Frame pointer (and any per-function features) builds of tail calling interpreter slower than expected

[Fidget-Spinner]

Bug report

Bug description:

When building with the normal interpreter (computed goto), the impact of frame pointers should only be around 2% geomean max.

However, it seems the impact on the tail calling interpreter is higher. This is likely because each bytecode handler on the tail calling interpreter is now a function, which means they each have a frame pointer prologue and epilogue.

There is a fix on Clang 21 and higher (though I don't know if GCC 16 has it): omit the frame pointer, but reserve it so it doesn't get clobbered only for the bytecode handlers, leave the rest building with frame pointers. The flags are "-fomit-frame-pointer -momit-leaf-frame-pointer -mreserve-frame-pointer-reg". The key observation is that since we tail call, we only need the first entry into a tail calling handler to have the frame pointer prologue (ie LABEL(start_frame) needs the prologue), but all other tail calling function pointers do not. This practically eliminates the entire prologue/epilogue overhead for the interpreter, while keeping register allocation good and not clobbering frame pointers. Unfortunately, clang does seem not have custom per-function attributes for this option, so we have to move the non-starter bytecode handlers to their own compilation unit (ie, move them to their own C file).

This seems to affect anything with per-function overhead as well, such as CET/BTI changes. We can fix those incrementally, as the approacha re the same.

Here are some geometric mean results on Sam's fastmark (pyperformance subset) from my laptop (i7-12700H x86-64) with clang 22:

• Baseline (FP on, tail calling interpreter): 0% slowdown
• FP off, tail calling interpreter: 2.7% speedup
• FP on, tail calling interpreter, reserve register patch (meowl + garbage + oiia): 2.5% speedup

TLDR: WIth this patch, the tail calling interpreter's overhead for frame pointer enabling drops from 2.7% to just 0.2%!!!! Frame pointers are practically free on the tail calling interpreter!!! Meanwhile, computed goto interpreter has a 1.5% hit (the i7-12700h machine is mine from PEP 831)

@pablogsal @markshannon @stratakis

CPython versions tested on:

CPython main branch

Operating systems tested on:

No response

[Fidget-Spinner]

Tested the patch with https://github.com/pablogsal/cpython-unwind/blob/main/run.py, and it works! The default tail calling build without the frame pointers crashes the script, while with frame pointers forced on and on my patched meowl frame pointer build, it works!