1*412f47f9SXin Li/* 2*412f47f9SXin Li * __strlen_aarch64_sve - compute the length of a string 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 (__strlen_aarch64_sve) 18*412f47f9SXin Li PTR_ARG (0) 19*412f47f9SXin Li setffr /* initialize FFR */ 20*412f47f9SXin Li ptrue p2.b /* all ones; loop invariant */ 21*412f47f9SXin Li mov x1, 0 /* initialize length */ 22*412f47f9SXin Li 23*412f47f9SXin Li /* Read a vector's worth of bytes, stopping on first fault. */ 24*412f47f9SXin Li .p2align 4 25*412f47f9SXin Li0: ldff1b z0.b, p2/z, [x0, x1] 26*412f47f9SXin Li rdffrs p0.b, p2/z 27*412f47f9SXin Li b.nlast 2f 28*412f47f9SXin Li 29*412f47f9SXin Li /* First fault did not fail: the whole vector is valid. 30*412f47f9SXin Li Avoid depending on the contents of FFR beyond the branch. */ 31*412f47f9SXin Li incb x1, all /* speculate increment */ 32*412f47f9SXin Li cmpeq p1.b, p2/z, z0.b, 0 /* loop if no zeros */ 33*412f47f9SXin Li b.none 0b 34*412f47f9SXin Li decb x1, all /* undo speculate */ 35*412f47f9SXin Li 36*412f47f9SXin Li /* Zero found. Select the bytes before the first and count them. */ 37*412f47f9SXin Li1: brkb p0.b, p2/z, p1.b 38*412f47f9SXin Li incp x1, p0.b 39*412f47f9SXin Li mov x0, x1 40*412f47f9SXin Li ret 41*412f47f9SXin Li 42*412f47f9SXin Li /* First fault failed: only some of the vector is valid. 43*412f47f9SXin Li Perform the comparison only on the valid bytes. */ 44*412f47f9SXin Li2: cmpeq p1.b, p0/z, z0.b, 0 45*412f47f9SXin Li b.any 1b 46*412f47f9SXin Li 47*412f47f9SXin Li /* No zero found. Re-init FFR, increment, and loop. */ 48*412f47f9SXin Li setffr 49*412f47f9SXin Li incp x1, p0.b 50*412f47f9SXin Li b 0b 51*412f47f9SXin Li 52*412f47f9SXin LiEND (__strlen_aarch64_sve) 53*412f47f9SXin Li 54*412f47f9SXin Li#endif 55*412f47f9SXin Li 56