Warning: This file is not a C or C++ file. It does not have highlighting.
| 1 | //===-- aarch64 implementation of memory function building blocks ---------===// |
|---|---|
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file provides aarch64 specific building blocks to compose memory |
| 10 | // functions. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | #ifndef LLVM_LIBC_SRC_STRING_MEMORY_UTILS_OP_AARCH64_H |
| 14 | #define LLVM_LIBC_SRC_STRING_MEMORY_UTILS_OP_AARCH64_H |
| 15 | |
| 16 | #include "src/__support/macros/attributes.h" // LIBC_INLINE |
| 17 | #include "src/__support/macros/config.h" // LIBC_NAMESPACE_DECL |
| 18 | #include "src/__support/macros/properties/architectures.h" |
| 19 | |
| 20 | #if defined(LIBC_TARGET_ARCH_IS_AARCH64) |
| 21 | |
| 22 | #include "src/__support/CPP/type_traits.h" // cpp::always_false |
| 23 | #include "src/__support/common.h" |
| 24 | #include "src/string/memory_utils/op_generic.h" |
| 25 | |
| 26 | #ifdef __ARM_NEON |
| 27 | #include <arm_neon.h> |
| 28 | |
| 29 | namespace LIBC_NAMESPACE_DECL { |
| 30 | namespace aarch64 { |
| 31 | |
| 32 | LIBC_INLINE_VAR constexpr bool kNeon = LLVM_LIBC_IS_DEFINED(__ARM_NEON); |
| 33 | |
| 34 | namespace neon { |
| 35 | |
| 36 | struct BzeroCacheLine { |
| 37 | static constexpr size_t SIZE = 64; |
| 38 | |
| 39 | LIBC_INLINE static void block(Ptr dst, uint8_t) { |
| 40 | #if __SIZEOF_POINTER__ == 4 |
| 41 | asm("dc zva, %w[dst]" : : [dst] "r"(dst) : "memory"); |
| 42 | #else |
| 43 | asm("dc zva, %[dst]" : : [dst] "r"(dst) : "memory"); |
| 44 | #endif |
| 45 | } |
| 46 | |
| 47 | LIBC_INLINE static void loop_and_tail(Ptr dst, uint8_t value, size_t count) { |
| 48 | size_t offset = 0; |
| 49 | do { |
| 50 | block(dst + offset, value); |
| 51 | offset += SIZE; |
| 52 | } while (offset < count - SIZE); |
| 53 | // Unaligned store, we can't use 'dc zva' here. |
| 54 | generic::Memset<generic_v512>::tail(dst, value, count); |
| 55 | } |
| 56 | }; |
| 57 | |
| 58 | LIBC_INLINE bool hasZva() { |
| 59 | uint64_t zva_val; |
| 60 | asm("mrs %[zva_val], dczid_el0" : [zva_val] "=r"(zva_val)); |
| 61 | // DC ZVA is permitted if DZP, bit [4] is zero. |
| 62 | // BS, bits [3:0] is log2 of the block count in words. |
| 63 | // So the next line checks whether the instruction is permitted and block |
| 64 | // count is 16 words (i.e. 64 bytes). |
| 65 | return (zva_val & 0b11111) == 0b00100; |
| 66 | } |
| 67 | |
| 68 | } // namespace neon |
| 69 | |
| 70 | /////////////////////////////////////////////////////////////////////////////// |
| 71 | // Bcmp |
| 72 | template <size_t Size> struct Bcmp { |
| 73 | static constexpr size_t SIZE = Size; |
| 74 | static constexpr size_t BlockSize = 32; |
| 75 | |
| 76 | LIBC_INLINE static const unsigned char *as_u8(CPtr ptr) { |
| 77 | return reinterpret_cast<const unsigned char *>(ptr); |
| 78 | } |
| 79 | |
| 80 | LIBC_INLINE static BcmpReturnType block(CPtr p1, CPtr p2) { |
