specific `optimizations' |
| 176 | obfuscate the code and in the end were slower than a simpler machine |
| 177 | independent approach. however, there are always some portability |
| 178 | considerations of some kind, and i have included some options |
| 179 | for varying numbers of register variables. |
| 180 | perhaps some will still regard the result as a mess! |
| 181 | |
| 182 | 1) i assume everything is byte addressable, although i don't actually |
| 183 | depend on the byte order, and that bytes are 8 bits. |
| 184 | i assume word pointers can be freely cast to and from char pointers. |
| 185 | note that 99% of C programs make these assumptions. |
| 186 | i always use unsigned char's if the high bit could be set. |
| 187 | 2) the typedef `word' means a 32 bit unsigned integral type. |
| 188 | if `unsigned long' is not 32 bits, change the typedef in desCore.h. |
| 189 | i assume sizeof(word) == 4 EVERYWHERE. |
| 190 | |
| 191 | the (worst-case) cost of my NOT doing endedness-specific optimizations |
| 192 | in the data loading and storing code surrounding the key iterations |
| 193 | is less than 12%. also, there is the added benefit that |
| 194 | the input and output work areas do not need to be word-aligned. |
| 195 | |
| 196 | |
| 197 | OPTIONAL performance optimizations |
| 198 | |
| 199 | 1) you should define one of `i386,' `vax,' `mc68000,' or `sparc,' |
| 200 | whichever one is closest to the capabilities of your machine. |
| 201 | see the start of desCode.h to see exactly what this selection implies. |
| 202 | note that if you select the wrong one, the des code will still work; |
| 203 | these are just performance tweaks. |
| 204 | 2) for those with functional `asm' keywords: you should change the |
| 205 | ROR and ROL macros to use machine rotate instructions if you have them. |
| 206 | this will save 2 instructions and a temporary per use, |
| 207 | or about 32 to 40 instructions per en/decryption. |
| 208 | note that gcc is smart enough to translate the ROL/R macros into |
| 209 | machine rotates! |
| 210 | |
| 211 | these optimizations are all rather persnickety, yet with them you should |
| 212 | be able to get performance equal to assembly-coding, except that: |
| 213 | 1) with the lack of a bit rotate operator in C, rotates have to be synthesized |
| 214 | from shifts. so access to `asm' will speed things up if your machine |
| 215 | has rotates, as explained above in (3) (not necessary if you use gcc). |
| 216 | 2) if your machine has less than 12 32-bit registers i doubt your compiler will |
| 217 | generate good code. |
| 218 | `i386' tries to configure the code for a 386 by only declaring 3 registers |
| 219 | (it appears that gcc can use ebx, esi and edi to hold register variables). |
| 220 | however, if you like assembly coding, the 386 does have 7 32-bit registers, |
| 221 | and if you use ALL of them, use `scaled by 8' address modes with displacement |
| 222 | and other tricks, you can get reasonable routines for DesQuickCore... with |
| 223 | about 250 instructions apiece. For DesSmall... it will help to rearrange |
| 224 | des_keymap, i.e., now the sbox # is the high part of the index and |
| 225 | the 6 bits of data is the low part; it helps to exchange these. |
| 226 | since i have no way to conveniently test it i have not provided my |
| 227 | shoehorned 386 version. note that with this release of desCore, gcc is able |
| 228 | to put everything in registers(!), and generate about 370 instructions apiece |
| 229 | for the DesQuickCore... routines! |
| 230 | |
| 231 | coding notes |
| 232 | |
| 233 | the en/decryption routines each use 6 necessary register variables, |
| 234 | with 4 being actively used at once during the inner iterations. |
| 235 | if you don't have 4 register variables get a new machine. |
| 236 | up to 8 more registers are used to hold constants in some configurations. |
| 237 | |
| 238 | i assume that the use of a constant is more expensive than using a register: |
| 239 | a) additionally, i have tried to put the larger constants in registers. |
| 240 | registering priority was by the following: |
| 241 | anything more than 12 bits (bad for RISC and CISC) |
| 242 | greater than 127 in value (can't use movq or byte immediate on CISC) |
| 243 | 9-127 (may not be able to use CISC shift immediate or add/sub quick), |
| 244 | 1-8 were never registered, being the cheapest constants. |
| 245 | b) the compiler may be too stupid to realize table and table+256 should |
| 246 | be assigned to different constant registers and instead repetitively |
| 247 | do the arithmetic, so i assign these to explicit `m' register variables |
| 248 | when possible and helpful. |
| 249 | |
| 250 | i assume that indexing is cheaper or equivalent to auto increment/decrement, |
| 251 | where the index is 7 bits unsigned or smaller. |
| 252 | this assumption is reversed for 68k and vax. |
| 253 | |
| 254 | i assume that addresses can be cheaply formed from two registers, |
| 255 | or from a register and a small constant. |
