1*4882a593Smuzhiyun/* 2*4882a593Smuzhiyun * arch/xtensa/lib/strnlen_user.S 3*4882a593Smuzhiyun * 4*4882a593Smuzhiyun * This file is subject to the terms and conditions of the GNU General 5*4882a593Smuzhiyun * Public License. See the file "COPYING" in the main directory of 6*4882a593Smuzhiyun * this archive for more details. 7*4882a593Smuzhiyun * 8*4882a593Smuzhiyun * Returns strnlen, including trailing zero terminator. 9*4882a593Smuzhiyun * Zero indicates error. 10*4882a593Smuzhiyun * 11*4882a593Smuzhiyun * Copyright (C) 2002 Tensilica Inc. 12*4882a593Smuzhiyun */ 13*4882a593Smuzhiyun 14*4882a593Smuzhiyun#include <linux/linkage.h> 15*4882a593Smuzhiyun#include <asm/asmmacro.h> 16*4882a593Smuzhiyun#include <asm/core.h> 17*4882a593Smuzhiyun 18*4882a593Smuzhiyun/* 19*4882a593Smuzhiyun * size_t __strnlen_user(const char *s, size_t len) 20*4882a593Smuzhiyun */ 21*4882a593Smuzhiyun 22*4882a593Smuzhiyun#ifdef __XTENSA_EB__ 23*4882a593Smuzhiyun# define MASK0 0xff000000 24*4882a593Smuzhiyun# define MASK1 0x00ff0000 25*4882a593Smuzhiyun# define MASK2 0x0000ff00 26*4882a593Smuzhiyun# define MASK3 0x000000ff 27*4882a593Smuzhiyun#else 28*4882a593Smuzhiyun# define MASK0 0x000000ff 29*4882a593Smuzhiyun# define MASK1 0x0000ff00 30*4882a593Smuzhiyun# define MASK2 0x00ff0000 31*4882a593Smuzhiyun# define MASK3 0xff000000 32*4882a593Smuzhiyun#endif 33*4882a593Smuzhiyun 34*4882a593Smuzhiyun# Register use: 35*4882a593Smuzhiyun# a2/ src 36*4882a593Smuzhiyun# a3/ len 37*4882a593Smuzhiyun# a4/ tmp 38*4882a593Smuzhiyun# a5/ mask0 39*4882a593Smuzhiyun# a6/ mask1 40*4882a593Smuzhiyun# a7/ mask2 41*4882a593Smuzhiyun# a8/ mask3 42*4882a593Smuzhiyun# a9/ tmp 43*4882a593Smuzhiyun# a10/ tmp 44*4882a593Smuzhiyun 45*4882a593Smuzhiyun.text 46*4882a593SmuzhiyunENTRY(__strnlen_user) 47*4882a593Smuzhiyun 48*4882a593Smuzhiyun abi_entry_default 49*4882a593Smuzhiyun # a2/ s, a3/ len 50*4882a593Smuzhiyun addi a4, a2, -4 # because we overincrement at the end; 51*4882a593Smuzhiyun # we compensate with load offsets of 4 52*4882a593Smuzhiyun movi a5, MASK0 # mask for byte 0 53*4882a593Smuzhiyun movi a6, MASK1 # mask for byte 1 54*4882a593Smuzhiyun movi a7, MASK2 # mask for byte 2 55*4882a593Smuzhiyun movi a8, MASK3 # mask for byte 3 56*4882a593Smuzhiyun bbsi.l a2, 0, .L1mod2 # if only 8-bit aligned 57*4882a593Smuzhiyun bbsi.l a2, 1, .L2mod4 # if only 16-bit aligned 58*4882a593Smuzhiyun 59*4882a593Smuzhiyun/* 60*4882a593Smuzhiyun * String is word-aligned. 61*4882a593Smuzhiyun */ 62*4882a593Smuzhiyun.Laligned: 63*4882a593Smuzhiyun srli a10, a3, 2 # number of loop iterations with 4B per loop 64*4882a593Smuzhiyun#if XCHAL_HAVE_LOOPS 65*4882a593Smuzhiyun loopnez a10, .Ldone 66*4882a593Smuzhiyun#else 67*4882a593Smuzhiyun beqz a10, .Ldone 68*4882a593Smuzhiyun slli a10, a10, 2 69*4882a593Smuzhiyun add a10, a10, a4 # a10 = end of last 4B chunk 70*4882a593Smuzhiyun#endif /* XCHAL_HAVE_LOOPS */ 71*4882a593Smuzhiyun.Loop: 72*4882a593SmuzhiyunEX(10f) l32i a9, a4, 4 # get next word of string 73*4882a593Smuzhiyun addi a4, a4, 4 # advance string pointer 74*4882a593Smuzhiyun bnone a9, a5, .Lz0 # if byte 0 is zero 75*4882a593Smuzhiyun bnone a9, a6, .Lz1 # if byte 1 is zero 76*4882a593Smuzhiyun bnone a9, a7, .Lz2 # if byte 2 is zero 77*4882a593Smuzhiyun bnone a9, a8, .Lz3 # if