25% regression in serde compile time

[dtolnay]

I narrowed this down to 1699293...c284f88. To reproduce:

rustup update nightly-2017-12-24 nightly-2017-12-25
git clone https://github.com/serde-rs/serde
cd serde/serde
git checkout v1.0.27
cargo update
cargo clean && time cargo +nightly-2017-12-24 build
cargo clean && time cargo +nightly-2017-12-25 build

For me nightly-2017-12-24 takes 7.4 seconds and nightly-2017-12-25 takes 9.2 seconds.

[arielb1]

A not-too-deep check reveals that trait selection might have slowed down. Seriously? (I didn't track it down to that commit, should probably bisect it to a single commit in that range).

[nikomatsakis]

triage: P-high

Good to know what happened.

[nikomatsakis]

I'll try to bisect.

[nikomatsakis]

Curious. I was not able to reproduce when building those commits from scratch, though I did see the perf difference with the nightly builds. I'll double-check. It could have to do with some of my config.toml options (e.g., debuginfo or debug-assertions), but that seems a bit surprising.

[nikomatsakis]

I tried with no config.toml at all and still got basically the same timing for those two commits: 11.6s in both cases.

[nikomatsakis]

cc @rust-lang/compiler -- can someone else try to reproduce from source?

[Zoxc]

I also built these two revisions from source. Both of my builds have the same timing of ~40 seconds for cargo build -j1. This is the same result as nightly-2017-12-25, while nightly-2017-12-24 builds in ~30 seconds.

[nikomatsakis]

@Zoxc thanks! Not entirely sure what to make of that.

[nikomatsakis]

cc @rust-lang/infra -- can we try using the PR-by-PR builds to narrow this down further or gain some insight? I'm not sure how to access those.