Optimise rust-based formulas by setting target-cpu

[Tokarak]

Verification

• This issue's title and/or description do not reference a single formula e.g. brew install wget. If they do, open an issue at https://github.com/Homebrew/homebrew-core/issues/new/choose instead.

Provide a detailed description of the proposed feature

Rust has optional support for building with certain cpu features enabled. For example, arm-based macs, I would imagine, use only apple-m1 or apple-m2 cpus. rustc --print cfg | grep target_feature | wc -l gives me 28 features, while the generic target only has neon.

Specific features can also be specified — this can be useful if it's feasible to take the common features of a set of cpus.

If a better union set of cpu features can be found than generic, that can be an improvement to the ecosystem.

What is the motivation for the feature?

This can increase peformance (sometimes marginally, other times significantly) for binaries built in Rust.

How will the feature be relevant to at least 90% of Homebrew users?

Mac users on official hardware should use only a small set of cpus. This means there is a high potential for improvement.

I'm worried about alienating the small number of hackingtosh and opencore-patcher users. Perphaps if the OS requires specific features, then those can be used. Otherwise, if a Hackingtosh can be built on any hardware, then it would seem that the only way to avoid alienation is to stick to the generic targets, or provide multiple builds (though I'm not looking to inflate the CI minutes).

This is particularly true for silicon based macs. AFAIK, there are no arm-based hackingtoshes, and there are only two, very similar cpu types. This could mean that it's completely safe to add these features cargo_args on silicon builds and distributions.

What alternatives to the feature have been considered?

• Builds for each official cpu target, with generic target as backup
• Lowest-common-denominator builds for official hardware. Generic backup may or may not be needed.
• Add this only where it is convenient (eg for arm-based targets)
• Add this to local build only (very achievable, but building wastes more power than unspecialised binaries ever will — alone, this is almost redundant)
• Don't fix what ain't broken

[carlocab]

For C/C++-based formulae, we set -march to the oldest supported architecture for a given OS version on Intel macOS. Currently that's nehalem, which is pretty old, but still an improvement over what you get when you omit -march (penryn). We don't do this on M1 macOS since Clang will default to targeting apple-m1 without any -march/-mcpu flags.

It seems reasonable to do something similar for Rust. How do you set the desired target-cpu with rustc?

[Bo98]

We don't do this on M1 macOS since Clang will default to targeting apple-m1 without any -march/-mcpu flags.

We will probably add flags at some point for the benefit of GCC. The flags are however different on GCC and Clang so it's trickier than Intel.

But yeah for Rust this seems like a good idea.

[alex]

FWIW starting with Rust 1.71 (coming out in 5.5 weeks) rust will default to targeting an M1 in macOS arm64 targets: rust-lang/rust#109899

[Bo98]

Ah excellent - we probably don't need to do anything here for arm64 then.

Not sure if there's anything worth doing for x86_64 or not.

[Tokarak]

Currently that's nehalem

This simplifies deciding which cpu to target.

FWIW starting with Rust 1.71 (coming out in 5.5 weeks) rust will default to targeting an M1 in macOS arm64 targets: rust-lang/rust#109899

That's nice to know, thanks.

nehalem is in the list of cpus supported by LLVM; I checked. The following flag needs to be added to cargo (x64 builds only):
-Ctarget-cpu=nehalem
Equivalently (but redundantly since Rust 1.71) for aarch64:
-Ctarget-cpu=apple-m1
To optimise local builds, -Ctarget-cpu=native can be passed. I don't know if homebrew differentiates between local and distribution builds. Also, for native AVX2 might need to be disabled by default; I saw many issues online saying it could make the binary slower.

If the flag can be conditionally added to std_cargo_args, that would mean no refactoring of formulas. I don't know much detail about the homebrew codebase, so maybe there is another way.

[carlocab]

I don't think passing a -C flag to cargo is correct. From cargo-install(1):

       -C PATH
           Changes the current working directory before executing any specified operations. This affects things like where cargo looks
           by default for the project manifest (Cargo.toml), as well as the directories searched for discovering .cargo/config.toml,
           for example.

Perhaps you need for this to be set in RUSTFLAGS?

[Tokarak]

Yes, you are right. -C passes a codegen flag to rustc.

For cargo, this option can be set in rustflags.

[MikeMcQuaid]

I don't know if homebrew differentiates between local and distribution builds.

We used to but intentionally don't any more to reduce the difference (and we don't support local builds anyway).

As a result,

-Ctarget-cpu=nehalem

sounds perfect.

[Tokarak]

RUSTFLAGS='-C target-cpu=nehalem' cargo build ...

[Tokarak]

I think I can self-assign this.

I could add a config flag to std_cargo_args based on the target triple (thus sorting out conditionals on cpu arch). A cargo config file will be more ergonomic at this point, but that can be left for a future PR.

If anyone has general advice or warnings of any pitfalls, please let me know.

[carlocab]

Off the top of my head, one important thing to note is that we'll want to use the same value returned from here instead of actually hard-coding nehalem or apple-m1:

brew/Library/Homebrew/extend/ENV/shared.rb

Lines 273 to 281 in 9d01e1d

	# @private
	sig { returns(Symbol) }
	def effective_arch
	if @build_bottle && @bottle_arch
	@bottle_arch.to_sym
	else
	Hardware.oldest_cpu
	end
	end

[MikeMcQuaid]

Sounds good, looking forward to the PR @Tokarak!

[Tokarak]

I looked at oldest_cpus, and it will be a nightmare to compatibilise with rust.
Current rust nightly behaviour defaults to oldest supported cpus

• Defaults to M1 on aarch64-apple-darwin
• Core2 on x86_64-apple-darwin
  • This is the current homebrew behaviour for Version < 10.14; it's easier not to change this in that case.

Given that Rust already minimaxs the cpu without version context, it may be best to create a function which returns rust flags with version context, similar to

brew/Library/Homebrew/extend/ENV/shared.rb

Lines 273 to 281 in 9d01e1d

	# @private
	sig { returns(Symbol) }
	def effective_arch
	if @build_bottle && @bottle_arch
	@bottle_arch.to_sym
	else
	Hardware.oldest_cpu
	end
	end

but without oldest_cpu which has a lot of complex logic which is incompatible with rustflags and is already handled by cargo. This allows to use useful default rust cpus (just return an empty string to not change rust flags!) Maintaining 5 lines of code shouldn't be an issue either, even though oldest_cpu and this function will have to be updated in parallel.

[MikeMcQuaid]

Closing to keep discussion in #15544.

Given that Rust already minimaxs the cpu without version context, it may be best to create a function which returns rust flags with version context, similar to

Seems reasonable 👍🏻