Auto merge of rust-lang#119396 - Nadrieril:intersection-tracking, r=<try>

Experiment: track overlapping ranges precisely

The `overlapping_range_endpoints` lint has false positives, e.g. rust-lang#117648. I expected that removing these false positives would have too much of a perf impact but never measured it. This PR is an experiment to see if the perf loss is manageable.

r? `@ghost`

Author: bors

compiler/rustc_pattern_analysis/src/lib.rs

@@ -121,8 +121,10 @@ pub fn analyze_match<'p, 'tcx>(
 
     let pat_column = PatternColumn::new(arms);
 
-    // Lint on ranges that overlap on their endpoints, which is likely a mistake.
-    lint_overlapping_range_endpoints(cx, &pat_column);
+    // Lint ranges that overlap on their endpoints, which is likely a mistake.
+    if !report.overlapping_range_endpoints.is_empty() {
+        lint_overlapping_range_endpoints(cx, &report.overlapping_range_endpoints);
+    }
 
     // Run the non_exhaustive_omitted_patterns lint. Only run on refutable patterns to avoid hitting
     // `if let`s. Only run if the match is exhaustive otherwise the error is redundant.

compiler/rustc_pattern_analysis/src/lints.rs

@@ -1,18 +1,13 @@
-use smallvec::SmallVec;
-
-use rustc_data_structures::captures::Captures;
-use rustc_middle::ty::{self, Ty};
+use rustc_middle::ty::Ty;
 use rustc_session::lint;
 use rustc_session::lint::builtin::NON_EXHAUSTIVE_OMITTED_PATTERNS;
-use rustc_span::Span;
 
-use crate::constructor::{IntRange, MaybeInfiniteInt};
 use crate::errors::{
     NonExhaustiveOmittedPattern, NonExhaustiveOmittedPatternLintOnArm, Overlap,
     OverlappingRangeEndpoints, Uncovered,
 };
 use crate::rustc::{
-    Constructor, DeconstructedPat, MatchArm, MatchCtxt, PlaceCtxt, RustcMatchCheckCtxt,
+    self, Constructor, DeconstructedPat, MatchArm, MatchCtxt, PlaceCtxt, RustcMatchCheckCtxt,
     SplitConstructorSet, WitnessPat,
 };
 use crate::TypeCx;
@@ -64,10 +59,6 @@ impl<'p, 'tcx> PatternColumn<'p, 'tcx> {
         pcx.ctors_for_ty().split(pcx, column_ctors)
     }
 
-    fn iter(&self) -> impl Iterator<Item = &'p DeconstructedPat<'p, 'tcx>> + Captures<'_> {
-        self.patterns.iter().copied()
-    }
-
     /// Does specialization: given a constructor, this takes the patterns from the column that match
     /// the constructor, and outputs their fields.
     /// This returns one column per field of the constructor. They usually all have the same length
@@ -212,77 +203,25 @@ pub(crate) fn lint_nonexhaustive_missing_variants<'a, 'p, 'tcx>(
     }
 }
 
