Take 2: Implement object-safety and dynamic dispatch for arbitrary_self_types
This replaces #50173. Over the months that that PR was open, we made a lot of changes to the way this was going to be implemented, and the long, meandering comment thread and commit history would have been confusing to people reading it in the future. So I decided to package everything up with new, straighforward commits and open a new PR.
Here are the main points. Please read the commit messages for details.
- To simplify codegen, we only support receivers that have the ABI of a pointer. That means they are builtin pointer types, or newtypes thereof.
- We introduce a new trait: `DispatchFromDyn<T>`, similar to `CoerceUnsized<T>`. `DispatchFromDyn` has extra requirements that `CoerceUnsized` does not: when you implement `DispatchFromDyn` for a struct, there cannot be any extra fields besides the field being coerced and `PhantomData` fields. This ensures that the struct's ABI is the same as a pointer.
- For a method's receiver (e.g. `self: Rc<Self>`) to be object-safe, it needs to have the following property:
- let `DynReceiver` be the receiver when `Self = dyn Trait`
- let `ConcreteReceiver` be the receiver when `Self = T`, where `T` is some unknown `Sized` type that implements `Trait`, and is the erased type of the trait object.
- `ConcreteReceiver` must implement `DispatchFromDyn<DynReceiver>`
In the case of `Rc<Self>`, this requires `Rc<T>: DispatchFromDyn<Rc<dyn Trait>>`
These rules are explained more thoroughly in the doc comment on `receiver_is_dispatchable` in object_safety.rs.
r? @nikomatsakis and @eddyb
cc @arielb1 @cramertj @withoutboats
Special thanks to @nikomatsakis for getting me un-stuck when implementing the object-safety checks, and @eddyb for helping with the codegen parts.
EDIT 2018-11-01: updated because CoerceSized has been replaced with DispatchFromDyn
rustc: improve E0669 span
E0669 refers to an operand that cannot be coerced into a single LLVM
value, unfortunately right now this uses the Span for the entire inline
assembly statement, which is less than ideal.
This commit preserves the Span from HIR, which lets us emit the error
using the Span for the operand itself in MIR.
r? @nagisa
cc/ @parched
universes refactor 3
Some more refactorings from my universe branch. These are getting a bit more "invasive" -- they start to plumb the universe information through the canonicalization process. As of yet though I don't **believe** this branch changes our behavior in any notable way, though I'm marking the branch as `WIP` to give myself a chance to verify this.
r? @scalexm
disallow `#[repr(C)] and `#[repr(packed)]` on structs implementing DispatchFromDyn because they will change the ABI from Scalar/ScalarPair to Aggregrate, resulting in an ICE during object-safety checks or codegen
Rename `CoerceSized` to `DispatchFromDyn`, and reverse the direction so that, for example, you write
```
impl<T: Unsize<U>, U> DispatchFromDyn<*const U> for *const T {}
```
instead of
```
impl<T: Unsize<U>, U> DispatchFromDyn<*const T> for *const U {}
```
this way the trait is really just a subset of `CoerceUnsized`.
The checks in object_safety.rs are updated for the new trait, and some documentation and method names in there are updated for the new trait name — e.g. `receiver_is_coercible` is now called `receiver_is_dispatchable`. Since the trait now works in the opposite direction, some code had to updated here for that too.
I did not update the error messages for invalid `CoerceSized` (now `DispatchFromDyn`) implementations, except to find/replace `CoerceSized` with `DispatchFromDyn`. Will ask for suggestions in the PR thread.
I’m not sure why these tests have different output now, but they do.
In all cases, the error message that is missing looks like this: “the
trait bound `dyn Trait: Trait` is not satisfied”
My guess is that the error message is going away because object-safety
now involves trait solving, and these extra error messages are no
longer leaking out.
Since the enums get optimized down to 1 byte long, the bits
set in the usize member don't align with the enums on big-endian
machines. Avoid this issue by shrinking the integer member to the
same size as the enums.
borrowck=migrate must look at parents of closures
This fixes the NLL migration mode (which is the default with edition=2018) to inspect all parents of a closure in addition to the closure itself when looking to see if AST-borrowck issues an error for the given code.
This should be a candidate for beta backport.
Fix#55492
back out bogus `Ok`-wrapping suggestion on `?` arm type mismatch
This suggestion was introduced in #51938 / 6cc78bf8d7 (while introducing different language for type errors coming from `?` rather than a `match`), but it has a lot of false-positives, and incorrect suggestions carry more badness than marginal good suggestions do goodness. I regret not doing this earlier. 😞Resolves#52537, resolves#54578.
r? @estebank
Rollup of 11 pull requests
Successful merges:
- #55148 (Implement FromStr for PathBuf)
- #55185 (path suggestions in Rust 2018 should point out the change in semantics)
- #55191 (Fix sub-variant doc display)
- #55199 (Impl items have generics)
- #55244 (Don't rerun MIR passes when inlining)
- #55252 (Add MaybeUninit::new)
- #55257 (Allow extern statics with an extern type)
- #55389 (Remove unnecessary mut in iterator.find_map documentation example, R…)
- #55406 (Update string.rs)
- #55412 (Fix an ICE in the min_const_fn analysis)
- #55421 (Add ManuallyDrop::take)
path suggestions in Rust 2018 should point out the change in semantics
Fixes#55130.
