| //@ revisions:rpass1 rpass2 | |
| //@ ignore-backends: gcc | |
| // This test case makes sure re-order the methods in a vtable will | |
| // trigger recompilation of codegen units that instantiate it. | |
| // | |
| // See https://github.com/rust-lang/rust/issues/89598 | |
| trait Foo { | |
| #[cfg(rpass1)] | |
| fn method1(&self) -> u32; | |
| fn method2(&self) -> u32; | |
| #[cfg(rpass2)] | |
| fn method1(&self) -> u32; | |
| } | |
| impl Foo for u32 { | |
| fn method1(&self) -> u32 { 17 } | |
| fn method2(&self) -> u32 { 42 } | |
| } | |
| fn main() { | |
| // Before #89598 was fixed, the vtable allocation would be cached during | |
| // a MIR optimization pass and then the codegen pass for the main object | |
| // file would not register a dependency on it (because of the missing | |
| // dep-tracking). | |
| // | |
| // In the rpass2 session, the main object file would not be re-compiled, | |
| // thus the mod1::foo(x) call would pass in an outdated vtable, while the | |
| // mod1 object would expect the new, re-ordered vtable, resulting in a | |
| // call to the wrong method. | |
| let x: &dyn Foo = &0u32; | |
| assert_eq!(mod1::foo(x), 17); | |
| } | |
| mod mod1 { | |
| pub(super) fn foo(x: &dyn super::Foo) -> u32 { | |
| x.method1() | |
| } | |
| } |