-/// Traverse the patterns to warn the user about ranges that overlap on their endpoints.
-#[instrument(level = "debug", skip(cx))]
 pub(crate) fn lint_overlapping_range_endpoints<'a, 'p, 'tcx>(
     cx: MatchCtxt<'a, 'p, 'tcx>,
-    column: &PatternColumn<'p, 'tcx>,
+    overlapping_range_endpoints: &[rustc::OverlappingRanges<'p, 'tcx>],
 ) {
-    let Some(ty) = column.head_ty(cx) else {
-        return;
-    };
-    let pcx = &PlaceCtxt::new_dummy(cx, ty);
-    let rcx: &RustcMatchCheckCtxt<'_, '_> = cx.tycx;
-
-    let set = column.analyze_ctors(pcx);
-
-    if matches!(ty.kind(), ty::Char | ty::Int(_) | ty::Uint(_)) {
-        let emit_lint = |overlap: &IntRange, this_span: Span, overlapped_spans: &[Span]| {
-            let overlap_as_pat = rcx.hoist_pat_range(overlap, ty);
-            let overlaps: Vec<_> = overlapped_spans
-                .iter()
-                .copied()
-                .map(|span| Overlap { range: overlap_as_pat.clone(), span })
-                .collect();
-            rcx.tcx.emit_spanned_lint(
-                lint::builtin::OVERLAPPING_RANGE_ENDPOINTS,
-                rcx.match_lint_level,
-                this_span,
-                OverlappingRangeEndpoints { overlap: overlaps, range: this_span },
-            );
-        };
-
-        // If two ranges overlapped, the split set will contain their intersection as a singleton.
-        let split_int_ranges = set.present.iter().filter_map(|c| c.as_int_range());
-        for overlap_range in split_int_ranges.clone() {
-            if overlap_range.is_singleton() {
-                let overlap: MaybeInfiniteInt = overlap_range.lo;
-                // Ranges that look like `lo..=overlap`.
-                let mut prefixes: SmallVec<[_; 1]> = Default::default();
-                // Ranges that look like `overlap..=hi`.
-                let mut suffixes: SmallVec<[_; 1]> = Default::default();
-                // Iterate on patterns that contained `overlap`.
-                for pat in column.iter() {
-                    let Constructor::IntRange(this_range) = pat.ctor() else { continue };
-                    let this_span = pat.data().unwrap().span;
-                    if this_range.is_singleton() {
-                        // Don't lint when one of the ranges is a singleton.
-                        continue;
-                    }
-                    if this_range.lo == overlap {
-                        // `this_range` looks like `overlap..=this_range.hi`; it overlaps with any
-                        // ranges that look like `lo..=overlap`.
-                        if !prefixes.is_empty() {
-                            emit_lint(overlap_range, this_span, &prefixes);
-                        }
-                        suffixes.push(this_span)
-                    } else if this_range.hi == overlap.plus_one() {
-                        // `this_range` looks like `this_range.lo..=overlap`; it overlaps with any
-                        // ranges that look like `overlap..=hi`.
-                        if !suffixes.is_empty() {
-                            emit_lint(overlap_range, this_span, &suffixes);
-                        }
-                        prefixes.push(this_span)
-                    }
-                }
-            }
-        }
-    } else {
-        // Recurse into the fields.
-        for ctor in set.present {
-            for col in column.specialize(pcx, &ctor) {
-                lint_overlapping_range_endpoints(cx, &col);
-            }
-        }
+    let rcx = cx.tycx;
+    for overlap in overlapping_range_endpoints {
+        let overlap_as_pat = rcx.hoist_pat_range(&overlap.overlaps_on, overlap.pat.ty());
+        let overlaps: Vec<_> = overlap
+            .overlaps_with
+            .iter()
+            .map(|pat| pat.data().unwrap().span)
+            .map(|span| Overlap { range: overlap_as_pat.clone(), span })
+            .collect();
+        let pat_span = overlap.pat.data().unwrap().span;
+        rcx.tcx.emit_spanned_lint(
+            lint::builtin::OVERLAPPING_RANGE_ENDPOINTS,
+            rcx.match_lint_level,
+            pat_span,
+            OverlappingRangeEndpoints { overlap: overlaps, range: pat_span },
+        );
     }
 }

compiler/rustc_pattern_analysis/src/rustc.rs

@@ -32,6 +32,8 @@ pub type DeconstructedPat<'p, 'tcx> =
     crate::pat::DeconstructedPat<'p, RustcMatchCheckCtxt<'p, 'tcx>>;
 pub type MatchArm<'p, 'tcx> = crate::MatchArm<'p, RustcMatchCheckCtxt<'p, 'tcx>>;
 pub type MatchCtxt<'a, 'p, 'tcx> = crate::MatchCtxt<'a, 'p, RustcMatchCheckCtxt<'p, 'tcx>>;
+pub type OverlappingRanges<'p, 'tcx> =
+    crate::usefulness::OverlappingRanges<'p, RustcMatchCheckCtxt<'p, 'tcx>>;
 pub(crate) type PlaceCtxt<'a, 'p, 'tcx> =
     crate::usefulness::PlaceCtxt<'a, 'p, RustcMatchCheckCtxt<'p, 'tcx>>;
 pub(crate) type SplitConstructorSet<'p, 'tcx> =

compiler/rustc_pattern_analysis/src/usefulness.rs

@@ -712,10 +712,11 @@
 //! I (Nadrieril) prefer to put new tests in `ui/pattern/usefulness` unless there's a specific
 //! reason not to, for example if they crucially depend on a particular feature like `or_patterns`.
 
