1/*
  2 * Copyright 2010 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 <linux/types.h>
 16#include <linux/string.h>
 17#include <linux/module.h>
 18#include <arch/chip.h>
 19
 20void *memset(void *s, int c, size_t n)
 21{
 22	uint32_t *out32;
 23	int n32;
 24	uint32_t v16, v32;
 25	uint8_t *out8 = s;
 26	int to_align32;
 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 < 3
 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 >= 3 so this won't write past the end. */
 58	while (((uintptr_t) out8 & 3) != 0) {
 59		*out8++ = c;
 60		--n;
 61	}
 62
 63	/* Align 'n'. */
 64	while (n & 3)
 65		out8[--n] = c;
 66
 67	out32 = (uint32_t *) out8;
 68	n32 = n >> 2;
 69
 70	/* Tile input byte out to 32 bits. */
 71	v16 = __insn_intlb(c, c);
 72	v32 = __insn_intlh(v16, v16);
 73
 74	/* This must be at least 8 or the following loop doesn't work. */
 75#define CACHE_LINE_SIZE_IN_WORDS (CHIP_L2_LINE_SIZE() / 4)
 76
 77	/* Determine how many words we need to emit before the 'out32'
 78	 * pointer becomes aligned modulo the cache line size.
 79	 */
 80	to_align32 =
 81		(-((uintptr_t)out32 >> 2)) & (CACHE_LINE_SIZE_IN_WORDS - 1);
 82
 83	/* Only bother aligning and using wh64 if there is at least
 84	 * one full cache line to process.  This check also prevents
 85	 * overrunning the end of the buffer with alignment words.
 86	 */
 87	if (to_align32 <= n32 - CACHE_LINE_SIZE_IN_WORDS) {
 88		int lines_left;
 89
 90		/* Align out32 mod the cache line size so we can use wh64. */
 91		n32 -= to_align32;
 92		for (; to_align32 != 0; to_align32--) {
 93			*out32 = v32;
 94			out32++;
 95		}
 96
 97		/* Use unsigned divide to turn this into a right shift. */
 98		lines_left = (unsigned)n32 / CACHE_LINE_SIZE_IN_WORDS;
 99
100		do {
101			/* Only wh64 a few lines at a time, so we don't
102			 * exceed the maximum number of victim lines.
103			 */
104			int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS())
105				  ? lines_left
106				  : CHIP_MAX_OUTSTANDING_VICTIMS());
107			uint32_t *wh = out32;
108			int i = x;
109			int j;
110
111			lines_left -= x;
112
113			do {
114				__insn_wh64(wh);
115				wh += CACHE_LINE_SIZE_IN_WORDS;
116			} while (--i);
117
118			for (j = x * (CACHE_LINE_SIZE_IN_WORDS / 4);
119			     j != 0; j--) {
120				*out32++ = v32;
121				*out32++ = v32;
122				*out32++ = v32;
123				*out32++ = v32;
124			}
125		} while (lines_left != 0);
126
127		/* We processed all full lines above, so only this many
128		 * words remain to be processed.
129		 */
130		n32 &= CACHE_LINE_SIZE_IN_WORDS - 1;
131	}
132
133	/* Now handle any leftover values. */
134	if (n32 != 0) {
135		do {
136			*out32 = v32;
137			out32++;
138		} while (--n32 != 0);
139	}
140
141	return s;
142}
143EXPORT_SYMBOL(memset);