Martin Stjernholm | c15e7e4 | 2020-12-02 22:50:53 +0000 | [diff] [blame] | 1 | /* chunkcopy.h -- fast chunk copy and set operations |
| 2 | * Copyright (C) 2017 ARM, Inc. |
Fairphone ODM | 25c12f5 | 2023-12-15 17:24:06 +0800 | [diff] [blame] | 3 | * Copyright 2017 The Chromium Authors |
Martin Stjernholm | c15e7e4 | 2020-12-02 22:50:53 +0000 | [diff] [blame] | 4 | * Use of this source code is governed by a BSD-style license that can be |
| 5 | * found in the Chromium source repository LICENSE file. |
| 6 | */ |
| 7 | |
| 8 | #ifndef CHUNKCOPY_H |
| 9 | #define CHUNKCOPY_H |
| 10 | |
| 11 | #include <stdint.h> |
| 12 | #include "zutil.h" |
| 13 | |
| 14 | #define Z_STATIC_ASSERT(name, assert) typedef char name[(assert) ? 1 : -1] |
| 15 | |
| 16 | #if __STDC_VERSION__ >= 199901L |
| 17 | #define Z_RESTRICT restrict |
| 18 | #else |
| 19 | #define Z_RESTRICT |
| 20 | #endif |
| 21 | |
| 22 | #if defined(__clang__) || defined(__GNUC__) || defined(__llvm__) |
| 23 | #define Z_BUILTIN_MEMCPY __builtin_memcpy |
| 24 | #else |
| 25 | #define Z_BUILTIN_MEMCPY zmemcpy |
| 26 | #endif |
| 27 | |
| 28 | #if defined(INFLATE_CHUNK_SIMD_NEON) |
| 29 | #include <arm_neon.h> |
| 30 | typedef uint8x16_t z_vec128i_t; |
| 31 | #elif defined(INFLATE_CHUNK_SIMD_SSE2) |
| 32 | #include <emmintrin.h> |
| 33 | typedef __m128i z_vec128i_t; |
| 34 | #else |
| 35 | #error chunkcopy.h inflate chunk SIMD is not defined for your build target |
| 36 | #endif |
| 37 | |
| 38 | /* |
Fairphone ODM | 25c12f5 | 2023-12-15 17:24:06 +0800 | [diff] [blame] | 39 | * Suppress MSan errors about copying uninitialized bytes (crbug.com/1376033). |
| 40 | */ |
| 41 | #define Z_DISABLE_MSAN |
| 42 | #if defined(__has_feature) |
| 43 | #if __has_feature(memory_sanitizer) |
| 44 | #undef Z_DISABLE_MSAN |
| 45 | #define Z_DISABLE_MSAN __attribute__((no_sanitize("memory"))) |
| 46 | #endif |
| 47 | #endif |
| 48 | |
| 49 | /* |
Martin Stjernholm | c15e7e4 | 2020-12-02 22:50:53 +0000 | [diff] [blame] | 50 | * chunk copy type: the z_vec128i_t type size should be exactly 128-bits |
| 51 | * and equal to CHUNKCOPY_CHUNK_SIZE. |
| 52 | */ |
| 53 | #define CHUNKCOPY_CHUNK_SIZE sizeof(z_vec128i_t) |
| 54 | |
| 55 | Z_STATIC_ASSERT(vector_128_bits_wide, |
| 56 | CHUNKCOPY_CHUNK_SIZE == sizeof(int8_t) * 16); |
| 57 | |
| 58 | /* |
| 59 | * Ask the compiler to perform a wide, unaligned load with a machine |
| 60 | * instruction appropriate for the z_vec128i_t type. |
| 61 | */ |
| 62 | static inline z_vec128i_t loadchunk( |
Fairphone ODM | 25c12f5 | 2023-12-15 17:24:06 +0800 | [diff] [blame] | 63 | const unsigned char FAR* s) Z_DISABLE_MSAN { |
Martin Stjernholm | c15e7e4 | 2020-12-02 22:50:53 +0000 | [diff] [blame] | 64 | z_vec128i_t v; |
| 65 | Z_BUILTIN_MEMCPY(&v, s, sizeof(v)); |
| 66 | return v; |
| 67 | } |
| 68 | |
| 69 | /* |
| 70 | * Ask the compiler to perform a wide, unaligned store with a machine |
| 71 | * instruction appropriate for the z_vec128i_t type. |
| 72 | */ |
| 73 | static inline void storechunk( |
| 74 | unsigned char FAR* d, |
| 75 | const z_vec128i_t v) { |
| 76 | Z_BUILTIN_MEMCPY(d, &v, sizeof(v)); |
| 77 | } |
| 78 | |
| 79 | /* |
| 80 | * Perform a memcpy-like operation, assuming that length is non-zero and that |
