I just noticed that this function also runs in every single subprocess. This would explain the 1min+ delays I saw with highly populated cache in phpstan/phpstan#5825 (my laptop has 16 logical cores, so this code took 16x the time).

Final commit produces a massive performance improvement for full analysis (33sec -> 13sec when analysing PocketMine with a 30k file cache on Windows).

Hi, if the directory is almost empty after 27df5cf then this isn't really needed, right?

@ondrejmirtes Not strictly true, there's still going to be immense drag for 2 days after updating, until PHPStan figures out that it can get rid of some files. In any case, if the cache gets overpopulated for some other reason, these changes will ensure that the problem doesn't come back again with such crippling impact (it made PHPStan basically unusable for me). While analysedPaths was the culprit for me, I've seen several other users with extremely bloated caches and I wouldn't bet that my issue is the same as theirs.

If you want to go minimal on this, please at the very least take 5b784ad.

shouldn't we push it event further, so the worker processes don't do additional IO, which the master process already took care of?

(e.g. all these autoloadFile, scanFiles, stubFiles parameter validations leads to IO which all worker do again and again, without further benefit)

see

• https://github.com/pmmp/phpstan-src/blob/bdccbfe1c9cf0e5d0e374386d72fe0f7d489030c/src/Command/CommandHelper.php#L83-L103
• https://github.com/pmmp/phpstan-src/blob/bdccbfe1c9cf0e5d0e374386d72fe0f7d489030c/src/Command/CommandHelper.php#L125-L128
• https://github.com/pmmp/phpstan-src/blob/bdccbfe1c9cf0e5d0e374386d72fe0f7d489030c/src/Command/CommandHelper.php#L161-L169
• https://github.com/pmmp/phpstan-src/blob/bdccbfe1c9cf0e5d0e374386d72fe0f7d489030c/src/Command/CommandHelper.php#L296-L304
• https://github.com/pmmp/phpstan-src/blob/bdccbfe1c9cf0e5d0e374386d72fe0f7d489030c/src/Command/CommandHelper.php#L306-L333
• ...

thats a lot of file IO, which can be rather slow (especially when done in parallel across several workers concurrently onto the same files)

Personally I do plan to explore what other stuff can be done to optimise this, but for now I'm taking it one step at a time. Once this problem is addressed it'll be more obvious from profiling what needs to be optimised. So far it's looking like the bulk of the overhead is coming from autoloading once this problem is subtracted.

To be clear, I'm only favouring evidence-driven optimisations here (stuff that I can easily observe with a profiler or by eye). Efforts can be better directed towards things that have maximum impact that way.

The next biggest optimisation step I explored (in the context of PocketMine, anyway) was to avoid re-analysing stubs every time PHPStan is run (see #730). But over half of the overhead there comes from regenerating Nette DI containers (at least on Windows), so it's possible the recent analysedPaths change might have reduced the need for that by improving container reusability.

shouldn't we push it event further, so the worker processes don't do additional IO, which the master process already took care of?

@staabm Please make these kinds of decisions only with benchmark backing it up. I think that the impact of those operations is negligible.

If you want to go minimal on this, please at the very least take 5b784ad.

Alright, I did that one: 0f6245b

As for the rest - I don't like doing nonsystemic changes like that. If the Symfony Finder is bad from performance perspective, we should do our own implementation with nice API and use it everywhere.

Thank you!

If the Symfony Finder is bad from performance perspective

It's bad in this specific context because it uses RecursiveDirectoryIterator when we know the depth of the directory tree will be 0. I suppose this could be optimised in Symfony itself by special-casing ->depth(0) to use FilesystemIterator instead of RecursiveDirectoryIterator.

It's bad in this specific context because it uses RecursiveDirectoryIterator when we know the depth of the directory tree will be 0.

I think we could use a files() on the finder, so it should not recurse

see https://symfony.com/doc/current/components/finder.html#files-or-directories

I think we could use a files() on the finder, so it should not recurse

Profile shows it still recurses in this case. (or at least, it's still calling RecursiveDirectoryIterator::hasChildren()).

In any case, this looks like a bug/performance issue that should be addressed in Symfony itself. Looking at the code I think it's probably not too difficult (a depth tracker would probably be good enough).

It looks like this is a bug in PHP itself. Consider the following code, which should be silent:

<?php
require 'vendor/autoload.php';
$iterator = new class(sys_get_temp_dir() . '/phpstan/cache/nette.configurator') extends \RecursiveDirectoryIterator{
	public function hasChildren($allowLinks = false){
		var_dump("called hasChildren");
		return parent::hasChildren($allowLinks);
	}
};
$iterator2 = new RecursiveIteratorIterator($iterator);
$iterator2->setMaxDepth(0);
foreach($iterator2 as $item){
	//var_dump($item);
}

instead outputs string(18) "called hasChildren" many thousands of times.

Submitted a bug report https://bugs.php.net/bug.php?id=81554.

@dktapps how did you measure/profile worker-prozess startup time?

I found a way to blackfire profile the worker process by patching the ProcessHelper with

diff --git a/src/Process/ProcessHelper.php b/src/Process/ProcessHelper.php
index 0e16484dd..10b18b478 100644
--- a/src/Process/ProcessHelper.php
+++ b/src/Process/ProcessHelper.php
@@ -27,6 +27,8 @@ class ProcessHelper
 $phpIni = php_ini_loaded_file();
 $phpCmd = $phpIni === false ? escapeshellarg(PHP_BINARY) : sprintf('%s -c %s', escapeshellarg(PHP_BINARY), escapeshellarg($phpIni));
+ $phpCmd = 'blackfire run --ignore-exit-status ' . $phpCmd;
+
 $processCommandArray = [
 $phpCmd,
 ];

that way I get one profile per worker.

@staabm I used xdebug. It generates a file per PHP process anyway.