Region inference: split results from RegionInferenceContext by amandasystems · Pull Request #151688 · rust-lang/rust

amandasystems · 2026-01-26T14:43:18Z

What?

This PR turns RegionInferenceContext into a builder that produces the inferred region values, InferredRegions. The resulting struct implements the public API of RegionInferenceContext and replaces it in consumers.rs. RegionInferenceContext::solve() now consumes (moves) the inference context. It is completely private to region inference.

Why?

RegionInferenceContext has become a huge dump for various values people want to access.
region_infer itself is a very large file that's difficult to find your way around.
Polonius wants to skip as much of region inference as possible, since most of it is redundant. This makes that possible
No method now need warnings about being called after region inference or else they'll panic
RegionInferenceContext now takes almost all of its fields by reference
We can probably drop at least some stuff earlier
Data dependencies during region inference are now made more explicit

Knock-on effects

Region constraint (blame) search is now independent of a RegionInferenceContext
Less data is retained after region inference. This may or may not lead to minor improvements in memory use.
RegionInferenceContext now almost exclusively contains references to values, as opposed to owning them. This addresses most of fn compute_closure_requirements_modulo_opaques shouldn't clone all its inputs #146079
This reduces the coupling between region inference and Polonius-next, making it possible to skip region inference if it is not needed to check universal regions or type tests.
region_infer gains two child modules and becomes a lot less of a behemoth.

Detailed overview of changes

new modules: region_infer::constraint_search and region_infer::universal_regions.
handle_placeholders now consumes less input and does constraint rewriting via mutable references
consumers.rs now has a InferenceResults instead of a RegionInferenceContext
The public methods eval_{equal, outlives} are moved to InferredRegions
Constraint (blame) search, MIR dumping, graphviz rendering, and checking universal regions now all get their own builder-style structs, and the methods for their respective tasks are moved onto them
The option for region inference results in ConsumerOptions is now called InferredRegions
calculate_borrows_out_of_scope_at_location of course now takes InferredRegions
LivenessValues for free regions is now initialised as live at all points along with the constraint rewriting during placeholder handling, as opposed to during construction of RegionInferenceContext
Some trace logging, especially during constraint search, is lost since the data required to print the logs was only necessary to print the logs and never used otherwise.

r? @lcnr

Closes #146079.

compiler/rustc_borrowck/src/region_infer/opaque_types/mod.rs

This ensures all of region inference is immutable, and makes every operation that requires region inference to have been executed to run explicitly require the results.

Since this collapses `RegionInferenceContext`, `::solve()` now consumes the inference context.

on `InferredRegions`.

rustbot · 2026-02-06T14:58:34Z

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

compiler/rustc_borrowck/src/region_infer/mod.rs

compiler/rustc_borrowck/src/diagnostics/find_use.rs

compiler/rustc_borrowck/src/diagnostics/opaque_types.rs

compiler/rustc_borrowck/src/diagnostics/region_name.rs

lcnr · 2026-02-12T17:09:02Z

compiler/rustc_borrowck/src/polonius/dump.rs

+    liveness: &LivenessValues,
    pass_where: PassWhere,
    out: &mut dyn io::Write,
 ) -> io::Result<()> {


pls use a consistent function param ordering between the functions here :<

lcnr · 2026-02-12T17:14:26Z

compiler/rustc_borrowck/src/region_infer/opaque_types/region_ctxt.rs

+        let (definitions, _has_placeholders) = region_definitions(
+            infcx,
+            universal_regions,
+            &mut constraints.liveness_constraints.clone(),


this is... very odd. You give a mutable reference to a clone?

please move the cloned thing into a let binding first

That's a quick hack workaround, I think we should talk about how to not have to do that, where to put region_definitions(), and whether the liveness constarints stuff is necessary in sync because it's pretty much the same thing.

lcnr · 2026-02-12T17:17:03Z

compiler/rustc_borrowck/src/region_infer/graphviz.rs

+    outlives_constraints: &'a OutlivesConstraintSet<'tcx>,
+    mut w: &mut dyn Write,
+) -> io::Result<()> {
+    dot::render(&RawConstraints { tcx, region_definitions, outlives_constraints }, &mut w)


why reference RawConstraints. That type should be Copy

It's because dot::render() takes its arguments by reference (and it wasn't written by me).

