1 /// This file includes the logic for exhaustiveness and usefulness checking for
2 /// pattern-matching. Specifically, given a list of patterns for a type, we can
4 /// (a) the patterns cover every possible constructor for the type [exhaustiveness]
5 /// (b) each pattern is necessary [usefulness]
7 /// The algorithm implemented here is a modified version of the one described in:
8 /// http://moscova.inria.fr/~maranget/papers/warn/index.html
9 /// However, to save future implementors from reading the original paper, we
10 /// summarise the algorithm here to hopefully save time and be a little clearer
11 /// (without being so rigorous).
13 /// The core of the algorithm revolves about a "usefulness" check. In particular, we
14 /// are trying to compute a predicate `U(P, p_{m + 1})` where `P` is a list of patterns
15 /// of length `m` for a compound (product) type with `n` components (we refer to this as
16 /// a matrix). `U(P, p_{m + 1})` represents whether, given an existing list of patterns
17 /// `p_1 ..= p_m`, adding a new pattern will be "useful" (that is, cover previously-
18 /// uncovered values of the type).
20 /// If we have this predicate, then we can easily compute both exhaustiveness of an
21 /// entire set of patterns and the individual usefulness of each one.
22 /// (a) the set of patterns is exhaustive iff `U(P, _)` is false (i.e., adding a wildcard
23 /// match doesn't increase the number of values we're matching)
24 /// (b) a pattern `p_i` is not useful if `U(P[0..=(i-1), p_i)` is false (i.e., adding a
25 /// pattern to those that have come before it doesn't increase the number of values
28 /// For example, say we have the following:
30 /// // x: (Option<bool>, Result<()>)
32 /// (Some(true), _) => {}
33 /// (None, Err(())) => {}
34 /// (None, Err(_)) => {}
37 /// Here, the matrix `P` is 3 x 2 (rows x columns).
43 /// We can tell it's not exhaustive, because `U(P, _)` is true (we're not covering
44 /// `[Some(false), _]`, for instance). In addition, row 3 is not useful, because
45 /// all the values it covers are already covered by row 2.
47 /// To compute `U`, we must have two other concepts.
48 /// 1. `S(c, P)` is a "specialized matrix", where `c` is a constructor (like `Some` or
49 /// `None`). You can think of it as filtering `P` to just the rows whose *first* pattern
50 /// can cover `c` (and expanding OR-patterns into distinct patterns), and then expanding
51 /// the constructor into all of its components.
52 /// The specialization of a row vector is computed by `specialize`.
54 /// It is computed as follows. For each row `p_i` of P, we have four cases:
55 /// 1.1. `p_(i,1) = c(r_1, .., r_a)`. Then `S(c, P)` has a corresponding row:
56 /// r_1, .., r_a, p_(i,2), .., p_(i,n)
57 /// 1.2. `p_(i,1) = c'(r_1, .., r_a')` where `c ≠ c'`. Then `S(c, P)` has no
58 /// corresponding row.
59 /// 1.3. `p_(i,1) = _`. Then `S(c, P)` has a corresponding row:
60 /// _, .., _, p_(i,2), .., p_(i,n)
61 /// 1.4. `p_(i,1) = r_1 | r_2`. Then `S(c, P)` has corresponding rows inlined from:
62 /// S(c, (r_1, p_(i,2), .., p_(i,n)))
63 /// S(c, (r_2, p_(i,2), .., p_(i,n)))
65 /// 2. `D(P)` is a "default matrix". This is used when we know there are missing
66 /// constructor cases, but there might be existing wildcard patterns, so to check the
67 /// usefulness of the matrix, we have to check all its *other* components.
68 /// The default matrix is computed inline in `is_useful`.
70 /// It is computed as follows. For each row `p_i` of P, we have three cases:
71 /// 1.1. `p_(i,1) = c(r_1, .., r_a)`. Then `D(P)` has no corresponding row.
72 /// 1.2. `p_(i,1) = _`. Then `D(P)` has a corresponding row:
73 /// p_(i,2), .., p_(i,n)
74 /// 1.3. `p_(i,1) = r_1 | r_2`. Then `D(P)` has corresponding rows inlined from:
75 /// D((r_1, p_(i,2), .., p_(i,n)))
76 /// D((r_2, p_(i,2), .., p_(i,n)))
78 /// Note that the OR-patterns are not always used directly in Rust, but are used to derive
79 /// the exhaustive integer matching rules, so they're written here for posterity.
81 /// The algorithm for computing `U`
82 /// -------------------------------
83 /// The algorithm is inductive (on the number of columns: i.e., components of tuple patterns).
84 /// That means we're going to check the components from left-to-right, so the algorithm
85 /// operates principally on the first component of the matrix and new pattern `p_{m + 1}`.
86 /// This algorithm is realised in the `is_useful` function.
88 /// Base case. (`n = 0`, i.e., an empty tuple pattern)
89 /// - If `P` already contains an empty pattern (i.e., if the number of patterns `m > 0`),
90 /// then `U(P, p_{m + 1})` is false.
91 /// - Otherwise, `P` must be empty, so `U(P, p_{m + 1})` is true.
93 /// Inductive step. (`n > 0`, i.e., whether there's at least one column
94 /// [which may then be expanded into further columns later])
95 /// We're going to match on the new pattern, `p_{m + 1}`.
96 /// - If `p_{m + 1} == c(r_1, .., r_a)`, then we have a constructor pattern.
97 /// Thus, the usefulness of `p_{m + 1}` can be reduced to whether it is useful when
98 /// we ignore all the patterns in `P` that involve other constructors. This is where
99 /// `S(c, P)` comes in:
100 /// `U(P, p_{m + 1}) := U(S(c, P), S(c, p_{m + 1}))`
101 /// This special case is handled in `is_useful_specialized`.
102 /// - If `p_{m + 1} == _`, then we have two more cases:
103 /// + All the constructors of the first component of the type exist within
104 /// all the rows (after having expanded OR-patterns). In this case:
105 /// `U(P, p_{m + 1}) := ∨(k ϵ constructors) U(S(k, P), S(k, p_{m + 1}))`
106 /// I.e., the pattern `p_{m + 1}` is only useful when all the constructors are
107 /// present *if* its later components are useful for the respective constructors
108 /// covered by `p_{m + 1}` (usually a single constructor, but all in the case of `_`).
109 /// + Some constructors are not present in the existing rows (after having expanded
110 /// OR-patterns). However, there might be wildcard patterns (`_`) present. Thus, we
111 /// are only really concerned with the other patterns leading with wildcards. This is
112 /// where `D` comes in:
113 /// `U(P, p_{m + 1}) := U(D(P), p_({m + 1},2), .., p_({m + 1},n))`
114 /// - If `p_{m + 1} == r_1 | r_2`, then the usefulness depends on each separately:
115 /// `U(P, p_{m + 1}) := U(P, (r_1, p_({m + 1},2), .., p_({m + 1},n)))
116 /// || U(P, (r_2, p_({m + 1},2), .., p_({m + 1},n)))`
118 /// Modifications to the algorithm
119 /// ------------------------------
120 /// The algorithm in the paper doesn't cover some of the special cases that arise in Rust, for
121 /// example uninhabited types and variable-length slice patterns. These are drawn attention to
122 /// throughout the code below. I'll make a quick note here about how exhaustive integer matching
123 /// is accounted for, though.
125 /// Exhaustive integer matching
126 /// ---------------------------
127 /// An integer type can be thought of as a (huge) sum type: 1 | 2 | 3 | ...
128 /// So to support exhaustive integer matching, we can make use of the logic in the paper for
129 /// OR-patterns. However, we obviously can't just treat ranges x..=y as individual sums, because
130 /// they are likely gigantic. So we instead treat ranges as constructors of the integers. This means
131 /// that we have a constructor *of* constructors (the integers themselves). We then need to work
132 /// through all the inductive step rules above, deriving how the ranges would be treated as
133 /// OR-patterns, and making sure that they're treated in the same way even when they're ranges.
134 /// There are really only four special cases here:
135 /// - When we match on a constructor that's actually a range, we have to treat it as if we would
137 /// + It turns out that we can simply extend the case for single-value patterns in
138 /// `specialize` to either be *equal* to a value constructor, or *contained within* a range
140 /// + When the pattern itself is a range, you just want to tell whether any of the values in
141 /// the pattern range coincide with values in the constructor range, which is precisely
143 /// Since when encountering a range pattern for a value constructor, we also use inclusion, it
144 /// means that whenever the constructor is a value/range and the pattern is also a value/range,
145 /// we can simply use intersection to test usefulness.
146 /// - When we're testing for usefulness of a pattern and the pattern's first component is a
148 /// + If all the constructors appear in the matrix, we have a slight complication. By default,
149 /// the behaviour (i.e., a disjunction over specialised matrices for each constructor) is
150 /// invalid, because we want a disjunction over every *integer* in each range, not just a
151 /// disjunction over every range. This is a bit more tricky to deal with: essentially we need
152 /// to form equivalence classes of subranges of the constructor range for which the behaviour
153 /// of the matrix `P` and new pattern `p_{m + 1}` are the same. This is described in more
154 /// detail in `split_grouped_constructors`.
