Forbid borrows and unsized types from being used as the type of a const generic under `adt_const_params`
Fixes#112219Fixes#112124Fixes#112125
### Motivation
Currently the `adt_const_params` feature allows writing `Foo<const N: [u8]>` this is entirely useless as it is not possible to write an expression which evaluates to a type that is not `Sized`. In order to actually use unsized types in const generics they are typically written as `const N: &[u8]` which *is* possible to provide a value of.
Unfortunately allowing the types of const parameters to contain references is non trivial (#120961) as it introduces a number of difficult questions about how equality of references in the type system should behave. References in the types of const generics is largely only useful for using unsized types in const generics.
This PR introduces a new feature gate `unsized_const_parameters` and moves support for `const N: [u8]` and `const N: &...` from `adt_const_params` into it. The goal here hopefully is to experiment with allowing `const N: [u8]` to work without references and then eventually completely forbid references in const generics.
Splitting this out into a new feature gate means that stabilization of `adt_const_params` does not have to resolve#120961 which is the only remaining "big" blocker for the feature. Remaining issues after this are a few ICEs and naming bikeshed for `ConstParamTy`.
### Implementation
The implementation is slightly subtle here as we would like to ensure that a stabilization of `adt_const_params` is forwards compatible with any outcome of `unsized_const_parameters`. This is inherently tricky as we do not support unstable trait implementations and we determine whether a type is valid as the type of a const parameter via a trait bound.
There are a few constraints here:
- We would like to *allow for the possibility* of adding a `Sized` supertrait to `ConstParamTy` in the event that we wind up opting to not support unsized types and instead requiring people to write the 'sized version', e.g. `const N: [u8; M]` instead of `const N: [u8]`.
- Crates should be able to enable `unsized_const_parameters` and write trait implementations of `ConstParamTy` for `!Sized` types without downstream crates that only enable `adt_const_params` being able to observe this (required for std to be able to `impl<T> ConstParamTy for [T]`
Ultimately the way this is accomplished is via having two traits (sad), `ConstParamTy` and `UnsizedConstParamTy`. Depending on whether `unsized_const_parameters` is enabled or not we change which trait is used to check whether a type is allowed to be a const parameter.
Long term (when stabilizing `UnsizedConstParamTy`) it should be possible to completely merge these traits (and derive macros), only having a single `trait ConstParamTy` and `macro ConstParamTy`.
Under `adt_const_params` it is now illegal to directly refer to `ConstParamTy` it is only used as an internal impl detail by `derive(ConstParamTy)` and checking const parameters are well formed. This is necessary in order to ensure forwards compatibility with all possible future directions for `feature(unsized_const_parameters)`.
Generally the intuition here should be that `ConstParamTy` is the stable trait that everything uses, and `UnsizedConstParamTy` is that plus unstable implementations (well, I suppose `ConstParamTy` isn't stable yet :P).
More accurate suggestion for `-> Box<dyn Trait>` or `-> impl Trait`
When encountering `-> Trait`, suggest `-> Box<dyn Trait>` (instead of `-> Box<Trait>`.
If there's a single returned type within the `fn`, suggest `-> impl Trait`.
When encountering `-> Trait`, suggest `-> Box<dyn Trait>` (instead of `-> Box<Trait>`.
If there's a single returned type within the `fn`, suggest `-> impl Trait`.
interpret: add sanity check in dyn upcast to double-check what codegen does
For dyn receiver calls, we already have two codepaths: look up the function to call by indexing into the vtable, or alternatively resolve the DefId given the dynamic type of the receiver. With debug assertions enabled, the interpreter does both and compares the results. (Without debug assertions we always use the vtable as it is simpler.)
This PR does the same for dyn trait upcasts. However, for casts *not* using the vtable is the easier thing to do, so now the vtable path is the debug-assertion-only path. In particular, there are cases where the vtable does not contain a pointer for upcasts but instead reuses the old pointer: when the supertrait vtable is a prefix of the larger vtable. We don't want to expose this optimization and detect UB if people do a transmute assuming this optimization, so we cannot in general use the vtable indexing path.
r? ``@oli-obk``
Represent type-level consts with new-and-improved `hir::ConstArg`
### Summary
This is a step toward `min_generic_const_exprs`. We now represent all const
generic arguments using an enum that differentiates between const *paths*
(temporarily just bare const params) and arbitrary anon consts that may perform
computations. This will enable us to cleanly implement the `min_generic_const_args`
plan of allowing the use of generics in paths used as const args, while
disallowing their use in arbitrary anon consts. Here is a summary of the salient
aspects of this change:
- Add `current_def_id_parent` to `LoweringContext`
This is needed to track anon const parents properly once we implement
`ConstArgKind::Path` (which requires moving anon const def-creation
outside of `DefCollector`).
