The post did not mention potential cost of splitting and masking these into 5 limbs instead of 4. Perhaps I am missing that in this proposed system, 256-bit int

The post did not mention potential cost of splitting and masking these into 5 limbs instead of 4. Perhaps I am missing that in this proposed system, 256-bit ints are always stored in 5 registers, then you only normalize when the result is needed? What are the costs of splitting/masking into 5-limbs and could the scaling impact the gains of parallel adds/subtracts meaningfully?