155 /// + If some constructors are missing from the matrix, it turns out we don't need to do
156 /// anything special (because we know none of the integers are actually wildcards: i.e., we
157 /// can't span wildcards using ranges).
159 use self::Constructor::*;
160 use self::Usefulness::*;
161 use self::WitnessPreference::*;
163 use rustc_data_structures::fx::FxHashMap;
164 use rustc_index::vec::Idx;
166 use super::{FieldPat, Pat, PatKind, PatRange};
167 use super::{PatternFoldable, PatternFolder, compare_const_vals};
169 use rustc::hir::def_id::DefId;
170 use rustc::hir::RangeEnd;
171 use rustc::ty::{self, Ty, TyCtxt, TypeFoldable, Const};
172 use rustc::ty::layout::{Integer, IntegerExt, VariantIdx, Size};
174 use rustc::mir::Field;
175 use rustc::mir::interpret::{ConstValue, Scalar, truncate, AllocId, Pointer};
176 use rustc::util::common::ErrorReported;
178 use syntax::attr::{SignedInt, UnsignedInt};
179 use syntax_pos::{Span, DUMMY_SP};
181 use arena::TypedArena;
183 use smallvec::{SmallVec, smallvec};
184 use std::cmp::{self, Ordering, min, max};
186 use std::iter::{FromIterator, IntoIterator};
187 use std::ops::RangeInclusive;
189 use std::convert::TryInto;
191 pub fn expand_pattern<'a, 'tcx>(cx: &MatchCheckCtxt<'a, 'tcx>, pat: Pat<'tcx>) -> &'a Pat<'tcx> {
192 cx.pattern_arena.alloc(LiteralExpander { tcx: cx.tcx }.fold_pattern(&pat))
195 struct LiteralExpander<'tcx> {
199 impl LiteralExpander<'tcx> {
200 /// Derefs `val` and potentially unsizes the value if `crty` is an array and `rty` a slice.
202 /// `crty` and `rty` can differ because you can use array constants in the presence of slice
203 /// patterns. So the pattern may end up being a slice, but the constant is an array. We convert
204 /// the array to a slice in that case.
205 fn fold_const_value_deref(
207 val: ConstValue<'tcx>,
208 // the pattern's pointee type
210 // the constant's pointee type
212 ) -> ConstValue<'tcx> {
213 debug!("fold_const_value_deref {:?} {:?} {:?}", val, rty, crty);
214 match (val, &crty.kind, &rty.kind) {
215 // the easy case, deref a reference
216 (ConstValue::Scalar(Scalar::Ptr(p)), x, y) if x == y => {
217 let alloc = self.tcx.alloc_map.lock().unwrap_memory(p.alloc_id);
223 // unsize array to slice if pattern is array but match value or other patterns are slice
224 (ConstValue::Scalar(Scalar::Ptr(p)), ty::Array(t, n), ty::Slice(u)) => {
227 data: self.tcx.alloc_map.lock().unwrap_memory(p.alloc_id),
228 start: p.offset.bytes().try_into().unwrap(),
229 end: n.eval_usize(self.tcx, ty::ParamEnv::empty()).try_into().unwrap(),
232 // fat pointers stay the same
233 | (ConstValue::Slice { .. }, _, _)
234 | (_, ty::Slice(_), ty::Slice(_))
235 | (_, ty::Str, ty::Str)
237 // FIXME(oli-obk): this is reachable for `const FOO: &&&u32 = &&&42;` being used
238 _ => bug!("cannot deref {:#?}, {} -> {}", val, crty, rty),
243 impl PatternFolder<'tcx> for LiteralExpander<'tcx> {
244 fn fold_pattern(&mut self, pat: &Pat<'tcx>) -> Pat<'tcx> {
245 debug!("fold_pattern {:?} {:?} {:?}", pat, pat.ty.kind, pat.kind);
246 match (&pat.ty.kind, &*pat.kind) {
249 &PatKind::Constant { value: Const {
251 ty: ty::TyS { kind: ty::Ref(_, crty, _), .. },
257 kind: box PatKind::Deref {
261 kind: box PatKind::Constant { value: self.tcx.mk_const(Const {
262 val: self.fold_const_value_deref(*val, rty, crty),
269 (_, &PatKind::Binding { subpattern: Some(ref s), .. }) => {
272 _ => pat.super_fold_with(self)
277 impl<'tcx> Pat<'tcx> {
278 fn is_wildcard(&self) -> bool {
280 PatKind::Binding { subpattern: None, .. } | PatKind::Wild =>
287 /// A 2D matrix. Nx1 matrices are very common, which is why `SmallVec[_; 2]`
288 /// works well for each row.
289 pub struct Matrix<'p, 'tcx>(Vec<SmallVec<[&'p Pat<'tcx>; 2]>>);
291 impl<'p, 'tcx> Matrix<'p, 'tcx> {
292 pub fn empty() -> Self {
296 pub fn push(&mut self, row: SmallVec<[&'p Pat<'tcx>; 2]>) {
301 /// Pretty-printer for matrices of patterns, example:
302 /// ++++++++++++++++++++++++++
304 /// ++++++++++++++++++++++++++
305 /// + true + [First] +
306 /// ++++++++++++++++++++++++++
307 /// + true + [Second(true)] +
308 /// ++++++++++++++++++++++++++
310 /// ++++++++++++++++++++++++++
311 /// + _ + [_, _, ..tail] +
312 /// ++++++++++++++++++++++++++
313 impl<'p, 'tcx> fmt::Debug for Matrix<'p, 'tcx> {
314 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
317 let &Matrix(ref m) = self;
318 let pretty_printed_matrix: Vec<Vec<String>> = m.iter().map(|row| {
319 row.iter().map(|pat| format!("{:?}", pat)).collect()
322 let column_count = m.iter().map(|row| row.len()).max().unwrap_or(0);
323 assert!(m.iter().all(|row| row.len() == column_count));
324 let column_widths: Vec<usize> = (0..column_count).map(|col| {
325 pretty_printed_matrix.iter().map(|row| row[col].len()).max().unwrap_or(0)
328 let total_width = column_widths.iter().cloned().sum::<usize>() + column_count * 3 + 1;
329 let br = "+".repeat(total_width);
330 write!(f, "{}\n", br)?;
331 for row in pretty_printed_matrix {
333 for (column, pat_str) in row.into_iter().enumerate() {
335 write!(f, "{:1$}", pat_str, column_widths[column])?;
339 write!(f, "{}\n", br)?;
345 impl<'p, 'tcx> FromIterator<SmallVec<[&'p Pat<'tcx>; 2]>> for Matrix<'p, 'tcx> {
346 fn from_iter<T>(iter: T) -> Self
347 where T: IntoIterator<Item=SmallVec<[&'p Pat<'tcx>; 2]>>
349 Matrix(iter.into_iter().collect())
353 pub struct MatchCheckCtxt<'a, 'tcx> {
354 pub tcx: TyCtxt<'tcx>,
355 /// The module in which the match occurs. This is necessary for
356 /// checking inhabited-ness of types because whether a type is (visibly)
357 /// inhabited can depend on whether it was defined in the current module or
358 /// not. E.g., `struct Foo { _private: ! }` cannot be seen to be empty
359 /// outside it's module and should not be matchable with an empty match
362 param_env: ty::ParamEnv<'tcx>,
363 pub pattern_arena: &'a TypedArena<Pat<'tcx>>,
364 pub byte_array_map: FxHashMap<*const Pat<'tcx>, Vec<&'a Pat<'tcx>>>,
367 impl<'a, 'tcx> MatchCheckCtxt<'a, 'tcx> {
368 pub fn create_and_enter<F, R>(
370 param_env: ty::ParamEnv<'tcx>,
375 F: for<'b> FnOnce(MatchCheckCtxt<'b, 'tcx>) -> R,
377 let pattern_arena = TypedArena::default();
383 pattern_arena: &pattern_arena,
384 byte_array_map: FxHashMap::default(),
388 fn is_uninhabited(&self, ty: Ty<'tcx>) -> bool {
389 if self.tcx.features().exhaustive_patterns {
390 self.tcx.is_ty_uninhabited_from(self.module, ty)
396 fn is_non_exhaustive_variant<'p>(&self, pattern: &'p Pat<'tcx>) -> bool {
397 match *pattern.kind {
398 PatKind::Variant { adt_def, variant_index, .. } => {
399 let ref variant = adt_def.variants[variant_index];
400 variant.is_field_list_non_exhaustive()
406 fn is_non_exhaustive_enum(&self, ty: Ty<'tcx>) -> bool {
408 ty::Adt(adt_def, ..) => adt_def.is_variant_list_non_exhaustive(),
413 fn is_local(&self, ty: Ty<'tcx>) -> bool {
415 ty::Adt(adt_def, ..) => adt_def.did.is_local(),
421 #[derive(Clone, Debug, PartialEq)]
422 enum Constructor<'tcx> {
423 /// The constructor of all patterns that don't vary by constructor,
424 /// e.g., struct patterns and fixed-length arrays.
