Memoize ExprHandler dispatch in MutatingScope::resolveType()

[staabm]

resolveType() re-scanned every tagged handler with a linear supports() sweep on each call - 2.3M supports() calls for 56K dispatches measured on a small self-analysis slice. NodeScopeResolver and TypeSpecifier already memoized this dispatch by Expr class in private copies; extract the memo into a shared ExprHandlerRegistry and use it in all three places.

before this PR:

➜  phpstan-src git:(2.2.x) php bin/phpstan clear-result-cache -q && time php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
php -d memory_limit=450M bin/phpstan analyze src/Analyser -q  
22.66s user 1.52s system 320% cpu 7.554 total
➜  phpstan-src git:(2.2.x) php bin/phpstan clear-result-cache -q && time php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
php -d memory_limit=450M bin/phpstan analyze src/Analyser -q  
22.62s user 1.53s system 318% cpu 7.573 total
➜  phpstan-src git:(2.2.x) php bin/phpstan clear-result-cache -q && time php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
php -d memory_limit=450M bin/phpstan analyze src/Analyser -q  
22.63s user 1.49s system 319% cpu 7.548 total

after this PR:

➜  phpstan-src git:(reg) php bin/phpstan clear-result-cache -q && time php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
php -d memory_limit=450M bin/phpstan analyze src/Analyser -q  
21.85s user 1.48s system 322% cpu 7.227 total
➜  phpstan-src git:(reg) php bin/phpstan clear-result-cache -q && time php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
php -d memory_limit=450M bin/phpstan analyze src/Analyser -q  
21.87s user 1.50s system 322% cpu 7.237 total
➜  phpstan-src git:(reg) php bin/phpstan clear-result-cache -q && time php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
php -d memory_limit=450M bin/phpstan analyze src/Analyser -q  
21.83s user 1.45s system 321% cpu 7.236 total

[ondrejmirtes]

I have a different version in the turbo-ext PR. Please take a look (it uses WeakMap) as I want to merge it soon.

If yours performed better it could be merged first but otherwise I'd prefer the other one.

[staabm]

best of 3 runs:

no weakmap

➜  phpstan-src git:(reg) hyperfine --prepare 'php bin/phpstan clear-result-cache -q' --runs 5 'php -d memory_limit=450M bin/phpstan analyze src/Analyser -q' -i
Benchmark 1: php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
  Time (mean ± σ):      7.587 s ±  0.040 s    [User: 21.813 s, System: 1.547 s]
  Range (min … max):    7.555 s …  7.652 s    5 runs

with weakmap

➜  phpstan-src git:(reg) ✗ hyperfine --prepare 'php bin/phpstan clear-result-cache -q' --runs 5 'php -d memory_limit=450M bin/phpstan analyze src/Analyser -q' -i
Benchmark 1: php -d memory_limit=450M bin/phpstan analyze src/Analyser -q
  Time (mean ± σ):      7.604 s ±  0.056 s    [User: 21.835 s, System: 1.577 s]
  Range (min … max):    7.545 s …  7.668 s    5 runs

I don't think the weakmap makes a consistent meaningful difference in my benchmark.

I run my benchmark only on src/Analyser atm, because otherwise because of thermal throttling numbers looks screwed.

---

in case we are fine in not using weakmap, we should revert the last commit in this PR before merging

[staabm]

diff of my impl vs. native acceleration PR: I made the key unique enough so it can also cache CallLike

[ondrejmirtes]

Yeah, please merge your no-WeakMap version (here), I'll rework the turbo-ext.