| 1 | /* |
| 2 | * strlen - calculate the length of 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 | #if __ARM_ARCH >= 6 && __ARM_ARCH_ISA_THUMB == 2 |
| 10 | |
| 11 | /* |
| 12 | Assumes: |
| 13 | ARMv6T2, AArch32 |
| 14 | |
| 15 | */ |
| 16 | |
| 17 | #include "../asmdefs.h" |
| 18 | |
| 19 | #ifdef __ARMEB__ |
| 20 | #define S2LO lsl |
| 21 | #define S2HI lsr |
| 22 | #else |
| 23 | #define S2LO lsr |
| 24 | #define S2HI lsl |
| 25 | #endif |
| 26 | |
| 27 | /* This code requires Thumb. */ |
| 28 | .thumb |
| 29 | .syntax unified |
| 30 | |
| 31 | /* Parameters and result. */ |
| 32 | #define srcin r0 |
| 33 | #define result r0 |
| 34 | |
| 35 | /* Internal variables. */ |
| 36 | #define src r1 |
| 37 | #define data1a r2 |
| 38 | #define data1b r3 |
| 39 | #define const_m1 r12 |
| 40 | #define const_0 r4 |
| 41 | #define tmp1 r4 /* Overlaps const_0 */ |
| 42 | #define tmp2 r5 |
| 43 | |
| 44 | ENTRY (__strlen_armv6t2) |
| 45 | pld [srcin, #0] |
| 46 | strd r4, r5, [sp, #-8]! |
| 47 | bic src, srcin, #7 |
| 48 | mvn const_m1, #0 |
| 49 | ands tmp1, srcin, #7 /* (8 - bytes) to alignment. */ |
| 50 | pld [src, #32] |
| 51 | bne.w L(misaligned8) |
| 52 | mov const_0, #0 |
| 53 | mov result, #-8 |
| 54 | L(loop_aligned): |
| 55 | /* Bytes 0-7. */ |
| 56 | ldrd data1a, data1b, [src] |
| 57 | pld [src, #64] |
| 58 | add result, result, #8 |
| 59 | L(start_realigned): |
| 60 | uadd8 data1a, data1a, const_m1 /* Saturating GE<0:3> set. */ |
| 61 | sel data1a, const_0, const_m1 /* Select based on GE<0:3>. */ |
| 62 | uadd8 data1b, data1b, const_m1 |
| 63 | sel data1b, data1a, const_m1 /* Only used if d1a == 0. */ |
| 64 | cbnz data1b, L(null_found) |
| 65 | |
| 66 | /* Bytes 8-15. */ |
| 67 | ldrd data1a, data1b, [src, #8] |
| 68 | uadd8 data1a, data1a, const_m1 /* Saturating GE<0:3> set. */ |
| 69 | add result, result, #8 |
| 70 | sel data1a, const_0, const_m1 /* Select based on GE<0:3>. */ |
| 71 | uadd8 data1b, data1b, const_m1 |
| 72 | sel data1b, data1a, const_m1 /* Only used if d1a == 0. */ |
| 73 | cbnz data1b, L(null_found) |
| 74 | |
| 75 | /* Bytes 16-23. */ |
| 76 | ldrd data1a, data1b, [src, #16] |
| 77 | uadd8 data1a, data1a, const_m1 /* Saturating GE<0:3> set. */ |
| 78 | add result, result, #8 |
| 79 | sel data1a, const_0, const_m1 /* Select based on GE<0:3>. */ |
| 80 | uadd8 data1b, data1b, const_m1 |
| 81 | sel data1b, data1a, const_m1 /* Only used if d1a == 0. */ |
| 82 | cbnz data1b, L(null_found) |
| 83 | |
| 84 | /* Bytes 24-31. */ |
| 85 | ldrd data1a, data1b, [src, #24] |
| 86 | add src, src, #32 |
| 87 | uadd8 data1a, data1a, const_m1 /* Saturating GE<0:3> set. */ |
| 88 | add result, result, #8 |
| 89 | sel data1a, const_0, const_m1 /* Select based on GE<0:3>. */ |
| 90 | uadd8 data1b, data1b, const_m1 |
| 91 | sel data1b, data1a, const_m1 /* Only used if d1a == 0. */ |
| 92 | cmp data1b, #0 |
| 93 | beq L(loop_aligned) |
| 94 | |
| 95 | L(null_found): |
| 96 | cmp data1a, #0 |
| 97 | itt eq |
| 98 | addeq result, result, #4 |
| 99 | moveq data1a, data1b |
| 100 | #ifndef __ARMEB__ |
| 101 | rev data1a, data1a |
| 102 | #endif |
| 103 | clz data1a, data1a |
| 104 | ldrd r4, r5, [sp], #8 |
| 105 | add result, result, data1a, lsr #3 /* Bits -> Bytes. */ |
| 106 | bx lr |
| 107 | |
| 108 | L(misaligned8): |
| 109 | ldrd data1a, data1b, [src] |
| 110 | and tmp2, tmp1, #3 |
| 111 | rsb result, tmp1, #0 |
| 112 | lsl tmp2, tmp2, #3 /* Bytes -> bits. */ |
| 113 | tst tmp1, #4 |
| 114 | pld [src, #64] |
| 115 | S2HI tmp2, const_m1, tmp2 |
| 116 | orn data1a, data1a, tmp2 |
| 117 | itt ne |
| 118 | ornne data1b, data1b, tmp2 |
| 119 | movne data1a, const_m1 |
| 120 | mov const_0, #0 |
| 121 | b L(start_realigned) |
| 122 | |
| 123 | END (__strlen_armv6t2) |
| 124 | |
| 125 | #endif /* __ARM_ARCH >= 6 && __ARM_ARCH_ISA_THUMB == 2 */ |
| 126 | |