gh-109587: Allow "precompiled" perf-trampolines to largely mitigate the cost of enabling perf-trampolines

[gsallam]

This pull request implements a proposal in #109587 to allow precompiled perf-trampolines. This would mitigate the costs of enabling perf-trampolines, such as the disk IO and memory overhead.

The proposal introduces two new C-API functions:

• PyUnstable_PerfTrampoline_CompileCode(): Creates a new trampoline and registers it.
• PyUnstable_PerfTrampoline_SetPersistAfterFork(): Controls whether to re-initialize trampolines in child processes.

These functions can be used by extension modules to initialize trampolines eagerly, after the application is "warmed up". This would make it possible to have perf-trampoline running in an always-enabled fashion.

Benefits for a forked multiprocess model:

• Reduces the CPU overhead of enabling perf-trampolines.
• Makes it possible to have perf-trampoline running in an always-enabled fashion.

• Issue: [proposal] Allow "precompiled" perf-trampolines to largely mitigate the cost of enabling perf-trampolines #109587

[brandtbucher]

Here too?

[czardoz]

@pablogsal , @gpshead Would love to hear your thoughts on this (whenever you get a chance to take a look) :)

[czardoz]

Ok I resolved the merge conflict, but I think the new MacOS failure is unrelated.

[pablogsal]

LGTM

Seems that we have some conflicts that we need to address before landing, @gsallam can you take a look?

[pablogsal]

@gsallam We still have some failures in the Windows and macOS CI. I think this is because the symbol doesn't exist in those cases:

_testinternalcapi.obj : error LNK2019: unresolved external symbol __imp_PyUnstable_PerfTrampoline_CompileCode referenced in function compile_perf_trampoline_entry [D:\a\cpython\cpython\PCbuild\_testinternalcapi.vcxproj] D:\a\cpython\cpython\PCbuild\amd64\_testinternalcapi_d.pyd : fatal error LNK1120: 1 unresolved externals [D:\a\cpython\cpython\PCbuild\_testinternalcapi.vcxproj] 0 Warning(s) 2 Error(s) 

[pablogsal]

We still have some failing tests it seems

[czardoz]

This approach is likely quite silly, but I don't have a Windows/Mac machine handy, so I'm kicking off these builds to test/debug.

[czardoz]

I will look at this over the weekend on my Windows box, this trial-and-error is not an optimal way :)

[pablogsal]

I will look at this over the weekend on my Windows box, this trial-and-error is not an optimal way :)

Don't worry, I pushed a commit that should fix it

[czardoz]

Thanks a lot @pablogsal! 😄

[encukou]

The PyUnstable functions added here are missing documentation.
@Yashp002 sent a PR to add docs: #143492