| 1 | |
| 2 | |
| 3 | * linux/arch/alpha/lib/memcpy.c
|
| 4 | *
|
| 5 | * Copyright (C) 1995 Linus Torvalds
|
| 6 |
|
| 7 |
|
| 8 | |
| 9 | * This is a reasonably optimized memcpy() routine.
|
| 10 |
|
| 11 |
|
| 12 | |
| 13 | * Note that the C code is written to be optimized into good assembly. However,
|
| 14 | * at this point gcc is unable to sanely compile "if (n >= 0)", resulting in a
|
| 15 | * explicit compare against 0 (instead of just using the proper "blt reg, xx" or
|
| 16 | * "bge reg, xx"). I hope alpha-gcc will be fixed to notice this eventually..
|
| 17 |
|
| 18 |
|
| 19 | #include <linux/types.h>
|
| 20 | #include <linux/export.h>
|
| 21 | #include <linux/string.h>
|
| 22 |
|
| 23 | |
| 24 | * This should be done in one go with ldq_u*2/mask/stq_u. Do it
|
| 25 | * with a macro so that we can fix it up later..
|
| 26 |
|
| 27 | #define ALIGN_DEST_TO8_UP(d,s,n) \
|
| 28 | while (d & 7) { \
|
| 29 | if (n <= 0) return; \
|
| 30 | n--; \
|
| 31 | *(char *) d = *(char *) s; \
|
| 32 | d++; s++; \
|
| 33 | }
|
| 34 | #define ALIGN_DEST_TO8_DN(d,s,n) \
|
| 35 | while (d & 7) { \
|
| 36 | if (n <= 0) return; \
|
| 37 | n--; \
|
| 38 | d--; s--; \
|
| 39 | *(char *) d = *(char *) s; \
|
| 40 | }
|
| 41 |
|
| 42 | |
| 43 | * This should similarly be done with ldq_u*2/mask/stq. The destination
|
| 44 | * is aligned, but we don't fill in a full quad-word
|
| 45 |
|
| 46 | #define DO_REST_UP(d,s,n) \
|
| 47 | while (n > 0) { \
|
| 48 | n--; \
|
| 49 | *(char *) d = *(char *) s; \
|
| 50 | d++; s++; \
|
| 51 | }
|
| 52 | #define DO_REST_DN(d,s,n) \
|
| 53 | while (n > 0) { \
|
| 54 | n--; \
|
| 55 | d--; s--; \
|
| 56 | *(char *) d = *(char *) s; \
|
| 57 | }
|
| 58 |
|
| 59 | |
| 60 | * This should be done with ldq/mask/stq. The source and destination are
|
| 61 | * aligned, but we don't fill in a full quad-word
|
| 62 |
|
| 63 | #define DO_REST_ALIGNED_UP(d,s,n) DO_REST_UP(d,s,n)
|
| 64 | #define DO_REST_ALIGNED_DN(d,s,n) DO_REST_DN(d,s,n)
|
| 65 |
|
| 66 | |
| 67 | * This does unaligned memory copies. We want to avoid storing to
|
| 68 | * an unaligned address, as that would do a read-modify-write cycle.
|
| 69 | * We also want to avoid double-reading the unaligned reads.
|
| 70 | *
|
| 71 | * Note the ordering to try to avoid load (and address generation) latencies.
|
| 72 |
|
| 73 | static inline void __memcpy_unaligned_up (unsigned long d, unsigned long s,
|
| 74 | long n)
|
| 75 | {
|
| 76 | ALIGN_DEST_TO8_UP(d,s,n);
|
| 77 | n -= 8;
|
| 78 | if (n >= 0) {
|
| 79 | unsigned long low_word, high_word;
|
| 80 | __asm__("ldq_u %0,%1":"=r" (low_word):"m" (*(unsigned long *) s));
|
| 81 | do {
|
| 82 | unsigned long tmp;
|
| 83 | __asm__("ldq_u %0,%1":"=r" (high_word):"m" (*(unsigned long *)(s+8)));
|
| 84 | n -= 8;
|
| 85 | __asm__("extql %1,%2,%0"
|
| 86 | :"=r" (low_word)
|
| 87 | :"r" (low_word), "r" (s));
|
| 88 | __asm__("extqh %1,%2,%0"
|
| 89 | :"=r" (tmp)
|
| 90 | :"r" (high_word), "r" (s));
|
| 91 | s += 8;
|
| 92 | *(unsigned long *) d = low_word | tmp;
|
| 93 | d += 8;
|
| 94 | low_word = high_word;
|
| 95 | } while (n >= 0);
|
| 96 | }
|
| 97 | n += 8;
|
| 98 | DO_REST_UP(d,s,n);
|
| 99 | }
|
| 100 |
|
| 101 | static inline void __memcpy_unaligned_dn (unsigned long d, unsigned long s,
|
| 102 | long n)
|
| 103 | {
|
| 104 |
|
| 105 | s += n;
|
| 106 | d += n;
|
| 107 | while (n--)
|
| 108 | * (char *) --d = * (char *) --s;
|
| 109 | }
|
| 110 |
|
| 111 | |
| 112 | * Hmm.. Strange. The __asm__ here is there to make gcc use an integer register
|
| 113 | * for the load-store. I don't know why, but it would seem that using a floating
|
| 114 | * point register for the move seems to slow things down (very small difference,
|
| 115 | * though).
|
| 116 | *
|
| 117 | * Note the ordering to try to avoid load (and address generation) latencies.
|
| 118 |
|
| 119 | static inline void __memcpy_aligned_up (unsigned long d, unsigned long s,
|
| 120 | long n)
|
| 121 | {
|
| 122 | ALIGN_DEST_TO8_UP(d,s,n);
|
| 123 | n -= 8;
|
| 124 | while (n >= 0) {
|
| 125 | unsigned long tmp;
|
| 126 | __asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
|
| 127 | n -= 8;
|
| 128 | s += 8;
|
| 129 | *(unsigned long *) d = tmp;
|
| 130 | d += 8;
|
| 131 | }
|
| 132 | n += 8;
|
| 133 | DO_REST_ALIGNED_UP(d,s,n);
|
| 134 | }
|
| 135 | static inline void __memcpy_aligned_dn (unsigned long d, unsigned long s,
|
| 136 | long n)
|
| 137 | {
|
| 138 | s += n;
|
| 139 | d += n;
|
| 140 | ALIGN_DEST_TO8_DN(d,s,n);
|
| 141 | n -= 8;
|
| 142 | while (n >= 0) {
|
| 143 | unsigned long tmp;
|
| 144 | s -= 8;
|
| 145 | __asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
|
| 146 | n -= 8;
|
| 147 | d -= 8;
|
| 148 | *(unsigned long *) d = tmp;
|
| 149 | }
|
| 150 | n += 8;
|
| 151 | DO_REST_ALIGNED_DN(d,s,n);
|
| 152 | }
|
| 153 |
|
| 154 | #undef memcpy
|
| 155 |
|
| 156 | void * memcpy(void * dest, const void *src, size_t n)
|
| 157 | {
|
| 158 | if (!(((unsigned long) dest ^ (unsigned long) src) & 7)) {
|
| 159 | __memcpy_aligned_up ((unsigned long) dest, (unsigned long) src,
|
| 160 | n);
|
| 161 | return dest;
|
| 162 | }
|
| 163 | __memcpy_unaligned_up ((unsigned long) dest, (unsigned long) src, n);
|
| 164 | return dest;
|
| 165 | }
|
| 166 | EXPORT_SYMBOL(memcpy);
|