Long delay printing errors with metadata

[erickt]

@tmandry suggested filing a new bug for #136516 (comment) since this is now reproducing on nightly.

On Fuchsia we've noticed a pretty significant complication time regression in Fuchsia with our starnix binary when reporting an errors, where sometimes it can take 10+ hours for Rust to exit. Unfortunately the starnix binary itself is 113K lines of code, and it depends on a whole swath of other dependencies, so it's taken about 3 months of time to decouple the code from Fuchsia and reduce it to this gist. With that, it takes about 5 minutes 30 seconds for Rust to exit with nightly-2026-01-08 on my linux workstation.

This seems to be related to .rmeta somehow, since removing --emit metadata causes Rust to exit after only a second or so.

@tmandry suggested using -Z time-passes -Z self-profile. I had to modify measureme to up the MAX_USER_VIRTUAL_STRING_ID to 100 billion, which got me a profile. Here's the summarize output. I've also attached the , but unfortunately it seems invalid. Perhaps it didn't work correctly with my patch to measureme? I also tried crox, but it produced the same odd graph as flamegraph. I'll keep working on it though.

Meta

rustc --version --verbose:

rustc 1.94.0-nightly (31cd367b9 2026-01-08)

Backtrace

See this gist.

[SpriteOvO]

@rustbot label +I-compiletime +perf-regression +T-compiler

[saethlin]

Is there anything helpful in a flamegraph rendered based on perf?

[erickt]

Good suggestion! I had much better luck with that. I haven't figured out how to get perf to generate an svg, but I did get this from perf script report flamegraph. Unfortunately I don't see an obvious red flag from it though.

[erickt]

I figured out how to get an SVG, in case it's easier to dig through. I also grabbed the first hundred or so lines of perf report -g in case that's easier to dig through.

[fmease]

Does passing rustc flag -Znext-solver have any effect?

[lqd]

Does passing rustc flag -Znext-solver have any effect?

AFAICT this repro looks to fail immediately with the new solver.

[erickt]

I reduced this down quite a bit more in this gist, now only 354 lines total, and now takes about a minute and a half to error out . It also was small enough to work with -Z self-profile without hacks, with this summarize output https://gist.github.com/erickt/d032ea57f43db949cb4575e043cc7a72, but I'm still getting 99% of the time being spent in <unknown> so it's still not that useful.

I can confirm it also fails immediately with -Znext-solver.

Also, one thing interesting to note is that the error goes away if I switch the edition to 2015, but occurs with 2021 and 2024.