| 1 | #pragma once | 
|---|
| 2 |  | 
|---|
| 3 | #include <string.h> | 
|---|
| 4 |  | 
|---|
| 5 | #ifdef __SSE2__ | 
|---|
| 6 | #include <emmintrin.h> | 
|---|
| 7 |  | 
|---|
| 8 |  | 
|---|
| 9 | /** memcpy function could work suboptimal if all the following conditions are met: | 
|---|
| 10 | * 1. Size of memory region is relatively small (approximately, under 50 bytes). | 
|---|
| 11 | * 2. Size of memory region is not known at compile-time. | 
|---|
| 12 | * | 
|---|
| 13 | * In that case, memcpy works suboptimal by following reasons: | 
|---|
| 14 | * 1. Function is not inlined. | 
|---|
| 15 | * 2. Much time/instructions are spend to process "tails" of data. | 
|---|
| 16 | * | 
|---|
| 17 | * There are cases when function could be implemented in more optimal way, with help of some assumptions. | 
|---|
| 18 | * One of that assumptions - ability to read and write some number of bytes after end of passed memory regions. | 
|---|
| 19 | * Under that assumption, it is possible not to implement difficult code to process tails of data and do copy always by big chunks. | 
|---|
| 20 | * | 
|---|
| 21 | * This case is typical, for example, when many small pieces of data are gathered to single contiguous piece of memory in a loop. | 
|---|
| 22 | * - because each next copy will overwrite excessive data after previous copy. | 
|---|
| 23 | * | 
|---|
| 24 | * Assumption that size of memory region is small enough allows us to not unroll the loop. | 
|---|
| 25 | * This is slower, when size of memory is actually big. | 
|---|
| 26 | * | 
|---|
| 27 | * Use with caution. | 
|---|
| 28 | */ | 
|---|
| 29 |  | 
|---|
| 30 | namespace detail | 
|---|
| 31 | { | 
|---|
| 32 | inline void memcpySmallAllowReadWriteOverflow15Impl(char * __restrict dst, const char * __restrict src, ssize_t n) | 
|---|
| 33 | { | 
|---|
| 34 | while (n > 0) | 
|---|
| 35 | { | 
|---|
| 36 | _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), | 
|---|
| 37 | _mm_loadu_si128(reinterpret_cast<const __m128i *>(src))); | 
|---|
| 38 |  | 
|---|
| 39 | dst += 16; | 
|---|
| 40 | src += 16; | 
|---|
| 41 | n -= 16; | 
|---|
| 42 | } | 
|---|
| 43 | } | 
|---|
| 44 | } | 
|---|
| 45 |  | 
|---|
| 46 | /** Works under assumption, that it's possible to read up to 15 excessive bytes after end of 'src' region | 
|---|
| 47 | *  and to write any garbage into up to 15 bytes after end of 'dst' region. | 
|---|
| 48 | */ | 
|---|
| 49 | inline void memcpySmallAllowReadWriteOverflow15(void * __restrict dst, const void * __restrict src, size_t n) | 
|---|
| 50 | { | 
|---|
| 51 | detail::memcpySmallAllowReadWriteOverflow15Impl(reinterpret_cast<char *>(dst), reinterpret_cast<const char *>(src), n); | 
|---|
| 52 | } | 
|---|
| 53 |  | 
|---|
| 54 | /** NOTE There was also a function, that assumes, that you could read any bytes inside same memory page of src. | 
|---|
| 55 | * This function was unused, and also it requires special handling for Valgrind and ASan. | 
|---|
| 56 | */ | 
|---|
| 57 |  | 
|---|
| 58 | #else    /// Implementation for other platforms. | 
|---|
| 59 |  | 
|---|
| 60 | inline void memcpySmallAllowReadWriteOverflow15(void * __restrict dst, const void * __restrict src, size_t n) | 
|---|
| 61 | { | 
|---|
| 62 | memcpy(dst, src, n); | 
|---|
| 63 | } | 
|---|
| 64 |  | 
|---|
| 65 | #endif | 
|---|
| 66 |  | 
|---|