> Is the 1 compare vs. 1 mask really a big deal? Is this mask operation the
> only reason the tables need to be power of two big?

If the only CPU you care about it a Pentium 4, maybe not.  It makes a
measurable difference on other platforms.

At any rate, it's easy to excuse `one instruction' here and `one
instruction' there (although in this case it depends on the platform),
but the system call path has to be optimal.