//! Trait solving using Chalk. use std::sync::{Arc, Mutex}; use chalk_ir::{cast::Cast, family::ChalkIr}; use hir_def::{expr::ExprId, DefWithBodyId, TraitId}; use log::debug; use ra_db::{impl_intern_key, salsa, CrateId}; use ra_prof::profile; use rustc_hash::FxHashSet; use crate::{db::HirDatabase, Crate, ImplBlock, TypeAlias}; use super::{Canonical, GenericPredicate, HirDisplay, ProjectionTy, TraitRef, Ty, TypeWalk}; use self::chalk::{from_chalk, ToChalk}; pub(crate) mod chalk; #[derive(Debug, Clone)] pub struct TraitSolver { krate: Crate, inner: Arc>>, } /// We need eq for salsa impl PartialEq for TraitSolver { fn eq(&self, other: &TraitSolver) -> bool { Arc::ptr_eq(&self.inner, &other.inner) } } impl Eq for TraitSolver {} impl TraitSolver { fn solve( &self, db: &impl HirDatabase, goal: &chalk_ir::UCanonical>>, ) -> Option> { let context = ChalkContext { db, krate: self.krate }; debug!("solve goal: {:?}", goal); let mut solver = match self.inner.lock() { Ok(it) => it, // Our cancellation works via unwinding, but, as chalk is not // panic-safe, we need to make sure to propagate the cancellation. // Ideally, we should also make chalk panic-safe. Err(_) => ra_db::Canceled::throw(), }; let solution = solver.solve(&context, goal); debug!("solve({:?}) => {:?}", goal, solution); solution } } /// This controls the maximum size of types Chalk considers. If we set this too /// high, we can run into slow edge cases; if we set it too low, Chalk won't /// find some solutions. const CHALK_SOLVER_MAX_SIZE: usize = 4; #[derive(Debug, Copy, Clone)] struct ChalkContext<'a, DB> { db: &'a DB, krate: Crate, } pub(crate) fn trait_solver_query( db: &(impl HirDatabase + salsa::Database), krate: Crate, ) -> TraitSolver { db.salsa_runtime().report_untracked_read(); // krate parameter is just so we cache a unique solver per crate let solver_choice = chalk_solve::SolverChoice::SLG { max_size: CHALK_SOLVER_MAX_SIZE }; debug!("Creating new solver for crate {:?}", krate); TraitSolver { krate, inner: Arc::new(Mutex::new(solver_choice.into_solver())) } } /// Collects impls for the given trait in the whole dependency tree of `krate`. pub(crate) fn impls_for_trait_query( db: &impl HirDatabase, krate: CrateId, trait_: TraitId, ) -> Arc<[ImplBlock]> { let mut impls = FxHashSet::default(); // We call the query recursively here. On the one hand, this means we can // reuse results from queries for different crates; on the other hand, this // will only ever get called for a few crates near the root of the tree (the // ones the user is editing), so this may actually be a waste of memory. I'm // doing it like this mainly for simplicity for now. for dep in db.crate_graph().dependencies(krate) { impls.extend(db.impls_for_trait(dep.crate_id, trait_).iter()); } let crate_impl_blocks = db.impls_in_crate(krate); impls.extend(crate_impl_blocks.lookup_impl_blocks_for_trait(trait_).map(ImplBlock::from)); impls.into_iter().collect() } /// A set of clauses that we assume to be true. E.g. if we are inside this function: /// ```rust /// fn foo(t: T) {} /// ``` /// we assume that `T: Default`. #[derive(Clone, Debug, PartialEq, Eq, Hash)] pub struct TraitEnvironment { pub predicates: Vec, } impl TraitEnvironment { /// Returns trait refs with the given self type which are supposed to hold /// in this trait env. E.g. if we are in `foo()`, this will /// find that `T: SomeTrait` if we call it for `T`. pub(crate) fn trait_predicates_for_self_ty<'a>( &'a self, ty: &'a Ty, ) -> impl Iterator + 'a { self.predicates.iter().filter_map(move |pred| match pred { GenericPredicate::Implemented(tr) if tr.self_ty() == ty => Some(tr), _ => None, }) } } /// Something (usually a goal), along with an environment. #[derive(Clone, Debug, PartialEq, Eq, Hash)] pub struct InEnvironment { pub environment: Arc, pub value: T, } impl InEnvironment { pub fn new(environment: Arc, value: T) -> InEnvironment { InEnvironment { environment, value } } } /// Something that needs to be proven (by Chalk) during type checking, e.g. that /// a certain type implements a certain trait. Proving the Obligation might /// result in additional information about inference variables. #[derive(Clone, Debug, PartialEq, Eq, Hash)] pub enum Obligation { /// Prove that a certain type implements a trait (the type is the `Self` type /// parameter to the `TraitRef`). Trait(TraitRef), Projection(ProjectionPredicate), } impl Obligation { pub fn from_predicate(predicate: GenericPredicate) -> Option { match predicate { GenericPredicate::Implemented(trait_ref) => Some(Obligation::Trait(trait_ref)), GenericPredicate::Projection(projection_pred) => { Some(Obligation::Projection(projection_pred)) } GenericPredicate::Error => None, } } } #[derive(Clone, Debug, PartialEq, Eq, Hash)] pub struct ProjectionPredicate { pub projection_ty: ProjectionTy, pub ty: Ty, } impl TypeWalk for ProjectionPredicate { fn walk(&self, f: &mut impl FnMut(&Ty)) { self.projection_ty.walk(f); self.ty.walk(f); } fn walk_mut_binders(&mut self, f: &mut impl FnMut(&mut Ty, usize), binders: usize) { self.projection_ty.walk_mut_binders(f, binders); self.ty.walk_mut_binders(f, binders); } } /// Solve a trait goal using Chalk. pub(crate) fn trait_solve_query( db: &impl HirDatabase, krate: Crate, goal: Canonical>, ) -> Option { let _p = profile("trait_solve_query"); debug!