store-buffers.cpp

void store_buffers() {
	static constexpr size_t items = 1ULL << 24;
	static constexpr size_t mask = items - 1;
	static constexpr size_t iterations = 1'000'000'000;
	std::vector<std::vector<unsigned char> > arrays{ 12, std::vector<unsigned char>(items, '\0') };
	::SetThreadPriority(::GetCurrentThread(), THREAD_PRIORITY_HIGHEST);

	__int32 unused[4];
	unsigned __int32 also_unused;
	{
		__cpuidex(unused, 0, 0);
		unsigned __int64 start = __rdtsc();

		size_t i = iterations;
		while(--i) {
			size_t slot = i & mask;
			unsigned char value = static_cast<unsigned char>(i & 0xff);
			arrays[0][slot] = value;
			arrays[1][slot] = value;
			arrays[2][slot] = value;
			arrays[3][slot] = value;
			arrays[4][slot] = value;
			arrays[5][slot] = value;
			arrays[6][slot] = value;
			arrays[7][slot] = value;
			arrays[8][slot] = value;
			arrays[9][slot] = value;
			arrays[10][slot] = value;
			arrays[11][slot] = value;
		}

		unsigned __int64 end = __rdtscp(&also_unused);
		__cpuidex(unused, 0, 0);

		std::cout << "combined: " << end - start << std::endl;
	}

	{
		__cpuidex(unused, 0, 0);
		unsigned __int64 start = __rdtsc();

		size_t i = iterations;
		while(--i) {
			size_t slot = i & mask;
			unsigned char value = static_cast<unsigned char>(i & 0xff);
			arrays[0][slot] = value;
			arrays[1][slot] = value;
			arrays[2][slot] = value;
			arrays[3][slot] = value;
			arrays[4][slot] = value;
			arrays[5][slot] = value;
		}

		i = iterations;
		while(--i) {
			size_t slot = i & mask;
			unsigned char value = static_cast<unsigned char>(i & 0xff);
			arrays[6][slot] = value;
			arrays[7][slot] = value;
			arrays[8][slot] = value;
			arrays[9][slot] = value;
			arrays[10][slot] = value;
			arrays[11][slot] = value;
		}

		unsigned __int64 end = __rdtscp(&also_unused);
		__cpuidex(unused, 0, 0);

		std::cout << "split:    " << end - start << std::endl;
	}
	::SetThreadPriority(::GetCurrentThread(), THREAD_PRIORITY_NORMAL);
}