| 1 | /* memcopy.h -- definitions for memory copy functions. Generic C version. |
| 2 | Copyright (C) 1991-2022 Free Software Foundation, Inc. |
| 3 | This file is part of the GNU C Library. |
| 4 | |
| 5 | The GNU C Library is free software; you can redistribute it and/or |
| 6 | modify it under the terms of the GNU Lesser General Public |
| 7 | License as published by the Free Software Foundation; either |
| 8 | version 2.1 of the License, or (at your option) any later version. |
| 9 | |
| 10 | The GNU C Library is distributed in the hope that it will be useful, |
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 13 | Lesser General Public License for more details. |
| 14 | |
| 15 | You should have received a copy of the GNU Lesser General Public |
| 16 | License along with the GNU C Library; if not, see |
| 17 | <https://www.gnu.org/licenses/>. */ |
| 18 | |
| 19 | #ifndef _MEMCOPY_H |
| 20 | #define _MEMCOPY_H 1 |
| 21 | |
| 22 | /* The strategy of the memory functions is: |
| 23 | |
| 24 | 1. Copy bytes until the destination pointer is aligned. |
| 25 | |
| 26 | 2. Copy words in unrolled loops. If the source and destination |
| 27 | are not aligned in the same way, use word memory operations, |
| 28 | but shift and merge two read words before writing. |
| 29 | |
| 30 | 3. Copy the few remaining bytes. |
| 31 | |
| 32 | This is fast on processors that have at least 10 registers for |
| 33 | allocation by GCC, and that can access memory at reg+const in one |
| 34 | instruction. |
| 35 | |
| 36 | I made an "exhaustive" test of this memmove when I wrote it, |
| 37 | exhaustive in the sense that I tried all alignment and length |
| 38 | combinations, with and without overlap. */ |
| 39 | |
| 40 | #include <sys/cdefs.h> |
| 41 | #include <endian.h> |
| 42 | #include <pagecopy.h> |
| 43 | |
| 44 | /* The macros defined in this file are: |
| 45 | |
| 46 | BYTE_COPY_FWD(dst_beg_ptr, src_beg_ptr, nbytes_to_copy) |
| 47 | |
| 48 | BYTE_COPY_BWD(dst_end_ptr, src_end_ptr, nbytes_to_copy) |
| 49 | |
| 50 | WORD_COPY_FWD(dst_beg_ptr, src_beg_ptr, nbytes_remaining, nbytes_to_copy) |
| 51 | |
| 52 | WORD_COPY_BWD(dst_end_ptr, src_end_ptr, nbytes_remaining, nbytes_to_copy) |
| 53 | |
| 54 | MERGE(old_word, sh_1, new_word, sh_2) |
| 55 | [I fail to understand. I feel stupid. --roland] |
| 56 | */ |
| 57 | |
| 58 | /* Type to use for aligned memory operations. |
| 59 | This should normally be the biggest type supported by a single load |
| 60 | and store. */ |
| 61 | #define op_t unsigned long int |
| 62 | #define OPSIZ (sizeof (op_t)) |
| 63 | |
| 64 | /* Type to use for unaligned operations. */ |
| 65 | typedef unsigned char byte; |
| 66 | |
| 67 | #if __BYTE_ORDER == __LITTLE_ENDIAN |
| 68 | #define MERGE(w0, sh_1, w1, sh_2) (((w0) >> (sh_1)) | ((w1) << (sh_2))) |
| 69 | #endif |
| 70 | #if __BYTE_ORDER == __BIG_ENDIAN |
| 71 | #define MERGE(w0, sh_1, w1, sh_2) (((w0) << (sh_1)) | ((w1) >> (sh_2))) |
| 72 | #endif |
| 73 | |
| 74 | /* Copy exactly NBYTES bytes from SRC_BP to DST_BP, |
| 75 | without any assumptions about alignment of the pointers. */ |
| 76 | #define BYTE_COPY_FWD(dst_bp, src_bp, nbytes) \ |
| 77 | do \ |
| 78 | { \ |
| 79 | size_t __nbytes = (nbytes); \ |
| 80 | while (__nbytes > 0) \ |
| 81 | { \ |
| 82 | byte __x = ((byte *) src_bp)[0]; \ |
| 83 | src_bp += 1; \ |
| 84 | __nbytes -= 1; \ |
| 85 | ((byte *) dst_bp)[0] = __x; \ |
| 86 | dst_bp += 1; \ |
| 87 | } \ |
| 88 | } while (0) |
| 89 | |
| 90 | /* Copy exactly NBYTES_TO_COPY bytes from SRC_END_PTR to DST_END_PTR, |
| 91 | beginning at the bytes right before the pointers and continuing towards |
| 92 | smaller addresses. Don't assume anything about alignment of the |
| 93 | pointers. */ |
| 94 | #define BYTE_COPY_BWD(dst_ep, src_ep, nbytes) \ |
| 95 | do \ |
| 96 | { \ |
| 97 | size_t __nbytes = (nbytes); \ |
| 98 | while (__nbytes > 0) \ |
| 99 | { \ |
| 100 | byte __x; \ |
| 101 | src_ep -= 1; \ |
| 102 | __x = ((byte *) src_ep)[0]; \ |
| 103 | dst_ep -= 1; \ |
| 104 | __nbytes -= 1; \ |
| 105 | ((byte *) dst_ep)[0] = __x; \ |
| 106 | } \ |
| 107 | } while (0) |
| 108 | |
| 109 | /* Copy *up to* NBYTES bytes from SRC_BP to DST_BP, with |
