shiftLeftBytes_mem.cpp

inline __m256i shiftLeftBytes_mem( const __m256i src, int i )
{
	assert( i >= 0 && i < 32 );
	// Align by 64 bytes so the complete array stays in a single cache line
	alignas( 64 ) std::array<uint8_t, 64> buffer;
	// Store zeros at offset 0
	_mm256_store_si256( ( __m256i* )buffer.data(), _mm256_setzero_si256() );
	// Store the source vector at offset 32
	_mm256_store_si256( ( __m256i* )( buffer.data() + 32 ), src );
	// Load back with the offset
	return _mm256_loadu_si256( ( const __m256i* )( buffer.data() + 32 - i ) );
}