429 ConstantValue(&'tcx ty::Const<'tcx>),
430 /// Ranges of literal values (`2..=5` and `2..5`).
431 ConstantRange(u128, u128, Ty<'tcx>, RangeEnd),
432 /// Array patterns of length n.
436 impl<'tcx> Constructor<'tcx> {
437 fn is_slice(&self) -> bool {
439 Slice { .. } => true,
444 fn variant_index_for_adt<'a>(
446 cx: &MatchCheckCtxt<'a, 'tcx>,
447 adt: &'tcx ty::AdtDef,
450 &Variant(id) => adt.variant_index_with_id(id),
452 assert!(!adt.is_enum());
455 &ConstantValue(c) => crate::const_eval::const_variant_index(cx.tcx, cx.param_env, c),
456 _ => bug!("bad constructor {:?} for adt {:?}", self, adt)
461 #[derive(Clone, Debug)]
462 pub enum Usefulness<'tcx> {
464 UsefulWithWitness(Vec<Witness<'tcx>>),
468 impl<'tcx> Usefulness<'tcx> {
469 fn is_useful(&self) -> bool {
477 #[derive(Copy, Clone, Debug)]
478 pub enum WitnessPreference {
483 #[derive(Copy, Clone, Debug)]
484 struct PatCtxt<'tcx> {
486 max_slice_length: u64,
489 /// A witness of non-exhaustiveness for error reporting, represented
490 /// as a list of patterns (in reverse order of construction) with
491 /// wildcards inside to represent elements that can take any inhabitant
492 /// of the type as a value.
494 /// A witness against a list of patterns should have the same types
495 /// and length as the pattern matched against. Because Rust `match`
496 /// is always against a single pattern, at the end the witness will
497 /// have length 1, but in the middle of the algorithm, it can contain
498 /// multiple patterns.
500 /// For example, if we are constructing a witness for the match against
502 /// struct Pair(Option<(u32, u32)>, bool);
504 /// match (p: Pair) {
505 /// Pair(None, _) => {}
506 /// Pair(_, false) => {}
510 /// We'll perform the following steps:
511 /// 1. Start with an empty witness
512 /// `Witness(vec![])`
513 /// 2. Push a witness `Some(_)` against the `None`
514 /// `Witness(vec![Some(_)])`
515 /// 3. Push a witness `true` against the `false`
516 /// `Witness(vec![Some(_), true])`
517 /// 4. Apply the `Pair` constructor to the witnesses
518 /// `Witness(vec![Pair(Some(_), true)])`
520 /// The final `Pair(Some(_), true)` is then the resulting witness.
521 #[derive(Clone, Debug)]
522 pub struct Witness<'tcx>(Vec<Pat<'tcx>>);
524 impl<'tcx> Witness<'tcx> {
525 pub fn single_pattern(self) -> Pat<'tcx> {
526 assert_eq!(self.0.len(), 1);
527 self.0.into_iter().next().unwrap()
530 fn push_wild_constructor<'a>(
532 cx: &MatchCheckCtxt<'a, 'tcx>,
533 ctor: &Constructor<'tcx>,
537 let sub_pattern_tys = constructor_sub_pattern_tys(cx, ctor, ty);
538 self.0.extend(sub_pattern_tys.into_iter().map(|ty| {
542 kind: box PatKind::Wild,
545 self.apply_constructor(cx, ctor, ty)
548 /// Constructs a partial witness for a pattern given a list of
549 /// patterns expanded by the specialization step.
551 /// When a pattern P is discovered to be useful, this function is used bottom-up
552 /// to reconstruct a complete witness, e.g., a pattern P' that covers a subset
553 /// of values, V, where each value in that set is not covered by any previously
554 /// used patterns and is covered by the pattern P'. Examples:
556 /// left_ty: tuple of 3 elements
557 /// pats: [10, 20, _] => (10, 20, _)
559 /// left_ty: struct X { a: (bool, &'static str), b: usize}
560 /// pats: [(false, "foo"), 42] => X { a: (false, "foo"), b: 42 }
561 fn apply_constructor<'a>(
563 cx: &MatchCheckCtxt<'a,'tcx>,
564 ctor: &Constructor<'tcx>,
568 let arity = constructor_arity(cx, ctor, ty);
570 let len = self.0.len() as u64;
571 let mut pats = self.0.drain((len - arity) as usize..).rev();
576 let pats = pats.enumerate().map(|(i, p)| {
578 field: Field::new(i),
583 if let ty::Adt(adt, substs) = ty.kind {
588 variant_index: ctor.variant_index_for_adt(cx, adt),
592 PatKind::Leaf { subpatterns: pats }
595 PatKind::Leaf { subpatterns: pats }
600 PatKind::Deref { subpattern: pats.nth(0).unwrap() }
603 ty::Slice(_) | ty::Array(..) => {
605 prefix: pats.collect(),
613 ConstantValue(value) => PatKind::Constant { value },
614 ConstantRange(lo, hi, ty, end) => PatKind::Range(PatRange {
615 lo: ty::Const::from_bits(cx.tcx, lo, ty::ParamEnv::empty().and(ty)),
616 hi: ty::Const::from_bits(cx.tcx, hi, ty::ParamEnv::empty().and(ty)),
635 /// This determines the set of all possible constructors of a pattern matching
636 /// values of type `left_ty`. For vectors, this would normally be an infinite set
637 /// but is instead bounded by the maximum fixed length of slice patterns in
638 /// the column of patterns being analyzed.
640 /// We make sure to omit constructors that are statically impossible. E.g., for
641 /// `Option<!>`, we do not include `Some(_)` in the returned list of constructors.
642 fn all_constructors<'a, 'tcx>(
643 cx: &mut MatchCheckCtxt<'a, 'tcx>,
645 ) -> Vec<Constructor<'tcx>> {
646 debug!("all_constructors({:?})", pcx.ty);
647 let ctors = match pcx.ty.kind {
649 [true, false].iter().map(|&b| {
650 ConstantValue(ty::Const::from_bool(cx.tcx, b))
653 ty::Array(ref sub_ty, len) if len.try_eval_usize(cx.tcx, cx.param_env).is_some() => {
654 let len = len.eval_usize(cx.tcx, cx.param_env);
655 if len != 0 && cx.is_uninhabited(sub_ty) {
661 // Treat arrays of a constant but unknown length like slices.
662 ty::Array(ref sub_ty, _) |
663 ty::Slice(ref sub_ty) => {
664 if cx.is_uninhabited(sub_ty) {
667 (0..pcx.max_slice_length+1).map(|length| Slice(length)).collect()
670 ty::Adt(def, substs) if def.is_enum() => {
673 !cx.tcx.features().exhaustive_patterns ||
674 !v.uninhabited_from(cx.tcx, substs, def.adt_kind()).contains(cx.tcx, cx.module)
676 .map(|v| Variant(v.def_id))
681 // The valid Unicode Scalar Value ranges.
682 ConstantRange('\u{0000}' as u128,
687 ConstantRange('\u{E000}' as u128,
688 '\u{10FFFF}' as u128,
695 let bits = Integer::from_attr(&cx.tcx, SignedInt(ity)).size().bits() as u128;
696 let min = 1u128 << (bits - 1);
698 vec![ConstantRange(min, max, pcx.ty, RangeEnd::Included)]
701 let size = Integer::from_attr(&cx.tcx, UnsignedInt(uty)).size();
702 let max = truncate(u128::max_value(), size);
703 vec![ConstantRange(0, max, pcx.ty, RangeEnd::Included)]
706 if cx.is_uninhabited(pcx.ty) {
716 fn max_slice_length<'p, 'a, 'tcx, I>(cx: &mut MatchCheckCtxt<'a, 'tcx>, patterns: I) -> u64
718 I: Iterator<Item = &'p Pat<'tcx>>,
721 // The exhaustiveness-checking paper does not include any details on
722 // checking variable-length slice patterns. However, they are matched
723 // by an infinite collection of fixed-length array patterns.
725 // Checking the infinite set directly would take an infinite amount
726 // of time. However, it turns out that for each finite set of
727 // patterns `P`, all sufficiently large array lengths are equivalent:
729 // Each slice `s` with a "sufficiently-large" length `l ≥ L` that applies
730 // to exactly the subset `Pₜ` of `P` can be transformed to a slice
731 // `sₘ` for each sufficiently-large length `m` that applies to exactly
732 // the same subset of `P`.
