Rollup merge of rust-lang#122644 - Nadrieril:complexity-tests, r=compiler-errors

pattern analysis: add a custom test harness

There are two features of the pattern analysis code that are hard to test: the newly-added pattern complexity limit, and the computation of arm intersections. This PR adds some crate-specific tests for that, including an unmaintainable but pretty macro to help construct patterns.

r? `@compiler-errors`

Author: GuillaumeGomez

Cargo.lock

@@ -4440,6 +4440,8 @@ dependencies = [
  "rustc_target",
  "smallvec",
  "tracing",
+ "tracing-subscriber",
+ "tracing-tree",
 ]
 
 [[package]]

compiler/rustc_pattern_analysis/Cargo.toml

@@ -22,6 +22,10 @@ smallvec = { version = "1.8.1", features = ["union"] }
 tracing = "0.1"
 # tidy-alphabetical-end
 
+[dev-dependencies]
+tracing-subscriber = { version = "0.3.3", default-features = false, features = ["fmt", "env-filter", "ansi"] }
+tracing-tree = "0.2.0"
+
 [features]
 default = ["rustc"]
 rustc = [

compiler/rustc_pattern_analysis/src/constructor.rs

@@ -819,6 +819,81 @@ impl<Cx: PatCx> Constructor<Cx> {
             }
         })
     }
+
+    pub(crate) fn fmt_fields(
+        &self,
+        f: &mut fmt::Formatter<'_>,
+        ty: &Cx::Ty,
+        mut fields: impl Iterator<Item = impl fmt::Debug>,
+    ) -> fmt::Result {
+        let mut first = true;
+        let mut start_or_continue = |s| {
+            if first {
+                first = false;
+                ""
+            } else {
+                s
+            }
+        };
+        let mut start_or_comma = || start_or_continue(", ");
+
+        match self {
+            Struct | Variant(_) | UnionField => {
+                Cx::write_variant_name(f, self, ty)?;
+                // Without `cx`, we can't know which field corresponds to which, so we can't
+                // get the names of the fields. Instead we just display everything as a tuple
+                // struct, which should be good enough.
+                write!(f, "(")?;
+                for p in fields {
+                    write!(f, "{}{:?}", start_or_comma(), p)?;
+                }
+                write!(f, ")")?;
+            }
+            // Note: given the expansion of `&str` patterns done in `expand_pattern`, we should
+            // be careful to detect strings here. However a string literal pattern will never
+            // be reported as a non-exhaustiveness witness, so we can ignore this issue.
+            Ref => {
+                write!(f, "&{:?}", &fields.next().unwrap())?;
+            }
+            Slice(slice) => {
+                write!(f, "[")?;
+                match slice.kind {
+                    SliceKind::FixedLen(_) => {
+                        for p in fields {
+                            write!(f, "{}{:?}", start_or_comma(), p)?;
+                        }
+                    }
+                    SliceKind::VarLen(prefix_len, _) => {
+                        for p in fields.by_ref().take(prefix_len) {
+                            write!(f, "{}{:?}", start_or_comma(), p)?;
+                        }
+                        write!(f, "{}..", start_or_comma())?;
+                        for p in fields {
+                            write!(f, "{}{:?}", start_or_comma(), p)?;
+                        }
+                    }
+                }
+                write!(f, "]")?;
+            }
+            Bool(b) => write!(f, "{b}")?,
+            // Best-effort, will render signed ranges incorrectly
+            IntRange(range) => write!(f, "{range:?}")?,
+            F32Range(lo, hi, end) => write!(f, "{lo}{end}{hi}")?,
+            F64Range(lo, hi, end) => write!(f, "{lo}{end}{hi}")?,
+            Str(value) => write!(f, "{value:?}")?,
+            Opaque(..) => write!(f, "<constant pattern>")?,
+            Or => {
+                for pat in fields {
+                    write!(f, "{}{:?}", start_or_continue(" | "), pat)?;
+                }
+            }
+            Never => write!(f, "!")?,
+            Wildcard | Missing | NonExhaustive | Hidden | PrivateUninhabited => {
+                write!(f, "_ : {:?}", ty)?
+            }
+        }
+        Ok(())
+    }
 }
 
