Set force-frame-pointers=yes on arm machines on arm linux

[jackkleeman]

To enable Parca - parca-dev/parca-agent#1805

[tillrohrmann]

Thanks for creating this PR @jackkleeman. Before merging it, what are the implications for Restate when enabling the frame pointers (like why isn't it on by default)?

[jackkleeman]

On a few platform eg apple (rust-lang/rust#85706) and openbsd the frame pointers are actually on by default. This field influences LLVM compiler config, which would otherwise only include frame pointers for some functions based on some factors eg whether they have variable size allocations:
https://github.com/llvm/llvm-project/blob/c0318f079a0c7a0558c29f04b4e190277f6f6687/llvm/lib/Target/X86/X86FrameLowering.cpp#L97

This discusses the perf implications of enabling frame pointers:
https://stackoverflow.com/questions/13006371/does-omitting-the-frame-pointers-really-have-a-positive-effect-on-performance-an
It seems perhaps some applications see a 5-10% slowdown.
This comment says that rustc and ripgrep saw negligible slowdowns:
rust-lang/rust#86196 (comment)

I can change to just enabling this for arm, given thats where we need it (and it already will be enabled on apple targets, so really the delta is just AWS arm servers). I also suspect we don't need this forever, as parca will probably work on getting the same info without frame pointers.

[tillrohrmann]

Thanks for the pointers @jackkleeman. I will try to see whether I can measure a difference in performance on arm64 on Linux. If not, then this PR should be good to be merged.

[tillrohrmann]

When testing it locally in my Docker environment, I couldn't measure a significant change in performance. Hence, +1 for merging it.

[slinkydeveloper]

@jackkleeman Have you checked we don't need to build the std library with nightly to enable frame pointers? I've seen this around: https://docs.rs/tracefp/latest/tracefp/

[jackkleeman]

I don't believe so, no. Let's see if Parca works with this