| 81 | if constexpr (Size == 16) { |
| 82 | auto _p1 = as_u8(p1); |
| 83 | auto _p2 = as_u8(p2); |
| 84 | uint8x16_t a = vld1q_u8(_p1); |
| 85 | uint8x16_t n = vld1q_u8(_p2); |
| 86 | uint8x16_t an = veorq_u8(a, n); |
| 87 | uint32x2_t an_reduced = vqmovn_u64(vreinterpretq_u64_u8(an)); |
| 88 | return vmaxv_u32(an_reduced); |
| 89 | } else if constexpr (Size == 32) { |
| 90 | auto _p1 = as_u8(p1); |
| 91 | auto _p2 = as_u8(p2); |
| 92 | uint8x16_t a = vld1q_u8(_p1); |
| 93 | uint8x16_t b = vld1q_u8(_p1 + 16); |
| 94 | uint8x16_t n = vld1q_u8(_p2); |
| 95 | uint8x16_t o = vld1q_u8(_p2 + 16); |
| 96 | uint8x16_t an = veorq_u8(a, n); |
| 97 | uint8x16_t bo = veorq_u8(b, o); |
| 98 | // anbo = (a ^ n) | (b ^ o). At least one byte is nonzero if there is |
| 99 | // a difference between the two buffers. We reduce this value down to 4 |
| 100 | // bytes in two steps. First, calculate the saturated move value when |
| 101 | // going from 2x64b to 2x32b. Second, compute the max of the 2x32b to get |
| 102 | // a single 32 bit nonzero value if a mismatch occurred. |
| 103 | uint8x16_t anbo = vorrq_u8(an, bo); |
| 104 | uint32x2_t anbo_reduced = vqmovn_u64(vreinterpretq_u64_u8(anbo)); |
| 105 | return vmaxv_u32(anbo_reduced); |
| 106 | } else if constexpr ((Size % BlockSize) == 0) { |
| 107 | for (size_t offset = 0; offset < Size; offset += BlockSize) |
| 108 | if (auto value = Bcmp<BlockSize>::block(p1 + offset, p2 + offset)) |
| 109 | return value; |
| 110 | } else { |
| 111 | static_assert(cpp::always_false<decltype(Size)>, "SIZE not implemented"); |
| 112 | } |
| 113 | return BcmpReturnType::zero(); |
| 114 | } |
| 115 | |
| 116 | LIBC_INLINE static BcmpReturnType tail(CPtr p1, CPtr p2, size_t count) { |
| 117 | return block(p1 + count - SIZE, p2 + count - SIZE); |
| 118 | } |
| 119 | |
| 120 | LIBC_INLINE static BcmpReturnType head_tail(CPtr p1, CPtr p2, size_t count) { |
| 121 | if constexpr (Size == 16) { |
| 122 | auto _p1 = as_u8(p1); |
| 123 | auto _p2 = as_u8(p2); |
| 124 | uint8x16_t a = vld1q_u8(_p1); |
| 125 | uint8x16_t b = vld1q_u8(_p1 + count - 16); |
| 126 | uint8x16_t n = vld1q_u8(_p2); |
| 127 | uint8x16_t o = vld1q_u8(_p2 + count - 16); |
| 128 | uint8x16_t an = veorq_u8(a, n); |
| 129 | uint8x16_t bo = veorq_u8(b, o); |
| 130 | // anbo = (a ^ n) | (b ^ o) |
| 131 | uint8x16_t anbo = vorrq_u8(an, bo); |
| 132 | uint32x2_t anbo_reduced = vqmovn_u64(vreinterpretq_u64_u8(anbo)); |
| 133 | return vmaxv_u32(anbo_reduced); |
| 134 | } else if constexpr (Size == 32) { |
| 135 | auto _p1 = as_u8(p1); |
| 136 | auto _p2 = as_u8(p2); |
| 137 | uint8x16_t a = vld1q_u8(_p1); |
| 138 | uint8x16_t b = vld1q_u8(_p1 + 16); |
| 139 | uint8x16_t c = vld1q_u8(_p1 + count - 16); |
| 140 | uint8x16_t d = vld1q_u8(_p1 + count - 32); |
| 141 | uint8x16_t n = vld1q_u8(_p2); |
| 142 | uint8x16_t o = vld1q_u8(_p2 + 16); |
| 143 | uint8x16_t p = vld1q_u8(_p2 + count - 16); |
| 144 | uint8x16_t q = vld1q_u8(_p2 + count - 32); |
