Gabe Black | 0e49905 | 2011-11-14 14:47:18 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (C) 1991,1992,1993,1997,1998,2003, 2005 Free Software Foundation, Inc. |
| 3 | * This file is part of the GNU C Library. |
| 4 | * Copyright (c) 2011 The Chromium OS Authors. |
| 5 | * |
Wolfgang Denk | d79de1d | 2013-07-08 09:37:19 +0200 | [diff] [blame] | 6 | * SPDX-License-Identifier: GPL-2.0+ |
Gabe Black | 0e49905 | 2011-11-14 14:47:18 +0000 | [diff] [blame] | 7 | */ |
| 8 | |
| 9 | /* From glibc-2.14, sysdeps/i386/memset.c */ |
| 10 | |
Gabe Black | 0e49905 | 2011-11-14 14:47:18 +0000 | [diff] [blame] | 11 | #include <linux/types.h> |
Masahiro Yamada | e722852 | 2014-11-26 16:00:58 +0900 | [diff] [blame] | 12 | #include <linux/compiler.h> |
| 13 | #include <asm/string.h> |
Gabe Black | 0e49905 | 2011-11-14 14:47:18 +0000 | [diff] [blame] | 14 | |
| 15 | typedef uint32_t op_t; |
| 16 | |
| 17 | void *memset(void *dstpp, int c, size_t len) |
| 18 | { |
| 19 | int d0; |
| 20 | unsigned long int dstp = (unsigned long int) dstpp; |
| 21 | |
| 22 | /* This explicit register allocation improves code very much indeed. */ |
| 23 | register op_t x asm("ax"); |
| 24 | |
| 25 | x = (unsigned char) c; |
| 26 | |
| 27 | /* Clear the direction flag, so filling will move forward. */ |
| 28 | asm volatile("cld"); |
| 29 | |
| 30 | /* This threshold value is optimal. */ |
| 31 | if (len >= 12) { |
| 32 | /* Fill X with four copies of the char we want to fill with. */ |
| 33 | x |= (x << 8); |
| 34 | x |= (x << 16); |
| 35 | |
| 36 | /* Adjust LEN for the bytes handled in the first loop. */ |
| 37 | len -= (-dstp) % sizeof(op_t); |
| 38 | |
| 39 | /* |
| 40 | * There are at least some bytes to set. No need to test for |
| 41 | * LEN == 0 in this alignment loop. |
| 42 | */ |
| 43 | |
| 44 | /* Fill bytes until DSTP is aligned on a longword boundary. */ |
| 45 | asm volatile( |
| 46 | "rep\n" |
| 47 | "stosb" /* %0, %2, %3 */ : |
| 48 | "=D" (dstp), "=c" (d0) : |
| 49 | "0" (dstp), "1" ((-dstp) % sizeof(op_t)), "a" (x) : |
| 50 | "memory"); |
| 51 | |
| 52 | /* Fill longwords. */ |
| 53 | asm volatile( |
| 54 | "rep\n" |
| 55 | "stosl" /* %0, %2, %3 */ : |
| 56 | "=D" (dstp), "=c" (d0) : |
| 57 | "0" (dstp), "1" (len / sizeof(op_t)), "a" (x) : |
| 58 | "memory"); |
| 59 | len %= sizeof(op_t); |
| 60 | } |
| 61 | |
| 62 | /* Write the last few bytes. */ |
| 63 | asm volatile( |
| 64 | "rep\n" |
| 65 | "stosb" /* %0, %2, %3 */ : |
| 66 | "=D" (dstp), "=c" (d0) : |
| 67 | "0" (dstp), "1" (len), "a" (x) : |
| 68 | "memory"); |
| 69 | |
| 70 | return dstpp; |
| 71 | } |
Graeme Russ | 1dd0b70 | 2011-12-27 22:46:39 +1100 | [diff] [blame] | 72 | |
| 73 | #define OP_T_THRES 8 |
| 74 | #define OPSIZ (sizeof(op_t)) |
| 75 | |
| 76 | #define BYTE_COPY_FWD(dst_bp, src_bp, nbytes) \ |
| 77 | do { \ |
| 78 | int __d0; \ |
| 79 | asm volatile( \ |
| 80 | /* Clear the direction flag, so copying goes forward. */ \ |
| 81 | "cld\n" \ |
| 82 | /* Copy bytes. */ \ |
| 83 | "rep\n" \ |
| 84 | "movsb" : \ |
| 85 | "=D" (dst_bp), "=S" (src_bp), "=c" (__d0) : \ |
| 86 | "0" (dst_bp), "1" (src_bp), "2" (nbytes) : \ |
| 87 | "memory"); \ |
| 88 | } while (0) |
| 89 | |
| 90 | #define WORD_COPY_FWD(dst_bp, src_bp, nbytes_left, nbytes) \ |
| 91 | do { \ |
| 92 | int __d0; \ |
| 93 | asm volatile( \ |
| 94 | /* Clear the direction flag, so copying goes forward. */ \ |
| 95 | "cld\n" \ |
| 96 | /* Copy longwords. */ \ |
| 97 | "rep\n" \ |
| 98 | "movsl" : \ |
| 99 | "=D" (dst_bp), "=S" (src_bp), "=c" (__d0) : \ |
| 100 | "0" (dst_bp), "1" (src_bp), "2" ((nbytes) / 4) : \ |
| 101 | "memory"); \ |
| 102 | (nbytes_left) = (nbytes) % 4; \ |
| 103 | } while (0) |
| 104 | |
| 105 | void *memcpy(void *dstpp, const void *srcpp, size_t len) |
| 106 | { |
| 107 | unsigned long int dstp = (long int)dstpp; |
| 108 | unsigned long int srcp = (long int)srcpp; |
| 109 | |
| 110 | /* Copy from the beginning to the end. */ |
| 111 | |
| 112 | /* If there not too few bytes to copy, use word copy. */ |
| 113 | if (len >= OP_T_THRES) { |
| 114 | /* Copy just a few bytes to make DSTP aligned. */ |
| 115 | len -= (-dstp) % OPSIZ; |
| 116 | BYTE_COPY_FWD(dstp, srcp, (-dstp) % OPSIZ); |
| 117 | |
| 118 | /* Copy from SRCP to DSTP taking advantage of the known |
| 119 | * alignment of DSTP. Number of bytes remaining is put |
| 120 | * in the third argument, i.e. in LEN. This number may |
| 121 | * vary from machine to machine. |
| 122 | */ |
| 123 | WORD_COPY_FWD(dstp, srcp, len, len); |
| 124 | |
| 125 | /* Fall out and copy the tail. */ |
| 126 | } |
| 127 | |
| 128 | /* There are just a few bytes to copy. Use byte memory operations. */ |
| 129 | BYTE_COPY_FWD(dstp, srcp, len); |
| 130 | |
| 131 | return dstpp; |
| 132 | } |