| 81 | * it's OK to overwrite at least CHUNKCOPY_CHUNK_SIZE bytes of output even if |
| 82 | * the length is shorter than this. |
| 83 | * |
| 84 | * It also guarantees that it will properly unroll the data if the distance |
| 85 | * between `out` and `from` is at least CHUNKCOPY_CHUNK_SIZE, which we rely on |
| 86 | * in chunkcopy_relaxed(). |
| 87 | * |
| 88 | * Aside from better memory bus utilisation, this means that short copies |
| 89 | * (CHUNKCOPY_CHUNK_SIZE bytes or fewer) will fall straight through the loop |
| 90 | * without iteration, which will hopefully make the branch prediction more |
| 91 | * reliable. |
| 92 | */ |
| 93 | static inline unsigned char FAR* chunkcopy_core( |
| 94 | unsigned char FAR* out, |
| 95 | const unsigned char FAR* from, |
Fairphone ODM | 25c12f5 | 2023-12-15 17:24:06 +0800 | [diff] [blame] | 96 | unsigned len) Z_DISABLE_MSAN { |
Martin Stjernholm | c15e7e4 | 2020-12-02 22:50:53 +0000 | [diff] [blame] | 97 | const int bump = (--len % CHUNKCOPY_CHUNK_SIZE) + 1; |
| 98 | storechunk(out, loadchunk(from)); |
| 99 | out += bump; |
| 100 | from += bump; |
| 101 | len /= CHUNKCOPY_CHUNK_SIZE; |
| 102 | while (len-- > 0) { |
| 103 | storechunk(out, loadchunk(from)); |
| 104 | out += CHUNKCOPY_CHUNK_SIZE; |
| 105 | from += CHUNKCOPY_CHUNK_SIZE; |
| 106 | } |
| 107 | return out; |
| 108 | } |
| 109 | |
| 110 | /* |
| 111 | * Like chunkcopy_core(), but avoid writing beyond of legal output. |
| 112 | * |
| 113 | * Accepts an additional pointer to the end of safe output. A generic safe |
| 114 | * copy would use (out + len), but it's normally the case that the end of the |
| 115 | * output buffer is beyond the end of the current copy, and this can still be |
| 116 | * exploited. |
| 117 | */ |
| 118 | static inline unsigned char FAR* chunkcopy_core_safe( |
| 119 | unsigned char FAR* out, |
| 120 | const unsigned char FAR* from, |
| 121 | unsigned len, |
| 122 | unsigned char FAR* limit) { |
| 123 | Assert(out + len <= limit, "chunk copy exceeds safety limit"); |
| 124 | if ((limit - out) < (ptrdiff_t)CHUNKCOPY_CHUNK_SIZE) { |
| 125 | const unsigned char FAR* Z_RESTRICT rfrom = from; |
| 126 | Assert((uintptr_t)out - (uintptr_t)from >= len, |
| 127 | "invalid restrict in chunkcopy_core_safe"); |
| 128 | Assert((uintptr_t)from - (uintptr_t)out >= len, |
| 129 | "invalid restrict in chunkcopy_core_safe"); |
| 130 | if (len & 8) { |
| 131 | Z_BUILTIN_MEMCPY(out, rfrom, 8); |
| 132 | out += 8; |
| 133 | rfrom += 8; |
| 134 | } |
| 135 | if (len & 4) { |
| 136 | Z_BUILTIN_MEMCPY(out, rfrom, 4); |
| 137 | out += 4; |
| 138 | rfrom += 4; |
| 139 | } |
| 140 | if (len & 2) { |
| 141 | Z_BUILTIN_MEMCPY(out, rfrom, 2); |
| 142 | out += 2; |
| 143 | rfrom += 2; |
| 144 | } |
| 145 | if (len & 1) { |
| 146 | *out++ = *rfrom++; |
| 147 | } |
| 148 | return out; |
| 149 | } |
| 150 | return chunkcopy_core(out, from, len); |
| 151 | } |
| 152 | |
| 153 | /* |
| 154 | * Perform short copies until distance can be rewritten as being at least |
| 155 | * CHUNKCOPY_CHUNK_SIZE. |
| 156 | * |
| 157 | * Assumes it's OK to overwrite at least the first 2*CHUNKCOPY_CHUNK_SIZE |
| 158 | * bytes of output even if the copy is shorter than this. This assumption |
| 159 | * holds within zlib inflate_fast(), which starts every iteration with at |
