Overengineered bit moving (right shifts only)

overengineered_bit_moving.cpp

uint64_t src  = start + size;
uint64_t dest = new_start + size;

if(src % 8 != 0) {
	uint8_t b = bytes[src >> 3] >> (8 - src % 8) & ((1UL << (src % 8)) - 1);

	if(dest % 8 == 0) {
		bytes[(dest >> 3) - 1] &= 0xFF << (src % 8);
		bytes[(dest >> 3) - 1] |= b;
		// std::cout << "1" << std::endl;
	} else if(dest % 8 >= src % 8) {
		bytes[dest >> 3] &= ~(((1UL << (src % 8)) - 1) << (8 - dest % 8));
		bytes[dest >> 3] |= b << (8 - dest % 8);
		// std::cout << "2" << std::endl;
	} else {
		bytes[dest >> 3] &= ((1UL << (8 - dest % 8)) - 1);
		bytes[dest >> 3] |= b << (8 - dest % 8);
		bytes[(dest >> 3) - 1] &= 0xFF << (src % 8 - dest % 8);
		bytes[(dest >> 3) - 1] |= b >> (dest % 8);
		// std::cout << "3" << std::endl;
	}

	dest -= src % 8;
	src -= src % 8;
}

// Destination bitmasks
const uint8_t dest_bits_right         = dest % 8;
const uint8_t dest_mask_right_noshift = ((1UL << dest_bits_right) - 1);
const uint8_t dest_mask_right         = dest_mask_right_noshift << (8 - dest_bits_right);
const uint8_t dest_mask_left          = (1UL << (8 - dest_bits_right)) - 1;

while(src - start != 0) {
	uint8_t b = bytes[(src >> 3) - 1];

	if(src - start > 8) {
		bytes[dest >> 3] &= dest_mask_left;
		bytes[dest >> 3] |= b << (8 - dest_bits_right);
		bytes[(dest >> 3) - 1] &= dest_mask_right;
		bytes[(dest >> 3) - 1] |= b >> dest_bits_right & dest_mask_left;
		src -= 8;
		dest -= 8;
	} else {
		if(dest % 8 == 0) {
			uint8_t mask = 0xFF << (src - start);
			bytes[(dest >> 3) - 1] &= mask;
			bytes[(dest >> 3) - 1] |= b & ((1UL << (src - start)) - 1);
		} else if(dest % 8 >= src - start) {
			uint8_t mask = ~(((1UL << (src - start)) - 1) << (8 - dest % 8));
			bytes[dest >> 3] &= mask;
			bytes[dest >> 3] |= (b & ((1UL << (src - start)) - 1)) << (8 - dest % 8);
		} else {
			bytes[dest >> 3] &= dest_mask_left;
			bytes[dest >> 3] |= b << (8 - dest_bits_right);
			bytes[(dest >> 3) - 1] &= 0xFF << (src - start - dest % 8);
			bytes[(dest >> 3) - 1] |= b >> (dest % 8) & ((1UL << (src - start - dest % 8)) - 1);
		}

		break;
	}
}