+use rustc_index::bit_set::BitSet;
 use smallvec::{smallvec, SmallVec};
 use std::fmt;
 
-use crate::constructor::{Constructor, ConstructorSet};
+use crate::constructor::{Constructor, ConstructorSet, IntRange};
 use crate::pat::{DeconstructedPat, WitnessPat};
 use crate::{Captures, MatchArm, MatchCtxt, TypeCx, TypedArena};
 
@@ -911,6 +912,10 @@ struct MatrixRow<'p, Cx: TypeCx> {
     /// [`compute_exhaustiveness_and_usefulness`] if the arm is found to be useful.
     /// This is reset to `false` when specializing.
     useful: bool,
+    /// Tracks which rows above this one have an intersection with this one, i.e. such that there is
+    /// a value that matches both rows.
+    /// FIXME: Because of relevancy we may miss some intersections.
+    intersects: BitSet<usize>,
 }
 
 impl<'p, Cx: TypeCx> MatrixRow<'p, Cx> {
@@ -938,6 +943,7 @@ impl<'p, Cx: TypeCx> MatrixRow<'p, Cx> {
             parent_row: self.parent_row,
             is_under_guard: self.is_under_guard,
             useful: false,
+            intersects: BitSet::new_empty(0), // Initialized in `Matrix::expand_and_push`.
         })
     }
 
@@ -955,6 +961,7 @@ impl<'p, Cx: TypeCx> MatrixRow<'p, Cx> {
             parent_row,
             is_under_guard: self.is_under_guard,
             useful: false,
+            intersects: BitSet::new_empty(0), // Initialized in `Matrix::expand_and_push`.
         }
     }
 }
@@ -991,13 +998,15 @@ struct Matrix<'p, Cx: TypeCx> {
 impl<'p, Cx: TypeCx> Matrix<'p, Cx> {
     /// Pushes a new row to the matrix. If the row starts with an or-pattern, this recursively
     /// expands it. Internal method, prefer [`Matrix::new`].
-    fn expand_and_push(&mut self, row: MatrixRow<'p, Cx>) {
+    fn expand_and_push(&mut self, mut row: MatrixRow<'p, Cx>) {
         if !row.is_empty() && row.head().is_or_pat() {
             // Expand nested or-patterns.
-            for new_row in row.expand_or_pat() {
+            for mut new_row in row.expand_or_pat() {
+                new_row.intersects = BitSet::new_empty(self.rows.len());
                 self.rows.push(new_row);
             }
         } else {
+            row.intersects = BitSet::new_empty(self.rows.len());
             self.rows.push(row);
         }
     }
@@ -1022,6 +1031,7 @@ impl<'p, Cx: TypeCx> Matrix<'p, Cx> {
                 parent_row: row_id, // dummy, we won't read it
                 is_under_guard: arm.has_guard,
                 useful: false,
+                intersects: BitSet::new_empty(row_id),
             };
             matrix.expand_and_push(v);
         }
@@ -1334,6 +1344,7 @@ impl<Cx: TypeCx> WitnessMatrix<Cx> {
 fn compute_exhaustiveness_and_usefulness<'a, 'p, Cx: TypeCx>(
     mcx: MatchCtxt<'a, 'p, Cx>,
     matrix: &mut Matrix<'p, Cx>,
+    overlapping_range_endpoints: &mut Vec<OverlappingRanges<'p, Cx>>,
     is_top_level: bool,
 ) -> WitnessMatrix<Cx> {
     debug_assert!(matrix.rows().all(|r| r.len() == matrix.column_count()));
@@ -1348,21 +1359,19 @@ fn compute_exhaustiveness_and_usefulness<'a, 'p, Cx: TypeCx>(
     let Some(ty) = matrix.head_ty(mcx) else {
         // The base case: there are no columns in the matrix. We are morally pattern-matching on ().
         // A row is useful iff it has no (unguarded) rows above it.
-        for row in matrix.rows_mut() {
-            // All rows are useful until they're not.
-            row.useful = true;
-            // When there's an unguarded row, the match is exhaustive and any subsequent row is not
-            // useful.
-            if !row.is_under_guard {
-                return WitnessMatrix::empty();
-            }
+        let mut useful = true; // Whether the next row is useful.
+        for (i, row) in matrix.rows_mut().enumerate() {
+            row.useful = useful;
+            row.intersects.insert_range(0..i);
+            // The next rows stays useful if this one is under a guard.
+            useful &= row.is_under_guard;
         }
-        // No (unguarded) rows, so the match is not exhaustive. We return a new witness unless
-        // irrelevant.
-        return if matrix.wildcard_row.relevant {
+        return if useful && matrix.wildcard_row.relevant {
+            // The wildcard row is useful; the match is non-exhaustive.
             WitnessMatrix::unit_witness()
         } else {
-            // We choose to not report anything here; see at the top for details.
+            // Either the match is exhaustive, or we choose not to report anything because of
+            // relevancy. See at the top for details.
             WitnessMatrix::empty()
         };
     };
@@ -1415,18 +1424,93 @@ fn compute_exhaustiveness_and_usefulness<'a, 'p, Cx: TypeCx>(
         let ctor_is_relevant = matches!(ctor, Constructor::Missing) || missing_ctors.is_empty();
         let mut spec_matrix = matrix.specialize_constructor(pcx, &ctor, ctor_is_relevant);
         let mut witnesses = ensure_sufficient_stack(|| {
-            compute_exhaustiveness_and_usefulness(mcx, &mut spec_matrix, false)
+            compute_exhaustiveness_and_usefulness(
+                mcx,
+                &mut spec_matrix,
+                overlapping_range_endpoints,
+                false,
+            )
         });
 
