1*412f47f9SXin Li/* 2*412f47f9SXin Li * memchr - find a character in a memory zone 3*412f47f9SXin Li * 4*412f47f9SXin Li * Copyright (c) 2018-2022, 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#include "asmdefs.h" 9*412f47f9SXin Li 10*412f47f9SXin Li#if __ARM_FEATURE_SVE 11*412f47f9SXin Li/* Assumptions: 12*412f47f9SXin Li * 13*412f47f9SXin Li * ARMv8-a, AArch64 14*412f47f9SXin Li * SVE Available. 15*412f47f9SXin Li */ 16*412f47f9SXin Li 17*412f47f9SXin LiENTRY (__memchr_aarch64_sve) 18*412f47f9SXin Li PTR_ARG (0) 19*412f47f9SXin Li SIZE_ARG (2) 20*412f47f9SXin Li dup z1.b, w1 /* duplicate c to a vector */ 21*412f47f9SXin Li setffr /* initialize FFR */ 22*412f47f9SXin Li mov x3, 0 /* initialize off */ 23*412f47f9SXin Li 24*412f47f9SXin Li .p2align 4 25*412f47f9SXin Li0: whilelo p1.b, x3, x2 /* make sure off < max */ 26*412f47f9SXin Li b.none 9f 27*412f47f9SXin Li 28*412f47f9SXin Li /* Read a vector's worth of bytes, bounded by max, 29*412f47f9SXin Li stopping on first fault. */ 30*412f47f9SXin Li ldff1b z0.b, p1/z, [x0, x3] 31*412f47f9SXin Li rdffrs p0.b, p1/z 32*412f47f9SXin Li b.nlast 2f 33*412f47f9SXin Li 34*412f47f9SXin Li /* First fault did not fail: the vector bounded by max is valid. 35*412f47f9SXin Li Avoid depending on the contents of FFR beyond the branch. */ 36*412f47f9SXin Li incb x3 /* speculate increment */ 37*412f47f9SXin Li cmpeq p2.b, p1/z, z0.b, z1.b /* search for c */ 38*412f47f9SXin Li b.none 0b 39*412f47f9SXin Li decb x3 /* undo speculate */ 40*412f47f9SXin Li 41*412f47f9SXin Li /* Found C. */ 42*412f47f9SXin Li1: brkb p2.b, p1/z, p2.b /* find the first c */ 43*412f47f9SXin Li add x0, x0, x3 /* form partial pointer */ 44*412f47f9SXin Li incp x0, p2.b /* form final pointer to c */ 45*412f47f9SXin Li ret 46*412f47f9SXin Li 47*412f47f9SXin Li /* First fault failed: only some of the vector is valid. 48*412f47f9SXin Li Perform the comparision only on the valid bytes. */ 49*412f47f9SXin Li2: cmpeq p2.b, p0/z, z0.b, z1.b 50*412f47f9SXin Li b.any 1b 51*412f47f9SXin Li 52*412f47f9SXin Li /* No C found. Re-init FFR, increment, and loop. */ 53*412f47f9SXin Li setffr 54*412f47f9SXin Li incp x3, p0.b 55*412f47f9SXin Li b 0b 56*412f47f9SXin Li 57*412f47f9SXin Li /* Found end of count. */ 58*412f47f9SXin Li9: mov x0, 0 /* return null */ 59*412f47f9SXin Li ret 60*412f47f9SXin Li 61*412f47f9SXin LiEND (__memchr_aarch64_sve) 62*412f47f9SXin Li 63*412f47f9SXin Li#endif 64*412f47f9SXin Li 65