diff options
Diffstat (limited to '')
-rw-r--r-- | src/test/ui/hrtb/issue-46989.rs | 40 |
1 files changed, 40 insertions, 0 deletions
diff --git a/src/test/ui/hrtb/issue-46989.rs b/src/test/ui/hrtb/issue-46989.rs new file mode 100644 index 000000000..4a09f4be1 --- /dev/null +++ b/src/test/ui/hrtb/issue-46989.rs @@ -0,0 +1,40 @@ +// Regression test for #46989: +// +// In the move to universes, this test started passing. +// It is not necessarily WRONG to do so, but it was a bit +// surprising. The reason that it passed is that when we were +// asked to prove that +// +// for<'a> fn(&'a i32): Foo +// +// we were able to use the impl below to prove +// +// fn(&'empty i32): Foo +// +// and then we were able to prove that +// +// fn(&'empty i32) = for<'a> fn(&'a i32) +// +// This last fact is somewhat surprising, but essentially "falls out" +// from handling variance correctly. In particular, consider the subtyping +// relations. First: +// +// fn(&'empty i32) <: for<'a> fn(&'a i32) +// +// This holds because -- intuitively -- a fn that takes a reference but doesn't use +// it can be given a reference with any lifetime. Similarly, the opposite direction: +// +// for<'a> fn(&'a i32) <: fn(&'empty i32) +// +// holds because 'a can be instantiated to 'empty. + +trait Foo {} + +impl<A> Foo for fn(A) {} + +fn assert_foo<T: Foo>() {} + +fn main() { + assert_foo::<fn(&i32)>(); + //~^ ERROR implementation of `Foo` is not general enough +} |