Check that a box expression's type is Sized

[jesyspa]

This resolves issue 87935.

This makes E0161 (move from an unsized rvalue) much less common. I've replaced the test to use this case, when a boxed dyn trait is passed by value, but that isn't an error when unsized_locals is enabled. I think it may be possible to get rid of E0161 entirely by checking that case earlier, but I'm not sure if that's desirable?

[rust-highfive]

r? @jackh726

(rust-highfive has picked a reviewer for you, use r? to override)

[jesyspa]

There's another possible implementation, where we issue E0161 for unsized box expressions even when unsized_locals are disabled. That approach keeps the checking in rustc_mir, but it currently emits duplicate errors for cases like rust fn foo(x: Box<[i32]>) { box *x; }

Given that the implementation in this PR gives a clearer error message, I think it's probably the better option.

[jackh726]

Why change the E0161 output?

[jackh726]

Thanks! Going to @bors rollup=never in case there are any subtle changes here that need to be bisected.

@bors r+

[jackh726]

@bors r+ rollup=never

[bors]

📌 Commit c75a930 has been approved by jackh726

[nbdd0121]

I wonder if it's worth a new obligation cause code if we want to remove box syntax eventually anyway 🤔

[jackh726]

Those are cheap :)

[bors]

⌛ Testing commit c75a930 with merge 1e3d632...

[bors]

☀️ Test successful - checks-actions
Approved by: jackh726
Pushing 1e3d632 to master...

[jesyspa]

Thanks for the review!

I wonder if it's worth a new obligation cause code if we want to remove box syntax eventually anyway thinking

I don't think this will complicate it, since check_box can be removed in its entirety when the syntax goes away. The cost should only be paid by code that uses box expressions.