diff --git a/src/SUMMARY.md b/src/SUMMARY.md index 9a98611860..9b59170de4 100644 --- a/src/SUMMARY.md +++ b/src/SUMMARY.md @@ -191,6 +191,7 @@ - [Sharing the trait solver with rust-analyzer](./solve/sharing-crates-with-rust-analyzer.md) - [`Unsize` and `CoerceUnsized` traits](./traits/unsize.md) - [Having separate `Trait` and `Projection` bounds](./traits/separate-projection-bounds.md) +- [Well-Formedness](./analysis/well-formed.md) - [Variance](./variance.md) - [Coherence checking](./coherence.md) - [HIR Type checking](./hir-typeck/summary.md) diff --git a/src/analysis/well-formed.md b/src/analysis/well-formed.md new file mode 100644 index 0000000000..07c0a99270 --- /dev/null +++ b/src/analysis/well-formed.md @@ -0,0 +1,163 @@ +# Well-Formedness of Items and Type-Level Terms + +The area of analysis that produces questions like "does `T: Debug` hold true for some data structure that uses T?" or "is this const generic parameter `const B: bool` being handed a value of the right type?" for the trait solver to answer. + +Items and Type-Level Terms[^terms][^terms-abbreviated] are "well formed" when they "follow rules" AKA "fulfill obligations" or "meet the necessary constraints." When we're doing a well-formedness check (wfck) we're usually concerned about if Trait obligations are met, but this also covers obligations of the types broadly, including making sure that the types of const generic terms type check. + +There are two different forms of well-formedness checking: + +- Type-Level Term[^terms] wfck (term-wfck). + - This is the primary subject of "Well-Formedness Checking." +- Items[^items] wfck (item-wfck). + - "Items wfck" can call into "Term-wfck" as Items contain Terms. + - Sometimes normalize their inner Terms first. + +Wfck is not "number of parameters" or "parameter type" checking (AKA "kind checking"), as we might see in languages like Haskell. Neither term-wfck nor item-wfck is concerned with if a type with 2 parameters has 1 or 3 types applied to it (assuming no defaults), or if a const generic parameter has a type applied to it. These kinds of problems will get handled during HIR-ty Lowering[^hir-ty-lower], not wfck. + +Wfck doesn't check or validate lifetimes, this is handled in [MIR](../borrow-check.md). + +## Type-Level Terms + +Type-Level Terms are the data structures that well-formedness checking is focused on analyzing. Item-wfck is one of the entry points where most term-wfck will get performed. + +### Obligations + +Term-wfck begins with generating the list of things that need to be true for the thing-being-wfck'd to be well-formed. + +These predicates end up being referred to as Obligations, Requirements, or Constraints. Preferred term is "obligations", as this matches the suffix of the type and the names of relevant functions. In future, this may be superseded by the new solver's term "Goal". + +The term-wfck module[^tlt-wf-module] contains an `obligations` function that takes type-level terms and returns `PredicateObligations`, a list of obligations that Type-Level Term must satisfy in order to be well-formed. The satisfaction of those obligations is performed by the Trait Solver[^trait-solver], and if they are satisfied then the term is Well-Formed. + +### Well-Formedness of Type-Level Terms + +Terms are Well-Formed when trait obligations within them are satisfied when passed to the trait solver. As an example, the following is not well-formed: + +```rust,ignore +Vec +--- +// Obligations to fulfill +Vec where T: Sized +Vec where str: Sized // This is not true, therefore the term is not well-formed. +``` + +During term-wfck we encounter the obligation for `Vec` that `T: Sized`. For `Vec` we encounter the obligation `str: Sized`, and as `str` is not `Sized` the term is not well-formed. + +### We Don't Need Normalization (Yet) + +[Normalization](../normalization.md) is the process of resolving [type aliases](../normalization.md#aliases) into their underlying type, this is because a type alias is considered well-formed if its underlying type is well-formed. The underlying type is undergoes well-formedness checking at most definition and instantiation sites. + +We don't need to perform normalization to perform term-wfck. + +### Const Generic Arguments + +Term-wfck is also responsible for getting "type-checking" obligations of const generic terms[^tyck-const-generics]. Let's look at the following use of const generics: + +```rust,ignore +fn use_const_generics() { /* ... */ } +// call site +use_const_generics::<6>(); +--- +// call site wfck obligations +const 6: usize +``` + +Applying term-wfck to the call site will provide us with the obligation `6: usize`. This obligation will be passed off to the trait solver just like any trait-style obligation, as the trait solver has more responsibilities than its name suggests. + +## Well-Formedness of Items + +Items[^items] are, generally speaking, "Things that get defined." Item-wfck[^item-wf-module] only happens at the signature