Fun with pclmulqdq(carryless multiply)

pclmulqdq.cpp

#include <cstdint>
#include <cstdio>
#include <bitset>
#include <immintrin.h>

int main()
{
	const std::uint64_t Bits
		= 0b00011000100000000001000000000010000000001000000100000001000000001;
	std::puts(("Bits:\t"+std::bitset<64>(Bits).to_string()).c_str());
	const __m128i Result = _mm_clmulepi64_si128(
		_mm_set_epi64x(0, Bits),
		_mm_set1_epi8(0xFF),
		0
	);
	const std::uint64_t Lo = _mm_cvtsi128_si64(Result);
	const std::uint64_t Hi = _mm_extract_epi64(Result,1);
	std::puts(("Lo:\t"+std::bitset<64>(Lo).to_string()).c_str());
	std::puts(("Hi:\t"+std::bitset<64>(Hi).to_string()).c_str());

	// Bit mask odds(mask odd set bits)
	std::puts(("BMO:\t"+std::bitset<64>(~(Hi) & Bits).to_string()).c_str());
	// Bit mask evens(mask even set bits)
	std::puts(("BME:\t"+std::bitset<64>(~(Lo) & Bits).to_string()).c_str());

	// Bit scan odd parity ranges(exclusive) 
	std::puts(("BSOP:\t"+std::bitset<64>((~Bits) & Lo).to_string()).c_str());
	// Bit scan even parity ranges(exclusive) 
	std::puts(("BSEP:\t"+std::bitset<64>((~Bits) & Hi).to_string()).c_str());
}

// Bits:   0011000100000000001000000000010000000001000000100000001000000001
// Lo:     1110111100000000000111111111110000000000111111100000000111111111
// Hi:     0001000011111111111000000000001111111111000000011111111000000000
// BMO:    0010000100000000000000000000010000000000000000100000000000000001
// BME:    0001000000000000001000000000000000000001000000000000001000000000
// BSOP:   1100111000000000000111111111100000000000111111000000000111111110
// BSEP:   0000000011111111110000000000001111111110000000011111110000000000