| 145 | uint8x16_t an = veorq_u8(a, n); |
| 146 | uint8x16_t bo = veorq_u8(b, o); |
| 147 | uint8x16_t cp = veorq_u8(c, p); |
| 148 | uint8x16_t dq = veorq_u8(d, q); |
| 149 | uint8x16_t anbo = vorrq_u8(an, bo); |
| 150 | uint8x16_t cpdq = vorrq_u8(cp, dq); |
| 151 | // abnocpdq = ((a ^ n) | (b ^ o)) | ((c ^ p) | (d ^ q)). Reduce this to |
| 152 | // a nonzero 32 bit value if a mismatch occurred. |
| 153 | uint64x2_t abnocpdq = vreinterpretq_u64_u8(anbo | cpdq); |
| 154 | uint32x2_t abnocpdq_reduced = vqmovn_u64(abnocpdq); |
| 155 | return vmaxv_u32(abnocpdq_reduced); |
| 156 | } else { |
| 157 | static_assert(cpp::always_false<decltype(Size)>, "SIZE not implemented"); |
| 158 | } |
| 159 | return BcmpReturnType::zero(); |
| 160 | } |
| 161 | |
| 162 | LIBC_INLINE static BcmpReturnType loop_and_tail(CPtr p1, CPtr p2, |
| 163 | size_t count) { |
| 164 | static_assert(Size > 1, "a loop of size 1 does not need tail"); |
| 165 | size_t offset = 0; |
| 166 | do { |
| 167 | if (auto value = block(p1 + offset, p2 + offset)) |
| 168 | return value; |
| 169 | offset += SIZE; |
| 170 | } while (offset < count - SIZE); |
| 171 | return tail(p1, p2, count); |
| 172 | } |
| 173 | }; |
| 174 | |
| 175 | } // namespace aarch64 |
| 176 | } // namespace LIBC_NAMESPACE_DECL |
| 177 | |
| 178 | #endif //__ARM_NEON |
| 179 | |
| 180 | namespace LIBC_NAMESPACE_DECL { |
| 181 | namespace generic { |
| 182 | |
| 183 | /////////////////////////////////////////////////////////////////////////////// |
| 184 | // Specializations for uint16_t |
| 185 | template <> struct cmp_is_expensive<uint16_t> : public cpp::false_type {}; |
| 186 | template <> LIBC_INLINE bool eq<uint16_t>(CPtr p1, CPtr p2, size_t offset) { |
| 187 | return load<uint16_t>(p1, offset) == load<uint16_t>(p2, offset); |
| 188 | } |
| 189 | template <> |
| 190 | LIBC_INLINE uint32_t neq<uint16_t>(CPtr p1, CPtr p2, size_t offset) { |
| 191 | return load<uint16_t>(p1, offset) ^ load<uint16_t>(p2, offset); |
| 192 | } |
| 193 | template <> |
| 194 | LIBC_INLINE MemcmpReturnType cmp<uint16_t>(CPtr p1, CPtr p2, size_t offset) { |
| 195 | return static_cast<int32_t>(load_be<uint16_t>(p1, offset)) - |
| 196 | static_cast<int32_t>(load_be<uint16_t>(p2, offset)); |
| 197 | } |
| 198 | |
| 199 | /////////////////////////////////////////////////////////////////////////////// |
| 200 | // Specializations for uint32_t |
| 201 | template <> struct cmp_is_expensive<uint32_t> : cpp::false_type {}; |
| 202 | template <> |
| 203 | LIBC_INLINE uint32_t neq<uint32_t>(CPtr p1, CPtr p2, size_t offset) { |
| 204 | return load<uint32_t>(p1, offset) ^ load<uint32_t>(p2, offset); |
| 205 | } |
| 206 | template <> |
| 207 | LIBC_INLINE MemcmpReturnType cmp<uint32_t>(CPtr p1, CPtr p2, size_t offset) { |
| 208 | const auto a = load_be<uint32_t>(p1, offset); |
| 209 | const auto b = load_be<uint32_t>(p2, offset); |
| 210 | return a > b ? 1 : a < b ? -1 : 0; |
| 211 | } |
| 212 | |
