Improve expression resolving of superglobals

[herndlm]

⚠️ This is a re-do of #2012 which caused performance issues. ⚠️

I tried to do some simple testing, but didn't notice much of a difference on where I ran this (a VM, don't have anything better right now, sorry):

before:

→ hyperfine --prepare 'php bin/phpstan clear-result-cache -q' 'php -d memory_limit=448M bin/phpstan'
Benchmark 1: php -d memory_limit=448M bin/phpstan
  Time (mean ± σ):     73.668 s ±  1.285 s    [User: 262.490 s, System: 3.837 s]
  Range (min … max):   72.091 s … 76.341 s    10 runs

after:

→ hyperfine --prepare 'php bin/phpstan clear-result-cache -q' 'php -d memory_limit=448M bin/phpstan'
Benchmark 1: php -d memory_limit=448M bin/phpstan
  Time (mean ± σ):     73.868 s ±  0.507 s    [User: 264.315 s, System: 3.717 s]
  Range (min … max):   73.323 s … 75.092 s    10 runs

it would be interesting if #2012 (comment) still happens and, if so, if we can get more details here to maybe improve it before I forget about this again.

[phpstan-bot]

This pull request has been marked as ready for review.

[ondrejmirtes]

I'm not a huge fan of this approach. It's going to pollute the output of MutatingScope::debug() (https://phpstan.org/r/e037be34-122b-4846-82f4-487dc0963845) and make it unreadable. Also the values are going to be in the scope and they are going to be mostly the same for all modern apps which isn't great.

Could we add these to the scope only when they are narrowed by some assignments or conditions? I think we already do a similar thing for PHP_VERSION_ID (maybe all constants). Thanks.

[herndlm]

Could we add these to the scope only when they are narrowed by some assignments or conditions? I think we already do a similar thing for PHP_VERSION_ID (maybe all constants). Thanks.

smart, why haven't I thought about this :) I pushed something, not sure if it's going in the right direction or not though.

[herndlm]

I have a feeling that the stuff accessing Scope::SUPERGLOBAL_VARIABLES and MutatingScope::isGlobalVariable() as well as the looping in getSuperglobalTypes() and getNativeSuperglobalTypes() can be somehow improved still. I could try refactoring it.

[staabm]

you could de-duplicate this method by passing in the expressionTypes via parameter

[herndlm]

right. I stole this basically from getNativeConstantTypes() vs getConstantTypes() which could be adapted in a similar way then. maybe in a follow-up refactor :)