| Andrew Scull | b4b6d4a | 2019-01-02 15:54:55 +0000 | [diff] [blame^] | 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * arch/openrisc/lib/memcpy.c |
| 4 | * |
| 5 | * Optimized memory copy routines for openrisc. These are mostly copied |
| 6 | * from ohter sources but slightly entended based on ideas discuassed in |
| 7 | * #openrisc. |
| 8 | * |
| 9 | * The word unroll implementation is an extension to the arm byte |
| 10 | * unrolled implementation, but using word copies (if things are |
| 11 | * properly aligned) |
| 12 | * |
| 13 | * The great arm loop unroll algorithm can be found at: |
| 14 | * arch/arm/boot/compressed/string.c |
| 15 | */ |
| 16 | |
| 17 | #include <linux/export.h> |
| 18 | |
| 19 | #include <linux/string.h> |
| 20 | |
| 21 | #ifdef CONFIG_OR1K_1200 |
| 22 | /* |
| 23 | * Do memcpy with word copies and loop unrolling. This gives the |
| 24 | * best performance on the OR1200 and MOR1KX archirectures |
| 25 | */ |
| 26 | void *memcpy(void *dest, __const void *src, __kernel_size_t n) |
| 27 | { |
| 28 | int i = 0; |
| 29 | unsigned char *d, *s; |
| 30 | uint32_t *dest_w = (uint32_t *)dest, *src_w = (uint32_t *)src; |
| 31 | |
| 32 | /* If both source and dest are word aligned copy words */ |
| 33 | if (!((unsigned int)dest_w & 3) && !((unsigned int)src_w & 3)) { |
| 34 | /* Copy 32 bytes per loop */ |
| 35 | for (i = n >> 5; i > 0; i--) { |
| 36 | *dest_w++ = *src_w++; |
| 37 | *dest_w++ = *src_w++; |
| 38 | *dest_w++ = *src_w++; |
| 39 | *dest_w++ = *src_w++; |
| 40 | *dest_w++ = *src_w++; |
| 41 | *dest_w++ = *src_w++; |
| 42 | *dest_w++ = *src_w++; |
| 43 | *dest_w++ = *src_w++; |
| 44 | } |
| 45 | |
| 46 | if (n & 1 << 4) { |
| 47 | *dest_w++ = *src_w++; |
| 48 | *dest_w++ = *src_w++; |
| 49 | *dest_w++ = *src_w++; |
| 50 | *dest_w++ = *src_w++; |
| 51 | } |
| 52 | |
| 53 | if (n & 1 << 3) { |
| 54 | *dest_w++ = *src_w++; |
| 55 | *dest_w++ = *src_w++; |
| 56 | } |
| 57 | |
| 58 | if (n & 1 << 2) |
| 59 | *dest_w++ = *src_w++; |
| 60 | |
| 61 | d = (unsigned char *)dest_w; |
| 62 | s = (unsigned char *)src_w; |
| 63 | |
| 64 | } else { |
| 65 | d = (unsigned char *)dest_w; |
| 66 | s = (unsigned char *)src_w; |
| 67 | |
| 68 | for (i = n >> 3; i > 0; i--) { |
| 69 | *d++ = *s++; |
| 70 | *d++ = *s++; |
| 71 | *d++ = *s++; |
| 72 | *d++ = *s++; |
| 73 | *d++ = *s++; |
| 74 | *d++ = *s++; |
| 75 | *d++ = *s++; |
| 76 | *d++ = *s++; |
| 77 | } |
| 78 | |
| 79 | if (n & 1 << 2) { |
| 80 | *d++ = *s++; |
| 81 | *d++ = *s++; |
| 82 | *d++ = *s++; |
| 83 | *d++ = *s++; |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | if (n & 1 << 1) { |
| 88 | *d++ = *s++; |
| 89 | *d++ = *s++; |
| 90 | } |
| 91 | |
| 92 | if (n & 1) |
| 93 | *d++ = *s++; |
| 94 | |
| 95 | return dest; |
| 96 | } |
| 97 | #else |
| 98 | /* |
| 99 | * Use word copies but no loop unrolling as we cannot assume there |
| 100 | * will be benefits on the archirecture |
| 101 | */ |
| 102 | void *memcpy(void *dest, __const void *src, __kernel_size_t n) |
| 103 | { |
| 104 | unsigned char *d = (unsigned char *)dest, *s = (unsigned char *)src; |
| 105 | uint32_t *dest_w = (uint32_t *)dest, *src_w = (uint32_t *)src; |
| 106 | |
| 107 | /* If both source and dest are word aligned copy words */ |
| 108 | if (!((unsigned int)dest_w & 3) && !((unsigned int)src_w & 3)) { |
| 109 | for (; n >= 4; n -= 4) |
| 110 | *dest_w++ = *src_w++; |
| 111 | } |
| 112 | |
| 113 | d = (unsigned char *)dest_w; |
| 114 | s = (unsigned char *)src_w; |
| 115 | |
| 116 | /* For remaining or if not aligned, copy bytes */ |
| 117 | for (; n >= 1; n -= 1) |
| 118 | *d++ = *s++; |
| 119 | |
| 120 | return dest; |
| 121 | |
| 122 | } |
| 123 | #endif |
| 124 | |
| 125 | EXPORT_SYMBOL(memcpy); |