same functions are not merged in library

[hellow554]

I tried this code:

pub fn a(c: char) -> bool {
    c == 's' || c == 'm' || c == 'h' || c == 'd' || c == 'w'
}

pub fn b(c: char) -> bool {
    matches!(c, 's' | 'm' | 'h' | 'd' | 'w')
}

When compiled with the latest stable version and look at the objdump, I see a combined version of a and b, because they are optimized to the same instructions and then merged in the library.
On nightly (already bisected) two copies are stored and it takes the double amount of space (3.9kIB vs 4.1kIB).

Here's a godbolt link: https://rust.godbolt.org/z/zWovojdEE

Version it worked on

It most recently worked on: 1.57.0
Doesn't work since: 63cc2bb cc #88243 @nikic

@rustbot modify labels: +regression-from-stable-to-nightly +I-heavy

[camelid]

It most recently worked on: 1.57.0
Doesn't work since: 63cc2bb cc #88243 @nikic

I'm confused: #88243 landed in 1.57.0.

[camelid]

Marking as regression-untriaged for now until we figure out which version this actually affects. That PR is in the latest stable (1.57.0), not nightly.

[nikic]

@camelid The change was reverted for the release.

The issue sounds like MergeFuncs doesn't run -- looking at https://github.com/llvm/llvm-project/blob/f178a05f220403f2a9d73c7640bfcc7dc2d7be72/llvm/lib/Passes/PassBuilderPipelines.cpp#L1194 I seem to have misscheduled the pass due to some unfortunate code structure: The passes are added in the middle of OptimizePM construction, but the OptimizePM is only actually scheduled afterwards, which means that MergeFuncs will run close to the start of the module pipeline, not the end. The same seems to have happened to IR outlining and hot-cold splitting.

[nikic]

Test: llvm/llvm-project@5b94037
Fix: https://reviews.llvm.org/D115098, llvm/llvm-project@ae7f468

[apiraino]

Assigning priority as discussed in the Zulip thread of the Prioritization Working Group.

@rustbot label -I-prioritize +P-medium

[MSxDOS]

This is fixed on nightly, apparently by #91657

[camelid]

Ok, closing then. That PR also added a test for this issue.