rustc_next_trait_solver/solve/eval_ctxt/canonical.rs
1//! Canonicalization is used to separate some goal from its context,
2//! throwing away unnecessary information in the process.
3//!
4//! This is necessary to cache goals containing inference variables
5//! and placeholders without restricting them to the current `InferCtxt`.
6//!
7//! Canonicalization is fairly involved, for more details see the relevant
8//! section of the [rustc-dev-guide][c].
9//!
10//! [c]: https://rustc-dev-guide.rust-lang.org/solve/canonicalization.html
11
12use std::iter;
13
14use rustc_index::IndexVec;
15use rustc_type_ir::data_structures::HashSet;
16use rustc_type_ir::inherent::*;
17use rustc_type_ir::relate::solver_relating::RelateExt;
18use rustc_type_ir::{
19 self as ty, Canonical, CanonicalVarValues, InferCtxtLike, Interner, TypeFoldable,
20};
21use tracing::{debug, instrument, trace};
22
23use crate::canonicalizer::Canonicalizer;
24use crate::delegate::SolverDelegate;
25use crate::resolve::eager_resolve_vars;
26use crate::solve::eval_ctxt::CurrentGoalKind;
27use crate::solve::{
28 CanonicalInput, CanonicalResponse, Certainty, EvalCtxt, ExternalConstraintsData, Goal,
29 MaybeCause, NestedNormalizationGoals, NoSolution, PredefinedOpaquesData, QueryInput,
30 QueryResult, Response, inspect, response_no_constraints_raw,
31};
32
33trait ResponseT<I: Interner> {
34 fn var_values(&self) -> CanonicalVarValues<I>;
35}
36
37impl<I: Interner> ResponseT<I> for Response<I> {
38 fn var_values(&self) -> CanonicalVarValues<I> {
39 self.var_values
40 }
41}
42
43impl<I: Interner, T> ResponseT<I> for inspect::State<I, T> {
44 fn var_values(&self) -> CanonicalVarValues<I> {
45 self.var_values
46 }
47}
48
49impl<D, I> EvalCtxt<'_, D>
50where
51 D: SolverDelegate<Interner = I>,
52 I: Interner,
53{
54 /// Canonicalizes the goal remembering the original values
55 /// for each bound variable.
56 pub(super) fn canonicalize_goal(
57 &self,
58 goal: Goal<I, I::Predicate>,
59 ) -> (Vec<I::GenericArg>, CanonicalInput<I, I::Predicate>) {
60 // We only care about one entry per `OpaqueTypeKey` here,
61 // so we only canonicalize the lookup table and ignore
62 // duplicate entries.
63 let opaque_types = self.delegate.clone_opaque_types_lookup_table();
64 let (goal, opaque_types) = eager_resolve_vars(self.delegate, (goal, opaque_types));
65
66 let mut orig_values = Default::default();
67 let canonical = Canonicalizer::canonicalize_input(
68 self.delegate,
69 &mut orig_values,
70 QueryInput {
71 goal,
72 predefined_opaques_in_body: self
73 .cx()
74 .mk_predefined_opaques_in_body(PredefinedOpaquesData { opaque_types }),
75 },
76 );
77 let query_input = ty::CanonicalQueryInput { canonical, typing_mode: self.typing_mode() };
78 (orig_values, query_input)
79 }
80
81 /// To return the constraints of a canonical query to the caller, we canonicalize:
82 ///
83 /// - `var_values`: a map from bound variables in the canonical goal to
84 /// the values inferred while solving the instantiated goal.
85 /// - `external_constraints`: additional constraints which aren't expressible
86 /// using simple unification of inference variables.
87 ///
88 /// This takes the `shallow_certainty` which represents whether we're confident
89 /// that the final result of the current goal only depends on the nested goals.
90 ///
91 /// In case this is `Certainy::Maybe`, there may still be additional nested goals
92 /// or inference constraints required for this candidate to be hold. The candidate
93 /// always requires all already added constraints and nested goals.