| 256 | for the 68000, the `two registers and small offset' form is used sparingly. |
| 257 | all index scaling is done explicitly - no hidden shifts by log2(sizeof). |
| 258 | |
| 259 | the code is written so that even a dumb compiler |
| 260 | should never need more than one hidden temporary, |
| 261 | increasing the chance that everything will fit in the registers. |
| 262 | KEEP THIS MORE SUBTLE POINT IN MIND IF YOU REWRITE ANYTHING. |
| 263 | (actually, there are some code fragments now which do require two temps, |
| 264 | but fixing it would either break the structure of the macros or |
| 265 | require declaring another temporary). |
| 266 | |
| 267 | |
| 268 | special efficient data format |
| 269 | |
| 270 | bits are manipulated in this arrangement most of the time (S7 S5 S3 S1): |
| 271 | 003130292827xxxx242322212019xxxx161514131211xxxx080706050403xxxx |
| 272 | (the x bits are still there, i'm just emphasizing where the S boxes are). |
| 273 | bits are rotated left 4 when computing S6 S4 S2 S0: |
| 274 | 282726252423xxxx201918171615xxxx121110090807xxxx040302010031xxxx |
| 275 | the rightmost two bits are usually cleared so the lower byte can be used |
| 276 | as an index into an sbox mapping table. the next two x'd bits are set |
| 277 | to various values to access different parts of the tables. |
| 278 | |
| 279 | |
| 280 | how to use the routines |
| 281 | |
| 282 | datatypes: |
| 283 | pointer to 8 byte area of type DesData |
| 284 | used to hold keys and input/output blocks to des. |
| 285 | |
| 286 | pointer to 128 byte area of type DesKeys |
| 287 | used to hold full 768-bit key. |
| 288 | must be long-aligned. |
| 289 | |
| 290 | DesQuickInit() |
| 291 | call this before using any other routine with `Quick' in its name. |
| 292 | it generates the special 64k table these routines need. |
| 293 | DesQuickDone() |
| 294 | frees this table |
| 295 | |
| 296 | DesMethod(m, k) |
| 297 | m points to a 128byte block, k points to an 8 byte des key |
| 298 | which must have odd parity (or -1 is returned) and which must |
| 299 | not be a (semi-)weak key (or -2 is returned). |
| 300 | normally DesMethod() returns 0. |
| 301 | m is filled in from k so that when one of the routines below |
| 302 | is called with m, the routine will act like standard des |
| 303 | en/decryption with the key k. if you use DesMethod, |
| 304 | you supply a standard 56bit key; however, if you fill in |
| 305 | m yourself, you will get a 768bit key - but then it won't |
| 306 | be standard. it's 768bits not 1024 because the least significant |
| 307 | two bits of each byte are not used. note that these two bits |
| 308 | will be set to magic constants which speed up the encryption/decryption |
| 309 | on some machines. and yes, each byte controls |
| 310 | a specific sbox during a specific iteration. |
| 311 | you really shouldn't use the 768bit format directly; i should |
| 312 | provide a routine that converts 128 6-bit bytes (specified in |
| 313 | S-box mapping order or something) into the right format for you. |
| 314 | this would entail some byte concatenation and rotation. |
| 315 | |
| 316 | Des{Small|Quick}{Fips|Core}{Encrypt|Decrypt}(d, m, s) |
| 317 | performs des on the 8 bytes at s into the 8 bytes at d. (d,s: char *). |
| 318 | uses m as a 768bit key as explained above. |
| 319 | the Encrypt|Decrypt choice is obvious. |
| 320 | Fips|Core determines whether a completely standard FIPS initial |
| 321 | and final permutation is done; if not, then the data is loaded |
| 322 | and stored in a nonstandard bit order (FIPS w/o IP/FP). |
| 323 | Fips slows down Quick by 10%, Small by 9%. |
| 324 | Small|Quick determines whether you use the normal routine |
| 325 | or the crazy quick one which gobbles up 64k more of memory. |
| 326 | Small is 50% slower then Quick, but Quick needs 32 times as much |
| 327 | memory. Quick is included for programs that do nothing but DES, |
| 328 | e.g., encryption filters, etc. |
| 329 | |
| 330 | |
| 331 | Getting it to compile on your machine |
| 332 | |
| 333 | there are no machine-dependencies in the code (see porting), |
| 334 | except perhaps the `now()' macro in desTest.c. |
| 335 | ALL generated tables are machine independent. |
| 336 | you should edit the Makefile with the appropriate optimization flags |
| 337 | for your compiler (MAX optimization). |
| 338 | |
| 339 | |
| 340 | Speeding up kerberos (and/or its des library) |
| 341 | |
| 342 | note that i have included a kerberos-compatible interface in desUtil.c |
| 343 | through the functions des_key_sched() and des_ecb_encrypt(). |
| 344 | to use these with kerberos or kerberos-compatible code put desCore.a |
| 345 | ahead of the kerberos-compatible library on your linker's command line. |
| 346 | you should not need to #include desCore.h; just include the header |
| 347 | file provided with the kerberos library. |
| 348 | |
| 349 | Other uses |
| 350 | |
| 351 | the macros in desCode.h would be very useful for putting inline des |
| 352 | functions in more complicated encryption routines. |