| 213 | /////////////////////////////////////////////////////////////////////////////// |
| 214 | // Specializations for uint64_t |
| 215 | template <> struct cmp_is_expensive<uint64_t> : cpp::false_type {}; |
| 216 | template <> |
| 217 | LIBC_INLINE uint32_t neq<uint64_t>(CPtr p1, CPtr p2, size_t offset) { |
| 218 | return load<uint64_t>(p1, offset) != load<uint64_t>(p2, offset); |
| 219 | } |
| 220 | template <> |
| 221 | LIBC_INLINE MemcmpReturnType cmp<uint64_t>(CPtr p1, CPtr p2, size_t offset) { |
| 222 | const auto a = load_be<uint64_t>(p1, offset); |
| 223 | const auto b = load_be<uint64_t>(p2, offset); |
| 224 | if (a != b) |
| 225 | return a > b ? 1 : -1; |
| 226 | return MemcmpReturnType::zero(); |
| 227 | } |
| 228 | |
| 229 | #if defined(__ARM_NEON) |
| 230 | |
| 231 | /////////////////////////////////////////////////////////////////////////////// |
| 232 | // Specializations for uint8x16_t |
| 233 | template <> struct is_vector<uint8x16_t> : cpp::true_type {}; |
| 234 | template <> struct cmp_is_expensive<uint8x16_t> : cpp::false_type {}; |
| 235 | template <> |
| 236 | LIBC_INLINE uint32_t neq<uint8x16_t>(CPtr p1, CPtr p2, size_t offset) { |
| 237 | for (size_t i = 0; i < 2; ++i) { |
| 238 | auto a = load<uint64_t>(p1, offset); |
| 239 | auto b = load<uint64_t>(p2, offset); |
| 240 | uint32_t cond = a != b; |
| 241 | if (cond) |
| 242 | return cond; |
| 243 | offset += sizeof(uint64_t); |
| 244 | } |
| 245 | return 0; |
| 246 | } |
| 247 | template <> |
| 248 | LIBC_INLINE MemcmpReturnType cmp<uint8x16_t>(CPtr p1, CPtr p2, size_t offset) { |
| 249 | for (size_t i = 0; i < 2; ++i) { |
| 250 | auto a = load_be<uint64_t>(p1, offset); |
| 251 | auto b = load_be<uint64_t>(p2, offset); |
| 252 | if (a != b) |
| 253 | return cmp_neq_uint64_t(a, b); |
| 254 | offset += sizeof(uint64_t); |
| 255 | } |
| 256 | return MemcmpReturnType::zero(); |
| 257 | } |
| 258 | |
| 259 | /////////////////////////////////////////////////////////////////////////////// |
| 260 | // Specializations for uint8x16x2_t |
| 261 | template <> struct is_vector<uint8x16x2_t> : cpp::true_type {}; |
| 262 | template <> struct cmp_is_expensive<uint8x16x2_t> : cpp::false_type {}; |
| 263 | template <> |
| 264 | LIBC_INLINE MemcmpReturnType cmp<uint8x16x2_t>(CPtr p1, CPtr p2, |
| 265 | size_t offset) { |
| 266 | for (size_t i = 0; i < 4; ++i) { |
| 267 | auto a = load_be<uint64_t>(p1, offset); |
| 268 | auto b = load_be<uint64_t>(p2, offset); |
| 269 | if (a != b) |
| 270 | return cmp_neq_uint64_t(a, b); |
| 271 | offset += sizeof(uint64_t); |
| 272 | } |
| 273 | return MemcmpReturnType::zero(); |
| 274 | } |
| 275 | |
| 276 | #endif // __ARM_NEON |
| 277 | |
| 278 | } // namespace generic |
| 279 | } // namespace LIBC_NAMESPACE_DECL |
| 280 | |
| 281 | #endif // LIBC_TARGET_ARCH_IS_AARCH64 |
| 282 | |
| 283 | #endif // LLVM_LIBC_SRC_STRING_MEMORY_UTILS_OP_AARCH64_H |
| 284 |
Warning: This file is not a C or C++ file. It does not have highlighting.