This commit extends existing path suggestions to link to documentation
on the changed semantics of `use` in Rust 2018.
Fix ordering of nested modules in non-mod.rs mods
Flatten relative offset into directory path before adding inline
(mod x { ... }) module names to the current directory path.
Fix#55094
Don't rerun MIR passes when inlining
Fixes#50411
r? @nikomatsakis
I updated your commit message with additional details. Let me know if any of that is incorrect. I also added the appropriate `compile-flags` directive to the test.
Thanks for you help on this!
cc @RalfJung related to your PR #55086
Partial implementation of uniform paths 2.0 to land before beta
Reimplementation of uniform paths using in-scope resolution rather than canaries is a minor breaking change due to stricter future-proofing, so it needs to be landed before beta or backported later.
I hope to implement at least something until beta so we have less to backport.
r? @Mark-Simulacrum
This suggestion was introduced in #51938 / 6cc78bf8d7 (while
introducing different language for type errors coming from `?` rather
than a `match`), but it has a lot of false-positives (as repeatedly
reported in Issues #52537, #52598, #54578, #55336), and incorrect
suggestions carry more badness than marginal good suggestions do
goodness. Just get rid of it (unless and until someone figures out how
to do it correctly).
Resolves#52537, resolves#54578.
We avoid an ICE by checking for an empty meta-item list before we
index into the meta-items, and leave commentary about where we'd like
to issue unused-attributes lints in the future. Note that empty lint
attributes are already accepted by the stable compiler; generalizing
this to weird reason-only lint attributes seems like the
conservative/consilient generalization.
Vadim Petrochenkov suggested this in review ("an error? just to be
conservative"), and it turns out to be convenient from the
implementer's perspective: in the initial proposed implementation (or
`HEAD~2`, as some might prefer to call it), we were doing an entire
whole iteration over the meta items just to find the reason (before
iterating over them to set the actual lint levels). This way, we can
just peek at the end rather than adding that extra loop (or
restructuring the existing code). The RFC doesn't seem to take a
position on this, and there's some precedent for restricting things to
be at the end of a sequence (we only allow `..` at the end of a struct
pattern, even if it would be possible to let it appear anywhere in the
sequence).
Implement by-value object safety
This PR implements **by-value object safety**, which is part of unsized rvalues #48055. That means, with `#![feature(unsized_locals)]`, you can call a method `fn foo(self, ...)` on trait objects. One aim of this is to enable `Box<FnOnce>` in the near future.
The difficulty here is this: when constructing a vtable for a trait `Foo`, we can't just put the function `<T as Foo>::foo` into the table. If `T` is no larger than `usize`, `self` is usually passed directly. However, as the caller of the vtable doesn't know the concrete `Self` type, we want a variant of `<T as Foo>::foo` where `self` is always passed by reference.
Therefore, when the compiler encounters such a method to be generated as a vtable entry, it produces a newly introduced instance called `InstanceDef::VtableShim(def_id)` (that wraps the original instance). the shim just derefs the receiver and calls the original method. We give different symbol names for the shims by appending `::{{vtable-shim}}` to the symbol path (and also adding vtable-shimness as an ingredient to the symbol hash).
r? @eddyb
This is just for the `reason =` name-value meta-item; the
`#[expect(lint)]` attribute also described in the RFC is a problem for
another day.
The place where we were directly calling `emit()` on a match block
(whose arms returned a mutable reference to a diagnostic-builder) was
admittedly cute, but no longer plausibly natural after adding the
if-let to the end of the `LintSource::Node` arm.
This regards #54503.
NLL: cast causes failure to promote to static
Fixes#55288. See commit messages for more details.
r? @oli-obk
cc @nikomatsakis
cc @pnkfelix
cc @RalfJung
Do not allow moving out of thread local under ast borrowck
AST borrowck failed to prevent moving out of a thread-local static.
This was broken. And it also (sometimes?) caused an ICE during drop elaboration.
Fix#47215Fix#54797
Handle bindings in substructure of patterns with type ascriptions
This attempts to follow the outline described by @nikomatsakis [here](https://github.com/rust-lang/rust/issues/47184#issuecomment-420041056). Its a bit more complicated than expected for two reasons:
1. In general it handles sets of type ascriptions, because such ascriptions can be nested within patterns
2. It has a separate types in the HAIR, `PatternTypeProjections` and `PatternTypeProjection`, which are analogues to the corresponding types in the MIR.
The main reason I added the new HAIR types was because I am worried that the current implementation is inefficent, and asymptotically so: It makes copies of vectors as it descends the patterns, even when those accumulated vectors are never used.
Longer term, I would like to used a linked tree structure for the `PatternTypeProjections` and `PatternTypeProjection`, and save the construction of standalone vectors for the MIR types. I didn't want to block landing this on that hypoethetical revision; but I figured I could at least make the future change easier by differentiating between the two types now.
Oh, one more thing: This doesn't attempt to handle `ref x` (in terms of ensuring that any necessary types are ascribed to `x` in that scenario as well). We should open an issue to investigate supporting that as well. But I didn't want to block this PR on that future work.
Fix#54570
