|
| 1 | +#if defined(__wasm_simd128__) && defined(__wasilibc_simd_string) |
| 2 | +// Skip Clang 19 and Clang 20 which have a bug (llvm/llvm-project#146574) |
| 3 | +// which results in an ICE when inline assembly is used with a vector result. |
| 4 | +#if __clang_major__ != 19 && __clang_major__ != 20 |
| 5 | + |
| 6 | +#include <stdint.h> |
| 7 | +#include <string.h> |
| 8 | +#include <wasm_simd128.h> |
| 9 | + |
| 10 | +#if !defined(__wasm_relaxed_simd__) || !defined(__RELAXED_FN_ATTRS) |
| 11 | +#define wasm_i8x16_relaxed_swizzle wasm_i8x16_swizzle |
| 12 | +#endif |
| 13 | + |
| 14 | +// SIMDized check which bytes are in a set (Geoff Langdale) |
| 15 | +// http://0x80.pl/notesen/2018-10-18-simd-byte-lookup.html |
| 16 | + |
| 17 | +// This is the same algorithm as truffle from Hyperscan: |
| 18 | +// https://github.com/intel/hyperscan/blob/v5.4.2/src/nfa/truffle.c#L64-L81 |
| 19 | +// https://github.com/intel/hyperscan/blob/v5.4.2/src/nfa/trufflecompile.cpp |
| 20 | + |
| 21 | +typedef struct { |
| 22 | + __u8x16 lo; |
| 23 | + __u8x16 hi; |
| 24 | +} __wasm_v128_bitmap256_t; |
| 25 | + |
| 26 | +__attribute__((always_inline)) |
| 27 | +static void __wasm_v128_setbit(__wasm_v128_bitmap256_t *bitmap, uint8_t i) { |
| 28 | + uint8_t hi_nibble = i >> 4; |
| 29 | + uint8_t lo_nibble = i & 0xf; |
| 30 | + bitmap->lo[lo_nibble] |= (uint8_t)(1u << (hi_nibble - 0)); |
| 31 | + bitmap->hi[lo_nibble] |= (uint8_t)(1u << (hi_nibble - 8)); |
| 32 | +} |
| 33 | + |
| 34 | +__attribute__((always_inline)) |
| 35 | +static v128_t __wasm_v128_chkbits(__wasm_v128_bitmap256_t bitmap, v128_t v) { |
| 36 | + v128_t hi_nibbles = wasm_u8x16_shr(v, 4); |
| 37 | + v128_t bitmask_lookup = wasm_u64x2_const_splat(0x8040201008040201); |
| 38 | + v128_t bitmask = wasm_i8x16_relaxed_swizzle(bitmask_lookup, hi_nibbles); |
| 39 | + |
| 40 | + v128_t indices_0_7 = v & wasm_u8x16_const_splat(0x8f); |
| 41 | + v128_t indices_8_15 = indices_0_7 ^ wasm_u8x16_const_splat(0x80); |
| 42 | + |
| 43 | + v128_t row_0_7 = wasm_i8x16_swizzle((v128_t)bitmap.lo, indices_0_7); |
| 44 | + v128_t row_8_15 = wasm_i8x16_swizzle((v128_t)bitmap.hi, indices_8_15); |
| 45 | + |
| 46 | + v128_t bitsets = row_0_7 | row_8_15; |
| 47 | + return bitsets & bitmask; |
| 48 | +} |
| 49 | + |
| 50 | +size_t strspn(const char *s, const char *c) |
| 51 | +{ |
| 52 | + // Note that reading before/after the allocation of a pointer is UB in |
| 53 | + // C, so inline assembly is used to generate the exact machine |
| 54 | + // instruction we want with opaque semantics to the compiler to avoid |
| 55 | + // the UB. |
| 56 | + uintptr_t align = (uintptr_t)s % sizeof(v128_t); |
| 57 | + uintptr_t addr = (uintptr_t)s - align; |
| 58 | + |
| 59 | + if (!c[0]) return 0; |
| 60 | + if (!c[1]) { |
| 61 | + v128_t vc = wasm_i8x16_splat(*c); |
| 62 | + for (;;) { |
| 63 | + v128_t v; |
| 64 | + __asm__( |
| 65 | + "local.get %1\n" |
| 66 | + "v128.load 0\n" |
| 67 | + "local.set %0\n" |
| 68 | + : "=r"(v) |
| 69 | + : "r"(addr) |
| 70 | + : "memory"); |
| 71 | + v128_t cmp = wasm_i8x16_eq(v, vc); |
| 72 | + // Bitmask is slow on AArch64, all_true is much faster. |
| 73 | + if (!wasm_i8x16_all_true(cmp)) { |
| 74 | + // Clear the bits corresponding to align (little-endian) |
| 75 | + // so we can count trailing zeros. |
| 76 | + int mask = (uint16_t)~wasm_i8x16_bitmask(cmp) >> align << align; |
| 77 | + // At least one bit will be set, unless align cleared them. |
| 78 | + // Knowing this helps the compiler if it unrolls the loop. |
| 79 | + __builtin_assume(mask || align); |
| 80 | + // If the mask became zero because of align, |
| 81 | + // it's as if we didn't find anything. |
| 82 | + if (mask) { |