| 160 | * least 258 bytes of output space available (258 being the maximum length |
| 161 | * output from a single token; see inffast.c). |
| 162 | */ |
| 163 | static inline unsigned char FAR* chunkunroll_relaxed( |
| 164 | unsigned char FAR* out, |
| 165 | unsigned FAR* dist, |
Fairphone ODM | 25c12f5 | 2023-12-15 17:24:06 +0800 | [diff] [blame] | 166 | unsigned FAR* len) Z_DISABLE_MSAN { |
Martin Stjernholm | c15e7e4 | 2020-12-02 22:50:53 +0000 | [diff] [blame] | 167 | const unsigned char FAR* from = out - *dist; |
| 168 | while (*dist < *len && *dist < CHUNKCOPY_CHUNK_SIZE) { |
| 169 | storechunk(out, loadchunk(from)); |
| 170 | out += *dist; |
| 171 | *len -= *dist; |
| 172 | *dist += *dist; |
| 173 | } |
| 174 | return out; |
| 175 | } |
| 176 | |
| 177 | #if defined(INFLATE_CHUNK_SIMD_NEON) |
| 178 | /* |
| 179 | * v_load64_dup(): load *src as an unaligned 64-bit int and duplicate it in |
| 180 | * every 64-bit component of the 128-bit result (64-bit int splat). |
| 181 | */ |
| 182 | static inline z_vec128i_t v_load64_dup(const void* src) { |
| 183 | return vcombine_u8(vld1_u8(src), vld1_u8(src)); |
| 184 | } |
| 185 | |
| 186 | /* |
| 187 | * v_load32_dup(): load *src as an unaligned 32-bit int and duplicate it in |
| 188 | * every 32-bit component of the 128-bit result (32-bit int splat). |
| 189 | */ |
| 190 | static inline z_vec128i_t v_load32_dup(const void* src) { |
| 191 | int32_t i32; |
| 192 | Z_BUILTIN_MEMCPY(&i32, src, sizeof(i32)); |
| 193 | return vreinterpretq_u8_s32(vdupq_n_s32(i32)); |
| 194 | } |
| 195 | |
| 196 | /* |
| 197 | * v_load16_dup(): load *src as an unaligned 16-bit int and duplicate it in |
| 198 | * every 16-bit component of the 128-bit result (16-bit int splat). |
| 199 | */ |
| 200 | static inline z_vec128i_t v_load16_dup(const void* src) { |
| 201 | int16_t i16; |
| 202 | Z_BUILTIN_MEMCPY(&i16, src, sizeof(i16)); |
| 203 | return vreinterpretq_u8_s16(vdupq_n_s16(i16)); |
| 204 | } |
| 205 | |
| 206 | /* |
| 207 | * v_load8_dup(): load the 8-bit int *src and duplicate it in every 8-bit |
| 208 | * component of the 128-bit result (8-bit int splat). |
| 209 | */ |
| 210 | static inline z_vec128i_t v_load8_dup(const void* src) { |
| 211 | return vld1q_dup_u8((const uint8_t*)src); |
| 212 | } |
| 213 | |
| 214 | /* |
| 215 | * v_store_128(): store the 128-bit vec in a memory destination (that might |
| 216 | * not be 16-byte aligned) void* out. |
| 217 | */ |
| 218 | static inline void v_store_128(void* out, const z_vec128i_t vec) { |
| 219 | vst1q_u8(out, vec); |
| 220 | } |
| 221 | |
| 222 | #elif defined(INFLATE_CHUNK_SIMD_SSE2) |
| 223 | /* |
| 224 | * v_load64_dup(): load *src as an unaligned 64-bit int and duplicate it in |
| 225 | * every 64-bit component of the 128-bit result (64-bit int splat). |
| 226 | */ |
| 227 | static inline z_vec128i_t v_load64_dup(const void* src) { |
| 228 | int64_t i64; |
| 229 | Z_BUILTIN_MEMCPY(&i64, src, sizeof(i64)); |
| 230 | return _mm_set1_epi64x(i64); |
| 231 | } |
| 232 | |
| 233 | /* |
| 234 | * v_load32_dup(): load *src as an unaligned 32-bit int and duplicate it in |
| 235 | * every 32-bit component of the 128-bit result (32-bit int splat). |
| 236 | */ |
| 237 | static inline z_vec128i_t v_load32_dup(const void* src) { |