734 // Because of that, each witness for reachability-checking from one
735 // of the sufficiently-large lengths can be transformed to an
736 // equally-valid witness from any other length, so we only have
737 // to check slice lengths from the "minimal sufficiently-large length"
740 // Note that the fact that there is a *single* `sₘ` for each `m`
741 // not depending on the specific pattern in `P` is important: if
742 // you look at the pair of patterns
745 // Then any slice of length ≥1 that matches one of these two
746 // patterns can be trivially turned to a slice of any
747 // other length ≥1 that matches them and vice-versa - for
748 // but the slice from length 2 `[false, true]` that matches neither
749 // of these patterns can't be turned to a slice from length 1 that
750 // matches neither of these patterns, so we have to consider
751 // slices from length 2 there.
753 // Now, to see that that length exists and find it, observe that slice
754 // patterns are either "fixed-length" patterns (`[_, _, _]`) or
755 // "variable-length" patterns (`[_, .., _]`).
757 // For fixed-length patterns, all slices with lengths *longer* than
758 // the pattern's length have the same outcome (of not matching), so
759 // as long as `L` is greater than the pattern's length we can pick
760 // any `sₘ` from that length and get the same result.
762 // For variable-length patterns, the situation is more complicated,
763 // because as seen above the precise value of `sₘ` matters.
765 // However, for each variable-length pattern `p` with a prefix of length
766 // `plₚ` and suffix of length `slₚ`, only the first `plₚ` and the last
767 // `slₚ` elements are examined.
769 // Therefore, as long as `L` is positive (to avoid concerns about empty
770 // types), all elements after the maximum prefix length and before
771 // the maximum suffix length are not examined by any variable-length
772 // pattern, and therefore can be added/removed without affecting
773 // them - creating equivalent patterns from any sufficiently-large
776 // Of course, if fixed-length patterns exist, we must be sure
777 // that our length is large enough to miss them all, so
778 // we can pick `L = max(FIXED_LEN+1 ∪ {max(PREFIX_LEN) + max(SUFFIX_LEN)})`
780 // for example, with the above pair of patterns, all elements
781 // but the first and last can be added/removed, so any
782 // witness of length ≥2 (say, `[false, false, true]`) can be
783 // turned to a witness from any other length ≥2.
785 let mut max_prefix_len = 0;
786 let mut max_suffix_len = 0;
787 let mut max_fixed_len = 0;
789 for row in patterns {
791 PatKind::Constant { value } => {
792 // extract the length of an array/slice from a constant
793 match (value.val, &value.ty.kind) {
794 (_, ty::Array(_, n)) => max_fixed_len = cmp::max(
796 n.eval_usize(cx.tcx, cx.param_env),
798 (ConstValue::Slice{ start, end, .. }, ty::Slice(_)) => max_fixed_len = cmp::max(
800 (end - start) as u64,
805 PatKind::Slice { ref prefix, slice: None, ref suffix } => {
806 let fixed_len = prefix.len() as u64 + suffix.len() as u64;
807 max_fixed_len = cmp::max(max_fixed_len, fixed_len);
809 PatKind::Slice { ref prefix, slice: Some(_), ref suffix } => {
810 max_prefix_len = cmp::max(max_prefix_len, prefix.len() as u64);
811 max_suffix_len = cmp::max(max_suffix_len, suffix.len() as u64);
817 cmp::max(max_fixed_len + 1, max_prefix_len + max_suffix_len)
820 /// An inclusive interval, used for precise integer exhaustiveness checking.
821 /// `IntRange`s always store a contiguous range. This means that values are
822 /// encoded such that `0` encodes the minimum value for the integer,
823 /// regardless of the signedness.
824 /// For example, the pattern `-128..=127i8` is encoded as `0..=255`.
825 /// This makes comparisons and arithmetic on interval endpoints much more
826 /// straightforward. See `signed_bias` for details.
828 /// `IntRange` is never used to encode an empty range or a "range" that wraps
829 /// around the (offset) space: i.e., `range.lo <= range.hi`.
831 struct IntRange<'tcx> {
832 pub range: RangeInclusive<u128>,
836 impl<'tcx> IntRange<'tcx> {
838 fn is_integral(ty: Ty<'_>) -> bool {
840 ty::Char | ty::Int(_) | ty::Uint(_) => true,
846 fn integral_size_and_signed_bias(tcx: TyCtxt<'tcx>, ty: Ty<'_>) -> Option<(Size, u128)> {
848 ty::Char => Some((Size::from_bytes(4), 0)),
850 let size = Integer::from_attr(&tcx, SignedInt(ity)).size();
851 Some((size, 1u128 << (size.bits() as u128 - 1)))
853 ty::Uint(uty) => Some((Integer::from_attr(&tcx, UnsignedInt(uty)).size(), 0)),
861 param_env: ty::ParamEnv<'tcx>,
863 ) -> Option<IntRange<'tcx>> {
864 if let Some((target_size, bias)) = Self::integral_size_and_signed_bias(tcx, value.ty) {
866 let val = if let ConstValue::Scalar(Scalar::Raw { data, size }) = value.val {
867 // For this specific pattern we can skip a lot of effort and go
868 // straight to the result, after doing a bit of checking. (We
869 // could remove this branch and just use the next branch, which
870 // is more general but much slower.)
871 Scalar::<()>::check_raw(data, size, target_size);
873 } else if let Some(val) = value.try_eval_bits(tcx, param_env, ty) {
874 // This is a more general form of the previous branch.
879 let val = val ^ bias;
880 Some(IntRange { range: val..=val, ty })
893 ) -> Option<IntRange<'tcx>> {
894 if Self::is_integral(ty) {
895 // Perform a shift if the underlying types are signed,
896 // which makes the interval arithmetic simpler.
897 let bias = IntRange::signed_bias(tcx, ty);
898 let (lo, hi) = (lo ^ bias, hi ^ bias);
899 // Make sure the interval is well-formed.
900 if lo > hi || lo == hi && *end == RangeEnd::Excluded {
903 let offset = (*end == RangeEnd::Excluded) as u128;
904 Some(IntRange { range: lo..=(hi - offset), ty })
913 param_env: ty::ParamEnv<'tcx>,
914 ctor: &Constructor<'tcx>,
915 ) -> Option<IntRange<'tcx>> {
916 // Floating-point ranges are permitted and we don't want
917 // to consider them when constructing integer ranges.
919 ConstantRange(lo, hi, ty, end) => Self::from_range(tcx, *lo, *hi, ty, end),
920 ConstantValue(val) => Self::from_const(tcx, param_env, val),
927 param_env: ty::ParamEnv<'tcx>,
929 ) -> Option<IntRange<'tcx>> {
932 box PatKind::Constant { value } => {
933 return Self::from_const(tcx, param_env, value);
935 box PatKind::Range(PatRange { lo, hi, end }) => {
936 return Self::from_range(
938 lo.eval_bits(tcx, param_env, lo.ty),
939 hi.eval_bits(tcx, param_env, hi.ty),
944 box PatKind::AscribeUserType { ref subpattern, .. } => {
952 // The return value of `signed_bias` should be XORed with an endpoint to encode/decode it.
953 fn signed_bias(tcx: TyCtxt<'tcx>, ty: Ty<'tcx>) -> u128 {
956 let bits = Integer::from_attr(&tcx, SignedInt(ity)).size().bits() as u128;
963 /// Converts a `RangeInclusive` to a `ConstantValue` or inclusive `ConstantRange`.
967 r: RangeInclusive<u128>,
968 ) -> Constructor<'tcx> {
969 let bias = IntRange::signed_bias(tcx, ty);
970 let (lo, hi) = r.into_inner();
972 let ty = ty::ParamEnv::empty().and(ty);
973 ConstantValue(ty::Const::from_bits(tcx, lo ^ bias, ty))
975 ConstantRange(lo ^ bias, hi ^ bias, ty, RangeEnd::Included)
979 /// Returns a collection of ranges that spans the values covered by `ranges`, subtracted
980 /// by the values covered by `self`: i.e., `ranges \ self` (in set notation).
984 param_env: ty::ParamEnv<'tcx>,
985 ranges: Vec<Constructor<'tcx>>,
986 ) -> Vec<Constructor<'tcx>> {
987 let ranges = ranges.into_iter().filter_map(|r| {
988 IntRange::from_ctor(tcx, param_env, &r).map(|i| i.range)
990 let mut remaining_ranges = vec![];
992 let (lo, hi) = self.range.into_inner();
993 for subrange in ranges {
994 let (subrange_lo, subrange_hi) = subrange.into_inner();
995 if lo > subrange_hi || subrange_lo > hi {
996 // The pattern doesn't intersect with the subrange at all,
997 // so the subrange remains untouched.
998 remaining_ranges.push(Self::range_to_ctor(tcx, ty, subrange_lo..=subrange_hi));
1000 if lo > subrange_lo {
1001 // The pattern intersects an upper section of the
1002 // subrange, so a lower section will remain.
1003 remaining_ranges.push(Self::range_to_ctor(tcx, ty, subrange_lo..=(lo - 1)));
1005 if hi < subrange_hi {
1006 // The pattern intersects a lower section of the
1007 // subrange, so an upper section will remain.
