Reproducibility regression caused by 57edf88

[jgalenson]

My nightly script that checks for reproducible builds of rustc failed this morning. Git bisect pointed me to 57edf88 (#75138).

With that commit, about 12 rlibs have differences in my parallel builds (librustc_privacy is one) and nothing in the diff of the binaries jumps out as an obvious culprit.

[jyn514]

The PR you linked to wasn't merged; I think you meant #75138.

[jgalenson]

Ah, yes, sorry about that. The commit should be correct, though.

[LeSeulArtichaut]

cc @jumbatm @oli-obk

[oli-obk]

oops, yea this is likely

rust/compiler/rustc_macros/src/session_diagnostic.rs

Line 621 in 71569e4

let args = referenced_fields.into_iter().map(|field: String| {

(not sure if there are other hashset/hashmap iterations happening, but that's the only one I found).

So the fix is twofold: First change all the HashSet and HashMap to FxHashSet and FxHashMap where possible, and second, collect to a Vec and sort that Vec before iterating over the set/map

[jumbatm]

Ah, my bad. I can take this.

@rustbot claim