mirror of
				https://sourceware.org/git/glibc.git
				synced 2025-10-28 23:34:53 +03:00 
			
		
		
		
	
		
			
				
	
	
		
			152 lines
		
	
	
		
			4.1 KiB
		
	
	
	
		
			C
		
	
	
	
	
	
			
		
		
	
	
			152 lines
		
	
	
		
			4.1 KiB
		
	
	
	
		
			C
		
	
	
	
	
	
| /* Copyright (C) 2011-2017 Free Software Foundation, Inc.
 | |
|    This file is part of the GNU C Library.
 | |
|    Contributed by Chris Metcalf <cmetcalf@tilera.com>, 2011.
 | |
| 
 | |
|    The GNU C Library is free software; you can redistribute it and/or
 | |
|    modify it under the terms of the GNU Lesser General Public
 | |
|    License as published by the Free Software Foundation; either
 | |
|    version 2.1 of the License, or (at your option) any later version.
 | |
| 
 | |
|    The GNU C Library is distributed in the hope that it will be useful,
 | |
|    but WITHOUT ANY WARRANTY; without even the implied warranty of
 | |
|    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 | |
|    Lesser General Public License for more details.
 | |
| 
 | |
|    You should have received a copy of the GNU Lesser General Public
 | |
|    License along with the GNU C Library.  If not, see
 | |
|    <http://www.gnu.org/licenses/>.  */
 | |
| 
 | |
| #include <string.h>
 | |
| #include <stdint.h>
 | |
| #include <arch/chip.h>
 | |
| 
 | |
| void * inhibit_loop_to_libcall
 | |
| __memset (void *s, int c, size_t n)
 | |
| {
 | |
|   uint32_t *out32;
 | |
|   int n32;
 | |
|   uint32_t v16, v32;
 | |
|   uint8_t *out8 = s;
 | |
|   int to_align32;
 | |
| 
 | |
|   /* Experimentation shows that a trivial tight loop is a win up until
 | |
|      around a size of 20, where writing a word at a time starts to win.  */
 | |
| #define BYTE_CUTOFF 20
 | |
| 
 | |
| #if BYTE_CUTOFF < 3
 | |
|   /* This must be at least at least this big, or some code later
 | |
|      on doesn't work.  */
 | |
| # error "BYTE_CUTOFF is too small."
 | |
| #endif
 | |
| 
 | |
|   if (n < BYTE_CUTOFF)
 | |
|     {
 | |
|       /* Strangely, this turns out to be the tightest way to write
 | |
|          this loop.  */
 | |
|       if (n != 0)
 | |
|         {
 | |
|           do
 | |
|             {
 | |
|               /* Strangely, combining these into one line performs worse.  */
 | |
|               *out8 = c;
 | |
|               out8++;
 | |
|             }
 | |
|           while (--n != 0);
 | |
|         }
 | |
| 
 | |
|       return s;
 | |
|     }
 | |
| 
 | |
|   /* Align 'out8'. We know n >= 3 so this won't write past the end. */
 | |
|   while (((uintptr_t) out8 & 3) != 0)
 | |
|     {
 | |
|       *out8++ = c;
 | |
|       --n;
 | |
|     }
 | |
| 
 | |
|   /* Align 'n'. */
 | |
|   while (n & 3)
 | |
|     out8[--n] = c;
 | |
| 
 | |
|   out32 = (uint32_t *) out8;
 | |
|   n32 = n >> 2;
 | |
| 
 | |
|   /* Tile input byte out to 32 bits. */
 | |
|   v16 = __insn_intlb (c, c);
 | |
|   v32 = __insn_intlh (v16, v16);
 | |
| 
 | |
|   /* This must be at least 8 or the following loop doesn't work. */
 | |
| #define CACHE_LINE_SIZE_IN_WORDS (CHIP_L2_LINE_SIZE() / 4)
 | |
| 
 | |
|   /* Determine how many words we need to emit before the 'out32'
 | |
|      pointer becomes aligned modulo the cache line size.  */
 | |
|   to_align32 = (-((uintptr_t) out32 >> 2)) & (CACHE_LINE_SIZE_IN_WORDS - 1);
 | |
| 
 | |
|   /* Only bother aligning and using wh64 if there is at least one full
 | |
|      cache line to process.  This check also prevents overrunning the
 | |
|      end of the buffer with alignment words.  */
 | |
|   if (to_align32 <= n32 - CACHE_LINE_SIZE_IN_WORDS)
 | |
|     {
 | |
|       int lines_left;
 | |
| 
 | |
|       /* Align out32 mod the cache line size so we can use wh64. */
 | |
|       n32 -= to_align32;
 | |
|       for (; to_align32 != 0; to_align32--)
 | |
|         {
 | |
|           *out32 = v32;
 | |
|           out32++;
 | |
|         }
 | |
| 
 | |
|       /* Use unsigned divide to turn this into a right shift. */
 | |
|       lines_left = (unsigned) n32 / CACHE_LINE_SIZE_IN_WORDS;
 | |
| 
 | |
|       do
 | |
|         {
 | |
|           /* Only wh64 a few lines at a time, so we don't exceed the
 | |
|              maximum number of victim lines.  */
 | |
|           int x = ((lines_left < CHIP_MAX_OUTSTANDING_VICTIMS ())? lines_left
 | |
|                    : CHIP_MAX_OUTSTANDING_VICTIMS ());
 | |
|           uint32_t *wh = out32;
 | |
|           int i = x;
 | |
|           int j;
 | |
| 
 | |
|           lines_left -= x;
 | |
| 
 | |
|           do
 | |
|             {
 | |
|               __insn_wh64 (wh);
 | |
|               wh += CACHE_LINE_SIZE_IN_WORDS;
 | |
|             }
 | |
|           while (--i);
 | |
| 
 | |
|           for (j = x * (CACHE_LINE_SIZE_IN_WORDS / 4); j != 0; j--)
 | |
|             {
 | |
|               *out32++ = v32;
 | |
|               *out32++ = v32;
 | |
|               *out32++ = v32;
 | |
|               *out32++ = v32;
 | |
|             }
 | |
|         }
 | |
|       while (lines_left != 0);
 | |
| 
 | |
|       /* We processed all full lines above, so only this many words
 | |
|          remain to be processed.  */
 | |
|       n32 &= CACHE_LINE_SIZE_IN_WORDS - 1;
 | |
|     }
 | |
| 
 | |
|   /* Now handle any leftover values. */
 | |
|   if (n32 != 0)
 | |
|     {
 | |
|       do
 | |
|         {
 | |
|           *out32 = v32;
 | |
|           out32++;
 | |
|         }
 | |
|       while (--n32 != 0);
 | |
|     }
 | |
| 
 | |
|   return s;
 | |
| }
 | |
| weak_alias (__memset, memset)
 | |
| libc_hidden_builtin_def (memset)
 |