Diesel new coherence failure: conflicting implementations of trait

[weiznich]

Code

I tried this code:

git clone https://github.com/diesel-rs/diesel/
cd diesel
git checkout b170af753cc3c6774f578aae1081711a09dfa2ac
cargo +nightly check -p diesel -F postgres -F 64-column-tables

I expected to see this happen: Code compiles fine

Instead, this happened: Code produces the following errors (Really long error, so I put it into a gist)

Version it worked on

It most recently worked on: Beta channel (1.83)

Version with regression

rustc --version --verbose:

rustc 1.84.0-nightly (da935398d 2024-10-19)
binary: rustc
commit-hash: da935398d582344c5b7689bd6632d8ec01b0c988
commit-date: 2024-10-19
host: x86_64-unknown-linux-gnu
release: 1.84.0-nightly
LLVM version: 19.1.1

Rustc-bisect points to a0c2aba, so cc @lcnr @compiler-errors (I'm sorry to come with the next issue there)

@rustbot modify labels: +regression-from-beta-to-nightly -regression-untriaged

[compiler-errors]

(I'm sorry to come with the next issue there)

It's ok! We specifically chose to wait until after the beta release to stabilize the new solver in coherence, to get plenty of time and exposure to issues like this. Thanks for reporting :D

[weiznich]

@Urgau I've already bisected it to a0c2aba (as written in the initial issue)

I also tried some things locally and interestingly this does not happen without the 64-column-tables feature. This feature only increases the tuple size diesel implements it's traits for, so it's quite surprising that rustc now reports that some traits conflict with each other while the same implementations work for smaller tuple sizes.

Additionally: It also might be meaningful to test with the 128-column-tables feature enabled afterwards to see if a potential fix works there as well.

[compiler-errors]

I've already bisected it to a0c2aba (as written in the initial issue)

That bisection here is not super useful. It's probably better to bisect it to when it started failing in the new trait solver specifically.

I've already done that, and it bisects somewhat unsurprisingly to #130617. cc @lcnr, we probably should adjust the bail heuristic here to account for more infer vars; it looks like diesel uses lots of those.