("trait_solve_query({})", goal.value.value.display(db)); if let Obligation::Projection(pred) = &goal.value.value { if let Ty::Bound(_) = &pred.projection_ty.parameters[0] { // Hack: don't ask Chalk to normalize with an unknown self type, it'll say that's impossible return Some(Solution::Ambig(Guidance::Unknown)); } } let canonical = goal.to_chalk(db).cast(); // We currently don't deal with universes (I think / hope they're not yet // relevant for our use cases?) let u_canonical = chalk_ir::UCanonical { canonical, universes: 1 }; let solution = db.trait_solver(krate).solve(db, &u_canonical); solution.map(|solution| solution_from_chalk(db, solution)) } fn solution_from_chalk( db: &impl HirDatabase, solution: chalk_solve::Solution, ) -> Solution { let convert_subst = |subst: chalk_ir::Canonical>| { let value = subst .value .parameters .into_iter() .map(|p| { let ty = match p { chalk_ir::Parameter(chalk_ir::ParameterKind::Ty(ty)) => from_chalk(db, ty), chalk_ir::Parameter(chalk_ir::ParameterKind::Lifetime(_)) => unimplemented!(), }; ty }) .collect(); let result = Canonical { value, num_vars: subst.binders.len() }; SolutionVariables(result) }; match solution { chalk_solve::Solution::Unique(constr_subst) => { let subst = chalk_ir::Canonical { value: constr_subst.value.subst, binders: constr_subst.binders, }; Solution::Unique(convert_subst(subst)) } chalk_solve::Solution::Ambig(chalk_solve::Guidance::Definite(subst)) => { Solution::Ambig(Guidance::Definite(convert_subst(subst))) } chalk_solve::Solution::Ambig(chalk_solve::Guidance::Suggested(subst)) => { Solution::Ambig(Guidance::Suggested(convert_subst(subst))) } chalk_solve::Solution::Ambig(chalk_solve::Guidance::Unknown) => { Solution::Ambig(Guidance::Unknown) } } } #[derive(Clone, Debug, PartialEq, Eq)] pub struct SolutionVariables(pub Canonical>); #[derive(Clone, Debug, PartialEq, Eq)] /// A (possible) solution for a proposed goal. pub enum Solution { /// The goal indeed holds, and there is a unique value for all existential /// variables. Unique(SolutionVariables), /// The goal may be provable in multiple ways, but regardless we may have some guidance /// for type inference. In this case, we don't return any lifetime /// constraints, since we have not "committed" to any particular solution /// yet. Ambig(Guidance), } #[derive(Clone, Debug, PartialEq, Eq)] /// When a goal holds ambiguously (e.g., because there are multiple possible /// solutions), we issue a set of *guidance* back to type inference. pub enum Guidance { /// The existential variables *must* have the given values if the goal is /// ever to hold, but that alone isn't enough to guarantee the goal will /// actually hold. Definite(SolutionVariables), /// There are multiple plausible values for the existentials, but the ones /// here are suggested as the preferred choice heuristically. These should /// be used for inference fallback only. Suggested(SolutionVariables), /// There's no useful information to feed back to type inference Unknown, } #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)] pub enum FnTrait { FnOnce, FnMut, Fn, } impl FnTrait { fn lang_item_name(self) -> &'static str { match self { FnTrait::FnOnce => "fn_once", FnTrait::FnMut => "fn_mut", FnTrait::Fn => "fn", } } } #[derive(Debug, Clone, PartialEq, Eq, Hash)] pub struct ClosureFnTraitImplData { def: DefWithBodyId, expr: ExprId, fn_trait: FnTrait, } /// An impl. Usually this comes from an impl block, but some built-in types get /// synthetic impls. #[derive(Debug, Clone, PartialEq, Eq, Hash)] pub enum Impl { /// A normal impl from an impl block. ImplBlock(ImplBlock), /// Closure types implement the Fn traits synthetically. ClosureFnTraitImpl(ClosureFnTraitImplData), } /// This exists just for Chalk, because our ImplIds are only unique per module. #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)] pub struct GlobalImplId(salsa::InternId); impl_intern_key!(GlobalImplId); /// An associated type value. Usually this comes from a `type` declaration /// inside an impl block, but for built-in impls we have to synthesize it. /// (We only need this because Chalk wants a unique ID for each of these.) #[derive(Debug, Clone, PartialEq, Eq, Hash)] pub enum AssocTyValue { /// A normal assoc type value from an impl block. TypeAlias(TypeAlias), /// The output type of the Fn trait implementation. ClosureFnTraitImplOutput(ClosureFnTraitImplData), } /// This exists just for Chalk, because it needs a unique ID for each associated /// type value in an impl (even synthetic ones). #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)] pub struct AssocTyValueId(salsa::InternId); impl_intern_key!(AssocTyValueId);