Regression in diesel with 2020-01-15 nightly: overflow evaluating requirement

[jebrosen]

rustc version: rustc 1.42.0-nightly (3291ae339 2020-01-15)

.../code/diesel/diesel (master=) ✖ RUSTFLAGS="--cap-lints=warn" cargo +nightly check
   Compiling proc-macro2 v1.0.7
   Compiling unicode-xid v0.2.0
   Compiling syn v1.0.13
   Compiling byteorder v1.3.2
   Compiling quote v1.0.2
   Compiling diesel_derives v1.4.1 (/home/jeb/code/diesel/diesel_derives)
    Checking diesel v1.4.3 (/home/jeb/code/diesel/diesel)
warning: use of deprecated item 'std::error::Error::description': use the Display impl or to_string()
(... trimmed many instances of this warning...)
warning: use of deprecated item 'std::error::Error::description': use the Display impl or to_string()
   --> diesel/src/migration/errors.rs:108:43
    |
108 |         write!(f, "Failed with: {}", self.description())
    |                                           ^^^^^^^^^^^

error[E0275]: overflow evaluating the requirement `<Self as query_dsl::limit_dsl::LimitDsl>::Output`

error: aborting due to previous error

For more information about this error, try `rustc --explain E0275`.
error: could not compile `diesel`.

To learn more, run the command again with --verbose.

Notably, there is no information about where the error actually is. rustc 1.42.0-nightly (31dd4f4ac 2020-01-13) was fine. I also see a passing CI run on diesel for nightly on commit 8a87b945b27.

EDIT: I did run this test on the master branch of diesel, but I first discovered the problem in diesel 1.4.3.

[RalfJung]

The same code (diesel 1.4.3) builds fine in stable, so marking as regression.

[weiznich]

There is a bisec in the linked diesel issue.

Bisect was able to find 6d0bb91 as the culprit, but it's a rollup.

[oli-obk]

Likely #67914 cc @Aaron1011

[Aaron1011]

It looks like we need to run normalization, not just fulfillment, in canonical mode.

I'll open a PR later today

[Aaron1011]

Unfortunately, normalization and projection don't have the equivalent of SelectionContext's "canonical mode" - they always report overflow errors. I'm somewhat hesitant to add such a mode, as I think this would be the only use-case.

The root issue here is that const-prop is very much "best effort" - it's okay to bail out if we're uncertain about something. However, trait selection and projection are very much not "best effort" - if we can't prove something, we tend to need to raise an error to avoid unsoundness. A lot of code is designed around this principle (e.g. normalize does not return a Result).