 #[derive(Debug, Clone, Copy)]

compiler/rustc_pattern_analysis/src/lib.rs

@@ -49,6 +49,12 @@ pub mod index {
         }
     }
 
+    impl<V> FromIterator<V> for IdxContainer<usize, V> {
+        fn from_iter<T: IntoIterator<Item = V>>(iter: T) -> Self {
+            Self(iter.into_iter().enumerate().collect())
+        }
+    }
+
     #[derive(Debug)]
     pub struct IdxSet<T>(pub rustc_hash::FxHashSet<T>);
     impl<T: Idx> IdxSet<T> {
@@ -120,7 +126,8 @@ pub trait PatCx: Sized + fmt::Debug {
     /// `DeconstructedPat`. Only invoqued when `pat.ctor()` is `Struct | Variant(_) | UnionField`.
     fn write_variant_name(
         f: &mut fmt::Formatter<'_>,
-        pat: &crate::pat::DeconstructedPat<Self>,
+        ctor: &crate::constructor::Constructor<Self>,
+        ty: &Self::Ty,
     ) -> fmt::Result;
 
     /// Raise a bug.

compiler/rustc_pattern_analysis/src/pat.rs

@@ -138,81 +138,11 @@ impl<Cx: PatCx> DeconstructedPat<Cx> {
 /// This is best effort and not good enough for a `Display` impl.
 impl<Cx: PatCx> fmt::Debug for DeconstructedPat<Cx> {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        let pat = self;
-        let mut first = true;
-        let mut start_or_continue = |s| {
-            if first {
-                first = false;
-                ""
-            } else {
-                s
-            }
-        };
-        let mut start_or_comma = || start_or_continue(", ");
-
         let mut fields: Vec<_> = (0..self.arity).map(|_| PatOrWild::Wild).collect();
         for ipat in self.iter_fields() {
             fields[ipat.idx] = PatOrWild::Pat(&ipat.pat);
         }
-
-        match pat.ctor() {
-            Struct | Variant(_) | UnionField => {
-                Cx::write_variant_name(f, pat)?;
-                // Without `cx`, we can't know which field corresponds to which, so we can't
-                // get the names of the fields. Instead we just display everything as a tuple
-                // struct, which should be good enough.
-                write!(f, "(")?;
-                for p in fields {
-                    write!(f, "{}", start_or_comma())?;
-                    write!(f, "{p:?}")?;
-                }
-                write!(f, ")")
-            }
-            // Note: given the expansion of `&str` patterns done in `expand_pattern`, we should
-            // be careful to detect strings here. However a string literal pattern will never
-            // be reported as a non-exhaustiveness witness, so we can ignore this issue.
-            Ref => {
-                write!(f, "&{:?}", &fields[0])
-            }
-            Slice(slice) => {
-                write!(f, "[")?;
-                match slice.kind {
-                    SliceKind::FixedLen(_) => {
-                        for p in fields {
-                            write!(f, "{}{:?}", start_or_comma(), p)?;
-                        }
-                    }
-                    SliceKind::VarLen(prefix_len, _) => {
-                        for p in &fields[..prefix_len] {
-                            write!(f, "{}{:?}", start_or_comma(), p)?;
-                        }
-                        write!(f, "{}", start_or_comma())?;
-                        write!(f, "..")?;
-                        for p in &fields[prefix_len..] {
-                            write!(f, "{}{:?}", start_or_comma(), p)?;
-                        }
-                    }
-                }
-                write!(f, "]")
-            }
-            Bool(b) => write!(f, "{b}"),
-            // Best-effort, will render signed ranges incorrectly
-            IntRange(range) => write!(f, "{range:?}"),
-            F32Range(lo, hi, end) => write!(f, "{lo}{end}{hi}"),
-            F64Range(lo, hi, end) => write!(f, "{lo}{end}{hi}"),
-            Str(value) => write!(f, "{value:?}"),
-            Opaque(..) => write!(f, "<constant pattern>"),
-            Or => {
-                for pat in fields {
-                    write!(f, "{}{:?}", start_or_continue(" | "), pat)?;
-                }
-                Ok(())
-            }
-            Never => write!(f, "!"),
-            Wildcard | Missing | NonExhaustive | Hidden | PrivateUninhabited => {
-                write!(f, "_ : {:?}", pat.ty())
-            }
-        }
+        self.ctor().fmt_fields(f, self.ty(), fields.into_iter())
     }
 }
 