1008 remaining_ranges.push(Self::range_to_ctor(tcx, ty, (hi + 1)..=subrange_hi));
1015 fn intersection(&self, other: &Self) -> Option<Self> {
1017 let (lo, hi) = (*self.range.start(), *self.range.end());
1018 let (other_lo, other_hi) = (*other.range.start(), *other.range.end());
1019 if lo <= other_hi && other_lo <= hi {
1020 Some(IntRange { range: max(lo, other_lo)..=min(hi, other_hi), ty })
1027 // A request for missing constructor data in terms of either:
1028 // - whether or not there any missing constructors; or
1029 // - the actual set of missing constructors.
1030 #[derive(PartialEq)]
1031 enum MissingCtorsInfo {
1036 // Used by `compute_missing_ctors`.
1037 #[derive(Debug, PartialEq)]
1038 enum MissingCtors<'tcx> {
1042 // Note that the Vec can be empty.
1043 Ctors(Vec<Constructor<'tcx>>),
1046 // When `info` is `MissingCtorsInfo::Ctors`, compute a set of constructors
1047 // equivalent to `all_ctors \ used_ctors`. When `info` is
1048 // `MissingCtorsInfo::Emptiness`, just determines if that set is empty or not.
1049 // (The split logic gives a performance win, because we always need to know if
1050 // the set is empty, but we rarely need the full set, and it can be expensive
1051 // to compute the full set.)
1052 fn compute_missing_ctors<'tcx>(
1053 info: MissingCtorsInfo,
1055 param_env: ty::ParamEnv<'tcx>,
1056 all_ctors: &Vec<Constructor<'tcx>>,
1057 used_ctors: &Vec<Constructor<'tcx>>,
1058 ) -> MissingCtors<'tcx> {
1059 let mut missing_ctors = vec![];
1061 for req_ctor in all_ctors {
1062 let mut refined_ctors = vec![req_ctor.clone()];
1063 for used_ctor in used_ctors {
1064 if used_ctor == req_ctor {
1065 // If a constructor appears in a `match` arm, we can
1066 // eliminate it straight away.
1067 refined_ctors = vec![]
1068 } else if let Some(interval) = IntRange::from_ctor(tcx, param_env, used_ctor) {
1069 // Refine the required constructors for the type by subtracting
1070 // the range defined by the current constructor pattern.
1071 refined_ctors = interval.subtract_from(tcx, param_env, refined_ctors);
1074 // If the constructor patterns that have been considered so far
1075 // already cover the entire range of values, then we the
1076 // constructor is not missing, and we can move on to the next one.
1077 if refined_ctors.is_empty() {
1081 // If a constructor has not been matched, then it is missing.
1082 // We add `refined_ctors` instead of `req_ctor`, because then we can
1083 // provide more detailed error information about precisely which
1084 // ranges have been omitted.
1085 if info == MissingCtorsInfo::Emptiness {
1086 if !refined_ctors.is_empty() {
1087 // The set is non-empty; return early.
1088 return MissingCtors::NonEmpty;
1091 missing_ctors.extend(refined_ctors);
1095 if info == MissingCtorsInfo::Emptiness {
1096 // If we reached here, the set is empty.
1099 MissingCtors::Ctors(missing_ctors)
1103 /// Algorithm from http://moscova.inria.fr/~maranget/papers/warn/index.html.
1104 /// The algorithm from the paper has been modified to correctly handle empty
1105 /// types. The changes are:
1106 /// (0) We don't exit early if the pattern matrix has zero rows. We just
1107 /// continue to recurse over columns.
1108 /// (1) all_constructors will only return constructors that are statically
1109 /// possible. E.g., it will only return `Ok` for `Result<T, !>`.
1111 /// This finds whether a (row) vector `v` of patterns is 'useful' in relation
1112 /// to a set of such vectors `m` - this is defined as there being a set of
1113 /// inputs that will match `v` but not any of the sets in `m`.
1115 /// All the patterns at each column of the `matrix ++ v` matrix must
1116 /// have the same type, except that wildcard (PatKind::Wild) patterns
1117 /// with type `TyErr` are also allowed, even if the "type of the column"
1118 /// is not `TyErr`. That is used to represent private fields, as using their
1119 /// real type would assert that they are inhabited.
1121 /// This is used both for reachability checking (if a pattern isn't useful in
1122 /// relation to preceding patterns, it is not reachable) and exhaustiveness
1123 /// checking (if a wildcard pattern is useful in relation to a matrix, the
1124 /// matrix isn't exhaustive).
1125 pub fn is_useful<'p, 'a, 'tcx>(
1126 cx: &mut MatchCheckCtxt<'a, 'tcx>,
1127 matrix: &Matrix<'p, 'tcx>,
1129 witness: WitnessPreference,
1130 ) -> Usefulness<'tcx> {
1131 let &Matrix(ref rows) = matrix;
1132 debug!("is_useful({:#?}, {:#?})", matrix, v);
1134 // The base case. We are pattern-matching on () and the return value is
1135 // based on whether our matrix has a row or not.
1136 // NOTE: This could potentially be optimized by checking rows.is_empty()
1137 // first and then, if v is non-empty, the return value is based on whether
1138 // the type of the tuple we're checking is inhabited or not.
1140 return if rows.is_empty() {
1142 ConstructWitness => UsefulWithWitness(vec![Witness(vec![])]),
1143 LeaveOutWitness => Useful,
1150 assert!(rows.iter().all(|r| r.len() == v.len()));
1153 // TyErr is used to represent the type of wildcard patterns matching
1154 // against inaccessible (private) fields of structs, so that we won't
1155 // be able to observe whether the types of the struct's fields are
1158 // If the field is truly inaccessible, then all the patterns
1159 // matching against it must be wildcard patterns, so its type
1162 // However, if we are matching against non-wildcard patterns, we
1163 // need to know the real type of the field so we can specialize
1164 // against it. This primarily occurs through constants - they
1165 // can include contents for fields that are inaccessible at the
1166 // location of the match. In that case, the field's type is
1167 // inhabited - by the constant - so we can just use it.
1169 // FIXME: this might lead to "unstable" behavior with macro hygiene
1170 // introducing uninhabited patterns for inaccessible fields. We
1171 // need to figure out how to model that.
1172 ty: rows.iter().map(|r| r[0].ty).find(|ty| !ty.references_error()).unwrap_or(v[0].ty),
1173 max_slice_length: max_slice_length(cx, rows.iter().map(|r| r[0]).chain(Some(v[0])))
1176 debug!("is_useful_expand_first_col: pcx={:#?}, expanding {:#?}", pcx, v[0]);
1178 if let Some(constructors) = pat_constructors(cx, v[0], pcx) {
1179 let is_declared_nonexhaustive = cx.is_non_exhaustive_variant(v[0]) && !cx.is_local(pcx.ty);
1180 debug!("is_useful - expanding constructors: {:#?}, is_declared_nonexhaustive: {:?}",
1181 constructors, is_declared_nonexhaustive);
1183 if is_declared_nonexhaustive {
1186 split_grouped_constructors(
1187 cx.tcx, cx.param_env, constructors, matrix, pcx.ty,
1188 ).into_iter().map(|c|
1189 is_useful_specialized(cx, matrix, v, c, pcx.ty, witness)
1190 ).find(|result| result.is_useful()).unwrap_or(NotUseful)
1193 debug!("is_useful - expanding wildcard");
1195 let used_ctors: Vec<Constructor<'_>> = rows.iter().flat_map(|row| {
1196 pat_constructors(cx, row[0], pcx).unwrap_or(vec![])
1198 debug!("used_ctors = {:#?}", used_ctors);
1199 // `all_ctors` are all the constructors for the given type, which
1200 // should all be represented (or caught with the wild pattern `_`).
1201 let all_ctors = all_constructors(cx, pcx);
1202 debug!("all_ctors = {:#?}", all_ctors);
1204 // `missing_ctors` is the set of constructors from the same type as the
1205 // first column of `matrix` that are matched only by wildcard patterns
1206 // from the first column.
1208 // Therefore, if there is some pattern that is unmatched by `matrix`,
1209 // it will still be unmatched if the first constructor is replaced by
1210 // any of the constructors in `missing_ctors`
1212 // However, if our scrutinee is *privately* an empty enum, we
1213 // must treat it as though it had an "unknown" constructor (in
1214 // that case, all other patterns obviously can't be variants)
1215 // to avoid exposing its emptyness. See the `match_privately_empty`
1216 // test for details.
1218 // FIXME: currently the only way I know of something can
1219 // be a privately-empty enum is when the exhaustive_patterns
1220 // feature flag is not present, so this is only
1221 // needed for that case.
1223 // Missing constructors are those that are not matched by any
1224 // non-wildcard patterns in the current column. We always determine if
1225 // the set is empty, but we only fully construct them on-demand,
1226 // because they're rarely used and can be big.
