PM managed GC

[dktapps]

This is a first look at #5628.

In a nutshell: Server MemoryManager now controls when GC is called. The mechanism by which it's activated remains the same as PHP's own, but instead of being called in random code paths & causing huge lag spikes, it gets called at a predictable location & can therefore be tracked in timings reports.

This required re-implementing PHP's garbage collection threshold algorithm, as it's not possible to reliably make PHP update its own internal threshold when GC isn't always enabled.

A couple of considerations I came across while working on this. It seems like PHP's GC algorithm is rather suboptimal for PM's use case.

• The average GC run in PM takes about 10-15ms. To my surprise, the biggest factor influencing performance is actually the depth of referenced objects. This means that anything referencing Server getting unref'd causes a hit to GC performance because Server will land in the GC root buffer, and Server is a gateway to everything.
• Long-lived complex objects like Server will frequently end up taking up space in the GC root buffer (whenever their refcounts go down), and then the GC wastes a ton of CPU time trying to scan them for dependencies. There's no real solution for this beyond requesting PHP to implement generations in the GC (A way to stop GC wasting CPU time on long-lived objects (GC generations or something else) php/php-src#17131).
• Acyclic objects flood the root buffer and cause the threshold to go sky-high. In itself, this doesn't affect the GC's performance that much (e.g. a Vector3 takes only tens of nanoseconds per object to dispose), but the filling up of the GC root buffer causes more frequent runs, causing complex objects to get repeatedly scanned as in the above point.

Follow-up

• Explore how much performance benefit we really get from the World's block cache. These cached objects tend to flood the GC root buffer, causing collection sweeps that collect barely any objects.
• After re-implementing the PHP native threshold algorithm, I realized it probably wasn't optimal for PM's use case. For example, the default root buffer size should probably be significantly larger (around 100k or so) to allow for the World block cache to not constantly trigger GC.

[dktapps]

Related issues:

php/php-src#17127 (Acyclic objects landing in GC root buffer)
php/php-src#17131 (No generational GC to avoid wasting CPU time on long-lived objects)