perf: Skip the first goto call in full construction (-35%)

[Marwes]

• perf: Skip the first goto call in full construction (-35%)

This is probably less significant when larger number of needles are used
but should still be a decent speed improvement.

 name                     old ns/iter  new ns/iter  diff ns/iter   diff %  speedup
 bench_construction       131,148      122,956            -8,192   -6.25%   x 1.07
 bench_full_construction  309,493      200,595          -108,898  -35.19%   x 1.54

• perf: Use for_each_transition in AcAutomaton::fill (-84%)

 name                     old ns/iter  new ns/iter  diff ns/iter   diff %  speedup
 bench_construction       167,415      25,804           -141,611  -84.59%   x 6.49
 bench_full_construction  275,170      133,253          -141,917  -51.57%   x 2.07

(The two benchmarks were done on different computers)

[Marwes]

This should be the last PR :) It does complicate things a bit more but it really helps when querying for the next state of Dense as many states (at least in these short benchmarks) only have distance of 1 to the first non-fail state.

However, if next_state needs to traverse more than one state then it should be increasingly likely that the memorization kicks so chances are the memoization gives a greater performance improvement than 12% in that case instead!

[Marwes]

Another thing that could have improved things further would be to actively update full_automaton as the automaton is searched (essentially doing path compression https://en.wikipedia.org/wiki/Disjoint-set_data_structure#Path_compression). Unfortunately, needing to query full_automation and checking for != FAIL_STATE on each loop to see if the state is already set instead of just checking si < current_si seems to negate any improvements of doing so.

[Marwes]

(*) Second to last PR, it is possible to improve (and complicate) the AcAutomaton a bit more as well (-24%)

[BurntSushi]

Thanks for this! I'm going to let this sit for at least a week to allow any additional changes you might have planned to accrue here in order to minimize churn and context switching.

[Marwes]

(*) Second to last PR, it is possible to improve (and complicate) the AcAutomaton a bit more as well (-24%)

This ended up not having that large of an impact so I scrapped it.

Instead I realized that for_each_transition can be used pervasively with some slight restructuring which it seems like LLVM really liked. As Dense states are just stored as Vec<(u8, StateIdx)> we avoid iterating over the entire 0...255 range and instead only iterate the few states in the Vec yielding a massive speedup in constructing AcAutomaton.

The only idea I have left to try is to re-apply the above idea as it could be more significant after this improvment but I don't have those sources accessible atm so that will have to wait.

[Marwes]

Rebased and added for_each_ok_transition so we don't have to rely on LLVM for the improvement in 2b9df73.

The only idea I have left to try is to re-apply the above idea as it could be more significant after this improvment but I don't have those sources accessible atm so that will have to wait.

Still no improvement from this so I am finished with all optimizations I am able to do here.

[Marwes]

ping @BurntSushi (this is done)

[BurntSushi]

Awesome! Thanks so much for this. Sorry it took a while. I needed to make time to carefully review this. I think I buy everything! I've merged this locally (to drop the unneeded CI commit---thank you for that! I fixed it elsewhere) and pushed it to crates.io in aho-corasick 0.6.8.

These speed improvements are really really nice. It actually makes a huge difference in a real test case I've been playing with: BurntSushi/ripgrep#838