summaryrefslogtreecommitdiff
path: root/src/lib/libcrypto/rc4/rc4_enc.c
diff options
context:
space:
mode:
authorbeck <>2000-03-19 11:13:58 +0000
committerbeck <>2000-03-19 11:13:58 +0000
commit796d609550df3a33fc11468741c5d2f6d3df4c11 (patch)
tree6c6d539061caa20372dad0ac4ddb1dfae2fbe7fe /src/lib/libcrypto/rc4/rc4_enc.c
parent5be3114c1fd7e0dfea1e38d3abb4cbba75244419 (diff)
downloadopenbsd-796d609550df3a33fc11468741c5d2f6d3df4c11.tar.gz
openbsd-796d609550df3a33fc11468741c5d2f6d3df4c11.tar.bz2
openbsd-796d609550df3a33fc11468741c5d2f6d3df4c11.zip
OpenSSL 0.9.5 merge
*warning* this bumps shared lib minors for libssl and libcrypto from 2.1 to 2.2 if you are using the ssl26 packages for ssh and other things to work you will need to get new ones (see ~beck/libsslsnap/<arch>) on cvs or ~beck/src-patent.tar.gz on cvs
Diffstat (limited to 'src/lib/libcrypto/rc4/rc4_enc.c')
-rw-r--r--src/lib/libcrypto/rc4/rc4_enc.c186
1 files changed, 185 insertions, 1 deletions
diff --git a/src/lib/libcrypto/rc4/rc4_enc.c b/src/lib/libcrypto/rc4/rc4_enc.c
index 3256bea8cc..d5f18a3a70 100644
--- a/src/lib/libcrypto/rc4/rc4_enc.c
+++ b/src/lib/libcrypto/rc4/rc4_enc.c
@@ -67,7 +67,7 @@
67 * Date: Wed, 14 Sep 1994 06:35:31 GMT 67 * Date: Wed, 14 Sep 1994 06:35:31 GMT
68 */ 68 */
69 69
70void RC4(RC4_KEY *key, unsigned long len, unsigned char *indata, 70void RC4(RC4_KEY *key, unsigned long len, const unsigned char *indata,
71 unsigned char *outdata) 71 unsigned char *outdata)
72 { 72 {
73 register RC4_INT *d; 73 register RC4_INT *d;
@@ -78,6 +78,190 @@ void RC4(RC4_KEY *key, unsigned long len, unsigned char *indata,
78 y=key->y; 78 y=key->y;
79 d=key->data; 79 d=key->data;
80 80
81#if defined(RC4_CHUNK)
82 /*
83 * The original reason for implementing this(*) was the fact that
84 * pre-21164a Alpha CPUs don't have byte load/store instructions
85 * and e.g. a byte store has to be done with 64-bit load, shift,
86 * and, or and finally 64-bit store. Peaking data and operating
87 * at natural word size made it possible to reduce amount of
88 * instructions as well as to perform early read-ahead without
89 * suffering from RAW (read-after-write) hazard. This resulted
90 * in ~40%(**) performance improvement on 21064 box with gcc.
91 * But it's not only Alpha users who win here:-) Thanks to the
92 * early-n-wide read-ahead this implementation also exhibits
93 * >40% speed-up on SPARC and 20-30% on 64-bit MIPS (depending
94 * on sizeof(RC4_INT)).
95 *
96 * (*) "this" means code which recognizes the case when input
97 * and output pointers appear to be aligned at natural CPU
98 * word boundary
99 * (**) i.e. according to 'apps/openssl speed rc4' benchmark,
100 * crypto/rc4/rc4speed.c exhibits almost 70% speed-up...
101 *
102 * Cavets.
103 *
104 * - RC4_CHUNK="unsigned long long" should be a #1 choice for
105 * UltraSPARC. Unfortunately gcc generates very slow code
106 * (2.5-3 times slower than one generated by Sun's WorkShop
107 * C) and therefore gcc (at least 2.95 and earlier) should
108 * always be told that RC4_CHUNK="unsigned long".
109 *
110 * <appro@fy.chalmers.se>
111 */
112
113# define RC4_STEP ( \
114 x=(x+1) &0xff, \
115 tx=d[x], \
116 y=(tx+y)&0xff, \
117 ty=d[y], \
118 d[y]=tx, \
119 d[x]=ty, \
120 (RC4_CHUNK)d[(tx+ty)&0xff]\
121 )
122
123 if ( ( ((unsigned long)indata & (sizeof(RC4_CHUNK)-1)) |
124 ((unsigned long)outdata & (sizeof(RC4_CHUNK)-1)) ) == 0 )
125 {
126 RC4_CHUNK ichunk,otp;
127 const union { long one; char little; } is_endian = {1};
128
129 /*
130 * I reckon we can afford to implement both endian
131 * cases and to decide which way to take at run-time
132 * because the machine code appears to be very compact
133 * and redundant 1-2KB is perfectly tolerable (i.e.
134 * in case the compiler fails to eliminate it:-). By
135 * suggestion from Terrel Larson <terr@terralogic.net>
136 * who also stands for the is_endian union:-)
137 *
138 * Special notes.