level for types and functions, including the methods and implementations. This doesn't happen for Free Type Aliases other than Const Generic Argument type checking. + +Items are a major entry point for performing term-wfck. Because Items contain Terms, item-wfck can invoke term-wfck. + +### We (Sometimes) Need Normalization + +Currently, there are places where normalization of an Item happens before its Terms have gone through wfck. This is considered problematic as this allows some terms to [bypass term-wfck entirely](https://github.com/rust-lang/rust/issues/100041). + +### Trivial Bounds + +Trivial bounds[^item-wf-global-bounds] are bounds that don't need any further normalization to go through wfck. These are also sometimes called Global Bounds. Consider the following: + +```rust,ignore +fn apartment_complex(block: T, name: String) where String: Clone { /* ... */ } +--- +String: Clone // Trivial bound! We don't have to wfck T or any sub-terms to know this holds. +// Maybe there's obligations on T but we don't care about them here. +// ... +``` + +This produces the trivial bound `String: Clone`. This is something we can check without instantiating any other information in this Item. We don't need to know any information about `T` to be able to make a judgment on the well-formedness of `String: Clone`. + +False trivial bounds are things like: + +```rust,ignore +fn apartment_simple(block: T, name: String) where String: Copy { /* ... */ } +--- +String: Copy // Trivial bound again, but this one is false! +``` + +Here we have a trivial bound that does not hold, because `String` is not `Copy`. + +## When We Don't Fully Do Well-Formedness Checking + +Well-formedness checking is not a coherent "stage" of type checking. It gets called from various contexts, and there are places where it gets skipped partially or entirely. + +### Trait Objects + +Trait objects of traits with where clauses / const generics do not undergo wfck until the type is coerced back into a concrete type. + +```rust,ignore +trait Trait {} +fn foo(_: &dyn Trait) {} +--- +// This doesn't end up being generated, because it happens within a trait object. +const N: usize +const B: bool +N = B // Substitution +// This fails once we coerce out of a trait object to a concrete type. +// But because we don't coerce, it passes wfck. +const B: usize + bool +``` + +The above shouldn't compile, and yet it does. `foo`s const generic argument is a `bool`, while `Trait`'s is a `usize`. But because the wfck of trait objects doesn't happen until coercion into a concrete type, the above compiles just fine. + +### Binders / Higher-Ranked Bounds + +HRBs also skip the wfck on their subjects. + +TODO: bit of background of why this doesn't happen. + +```rust,ignore +let _: for<'a> fn(Vec<[&'a ()]>); +--- +// This doesn't end up being generated, because it happens within a HRB +[&'a ()]: Sized // slices aren't sized. +``` + +### Free Type Aliases + +The rhs of Free Type Aliases[^fta] do not go through a full wfck at the definition site, with the exception of shallowly "type checking" const generic parameters of the rhs. + +This means the following _currently_ passes type checking, assuming you don't actually use it in a non-FTA Item: + +```rust,ignore +type WorksButShouldNot = Vec; +--- +// This should fail! But we skip the rhs of free type aliases +str: Sized +``` + +This shouldn't work, as `T: Sized`, `str: Sized` being implied by `Vec`, but because the rhs of a free type alias doesn't go through well-formedness checking unless it's used this doesn't error. + +[^trait-solver]: Despite being called a trait solver, it solves other things too[^boxy]. +[^fta]: Type aliases not associated with anything, i.e. a module-level `type Alias = Vec;`. +[^boxy]: Boxy said so. TODO: don't have any of these references :) +[^items]: "Definition" style things in rust, See the [glossary](../appendix/glossary.md) +[^terms]: Type expressions? TODO: this needs to be nailed down, and maybe inserted into the glossary. +[^terms-abbreviated]: Abbreviated as "Terms" on this page in some areas. +[^hir-ty-lower]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_hir_analysis/hir_ty_lowering/index.html +[^tlt-wf-module]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_trait_selection/traits/wf/index.html +[^item-wf-module]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_hir_analysis/check/wfcheck/index.html +[^wf-ctx-construction]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_hir_analysis/check/wfcheck/fn.enter_wf_checking_ctxt.html +[^item-wf-ctx]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_hir_analysis/check/wfcheck/struct.WfCheckingCtxt.html +[^item-wf-global-bounds]: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_hir_analysis/check/wfcheck/struct.WfCheckingCtxt.html#method.check_false_global_bounds +[^tyck-const-generics]: https://rustc-dev-guide.rust-lang.org/const-generics.html#checking-types-of-const-arguments \ No newline at end of file