| 238 | int32_t i32; |
| 239 | Z_BUILTIN_MEMCPY(&i32, src, sizeof(i32)); |
| 240 | return _mm_set1_epi32(i32); |
| 241 | } |
| 242 | |
| 243 | /* |
| 244 | * v_load16_dup(): load *src as an unaligned 16-bit int and duplicate it in |
| 245 | * every 16-bit component of the 128-bit result (16-bit int splat). |
| 246 | */ |
| 247 | static inline z_vec128i_t v_load16_dup(const void* src) { |
| 248 | int16_t i16; |
| 249 | Z_BUILTIN_MEMCPY(&i16, src, sizeof(i16)); |
| 250 | return _mm_set1_epi16(i16); |
| 251 | } |
| 252 | |
| 253 | /* |
| 254 | * v_load8_dup(): load the 8-bit int *src and duplicate it in every 8-bit |
| 255 | * component of the 128-bit result (8-bit int splat). |
| 256 | */ |
| 257 | static inline z_vec128i_t v_load8_dup(const void* src) { |
| 258 | return _mm_set1_epi8(*(const char*)src); |
| 259 | } |
| 260 | |
| 261 | /* |
| 262 | * v_store_128(): store the 128-bit vec in a memory destination (that might |
| 263 | * not be 16-byte aligned) void* out. |
| 264 | */ |
| 265 | static inline void v_store_128(void* out, const z_vec128i_t vec) { |
| 266 | _mm_storeu_si128((__m128i*)out, vec); |
| 267 | } |
| 268 | #endif |
| 269 | |
| 270 | /* |
| 271 | * Perform an overlapping copy which behaves as a memset() operation, but |
| 272 | * supporting periods other than one, and assume that length is non-zero and |
| 273 | * that it's OK to overwrite at least CHUNKCOPY_CHUNK_SIZE*3 bytes of output |
| 274 | * even if the length is shorter than this. |
| 275 | */ |
| 276 | static inline unsigned char FAR* chunkset_core( |
| 277 | unsigned char FAR* out, |
| 278 | unsigned period, |
| 279 | unsigned len) { |
| 280 | z_vec128i_t v; |
| 281 | const int bump = ((len - 1) % sizeof(v)) + 1; |
| 282 | |
| 283 | switch (period) { |
| 284 | case 1: |
| 285 | v = v_load8_dup(out - 1); |
| 286 | v_store_128(out, v); |
| 287 | out += bump; |
| 288 | len -= bump; |
| 289 | while (len > 0) { |
| 290 | v_store_128(out, v); |
| 291 | out += sizeof(v); |
| 292 | len -= sizeof(v); |
| 293 | } |
| 294 | return out; |
| 295 | case 2: |
| 296 | v = v_load16_dup(out - 2); |
| 297 | v_store_128(out, v); |
| 298 | out += bump; |
| 299 | len -= bump; |
| 300 | if (len > 0) { |
| 301 | v = v_load16_dup(out - 2); |
| 302 | do { |
| 303 | v_store_128(out, v); |
| 304 | out += sizeof(v); |
| 305 | len -= sizeof(v); |
| 306 | } while (len > 0); |
| 307 | } |
| 308 | return out; |
| 309 | case 4: |
| 310 | v = v_load32_dup(out - 4); |
| 311 | v_store_128(out, v); |
| 312 | out += bump; |
| 313 | len -= bump; |
| 314 | if (len > 0) { |
| 315 | v = v_load32_dup(out - 4); |
| 316 | do { |
| 317 | v_store_128(out, v); |
| 318 | out += sizeof(v); |
| 319 | len -= sizeof(v); |
| 320 | } while (len > 0); |
| 321 | } |
| 322 | return out; |
| 323 | case 8: |
| 324 | v = v_load64_dup(out - 8); |
| 325 | v_store_128(out, v); |
| 326 | out += bump; |
| 327 | len -= bump; |
| 328 | if (len > 0) { |
| 329 | v = v_load64_dup(out - 8); |
| 330 | do { |
| 331 | v_store_128(out, v); |
| 332 | out += sizeof(v); |
| 333 | len -= sizeof(v); |
| 334 | } while (len > 0); |
| 335 | } |
| 336 | return out; |
| 337 | } |
| 338 | out = chunkunroll_relaxed(out, &period, &len); |
| 339 | return chunkcopy_core(out, out - period, len); |
