blob: 795b9e5b9e6ab88e07641189eaba1c3110a7cba4 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001#ifndef __ASM_ARM_UNALIGNED_H
2#define __ASM_ARM_UNALIGNED_H
3
4#include <asm/types.h>
5
Lennert Buytenhek17b602b2006-09-21 02:24:38 +01006extern int __bug_unaligned_x(const void *ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007
8/*
9 * What is the most efficient way of loading/storing an unaligned value?
10 *
11 * That is the subject of this file. Efficiency here is defined as
12 * minimum code size with minimum register usage for the common cases.
13 * It is currently not believed that long longs are common, so we
14 * trade efficiency for the chars, shorts and longs against the long
15 * longs.
16 *
17 * Current stats with gcc 2.7.2.2 for these functions:
18 *
19 * ptrsize get: code regs put: code regs
20 * 1 1 1 1 2
21 * 2 3 2 3 2
22 * 4 7 3 7 3
23 * 8 20 6 16 6
24 *
25 * gcc 2.95.1 seems to code differently:
26 *
27 * ptrsize get: code regs put: code regs
28 * 1 1 1 1 2
29 * 2 3 2 3 2
30 * 4 7 4 7 4
31 * 8 19 8 15 6
32 *
33 * which may or may not be more efficient (depending upon whether
34 * you can afford the extra registers). Hopefully the gcc 2.95
35 * is inteligent enough to decide if it is better to use the
36 * extra register, but evidence so far seems to suggest otherwise.
37 *
38 * Unfortunately, gcc is not able to optimise the high word
39 * out of long long >> 32, or the low word from long long << 32
40 */
41
42#define __get_unaligned_2_le(__p) \
43 (__p[0] | __p[1] << 8)
44
45#define __get_unaligned_2_be(__p) \
46 (__p[0] << 8 | __p[1])
47
48#define __get_unaligned_4_le(__p) \
49 (__p[0] | __p[1] << 8 | __p[2] << 16 | __p[3] << 24)
50
51#define __get_unaligned_4_be(__p) \
52 (__p[0] << 24 | __p[1] << 16 | __p[2] << 8 | __p[3])
53
Lennert Buytenhek17b602b2006-09-21 02:24:38 +010054#define __get_unaligned_8_le(__p) \
55 ((unsigned long long)__get_unaligned_4_le((__p+4)) << 32 | \
56 __get_unaligned_4_le(__p))
57
58#define __get_unaligned_8_be(__p) \
59 ((unsigned long long)__get_unaligned_4_be(__p) << 32 | \
60 __get_unaligned_4_be((__p+4)))
61
62#define __get_unaligned_le(ptr) \
63 ({ \
64 const __u8 *__p = (const __u8 *)(ptr); \
65 __builtin_choose_expr(sizeof(*(ptr)) == 1, *__p, \
66 __builtin_choose_expr(sizeof(*(ptr)) == 2, __get_unaligned_2_le(__p), \
67 __builtin_choose_expr(sizeof(*(ptr)) == 4, __get_unaligned_4_le(__p), \
68 __builtin_choose_expr(sizeof(*(ptr)) == 8, __get_unaligned_8_le(__p), \
69 (void)__bug_unaligned_x(__p))))); \
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 })
71
Lennert Buytenhek17b602b2006-09-21 02:24:38 +010072#define __get_unaligned_be(ptr) \
73 ({ \
74 const __u8 *__p = (const __u8 *)(ptr); \
75 __builtin_choose_expr(sizeof(*(ptr)) == 1, *__p, \
76 __builtin_choose_expr(sizeof(*(ptr)) == 2, __get_unaligned_2_be(__p), \
77 __builtin_choose_expr(sizeof(*(ptr)) == 4, __get_unaligned_4_be(__p), \
78 __builtin_choose_expr(sizeof(*(ptr)) == 8, __get_unaligned_8_be(__p), \
79 (void)__bug_unaligned_x(__p))))); \
Linus Torvalds1da177e2005-04-16 15:20:36 -070080 })
81
82
83static inline void __put_unaligned_2_le(__u32 __v, register __u8 *__p)
84{
85 *__p++ = __v;
86 *__p++ = __v >> 8;
87}
88
89static inline void __put_unaligned_2_be(__u32 __v, register __u8 *__p)
90{
91 *__p++ = __v >> 8;
92 *__p++ = __v;
93}
94
95static inline void __put_unaligned_4_le(__u32 __v, register __u8 *__p)
96{
97 __put_unaligned_2_le(__v >> 16, __p + 2);
98 __put_unaligned_2_le(__v, __p);
99}
100
101static inline void __put_unaligned_4_be(__u32 __v, register __u8 *__p)
102{
103 __put_unaligned_2_be(__v >> 16, __p);
104 __put_unaligned_2_be(__v, __p + 2);
105}
106
107static inline void __put_unaligned_8_le(const unsigned long long __v, register __u8 *__p)
108{
109 /*
110 * tradeoff: 8 bytes of stack for all unaligned puts (2
111 * instructions), or an extra register in the long long
112 * case - go for the extra register.
113 */
114 __put_unaligned_4_le(__v >> 32, __p+4);
115 __put_unaligned_4_le(__v, __p);
116}
117
118static inline void __put_unaligned_8_be(const unsigned long long __v, register __u8 *__p)
119{
120 /*
121 * tradeoff: 8 bytes of stack for all unaligned puts (2
122 * instructions), or an extra register in the long long
123 * case - go for the extra register.
124 */
125 __put_unaligned_4_be(__v >> 32, __p);
126 __put_unaligned_4_be(__v, __p+4);
127}
128
129/*
130 * Try to store an unaligned value as efficiently as possible.
131 */
132#define __put_unaligned_le(val,ptr) \
133 ({ \
134 switch (sizeof(*(ptr))) { \
135 case 1: \
136 *(ptr) = (val); \
137 break; \
138 case 2: __put_unaligned_2_le((val),(__u8 *)(ptr)); \
139 break; \
140 case 4: __put_unaligned_4_le((val),(__u8 *)(ptr)); \
141 break; \
142 case 8: __put_unaligned_8_le((val),(__u8 *)(ptr)); \
143 break; \
144 default: __bug_unaligned_x(ptr); \
145 break; \
146 } \
147 (void) 0; \
148 })
149
150#define __put_unaligned_be(val,ptr) \
151 ({ \
152 switch (sizeof(*(ptr))) { \
153 case 1: \
154 *(ptr) = (val); \
155 break; \
156 case 2: __put_unaligned_2_be((val),(__u8 *)(ptr)); \
157 break; \
158 case 4: __put_unaligned_4_be((val),(__u8 *)(ptr)); \
159 break; \
160 case 8: __put_unaligned_8_be((val),(__u8 *)(ptr)); \
161 break; \
162 default: __bug_unaligned_x(ptr); \
163 break; \
164 } \
165 (void) 0; \
166 })
167
168/*
169 * Select endianness
170 */
171#ifndef __ARMEB__
172#define get_unaligned __get_unaligned_le
173#define put_unaligned __put_unaligned_le
174#else
175#define get_unaligned __get_unaligned_be
176#define put_unaligned __put_unaligned_be
177#endif
178
179#endif