Correctly register lifetime constraints from opaque types from canonical queries called from nll #123035
Conversation
…cal queries called from nll
|
r? @Nadrieril rustbot has assigned @Nadrieril. Use |
rustbot
added
S-waiting-on-review
| _ => { | ||
| if a != b { | ||
| span_bug!( | ||
| cause.span(), | ||
| "self-referential opaque type can only differ in lifetime parameters, but got {a} != {b}" | ||
| ) | ||
| } | ||
| } |
There was a problem hiding this comment.
This is super hacky, but at least it's a high signal ICE instead of some hard-to-debug issues. We also can't work on the FIXME before we get rid of DefiningAnchor, so I think we should just roll with this for now and address it later.
| for (key, hidden) in opaques { | ||
| obligations.extend( | ||
| self.at(cause, param_env) | ||
| .eq( | ||
| DefineOpaqueTypes::Yes, | ||
| Ty::new_opaque(self.tcx, key.def_id.into(), key.args), | ||
| hidden, | ||
| )? | ||
| .obligations, | ||
| ) | ||
| } |
There was a problem hiding this comment.
all non-nll query callers still need their opaque types registered. After #122077 this will become a simple register_hidden_type call
| } | ||
| } else { | ||
| obligations.extend( | ||
| self.handle_opaque_type( |
There was a problem hiding this comment.
After #122077 this will become a register_hidden_type
|
hmm... I am only now realizing that So maybe this isn't the way to go after all. But it also feels very hacky to just wrap this in another |
oli-obk
changed the title
|
Yea no, this doesn't fix everything and is horrible on top of it. I found a simpler solution that I'll do directly in #122077 |
Unblocks #122077
in #122077 we had the fun situation where a
Opaque<'1, '2>got a hidden typeOpaque<'3, '4>registered (before that situation just didn't occur). Sincequery_response_instantiation_guessalways equated opaque types with their hidden type (to let the type relation logic reregister the hidden type if applicable), during #122077 that started registering lifetime relations in the InferCtxt. Nll needs to take them out and tell borrowck about them, but that never happened, so borrowck started ICEing.My next step was to stop equating opaque types to their hidden type, as we now always know that we can just register the hidden type in the query caller's InferCtxt. But then the opaque type's lifetimes never got resolved, even though the hidden type's lifetimes got resolved. That then errors out because there's an opaque type whose generic params do not refer to generic params of the item currently being borrowcked.
So my solution for now is to explicitly register lifetime relations in nll.