1227 let cheap_missing_ctors = compute_missing_ctors(
1228 MissingCtorsInfo::Emptiness, cx.tcx, cx.param_env, &all_ctors, &used_ctors,
1231 let is_privately_empty = all_ctors.is_empty() && !cx.is_uninhabited(pcx.ty);
1232 let is_declared_nonexhaustive = cx.is_non_exhaustive_enum(pcx.ty) && !cx.is_local(pcx.ty);
1233 debug!("cheap_missing_ctors={:#?} is_privately_empty={:#?} is_declared_nonexhaustive={:#?}",
1234 cheap_missing_ctors, is_privately_empty, is_declared_nonexhaustive);
1236 // For privately empty and non-exhaustive enums, we work as if there were an "extra"
1237 // `_` constructor for the type, so we can never match over all constructors.
1238 let is_non_exhaustive = is_privately_empty || is_declared_nonexhaustive ||
1239 (pcx.ty.is_ptr_sized_integral() && !cx.tcx.features().precise_pointer_size_matching);
1241 if cheap_missing_ctors == MissingCtors::Empty && !is_non_exhaustive {
1242 split_grouped_constructors(cx.tcx, cx.param_env, all_ctors, matrix, pcx.ty)
1243 .into_iter().map(|c| is_useful_specialized(cx, matrix, v, c, pcx.ty, witness))
1244 .find(|result| result.is_useful())
1245 .unwrap_or(NotUseful)
1247 let matrix = rows.iter().filter_map(|r| {
1248 if r[0].is_wildcard() {
1249 Some(SmallVec::from_slice(&r[1..]))
1254 match is_useful(cx, &matrix, &v[1..], witness) {
1255 UsefulWithWitness(pats) => {
1257 // In this case, there's at least one "free"
1258 // constructor that is only matched against by
1259 // wildcard patterns.
1261 // There are 2 ways we can report a witness here.
1262 // Commonly, we can report all the "free"
1263 // constructors as witnesses, e.g., if we have:
1266 // enum Direction { N, S, E, W }
1267 // let Direction::N = ...;
1270 // we can report 3 witnesses: `S`, `E`, and `W`.
1272 // However, there are 2 cases where we don't want
1273 // to do this and instead report a single `_` witness:
1275 // 1) If the user is matching against a non-exhaustive
1276 // enum, there is no point in enumerating all possible
1277 // variants, because the user can't actually match
1278 // against them himself, e.g., in an example like:
1280 // let err: io::ErrorKind = ...;
1282 // io::ErrorKind::NotFound => {},
1285 // we don't want to show every possible IO error,
1286 // but instead have `_` as the witness (this is
1287 // actually *required* if the user specified *all*
1288 // IO errors, but is probably what we want in every
1291 // 2) If the user didn't actually specify a constructor
1292 // in this arm, e.g., in
1294 // let x: (Direction, Direction, bool) = ...;
1295 // let (_, _, false) = x;
1297 // we don't want to show all 16 possible witnesses
1298 // `(<direction-1>, <direction-2>, true)` - we are
1299 // satisfied with `(_, _, true)`. In this case,
1300 // `used_ctors` is empty.
1301 let new_witnesses = if is_non_exhaustive || used_ctors.is_empty() {
1302 // All constructors are unused. Add wild patterns
1303 // rather than each individual constructor.
1304 pats.into_iter().map(|mut witness| {
1305 witness.0.push(Pat {
1308 kind: box PatKind::Wild,
1313 let expensive_missing_ctors = compute_missing_ctors(
1314 MissingCtorsInfo::Ctors, cx.tcx, cx.param_env, &all_ctors, &used_ctors,
1316 if let MissingCtors::Ctors(missing_ctors) = expensive_missing_ctors {
1317 pats.into_iter().flat_map(|witness| {
1318 missing_ctors.iter().map(move |ctor| {
1319 // Extends the witness with a "wild" version of this
1320 // constructor, that matches everything that can be built with
1321 // it. For example, if `ctor` is a `Constructor::Variant` for
1322 // `Option::Some`, this pushes the witness for `Some(_)`.
1323 witness.clone().push_wild_constructor(cx, ctor, pcx.ty)
1327 bug!("cheap missing ctors")
1330 UsefulWithWitness(new_witnesses)
1338 /// A shorthand for the `U(S(c, P), S(c, q))` operation from the paper. I.e., `is_useful` applied
1339 /// to the specialised version of both the pattern matrix `P` and the new pattern `q`.
1340 fn is_useful_specialized<'p, 'a, 'tcx>(
1341 cx: &mut MatchCheckCtxt<'a, 'tcx>,
1342 &Matrix(ref m): &Matrix<'p, 'tcx>,
1344 ctor: Constructor<'tcx>,
1346 witness: WitnessPreference,
1347 ) -> Usefulness<'tcx> {
1348 debug!("is_useful_specialized({:#?}, {:#?}, {:?})", v, ctor, lty);
1349 let sub_pat_tys = constructor_sub_pattern_tys(cx, &ctor, lty);
1350 let wild_patterns_owned: Vec<_> = sub_pat_tys.iter().map(|ty| {
1354 kind: box PatKind::Wild,
1357 let wild_patterns: Vec<_> = wild_patterns_owned.iter().collect();
1358 let matrix = Matrix(
1360 .filter_map(|r| specialize(cx, &r, &ctor, &wild_patterns))
1363 match specialize(cx, v, &ctor, &wild_patterns) {
1364 Some(v) => match is_useful(cx, &matrix, &v, witness) {
1365 UsefulWithWitness(witnesses) => UsefulWithWitness(
1366 witnesses.into_iter()
1367 .map(|witness| witness.apply_constructor(cx, &ctor, lty))
1376 /// Determines the constructors that the given pattern can be specialized to.
1378 /// In most cases, there's only one constructor that a specific pattern
1379 /// represents, such as a specific enum variant or a specific literal value.
1380 /// Slice patterns, however, can match slices of different lengths. For instance,
1381 /// `[a, b, ..tail]` can match a slice of length 2, 3, 4 and so on.
1383 /// Returns `None` in case of a catch-all, which can't be specialized.
1384 fn pat_constructors<'tcx>(cx: &mut MatchCheckCtxt<'_, 'tcx>,
1387 -> Option<Vec<Constructor<'tcx>>>
1390 PatKind::AscribeUserType { ref subpattern, .. } =>
1391 pat_constructors(cx, subpattern, pcx),
1392 PatKind::Binding { .. } | PatKind::Wild => None,
1393 PatKind::Leaf { .. } | PatKind::Deref { .. } => Some(vec![Single]),
1394 PatKind::Variant { adt_def, variant_index, .. } => {
1395 Some(vec![Variant(adt_def.variants[variant_index].def_id)])
1397 PatKind::Constant { value } => Some(vec![ConstantValue(value)]),
1398 PatKind::Range(PatRange { lo, hi, end }) =>
1399 Some(vec![ConstantRange(
1400 lo.eval_bits(cx.tcx, cx.param_env, lo.ty),
1401 hi.eval_bits(cx.tcx, cx.param_env, hi.ty),
1405 PatKind::Array { .. } => match pcx.ty.kind {
1406 ty::Array(_, length) => Some(vec![
1407 Slice(length.eval_usize(cx.tcx, cx.param_env))
1409 _ => span_bug!(pat.span, "bad ty {:?} for array pattern", pcx.ty)
1411 PatKind::Slice { ref prefix, ref slice, ref suffix } => {
1412 let pat_len = prefix.len() as u64 + suffix.len() as u64;
1413 if slice.is_some() {
1414 Some((pat_len..pcx.max_slice_length+1).map(Slice).collect())
1416 Some(vec![Slice(pat_len)])
1419 PatKind::Or { .. } => {
1420 bug!("support for or-patterns has not been fully implemented yet.");
1425 /// This computes the arity of a constructor. The arity of a constructor
1426 /// is how many subpattern patterns of that constructor should be expanded to.
1428 /// For instance, a tuple pattern `(_, 42, Some([]))` has the arity of 3.
1429 /// A struct pattern's arity is the number of fields it contains, etc.
1430 fn constructor_arity(cx: &MatchCheckCtxt<'a, 'tcx>, ctor: &Constructor<'tcx>, ty: Ty<'tcx>) -> u64 {
1431 debug!("constructor_arity({:#?}, {:?})", ctor, ty);
1433 ty::Tuple(ref fs) => fs.len() as u64,
1434 ty::Slice(..) | ty::Array(..) => match *ctor {
1435 Slice(length) => length,
1436 ConstantValue(_) => 0,
1437 _ => bug!("bad slice pattern {:?} {:?}", ctor, ty)
1440 ty::Adt(adt, _) => {
1441 adt.variants[ctor.variant_index_for_adt(cx, adt)].fields.len() as u64
1447 /// This computes the types of the sub patterns that a constructor should be
1450 /// For instance, a tuple pattern (43u32, 'a') has sub pattern types [u32, char].