byte 3 is zero 78*4882a593Smuzhiyun#if !XCHAL_HAVE_LOOPS 79*4882a593Smuzhiyun blt a4, a10, .Loop 80*4882a593Smuzhiyun#endif 81*4882a593Smuzhiyun 82*4882a593Smuzhiyun.Ldone: 83*4882a593SmuzhiyunEX(10f) l32i a9, a4, 4 # load 4 bytes for remaining checks 84*4882a593Smuzhiyun 85*4882a593Smuzhiyun bbci.l a3, 1, .L100 86*4882a593Smuzhiyun # check two more bytes (bytes 0, 1 of word) 87*4882a593Smuzhiyun addi a4, a4, 2 # advance string pointer 88*4882a593Smuzhiyun bnone a9, a5, .Lz0 # if byte 0 is zero 89*4882a593Smuzhiyun bnone a9, a6, .Lz1 # if byte 1 is zero 90*4882a593Smuzhiyun.L100: 91*4882a593Smuzhiyun bbci.l a3, 0, .L101 92*4882a593Smuzhiyun # check one more byte (byte 2 of word) 93*4882a593Smuzhiyun # Actually, we don't need to check. Zero or nonzero, we'll add one. 94*4882a593Smuzhiyun # Do not add an extra one for the NULL terminator since we have 95*4882a593Smuzhiyun # exhausted the original len parameter. 96*4882a593Smuzhiyun addi a4, a4, 1 # advance string pointer 97*4882a593Smuzhiyun.L101: 98*4882a593Smuzhiyun sub a2, a4, a2 # compute length 99*4882a593Smuzhiyun abi_ret_default 100*4882a593Smuzhiyun 101*4882a593Smuzhiyun# NOTE that in several places below, we point to the byte just after 102*4882a593Smuzhiyun# the zero byte in order to include the NULL terminator in the count. 103*4882a593Smuzhiyun 104*4882a593Smuzhiyun.Lz3: # byte 3 is zero 105*4882a593Smuzhiyun addi a4, a4, 3 # point to zero byte 106*4882a593Smuzhiyun.Lz0: # byte 0 is zero 107*4882a593Smuzhiyun addi a4, a4, 1 # point just beyond zero byte 108*4882a593Smuzhiyun sub a2, a4, a2 # subtract to get length 109*4882a593Smuzhiyun abi_ret_default 110*4882a593Smuzhiyun.Lz1: # byte 1 is zero 111*4882a593Smuzhiyun addi a4, a4, 1+1 # point just beyond zero byte 112*4882a593Smuzhiyun sub a2, a4, a2 # subtract to get length 113*4882a593Smuzhiyun abi_ret_default 114*4882a593Smuzhiyun.Lz2: # byte 2 is zero 115*4882a593Smuzhiyun addi a4, a4, 2+1 # point just beyond zero byte 116*4882a593Smuzhiyun sub a2, a4, a2 # subtract to get length 117*4882a593Smuzhiyun abi_ret_default 118*4882a593Smuzhiyun 119*4882a593Smuzhiyun.L1mod2: # address is odd 120*4882a593SmuzhiyunEX(10f) l8ui a9, a4, 4 # get byte 0 121*4882a593Smuzhiyun addi a4, a4, 1 # advance string pointer 122*4882a593Smuzhiyun beqz a9, .Lz3 # if byte 0 is zero 123*4882a593Smuzhiyun bbci.l a4, 1, .Laligned # if string pointer is now word-aligned 124*4882a593Smuzhiyun 125*4882a593Smuzhiyun.L2mod4: # address is 2 mod 4 126*4882a593Smuzhiyun addi a4, a4, 2 # advance ptr for aligned access 127*4882a593SmuzhiyunEX(10f) l32i a9, a4, 0 # get word with first two bytes of string 128*4882a593Smuzhiyun bnone a9, a7, .Lz2 # if byte 2 (of word, not string) is zero 129*4882a593Smuzhiyun bany a9, a8, .Laligned # if byte 3 (of word, not string) is nonzero 130*4882a593Smuzhiyun # byte 3 is zero 131*4882a593Smuzhiyun addi a4, a4, 3+1 # point just beyond zero byte 132*4882a593Smuzhiyun sub a2, a4, a2 # subtract to get length 133*4882a593Smuzhiyun abi_ret_default 134*4882a593Smuzhiyun 135*4882a593SmuzhiyunENDPROC(__strnlen_user) 136*4882a593Smuzhiyun 137*4882a593Smuzhiyun .section .fixup, "ax" 138*4882a593Smuzhiyun .align 4 139*4882a593Smuzhiyun10: 140*4882a593Smuzhiyun movi a2, 0 141*4882a593Smuzhiyun abi_ret_default 142