| 83 | + // Find the offset of the first one bit (little-endian). |
| 84 | + return addr - (uintptr_t)s + __builtin_ctz(mask); |
| 85 | + } |
| 86 | + } |
| 87 | + align = 0; |
| 88 | + addr += sizeof(v128_t); |
| 89 | + } |
| 90 | + } |
| 91 | + |
| 92 | + __wasm_v128_bitmap256_t bitmap = {}; |
| 93 | + |
| 94 | + for (; *c; c++) { |
| 95 | + // Terminator IS NOT on the bitmap. |
| 96 | + __wasm_v128_setbit(&bitmap, (uint8_t)*c); |
| 97 | + } |
| 98 | + |
| 99 | + for (;;) { |
| 100 | + v128_t v; |
| 101 | + __asm__( |
| 102 | + "local.get %1\n" |
| 103 | + "v128.load 0\n" |
| 104 | + "local.set %0\n" |
| 105 | + : "=r"(v) |
| 106 | + : "r"(addr) |
| 107 | + : "memory"); |
| 108 | + v128_t found = __wasm_v128_chkbits(bitmap, v); |
| 109 | + // Bitmask is slow on AArch64, all_true is much faster. |
| 110 | + if (!wasm_i8x16_all_true(found)) { |
| 111 | + v128_t cmp = wasm_i8x16_eq(found, (v128_t){}); |
| 112 | + // Clear the bits corresponding to align (little-endian) |
| 113 | + // so we can count trailing zeros. |
| 114 | + int mask = wasm_i8x16_bitmask(cmp) >> align << align; |
| 115 | + // At least one bit will be set, unless align cleared them. |
| 116 | + // Knowing this helps the compiler if it unrolls the loop. |
| 117 | + __builtin_assume(mask || align); |
| 118 | + // If the mask became zero because of align, |
| 119 | + // it's as if we didn't find anything. |
| 120 | + if (mask) { |
| 121 | + // Find the offset of the first one bit (little-endian). |
| 122 | + return addr - (uintptr_t)s + __builtin_ctz(mask); |
| 123 | + } |
| 124 | + } |
| 125 | + align = 0; |
| 126 | + addr += sizeof(v128_t); |
| 127 | + } |
| 128 | +} |
| 129 | + |
| 130 | +size_t strcspn(const char *s, const char *c) |
| 131 | +{ |
| 132 | + if (!c[0] || !c[1]) return __strchrnul(s, *c) - s; |
| 133 | + |
| 134 | + // Note that reading before/after the allocation of a pointer is UB in |
| 135 | + // C, so inline assembly is used to generate the exact machine |
| 136 | + // instruction we want with opaque semantics to the compiler to avoid |
| 137 | + // the UB. |
| 138 | + uintptr_t align = (uintptr_t)s % sizeof(v128_t); |
| 139 | + uintptr_t addr = (uintptr_t)s - align; |
| 140 | + |
| 141 | + __wasm_v128_bitmap256_t bitmap = {}; |
| 142 | + |
| 143 | + do { |
| 144 | + // Terminator IS on the bitmap. |
| 145 | + __wasm_v128_setbit(&bitmap, (uint8_t)*c); |
| 146 | + } while (*c++); |
| 147 | + |
| 148 | + for (;;) { |
| 149 | + v128_t v; |
| 150 | + __asm__( |
| 151 | + "local.get %1\n" |
| 152 | + "v128.load 0\n" |
| 153 | + "local.set %0\n" |
| 154 | + : "=r"(v) |
| 155 | + : "r"(addr) |
| 156 | + : "memory"); |
| 157 | + v128_t found = __wasm_v128_chkbits(bitmap, v); |
| 158 | + // Bitmask is slow on AArch64, any_true is much faster. |
| 159 | + if (wasm_v128_any_true(found)) { |
| 160 | + v128_t cmp = wasm_i8x16_eq(found, (v128_t){}); |
| 161 | + // Clear the bits corresponding to align (little-endian) |
| 162 | + // so we can count trailing zeros. |
| 163 | + int mask = (uint16_t)~wasm_i8x16_bitmask(cmp) >> align << align; |
| 164 | + // At least one bit will be set, unless align cleared them. |
| 165 | + // Knowing this helps the compiler if it unrolls the loop. |
| 166 | + __builtin_assume(mask || align); |
| 167 | + // If the mask became zero because of align, |
| 168 | + // it's as if we didn't find anything. |
| 169 | + if (mask) { |
| 170 | + // Find the offset of the first one bit (little-endian). |
| 171 | + return addr - (uintptr_t)s + __builtin_ctz(mask); |
| 172 | + } |
| 173 | + } |
| 174 | + align = 0; |
| 175 | + addr += sizeof(v128_t); |
| 176 | + } |
| 177 | +} |
| 178 | + |
| 179 | +#endif |
| 180 | +#endif |
0 commit comments