| 340 | } |
| 341 | |
| 342 | /* |
| 343 | * Perform a memcpy-like operation, but assume that length is non-zero and that |
| 344 | * it's OK to overwrite at least CHUNKCOPY_CHUNK_SIZE bytes of output even if |
| 345 | * the length is shorter than this. |
| 346 | * |
| 347 | * Unlike chunkcopy_core() above, no guarantee is made regarding the behaviour |
| 348 | * of overlapping buffers, regardless of the distance between the pointers. |
| 349 | * This is reflected in the `restrict`-qualified pointers, allowing the |
| 350 | * compiler to re-order loads and stores. |
| 351 | */ |
| 352 | static inline unsigned char FAR* chunkcopy_relaxed( |
| 353 | unsigned char FAR* Z_RESTRICT out, |
| 354 | const unsigned char FAR* Z_RESTRICT from, |
| 355 | unsigned len) { |
| 356 | Assert((uintptr_t)out - (uintptr_t)from >= len, |
| 357 | "invalid restrict in chunkcopy_relaxed"); |
| 358 | Assert((uintptr_t)from - (uintptr_t)out >= len, |
| 359 | "invalid restrict in chunkcopy_relaxed"); |
| 360 | return chunkcopy_core(out, from, len); |
| 361 | } |
| 362 | |
| 363 | /* |
| 364 | * Like chunkcopy_relaxed(), but avoid writing beyond of legal output. |
| 365 | * |
| 366 | * Unlike chunkcopy_core_safe() above, no guarantee is made regarding the |
| 367 | * behaviour of overlapping buffers, regardless of the distance between the |
| 368 | * pointers. This is reflected in the `restrict`-qualified pointers, allowing |
| 369 | * the compiler to re-order loads and stores. |
| 370 | * |
| 371 | * Accepts an additional pointer to the end of safe output. A generic safe |
| 372 | * copy would use (out + len), but it's normally the case that the end of the |
| 373 | * output buffer is beyond the end of the current copy, and this can still be |
| 374 | * exploited. |
| 375 | */ |
| 376 | static inline unsigned char FAR* chunkcopy_safe( |
| 377 | unsigned char FAR* out, |
| 378 | const unsigned char FAR* Z_RESTRICT from, |
| 379 | unsigned len, |
| 380 | unsigned char FAR* limit) { |
| 381 | Assert(out + len <= limit, "chunk copy exceeds safety limit"); |
| 382 | Assert((uintptr_t)out - (uintptr_t)from >= len, |
| 383 | "invalid restrict in chunkcopy_safe"); |
| 384 | Assert((uintptr_t)from - (uintptr_t)out >= len, |
| 385 | "invalid restrict in chunkcopy_safe"); |
| 386 | |
| 387 | return chunkcopy_core_safe(out, from, len, limit); |
| 388 | } |
| 389 | |
| 390 | /* |
| 391 | * Perform chunky copy within the same buffer, where the source and destination |
| 392 | * may potentially overlap. |
| 393 | * |
| 394 | * Assumes that len > 0 on entry, and that it's safe to write at least |
| 395 | * CHUNKCOPY_CHUNK_SIZE*3 bytes to the output. |
| 396 | */ |
| 397 | static inline unsigned char FAR* chunkcopy_lapped_relaxed( |
| 398 | unsigned char FAR* out, |
| 399 | unsigned dist, |
| 400 | unsigned len) { |
| 401 | if (dist < len && dist < CHUNKCOPY_CHUNK_SIZE) { |
| 402 | return chunkset_core(out, dist, len); |
| 403 | } |
| 404 | return chunkcopy_core(out, out - dist, len); |
| 405 | } |
| 406 | |
| 407 | /* |
| 408 | * Behave like chunkcopy_lapped_relaxed(), but avoid writing beyond of legal |
| 409 | * output. |
| 410 | * |
| 411 | * Accepts an additional pointer to the end of safe output. A generic safe |
