Add binary size measurement for multiple targets to perf

[thomcc]

We should add binary size measurement for a small set of target/program combinations to the perf suite. This involves choosing which targets and programs are relevant, although perhaps that should be informed by the work in #2, and possibly just done in these comments (or by the PR that implements support).

This is an action item from the meeting on 2023-11-30 and was discussed a bit around here: https://rust-lang.zulipchat.com/#narrow/stream/405744-wg-binary-size/topic/Meeting.202023-11-30/near/405210915.

[tgross35]

Created rust-lang/rustc-perf#1760

Any ideas of sample programs to build for the no_std or WASM targets?

[Kobzol]

We already measure binary size for all compile time benchmarks in the suite, and we have two specific benchmarks that optimize for small size (helloworld-tiny and ripgrep-tiny). Do you have any other specific things that you would like to add to the suite?

[tgross35]

At the meeting it came up that results for multiple targets would be nice, since there is really no reference for anything other than x86 linux. I am sure different targets is on the wishlist for runtime performance too, but being able to cross compile without running should hopefully make only reporting size relatively easy.

I think we would also like some no_std crates. But that should be easy enough for anyone to submit if there is support for build-only crates.

[Kobzol]

All our compile-time benchmarks (so most of our suite) are build only, but I suppose that you meant "we only care about the final artifact, not about compilation time". Right, that makes sense, so the real goal is to have an ability to cross-compile. I guess that we would need Docker/cross for the interesting targets (Windows, macOS)? I also wonder if the binary size tends to be the same when you cross-compile vs when you compile natively.

no_std should be easy to do, we can just add a new benchmark that is no_std :) In fact we might already have some, cranelift-codegen maybe.