Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | #ifndef __ASM_ARM_UNALIGNED_H |
2 | #define __ASM_ARM_UNALIGNED_H | |
3 | ||
4 | #include <asm/types.h> | |
5 | ||
6 | extern int __bug_unaligned_x(void *ptr); | |
7 | ||
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_4_le(__p) \ | |
46 | (__p[0] | __p[1] << 8 | __p[2] << 16 | __p[3] << 24) | |
47 | ||
48 | #define __get_unaligned_le(ptr) \ | |
49 | ({ \ | |
50 | __typeof__(*(ptr)) __v; \ | |
51 | __u8 *__p = (__u8 *)(ptr); \ | |
52 | switch (sizeof(*(ptr))) { \ | |
53 | case 1: __v = *(ptr); break; \ | |
54 | case 2: __v = __get_unaligned_2_le(__p); break; \ | |
55 | case 4: __v = __get_unaligned_4_le(__p); break; \ | |
56 | case 8: { \ | |
57 | unsigned int __v1, __v2; \ | |
58 | __v2 = __get_unaligned_4_le((__p+4)); \ | |
59 | __v1 = __get_unaligned_4_le(__p); \ | |
60 | __v = ((unsigned long long)__v2 << 32 | __v1); \ | |
61 | } \ | |
62 | break; \ | |
63 | default: __v = __bug_unaligned_x(__p); break; \ | |
64 | } \ | |
65 | __v; \ | |
66 | }) | |
67 | ||
68 | static inline void __put_unaligned_2_le(__u32 __v, register __u8 *__p) | |
69 | { | |
70 | *__p++ = __v; | |
71 | *__p++ = __v >> 8; | |
72 | } | |
73 | ||
74 | static inline void __put_unaligned_4_le(__u32 __v, register __u8 *__p) | |
75 | { | |
76 | __put_unaligned_2_le(__v >> 16, __p + 2); | |
77 | __put_unaligned_2_le(__v, __p); | |
78 | } | |
79 | ||
80 | static inline void __put_unaligned_8_le(const unsigned long long __v, register __u8 *__p) | |
81 | { | |
82 | /* | |
83 | * tradeoff: 8 bytes of stack for all unaligned puts (2 | |
84 | * instructions), or an extra register in the long long | |
85 | * case - go for the extra register. | |
86 | */ | |
87 | __put_unaligned_4_le(__v >> 32, __p+4); | |
88 | __put_unaligned_4_le(__v, __p); | |
89 | } | |
90 | ||
91 | /* | |
92 | * Try to store an unaligned value as efficiently as possible. | |
93 | */ | |
94 | #define __put_unaligned_le(val,ptr) \ | |
95 | ({ \ | |
96 | switch (sizeof(*(ptr))) { \ | |
97 | case 1: \ | |
98 | *(ptr) = (val); \ | |
99 | break; \ | |
100 | case 2: __put_unaligned_2_le((val),(__u8 *)(ptr)); \ | |
101 | break; \ | |
102 | case 4: __put_unaligned_4_le((val),(__u8 *)(ptr)); \ | |
103 | break; \ | |
104 | case 8: __put_unaligned_8_le((val),(__u8 *)(ptr)); \ | |
105 | break; \ | |
106 | default: __bug_unaligned_x(ptr); \ | |
107 | break; \ | |
108 | } \ | |
109 | (void) 0; \ | |
110 | }) | |
111 | ||
112 | /* | |
113 | * Select endianness | |
114 | */ | |
115 | #define get_unaligned __get_unaligned_le | |
116 | #define put_unaligned __put_unaligned_le | |
117 | ||
118 | #endif |