Slow coherence checking and type checking.

[coolcatcoder]

I was working on creating a crate and encountered some unexpectedly long compile times.
I don't have a minimal reproduction, but I have expanded my macros, and copied the code into a playground.

It occurs on all channels. It is still pretty bad even with the parallel front-end enabled. According to time-passes, the biggest culprits were coherence checking and type checking.

time:   6.038; rss:   80MB ->  272MB ( +192MB)  coherence_checking
time:   6.084; rss:   80MB ->  274MB ( +195MB)  type_check_crate

My use case is clearly atypical, so I will understand if you consider this frivolous and close it. (Although, when I searched to see if my issue had already been reported, I did notice submitted situations that were arguably even less likely to be encountered by anyone.)

[moxian]

Bisection: While this code has always been slow, it took an extra ~30% comptime hit in nightly-2024-02-15.

//  nightly-2024-02-15
time:   0.006; rss:   34MB ->   38MB (   +4MB)  coherence_checking
time:  12.322; rss:   38MB ->  236MB ( +198MB)  wf_checking
time:  12.400; rss:   31MB ->  240MB ( +208MB)  type_check_crate

// nightly-2024-02-14
time:   0.006; rss:   34MB ->   38MB (   +4MB)  coherence_checking
time:   9.725; rss:   38MB ->  220MB ( +182MB)  wf_checking
time:   9.802; rss:   31MB ->  224MB ( +192MB)  type_check_crate

Unfortunately that nightly has several type-level PRs landing, making it hard to pinpoint regression precisely, but #118882 sounds most likely

cargo-bisect-rustc output

---

Regression in nightly-2024-02-15

---

converted 2024-02-14 to a84bb95
converted 2024-02-15 to ee9c7c9
get_commits_between returning commits, len: 9

• commit[0] 2024-02-13: Auto merge of Rollup of 8 pull requests #121036 - matthiaskrgr:rollup-ul05q8e, r=matthiaskrgr
• commit[1] 2024-02-14: Auto merge of Ignore own item bounds in parent alias types in for_each_item_bound #120942 - compiler-errors:deep-assoc-hang, r=lcnr
• commit[2] 2024-02-14: Auto merge of Rollup of 8 pull requests #121055 - matthiaskrgr:rollup-bzn5sda, r=matthiaskrgr
• commit[3] 2024-02-14: Auto merge of cargo update #120454 - clubby789:cargo-update, r=Nilstrieb
• commit[4] 2024-02-14: Auto merge of Fully stop using the HIR in trait impl checks #121018 - oli-obk:impl_unsafety, r=TaKO8Ki
• commit[5] 2024-02-14: Auto merge of Optimize away poison guards when std is built with panic=abort #100603 - tmandry:zst-guards, r=dtolnay
• commit[6] 2024-02-14: Auto merge of Rollup of 13 pull requests #121078 - oli-obk:rollup-p11zsav, r=oli-obk
• commit[7] 2024-02-14: Auto merge of Rollup of 6 pull requests #121086 - GuillaumeGomez:rollup-y82fs9y, r=GuillaumeGomez
• commit[8] 2024-02-14: Auto merge of Continue compilation after check_mod_type_wf errors #120847 - oli-obk:track_errors9, r=compiler-errors

Edit: and also another ~20% hit several weeks prior in nightly-2024-01-18. Likely #120019

// nightly-2024-01-17
time:   0.007; rss:   34MB ->   38MB (   +4MB)  coherence_checking
time:   8.118; rss:   38MB ->  206MB ( +168MB)  wf_checking
time:   8.207; rss:   32MB ->  210MB ( +178MB)  type_check_crate

cargo-bisect-rustc out

********************************************************************************
Regression in nightly-2024-01-18
********************************************************************************
converted 2024-01-17 to 098d4fd74c078b12bfc2e9438a2a04bc18b393bc
converted 2024-01-18 to 6ae4cfbbb080cafea7f6be48ce47678ee057352c
get_commits_between returning commits, len: 9
  commit[0] 2024-01-16: Auto merge of #119977 - Mark-Simulacrum:defid-cache, r=cjgillot
  commit[1] 2024-01-17: Auto merge of #120036 - weihanglo:update-cargo, r=weihanglo
  commit[2] 2024-01-17: Auto merge of #120019 - lcnr:fn-wf, r=BoxyUwU
  commit[3] 2024-01-17: Auto merge of #119651 - novafacing:proc_macro_c_str_literals, r=Amanieu
  commit[4] 2024-01-17: Auto merge of #119922 - nnethercote:fix-Diag-code-is_lint, r=oli-obk
  commit[5] 2024-01-17: Auto merge of #119111 - michaelwoerister:measureme-11, r=Mark-Simulacrum
  commit[6] 2024-01-17: Auto merge of #119670 - cjgillot:gvn-arithmetic, r=oli-obk
  commit[7] 2024-01-17: Auto merge of #119930 - Urgau:check-cfg-empty-values-means-empty, r=petrochenkov
  commit[8] 2024-01-17: Auto merge of #118708 - davidtwco:target-tier-assembly-test, r=Mark-Simulacrum

@rustbot label: +T-compiler +T-types +A-coherence +I-compiletime -needs-triage +A-trait-system

[moxian]

Smaller:

use core::{marker::PhantomData, ops::Deref};

struct DerefOnly<T: ?Sized>(PhantomData<T>);

impl<T: Deref> Deref for DerefOnly<T> {
    type Target = DerefOnly<T::Target>;

    fn deref(&self) -> &Self::Target {
        &DerefOnly(PhantomData)
    }
}

struct Empty;

impl DerefOnly<Empty> {
    fn whatever(&self) {}
}

// copy-paste the function a couple of times if your pc is too fast to notice
fn foo_1() {
    DerefOnly(PhantomData::<
  &&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&Empty,
        >).whatever();
}
fn foo_2() {
    DerefOnly(PhantomData::<
  &&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&Empty,
        >).whatever();
}

fn main() {}

What's happening is we create a let x = DerefOnly{0: PhantomData<&&..&&Empty>}, and try to call a method .whatever() on it. The type we have at hand doesn't have said method but the type x.deref().deref()......deref() does. And it's this deref chain that's taking us some time to traverse. I think... I'm by no means an expert.