Description

Status: Investigation only — no fix implemented yet.

Regex pattern a(b.*?c)?d fails to match abccd in .NET 10, a regression from .NET 9. The bug occurs when a lazy quantifier (*?) with min=0 using Setlazy/Notonelazy (character class or negated-char loops) is inside an optional group (...)?.

var regex = new Regex(@"a(b.*?c)?d");
var match = regex.Match("abccd");
// .NET 9: match.Success == true
// .NET 10: match.Success == false

Investigation findings

• Confirmed reproduction: Setlazy/Notonelazy with M=0 inside optional groups fail; Onelazy with M=0 works; M≥1 works
• Bytecode analysis: The failing pattern's compiled bytecodes contain Setjump/Forejump (atomic group markers) wrapping the outer ? loop, preventing backtracking into the inner .*?
• Root cause area: FindAndMakeLoopsAtomic in RegexNode.cs, introduced by commit c0f4c7dd39 ("Support auto-atomicity on {lazy} loops", Support auto-atomicity on {lazy} loops #117943). CanBeMadeAtomic for general Loop nodes checks only the loop body's first/last character classes against the subsequent node, but doesn't account for characters matchable by internal lazy quantifiers. The MayContainBacktracking guard (Fix incorrect atomic loop optimization when body contains backtracking #124254) should block this but the atomic wrapper still appears in bytecodes through a path not yet identified.

What remains

• Pinpoint the exact code path inserting the atomic wrapper despite MayContainBacktracking returning true
• Implement the fix
• Add test cases to Regex.Match.Tests.cs
• Verify existing tests pass

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.