[NativeAOT] Fix floating pointer register unwinding on Arm32

[jkotas]

NativeAOT specific register window implementation has to match the one in libunwind.

Fixes #118942.

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[dotnet-policy-service]

Tagging subscribers to this area: @agocke, @MichalStrehovsky, @jkotas
See info in area-owners.md if you want to be subscribed.

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[am11]

Fixes #118942.

[MichalStrehovsky]

Do you know what was the exact condition that was triggering this? If it's something sane, I wonder if it would be worth a regression test.

If we weren't uploading XML files with test results into AzDo and AzDo wasn't rejecting the durations, we'd basically never hit this in any of our testing.

[jkotas]

Do you know what was the exact condition that was triggering this?

The distilled repro would be a chain of multiple methods that interleave floating-point computation and exception handling in a specific shape, something like this:

void M1()
{
    double mustBeUsedMultipleTimesToGetEnregisteredIntoCalledSavedFPRegister;
    mustBeUsedMultipleTimesToGetEnregisteredIntoCalledSavedFPRegister = some computation;
    M2();
    do more stuff with mustBeUsedMultipleTimesToGetEnregisteredIntoCalledSavedFPRegister;
}

[NoInline]
void M2()
{
    // This method should not use any floating point
    try
    {
        M3();
    }
    catch
    {
    }
}

[NoInline]
void M3()
{
    // The JIT must enregister this into same register as the local variable in M1()
    double mustBeUsedMultipleTimesToGetEnregisteredIntoCalledSavedFPRegister;
    mustBeUsedMultipleTimesToGetEnregisteredIntoCalledSavedFPRegister = some other computation;
    M4();`
    do more stuff with mustBeUsedMultipleTimesToGetEnregisteredIntoCalledSavedFPRegister;
}

[NoInline]
void M4()
{
    throw new Exception();
}

I can try to add a test like this.

[janvorli]

LGTM, thank you!

[jkotas]

I can try to add a test like this.

It turns out we have multiple tests like that, but they are non-effective for various reasons. I am just going to fix one of the existing tests do what it meant to test.

[jkotas]

Pushed test updated. #119108 validated that the updated test fails on both linux-arm and linux-arm64 without the fixes.

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[jkotas]

/ba-g infrastructure timeout

[jkotas]

/backport to release/10.0

[github-actions]

Started backporting to release/10.0: https://github.com/dotnet/runtime/actions/runs/17276805531