Assertion failed 'block->countOfInEdges() > 0' during 'Remove empty finally' with infinite loop in finally

[jakobbotsch]

public static void Main()
{
    try
    {
        System.Console.WriteLine("A");
    }
    finally
    {
        while (true)
        {
        }
    }
}

Assert failure(PID 14880 [0x00003a20], Thread: 23380 [0x5b54]): Assertion failed 'block->countOfInEdges() > 0' in 'Program:Main()' during 'Remove empty finally' (IL size 16)

    File: D:\dev\dotnet\runtime\src\coreclr\jit\fgflow.cpp Line: 280
    Image: D:\dev\dotnet\runtime\artifacts\tests\coreclr\windows.x64.Debug\Tests\Core_Root\corerun.exe

[JulieLeeMSFT]

@kunalspathak PTAL.

[kunalspathak]

I have seen cases where roslyn would add unreachable infinite loop inside finally - dotnet/roslyn#51205

[EgorBo]

I also saw in the past that until we calculate preds we don't maintain bbRefs 100% accurate in the initial phases

[kunalspathak]

So it is not directly related to dotnet/roslyn#51205 but has similar effects.

Roslyn duplicates the infinite loop from inside the finally block to the outside and add a jump to it leaving the infinite loop inside finally as it-is and unreachable. Of course, jit need to be resilient to such ILs, but may be there is room to improve in roslyn as well - @jaredpar ?

IL to import:
IL_0000  72 33 11 00 70    ldstr        0x70001133
IL_0005  28 3c 00 00 0a    call         0xA00003C
IL_000a  de 02             leave.s      2 (IL_000e)
IL_000c  2b fe             br.s         -2 (IL_000c)
IL_000e  2b fe             br.s         -2 (IL_000e)

[jaredpar]

Roslyn duplicates the infinite loop from inside the finally block to the outside and add a jump to it leaving the infinite loop inside finally as it-is and unreachable.

Making sure we're on the same page: the loop outside the finally is the unreachable one correct?

may be there is room to improve in roslyn as well

Possibly. I'm a bit surprised we are duplicating here, not sure why we'd be doing that when a loop is inside a finally. @cston any intuition here?

[kunalspathak]

the loop outside the finally is the unreachable one correct?

No, the loop inside the finally is unreachable. We directly jump to the loop outside on IL_000e

[jaredpar]

Yes but won't an exception jump to the finally block?