Since the bypass case is quite unlikely, especially when doing merged sigmap/level coding,
it's faster to use a branch than a cmov.