While reviewing @MattStephanson's implementation of #4169 / WG21-P0952R2, @statementreply noted in #4740 (comment) :

_Sx doesn't need to be wider than _Result_uint_type when _Rx is a power of 2, as is noted in https://eel.is/c++draft/rand.util.canonical#7 . The $i$-th random value could be shifted by $i \log_2 R - \log_2 x$ bits instead (the 0th random value is shifted right).