- Create `hir::ConstArgKind` enum with `Path` and `Anon` variants. Use it in the
existing `hir::ConstArg` struct, replacing the previous `hir::AnonConst` field.
- Use `ConstArg` for all instances of const args. Specifically, use it instead
of `AnonConst` for assoc item constraints, array lengths, and const param
defaults.
- Some `ast::AnonConst`s now have their `DefId`s created in
rustc_ast_lowering rather than `DefCollector`. This is because in some
cases they will end up becoming a `ConstArgKind::Path` instead, which
has no `DefId`. We have to solve this in a hacky way where we guess
whether the `AnonConst` could end up as a path const since we can't
know for sure until after name resolution (`N` could refer to a free
const or a nullary struct). If it has no chance as being a const
param, then we create a `DefId` in `DefCollector` -- otherwise we
decide during ast_lowering. This will have to be updated once all path
consts use `ConstArgKind::Path`.
- We explicitly use `ConstArgHasType` for array lengths, rather than
implicitly relying on anon const type feeding -- this is due to the
addition of `ConstArgKind::Path`.
- Some tests have their outputs changed, but the changes are for the
most part minor (including removing duplicate or almost-duplicate
errors). One test now ICEs, but it is for an incomplete, unstable
feature and is now tracked at https://github.com/rust-lang/rust/issues/127009.
### Followup items post-merge
- Use `ConstArgKind::Path` for all const paths, not just const params.
- Fix (no github dont close this issue) #127009
- If a path in generic args doesn't resolve as a type, try to resolve as a const
instead (do this in rustc_resolve). Then remove the special-casing from
`rustc_ast_lowering`, so that all params will automatically be lowered as
`ConstArgKind::Path`.
- (?) Consider making `const_evaluatable_unchecked` a hard error, or at least
trying it in crater
r? `@BoxyUwU`
This is a very large commit since a lot needs to be changed in order to
make the tests pass. The salient changes are:
- `ConstArgKind` gets a new `Path` variant, and all const params are now
represented using it. Non-param paths still use `ConstArgKind::Anon`
to prevent this change from getting too large, but they will soon use
the `Path` variant too.
- `ConstArg` gets a distinct `hir_id` field and its own variant in
`hir::Node`. This affected many parts of the compiler that expected
the parent of an `AnonConst` to be the containing context (e.g., an
array repeat expression). They have been changed to check the
"grandparent" where necessary.
- Some `ast::AnonConst`s now have their `DefId`s created in
rustc_ast_lowering rather than `DefCollector`. This is because in some
cases they will end up becoming a `ConstArgKind::Path` instead, which
has no `DefId`. We have to solve this in a hacky way where we guess
whether the `AnonConst` could end up as a path const since we can't
know for sure until after name resolution (`N` could refer to a free
const or a nullary struct). If it has no chance as being a const
param, then we create a `DefId` in `DefCollector` -- otherwise we
decide during ast_lowering. This will have to be updated once all path
consts use `ConstArgKind::Path`.
- We explicitly use `ConstArgHasType` for array lengths, rather than
implicitly relying on anon const type feeding -- this is due to the
addition of `ConstArgKind::Path`.
- Some tests have their outputs changed, but the changes are for the
most part minor (including removing duplicate or almost-duplicate
errors). One test now ICEs, but it is for an incomplete, unstable
feature and is now tracked at #127009.
Invert infer `error_reporting` mod struture
Parallel change to #127493, which moves `rustc_infer::infer::error_reporting` to `rustc_infer::error_reporting::infer`. After this, we should just be able to merge this into `rustc_trait_selection::error_reporting::infer`, and pull down `TypeErrCtxt` into that crate. 👍
r? lcnr
This commit changes the error reporting mechanism for not implemented
traits to skip impl marked as `#[diagnostic::do_not_recommend]` in the
help part of the error message ("the following other types implement
trait `Foo`:"). The main use case here is to allow crate authors to skip
non-meaningful confusing suggestions. A common example for this are
fully generic impls on tuples.
