Use the same crc64 that we already use for the symbol table hash as
the perfect hash function prehash. We appear to get radically faster
convergence this way, and the crc64 is probably *faster*, since the
table likely to be resident in memory.
Combining arithmetric (add) and bitwise (xor) mixing seems to give
better result than either.
With the new prehash function, we find a valid hash much quicker.
rand() in Perl can vary between platforms, so don't use it. Instead,
remove a completely pointless level of indirection (it introduced a
permutation which cancelled itself out) and provide a canned set of
random numbers for the rest. This guarantees we will always use the
same numbers.
Speed up the rejection of unhashed values (typically identifiers) by
filling unused hash slots with a large value (but not so large that
it is likely to overflow.) This means those values will be rejected
already by the range check, not needing strcmp().
Since we fold the f- and g-functions together, if we guarantee that g is
bipartite, we can make g twice the size of f without cost. This greatly
improves the odds of generating a smaller hash.
Finish the perfect hash tokenizer, and actually enable it.
Move stdscan() et al to a separate file, since it's not needed in any
of the clients of nasmlib other than nasm itself.
Run make alldeps.