         // Transform witnesses for `spec_matrix` into witnesses for `matrix`.
         witnesses.apply_constructor(pcx, &missing_ctors, &ctor, report_individual_missing_ctors);
         // Accumulate the found witnesses.
         ret.extend(witnesses);
 
-        // A parent row is useful if any of its children is.
-        for child_row in spec_matrix.rows() {
-            let parent_row = &mut matrix.rows[child_row.parent_row];
-            parent_row.useful = parent_row.useful || child_row.useful;
+        for (child_row_id, child_row) in spec_matrix.rows().enumerate() {
+            let parent_row_id = child_row.parent_row;
+            let parent_row = &mut matrix.rows[parent_row_id];
+            // A parent row is useful if any of its children is.
+            parent_row.useful |= child_row.useful;
+            for child_intersects in child_row.intersects.iter() {
+                // Convert the intersecting ids into ids for the parent matrix.
+                let parent_intersects = spec_matrix.rows[child_intersects].parent_row;
+                debug!("child row {child_row_id} intersects with child row {child_intersects}");
+                debug!("parent row {parent_row_id} intersects with parent row {parent_intersects}");
+                if parent_intersects != parent_row_id {
+                    // self-intersect can happen with or-patterns
+                    parent_row.intersects.insert(parent_intersects);
+                }
+            }
+        }
+
+        // Detect overlapping ranges.
+        if let Constructor::IntRange(overlap_range) = ctor {
+            // If two ranges overlap on their endpoing, that endpoint will show up as a singleton in
+            // the splitted set.
+            if overlap_range.is_singleton() {
+                let overlap = overlap_range.lo;
+                // Ranges that look like `lo..=overlap`.
+                let mut prefixes: SmallVec<[_; 1]> = Default::default();
+                // Ranges that look like `overlap..=hi`.
+                let mut suffixes: SmallVec<[_; 1]> = Default::default();
+                // Iterate on patterns that contained `overlap`.
+                for (row_id, row) in matrix.rows().enumerate() {
+                    let pat = row.head();
+                    let Constructor::IntRange(this_range) = pat.ctor() else { continue };
+                    // Don't lint when one of the ranges is a singleton.
+                    if this_range.is_singleton() {
+                        continue;
+                    }
+                    if this_range.lo == overlap {
+                        // `this_range` looks like `overlap..=this_range.hi`; it overlaps with any
+                        // ranges that look like `lo..=overlap`.
+                        if !prefixes.is_empty() {
+                            let overlaps_with: Vec<_> = prefixes
+                                .iter()
+                                .filter(|&&other_row_id| row.intersects.contains(other_row_id))
+                                .map(|&other_row_id| matrix.rows[other_row_id].head())
+                                .collect();
+                            if !overlaps_with.is_empty() {
+                                overlapping_range_endpoints.push(OverlappingRanges {
+                                    pat,
+                                    overlaps_on: overlap_range,
+                                    overlaps_with,
+                                });
+                            }
+                        }
+                        suffixes.push(row_id)
+                    } else if this_range.hi == overlap.plus_one() {
+                        // `this_range` looks like `this_range.lo..=overlap`; it overlaps with any
+                        // ranges that look like `overlap..=hi`.
+                        if !suffixes.is_empty() {
+                            let overlaps_with: Vec<_> = suffixes
+                                .iter()
+                                .filter(|&&other_row_id| row.intersects.contains(other_row_id))
+                                .map(|&other_row_id| matrix.rows[other_row_id].head())
+                                .collect();
+                            if !overlaps_with.is_empty() {
+                                overlapping_range_endpoints.push(OverlappingRanges {
+                                    pat,
+                                    overlaps_on: overlap_range,
+                                    overlaps_with,
+                                });
+                            }
+                        }
+                        prefixes.push(row_id)
+                    }
+                }
+            }
         }
     }
 
