Forgone caching in cycles caused much overflow in trait solving

[pnkfelix]

We believe PR #60444 injected a number of regressions of the form "overflow evaluating requirement [...]".

(PR #60444 also injected a compile-time performance regression; that arguably distinct issue is tracked in issue #60846.)

This issue is meant to collect the cases where there was some sort of "overflow evaluating requirement" that was injected by PR #60444. (It is possible that many (or even all) such cases are resolved by follow-up PR #61754. I'll be beta-nominating that PR on its own so that we can evaluate that question in the compiler-team meeting.)

Update: adding list of cases to investigate:

• regression: overflow evaluating, Send/Sync? #61472 -- resolved by PR create a "provisional cache" to restore performance in the case of cycles #61754
• rust-analyzer (see comment below)
• Rustdoc recursion limit issue #62059
• ICE: Rust spins when referencing associated types in where clause #62430

[pnkfelix]

In case its not clear from my description: It is possible that this issue is entirely resolved by PR #61754, at least on the nightly channel. This umbrella issue is meant to be a place where I can make a list of all the relevant regressions and evaluate whether PR #61754 fixed them, which may inform a future decision on whether or not to beta-backport that PR.

[pnkfelix]

Hmm, I may have been overly hasty in creating a fresh umbrella issue to track "all" the regressions here. Now that I look again at the list, and at the cross-references to PR #60444, it seems like issue #61472 may be the only issue on the Rust repo here? (I do believe that rust-lang/rust-analyzer#1283 may be another instance of this issue, but that isn't an issue on the rust-lang/rust repo, so it seems a little silly to try to include it on this list...)

[matklad]

I've tried rustc 1.37.0-nightly (04a3dd8a8 2019-06-18) and it still takes unreasonable time to compile rust-analyzer (this commit specifically). I don't know if the compilation succeeds or fails: it has been going on for 10 minutes already.

[matklad]

It overflowed half an hour later

[pnkfelix]

Note: The compile-time performance regression was/is meant to be tracked by issue #60846. (And maybe we need to open a fresh such issue, since #60846 was closed by PR #61754 )

This issue is just meant to track "overflow evaluating requirement" diagnostics that we believe were injected by #60444.

[pnkfelix]

triage: P-high, removing nomination.

[pnkfelix]

Unassigning self; I still think this is important, but I'm not going to be able to address it, at least not until I return from leave in September.

[nagisa]

Pre-triage visit: Doubt this will be looked at before vacation spree ends, but if you want to give it a shot, by all means do so.

[pnkfelix]

visiting for triage. Downgrading to P-medium and re-assigning back to self.