implement faster HashMap

[ericvergnaud]

new faster HashMap

[ericvergnaud]

ignore unrelated php failure

[kaby76]

For this PR, there is a 40% speed up. (2.8537 +- SD 0.0124 vs 2.0492 +- SD 0.033).

[kaby76]

(There may be other optimizations that Antlr4ng does, but I haven't gotten that far.)

[ericvergnaud]

So we’re talking 60-65% cumulated improvements ? Whoosh!!!Envoyé de mon iPhoneLe 10 mars 2024 à 23:32, Ken Domino ***@***.***> a écrit : (There may be other optimizations that Antlr4ng does, but I haven't gotten that far.) —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***>

[kaby76]

Yes, something like ~60%. There are probably other optimizations to look at.

At some point, can we get a release of 4.13.2?? No hurries though, after other fixes are considered. I did not like this 4.13.1-patch1 release because that breaks my testing code, and I don't think GitHub Dependabot even considers this a release.

[kaby76]

This PR passes all grammars-v4 tests.

[ericvergnaud]

At some point, can we get a release of 4.13.2?? No hurries though, after other fixes are considered. I did not like this 4.13.1-patch1 release because that breaks my testing code, and I don't think GitHub Dependabot even considers this a release.

That's a question for @parrt.

[ericvergnaud]

@kaby76 it's be interesting to run your performance benchmark on the merged dev version

[kaby76]

The runtime for the tip of dev is better.

But, it is still missing something as Antlr4ng is faster. I am still surprised the hash value caching didn't help. I must have done something wrong, or it should have been applied somewhere else. Looking....

[ericvergnaud]

antlr4ng also implements specialized constructors which avoid testing instance types. It might be worth doing this for the objects that are instantiated millions of times (such as ATNConfig). But we might be reaching a point where profiling becomes necessary to better understand the remaining bottlenecks.
Sam mentioned that the reason for not caching ATNConfig hashCode is memory footprint.

[ericvergnaud]

But, it is still missing something as Antlr4ng is faster. I am still surprised the hash value caching didn't help. I must have done something wrong, or it should have been applied somewhere else. Looking....

I can revise and rebase that branch if you're willing to double-check that...

[kaby76]

I'm profiling the CSharp target because your PR changed one file in that target. (IMO, C# has far-superior profiling tools to anything I've seen in all other targets.)