@@ -295,7 +225,6 @@ impl<'p, Cx: PatCx> fmt::Debug for PatOrWild<'p, Cx> {
 
 /// Same idea as `DeconstructedPat`, except this is a fictitious pattern built up for diagnostics
 /// purposes. As such they don't use interning and can be cloned.
-#[derive(Debug)]
 pub struct WitnessPat<Cx: PatCx> {
     ctor: Constructor<Cx>,
     pub(crate) fields: Vec<WitnessPat<Cx>>,
@@ -353,3 +282,10 @@ impl<Cx: PatCx> WitnessPat<Cx> {
         self.fields.iter()
     }
 }
+
+/// This is best effort and not good enough for a `Display` impl.
+impl<Cx: PatCx> fmt::Debug for WitnessPat<Cx> {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        self.ctor().fmt_fields(f, self.ty(), self.fields.iter())
+    }
+}

compiler/rustc_pattern_analysis/src/rustc.rs

@@ -874,13 +874,14 @@ impl<'p, 'tcx: 'p> PatCx for RustcPatCtxt<'p, 'tcx> {
 
     fn write_variant_name(
         f: &mut fmt::Formatter<'_>,
-        pat: &crate::pat::DeconstructedPat<Self>,
+        ctor: &crate::constructor::Constructor<Self>,
+        ty: &Self::Ty,
     ) -> fmt::Result {
-        if let ty::Adt(adt, _) = pat.ty().kind() {
+        if let ty::Adt(adt, _) = ty.kind() {
             if adt.is_box() {
                 write!(f, "Box")?
             } else {
-                let variant = adt.variant(Self::variant_index_for_adt(pat.ctor(), *adt));
+                let variant = adt.variant(Self::variant_index_for_adt(ctor, *adt));
                 write!(f, "{}", variant.name)?;
             }
         }

compiler/rustc_pattern_analysis/src/usefulness.rs

@@ -1042,7 +1042,7 @@ struct MatrixRow<'p, Cx: PatCx> {
     is_under_guard: bool,
     /// When we specialize, we remember which row of the original matrix produced a given row of the
     /// specialized matrix. When we unspecialize, we use this to propagate usefulness back up the
-    /// callstack.
+    /// callstack. On creation, this stores the index of the original match arm.
     parent_row: usize,
     /// False when the matrix is just built. This is set to `true` by
     /// [`compute_exhaustiveness_and_usefulness`] if the arm is found to be useful.
@@ -1163,10 +1163,10 @@ impl<'p, Cx: PatCx> Matrix<'p, Cx> {
             place_info: smallvec![place_info],
             wildcard_row_is_relevant: true,
         };
-        for (row_id, arm) in arms.iter().enumerate() {
+        for (arm_id, arm) in arms.iter().enumerate() {
             let v = MatrixRow {
                 pats: PatStack::from_pattern(arm.pat),
-                parent_row: row_id, // dummy, we don't read it
+                parent_row: arm_id,
                 is_under_guard: arm.has_guard,
                 useful: false,
                 intersects: BitSet::new_empty(0), // Initialized in `Matrix::expand_and_push`.
@@ -1738,6 +1738,9 @@ pub struct UsefulnessReport<'p, Cx: PatCx> {
     /// 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>>,
+    /// For each arm, a set of indices of arms above it that have non-empty intersection, i.e. there
+    /// is a value matched by both arms. This may miss real intersections.
+    pub arm_intersections: Vec<BitSet<usize>>,
 }
 
 /// Computes whether a match is exhaustive and which of its arms are useful.
@@ -1769,5 +1772,19 @@ pub fn compute_match_usefulness<'p, Cx: PatCx>(
         })
         .collect();
 
-    Ok(UsefulnessReport { arm_usefulness, non_exhaustiveness_witnesses })
+    let mut arm_intersections: Vec<_> =
+        arms.iter().enumerate().map(|(i, _)| BitSet::new_empty(i)).collect();
+    for row in matrix.rows() {
+        let arm_id = row.parent_row;
+        for intersection in row.intersects.iter() {
+            // Convert the matrix row ids into arm ids (they can differ because we expand or-patterns).
+            let arm_intersection = matrix.rows[intersection].parent_row;
+            // Note: self-intersection can happen with or-patterns.
+            if arm_intersection != arm_id {
+                arm_intersections[arm_id].insert(arm_intersection);
+            }
+        }
+    }
+
+    Ok(UsefulnessReport { arm_usefulness, non_exhaustiveness_witnesses, arm_intersections })
 }