| caab277b |
1 | /* SPDX-License-Identifier: GPL-2.0-only */ |
| 0a42cb0a |
2 | /* |
| 3 | * Copyright (C) 2013 ARM Ltd. |
| 4 | * Copyright (C) 2013 Linaro. |
| 5 | * |
| 6 | * This code is based on glibc cortex strings work originally authored by Linaro |
| 0a42cb0a |
7 | * be found @ |
| 8 | * |
| 9 | * http://bazaar.launchpad.net/~linaro-toolchain-dev/cortex-strings/trunk/ |
| 10 | * files/head:/src/aarch64/ |
| 0a42cb0a |
11 | */ |
| 12 | |
| 13 | #include <linux/linkage.h> |
| 14 | #include <asm/assembler.h> |
| 15 | |
| 16 | /* |
| 17 | * calculate the length of a string |
| 18 | * |
| 19 | * Parameters: |
| 20 | * x0 - const string pointer |
| 21 | * Returns: |
| 22 | * x0 - the return length of specific string |
| 23 | */ |
| 24 | |
| 25 | /* Arguments and results. */ |
| 26 | srcin .req x0 |
| 27 | len .req x0 |
| 28 | |
| 29 | /* Locals and temporaries. */ |
| 30 | src .req x1 |
| 31 | data1 .req x2 |
| 32 | data2 .req x3 |
| 33 | data2a .req x4 |
| 34 | has_nul1 .req x5 |
| 35 | has_nul2 .req x6 |
| 36 | tmp1 .req x7 |
| 37 | tmp2 .req x8 |
| 38 | tmp3 .req x9 |
| 39 | tmp4 .req x10 |
| 40 | zeroones .req x11 |
| 41 | pos .req x12 |
| 42 | |
| 43 | #define REP8_01 0x0101010101010101 |
| 44 | #define REP8_7f 0x7f7f7f7f7f7f7f7f |
| 45 | #define REP8_80 0x8080808080808080 |
| 46 | |
| 19a2ca0f |
47 | WEAK(strlen) |
| 0a42cb0a |
48 | mov zeroones, #REP8_01 |
| 49 | bic src, srcin, #15 |
| 50 | ands tmp1, srcin, #15 |
| 51 | b.ne .Lmisaligned |
| 52 | /* |
| 53 | * NUL detection works on the principle that (X - 1) & (~X) & 0x80 |
| 54 | * (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and |
| 55 | * can be done in parallel across the entire word. |
| 56 | */ |
| 57 | /* |
| 58 | * The inner loop deals with two Dwords at a time. This has a |
| 59 | * slightly higher start-up cost, but we should win quite quickly, |
| 60 | * especially on cores with a high number of issue slots per |
| 61 | * cycle, as we get much better parallelism out of the operations. |
| 62 | */ |
| 63 | .Lloop: |
| 64 | ldp data1, data2, [src], #16 |
| 65 | .Lrealigned: |
| 66 | sub tmp1, data1, zeroones |
| 67 | orr tmp2, data1, #REP8_7f |
| 68 | sub tmp3, data2, zeroones |
| 69 | orr tmp4, data2, #REP8_7f |
| 70 | bic has_nul1, tmp1, tmp2 |
| 71 | bics has_nul2, tmp3, tmp4 |
| 72 | ccmp has_nul1, #0, #0, eq /* NZCV = 0000 */ |
| 73 | b.eq .Lloop |
| 74 | |
| 75 | sub len, src, srcin |
| 76 | cbz has_nul1, .Lnul_in_data2 |
| 77 | CPU_BE( mov data2, data1 ) /*prepare data to re-calculate the syndrome*/ |
| 78 | sub len, len, #8 |
| 79 | mov has_nul2, has_nul1 |
| 80 | .Lnul_in_data2: |
| 81 | /* |
| 82 | * For big-endian, carry propagation (if the final byte in the |
| 83 | * string is 0x01) means we cannot use has_nul directly. The |
| 84 | * easiest way to get the correct byte is to byte-swap the data |
| 85 | * and calculate the syndrome a second time. |
| 86 | */ |
| 87 | CPU_BE( rev data2, data2 ) |
| 88 | CPU_BE( sub tmp1, data2, zeroones ) |
| 89 | CPU_BE( orr tmp2, data2, #REP8_7f ) |
| 90 | CPU_BE( bic has_nul2, tmp1, tmp2 ) |
| 91 | |
| 92 | sub len, len, #8 |
| 93 | rev has_nul2, has_nul2 |
| 94 | clz pos, has_nul2 |
| 95 | add len, len, pos, lsr #3 /* Bits to bytes. */ |
| 96 | ret |
| 97 | |
| 98 | .Lmisaligned: |
| 99 | cmp tmp1, #8 |
| 100 | neg tmp1, tmp1 |
| 101 | ldp data1, data2, [src], #16 |
| 102 | lsl tmp1, tmp1, #3 /* Bytes beyond alignment -> bits. */ |
| 103 | mov tmp2, #~0 |
| 104 | /* Big-endian. Early bytes are at MSB. */ |
| 105 | CPU_BE( lsl tmp2, tmp2, tmp1 ) /* Shift (tmp1 & 63). */ |
| 106 | /* Little-endian. Early bytes are at LSB. */ |
| 107 | CPU_LE( lsr tmp2, tmp2, tmp1 ) /* Shift (tmp1 & 63). */ |
| 108 | |
| 109 | orr data1, data1, tmp2 |
| 110 | orr data2a, data2, tmp2 |
| 111 | csinv data1, data1, xzr, le |
| 112 | csel data2, data2, data2a, le |
| 113 | b .Lrealigned |
| 20791846 |
114 | ENDPIPROC(strlen) |
| ac0e8c72 |
115 | EXPORT_SYMBOL_NOKASAN(strlen) |
projects / thirdparty / linux.git / blame