about summary refs log tree commit diff
diff options
context:
space:
mode:
authorTrevor Gross <t.gross35@gmail.com>2024-07-17 19:53:28 -0500
committerGitHub <noreply@github.com>2024-07-17 19:53:28 -0500
commitc36a39cd1ff78ec1336264d5a109ea96ae6642d4 (patch)
treea6e73e45e7b81a3969ddeb498270f8e7d972ea70
parent3c4f820c5bf6627faf7c5e4bb1837ad47275b2b6 (diff)
parentf9c0d3370f04d08cda2da268a54bc543eb310fc7 (diff)
downloadrust-c36a39cd1ff78ec1336264d5a109ea96ae6642d4.tar.gz
rust-c36a39cd1ff78ec1336264d5a109ea96ae6642d4.zip
Rollup merge of #127859 - RalfJung:ptr-dyn-metadata, r=scottmcm
ptr::metadata: avoid references to extern types

References to `extern types` are somewhat dubious entities, since generally we say that references must be dereferenceable for their size as determined via `size_of_val`, but with `extern type` that is an ill-defined statement. I'd like to make Miri warn for such cases since it interacts poorly with Stacked Borrows. To avoid warnings people can't fix, this requires not using references to `extern type` in the standard library, and I think `DynMetadata` is the only currently remaining use. so this changes `DynMetadata` to use a NonNull raw pointer instead. Given that the alignment was 1, this shouldn't really change anything meaningful.

I also updated a comment added by `@scottmcm` in https://github.com/rust-lang/rust/pull/125479, since I think the old comment is wrong. The `DynMetadata` type itself is not special, it is a normal aggregate. But computing field types for wide pointers (including references) is special.
-rw-r--r--library/core/src/ptr/metadata.rs16
1 files changed, 10 insertions, 6 deletions
diff --git a/library/core/src/ptr/metadata.rs b/library/core/src/ptr/metadata.rs
index eb86bf66206..06f205c0f26 100644
--- a/library/core/src/ptr/metadata.rs
+++ b/library/core/src/ptr/metadata.rs
@@ -5,6 +5,7 @@ use crate::hash::{Hash, Hasher};
 use crate::intrinsics::aggregate_raw_ptr;
 use crate::intrinsics::ptr_metadata;
 use crate::marker::Freeze;
+use crate::ptr::NonNull;
 
 /// Provides the pointer metadata type of any pointed-to type.
 ///
@@ -153,7 +154,7 @@ pub const fn from_raw_parts_mut<T: ?Sized>(
 /// compare equal (since identical vtables can be deduplicated within a codegen unit).
 #[lang = "dyn_metadata"]
 pub struct DynMetadata<Dyn: ?Sized> {
-    _vtable_ptr: &'static VTable,
+    _vtable_ptr: NonNull<VTable>,
     _phantom: crate::marker::PhantomData<Dyn>,
 }
 
@@ -166,15 +167,18 @@ extern "C" {
 }
 
 impl<Dyn: ?Sized> DynMetadata<Dyn> {
-    /// One of the things that rustc_middle does with this being a lang item is
-    /// give it `FieldsShape::Primitive`, which means that as far as codegen can
-    /// tell, it *is* a reference, and thus doesn't have any fields.
-    /// That means we can't use field access, and have to transmute it instead.
+    /// When `DynMetadata` appears as the metadata field of a wide pointer, the rustc_middle layout
+    /// computation does magic and the resulting layout is *not* a `FieldsShape::Aggregate`, instead
+    /// it is a `FieldsShape::Primitive`. This means that the same type can have different layout
+    /// depending on whether it appears as the metadata field of a wide pointer or as a stand-alone
+    /// type, which understandably confuses codegen and leads to ICEs when trying to project to a
+    /// field of `DynMetadata`. To work around that issue, we use `transmute` instead of using a
+    /// field projection.
     #[inline]
     fn vtable_ptr(self) -> *const VTable {
         // SAFETY: this layout assumption is hard-coded into the compiler.
         // If it's somehow not a size match, the transmute will error.
-        unsafe { crate::mem::transmute::<Self, &'static VTable>(self) }
+        unsafe { crate::mem::transmute::<Self, *const VTable>(self) }
     }
 
     /// Returns the size of the type associated with this vtable.