Add a regression test for #44415

[dsprenkels]

This PR adds a regression test for issue #44415.

Fixes #44415.

[rust-highfive]

r? @alexcrichton

(rust_highfive has picked a reviewer for you, use r? to override)

[dsprenkels]

Turns out I accidentally closed this PR while cleaning up my fork.

[alexcrichton]

Thanks @dsprenkels!

For us though // error-pattern is pretty old at this point, perhaps //~ ERROR: ... could be used instead? That way we can assert both the error message and the location of the error message!

[dsprenkels]

Oh yeah, I forgot about that. I will fix it accordingly.

[dsprenkels]

Hmmm. It turns out harder than expected, because the error does not trigger in the file itself, but in mem.rs.

error[E0391]: unsupported cyclic reference between types/traits detected
   --> /scratch/dsprenkels/rust/src/libcore/mem.rs:315:14
    |
315 |     unsafe { intrinsics::size_of::<T>() }
    |              ^^^^^^^^^^^^^^^^^^^^^^^^^^ cyclic reference
    |
note: the cycle begins when computing layout of `Foo`...
   --> /scratch/dsprenkels/rust/src/libcore/mem.rs:315:14
    |
315 |     unsafe { intrinsics::size_of::<T>() }
    |              ^^^^^^^^^^^^^^^^^^^^^^^^^^
note: ...which then requires const-evaluating `Foo::{{initializer}}`...
   --> issue-44415.rs:14:17
    |
14  |     bytes: [u8; std::mem::size_of::<Foo>()],
    |                 ^^^^^^^^^^^^^^^^^^^^^^^^^^
    = note: ...which then again requires computing layout of `Foo`, completing the cycle.

error: aborting due to previous error

So the //~ ERROR [...] pattern does not trigger on the correct line in issue-44415.rs. intrinsics::size_of<T>() is an intrinsic and can probably not be defined inline.

Is it possible to specify an //~ ERROR [...] pattern for an error that occurs in another file? In any case, it looks like const-size_of-cycle.rs may have had the same problem. (The PR has been adapted from this test.) This test uses // error-pattern. Shall I keep using // error-pattern?

[kennytm]

Maybe try to make it a UI test instead of compile-fail test?

[dsprenkels]

Yes I could do this. However, we would have to update the test every time src/libcore/mem.rs changes.

Instead, I chose to redefine the size_of function inside the test itself, such that we actually can trigger on the error (because the file is correct).

[dsprenkels]

Btw. I propose that if we adopt any of the solutions that are not // error-pattern, that we will also update const-size_of-cycle.rs.

[emilyalbini]

@alexcrichton ping from triage!

[alexcrichton]

@bors: r+

Thanks!

[bors]

📌 Commit f641ac6 has been approved by alexcrichton