94 #[instrument(level = "trace", skip(self), ret)]
95 pub(in crate::solve) fn evaluate_added_goals_and_make_canonical_response(
96 &mut self,
97 shallow_certainty: Certainty,
98 ) -> QueryResult<I> {
99 self.inspect.make_canonical_response(shallow_certainty);
100
101 let goals_certainty = self.try_evaluate_added_goals()?;
102 assert_eq!(
103 self.tainted,
104 Ok(()),
105 "EvalCtxt is tainted -- nested goals may have been dropped in a \
106 previous call to `try_evaluate_added_goals!`"
107 );
108
109 // We only check for leaks from universes which were entered inside
110 // of the query.
111 self.delegate.leak_check(self.max_input_universe).map_err(|NoSolution| {
112 trace!("failed the leak check");
113 NoSolution
114 })?;
115
116 let (certainty, normalization_nested_goals) =
117 match (self.current_goal_kind, shallow_certainty) {
118 // When normalizing, we've replaced the expected term with an unconstrained
119 // inference variable. This means that we dropped information which could
120 // have been important. We handle this by instead returning the nested goals
121 // to the caller, where they are then handled. We only do so if we do not
122 // need to recompute the `NormalizesTo` goal afterwards to avoid repeatedly
123 // uplifting its nested goals. This is the case if the `shallow_certainty` is
124 // `Certainty::Yes`.
125 (CurrentGoalKind::NormalizesTo, Certainty::Yes) => {
126 let goals = std::mem::take(&mut self.nested_goals);
127 // As we return all ambiguous nested goals, we can ignore the certainty
128 // returned by `self.try_evaluate_added_goals()`.
129 if goals.is_empty() {
130 assert!(matches!(goals_certainty, Certainty::Yes));
131 }
132 (
133 Certainty::Yes,
134 NestedNormalizationGoals(
135 goals.into_iter().map(|(s, g, _)| (s, g)).collect(),
136 ),
137 )
138 }
139 _ => {
140 let certainty = shallow_certainty.and(goals_certainty);
141 (certainty, NestedNormalizationGoals::empty())
142 }
143 };
144
145 if let Certainty::Maybe(cause @ MaybeCause::Overflow { keep_constraints: false, .. }) =
146 certainty
147 {
148 // If we have overflow, it's probable that we're substituting a type
149 // into itself infinitely and any partial substitutions in the query
150 // response are probably not useful anyways, so just return an empty
151 // query response.
152 //
153 // This may prevent us from potentially useful inference, e.g.
154 // 2 candidates, one ambiguous and one overflow, which both
155 // have the same inference constraints.
156 //
157 // Changing this to retain some constraints in the future
158 // won't be a breaking change, so this is good enough for now.
159 return Ok(self.make_ambiguous_response_no_constraints(cause));
160 }
161
162 let external_constraints =
163 self.compute_external_query_constraints(certainty, normalization_nested_goals);
164 let (var_values, mut external_constraints) =
165 eager_resolve_vars(self.delegate, (self.var_values, external_constraints));
166
167 // Remove any trivial or duplicated region constraints once we've resolved regions
168 let mut unique = HashSet::default();
169 external_constraints.region_constraints.retain(|outlives| {
170 outlives.0.as_region().is_none_or(|re| re != outlives.1) && unique.insert(*outlives)
171 });
172
173 let canonical = Canonicalizer::canonicalize_response(
174 self.delegate,
175 self.max_input_universe,
176 &mut Default::default(),
177 Response {
178 var_values,
179 certainty,
180 external_constraints: self.cx().mk_external_constraints(external_constraints),
181 },
182 );
183
184 // HACK: We bail with overflow if the response would have too many non-region
185 // inference variables. This tends to only happen if we encounter a lot of
186 // ambiguous alias types which get replaced with fresh inference variables
187 // during generalization. This prevents hangs caused by an exponential blowup,
188 // see tests/ui/traits/next-solver/coherence-alias-hang.rs.
189 match self.current_goal_kind {
190 // We don't do so for `NormalizesTo` goals as we erased the expected term and
191 // bailing with overflow here would prevent us from detecting a type-mismatch,
192 // causing a coherence error in diesel, see #131969. We still bail with overflow
193 // when later returning from the parent AliasRelate goal.
