157 lines · plain
1/*2 * strnlen - calculate the length of a string with limit.3 *4 * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.5 * See https://llvm.org/LICENSE.txt for license information.6 * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception7 */8 9/* Assumptions:10 *11 * ARMv8-a, AArch6412 */13 14#include "../asmdefs.h"15 16/* Arguments and results. */17#define srcin x018#define len x019#define limit x120 21/* Locals and temporaries. */22#define src x223#define data1 x324#define data2 x425#define data2a x526#define has_nul1 x627#define has_nul2 x728#define tmp1 x829#define tmp2 x930#define tmp3 x1031#define tmp4 x1132#define zeroones x1233#define pos x1334#define limit_wd x1435 36#define REP8_01 0x010101010101010137#define REP8_7f 0x7f7f7f7f7f7f7f7f38#define REP8_80 0x808080808080808039 40 .text41 .p2align 642L(start):43 /* Pre-pad to ensure critical loop begins an icache line. */44 .rep 745 nop46 .endr47 /* Put this code here to avoid wasting more space with pre-padding. */48L(hit_limit):49 mov len, limit50 ret51 52ENTRY_ALIGN (__strnlen_aarch64, 0)53 cbz limit, L(hit_limit)54 mov zeroones, #REP8_0155 bic src, srcin, #1556 ands tmp1, srcin, #1557 b.ne L(misaligned)58 /* Calculate the number of full and partial words -1. */59 sub limit_wd, limit, #1 /* Limit != 0, so no underflow. */60 lsr limit_wd, limit_wd, #4 /* Convert to Qwords. */61 62 /* NUL detection works on the principle that (X - 1) & (~X) & 0x8063 (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and64 can be done in parallel across the entire word. */65 /* The inner loop deals with two Dwords at a time. This has a66 slightly higher start-up cost, but we should win quite quickly,67 especially on cores with a high number of issue slots per68 cycle, as we get much better parallelism out of the operations. */69 70 /* Start of critical section -- keep to one 64Byte cache line. */71L(loop):72 ldp data1, data2, [src], #1673L(realigned):74 sub tmp1, data1, zeroones75 orr tmp2, data1, #REP8_7f76 sub tmp3, data2, zeroones77 orr tmp4, data2, #REP8_7f78 bic has_nul1, tmp1, tmp279 bic has_nul2, tmp3, tmp480 subs limit_wd, limit_wd, #181 orr tmp1, has_nul1, has_nul282 ccmp tmp1, #0, #0, pl /* NZCV = 0000 */83 b.eq L(loop)84 /* End of critical section -- keep to one 64Byte cache line. */85 86 orr tmp1, has_nul1, has_nul287 cbz tmp1, L(hit_limit) /* No null in final Qword. */88 89 /* We know there's a null in the final Qword. The easiest thing90 to do now is work out the length of the string and return91 MIN (len, limit). */92 93 sub len, src, srcin94 cbz has_nul1, L(nul_in_data2)95#ifdef __AARCH64EB__96 mov data2, data197#endif98 sub len, len, #899 mov has_nul2, has_nul1100L(nul_in_data2):101#ifdef __AARCH64EB__102 /* For big-endian, carry propagation (if the final byte in the103 string is 0x01) means we cannot use has_nul directly. The104 easiest way to get the correct byte is to byte-swap the data105 and calculate the syndrome a second time. */106 rev data2, data2107 sub tmp1, data2, zeroones108 orr tmp2, data2, #REP8_7f109 bic has_nul2, tmp1, tmp2110#endif111 sub len, len, #8112 rev has_nul2, has_nul2113 clz pos, has_nul2114 add len, len, pos, lsr #3 /* Bits to bytes. */115 cmp len, limit116 csel len, len, limit, ls /* Return the lower value. */117 ret118 119L(misaligned):120 /* Deal with a partial first word.121 We're doing two things in parallel here;122 1) Calculate the number of words (but avoiding overflow if123 limit is near ULONG_MAX) - to do this we need to work out124 limit + tmp1 - 1 as a 65-bit value before shifting it;125 2) Load and mask the initial data words - we force the bytes126 before the ones we are interested in to 0xff - this ensures127 early bytes will not hit any zero detection. */128 sub limit_wd, limit, #1129 neg tmp4, tmp1130 cmp tmp1, #8131 132 and tmp3, limit_wd, #15133 lsr limit_wd, limit_wd, #4134 mov tmp2, #~0135 136 ldp data1, data2, [src], #16137 lsl tmp4, tmp4, #3 /* Bytes beyond alignment -> bits. */138 add tmp3, tmp3, tmp1139 140#ifdef __AARCH64EB__141 /* Big-endian. Early bytes are at MSB. */142 lsl tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */143#else144 /* Little-endian. Early bytes are at LSB. */145 lsr tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */146#endif147 add limit_wd, limit_wd, tmp3, lsr #4148 149 orr data1, data1, tmp2150 orr data2a, data2, tmp2151 152 csinv data1, data1, xzr, le153 csel data2, data2, data2a, le154 b L(realigned)155 156END (__strnlen_aarch64)157