Re: [PATCH] More speedups, this time in build_trtable

[Jakub Jelinek <jakub at redhat dot com>]

On Tue, Dec 16, 2003 at 02:37:35PM +0100, Paolo Bonzini wrote:
> After Jakub's patch that disabled transit_state_sb, build_trtable 

transit_state_sb removal has nothing to do with this FYI.

> frequently shows up at the top of the profiles.  This patch divides the 
> inner loop in two (for state->word_trtable == 0 or == 1 respectively) 
> and rewrites the iteration on bitsets so that it does not iterate on  
> empty items.  The result is a 20% improvement in the time needed to run 
> the sed testsuite on my machine, with a top improvement of 2x on 
> madding.sed (which is a search for a very long string).

In what locale?

Will not your patch set word_trtable = 1 even when without the patch it
would be clear?  That would be a pessimization (a DFA state is built once
with build_trtable, but can be used many times during searching and when
word_trtable is set, transit_state needs to acquire context).
E.g. with a(\b[abc]|[def]).  Perhaps it would be enough if just
group_nodes_into_DFAstates never created empty dests_ch[] bitsets.

	Jakub