194 CurrentGoalKind::NormalizesTo => {}
195 CurrentGoalKind::Misc | CurrentGoalKind::CoinductiveTrait => {
196 let num_non_region_vars = canonical
197 .variables
198 .iter()
199 .filter(|c| !c.is_region() && c.is_existential())
200 .count();
201 if num_non_region_vars > self.cx().recursion_limit() {
202 debug!(?num_non_region_vars, "too many inference variables -> overflow");
203 return Ok(self.make_ambiguous_response_no_constraints(MaybeCause::Overflow {
204 suggest_increasing_limit: true,
205 keep_constraints: false,
206 }));
207 }
208 }
209 }
210
211 Ok(canonical)
212 }
213
214 /// Constructs a totally unconstrained, ambiguous response to a goal.
215 ///
216 /// Take care when using this, since often it's useful to respond with
217 /// ambiguity but return constrained variables to guide inference.
218 pub(in crate::solve) fn make_ambiguous_response_no_constraints(
219 &self,
220 maybe_cause: MaybeCause,
221 ) -> CanonicalResponse<I> {
222 response_no_constraints_raw(
223 self.cx(),
224 self.max_input_universe,
225 self.variables,
226 Certainty::Maybe(maybe_cause),
227 )
228 }
229
230 /// Computes the region constraints and *new* opaque types registered when
231 /// proving a goal.
232 ///
233 /// If an opaque was already constrained before proving this goal, then the
234 /// external constraints do not need to record that opaque, since if it is
235 /// further constrained by inference, that will be passed back in the var
236 /// values.
237 #[instrument(level = "trace", skip(self), ret)]
238 fn compute_external_query_constraints(
239 &self,
240 certainty: Certainty,
241 normalization_nested_goals: NestedNormalizationGoals<I>,
242 ) -> ExternalConstraintsData<I> {
243 // We only return region constraints once the certainty is `Yes`. This
244 // is necessary as we may drop nested goals on ambiguity, which may result
245 // in unconstrained inference variables in the region constraints. It also
246 // prevents us from emitting duplicate region constraints, avoiding some
247 // unnecessary work. This slightly weakens the leak check in case it uses
248 // region constraints from an ambiguous nested goal. This is tested in both
249 // `tests/ui/higher-ranked/leak-check/leak-check-in-selection-5-ambig.rs` and
250 // `tests/ui/higher-ranked/leak-check/leak-check-in-selection-6-ambig-unify.rs`.
251 let region_constraints = if certainty == Certainty::Yes {
252 self.delegate.make_deduplicated_outlives_constraints()
253 } else {
254 Default::default()
255 };
256
257 // We only return *newly defined* opaque types from canonical queries.
258 //
259 // Constraints for any existing opaque types are already tracked by changes
260 // to the `var_values`.
261 let opaque_types = self
262 .delegate
263 .clone_opaque_types_added_since(self.initial_opaque_types_storage_num_entries);
264
265 ExternalConstraintsData { region_constraints, opaque_types, normalization_nested_goals }
266 }
267
268 /// After calling a canonical query, we apply the constraints returned
269 /// by the query using this function.
270 ///
271 /// This happens in three steps:
272 /// - we instantiate the bound variables of the query response
273 /// - we unify the `var_values` of the response with the `original_values`
274 /// - we apply the `external_constraints` returned by the query, returning
275 /// the `normalization_nested_goals`
276 pub(super) fn instantiate_and_apply_query_response(
277 &mut self,
278 param_env: I::ParamEnv,
279 original_values: &[I::GenericArg],
280 response: CanonicalResponse<I>,
281 ) -> (NestedNormalizationGoals<I>, Certainty) {
282 let instantiation = Self::compute_query_response_instantiation_values(
283 self.delegate,
284 &original_values,
285 &response,
286 self.origin_span,
287 );
288
289 let Response { var_values, external_constraints, certainty } =
290 self.delegate.instantiate_canonical(response, instantiation);
291
292 Self::unify_query_var_values(
293 self.delegate,
294 param_env,
295 &original_values,
296 var_values,
297 self.origin_span,
298 );
299
300 let ExternalConstraintsData {
301 region_constraints,
302 opaque_types,
303 normalization_nested_goals,
304 } = &*external_constraints;
305
306 self.register_region_constraints(region_constraints);
307 self.register_new_opaque_types(opaque_types);
308
309 (normalization_nested_goals.clone(), certainty)
310 }
311
312 /// This returns the canonical variable values to instantiate the bound variables of
313 /// the canonical response. This depends on the `original_values` for the
314 /// bound variables.
