x86_64 performance patch.
[openssl.git] / crypto / rc5 / rc5_locl.h
1 /* crypto/rc5/rc5_locl.h */
2 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
3  * All rights reserved.
4  *
5  * This package is an SSL implementation written
6  * by Eric Young (eay@cryptsoft.com).
7  * The implementation was written so as to conform with Netscapes SSL.
8  * 
9  * This library is free for commercial and non-commercial use as long as
10  * the following conditions are aheared to.  The following conditions
11  * apply to all code found in this distribution, be it the RC4, RSA,
12  * lhash, DES, etc., code; not just the SSL code.  The SSL documentation
13  * included with this distribution is covered by the same copyright terms
14  * except that the holder is Tim Hudson (tjh@cryptsoft.com).
15  * 
16  * Copyright remains Eric Young's, and as such any Copyright notices in
17  * the code are not to be removed.
18  * If this package is used in a product, Eric Young should be given attribution
19  * as the author of the parts of the library used.
20  * This can be in the form of a textual message at program startup or
21  * in documentation (online or textual) provided with the package.
22  * 
23  * Redistribution and use in source and binary forms, with or without
24  * modification, are permitted provided that the following conditions
25  * are met:
26  * 1. Redistributions of source code must retain the copyright
27  *    notice, this list of conditions and the following disclaimer.
28  * 2. Redistributions in binary form must reproduce the above copyright
29  *    notice, this list of conditions and the following disclaimer in the
30  *    documentation and/or other materials provided with the distribution.
31  * 3. All advertising materials mentioning features or use of this software
32  *    must display the following acknowledgement:
33  *    "This product includes cryptographic software written by
34  *     Eric Young (eay@cryptsoft.com)"
35  *    The word 'cryptographic' can be left out if the rouines from the library
36  *    being used are not cryptographic related :-).
37  * 4. If you include any Windows specific code (or a derivative thereof) from 
38  *    the apps directory (application code) you must include an acknowledgement:
39  *    "This product includes software written by Tim Hudson (tjh@cryptsoft.com)"
40  * 
41  * THIS SOFTWARE IS PROVIDED BY ERIC YOUNG ``AS IS'' AND
42  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
43  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
44  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
45  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
46  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
47  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
48  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
49  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
50  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
51  * SUCH DAMAGE.
52  * 
53  * The licence and distribution terms for any publically available version or
54  * derivative of this code cannot be changed.  i.e. this code cannot simply be
55  * copied and put under another distribution licence
56  * [including the GNU Public Licence.]
57  */
58
59 #include <stdlib.h>
60
61 #undef c2l
62 #define c2l(c,l)        (l =((unsigned long)(*((c)++)))    , \
63                          l|=((unsigned long)(*((c)++)))<< 8L, \
64                          l|=((unsigned long)(*((c)++)))<<16L, \
65                          l|=((unsigned long)(*((c)++)))<<24L)
66
67 /* NOTE - c is not incremented as per c2l */
68 #undef c2ln
69 #define c2ln(c,l1,l2,n) { \
70                         c+=n; \
71                         l1=l2=0; \
72                         switch (n) { \
73                         case 8: l2 =((unsigned long)(*(--(c))))<<24L; \
74                         case 7: l2|=((unsigned long)(*(--(c))))<<16L; \
75                         case 6: l2|=((unsigned long)(*(--(c))))<< 8L; \
76                         case 5: l2|=((unsigned long)(*(--(c))));     \
77                         case 4: l1 =((unsigned long)(*(--(c))))<<24L; \
78                         case 3: l1|=((unsigned long)(*(--(c))))<<16L; \
79                         case 2: l1|=((unsigned long)(*(--(c))))<< 8L; \
80                         case 1: l1|=((unsigned long)(*(--(c))));     \
81                                 } \
82                         }
83
84 #undef l2c
85 #define l2c(l,c)        (*((c)++)=(unsigned char)(((l)     )&0xff), \
86                          *((c)++)=(unsigned char)(((l)>> 8L)&0xff), \
87                          *((c)++)=(unsigned char)(((l)>>16L)&0xff), \
88                          *((c)++)=(unsigned char)(((l)>>24L)&0xff))
89
90 /* NOTE - c is not incremented as per l2c */
91 #undef l2cn
92 #define l2cn(l1,l2,c,n) { \
93                         c+=n; \
94                         switch (n) { \
95                         case 8: *(--(c))=(unsigned char)(((l2)>>24L)&0xff); \
96                         case 7: *(--(c))=(unsigned char)(((l2)>>16L)&0xff); \
97                         case 6: *(--(c))=(unsigned char)(((l2)>> 8L)&0xff); \