lcnr · 2026-02-12T17:17:29Z

compiler/rustc_borrowck/src/region_infer/graphviz.rs

+    dot::render(&SccConstraints { tcx, region_definitions, constraint_sccs, nodes_per_scc }, &mut w)
 }

 struct RawConstraints<'a, 'tcx> {


make this copy and take it by value

lcnr · 2026-02-12T17:17:41Z

compiler/rustc_borrowck/src/region_infer/graphviz.rs

    nodes_per_scc: IndexVec<ConstraintSccIndex, Vec<RegionVid>>,
+    region_definitions: &'a IndexVec<RegionVid, RegionDefinition<'tcx>>,
+    constraint_sccs: &'a ConstraintSccs,
 }


lcnr · 2026-02-12T17:21:57Z

compiler/rustc_borrowck/src/region_infer/mod.rs

+            outlives_requirements.as_mut(),
+            &mut errors_buffer,
+            type_tests,
+        );


hmm, I would prefer check_type_tests to stay on the region inference context 🤔

they are a part of region inference to me

or like, the RegionInferenceCtxt should either not exist at all anymore, or still be used for the region computations 🤔 maybe chat about it in sync

lcnr · 2026-02-12T17:25:53Z

compiler/rustc_borrowck/src/handle_placeholders.rs

+        if let NllRegionVariableOrigin::FreeRegion = origin {
+            // Add all nodes in the CFG to liveness constraints for free regions
+            liveness_constraints.add_all_points(rvid);
+        }


this is a functional change? 🤔

It should not be, this always happens during region inference, I just moved it slightly earlier.

lcnr

scc_values: InferredRegions feels wrong. Maybe just computed_regions: ComputedRegions, regions: InferredRegions

region_definitions why mutable liveness constraints

MirDumper::dump_mir why is scc_values an fn param instead of a field?

View changes since this review

lcnr · 2026-02-12T17:29:27Z

compiler/rustc_borrowck/src/nll.rs

+    // These clones can more or less be avoided,
+    // since they are probably idempotent.


what exactly does this comment mean? It's confusing to me

It's a grammar error; what I meant was that the clones make the operations local since it operates on cloned values as opposed to shared ones, but the operations being performed are probably idempotent so doing them twice (in case of shared values) is likely fine. In other words, just dropping the clones will probably do the right thing.

lcnr · 2026-02-12T17:33:21Z

Can you rebase, I will then do a perf run and locally look at the docs after this PR as that should make it clearer how the API has changed

compiler/rustc_borrowck/src/region_infer/mod.rs

rustbot assigned lcnr Jan 26, 2026

rustbot added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. labels Jan 26, 2026

This comment has been minimized.

Sign in to view

amandasystems force-pushed the split-region-inference branch from 6befd69 to 13a57f6 Compare January 26, 2026 14:50

This comment has been minimized.

Sign in to view

amandasystems force-pushed the split-region-inference branch from 13a57f6 to 00e8121 Compare January 26, 2026 16:12

lcnr reviewed Jan 27, 2026

View reviewed changes

compiler/rustc_borrowck/src/region_infer/opaque_types/mod.rs Outdated Show resolved Hide resolved

amandasystems force-pushed the split-region-inference branch from 00e8121 to 454f98a Compare January 28, 2026 22:21

This comment has been minimized.

Sign in to view

lcnr reviewed Feb 3, 2026

View reviewed changes

compiler/rustc_borrowck/src/region_infer/opaque_types/mod.rs Outdated Show resolved Hide resolved

amandasystems added 7 commits February 6, 2026 14:37

Region inference: split results from RegionInferenceContext

0832b56

This ensures all of region inference is immutable, and makes every operation that requires region inference to have been executed to run explicitly require the results.

Work around the clones

0973817

Split out all the SCC components into InferredRegions.

c92bbfd

Since this collapses `RegionInferenceContext`, `::solve()` now consumes the inference context.

Re-implement the public interface of RegionInferenceContext

ba9fe00

on `InferredRegions`.

Address review comment

49ea7dd

Restructure to avoid opaque types cloning

540582e

Update consumer options

174023c

amandasystems force-pushed the split-region-inference branch from 4c01f95 to 174023c Compare February 6, 2026 14:58

This comment has been minimized.

Sign in to view

amandasystems added 2 commits February 9, 2026 15:36

Remove the final references to RegionInferenceContext from nll

1dea889

Combine RegionInferenceContext::new() and ::solve().

e647018