This is the initial split to decouple AVX256 composed operations from their MMX/SSE counterparts. This is to work around the subtle differences with AVX/SSE zext/insert behaviour.
SupportsSaturatingRoundingShifts