1/* 2 * arch/xtensa/lib/strnlen_user.S 3 * 4 * This file is subject to the terms and conditions of the GNU General 5 * Public License. See the file "COPYING" in the main directory of 6 * this archive for more details. 7 * 8 * Returns strnlen, including trailing zero terminator. 9 * Zero indicates error. 10 * 11 * Copyright (C) 2002 Tensilica Inc. 12 */ 13 14#include <xtensa/coreasm.h> 15 16/* Load or store instructions that may cause exceptions use the EX macro. */ 17 18#define EX(insn,reg1,reg2,offset,handler) \ 199: insn reg1, reg2, offset; \ 20 .section __ex_table, "a"; \ 21 .word 9b, handler; \ 22 .previous 23 24/* 25 * size_t __strnlen_user(const char *s, size_t len) 26 */ 27.text 28.begin literal 29.align 4 30.Lmask0: 31 .byte 0xff, 0x00, 0x00, 0x00 32.Lmask1: 33 .byte 0x00, 0xff, 0x00, 0x00 34.Lmask2: 35 .byte 0x00, 0x00, 0xff, 0x00 36.Lmask3: 37 .byte 0x00, 0x00, 0x00, 0xff 38.end literal 39 40# Register use: 41# a2/ src 42# a3/ len 43# a4/ tmp 44# a5/ mask0 45# a6/ mask1 46# a7/ mask2 47# a8/ mask3 48# a9/ tmp 49# a10/ tmp 50 51.align 4 52.global __strnlen_user 53.type __strnlen_user,@function 54__strnlen_user: 55 entry sp, 16 # minimal stack frame 56 # a2/ s, a3/ len 57 addi a4, a2, -4 # because we overincrement at the end; 58 # we compensate with load offsets of 4 59 l32r a5, .Lmask0 # mask for byte 0 60 l32r a6, .Lmask1 # mask for byte 1 61 l32r a7, .Lmask2 # mask for byte 2 62 l32r a8, .Lmask3 # mask for byte 3 63 bbsi.l a2, 0, .L1mod2 # if only 8-bit aligned 64 bbsi.l a2, 1, .L2mod4 # if only 16-bit aligned 65 66/* 67 * String is word-aligned. 68 */ 69.Laligned: 70 srli a10, a3, 2 # number of loop iterations with 4B per loop 71#if XCHAL_HAVE_LOOPS 72 loopnez a10, .Ldone 73#else 74 beqz a10, .Ldone 75 slli a10, a10, 2 76 add a10, a10, a4 # a10 = end of last 4B chunk 77#endif /* XCHAL_HAVE_LOOPS */ 78.Loop: 79 EX(l32i, a9, a4, 4, lenfixup) # get next word of string 80 addi a4, a4, 4 # advance string pointer 81 bnone a9, a5, .Lz0 # if byte 0 is zero 82 bnone a9, a6, .Lz1 # if byte 1 is zero 83 bnone a9, a7, .Lz2 # if byte 2 is zero 84 bnone a9, a8, .Lz3 # if byte 3 is zero 85#if !XCHAL_HAVE_LOOPS 86 blt a4, a10, .Loop 87#endif 88 89.Ldone: 90 EX(l32i, a9, a4, 4, lenfixup) # load 4 bytes for remaining checks 91 92 bbci.l a3, 1, .L100 93 # check two more bytes (bytes 0, 1 of word) 94 addi a4, a4, 2 # advance string pointer 95 bnone a9, a5, .Lz0 # if byte 0 is zero 96 bnone a9, a6, .Lz1 # if byte 1 is zero 97.L100: 98 bbci.l a3, 0, .L101 99 # check one more byte (byte 2 of word) 100 # Actually, we don't need to check. Zero or nonzero, we'll add one. 101 # Do not add an extra one for the NULL terminator since we have 102 # exhausted the original len parameter. 103 addi a4, a4, 1 # advance string pointer 104.L101: 105 sub a2, a4, a2 # compute length 106 retw 107 108# NOTE that in several places below, we point to the byte just after 109# the zero byte in order to include the NULL terminator in the count. 110 111.Lz3: # byte 3 is zero 112 addi a4, a4, 3 # point to zero byte 113.Lz0: # byte 0 is zero 114 addi a4, a4, 1 # point just beyond zero byte 115 sub a2, a4, a2 # subtract to get length 116 retw 117.Lz1: # byte 1 is zero 118 addi a4, a4, 1+1 # point just beyond zero byte 119 sub a2, a4, a2 # subtract to get length 120 retw 121.Lz2: # byte 2 is zero 122 addi a4, a4, 2+1 # point just beyond zero byte 123 sub a2, a4, a2 # subtract to get length 124 retw 125 126.L1mod2: # address is odd 127 EX(l8ui, a9, a4, 4, lenfixup) # get byte 0 128 addi a4, a4, 1 # advance string pointer 129 beqz a9, .Lz3 # if byte 0 is zero 130 bbci.l a4, 1, .Laligned # if string pointer is now word-aligned 131 132.L2mod4: # address is 2 mod 4 133 addi a4, a4, 2 # advance ptr for aligned access 134 EX(l32i, a9, a4, 0, lenfixup) # get word with first two bytes of string 135 bnone a9, a7, .Lz2 # if byte 2 (of word, not string) is zero 136 bany a9, a8, .Laligned # if byte 3 (of word, not string) is nonzero 137 # byte 3 is zero 138 addi a4, a4, 3+1 # point just beyond zero byte 139 sub a2, a4, a2 # subtract to get length 140 retw 141 142 .section .fixup, "ax" 143 .align 4 144lenfixup: 145 movi a2, 0 146 retw 147 148