| 110 | the assumption that DST_BP is aligned on an OPSIZ multiple. If |
| 111 | not all bytes could be easily copied, store remaining number of bytes |
| 112 | in NBYTES_LEFT, otherwise store 0. */ |
| 113 | extern void _wordcopy_fwd_aligned (long int, long int, size_t) |
| 114 | attribute_hidden __THROW; |
| 115 | extern void _wordcopy_fwd_dest_aligned (long int, long int, size_t) |
| 116 | attribute_hidden __THROW; |
| 117 | #define WORD_COPY_FWD(dst_bp, src_bp, nbytes_left, nbytes) \ |
| 118 | do \ |
| 119 | { \ |
| 120 | if (src_bp % OPSIZ == 0) \ |
| 121 | _wordcopy_fwd_aligned (dst_bp, src_bp, (nbytes) / OPSIZ); \ |
| 122 | else \ |
| 123 | _wordcopy_fwd_dest_aligned (dst_bp, src_bp, (nbytes) / OPSIZ); \ |
| 124 | src_bp += (nbytes) & -OPSIZ; \ |
| 125 | dst_bp += (nbytes) & -OPSIZ; \ |
| 126 | (nbytes_left) = (nbytes) % OPSIZ; \ |
| 127 | } while (0) |
| 128 | |
| 129 | /* Copy *up to* NBYTES_TO_COPY bytes from SRC_END_PTR to DST_END_PTR, |
| 130 | beginning at the words (of type op_t) right before the pointers and |
| 131 | continuing towards smaller addresses. May take advantage of that |
| 132 | DST_END_PTR is aligned on an OPSIZ multiple. If not all bytes could be |
| 133 | easily copied, store remaining number of bytes in NBYTES_REMAINING, |
| 134 | otherwise store 0. */ |
| 135 | extern void _wordcopy_bwd_aligned (long int, long int, size_t) |
| 136 | attribute_hidden __THROW; |
| 137 | extern void _wordcopy_bwd_dest_aligned (long int, long int, size_t) |
| 138 | attribute_hidden __THROW; |
| 139 | #define WORD_COPY_BWD(dst_ep, src_ep, nbytes_left, nbytes) \ |
| 140 | do \ |
| 141 | { \ |
| 142 | if (src_ep % OPSIZ == 0) \ |
| 143 | _wordcopy_bwd_aligned (dst_ep, src_ep, (nbytes) / OPSIZ); \ |
| 144 | else \ |
| 145 | _wordcopy_bwd_dest_aligned (dst_ep, src_ep, (nbytes) / OPSIZ); \ |
| 146 | src_ep -= (nbytes) & -OPSIZ; \ |
| 147 | dst_ep -= (nbytes) & -OPSIZ; \ |
| 148 | (nbytes_left) = (nbytes) % OPSIZ; \ |
| 149 | } while (0) |
| 150 | |
| 151 | /* The macro PAGE_COPY_FWD_MAYBE (dstp, srcp, nbytes_left, nbytes) is invoked |
| 152 | like WORD_COPY_FWD et al. The pointers should be at least word aligned. |
| 153 | This will check if virtual copying by pages can and should be done and do it |
| 154 | if so. The pointers will be aligned to PAGE_SIZE bytes. The macro requires |
| 155 | that pagecopy.h defines at least PAGE_COPY_THRESHOLD to 0. If |
| 156 | PAGE_COPY_THRESHOLD is non-zero, the header must also define PAGE_COPY_FWD |
| 157 | and PAGE_SIZE. |
| 158 | */ |
| 159 | #if PAGE_COPY_THRESHOLD |
| 160 | |
| 161 | # include <assert.h> |
| 162 | |
| 163 | # define PAGE_COPY_FWD_MAYBE(dstp, srcp, nbytes_left, nbytes) \ |
| 164 | do \ |
| 165 | { \ |
| 166 | if ((nbytes) >= PAGE_COPY_THRESHOLD \ |
| 167 | && PAGE_OFFSET ((dstp) - (srcp)) == 0) \ |
| 168 | { \ |
| 169 | /* The amount to copy is past the threshold for copying \ |
| 170 | pages virtually with kernel VM operations, and the \ |
| 171 | source and destination addresses have the same alignment. */ \ |
| 172 | size_t nbytes_before = PAGE_OFFSET (-(dstp)); \ |
| 173 | if (nbytes_before != 0) \ |
| 174 | { \ |
| 175 | /* First copy the words before the first page boundary. */ \ |
| 176 | WORD_COPY_FWD (dstp, srcp, nbytes_left, nbytes_before); \ |
| 177 | assert (nbytes_left == 0); \ |
| 178 | nbytes -= nbytes_before; \ |
| 179 | } \ |
| 180 | PAGE_COPY_FWD (dstp, srcp, nbytes_left, nbytes); \ |
| 181 | } \ |
| 182 | } while (0) |
| 183 | |
| 184 | /* The page size is always a power of two, so we can avoid modulo division. */ |
| 185 | # define PAGE_OFFSET(n) ((n) & (PAGE_SIZE - 1)) |
| 186 | |
| 187 | #else |
| 188 | |
| 189 | # define PAGE_COPY_FWD_MAYBE(dstp, srcp, nbytes_left, nbytes) /* nada */ |
| 190 | |
| 191 | #endif |
| 192 | |
| 193 | /* Threshold value for when to enter the unrolled loops. */ |
| 194 | #define OP_T_THRES 16 |
| 195 | |
| 196 | /* Set to 1 if memcpy is safe to use for forward-copying memmove with |
| 197 | overlapping addresses. This is 0 by default because memcpy implementations |
| 198 | are generally not safe for overlapping addresses. */ |
| 199 | #define MEMCPY_OK_FOR_FWD_MEMMOVE 0 |
| 200 | |
| 201 | #endif /* memcopy.h */ |
| 202 | |