[arm32/linux] Assert failure: dwSrcIndex > dwLastSrcIndex in ABI stress

[jakobbotsch]

pinvoke stress on arm32 fails with an assert failure:

Assert failure(PID 46 [0x0000002e], Thread: 46 [0x002e]): dwSrcIndex > dwLastSrcIndex
    File: /__w/2/s/src/vm/arm/stubs.cpp Line: 1558
    Image: /root/helix/work/correlation/corerun

/root/helix/work/workitem/JIT/Stress/ABI/pinvokes_do/pinvokes_do.sh: line 271:    46 Aborted                 $LAUNCHER $ExePath "${CLRTestExecutionArguments[@]}"

Return code:      1
Raw output file:      /root/helix/work/workitem/JIT/Stress/Reports/JIT.Stress/ABI/pinvokes_do/pinvokes_do.output.txt
Raw output:
BEGIN EXECUTION
/root/helix/work/correlation/corerun pinvokes_do.exe '--pinvokes' '--num-calls' '1000' '--no-ctrlc-summary'
Stressing pinvokes
OSVersion: Unix 4.4.38.0
OSArchitecture: Arm
ProcessArchitecture: Arm
Selecting armhf ABI.
50 callers done
100 callers done
Expected: 100
Actual: 134
END EXECUTION - FAILED
Test Harness Exitcode is : 1
To run the test:
> set CORE_ROOT=/root/helix/work/correlation
> /root/helix/work/workitem/JIT/Stress/ABI/pinvokes_do/pinvokes_do.sh

Small self-contained example:

using System;
using System.Runtime.InteropServices;

public class Program
{
    public static void Main()
    {
        CalleeDelegate callee = Callee;
        callee(default, default, default, default, default, default);
        GC.KeepAlive(callee);
    }

    private static int Callee(Hfa1 a1, S1P a2, Hfa2 a3, Hfa2 a4, S9U a5, S11U a6)
    {
        return 0;
    }

    [UnmanagedFunctionPointer(CallingConvention.Cdecl)]
    private delegate int CalleeDelegate(Hfa1 a1, S1P a2, Hfa2 a3, Hfa2 a4, S9U a5, S11U a6);

    public struct S1P { public byte F0; }
    public struct S9U { public byte F0, F1, F2, F3, F4, F5, F6, F7, F8; }
    public struct S11U { public byte F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10; }
    public struct Hfa1 { public float F0, F1; }
    public struct Hfa2 { public double F0, F1, F2, F3; }
}

category:correctness
theme:pinvoke
skill-level:intermediate
cost:medium

[BruceForstall]

I presume this depends on dotnet/coreclr#26090

[jakobbotsch]

Yes indeed, I opened this to have something to reference when excluding that test for arm32.

[jakobbotsch]

The test that found this is currently disabled on ARM32:
https://github.com/dotnet/coreclr/blob/fdb2bc4233bf93e44bda4b1ae3b1432e80c9f0da/tests/issues.targets#L232-L234

As part of a fix to this issue this test should be enabled.

This test will run through a lot of different callers. To make a smaller test case, you can ask the test to run only a single caller using the arg --caller-index 1234. However since this is an assertion failure the test output does not show exactly which caller it is (only that it is in the 100-150 area). So to find the proper index the following can be done:

1. Run with verbose, i.e. --pinvokes --num-calls 1000 --verbose
2. It will give a lot of output. Once it asserts, scroll up to see the name of the caller that was being tested. It should be something like ABIStress_Pinvoker1234
3. Rerun the test with only this caller, i.e. something like --pinvokes --caller-index 1234. This should reproduce the problem without going through a bunch of other callers first.

[BruceForstall]

The assert is in StubLinkerCPU::EmitShuffleThunk()

[jakobbotsch]

Maybe this could be similar to #13225?

[jakobbotsch]

Should this be assigned to someone else @BruceForstall? Since this looks like a VM bug.