| 412 | * copy would use (out + len), but it's normally the case that the end of the |
| 413 | * output buffer is beyond the end of the current copy, and this can still be |
| 414 | * exploited. |
| 415 | */ |
| 416 | static inline unsigned char FAR* chunkcopy_lapped_safe( |
| 417 | unsigned char FAR* out, |
| 418 | unsigned dist, |
| 419 | unsigned len, |
| 420 | unsigned char FAR* limit) { |
| 421 | Assert(out + len <= limit, "chunk copy exceeds safety limit"); |
| 422 | if ((limit - out) < (ptrdiff_t)(3 * CHUNKCOPY_CHUNK_SIZE)) { |
| 423 | /* TODO(cavalcantii): try harder to optimise this */ |
| 424 | while (len-- > 0) { |
| 425 | *out = *(out - dist); |
| 426 | out++; |
| 427 | } |
| 428 | return out; |
| 429 | } |
| 430 | return chunkcopy_lapped_relaxed(out, dist, len); |
| 431 | } |
| 432 | |
| 433 | /* TODO(cavalcanti): see crbug.com/1110083. */ |
| 434 | static inline unsigned char FAR* chunkcopy_safe_ugly(unsigned char FAR* out, |
| 435 | unsigned dist, |
| 436 | unsigned len, |
| 437 | unsigned char FAR* limit) { |
| 438 | #if defined(__GNUC__) && !defined(__clang__) |
| 439 | /* Speed is the same as using chunkcopy_safe |
| 440 | w/ GCC on ARM (tested gcc 6.3 and 7.5) and avoids |
| 441 | undefined behavior. |
| 442 | */ |
| 443 | return chunkcopy_core_safe(out, out - dist, len, limit); |
| 444 | #elif defined(__clang__) && defined(ARMV8_OS_ANDROID) && !defined(__aarch64__) |
| 445 | /* Seems to perform better on 32bit (i.e. Android). */ |
| 446 | return chunkcopy_core_safe(out, out - dist, len, limit); |
| 447 | #else |
| 448 | /* Seems to perform better on 64bit. */ |
| 449 | return chunkcopy_lapped_safe(out, dist, len, limit); |
| 450 | #endif |
| 451 | } |
| 452 | |
| 453 | /* |
| 454 | * The chunk-copy code above deals with writing the decoded DEFLATE data to |
| 455 | * the output with SIMD methods to increase decode speed. Reading the input |
| 456 | * to the DEFLATE decoder with a wide, SIMD method can also increase decode |
| 457 | * speed. This option is supported on little endian machines, and reads the |
| 458 | * input data in 64-bit (8 byte) chunks. |
| 459 | */ |
| 460 | |
| 461 | #ifdef INFLATE_CHUNK_READ_64LE |
| 462 | /* |
| 463 | * Buffer the input in a uint64_t (8 bytes) in the wide input reading case. |
| 464 | */ |
| 465 | typedef uint64_t inflate_holder_t; |
| 466 | |
| 467 | /* |
| 468 | * Ask the compiler to perform a wide, unaligned load of a uint64_t using a |
| 469 | * machine instruction appropriate for the uint64_t type. |
| 470 | */ |
| 471 | static inline inflate_holder_t read64le(const unsigned char FAR *in) { |
| 472 | inflate_holder_t input; |
| 473 | Z_BUILTIN_MEMCPY(&input, in, sizeof(input)); |
| 474 | return input; |
| 475 | } |
| 476 | #else |
| 477 | /* |
| 478 | * Otherwise, buffer the input bits using zlib's default input buffer type. |
| 479 | */ |
| 480 | typedef unsigned long inflate_holder_t; |
| 481 | |
| 482 | #endif /* INFLATE_CHUNK_READ_64LE */ |
| 483 | |
| 484 | #undef Z_STATIC_ASSERT |
| 485 | #undef Z_RESTRICT |
| 486 | #undef Z_BUILTIN_MEMCPY |
Fairphone ODM | 25c12f5 | 2023-12-15 17:24:06 +0800 | [diff] [blame] | 487 | #undef Z_DISABLE_MSAN |
Martin Stjernholm | c15e7e4 | 2020-12-02 22:50:53 +0000 | [diff] [blame] | 488 | |
| 489 | #endif /* CHUNKCOPY_H */ |