nordic-dev.net/rust
nordic-dev.net/rust
2
0
Fork 0
mirror of https://github.com/rust-lang/rust.git synced 2024-11-22 06:44:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add a debug assertion in codegen that unsize casts of the same principal trait def id are truly NOPs

Browse Source
This commit is contained in:
Michael Goulet 2024-09-23 13:52:02 -04:00
parent 8fc8e03150
commit 3209943604
3 changed files with 38 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
compiler/rustc_codegen_cranelift/src/unsize.rs
View File

@ -34,7 +34,22 @@ pub(crate) fn unsized_info<'tcx>(
let old_info =
old_info.expect("unsized_info: missing old info for trait upcasting coercion");
if data_a.principal_def_id() == data_b.principal_def_id() {
// A NOP cast that doesn't actually change anything, should be allowed even with invalid vtables.
// Codegen takes advantage of the additional assumption, where if the
// principal trait def id of what's being casted doesn't change,
// then we don't need to adjust the vtable at all. This
// corresponds to the fact that `dyn Tr<A>: Unsize<dyn Tr<B>>`
// requires that `A = B`; we don't allow *upcasting* objects
// between the same trait with different args. If we, for
// some reason, were to relax the `Unsize` trait, it could become
// unsound, so let's assert here that the trait refs are *equal*.
//
// We can use `assert_eq` because the binders should have been anonymized,
// and because higher-ranked equality now requires the binders are equal.
debug_assert_eq!(
data_a.principal(),
data_b.principal(),
"NOP unsize vtable changed principal trait ref: {data_a} -> {data_b}"
);
return old_info;
}

24
compiler/rustc_codegen_ssa/src/base.rs
View File

@ -125,8 +125,28 @@ fn unsized_info<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
let old_info =
old_info.expect("unsized_info: missing old info for trait upcasting coercion");
if data_a.principal_def_id() == data_b.principal_def_id() {
// A NOP cast that doesn't actually change anything, should be allowed even with
// invalid vtables.
// Codegen takes advantage of the additional assumption, where if the
// principal trait def id of what's being casted doesn't change,
// then we don't need to adjust the vtable at all. This
// corresponds to the fact that `dyn Tr<A>: Unsize<dyn Tr<B>>`
// requires that `A = B`; we don't allow *upcasting* objects
// between the same trait with different args. If we, for
// some reason, were to relax the `Unsize` trait, it could become
// unsound, so let's assert here that the trait refs are *equal*.
//
// We can use `assert_eq` because the binders should have been anonymized,
// and because higher-ranked equality now requires the binders are equal.
debug_assert_eq!(
data_a.principal(),
data_b.principal(),
"NOP unsize vtable changed principal trait ref: {data_a} -> {data_b}"
);
// A NOP cast that doesn't actually change anything, let's avoid any
// unnecessary work. This relies on the assumption that if the principal
// traits are equal, then the associated type bounds (`dyn Trait<Assoc=T>`)
// are also equal, which is ensured by the fact that normalization is
// a function and we do not allow overlapping impls.
return old_info;
}

15
compiler/rustc_mir_transform/src/validate.rs
View File

@ -1233,21 +1233,6 @@ impl<'a, 'tcx> Visitor<'tcx> for TypeChecker<'a, 'tcx> {
)) {
self.fail(location, format!("Unsize coercion, but `{op_ty}` isn't coercible to `{target_type}`"));
}
// FIXME: Codegen has an additional assumption, where if the
// principal trait def id of what's being casted doesn't change,
// then we don't need to adjust the vtable at all. This
// corresponds to the fact that `dyn Tr<A>: Unsize<dyn Tr<B>>`
// requires that `A = B`; we don't allow *upcasting* objects
// between the same trait with different args. Nothing actually
// validates this, though. While it's true right now, if we for
// some reason were to relax the `Unsize` trait, it could become
// unsound. We should eventually validate that, but it would
// require peeling `&Box<Struct<.., dyn Tr<A>, ..>>` down to
// the trait object that's being unsized, and that's rather
// annoying, and also it would need to be opportunistic since
// this MIR is not yet fully monomorphized, so we may bottom
// out in an alias or a projection or something.
}
CastKind::PointerCoercion(PointerCoercion::DynStar, _) => {
// FIXME(dyn-star): make sure nothing needs to be done here.