@@ -1452,13 +1536,22 @@ pub enum Usefulness<'p, Cx: TypeCx> {
     Redundant,
 }
 
+/// Indicates whether or not a given arm is useful.
+#[derive(Clone, Debug)]
+pub struct OverlappingRanges<'p, Cx: TypeCx> {
+    pub pat: &'p DeconstructedPat<'p, Cx>,
+    pub overlaps_on: IntRange,
+    pub overlaps_with: Vec<&'p DeconstructedPat<'p, Cx>>,
+}
+
 /// The output of checking a match for exhaustiveness and arm usefulness.
 pub struct UsefulnessReport<'p, Cx: TypeCx> {
     /// For each arm of the input, whether that arm is useful after the arms above it.
     pub arm_usefulness: Vec<(MatchArm<'p, Cx>, Usefulness<'p, Cx>)>,
     /// If the match is exhaustive, this is empty. If not, this contains witnesses for the lack of
     /// exhaustiveness.
     pub non_exhaustiveness_witnesses: Vec<WitnessPat<Cx>>,
+    pub overlapping_range_endpoints: Vec<OverlappingRanges<'p, Cx>>,
 }
 
 /// Computes whether a match is exhaustive and which of its arms are useful.
@@ -1469,8 +1562,14 @@ pub fn compute_match_usefulness<'p, Cx: TypeCx>(
     scrut_ty: Cx::Ty,
     scrut_validity: ValidityConstraint,
 ) -> UsefulnessReport<'p, Cx> {
+    let mut overlapping_range_endpoints = Vec::new();
     let mut matrix = Matrix::new(cx.wildcard_arena, arms, scrut_ty, scrut_validity);
-    let non_exhaustiveness_witnesses = compute_exhaustiveness_and_usefulness(cx, &mut matrix, true);
+    let non_exhaustiveness_witnesses = compute_exhaustiveness_and_usefulness(
+        cx,
+        &mut matrix,
+        &mut overlapping_range_endpoints,
+        true,
+    );
 
     let non_exhaustiveness_witnesses: Vec<_> = non_exhaustiveness_witnesses.single_column();
     let arm_usefulness: Vec<_> = arms
@@ -1487,5 +1586,5 @@ pub fn compute_match_usefulness<'p, Cx: TypeCx>(
             (arm, usefulness)
         })
         .collect();
-    UsefulnessReport { arm_usefulness, non_exhaustiveness_witnesses }
+    UsefulnessReport { arm_usefulness, non_exhaustiveness_witnesses, overlapping_range_endpoints }
 }