1*412f47f9SXin Li/* 2*412f47f9SXin Li * strnlen - calculate the length of a string with limit. 3*412f47f9SXin Li * 4*412f47f9SXin Li * Copyright (c) 2020-2022, Arm Limited. 5*412f47f9SXin Li * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception 6*412f47f9SXin Li */ 7*412f47f9SXin Li 8*412f47f9SXin Li/* Assumptions: 9*412f47f9SXin Li * 10*412f47f9SXin Li * ARMv8-a, AArch64, Advanced SIMD. 11*412f47f9SXin Li * MTE compatible. 12*412f47f9SXin Li */ 13*412f47f9SXin Li 14*412f47f9SXin Li#include "asmdefs.h" 15*412f47f9SXin Li 16*412f47f9SXin Li#define srcin x0 17*412f47f9SXin Li#define cntin x1 18*412f47f9SXin Li#define result x0 19*412f47f9SXin Li 20*412f47f9SXin Li#define src x2 21*412f47f9SXin Li#define synd x3 22*412f47f9SXin Li#define shift x4 23*412f47f9SXin Li#define tmp x4 24*412f47f9SXin Li#define cntrem x5 25*412f47f9SXin Li 26*412f47f9SXin Li#define qdata q0 27*412f47f9SXin Li#define vdata v0 28*412f47f9SXin Li#define vhas_chr v1 29*412f47f9SXin Li#define vend v2 30*412f47f9SXin Li#define dend d2 31*412f47f9SXin Li 32*412f47f9SXin Li/* 33*412f47f9SXin Li Core algorithm: 34*412f47f9SXin Li Process the string in 16-byte aligned chunks. Compute a 64-bit mask with 35*412f47f9SXin Li four bits per byte using the shrn instruction. A count trailing zeros then 36*412f47f9SXin Li identifies the first zero byte. */ 37*412f47f9SXin Li 38*412f47f9SXin LiENTRY (__strnlen_aarch64) 39*412f47f9SXin Li PTR_ARG (0) 40*412f47f9SXin Li SIZE_ARG (1) 41*412f47f9SXin Li bic src, srcin, 15 42*412f47f9SXin Li cbz cntin, L(nomatch) 43*412f47f9SXin Li ld1 {vdata.16b}, [src] 44*412f47f9SXin Li cmeq vhas_chr.16b, vdata.16b, 0 45*412f47f9SXin Li lsl shift, srcin, 2 46*412f47f9SXin Li shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */ 47*412f47f9SXin Li fmov synd, dend 48*412f47f9SXin Li lsr synd, synd, shift 49*412f47f9SXin Li cbz synd, L(start_loop) 50*412f47f9SXin LiL(finish): 51*412f47f9SXin Li rbit synd, synd 52*412f47f9SXin Li clz synd, synd 53*412f47f9SXin Li lsr result, synd, 2 54*412f47f9SXin Li cmp cntin, result 55*412f47f9SXin Li csel result, cntin, result, ls 56*412f47f9SXin Li ret 57*412f47f9SXin Li 58*412f47f9SXin LiL(nomatch): 59*412f47f9SXin Li mov result, cntin 60*412f47f9SXin Li ret 61*412f47f9SXin Li 62*412f47f9SXin LiL(start_loop): 63*412f47f9SXin Li sub tmp, src, srcin 64*412f47f9SXin Li add tmp, tmp, 17 65*412f47f9SXin Li subs cntrem, cntin, tmp 66*412f47f9SXin Li b.lo L(nomatch) 67*412f47f9SXin Li 68*412f47f9SXin Li /* Make sure that it won't overread by a 16-byte chunk */ 69*412f47f9SXin Li tbz cntrem, 4, L(loop32_2) 70*412f47f9SXin Li sub src, src, 16 71*412f47f9SXin Li .p2align 5 72*412f47f9SXin LiL(loop32): 73*412f47f9SXin Li ldr qdata, [src, 32]! 74*412f47f9SXin Li cmeq vhas_chr.16b, vdata.16b, 0 75*412f47f9SXin Li umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ 76*412f47f9SXin Li fmov synd, dend 77*412f47f9SXin Li cbnz synd, L(end) 78*412f47f9SXin LiL(loop32_2): 79*412f47f9SXin Li ldr qdata, [src, 16] 80*412f47f9SXin Li subs cntrem, cntrem, 32 81*412f47f9SXin Li cmeq vhas_chr.16b, vdata.16b, 0 82*412f47f9SXin Li b.lo L(end_2) 83*412f47f9SXin Li umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ 84*412f47f9SXin Li fmov synd, dend 85*412f47f9SXin Li cbz synd, L(loop32) 86*412f47f9SXin LiL(end_2): 87*412f47f9SXin Li add src, src, 16 88*412f47f9SXin LiL(end): 89*412f47f9SXin Li shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */ 90*412f47f9SXin Li sub result, src, srcin 91*412f47f9SXin Li fmov synd, dend 92*412f47f9SXin Li#ifndef __AARCH64EB__ 93*412f47f9SXin Li rbit synd, synd 94*412f47f9SXin Li#endif 95*412f47f9SXin Li clz synd, synd 96*412f47f9SXin Li add result, result, synd, lsr 2 97*412f47f9SXin Li cmp cntin, result 98*412f47f9SXin Li csel result, cntin, result, ls 99*412f47f9SXin Li ret 100*412f47f9SXin Li 101*412f47f9SXin LiEND (__strnlen_aarch64) 102*412f47f9SXin Li 103