gh-158847: Avoid enumerating cycles when choosing pegen recursion leaders

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: jonbaldie

Lib/test/test_peg_generator/test_pegen.py

@@ -12,7 +12,9 @@
 with test_tools.imports_under_tool("peg_generator"):
     from pegen.grammar_parser import GeneratedParser as GrammarParser
     from pegen.testutil import parse_string, generate_parser, make_parser
-    from pegen.grammar import GrammarVisitor, GrammarError, Grammar, RuleKind
+    from pegen.grammar import (
+        GrammarVisitor, GrammarError, Grammar, NameLeaf, RuleKind,
+    )
     from pegen.grammar_visualizer import ASTGrammarPrinter
     from pegen.parser import Parser
     from pegen.parser_generator import compute_nullables, compute_left_recursives
@@ -751,6 +753,56 @@ def test_opt_sequence(self) -> None:
         # of a line in the generated source. See bpo-41044
         make_parser(grammar)
 
+    def test_left_recursion_leader_order(self) -> None:
+        grammar = parse_string("""
+        start: zeta NEWLINE
+        zeta: alpha '+' | NUMBER
+        alpha: zeta '-' | NUMBER
+        """, GrammarParser)
+        PythonParserGenerator(grammar, io.StringIO())
+        self.assertTrue(grammar.rules["alpha"].leader)
+        self.assertFalse(grammar.rules["zeta"].leader)
+
+    def test_large_left_recursive_grammar(self) -> None:
+        size = 12
+        lines = ["start: r0 NEWLINE ENDMARKER"]
+        for i in range(size):
+            children = list(range(i + 1, size))
+            if i:
+                children.append(0)
+            alternatives = [f"r{j} '+'" for j in children] + ["NUMBER"]
+            lines.append(f"r{i}: " + " | ".join(alternatives))
+        parser_class = make_parser("\n".join(lines) + "\n")
+        node = parse_string("1\n", parser_class)
+        self.assertEqual(node[0].string, "1")
+
+    def test_left_recursion_analysis_work(self) -> None:
+        class CountedName(str):
+            comparisons = 0
+            __hash__ = str.__hash__
+
+            def __eq__(self, other):
+                type(self).comparisons += 1
+                return super().__eq__(other)
+
+        size = 8
+        lines = ["start: r0 NEWLINE ENDMARKER"]
+        for i in range(size):
+            alternatives = [
+                f"r{j} '+'" for j in range(size) if j != i
+            ] + ["NUMBER"]
+            lines.append(f"r{i}: " + " | ".join(alternatives))
+        grammar = parse_string("\n".join(lines) + "\n", GrammarParser)
+        for rule in grammar.rules.values():
+            for alt in rule.rhs.alts:
+                for item in alt.items:
+                    if isinstance(item.item, NameLeaf):
+                        item.item.value = CountedName(item.item.value)
+        CountedName.comparisons = 0
+        with self.assertRaisesRegex(ValueError, "no leadership candidate"):
+            PythonParserGenerator(grammar, io.StringIO())
+        self.assertLess(CountedName.comparisons, 2_000)
+
     def test_left_recursion_too_complex(self) -> None:
         grammar = """
         start: foo

Misc/NEWS.d/next/Tools-Demos/2026-10-05-14-04-06.gh-issue-158847.Xbj0dg.rst

@@ -0,0 +1,2 @@
+Pegen now picks the leader of a left-recursive rule cycle in polynomial time,
+instead of enumerating every path through the cycle.

Tools/peg_generator/pegen/grammar_analysis.py

@@ -115,19 +115,15 @@ def compute_left_recursives(
         if len(scc) > 1:
             for name in scc:
                 rules[name].left_recursive = True
-            # Try to find a leader such that all cycles go through it.
-            leaders = set(scc)
-            for start in scc:
-                for cycle in sccutils.find_cycles_in_scc(graph, scc, start):
-                    # print("Cycle:", " -> ".join(cycle))
-                    leaders -= scc - set(cycle)
-                    if not leaders:
-                        raise ValueError(
-                            f"SCC {scc} has no leadership candidate (no element is included in all cycles)"
-                        )
-            # print("Leaders:", leaders)
-            leader = min(leaders)  # Pick an arbitrary leader from the candidates.
-            rules[leader].leader = True
+            # A leader lies in every cycle, so removing it must leave a DAG.
+            for leader in sorted(scc):
+                if sccutils.is_acyclic(graph, scc - {leader}):
+                    rules[leader].leader = True
+                    break
+            else:
+                raise ValueError(
+                    f"SCC {scc} has no leadership candidate (no element is included in all cycles)"
+                )
         else:
             name = min(scc)  # The only element.
             if name in graph[name]:

Tools/peg_generator/pegen/sccutils.py

@@ -1,6 +1,6 @@
 # Adapted from mypy (mypy/build.py) under the MIT license.
 
-from collections.abc import Iterable, Iterator, Set
+from collections.abc import Iterator, Set
 
 
 def strongly_connected_components(
@@ -49,32 +49,22 @@ def dfs(v: str) -> Iterator[set[str]]:
             yield from dfs(v)
 
 
-def find_cycles_in_scc(
-    graph: dict[str, Set[str]], scc: Set[str], start: str
-) -> Iterable[list[str]]:
-    """Find cycles in SCC emanating from start.
-
-    Yields lists of the form ['A', 'B', 'C', 'A'], which means there's
-    a path from A -> B -> C -> A.  The first item is always the start
-    argument, but the last item may be another element, e.g.  ['A',
-    'B', 'C', 'B'] means there's a path from A to B and there's a
-    cycle from B to C and back.
-    """
-    # Basic input checks.
-    assert start in scc, (start, scc)
-    assert scc <= graph.keys(), scc - graph.keys()
-
-    # Reduce the graph to nodes in the SCC.
-    graph = {src: {dst for dst in dsts if dst in scc} for src, dsts in graph.items() if src in scc}
-    assert start in graph
-
-    # Recursive helper that yields cycles.
-    def dfs(node: str, path: list[str]) -> Iterator[list[str]]:
-        if node in path:
-            yield path + [node]
-            return
-        path = path + [node]  # TODO: Make this not quadratic.
-        for child in graph[node]:
-            yield from dfs(child, path)
-
-    yield from dfs(start, [])
+def is_acyclic(graph: dict[str, Set[str]], vertices: Set[str]) -> bool:
+    """Check the subgraph induced by vertices using a topological sort."""
+    indegree = dict.fromkeys(vertices, 0)
+    for src in vertices:
+        for dst in graph[src]:
+            if dst in vertices:
+                indegree[dst] += 1
+
+    ready = [node for node, degree in indegree.items() if degree == 0]
+    processed = 0
+    while ready:
+        src = ready.pop()
+        processed += 1
+        for dst in graph[src]:
+            if dst in vertices:
+                indegree[dst] -= 1
+                if indegree[dst] == 0:
+                    ready.append(dst)
+    return processed == len(vertices)