Vector.IsHardwareAccelerated result differs when called directly and called via reflection

[AndyAyersMS]

using System;
using System.Reflection;
using System.Numerics;

class X
{
    public static int Main()
    {
        bool hw1 = Vector.IsHardwareAccelerated;

        Assembly spc = typeof(object).Assembly;
        Type vecType = spc.GetType("System.Numerics.Vector");
        MethodInfo method = vecType.GetMethod("get_IsHardwareAccelerated");
        bool hw2 = (bool) method.Invoke(null, null);

        Console.WriteLine($"Direct call to Vector.IsHardwareAccelerated returns {hw1}");
        Console.WriteLine($"Reflection call to Vector.IsHardwareAccelerated returns {hw2}");

        return (hw1 == hw2 ? 100 : -1);
    }
}

results in

Direct call to Vector.IsHardwareAccelerated returns True
Reflection call to Vector.IsHardwareAccelerated returns False

Seems like we're not recognizing this as an intrinsic when it is jitted on its own.

category:cq
theme:vector-codegen
skill-level:intermediate
cost:medium

[AndyAyersMS]

cc @dotnet/jit-contrib @tannergooding

[AndyAyersMS]

Saw this when looking at PMI results, where we force this method to be jitted.

[tannergooding]

CC. @CarolEidt

This has always been like this as the methods aren't recursive (and we didn't actually add the mustExpand recursion support until HWIntrinsics in 3.0).

It shouldn't be particularly hard to fix, however.

[tannergooding]

I have the fix for this done locally however since IsHardwareAccelerated lives in Vector (rather than Vector<T>), it is included in R2R.

I though this just needed to be fTreatAsRegularMethodCall = TRUE in zapinfo, but then it isn't recognized as intrinsic at runtime and so it fails if R2R is enabled (the call ends up being recursive and not recognized as an intrinsic).

[tannergooding]

Moving this to future for the time being as the existing behavior goes back to .NET Framework and so I don't think it is particularly critical that we get it fixed for .NET 5.

I think I would need input from @davidwrighton or @jkotas as to what is missing for R2R to start working correctly with IsHardwareAccelerated being recursive.
As best as I can tell, there isn't any difference between it and the IsSupported checks in terms of attribution or handling, but I'm likely missing a check somewhere.

[tannergooding]

The changes I had prepped are here, for reference: tannergooding@b6b555f

[jkotas]

One problem that I can see in your delta is that the mustExpand expansion will kick in only for FEATURE_HW_INTRINSICS and so it will not work on ARM. I think it needs to be unconditional outside FEATURE_HW_INTRINSICS ifdef like the NI_IsSupported_True / NI_IsSupported_False for regular hardware intrinsics.

Also, you may need a matching change in Mono.

The zapper side of your delta looks fine.

[tannergooding]

Moving to 8.0.0. This is in PR but not high priority and is dependent on someone from @dotnet/crossgen-contrib following up on the potential issue with the ISA flags