intentional incompleteness

[lcnr]

Trait solving has to be complete during coherence to be sound.

We however have a lot of stable code which depends on incompleteness. The new solver is intentionally incomplete when run outside of coherence.

try_merge_candidates

For trait and project goals, we fall back to only considering ParamEnv and AliasBound candidates if trying to merge all candidates failed.

This is similar to the behavior of the old solver. TODO: Do we incorrectly guide GAT substs inference rn, I think we may be missing a check there?

Why do we have to be incomplete here

TODO

AliasRelate

We prefer to relate the generic arguments of aliases over normalization in case of ambiguity.

Why do we have to be incomplete here

TODO

[lcnr]

Relating generic arguments of aliases together with lazy norm results in the following test to pass:

https://github.com/rust-lang/rust/blob/be72f2587c91579406117f99fa332383d66b7dcd/tests/ui/traits/new-solver/alias_eq_substs_eq_not_intercrate.rs#L10-L22

[lcnr]

split into

• Incompleteness via bidirectional normalizes-to #25
• Incompleteness in preferring subst-relate candidate in alias-relate goals #26
• incompleteness when merging trait candidates #45