Alternate fix for issue #24085 #24404
Conversation
Insert an inference variable before performing trait selection for builtin traits. This prevents overly-restrictive region constraints that lead to later inference failures.
|
Hmm, I definitely prefer this version. I was just planning to checkout your other branch and dig more deeply and try to make sure I understood what was happening! I'll do that anyway. |
|
OK, so I dug into this test. I think the bug has to do with one of the darker corners of region inference (the "givens" field, specifically). I have a fix that extends the current hack, but it's really getting high time to rewrite the region inference code -- and in particular the way it handles closures. (I've also been sketching a design of that that I will try to write-up and prototype, but that's a separate thing...) |
|
@nikomatsakis @bkoropoff how do you feel about this patch today? |
|
I think we should close this in favor of #24527. I've been criminally negligent in not benchmarking and rebasing that. I'll try to get to that today. Sorry @bkoropoff :( |
This implements my suggestion in #24360 of throwing inference variables at the problem instead. It works, though at the cost of making inference do more work.
r? @nikomatsakis