nondeterminism from codegen-units

[bmwiedemann]

While working on reproducible builds for openSUSE (sponsored by the NLnet NGI0 fund), I found that
compiling various rust applications with the default of codegen-units=16, it produced nondeterministic variations in resulting binaries. e.g.
pop-os/launcher#230

IMHO, the default settings should be made to behave deterministically. Either make codegen-units=1 the default or ensure that codegen-units=16 can produce deterministic results.

In the unlikely event that this is not possible, it should clearly be documented how to achieve reproducible builds, not only in
https://doc.rust-lang.org/rustc/codegen-options/index.html#codegen-units

This is about rust1.79 but started much earlier. At least with 1.73

[kpcyrd]

Arch Linux is currently also having issues with the rustc and rustdoc binaries being unreproducible:

https://web.archive.org/web/20240805092434/https://reproducible.archlinux.org/api/v0/builds/649842/diffoscope

[saethlin]

IMHO, the default settings should be made to behave deterministically. Either make codegen-units=1 the default or ensure that codegen-units=16 can produce deterministic results.

CGU partitioning is very deliberately designed to be deterministic.

[saethlin]

@bmwiedemann It's unclear to me exactly what you are trying to report. In llvm/llvm-project#72206 you just generally refer to " different binaries in every build, even when trying to run builds as similar as possible."

But in pop-os/launcher#230 you are pointing out that changing the build environment but not the build settings changes the build. That has a fair chance of being a different rustc bug than the issue you filed on LLVM.

So I do not think it is accurate to say this started in 1.73.

@kpcyrd What are you trying to report? All you linked is a diff from a system that is completely unfamiliar, with no enlightening context. I can't tell whether you're just leaving a "seeing this too" comment (which is entirely unnecessary), or if you're trying to provide additional data for debugging, or if you're trying to report a third variation.

[workingjubilee]

Based on the response to @bmwiedemann's original report, binaries are reproducible based on codegen units if you use the same number of codegen units. Thus, closing as already-completed.

@kpcyrd "I'm having trouble making these binaries reproducible" doesn't distinguish your report from any other extant A-reproducibility issue on the tracker, so please open a new issue with an actual report. Build logs are not an issue report. Even if they were, each distribution uses a wildly different format so it is not possible for any lessons to be learned and applied from distro to distro. The day SUSE and Arch and Debian and everyone else standardize on using exactly the same tooling with exactly the same configuration is nowhere in sight.