Investigate usage of BOLT

[deeprobin]

Description

BOLT was a Facebook research project that has since been merged into LLVM.

It is a post-link optimizer that optimizes the binary based on a profile, as the binary layout can have a certain influence on performance.

We could possibly try this out in the context of PHP - there are now more and more open source projects that use BOLT.

For example, rustc, the Rust compiler has implemented this in its CI pipeline (3-5% performance improvement).
There was also a blog article for Postgres. Even more performance was achieved there.

I could well imagine that BOLT would do PHP a lot of good.
However, BOLT is only supported for Linux binaries - whether Windows & *BSD support is coming is questionable.

What do you think? Should we try it out in the context of PHP?

[devnexen]

For example, rustc, the Rust compiler has implemented this in its CI pipeline (3-5% performance improvement). There was also a blog article for Postgres. Even more performance was achieved there.

To note also binaries size increase noticeably (I do not know how much for rustc however I ve tried few times and for one the binaries were 5 times bigger while having also similar performance increase (< 5 %)).

[NattyNarwhal]

We're not shipping binaries except for Windows, so I don't know what this would entail other than testing it in CI.

[devnexen]

Sure... was just highlighting benefits vs costs.

[deeprobin]

@devnexen But is storage / binary size a big problem today?
The PHP interpreter is not really an application that runs very often on IoT devices where the storage is actually that relevant.

Personally, I would consider performance to be more important than binary size.
If there are significant differences in the binary size, we could consider whether it might make sense to publish a kind of “php-slim”.

But I think it might be worth trying BOLT and benchmark the performance differences.

[devnexen]

Oh I do not disagree

[iluuu1994]

But is storage / binary size a big problem today?

Maybe not, but note that binary size can also hurt instruction cache. So, it's not a given that performance actually improves. It heavily depends on what code paths are hit. We already signal this to compilers using the __builtin_expect intrinsic. For many branches, this will depend on your individual PHP application. This poses the question: On what application(s) would we profile PHP? There is no good answer to this. So, if you'd like to make use of PGO (profile-guided optimization), you'll most likely need to profile and build PHP for your individual use-case.

So, I don't believe there's anything we should do here.

Also note that PHP for everything but Windows is compiled and distributed by distributions. So, this is not something we can decide anyway.

[github-actions]

No feedback was provided. The issue is being suspended because we assume that you are no longer experiencing the problem. If this is not the case and you are able to provide the information that was requested earlier, please do so. Thank you.