Don't over-constrain projections in generic method signatures #92728
Conversation
rustbot
added
the
T-compiler
rust-highfive
added
the
S-waiting-on-review
| self.depth, | ||
| &mut self.obligations, | ||
| ); | ||
| let normalized_ty = if self.selcx.normalization_mode.eager_inference_replacement { |
There was a problem hiding this comment.
This is the same as to what is done in #90887, but puts the flag on SelectionContext instead of AssocTypeNormalizer
| let any_instantiations = infcx.any_instantiations(&snapshot); | ||
|
|
||
| if any_instantiations && !selcx.normalization_mode.allow_infer_constraint_during_projection |
There was a problem hiding this comment.
Now that I think about it, this could probably be done during candidate assembly.
| let mut selcx = traits::SelectionContext::new(self); | ||
| selcx.normalization_mode.eager_inference_replacement = false; | ||
| let mut normalize_obligations = Vec::new(); | ||
|
|
There was a problem hiding this comment.
The problem with doing normalizing like this, is that maybe one type can be normalized, but the other can't. We probably should allow coercing by relating the projection itself.
Another problem with doing it like this: what happens if the result of a normalization is itself a projection. You could imagine that we might want to be able to coerce to any types in the "chain" of normalizations from either side. But we can't just relate types, because we also allow e.g. unsizing. Otherwise, lazy norm might be able to solve this.
Imo, the other half of this PR (not constraining inference vars when confirming a method call) is "more correct" than this one. There might be some version of this change that can be more conservative, where we could just require a ::<T, U> to be specified.
| _ => a, | ||
| } | ||
| } | ||
| _ => traits::normalize_with_depth_to( |
There was a problem hiding this comment.
Problem here: We don't have a way to "project_and_unify_type" for generalized type (essentially a Relate instead of a Fold). This hack fails as soon as we wrap the projection in an adt, for example.
| @@ -0,0 +1,32 @@ | |||
| // check-pass | |||
There was a problem hiding this comment.
This is the relevant test. (Don't mind the no_core MCVE; I made this because otherwise the logs where too long)
|
Tried to rebase, but my fix no longer works. Have to think about this more I guess. |
|
☔ The latest upstream changes (presumably #95333) made this pull request unmergeable. Please resolve the merge conflicts. |
jackh726
added
S-experimental
rustbot
added
the
S-inactive
Fixes #91762
This is likely not the "right" way to do this. Just opening for thoughts
r? @nikomatsakis