| 1 | /* |
| 2 | * strchrnul - find a character or nul in a string |
| 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-exception |
| 7 | */ |
| 8 | |
| 9 | /* Assumptions: |
| 10 | * |
| 11 | * ARMv8-a, AArch64 |
| 12 | * Neon Available. |
| 13 | */ |
| 14 | |
| 15 | #include "../asmdefs.h" |
| 16 | |
| 17 | /* Arguments and results. */ |
| 18 | #define srcin x0 |
| 19 | #define chrin w1 |
| 20 | |
| 21 | #define result x0 |
| 22 | |
| 23 | #define src x2 |
| 24 | #define tmp1 x3 |
| 25 | #define wtmp2 w4 |
| 26 | #define tmp3 x5 |
| 27 | |
| 28 | #define vrepchr v0 |
| 29 | #define vdata1 v1 |
| 30 | #define vdata2 v2 |
| 31 | #define vhas_nul1 v3 |
| 32 | #define vhas_nul2 v4 |
| 33 | #define vhas_chr1 v5 |
| 34 | #define vhas_chr2 v6 |
| 35 | #define vrepmask v7 |
| 36 | #define vend1 v16 |
| 37 | |
| 38 | /* Core algorithm. |
| 39 | |
| 40 | For each 32-byte hunk we calculate a 64-bit syndrome value, with |
| 41 | two bits per byte (LSB is always in bits 0 and 1, for both big |
| 42 | and little-endian systems). For each tuple, bit 0 is set iff |
| 43 | the relevant byte matched the requested character or nul. Since the |
| 44 | bits in the syndrome reflect exactly the order in which things occur |
| 45 | in the original string a count_trailing_zeros() operation will |
| 46 | identify exactly which byte is causing the termination. */ |
| 47 | |
| 48 | /* Locals and temporaries. */ |
| 49 | |
| 50 | ENTRY (__strchrnul_aarch64) |
| 51 | /* Magic constant 0x40100401 to allow us to identify which lane |
| 52 | matches the termination condition. */ |
| 53 | mov wtmp2, #0x0401 |
| 54 | movk wtmp2, #0x4010, lsl #16 |
| 55 | dup vrepchr.16b, chrin |
| 56 | bic src, srcin, #31 /* Work with aligned 32-byte hunks. */ |
| 57 | dup vrepmask.4s, wtmp2 |
| 58 | ands tmp1, srcin, #31 |
| 59 | b.eq L(loop) |
| 60 | |
| 61 | /* Input string is not 32-byte aligned. Rather than forcing |
| 62 | the padding bytes to a safe value, we calculate the syndrome |
| 63 | for all the bytes, but then mask off those bits of the |
| 64 | syndrome that are related to the padding. */ |
| 65 | ld1 {vdata1.16b, vdata2.16b}, [src], #32 |
| 66 | neg tmp1, tmp1 |
| 67 | cmeq vhas_nul1.16b, vdata1.16b, #0 |
| 68 | cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b |
| 69 | cmeq vhas_nul2.16b, vdata2.16b, #0 |
| 70 | cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b |
| 71 | orr vhas_chr1.16b, vhas_chr1.16b, vhas_nul1.16b |
| 72 | orr vhas_chr2.16b, vhas_chr2.16b, vhas_nul2.16b |
| 73 | and vhas_chr1.16b, vhas_chr1.16b, vrepmask.16b |
| 74 | and vhas_chr2.16b, vhas_chr2.16b, vrepmask.16b |
| 75 | lsl tmp1, tmp1, #1 |
| 76 | addp vend1.16b, vhas_chr1.16b, vhas_chr2.16b // 256->128 |
| 77 | mov tmp3, #~0 |
| 78 | addp vend1.16b, vend1.16b, vend1.16b // 128->64 |
| 79 | lsr tmp1, tmp3, tmp1 |
| 80 | |
| 81 | mov tmp3, vend1.d[0] |
| 82 | bic tmp1, tmp3, tmp1 // Mask padding bits. |
| 83 | cbnz tmp1, L(tail) |
| 84 | |
| 85 | L(loop): |
| 86 | ld1 {vdata1.16b, vdata2.16b}, [src], #32 |
| 87 | cmeq vhas_nul1.16b, vdata1.16b, #0 |
| 88 | cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b |
| 89 | cmeq vhas_nul2.16b, vdata2.16b, #0 |
| 90 | cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b |
| 91 | /* Use a fast check for the termination condition. */ |
| 92 | orr vhas_chr1.16b, vhas_nul1.16b, vhas_chr1.16b |
| 93 | orr vhas_chr2.16b, vhas_nul2.16b, vhas_chr2.16b |
| 94 | orr vend1.16b, vhas_chr1.16b, vhas_chr2.16b |
| 95 | addp vend1.2d, vend1.2d, vend1.2d |
| 96 | mov tmp1, vend1.d[0] |
| 97 | cbz tmp1, L(loop) |
| 98 | |
| 99 | /* Termination condition found. Now need to establish exactly why |
| 100 | we terminated. */ |
| 101 | and vhas_chr1.16b, vhas_chr1.16b, vrepmask.16b |
| 102 | and vhas_chr2.16b, vhas_chr2.16b, vrepmask.16b |
| 103 | addp vend1.16b, vhas_chr1.16b, vhas_chr2.16b // 256->128 |
| 104 | addp vend1.16b, vend1.16b, vend1.16b // 128->64 |
| 105 | |
| 106 | mov tmp1, vend1.d[0] |
| 107 | L(tail): |
| 108 | /* Count the trailing zeros, by bit reversing... */ |
| 109 | rbit tmp1, tmp1 |
| 110 | /* Re-bias source. */ |
| 111 | sub src, src, #32 |
| 112 | clz tmp1, tmp1 /* ... and counting the leading zeros. */ |
| 113 | /* tmp1 is twice the offset into the fragment. */ |
| 114 | add result, src, tmp1, lsr #1 |
| 115 | ret |
| 116 | |
| 117 | END (__strchrnul_aarch64) |
| 118 | |