315 fn compute_query_response_instantiation_values<T: ResponseT<I>>(
316 delegate: &D,
317 original_values: &[I::GenericArg],
318 response: &Canonical<I, T>,
319 span: I::Span,
320 ) -> CanonicalVarValues<I> {
321 // FIXME: Longterm canonical queries should deal with all placeholders
322 // created inside of the query directly instead of returning them to the
323 // caller.
324 let prev_universe = delegate.universe();
325 let universes_created_in_query = response.max_universe.index();
326 for _ in 0..universes_created_in_query {
327 delegate.create_next_universe();
328 }
329
330 let var_values = response.value.var_values();
331 assert_eq!(original_values.len(), var_values.len());
332
333 // If the query did not make progress with constraining inference variables,
334 // we would normally create a new inference variables for bound existential variables
335 // only then unify this new inference variable with the inference variable from
336 // the input.
337 //
338 // We therefore instantiate the existential variable in the canonical response with the
339 // inference variable of the input right away, which is more performant.
340 let mut opt_values = IndexVec::from_elem_n(None, response.variables.len());
341 for (original_value, result_value) in
342 iter::zip(original_values, var_values.var_values.iter())
343 {
344 match result_value.kind() {
345 ty::GenericArgKind::Type(t) => {
346 if let ty::Bound(debruijn, b) = t.kind() {
347 assert_eq!(debruijn, ty::INNERMOST);
348 opt_values[b.var()] = Some(*original_value);
349 }
350 }
351 ty::GenericArgKind::Lifetime(r) => {
352 if let ty::ReBound(debruijn, br) = r.kind() {
353 assert_eq!(debruijn, ty::INNERMOST);
354 opt_values[br.var()] = Some(*original_value);
355 }
356 }
357 ty::GenericArgKind::Const(c) => {
358 if let ty::ConstKind::Bound(debruijn, bv) = c.kind() {
359 assert_eq!(debruijn, ty::INNERMOST);
360 opt_values[bv.var()] = Some(*original_value);
361 }
362 }
363 }
364 }
365
366 let var_values = delegate.cx().mk_args_from_iter(
367 response.variables.iter().enumerate().map(|(index, var_kind)| {
368 if var_kind.universe() != ty::UniverseIndex::ROOT {
369 // A variable from inside a binder of the query. While ideally these shouldn't
370 // exist at all (see the FIXME at the start of this method), we have to deal with
371 // them for now.
372 delegate.instantiate_canonical_var_with_infer(var_kind, span, |idx| {
373 prev_universe + idx.index()
374 })
375 } else if var_kind.is_existential() {
376 // As an optimization we sometimes avoid creating a new inference variable here.
377 //
378 // All new inference variables we create start out in the current universe of the caller.
379 // This is conceptually wrong as these inference variables would be able to name
380 // more placeholders then they should be able to. However the inference variables have
381 // to "come from somewhere", so by equating them with the original values of the caller
382 // later on, we pull them down into their correct universe again.
383 if let Some(v) = opt_values[ty::BoundVar::from_usize(index)] {
384 v
385 } else {
386 delegate
387 .instantiate_canonical_var_with_infer(var_kind, span, |_| prev_universe)
388 }
389 } else {
390 // For placeholders which were already part of the input, we simply map this
391 // universal bound variable back the placeholder of the input.
392 original_values[var_kind.expect_placeholder_index()]
393 }
394 }),
395 );
396
397 CanonicalVarValues { var_values }
398 }
399
400 /// Unify the `original_values` with the `var_values` returned by the canonical query..
401 ///
402 /// This assumes that this unification will always succeed. This is the case when
403 /// applying a query response right away. However, calling a canonical query, doing any
404 /// other kind of trait solving, and only then instantiating the result of the query
405 /// can cause the instantiation to fail. This is not supported and we ICE in this case.
406 ///
407 /// We always structurally instantiate aliases. Relating aliases needs to be different
408 /// depending on whether the alias is *rigid* or not. We're only really able to tell
409 /// whether an alias is rigid by using the trait solver. When instantiating a response
410 /// from the solver we assume that the solver correctly handled aliases and therefore
411 /// always relate them structurally here.
