Rollup merge of #123067 - Nadrieril:always-simplify-or, r=oli-obk

match lowering: consistently merge simple or-patterns

There are two places where we expand or-patterns in match lowering: the main one is `test_candidates_with_or`, and there's one in `match_candidates` that's an optimization for the simple case where the whole pattern is just one or-pattern.

To reduce duplication, we merge or-pattern alternatives into a single block when possible, but we only to that in `test_candidates_with_or`. This PR fixes this oversight and merges them in `match_candidates` too.

This is a part of splitting up #122046 into smaller bits.

Author: matthiaskrgr

compiler/rustc_mir_build/src/build/matches/mod.rs

@@ -1006,6 +1006,10 @@ struct Candidate<'pat, 'tcx> {
     /// If the candidate matches, bindings and ascriptions must be established.
     extra_data: PatternExtraData<'tcx>,
 
+    /// If we filled `self.subcandidate`, we store here the span of the or-pattern they came from.
+    // Invariant: it is `None` iff `subcandidates.is_empty()`.
+    or_span: Option<Span>,
+
     /// The block before the `bindings` have been established.
     pre_binding_block: Option<BasicBlock>,
     /// The pre-binding block of the next candidate.
@@ -1028,6 +1032,7 @@ impl<'tcx, 'pat> Candidate<'pat, 'tcx> {
             extra_data: flat_pat.extra_data,
             has_guard,
             subcandidates: Vec::new(),
+            or_span: None,
             otherwise_block: None,
             pre_binding_block: None,
             next_candidate_pre_binding_block: None,
@@ -1277,7 +1282,8 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
                 //
                 // only generates a single switch.
                 candidate.subcandidates = self.create_or_subcandidates(pats, candidate.has_guard);
-                candidate.match_pairs.pop();
+                let first_match_pair = candidate.match_pairs.pop().unwrap();
+                candidate.or_span = Some(first_match_pair.pattern.span);
                 split_or_candidate = true;
             }
         }
@@ -1287,8 +1293,7 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
                 // At least one of the candidates has been split into subcandidates.
                 // We need to change the candidate list to include those.
                 let mut new_candidates = Vec::new();
-
-                for candidate in candidates {
+                for candidate in candidates.iter_mut() {
                     candidate.visit_leaves(|leaf_candidate| new_candidates.push(leaf_candidate));
                 }
                 self.match_simplified_candidates(
@@ -1298,6 +1303,10 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
                     otherwise_block,
                     &mut *new_candidates,
                 );
+
+                for candidate in candidates {
+                    self.merge_trivial_subcandidates(candidate);
+                }
             } else {
                 self.match_simplified_candidates(
                     span,
@@ -1531,16 +1540,13 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
             &mut or_candidate_refs,
         );
         candidate.subcandidates = or_candidates;
-        self.merge_trivial_subcandidates(candidate, self.source_info(or_span));
+        candidate.or_span = Some(or_span);
+        self.merge_trivial_subcandidates(candidate);
     }
 
     /// Try to merge all of the subcandidates of the given candidate into one.
     /// This avoids exponentially large CFGs in cases like `(1 | 2, 3 | 4, ...)`.
-    fn merge_trivial_subcandidates(
-        &mut self,
-        candidate: &mut Candidate<'_, 'tcx>,
-        source_info: SourceInfo,
-    ) {
+    fn merge_trivial_subcandidates(&mut self, candidate: &mut Candidate<'_, 'tcx>) {
         if candidate.subcandidates.is_empty() || candidate.has_guard {
             // FIXME(or_patterns; matthewjasper) Don't give up if we have a guard.
             return;
@@ -1550,7 +1556,7 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
 
         // Not `Iterator::all` because we don't want to short-circuit.
         for subcandidate in &mut candidate.subcandidates {
-            self.merge_trivial_subcandidates(subcandidate, source_info);
+            self.merge_trivial_subcandidates(subcandidate);
 
             // FIXME(or_patterns; matthewjasper) Try to be more aggressive here.
             can_merge &=
@@ -1559,6 +1565,8 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
 
         if can_merge {
             let any_matches = self.cfg.start_new_block();
+            let or_span = candidate.or_span.take().unwrap();
+            let source_info = self.source_info(or_span);
             for subcandidate in mem::take(&mut candidate.subcandidates) {
                 let or_block = subcandidate.pre_binding_block.unwrap();
                 self.cfg.goto(or_block, source_info, any_matches);

compiler/rustc_mir_build/src/build/matches/simplify.rs

@@ -82,7 +82,8 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
                     &*candidate.match_pairs
                 {
                     candidate.subcandidates = self.create_or_subcandidates(pats, has_guard);
-                    candidate.match_pairs.pop();
+                    let first_match_pair = candidate.match_pairs.pop().unwrap();
+                    candidate.or_span = Some(first_match_pair.pattern.span);
                 }
                 candidate
             })

tests/mir-opt/issues/issue_75439.foo.MatchBranchSimplification.diff

@@ -26,18 +26,20 @@
           _3 = _1;
           _2 = move _3 as [u32; 4] (Transmute);
           StorageDead(_3);
-          switchInt(_2[0 of 4]) -> [0: bb1, otherwise: bb6];
+          switchInt(_2[0 of 4]) -> [0: bb1, otherwise: bb4];
       }
 
       bb1: {
-          switchInt(_2[1 of 4]) -> [0: bb2, otherwise: bb6];
+          switchInt(_2[1 of 4]) -> [0: bb2, otherwise: bb4];
       }
 
       bb2: {
-          switchInt(_2[2 of 4]) -> [0: bb4, 4294901760: bb5, otherwise: bb6];
+          switchInt(_2[2 of 4]) -> [0: bb3, 4294901760: bb3, otherwise: bb4];
       }
 
       bb3: {
+          StorageLive(_4);
+          _4 = _2[3 of 4];
           StorageLive(_5);
           StorageLive(_6);
           _6 = _4;
@@ -46,27 +48,15 @@
           _0 = Option::<[u8; 4]>::Some(move _5);
           StorageDead(_5);
           StorageDead(_4);
-          goto -> bb7;
+          goto -> bb5;
       }
 
       bb4: {
-          StorageLive(_4);
-          _4 = _2[3 of 4];
-          goto -> bb3;
-      }
-  
-      bb5: {
-          StorageLive(_4);
-          _4 = _2[3 of 4];
-          goto -> bb3;
-      }
-  
-      bb6: {
           _0 = Option::<[u8; 4]>::None;
-          goto -> bb7;
+          goto -> bb5;
       }
 
-      bb7: {
+      bb5: {
           StorageDead(_2);
           return;
       }