Uplift elaboration into `rustc_type_ir`
Allows us to deduplicate and consolidate elaboration (including these stupid elaboration duplicate fns i added for pretty printing like 3 years ago) so I'm pretty hyped about this change :3
r? lcnr
Make `can_eq` process obligations (almost) everywhere
Move `can_eq` to an extension trait on `InferCtxt` in `rustc_trait_selection`, and change it so that it processes obligations. This should strengthen it to be more accurate in some cases, but is most important for the new trait solver which delays relating aliases to `AliasRelate` goals. Without this, we always basically just return true when passing aliases to `can_eq`, which can lead to weird errors, for example #127149.
I'm not actually certain if we should *have* `can_eq` be called on the good path. In cases where we need `can_eq`, we probably should just be using a regular probe.
Fixes#127149
r? lcnr
Don't try to label `ObligationCauseCode::CompareImplItem` for an RPITIT, since it has no name
The old (current) trait solver has a limitation that when a where clause in param-env must be normalized using the same where clause, then we get spurious errors in `normalize_param_env_or_error`. I don't think there's an issue tracking it, but it's the root cause for many of the "fixed-by-next-solver" labeled issues.
Specifically, these errors may occur when checking predicate entailment of the GAT that comes out of desugaring RPITITs. Since we use `ObligationCauseCode::CompareImplItem` for these predicates, we try calling `item_name` on an RPITIT which fails, since the RPITIT has no name.
We simply suppress this logic when we're reporting a predicate entailment error for an RPITIT. RPITITs should never have predicate entailment errors, *by construction*, but they may due to this bug in the old solver.
Addresses the ICE in #127331, though doesn't fix the underlying issue (which is fundamental to the old solver).
r? types
Use `ControlFlow` results for visitors that are only looking for a single value
These visitors all had a `Option<Value>` or `bool` field, that, once set, was never unset or modified again. They have been refactored by removing the field and returning `ControlFlow` directly from the visitor
Re-implement a type-size based limit
r? lcnr
This PR reintroduces the type length limit added in #37789, which was accidentally made practically useless by the caching changes to `Ty::walk` in #72412, which caused the `walk` function to no longer walk over identical elements.
Hitting this length limit is not fatal unless we are in codegen -- so it shouldn't affect passes like the mir inliner which creates potentially very large types (which we observed, for example, when the new trait solver compiles `itertools` in `--release` mode).
This also increases the type length limit from `1048576 == 2 ** 20` to `2 ** 24`, which covers all of the code that can be reached with craterbot-check. Individual crates can increase the length limit further if desired.
Perf regression is mild and I think we should accept it -- reinstating this limit is important for the new trait solver and to make sure we don't accidentally hit more type-size related regressions in the future.
Fixes#125460
Actually report normalization-based type errors correctly for alias-relate obligations in new solver
We have some special casing to report type mismatch errors that come from projection predicates, but we don't do that for alias-relate obligations. This PR implements that. There's a bit of code duplication, but 🤷
Best reviewed without whitespace.
r? lcnr
Check alias args for WF even if they have escaping bound vars
#### What
This PR stops skipping arguments of aliases if they have escaping bound vars, instead recursing into them and only discarding the resulting obligations referencing bounds vars.
#### An example:
From the test:
```
trait Trait {
type Gat<U: ?Sized>;
}
fn test<T>(f: for<'a> fn(<&'a T as Trait>::Gat<&'a [str]>)) where for<'a> &'a T: Trait {}
//~^ ERROR the size for values of type `[()]` cannot be known at compilation time
fn main() {}
```
We now prove that `str: Sized` in order for `&'a [str]` to be well-formed. We were previously unconditionally skipping over `&'a [str]` as it referenced a buond variable. We now recurse into it and instead only discard the `[str]: 'a` obligation because of the escaping bound vars.
#### Why?
This is a change that improves consistency about proving well-formedness earlier in the pipeline, which is necessary for future work on where-bounds in binders and correctly handling higher-ranked implied bounds. I don't expect this to fix any unsoundness.
#### What doesn't it fix?
Specifically, this doesn't check projection predicates' components are well-formed, because there are too many regressions: https://github.com/rust-lang/rust/pull/123737#issuecomment-2052198478