1451 fn constructor_sub_pattern_tys<'a, 'tcx>(
1452 cx: &MatchCheckCtxt<'a, 'tcx>,
1453 ctor: &Constructor<'tcx>,
1455 ) -> Vec<Ty<'tcx>> {
1456 debug!("constructor_sub_pattern_tys({:#?}, {:?})", ctor, ty);
1458 ty::Tuple(ref fs) => fs.into_iter().map(|t| t.expect_ty()).collect(),
1459 ty::Slice(ty) | ty::Array(ty, _) => match *ctor {
1460 Slice(length) => (0..length).map(|_| ty).collect(),
1461 ConstantValue(_) => vec![],
1462 _ => bug!("bad slice pattern {:?} {:?}", ctor, ty)
1464 ty::Ref(_, rty, _) => vec![rty],
1465 ty::Adt(adt, substs) => {
1467 // Use T as the sub pattern type of Box<T>.
1468 vec![substs.type_at(0)]
1470 adt.variants[ctor.variant_index_for_adt(cx, adt)].fields.iter().map(|field| {
1471 let is_visible = adt.is_enum()
1472 || field.vis.is_accessible_from(cx.module, cx.tcx);
1474 let ty = field.ty(cx.tcx, substs);
1476 // If the field type returned is an array of an unknown
1477 // size return an TyErr.
1479 if len.try_eval_usize(cx.tcx, cx.param_env).is_none() =>
1484 // Treat all non-visible fields as TyErr. They
1485 // can't appear in any other pattern from
1486 // this match (because they are private),
1487 // so their type does not matter - but
1488 // we don't want to know they are
1499 // checks whether a constant is equal to a user-written slice pattern. Only supports byte slices,
1500 // meaning all other types will compare unequal and thus equal patterns often do not cause the
1501 // second pattern to lint about unreachable match arms.
1502 fn slice_pat_covered_by_const<'tcx>(
1505 const_val: &'tcx ty::Const<'tcx>,
1506 prefix: &[Pat<'tcx>],
1507 slice: &Option<Pat<'tcx>>,
1508 suffix: &[Pat<'tcx>],
1509 param_env: ty::ParamEnv<'tcx>,
1510 ) -> Result<bool, ErrorReported> {
1511 let data: &[u8] = match (const_val.val, &const_val.ty.kind) {
1512 (ConstValue::ByRef { offset, alloc, .. }, ty::Array(t, n)) => {
1513 assert_eq!(*t, tcx.types.u8);
1514 let n = n.eval_usize(tcx, param_env);
1515 let ptr = Pointer::new(AllocId(0), offset);
1516 alloc.get_bytes(&tcx, ptr, Size::from_bytes(n)).unwrap()
1518 (ConstValue::Slice { data, start, end }, ty::Slice(t)) => {
1519 assert_eq!(*t, tcx.types.u8);
1520 let ptr = Pointer::new(AllocId(0), Size::from_bytes(start as u64));
1521 data.get_bytes(&tcx, ptr, Size::from_bytes((end - start) as u64)).unwrap()
1523 // FIXME(oli-obk): create a way to extract fat pointers from ByRef
1524 (_, ty::Slice(_)) => return Ok(false),
1526 "slice_pat_covered_by_const: {:#?}, {:#?}, {:#?}, {:#?}",
1527 const_val, prefix, slice, suffix,
1531 let pat_len = prefix.len() + suffix.len();
1532 if data.len() < pat_len || (slice.is_none() && data.len() > pat_len) {
1537 data[..prefix.len()].iter().zip(prefix).chain(
1538 data[data.len()-suffix.len()..].iter().zip(suffix))
1541 box PatKind::Constant { value } => {
1542 let b = value.eval_bits(tcx, param_env, pat.ty);
1543 assert_eq!(b as u8 as u128, b);
1555 // Whether to evaluate a constructor using exhaustive integer matching. This is true if the
1556 // constructor is a range or constant with an integer type.
1557 fn should_treat_range_exhaustively(tcx: TyCtxt<'tcx>, ctor: &Constructor<'tcx>) -> bool {
1558 let ty = match ctor {
1559 ConstantValue(value) => value.ty,
1560 ConstantRange(_, _, ty, _) => ty,
1563 if let ty::Char | ty::Int(_) | ty::Uint(_) = ty.kind {
1564 !ty.is_ptr_sized_integral() || tcx.features().precise_pointer_size_matching
1570 /// For exhaustive integer matching, some constructors are grouped within other constructors
1571 /// (namely integer typed values are grouped within ranges). However, when specialising these
1572 /// constructors, we want to be specialising for the underlying constructors (the integers), not
1573 /// the groups (the ranges). Thus we need to split the groups up. Splitting them up naïvely would
1574 /// mean creating a separate constructor for every single value in the range, which is clearly
1575 /// impractical. However, observe that for some ranges of integers, the specialisation will be
1576 /// identical across all values in that range (i.e., there are equivalence classes of ranges of
1577 /// constructors based on their `is_useful_specialized` outcome). These classes are grouped by
1578 /// the patterns that apply to them (in the matrix `P`). We can split the range whenever the
1579 /// patterns that apply to that range (specifically: the patterns that *intersect* with that range)
1581 /// Our solution, therefore, is to split the range constructor into subranges at every single point
1582 /// the group of intersecting patterns changes (using the method described below).
1583 /// And voilà! We're testing precisely those ranges that we need to, without any exhaustive matching
1584 /// on actual integers. The nice thing about this is that the number of subranges is linear in the
1585 /// number of rows in the matrix (i.e., the number of cases in the `match` statement), so we don't
1586 /// need to be worried about matching over gargantuan ranges.
1588 /// Essentially, given the first column of a matrix representing ranges, looking like the following:
1590 /// |------| |----------| |-------| ||
1591 /// |-------| |-------| |----| ||
1594 /// We split the ranges up into equivalence classes so the ranges are no longer overlapping:
1596 /// |--|--|||-||||--||---|||-------| |-|||| ||
1598 /// The logic for determining how to split the ranges is fairly straightforward: we calculate
1599 /// boundaries for each interval range, sort them, then create constructors for each new interval
1600 /// between every pair of boundary points. (This essentially sums up to performing the intuitive
1601 /// merging operation depicted above.)
1602 fn split_grouped_constructors<'p, 'tcx>(
1604 param_env: ty::ParamEnv<'tcx>,
1605 ctors: Vec<Constructor<'tcx>>,
1606 &Matrix(ref m): &Matrix<'p, 'tcx>,
1608 ) -> Vec<Constructor<'tcx>> {
1609 let mut split_ctors = Vec::with_capacity(ctors.len());
1611 for ctor in ctors.into_iter() {
1613 // For now, only ranges may denote groups of "subconstructors", so we only need to
1614 // special-case constant ranges.
1615 ConstantRange(..) if should_treat_range_exhaustively(tcx, &ctor) => {
1616 // We only care about finding all the subranges within the range of the constructor
1617 // range. Anything else is irrelevant, because it is guaranteed to result in
1618 // `NotUseful`, which is the default case anyway, and can be ignored.
1619 let ctor_range = IntRange::from_ctor(tcx, param_env, &ctor).unwrap();
1621 /// Represents a border between 2 integers. Because the intervals spanning borders
1622 /// must be able to cover every integer, we need to be able to represent
1623 /// 2^128 + 1 such borders.
1624 #[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
1630 // A function for extracting the borders of an integer interval.
1631 fn range_borders(r: IntRange<'_>) -> impl Iterator<Item = Border> {
1632 let (lo, hi) = r.range.into_inner();
1633 let from = Border::JustBefore(lo);
1634 let to = match hi.checked_add(1) {
1635 Some(m) => Border::JustBefore(m),
1636 None => Border::AfterMax,
1638 vec![from, to].into_iter()
1641 // `borders` is the set of borders between equivalence classes: each equivalence
1642 // class lies between 2 borders.
1643 let row_borders = m.iter()
1644 .flat_map(|row| IntRange::from_pat(tcx, param_env, row[0]))
1645 .flat_map(|range| ctor_range.intersection(&range))
1646 .flat_map(|range| range_borders(range));
1647 let ctor_borders = range_borders(ctor_range.clone());
1648 let mut borders: Vec<_> = row_borders.chain(ctor_borders).collect();
1649 borders.sort_unstable();
1651 // We're going to iterate through every pair of borders, making sure that each
1652 // represents an interval of nonnegative length, and convert each such interval
1653 // into a constructor.
1654 for IntRange { range, .. } in borders.windows(2).filter_map(|window| {
1655 match (window[0], window[1]) {
1656 (Border::JustBefore(n), Border::JustBefore(m)) => {
1658 Some(IntRange { range: n..=(m - 1), ty })
1663 (Border::JustBefore(n), Border::AfterMax) => {
1664 Some(IntRange { range: n..=u128::MAX, ty })
1666 (Border::AfterMax, _) => None,
1669 split_ctors.push(IntRange::range_to_ctor(tcx, ty, range));
1672 // Any other constructor can be used unchanged.