98                         case 5: *(--(c))=(unsigned char)(((l2)     )&0xff); \
99                         case 4: *(--(c))=(unsigned char)(((l1)>>24L)&0xff); \
100                         case 3: *(--(c))=(unsigned char)(((l1)>>16L)&0xff); \
101                         case 2: *(--(c))=(unsigned char)(((l1)>> 8L)&0xff); \
102                         case 1: *(--(c))=(unsigned char)(((l1)     )&0xff); \
103                                 } \
104                         }
105
106 /* NOTE - c is not incremented as per n2l */
107 #define n2ln(c,l1,l2,n) { \
108                         c+=n; \
109                         l1=l2=0; \
110                         switch (n) { \
111                         case 8: l2 =((unsigned long)(*(--(c))))    ; \
112                         case 7: l2|=((unsigned long)(*(--(c))))<< 8; \
113                         case 6: l2|=((unsigned long)(*(--(c))))<<16; \
114                         case 5: l2|=((unsigned long)(*(--(c))))<<24; \
115                         case 4: l1 =((unsigned long)(*(--(c))))    ; \
116                         case 3: l1|=((unsigned long)(*(--(c))))<< 8; \
117                         case 2: l1|=((unsigned long)(*(--(c))))<<16; \
118                         case 1: l1|=((unsigned long)(*(--(c))))<<24; \
119                                 } \
120                         }
121
122 /* NOTE - c is not incremented as per l2n */
123 #define l2nn(l1,l2,c,n) { \
124                         c+=n; \
125                         switch (n) { \
126                         case 8: *(--(c))=(unsigned char)(((l2)    )&0xff); \
127                         case 7: *(--(c))=(unsigned char)(((l2)>> 8)&0xff); \
128                         case 6: *(--(c))=(unsigned char)(((l2)>>16)&0xff); \
129                         case 5: *(--(c))=(unsigned char)(((l2)>>24)&0xff); \
130                         case 4: *(--(c))=(unsigned char)(((l1)    )&0xff); \
131                         case 3: *(--(c))=(unsigned char)(((l1)>> 8)&0xff); \
132                         case 2: *(--(c))=(unsigned char)(((l1)>>16)&0xff); \
133                         case 1: *(--(c))=(unsigned char)(((l1)>>24)&0xff); \
134                                 } \
135                         }
136
137 #undef n2l
138 #define n2l(c,l)        (l =((unsigned long)(*((c)++)))<<24L, \
139                          l|=((unsigned long)(*((c)++)))<<16L, \
140                          l|=((unsigned long)(*((c)++)))<< 8L, \
141                          l|=((unsigned long)(*((c)++))))
142
143 #undef l2n
144 #define l2n(l,c)        (*((c)++)=(unsigned char)(((l)>>24L)&0xff), \
145                          *((c)++)=(unsigned char)(((l)>>16L)&0xff), \
146                          *((c)++)=(unsigned char)(((l)>> 8L)&0xff), \
147                          *((c)++)=(unsigned char)(((l)     )&0xff))
148
149 #if defined(OPENSSL_SYS_WIN32) && defined(_MSC_VER)
150 #define ROTATE_l32(a,n)     _lrotl(a,n)
151 #define ROTATE_r32(a,n)     _lrotr(a,n)
152 #elif defined(__GNUC__) && __GNUC__>=2 && !defined(NO_ASM) && !defined(NO_INLINE_ASM)
153 # if defined(__i386) || defined(__i386__) || defined(__x86_64) || defined(__x86_64__)
154 #  define ROTATE_l32(a,n)       ({ register unsigned int ret;   \
155                                         asm ("roll %%cl,%0"     \
156                                                 : "=r"(ret)     \
157                                                 : "c"(n),"0"(a) \
158                                                 : "cc");        \
159                                         ret;                    \
160                                 })
161 #  define ROTATE_r32(a,n)       ({ register unsigned int ret;   \
162                                         asm ("rorl %%cl,%0"     \
163                                                 : "=r"(ret)     \
164                                                 : "c"(n),"0"(a) \
165                                                 : "cc");        \
166                                         ret;                    \
167                                 })
168 # endif
169 #else
170 #define ROTATE_l32(a,n)     (((a)<<(n&0x1f))|(((a)&0xffffffff)>>(32-(n&0x1f))))
171 #define ROTATE_r32(a,n)     (((a)<<(32-(n&0x1f)))|(((a)&0xffffffff)>>(n&0x1f)))
172 #endif
173
174 #define RC5_32_MASK     0xffffffffL
175
176 #define RC5_16_P        0xB7E1
177 #define RC5_16_Q        0x9E37
178 #define RC5_32_P        0xB7E15163L
179 #define RC5_32_Q        0x9E3779B9L
180 #define RC5_64_P        0xB7E151628AED2A6BLL
181 #define RC5_64_Q        0x9E3779B97F4A7C15LL
182
183 #define E_RC5_32(a,b,s,n) \
184         a^=b; \
185         a=ROTATE_l32(a,b); \
186         a+=s[n]; \
187         a&=RC5_32_MASK; \
188         b^=a; \
189         b=ROTATE_l32(b,a); \
190         b+=s[n+1]; \
191         b&=RC5_32_MASK;
192
193 #define D_RC5_32(a,b,s,n) \
194         b-=s[n+1]; \
195         b&=RC5_32_MASK; \
196         b=ROTATE_r32(b,a); \
197         b^=a; \
198         a-=s[n]; \
199         a&=RC5_32_MASK; \
200         a=ROTATE_r32(a,b); \
201         a^=b;
202
203
204