| 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 |  | 
 | 15 | #include <arch/chip.h> | 
 | 16 |  | 
 | 17 | #include <linux/types.h> | 
 | 18 | #include <linux/string.h> | 
 | 19 | #include <linux/module.h> | 
 | 20 |  | 
 | 21 | #undef memset | 
 | 22 |  | 
 | 23 | void *memset(void *s, int c, size_t n) | 
 | 24 | { | 
 | 25 | 	uint64_t *out64; | 
 | 26 | 	int n64, to_align64; | 
 | 27 | 	uint64_t v64; | 
 | 28 | 	uint8_t *out8 = s; | 
 | 29 |  | 
 | 30 | 	/* Experimentation shows that a trivial tight loop is a win up until | 
 | 31 | 	 * around a size of 20, where writing a word at a time starts to win. | 
 | 32 | 	 */ | 
 | 33 | #define BYTE_CUTOFF 20 | 
 | 34 |  | 
 | 35 | #if BYTE_CUTOFF < 7 | 
 | 36 | 	/* This must be at least at least this big, or some code later | 
 | 37 | 	 * on doesn't work. | 
 | 38 | 	 */ | 
 | 39 | #error "BYTE_CUTOFF is too small" | 
 | 40 | #endif | 
 | 41 |  | 
 | 42 | 	if (n < BYTE_CUTOFF) { | 
 | 43 | 		/* Strangely, this turns out to be the tightest way to | 
 | 44 | 		 * write this loop. | 
 | 45 | 		 */ | 
 | 46 | 		if (n != 0) { | 
 | 47 | 			do { | 
 | 48 | 				/* Strangely, combining these into one line | 
 | 49 | 				 * performs worse. | 
 | 50 | 				 */ | 
 | 51 | 				*out8 = c; | 
 | 52 | 				out8++; | 
 | 53 | 			} while (--n != 0); | 
 | 54 | 		} | 
 | 55 |  | 
 | 56 | 		return s; | 
 | 57 | 	} | 
 | 58 |  | 
 | 59 | 	/* Align 'out8'. We know n >= 7 so this won't write past the end. */ | 
 | 60 | 	while (((uintptr_t) out8 & 7) != 0) { | 
 | 61 | 		*out8++ = c; | 
 | 62 | 		--n; | 
 | 63 | 	} | 
 | 64 |  | 
 | 65 | 	/* Align 'n'. */ | 
 | 66 | 	while (n & 7) | 
 | 67 | 		out8[--n] = c; | 
 | 68 |  | 
 | 69 | 	out64 = (uint64_t *) out8; | 
 | 70 | 	n64 = n >> 3; | 
 | 71 |  | 
 | 72 | 	/* Tile input byte out to 64 bits. */ | 
 | 73 | 	/* KLUDGE */ | 
 | 74 | 	v64 = 0x0101010101010101ULL * (uint8_t)c; | 
 | 75 |  | 
 | 76 | 	/* This must be at least 8 or the following loop doesn't work. */ | 
 | 77 | #define CACHE_LINE_SIZE_IN_DOUBLEWORDS (CHIP_L2_LINE_SIZE() / 8) | 
 | 78 |  | 
 | 79 | 	/* Determine how many words we need to emit before the 'out32' | 
 | 80 | 	 * pointer becomes aligned modulo the cache line size. | 
 | 81 | 	 */ | 
 | 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 | 	 */ | 
 | 89 | 	if (to_align64 <= n64 - CACHE_LINE_SIZE_IN_DOUBLEWORDS) { | 
 | 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 | 			*out64 = v64; | 
 | 96 | 			out64++; | 
 | 97 | 		} | 
 | 98 |  | 
 | 99 | 		/* Use unsigned divide to turn this into a right shift. */ | 
 | 100 | 		lines_left = (unsigned)n64 / CACHE_LINE_SIZE_IN_DOUBLEWORDS; | 
 | 101 |  | 
 | 102 | 		do { | 
 | 103 | 			/* Only wh64 a few lines at a time, so we don't | 
 | 104 | 			 * exceed the maximum number of victim lines. | 
 | 105 | 			 */ | 
 | 106 | 			int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS()) | 
 | 107 | 				  ? 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 | 				__insn_wh64(wh); | 
 | 117 | 				wh += CACHE_LINE_SIZE_IN_DOUBLEWORDS; | 
 | 118 | 			} while (--i); | 
 | 119 |  | 
 | 120 | 			for (j = x * (CACHE_LINE_SIZE_IN_DOUBLEWORDS / 4); | 
 | 121 | 			     j != 0; j--) { | 
 | 122 | 				*out64++ = v64; | 
 | 123 | 				*out64++ = v64; | 
 | 124 | 				*out64++ = v64; | 
 | 125 | 				*out64++ = v64; | 
 | 126 | 			} | 
 | 127 | 		} while (lines_left != 0); | 
 | 128 |  | 
 | 129 | 		/* We processed all full lines above, so only this many | 
 | 130 | 		 * words remain to be processed. | 
 | 131 | 		 */ | 
 | 132 | 		n64 &= CACHE_LINE_SIZE_IN_DOUBLEWORDS - 1; | 
 | 133 | 	} | 
 | 134 |  | 
 | 135 | 	/* Now handle any leftover values. */ | 
 | 136 | 	if (n64 != 0) { | 
 | 137 | 		do { | 
 | 138 | 			*out64 = v64; | 
 | 139 | 			out64++; | 
 | 140 | 		} while (--n64 != 0); | 
 | 141 | 	} | 
 | 142 |  | 
 | 143 | 	return s; | 
 | 144 | } | 
 | 145 | EXPORT_SYMBOL(memset); |