resolve: mark it undetermined if single import is not has any bindings

[bvanjoi]

• Fixes ICE: inconsistent resolution for an import #124490
• Fixes ICE:assertion failed: import.imported_module.get().is_none() #125013

This issue arises from incorrect resolution updates, for example:

mod a {
    pub mod b {
        pub mod c {}
    }
}

use a::*;

use b::c;
use c as b;

fn main() {}

1. In the first loop, binding (root, b) is refer to root::a::b due to use a::*.
   1. However, binding (root, c) isn't defined by use b::c during this stage because use c as b falls under the single_imports of (root, b), where the imported_module hasn't been computed yet. This results in marking the path_res for b as Indeterminate.
   2. Then, the imported_module for use c as b will be recorded.
2. In the second loop, use b::c will be processed again:
   1. Firstly, it attempts to find the path_res for (root, b).
   2. It will iterate through the single_imports of use b::c, encounter use c as b, attempt to resolve c in root, and ultimately return Err(Undetermined), thus passing the iterator.
   3. Use the binding (root, b) -> root::a::b introduced by use a::* and ultimately return root::a::b as the path_res of b.
   4. Then define the binding (root, c) -> root::a::b::c.
3. Then process use c as b, update the resolution for (root, b) to refer to root::a::b::c, ultimately causing inconsistency.

In my view, step 2.2 has an issue where it should exit early, similar to the behavior when there's no imported_module. Therefore, I've added an attribute called indeterminate to ImportData. This will help us handle only those single imports that have at least one determined binding.

r? @petrochenkov

[bvanjoi]

#124840 (comment)

CI failed, so mark it as a draft.

reduced:

mod a {
    pub(crate) use crate::S;
}
mod b {
    pub struct S;
}
use b::*;
use self::a::S;

fn main() {}

[bvanjoi]

Update: I've changed that it now only returns Undetermined when resolving modules.

I'm not sure if there's a better solution or potential edge cases. Regardless, since CI is passing, mark it as ready for review.

[petrochenkov]

2. It will iterate through the single_imports of use b::c, encounter use c as b, attempt to resolve c in root, and ultimately return Err(Undetermined), thus passing the iterator.

I don't understand, both Err(Undetermined) and the new case introduced in this PR result in

return Err((Undetermined, Weak::No))

and neither of them is "passing the iterator" (if I understand it correctly).

Does resolve_ident_in_module(...) returns Err(Determined) or Ok(binding) maybe, even if single_import.indeterminate is true?
It's not clear to me how this can happen.

[bvanjoi]

Does resolve_ident_in_module(...) returns Err(Determined)

Yep, it will return Err(Determined) for resolve_ident_in_module(c, root). It might be a typo..

[bvanjoi]

@rustbot ready

[petrochenkov]

The PR is now looking nice after the cleanup, but I still don't understand why it works :D

Why is "all bindings undetermined" able to correctly reject something that resolve_ident_in_module is not?
I would expect resolve_ident_in_module to return Undetermined without this additional intervention, that feels pretty arbitrary (especially with the module().is_some() condition).
@rustbot author

[bvanjoi]

The reason why resolve_ident_in_module(c, root) returns Err(Determined) is because it skips the iteration of single_imports. More specifically, the single_import.vis value is None because it was already taken at the entry point. This results in hitting a continue.

[bvanjoi]

Also, I tried moving the single_import.imported_module.get() condition before single_import.vis.get(), but ./x.py build failed to compile. Therefore, I've decided not to solve this problem in this manner. :E

[petrochenkov]

I suspect that resolver may be stuck on some valid code after this change, let's check.
@bors try

[petrochenkov]

r=me with the comment #124840 (comment).
@rustbot author

[bvanjoi]

@rustbot ready

[petrochenkov]

Thanks!
@bors r+

[bors]

📌 Commit f67a0eb has been approved by petrochenkov

It is now in the queue for this repository.

[fmease]

bors sleepy @bors r-