dark/zerotiertspu
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Some micro-optimizations, including a memcpy that is faster than Linux for most distro/compiler versions.

Browse source
This commit is contained in:
Adam Ierymenko 2017-12-15 11:03:20 -08:00
parent cb8e7b4d5e
commit 0608704626
22 changed files with 163 additions and 81 deletions
Download patch file Download diff file Expand all files Collapse all files

38
node/Utils.hpp
View file

@ -40,6 +40,44 @@
#include "Constants.hpp"
// So it's 2017 and this still helps on most Linux versions. It shouldn't but it does. Go figure.
#if defined(__LINUX__) && ((defined(_MSC_VER) || defined(__GNUC__)) && (defined(__amd64) || defined(__amd64__) || defined(__x86_64) || defined(__x86_64__) || defined(__AMD64) || defined(__AMD64__) || defined(_M_X64)))
#include <emmintrin.h>
static inline void ZT_FAST_MEMCPY(void *a,const void *b,unsigned long k)
{
char *aa = reinterpret_cast<char *>(a);
const char *bb = reinterpret_cast<const char *>(b);
while (likely(k >= 128)) {
__m128i t1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
__m128i t2 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
__m128i t3 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
__m128i t4 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t1); aa += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t2); aa += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t3); aa += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t4); aa += 16;
__m128i t5 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
__m128i t6 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
__m128i t7 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
__m128i t8 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t5); aa += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t6); aa += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t7); aa += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t8); aa += 16;
k -= 128;
}
while (likely(k >= 16)) {
__m128i t1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(bb)); bb += 16;
_mm_storeu_si128(reinterpret_cast<__m128i *>(aa),t1); aa += 16;
k -= 16;
}
for(unsigned long i=0;i<k;++i)
aa[i] = bb[i];
}
#else
#define ZT_FAST_MEMCPY(a,b,c) memcpy(a,b,c)
#endif
namespace ZeroTier {
/**