Regex evaluation bug - discrepancy between compiled and non-compiled regex

[lahma]

Description

Running same Regex as compiled and non-compiled version is producing different outcome.

Reproduction Steps

using System.Text.RegularExpressions;

namespace BugReport;

public class RegexTests
{
    [Fact]
    public void CompiledRegexShouldProduceSameResultAsNonCompiled()
    {
        const string Pattern = @"(.*?)a(?!(a+)b\2c)\2(.*)";
        var nonCompiled = new Regex(Pattern, RegexOptions.None);
        var compiled = new Regex(Pattern, RegexOptions.Compiled);

        const string Input = "baaabaac";
        const int GroupNumber = 2;

        Assert.Equal(nonCompiled.Match(Input).Groups[GroupNumber].Value, compiled.Match(Input).Groups[GroupNumber].Value);
    }
}

Expected behavior

Group capture is equal, in this case value should be empty.

Actual behavior

Xunit.Sdk.EqualException
Assert.Equal() Failure: Strings differ
Expected: ""
Actual:   "a"
           ↑ (pos 0)
   at CompiledRegexShouldProduceSameResultAsNonCompiled() in Test.cs:line 18
   at System.RuntimeMethodHandle.InvokeMethod(Object target, Void** arguments, Signature sig, Boolean isConstructor)
   at System.Reflection.MethodBaseInvoker.InvokeWithNoArgs(Object obj, BindingFlags invokeAttr)

Regression?

No response

Known Workarounds

No response

Configuration

Runnin latest .NET, Windows 11.

❯ dotnet --version
8.0.101

Other information

No response

[ghost]

Tagging subscribers to this area: @dotnet/area-system-text-regularexpressions
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Description

Running same code as compiled and non-compiled version is producing different outcome.

Reproduction Steps

using System.Text.RegularExpressions;

namespace BugReport;

public class RegexTests
{
    [Fact]
    public void CompiledRegexShouldProduceSameResultAsNonCompiled()
    {
        const string Pattern = @"(.*?)a(?!(a+)b\2c)\2(.*)";
        var nonCompiled = new Regex(Pattern, RegexOptions.None);
        var compiled = new Regex(Pattern, RegexOptions.Compiled);

        const string Input = "baaabaac";
        const int GroupNumber = 2;

        Assert.Equal(nonCompiled.Match(Input).Groups[GroupNumber].Value, compiled.Match(Input).Groups[GroupNumber].Value);
    }
}

Expected behavior

Group capture is equal.

Actual behavior

Xunit.Sdk.EqualException
Assert.Equal() Failure: Strings differ
Expected: ""
Actual:   "a"
           ↑ (pos 0)
   at CompiledRegexShouldProduceSameResultAsNonCompiled() in Test.cs:line 18
   at System.RuntimeMethodHandle.InvokeMethod(Object target, Void** arguments, Signature sig, Boolean isConstructor)
   at System.Reflection.MethodBaseInvoker.InvokeWithNoArgs(Object obj, BindingFlags invokeAttr)

Regression?

No response

Known Workarounds

No response

Configuration

Runnin latest .NET, Windows 11.

❯ dotnet --version
8.0.101

Other information

No response

Author:	lahma
Assignees:	-
Labels:	area-System.Text.RegularExpressions
Milestone:	-

[stephentoub]

Thanks for the helpful repro.

Here's a slightly simpler one based on yours:

using System.Text.RegularExpressions;

const string Pattern = @"(?!(b)b)\1";
const string Input = "ba";

var nonCompiled = new Regex(Pattern, RegexOptions.None);
var compiled = new Regex(Pattern, RegexOptions.Compiled);

Console.WriteLine(nonCompiled.Match(Input).Success);
Console.WriteLine(compiled.Match(Input).Success);

The issue appears to be in how compiled (and source generated) regexes are handling capture groups inside of negative lookarounds. They're not uncapturing when exiting the construct, so whereas the backreference doesn't end up matching in the interpreter (because there's no capture to match), it does end up matching in the compiled regex because the capture is still there and matches.

[stephentoub]

@lahma, given that a backreference to a capture inside a negative lookahead from outside that lookahead will never match, can you speak to the pattern you were using that encountered this? Was it just a test, or was the pattern actually trying to do something useful with that backreference?

[lahma]

Well hard to say if we can call this a real world scenario, as we are talking about my arch enemy after all. I encountered this when running ECMAScript test suite and testing generated Regex instances in compiled mode when trying to optimize Jint. Here's the actual test case.

[stephentoub]

Thanks. That's what I figured it was.

[lahma]

Adding a note that when run under net462 and net6.0 the compiled and non-compiled versions work the same way, both net7.0 and net8.0 have regressed.

[stephentoub]

Adding a note that when run under net462 and net6.0 the compiled and non-compiled versions work the same way, both net7.0 and net8.0 have regressed.

Yes, we rewrote the compiler in 7.