Switch to LLVM/bolt

[h-vetinari]

Hey all!

Since BOLT is now included as of LLVM 14, I wanted to open an issue to track updating the bolt version being used here to the one in LLVM. It will make eventual packaging for conda(-forge) a fair bit easier**. 🙃

** even more so if pyston were to rely on released versions of LLVM rather than through submodules, but that's a different subject.

[kmod]

That'd be great! We just updated to LLVM 13.0.1 (the latest release) so updating to LLVM 14 will be easier now.

We have the ability now to use a system-provided LLVM, which we use for everything except development. The two things that I can think of that would have to be done to get rid of the submodule:

• We would need to change all references from clang, llvm-link, etc, to clang-13, etc
• We'd need a strategy for when we run into bolt bugs, which are typically fixed in their main but not their release. This has been frequent enough that I expect it to happen again.

Perhaps we can get rid of the llvm submodule (solve the first issue) and keep the bolt submodule (punt on the second issue)?

[h-vetinari]

Cool!

We'd need a strategy for when we run into bolt bugs, which are typically fixed in their main but not their release. This has been frequent enough that I expect it to happen again.

I expect (naïvely...?) bolt to stabilize in terms of bug rate, but that's something to be seen. In this context, it's worth noting thought that conda-forge would be able to carry relevant (bolt-)patches without waiting for the newest LLVM release.

Perhaps we can get rid of the llvm submodule (solve the first issue) and keep the bolt submodule (punt on the second issue)?

I'm not suggesting to do it right away. It would already be an improvement if the bolt submodule here were to point to LLVM/bolt. And great if llvm itself can "go" as well! :)