Re: reducing the footprint of ScanKeyword (was Re: Large writablevariables)

On Sun, Dec 16, 2018 at 11:50:15AM -0500, John Naylor wrote:
> A few months ago I was looking into faster search algorithms for
> ScanKeywordLookup(), so this is interesting to me. While an optimal
> full replacement would be a lot of work, the above ideas are much less
> invasive and would still have some benefit. Unless anyone intends to
> work on this, I'd like to flesh out the offset-into-giant-string
> approach a bit further:

Hello John,
I was pointed at your patch on IRC and decided to look into adding my
own pieces. What I can provide you is a fast perfect hash function
generator.  I've attached a sample hash function based on the current
main keyword list. hash() essentially gives you the number of the only
possible match, a final strcmp/memcmp is still necessary to verify that
it is an actual keyword though. The |0x20 can be dropped if all cases
have pre-lower-cased the input already. This would replace the binary
search in the lookup functions. Returning offsets directly would be easy
as well. That allows writing a single string where each entry is prefixed
with a type mask, the token id, the length of the keyword and the actual
keyword text. Does that sound useful to you?

Joerg