cargo miri test no longer correctly detects tests

[apoelstra]

Since rust-lang/rust#53410 the #[test] attribute has become a macro and is not visible in the HIR as an attribute. This means that the code detecting test cases never triggers so any #[test]-labeled functions are ignored.

CI passes because the effect of this is that the tests simply aren't run, so of course they don't fail :).

I am investigating the effects of the compiler PR. Empirically it looks like functions that used to have a test attribute in the HIR now have a main attribute, but I'm not sure if that's happening on purpose and if that's the correct way for us to detect test cases now.

[apoelstra]

After running with #482 I see that unit tests in libraries don't have the main attribute on them. In fact they have no attributes at all :(

[RalfJung]

Somehow the test binary has to work, find the tests, run them. How does it do that?

[apoelstra]

It looks like it looks for the rustc_test_marker attribute but for some reason in my debug dumps I didn't see this.Maybe it's on a different item than the function itself or something? Maybe it's lost going from AST -> HIR (doesn't sound plausible). I will investigate.

[apoelstra]

Ah, yep, the rustc_test_marker attribute is applied to a new item with type that looks like Const(type(test::TestDescAndFn), BodyId { node_id: NodeId(16081) }), rather than something like Fn(FnDecl).

[apoelstra]

The TestDescAndFn type is defined here: https://github.com/rust-lang/rust/blob/master/src/libtest/lib.rs#L239-L243

Its second field is an enum, which for ordinary test cases is equal to TestFn::StaticTestFn(func), and we want to run the MIR of func. But I don't know how to extract it from any of these data structures without nesting a dozen if-statements pulling everything apart.

[RalfJung]

rust-lang/rust#56243 makes some progress towards this: We still run into rust-lang/rust#54957, but other than that, we can actually run the libtest harness in miri.

[Gankra]

Can you just disable the btree-breaking check for a bit?

[RalfJung]

That would mean disabling Stacked Borrows. One can run with -Zmiri-disable-validation to achieve that, but then also miri checks way fewer things than it should check.

I can try disabling it inside the BTree module (there's a whiteliist), but that could fail to work because it means references created by BTree are not properly tracked.