GC improvement

[dstogov]

No description provided.

[nikic]

I will review and run benchmarks on this today in the evening.

[nikic]

Here are my benchmark results, comparing three different workloads between master (OLD) and this PR (NEW). The first workload currently completely trashes the GC implementation on master, the other two are more lightweight.

Here the GC "disabled" lines mean disabled at runtime, so roots are still collected.

// Very, very, very many objects
GC       |    OLD |   NEW
disabled |  1.32s | 1.50s
enabled  | 12.75s | 2.32s

// Very many objects
GC       |    OLD |   NEW
disabled |  0.87s | 0.87s
enabled  |  1.48s | 0.94s

// Less many objects
GC       |    OLD |   NEW
disabled |  1.65s | 1.62s
enabled  |  1.75s | 1.62s

The most interesting is probably the first result:

• With GC enabled, this is still 1.5x slower than with GC disabled (previously it was nearly 10x slower). This is because the (linear) GC threshold backoff happens too slowly for this case. I think this is fine to start with.
• With GC (runtime) disabled, we have a slowdown from the old to the new implementation (1.32s to 1.50s). perf shows that 10% of the time is spent inside gc_remove_compressed. So this workload happens to use enough objects that the compression scheme becomes a problem...

For the compression, I wonder if the current scheme that separates compressed & uncompressed addresses is really worthwhile. I would suggest to always mask the address (something like this: https://gist.github.com/nikic/515dc2dfdc4912cee5c6e1fb17a4d276). This means that we always have to do a root->ref != ref check when removing a root, but on the other hand it saves a bunch of checks in other places and also resolves the gc_remove_compressed performance issue mentioned above. Looking at perf, always doing this check does not seem problematic.

In any case, I really like this implementation, dropping the linked lists makes this a lot nicer. It's a great improvement, let's land it!

[nikic]

nit: poissible -> possible

[nikic]

nit: Can be just * instead of *=, as GC_G(buf_size) is assigned below.

[nikic]

nit: Indentation

[nikic]

The GC_GET_PTR here can be dropped. At least for me the mask is not optimized away.

[nikic]

I think this function is no longer used.

[nikic]

We could explicitly set gc_protected on unclean shutdown and save one check here.

[dstogov]

Right. I already thought about this. Lets do this after merge.

[nikic]

last looks unused here.

[dstogov]

@nikic thank you for review and benchmarks. Really impressive :)

You may play with GC_THRESHOLD adoption after the merge.

I'll think about compression once again.
In general, I like your patch.
It didn't work out of the box (you forgot about GC_BITS).
We have to perform (GC_GET_PTR(root->ref) == ref) check.
For me, your implementation is a bit "slower" (according to callgrind)

I hope, I'll merge this by tomorrow evening.

[nikic]

You're right, I missed the GC_GET_PTR(). I guess we need it in this line as well, it's just less likely to hit here.

[dstogov]

I've found a compromise, that keep good performance an makes small overhead only for apps with big data sets - (GC_G(first_unused) >= GC_MAX_UNCOMPRESSED).

[frederikbosch]

Amazing work!

[kelunik]

These recent GC changes cause segfaults, see https://bugs.php.net/bug.php?id=76050.

[dstogov]

Thanks for catching.
I've reproduced and understood the reason.
@nikic could you also take a look into bug report.