139 *
140 * - is_endian is declared automatic as doing otherwise
141 * (declaring static) prevents gcc from eliminating
142 * the redundant code;
143 * - compilers (those I've tried) don't seem to have
144 * problems eliminating either the operators guarded
145 * by "if (sizeof(RC4_CHUNK)==8)" or the condition
146 * expressions themselves so I've got 'em to replace
147 * corresponding #ifdefs from the previous version;
148 * - I chose to let the redundant switch cases when
149 * sizeof(RC4_CHUNK)!=8 be (were also #ifdefed
150 * before);
151 * - in case you wonder "&(sizeof(RC4_CHUNK)*8-1)" in
152 * [LB]ESHFT guards against "shift is out of range"
153 * warnings when sizeof(RC4_CHUNK)!=8
154 *
155 * <appro@fy.chalmers.se>
156 */
157 if (!is_endian.little)
158 { /* BIG-ENDIAN CASE */
159# define BESHFT(c) (((sizeof(RC4_CHUNK)-(c)-1)*8)&(sizeof(RC4_CHUNK)*8-1))
160 for (;len&-sizeof(RC4_CHUNK);len-=sizeof(RC4_CHUNK))
161 {
162 ichunk = *(RC4_CHUNK *)indata;
163 otp = RC4_STEP<<BESHFT(0);
164 otp |= RC4_STEP<<BESHFT(1);
165 otp |= RC4_STEP<<BESHFT(2);
166 otp |= RC4_STEP<<BESHFT(3);
167 if (sizeof(RC4_CHUNK)==8)
168 {
169 otp |= RC4_STEP<<BESHFT(4);
170 otp |= RC4_STEP<<BESHFT(5);
171 otp |= RC4_STEP<<BESHFT(6);
172 otp |= RC4_STEP<<BESHFT(7);
173 }
174 *(RC4_CHUNK *)outdata = otp^ichunk;
175 indata += sizeof(RC4_CHUNK);
176 outdata += sizeof(RC4_CHUNK);
177 }
178 if (len)
179 {
180 RC4_CHUNK mask=(RC4_CHUNK)-1, ochunk;
181
182 ichunk = *(RC4_CHUNK *)indata;
183 ochunk = *(RC4_CHUNK *)outdata;
184 otp = 0;
185 i = BESHFT(0);
186 mask <<= (sizeof(RC4_CHUNK)-len)<<3;
187 switch (len&(sizeof(RC4_CHUNK)-1))
188 {
189 case 7: otp = RC4_STEP<<i, i-=8;
190 case 6: otp |= RC4_STEP<<i, i-=8;
191 case 5: otp |= RC4_STEP<<i, i-=8;
192 case 4: otp |= RC4_STEP<<i, i-=8;
193 case 3: otp |= RC4_STEP<<i, i-=8;
194 case 2: otp |= RC4_STEP<<i, i-=8;
195 case 1: otp |= RC4_STEP<<i, i-=8;
196 case 0: ; /*
197 * it's never the case,
198 * but it has to be here
199 * for ultrix?
200 */
201 }
202 ochunk &= ~mask;
203 ochunk |= (otp^ichunk) & mask;
204 *(RC4_CHUNK *)outdata = ochunk;
205 }
206 key->x=x;
207 key->y=y;
208 return;
209 }
210 else
211 { /* LITTLE-ENDIAN CASE */
212# define LESHFT(c) (((c)*8)&(sizeof(RC4_CHUNK)*8-1))
213 for (;len&-sizeof(RC4_CHUNK);len-=sizeof(RC4_CHUNK))
214 {
215 ichunk = *(RC4_CHUNK *)indata;
216 otp = RC4_STEP;
217 otp |= RC4_STEP<<8;
218 otp |= RC4_STEP<<16;
219 otp |= RC4_STEP<<24;
220 if (sizeof(RC4_CHUNK)==8)
221 {
222 otp |= RC4_STEP<<LESHFT(4);
223 otp |= RC4_STEP<<LESHFT(5);
224 otp |= RC4_STEP<<LESHFT(6);
225 otp |= RC4_STEP<<LESHFT(7);
226 }
227 *(RC4_CHUNK *)outdata = otp^ichunk;
228 indata += sizeof(RC4_CHUNK);
229 outdata += sizeof(RC4_CHUNK);
230 }
231 if (len)
232 {
233 RC4_CHUNK mask=(RC4_CHUNK)-1, ochunk;
234
235 ichunk = *(RC4_CHUNK *)indata;
236 ochunk = *(RC4_CHUNK *)outdata;
237 otp = 0;
238 i = 0;
239 mask >>= (sizeof(RC4_CHUNK)-len)<<3;
240 switch (len&(sizeof(RC4_CHUNK)-1))
241 {
242 case 7: otp = RC4_STEP, i+=8;
243 case 6: otp |= RC4_STEP<<i, i+=8;
244 case 5: otp |= RC4_STEP<<i, i+=8;
245 case 4: otp |= RC4_STEP<<i, i+=8;
246 case 3: otp |= RC4_STEP<<i, i+=8;
247 case 2: otp |= RC4_STEP<<i, i+=8;
248 case 1: otp |= RC4_STEP<<i, i+=8;
249 case 0: ; /*
250 * it's never the case,
251 * but it has to be here
252 * for ultrix?
253 */
254 }
255 ochunk &= ~mask;
256 ochunk |= (otp^ichunk) & mask;
257 *(RC4_CHUNK *)outdata = ochunk;
258 }
259 key->x=x;
260 key->y=y;
261 return;
262 }
263 }
264#endif
81#define LOOP(in,out) \ 265#define LOOP(in,out) \
82 x=((x+1)&0xff); \ 266 x=((x+1)&0xff); \
83 tx=d[x]; \ 267 tx=d[x]; \
generated by cgit v1.2.3-55-g6feb (git 2.39.1) at 2025-04-25 21:56:08 +0000