Chris Metcalf | 18aecc2 | 2011-05-04 14:38:26 -0400 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2011 Tilera Corporation. All Rights Reserved. |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or |
| 5 | * modify it under the terms of the GNU General Public License |
| 6 | * as published by the Free Software Foundation, version 2. |
| 7 | * |
| 8 | * This program is distributed in the hope that it will be useful, but |
| 9 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 10 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or |
| 11 | * NON INFRINGEMENT. See the GNU General Public License for |
| 12 | * more details. |
| 13 | */ |
| 14 | |
Chris Metcalf | 18aecc2 | 2011-05-04 14:38:26 -0400 | [diff] [blame] | 15 | #include <linux/types.h> |
| 16 | #include <linux/string.h> |
| 17 | #include <linux/module.h> |
Chris Metcalf | c53c70a | 2013-08-01 15:52:17 -0400 | [diff] [blame] | 18 | #include <arch/chip.h> |
| 19 | #include "string-endian.h" |
Chris Metcalf | 18aecc2 | 2011-05-04 14:38:26 -0400 | [diff] [blame] | 20 | |
| 21 | void *memset(void *s, int c, size_t n) |
| 22 | { |
| 23 | uint64_t *out64; |
| 24 | int n64, to_align64; |
| 25 | uint64_t v64; |
| 26 | uint8_t *out8 = s; |
| 27 | |
| 28 | /* Experimentation shows that a trivial tight loop is a win up until |
| 29 | * around a size of 20, where writing a word at a time starts to win. |
| 30 | */ |
| 31 | #define BYTE_CUTOFF 20 |
| 32 | |
| 33 | #if BYTE_CUTOFF < 7 |
| 34 | /* This must be at least at least this big, or some code later |
| 35 | * on doesn't work. |
| 36 | */ |
| 37 | #error "BYTE_CUTOFF is too small" |
| 38 | #endif |
| 39 | |
| 40 | if (n < BYTE_CUTOFF) { |
| 41 | /* Strangely, this turns out to be the tightest way to |
| 42 | * write this loop. |
| 43 | */ |
| 44 | if (n != 0) { |
| 45 | do { |
| 46 | /* Strangely, combining these into one line |
| 47 | * performs worse. |
| 48 | */ |
| 49 | *out8 = c; |
| 50 | out8++; |
| 51 | } while (--n != 0); |
| 52 | } |
| 53 | |
| 54 | return s; |
| 55 | } |
| 56 | |
| 57 | /* Align 'out8'. We know n >= 7 so this won't write past the end. */ |
| 58 | while (((uintptr_t) out8 & 7) != 0) { |
| 59 | *out8++ = c; |
| 60 | --n; |
| 61 | } |
| 62 | |
| 63 | /* Align 'n'. */ |
| 64 | while (n & 7) |
| 65 | out8[--n] = c; |
| 66 | |
| 67 | out64 = (uint64_t *) out8; |
| 68 | n64 = n >> 3; |
| 69 | |
| 70 | /* Tile input byte out to 64 bits. */ |
Chris Metcalf | c53c70a | 2013-08-01 15:52:17 -0400 | [diff] [blame] | 71 | v64 = copy_byte(c); |
Chris Metcalf | 18aecc2 | 2011-05-04 14:38:26 -0400 | [diff] [blame] | 72 | |
| 73 | /* This must be at least 8 or the following loop doesn't work. */ |
| 74 | #define CACHE_LINE_SIZE_IN_DOUBLEWORDS (CHIP_L2_LINE_SIZE() / 8) |
| 75 | |
| 76 | /* Determine how many words we need to emit before the 'out32' |
| 77 | * pointer becomes aligned modulo the cache line size. |
| 78 | */ |
| 79 | to_align64 = (-((uintptr_t)out64 >> 3)) & |
| 80 | (CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1); |
| 81 | |
| 82 | /* Only bother aligning and using wh64 if there is at least |
| 83 | * one full cache line to process. This check also prevents |
| 84 | * overrunning the end of the buffer with alignment words. |
| 85 | */ |
| 86 | if (to_align64 <= n64 - CACHE_LINE_SIZE_IN_DOUBLEWORDS) { |
| 87 | int lines_left; |
| 88 | |
| 89 | /* Align out64 mod the cache line size so we can use wh64. */ |
| 90 | n64 -= to_align64; |
| 91 | for (; to_align64 != 0; to_align64--) { |
| 92 | *out64 = v64; |
| 93 | out64++; |
| 94 | } |
| 95 | |
| 96 | /* Use unsigned divide to turn this into a right shift. */ |
| 97 | lines_left = (unsigned)n64 / CACHE_LINE_SIZE_IN_DOUBLEWORDS; |
| 98 | |
| 99 | do { |
| 100 | /* Only wh64 a few lines at a time, so we don't |
| 101 | * exceed the maximum number of victim lines. |
| 102 | */ |
| 103 | int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS()) |
| 104 | ? lines_left |
| 105 | : CHIP_MAX_OUTSTANDING_VICTIMS()); |
| 106 | uint64_t *wh = out64; |
| 107 | int i = x; |
| 108 | int j; |
| 109 | |
| 110 | lines_left -= x; |
| 111 | |
| 112 | do { |
| 113 | __insn_wh64(wh); |
| 114 | wh += CACHE_LINE_SIZE_IN_DOUBLEWORDS; |
| 115 | } while (--i); |
| 116 | |
| 117 | for (j = x * (CACHE_LINE_SIZE_IN_DOUBLEWORDS / 4); |
| 118 | j != 0; j--) { |
| 119 | *out64++ = v64; |
| 120 | *out64++ = v64; |
| 121 | *out64++ = v64; |
| 122 | *out64++ = v64; |
| 123 | } |
| 124 | } while (lines_left != 0); |
| 125 | |
| 126 | /* We processed all full lines above, so only this many |
| 127 | * words remain to be processed. |
| 128 | */ |
| 129 | n64 &= CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1; |
| 130 | } |
| 131 | |
| 132 | /* Now handle any leftover values. */ |
| 133 | if (n64 != 0) { |
| 134 | do { |
| 135 | *out64 = v64; |
| 136 | out64++; |
| 137 | } while (--n64 != 0); |
| 138 | } |
| 139 | |
| 140 | return s; |
| 141 | } |
| 142 | EXPORT_SYMBOL(memset); |