412 #[instrument(level = "trace", skip(delegate))]
413 fn unify_query_var_values(
414 delegate: &D,
415 param_env: I::ParamEnv,
416 original_values: &[I::GenericArg],
417 var_values: CanonicalVarValues<I>,
418 span: I::Span,
419 ) {
420 assert_eq!(original_values.len(), var_values.len());
421
422 for (&orig, response) in iter::zip(original_values, var_values.var_values.iter()) {
423 let goals =
424 delegate.eq_structurally_relating_aliases(param_env, orig, response, span).unwrap();
425 assert!(goals.is_empty());
426 }
427 }
428
429 fn register_region_constraints(
430 &mut self,
431 outlives: &[ty::OutlivesPredicate<I, I::GenericArg>],
432 ) {
433 for &ty::OutlivesPredicate(lhs, rhs) in outlives {
434 match lhs.kind() {
435 ty::GenericArgKind::Lifetime(lhs) => self.register_region_outlives(lhs, rhs),
436 ty::GenericArgKind::Type(lhs) => self.register_ty_outlives(lhs, rhs),
437 ty::GenericArgKind::Const(_) => panic!("const outlives: {lhs:?}: {rhs:?}"),
438 }
439 }
440 }
441
442 fn register_new_opaque_types(&mut self, opaque_types: &[(ty::OpaqueTypeKey<I>, I::Ty)]) {
443 for &(key, ty) in opaque_types {
444 let prev = self.delegate.register_hidden_type_in_storage(key, ty, self.origin_span);
445 // We eagerly resolve inference variables when computing the query response.
446 // This can cause previously distinct opaque type keys to now be structurally equal.
447 //
448 // To handle this, we store any duplicate entries in a separate list to check them
449 // at the end of typeck/borrowck. We could alternatively eagerly equate the hidden
450 // types here. However, doing so is difficult as it may result in nested goals and
451 // any errors may make it harder to track the control flow for diagnostics.
452 if let Some(prev) = prev {
453 self.delegate.add_duplicate_opaque_type(key, prev, self.origin_span);
454 }
455 }
456 }
457}
458
459/// Used by proof trees to be able to recompute intermediate actions while
460/// evaluating a goal. The `var_values` not only include the bound variables
461/// of the query input, but also contain all unconstrained inference vars
462/// created while evaluating this goal.
463pub(in crate::solve) fn make_canonical_state<D, T, I>(
464 delegate: &D,
465 var_values: &[I::GenericArg],
466 max_input_universe: ty::UniverseIndex,
467 data: T,
468) -> inspect::CanonicalState<I, T>
469where
470 D: SolverDelegate<Interner = I>,
471 I: Interner,
472 T: TypeFoldable<I>,
473{
474 let var_values = CanonicalVarValues { var_values: delegate.cx().mk_args(var_values) };
475 let state = inspect::State { var_values, data };
476 let state = eager_resolve_vars(delegate, state);
477 Canonicalizer::canonicalize_response(delegate, max_input_universe, &mut vec![], state)
478}
479
480// FIXME: needs to be pub to be accessed by downstream
481// `rustc_trait_selection::solve::inspect::analyse`.
482pub fn instantiate_canonical_state<D, I, T: TypeFoldable<I>>(
483 delegate: &D,
484 span: I::Span,
485 param_env: I::ParamEnv,
486 orig_values: &mut Vec<I::GenericArg>,
487 state: inspect::CanonicalState<I, T>,
488) -> T
489where
490 D: SolverDelegate<Interner = I>,
491 I: Interner,
492{
493 // In case any fresh inference variables have been created between `state`
494 // and the previous instantiation, extend `orig_values` for it.
495 orig_values.extend(
496 state.value.var_values.var_values.as_slice()[orig_values.len()..]
497 .iter()
498 .map(|&arg| delegate.fresh_var_for_kind_with_span(arg, span)),
499 );
500
501 let instantiation =
502 EvalCtxt::compute_query_response_instantiation_values(delegate, orig_values, &state, span);
503
504 let inspect::State { var_values, data } = delegate.instantiate_canonical(state, instantiation);
505
506 EvalCtxt::unify_query_var_values(delegate, param_env, orig_values, var_values, span);
507 data
508}