x86-64 microarchitecture level

[nagayev]

Feature or enhancement

Proposal:

We can compile Python only for modern processors to achieve a performance speedup.
Using a level 2 compiler, we can take advantage of SSE3 instructions, for example (https://en.wikipedia.org/wiki/X86-64#Microarchitecture_levels), and we still have the option to run Python on older hardware (a 2008 or newer processor is required)."
Also Windows 11 requires theese instructions (https://www.tomshardware.com/software/windows/microsoft-updates-windows-11-24h2-requirements-cpu-must-support-sse42-or-the-os-will-not-boot), RHEL 10 use v3:
https://docs.redhat.com/en/documentation/red_hat_enterprise_linux/10-beta/html/10.0_beta_release_notes/architectures and
Fedora discussion https://fedoraproject.org/wiki/Changes/Optimized_Binaries_for_the_AMD64_Architecture_v2

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

[mdboom]

@nagayev: Have you done any measurement of the benefits of these instructions on CPython using pyperformance?

[eli-schwartz]

I'm not certain I understand the purpose of this ticket. python.org doesn't distribute Linux binaries but if it did then they would need to work on more systems than just Fedora, and people frequently use Linux to revitalize older hardware which may not support microarchitecture -v2 or -v3 levels. Effectively dropping support for all those users is not a great experience, much less because the result is successfully downloading cpython, then running it and getting a SIGILL.

For Windows I'm not sure the distinction can be usable since I don't think that MSVC's cl.exe supports microarchitecture groupings.

[Fidget-Spinner]

I agree with Eli here. This would be a breaking change, and would either need:

1. The supported platforms (PEP 11) to update.
   OR
2. A PEP.

Anyways, I don't think this matters, because distributions choose to package/support their own CPython.

Thus, I'm closing this issue.

[eli-schwartz]

It would be interesting to compile with optimizations that align with the minimum supported hardware for Windows versions currently supported by CPython (currently that is Windows 8 and on, I believe, not Windows 11?) but knowing exactly which flags are being proposed for MSVC would help a lot -- it's not the same as the flags for gcc / clang.

It would also be fiddly to do as you want to pass them only for official binary releases and not restrict what people building from source can do.