1673 _ => split_ctors.push(ctor),
1680 fn constructor_covered_by_range<'tcx>(
1682 param_env: ty::ParamEnv<'tcx>,
1683 ctor: &Constructor<'tcx>,
1685 ) -> Result<bool, ErrorReported> {
1686 let (from, to, end, ty) = match pat.kind {
1687 box PatKind::Constant { value } => (value, value, RangeEnd::Included, value.ty),
1688 box PatKind::Range(PatRange { lo, hi, end }) => (lo, hi, end, lo.ty),
1689 _ => bug!("`constructor_covered_by_range` called with {:?}", pat),
1691 trace!("constructor_covered_by_range {:#?}, {:#?}, {:#?}, {}", ctor, from, to, ty);
1692 let cmp_from = |c_from| compare_const_vals(tcx, c_from, from, param_env, ty)
1693 .map(|res| res != Ordering::Less);
1694 let cmp_to = |c_to| compare_const_vals(tcx, c_to, to, param_env, ty);
1695 macro_rules! some_or_ok {
1699 None => return Ok(false), // not char or int
1704 ConstantValue(value) => {
1705 let to = some_or_ok!(cmp_to(value));
1706 let end = (to == Ordering::Less) ||
1707 (end == RangeEnd::Included && to == Ordering::Equal);
1708 Ok(some_or_ok!(cmp_from(value)) && end)
1710 ConstantRange(from, to, ty, RangeEnd::Included) => {
1711 let to = some_or_ok!(cmp_to(ty::Const::from_bits(
1714 ty::ParamEnv::empty().and(ty),
1716 let end = (to == Ordering::Less) ||
1717 (end == RangeEnd::Included && to == Ordering::Equal);
1718 Ok(some_or_ok!(cmp_from(ty::Const::from_bits(
1721 ty::ParamEnv::empty().and(ty),
1724 ConstantRange(from, to, ty, RangeEnd::Excluded) => {
1725 let to = some_or_ok!(cmp_to(ty::Const::from_bits(
1728 ty::ParamEnv::empty().and(ty)
1730 let end = (to == Ordering::Less) ||
1731 (end == RangeEnd::Excluded && to == Ordering::Equal);
1732 Ok(some_or_ok!(cmp_from(ty::Const::from_bits(
1735 ty::ParamEnv::empty().and(ty)))
1743 fn patterns_for_variant<'p, 'tcx>(
1744 subpatterns: &'p [FieldPat<'tcx>],
1745 wild_patterns: &[&'p Pat<'tcx>])
1746 -> SmallVec<[&'p Pat<'tcx>; 2]>
1748 let mut result = SmallVec::from_slice(wild_patterns);
1750 for subpat in subpatterns {
1751 result[subpat.field.index()] = &subpat.pattern;
1754 debug!("patterns_for_variant({:#?}, {:#?}) = {:#?}", subpatterns, wild_patterns, result);
1758 /// This is the main specialization step. It expands the first pattern in the given row
1759 /// into `arity` patterns based on the constructor. For most patterns, the step is trivial,
1760 /// for instance tuple patterns are flattened and box patterns expand into their inner pattern.
1762 /// OTOH, slice patterns with a subslice pattern (..tail) can be expanded into multiple
1763 /// different patterns.
1764 /// Structure patterns with a partial wild pattern (Foo { a: 42, .. }) have their missing
1765 /// fields filled with wild patterns.
1766 fn specialize<'p, 'a: 'p, 'tcx>(
1767 cx: &mut MatchCheckCtxt<'a, 'tcx>,
1768 r: &[&'p Pat<'tcx>],
1769 constructor: &Constructor<'tcx>,
1770 wild_patterns: &[&'p Pat<'tcx>],
1771 ) -> Option<SmallVec<[&'p Pat<'tcx>; 2]>> {
1774 let head = match *pat.kind {
1775 PatKind::AscribeUserType { ref subpattern, .. } => {
1776 specialize(cx, ::std::slice::from_ref(&subpattern), constructor, wild_patterns)
1779 PatKind::Binding { .. } | PatKind::Wild => {
1780 Some(SmallVec::from_slice(wild_patterns))
1783 PatKind::Variant { adt_def, variant_index, ref subpatterns, .. } => {
1784 let ref variant = adt_def.variants[variant_index];
1785 Some(Variant(variant.def_id))
1786 .filter(|variant_constructor| variant_constructor == constructor)
1787 .map(|_| patterns_for_variant(subpatterns, wild_patterns))
1790 PatKind::Leaf { ref subpatterns } => {
1791 Some(patterns_for_variant(subpatterns, wild_patterns))
1794 PatKind::Deref { ref subpattern } => {
1795 Some(smallvec![subpattern])
1798 PatKind::Constant { value } if constructor.is_slice() => {
1799 // We extract an `Option` for the pointer because slices of zero
1800 // elements don't necessarily point to memory, they are usually
1801 // just integers. The only time they should be pointing to memory
1802 // is when they are subslices of nonzero slices.
1803 let (alloc, offset, n, ty) = match value.ty.kind {
1804 ty::Array(t, n) => {
1806 ConstValue::ByRef { offset, alloc, .. } => (
1809 n.eval_usize(cx.tcx, cx.param_env),
1814 "array pattern is {:?}", value,
1820 ConstValue::Slice { data, start, end } => (
1822 Size::from_bytes(start as u64),
1823 (end - start) as u64,
1826 ConstValue::ByRef { .. } => {
1827 // FIXME(oli-obk): implement `deref` for `ConstValue`
1832 "slice pattern constant must be scalar pair but is {:?}",
1839 "unexpected const-val {:?} with ctor {:?}",
1844 if wild_patterns.len() as u64 == n {
1845 // convert a constant slice/array pattern to a list of patterns.
1846 let layout = cx.tcx.layout_of(cx.param_env.and(ty)).ok()?;
1847 let ptr = Pointer::new(AllocId(0), offset);
1849 let ptr = ptr.offset(layout.size * i, &cx.tcx).ok()?;
1850 let scalar = alloc.read_scalar(
1851 &cx.tcx, ptr, layout.size,
1853 let scalar = scalar.not_undef().ok()?;
1854 let value = ty::Const::from_scalar(cx.tcx, scalar, ty);
1858 kind: box PatKind::Constant { value },
1860 Some(&*cx.pattern_arena.alloc(pattern))
1867 PatKind::Constant { .. } |
1868 PatKind::Range { .. } => {
1869 // If the constructor is a:
1870 // - Single value: add a row if the pattern contains the constructor.
1871 // - Range: add a row if the constructor intersects the pattern.
1872 if should_treat_range_exhaustively(cx.tcx, constructor) {
1873 match (IntRange::from_ctor(cx.tcx, cx.param_env, constructor),
1874 IntRange::from_pat(cx.tcx, cx.param_env, pat)) {
1875 (Some(ctor), Some(pat)) => {
1876 ctor.intersection(&pat).map(|_| {
1877 let (pat_lo, pat_hi) = pat.range.into_inner();
1878 let (ctor_lo, ctor_hi) = ctor.range.into_inner();
1879 assert!(pat_lo <= ctor_lo && ctor_hi <= pat_hi);
1886 // Fallback for non-ranges and ranges that involve
1887 // floating-point numbers, which are not conveniently handled
1888 // by `IntRange`. For these cases, the constructor may not be a
1889 // range so intersection actually devolves into being covered
1891 match constructor_covered_by_range(cx.tcx, cx.param_env, constructor, pat) {
1892 Ok(true) => Some(smallvec![]),
1893 Ok(false) | Err(ErrorReported) => None,
1898 PatKind::Array { ref prefix, ref slice, ref suffix } |
1899 PatKind::Slice { ref prefix, ref slice, ref suffix } => {
1900 match *constructor {
1902 let pat_len = prefix.len() + suffix.len();
1903 if let Some(slice_count) = wild_patterns.len().checked_sub(pat_len) {
1904 if slice_count == 0 || slice.is_some() {
1905 Some(prefix.iter().chain(
1906 wild_patterns.iter().map(|p| *p)
1909 .chain(suffix.iter())
1918 ConstantValue(cv) => {
1919 match slice_pat_covered_by_const(
1920 cx.tcx, pat.span, cv, prefix, slice, suffix, cx.param_env,
1922 Ok(true) => Some(smallvec![]),
1924 Err(ErrorReported) => None
1927 _ => span_bug!(pat.span,
1928 "unexpected ctor {:?} for slice pat", constructor)
1932 PatKind::Or { .. } => {
1933 bug!("support for or-patterns has not been fully implemented yet.");
1936 debug!("specialize({:#?}, {:#?}) = {:#?}", r[0], wild_patterns, head);
1938 head.map(|mut head| {
1939 head.extend_from_slice(&r[1 ..]);