xf.li | bfc6e71 | 2025-02-07 01:54:34 -0800 | [diff] [blame^] | 1 | /* Copyright (C) 2011-2016 Free Software Foundation, Inc. |
| 2 | This file is part of the GNU C Library. |
| 3 | Contributed by Chris Metcalf <cmetcalf@tilera.com>, 2011. |
| 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 | <http://www.gnu.org/licenses/>. */ |
| 18 | |
| 19 | #include <arch/chip.h> |
| 20 | #include <string.h> |
| 21 | #include <stdint.h> |
| 22 | #include "string-endian.h" |
| 23 | |
| 24 | void * inhibit_loop_to_libcall |
| 25 | __memset (void *s, int c, size_t n) |
| 26 | { |
| 27 | uint64_t *out64; |
| 28 | int n64, to_align64; |
| 29 | uint64_t v64; |
| 30 | uint8_t *out8 = s; |
| 31 | |
| 32 | /* Experimentation shows that a trivial tight loop is a win up until |
| 33 | around a size of 20, where writing a word at a time starts to win. */ |
| 34 | #define BYTE_CUTOFF 20 |
| 35 | |
| 36 | #if BYTE_CUTOFF < 7 |
| 37 | /* This must be at least at least this big, or some code later |
| 38 | on doesn't work. */ |
| 39 | # error "BYTE_CUTOFF is too small." |
| 40 | #endif |
| 41 | |
| 42 | if (n < BYTE_CUTOFF) |
| 43 | { |
| 44 | /* Strangely, this turns out to be the tightest way to write |
| 45 | this loop. */ |
| 46 | if (n != 0) |
| 47 | { |
| 48 | do |
| 49 | { |
| 50 | /* Strangely, combining these into one line performs worse. */ |
| 51 | *out8 = c; |
| 52 | out8++; |
| 53 | } |
| 54 | while (--n != 0); |
| 55 | } |
| 56 | |
| 57 | return s; |
| 58 | } |
| 59 | |
| 60 | /* Align 'out8'. We know n >= 7 so this won't write past the end. */ |
| 61 | while (((uintptr_t) out8 & 7) != 0) |
| 62 | { |
| 63 | *out8++ = c; |
| 64 | --n; |
| 65 | } |
| 66 | |
| 67 | /* Align 'n'. */ |
| 68 | while (n & 7) |
| 69 | out8[--n] = c; |
| 70 | |
| 71 | out64 = (uint64_t *) out8; |
| 72 | n64 = n >> 3; |
| 73 | |
| 74 | /* Tile input byte out to 64 bits. */ |
| 75 | v64 = copy_byte(c); |
| 76 | |
| 77 | /* This must be at least 8 or the following loop doesn't work. */ |
| 78 | #define CACHE_LINE_SIZE_IN_DOUBLEWORDS (CHIP_L2_LINE_SIZE() / 8) |
| 79 | |
| 80 | /* Determine how many words we need to emit before the 'out32' |
| 81 | pointer becomes aligned modulo the cache line size. */ |
| 82 | to_align64 = (-((uintptr_t) out64 >> 3)) & |
| 83 | (CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1); |
| 84 | |
| 85 | /* Only bother aligning and using wh64 if there is at least |
| 86 | one full cache line to process. This check also prevents |
| 87 | overrunning the end of the buffer with alignment words. */ |
| 88 | if (to_align64 <= n64 - CACHE_LINE_SIZE_IN_DOUBLEWORDS) |
| 89 | { |
| 90 | int lines_left; |
| 91 | |
| 92 | /* Align out64 mod the cache line size so we can use wh64. */ |
| 93 | n64 -= to_align64; |
| 94 | for (; to_align64 != 0; to_align64--) |
| 95 | { |
| 96 | *out64 = v64; |
| 97 | out64++; |
| 98 | } |
| 99 | |
| 100 | /* Use unsigned divide to turn this into a right shift. */ |
| 101 | lines_left = (unsigned) n64 / CACHE_LINE_SIZE_IN_DOUBLEWORDS; |
| 102 | |
| 103 | do |
| 104 | { |
| 105 | /* Only wh64 a few lines at a time, so we don't exceed the |
| 106 | maximum number of victim lines. */ |
| 107 | int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS ()) ? lines_left |
| 108 | : CHIP_MAX_OUTSTANDING_VICTIMS ()); |
| 109 | uint64_t *wh = out64; |
| 110 | int i = x; |
| 111 | int j; |
| 112 | |
| 113 | lines_left -= x; |
| 114 | |
| 115 | do |
| 116 | { |
| 117 | __insn_wh64 (wh); |
| 118 | wh += CACHE_LINE_SIZE_IN_DOUBLEWORDS; |
| 119 | } |
| 120 | while (--i); |
| 121 | |
| 122 | for (j = x * (CACHE_LINE_SIZE_IN_DOUBLEWORDS / 4); j != 0; j--) |
| 123 | { |
| 124 | *out64++ = v64; |
| 125 | *out64++ = v64; |
| 126 | *out64++ = v64; |
| 127 | *out64++ = v64; |
| 128 | } |
| 129 | } |
| 130 | while (lines_left != 0); |
| 131 | |
| 132 | /* We processed all full lines above, so only this many |
| 133 | words remain to be processed. */ |
| 134 | n64 &= CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1; |
| 135 | } |
| 136 | |
| 137 | /* Now handle any leftover values. */ |
| 138 | if (n64 != 0) |
| 139 | { |
| 140 | do |
| 141 | { |
| 142 | *out64 = v64; |
| 143 | out64++; |
| 144 | } |
| 145 | while (--n64 != 0); |
| 146 | } |
| 147 | |
| 148 | return s; |
| 149 | } |
| 150 | weak_alias (__memset, memset) |
| 151 | libc_hidden_builtin_def (memset) |