Explore if the upcoming XCode 16.3 supports tail calling interpreters

[Fidget-Spinner]

Feature or enhancement

Proposal:

Xcode 16.3 seems to be getting a mixture of LLVM 17-19 features. It may be possible preserve_none ends up in there. If so, it would significantly boost interpreter performance if we can enable it.

https://developer.apple.com/documentation/xcode-release-notes/xcode-16_3-release-notes

We should play around with Xcode 16.3 and at least verify that it can/cannot support [[clang::musttail]] and __attribute__((preserve_none)).

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

[Fidget-Spinner]

Internal reports suggests this might be possible. Will let Ned know once Xcode 16.3 is released proper.

[ronaldoussoren]

The current Xcode beta (16.3 build 16E5104o) builds with --with-tail-call-interp and passes the test suite (e53d105).

The compiler claims to be Apple clang version 17.0.0 (clang-1700.0.9.2).

All of this on an M1 system, I'm assuming that this will also work for x86_64.

[Fidget-Spinner]

@ned-deily It seems Xcode 16.3 supports this and gets the best performance on this configuration. See https://github.com/facebookexperimental/free-threading-benchmarking/tree/main/results/bm-20250403-3.14.0a6+-0227a6d-NOGIL for benchmarks on free threaded build.

Can you please pass --with-tail-call-interp to the 3.14 macos release builds?

[Fidget-Spinner]

On second thought, it seems to me Xcode 16.3 might have the computed goto bug which will worsen performance on everything prior to 3.14.

Ned: you might want to stick to Xcode 16.2 for 3.13 and below, while using 16.3 for 3.14. Or if thats too much of a hassle, just 16.2 for everything while noting that the perf on 3.14 might be bad.

This is until xcode releases a clang with the computed goto patch.

[hugovk]

@Fidget-Spinner Is this still a release blocker?

[ned-deily]

No, it's not a release blocker. I'm looking at the various Xcode options for the installer builds but we have viable choices.

[ned-deily]

An update: I wasn't able to complete Xcode testing prior to today's releases so I chose to stick with the build tools used for the previous releases: Xcode 15.4 for 3.13.3 and Xcode 16.2 (using computed gotos) for 3.14.0a7. Very preliminary results suggested that those did not have any negative surprises. Now that the releases are complete, I will finish the testing across a wider range of options and summarize the results in the coming days.

[Fidget-Spinner]

@ned-deily were you able to finish your testing?