JIT: Invalid result with SVE intrinsics

[jakobbotsch]

// Generated by Fuzzlyn v3.3 on 2026-02-15 17:21:42
// Run on Arm64 Linux
// Seed: 103488890012433678-vectort,vector64,vector128,armadvsimd,armadvsimdarm64,armaes,armarmbase,armarmbasearm64,armcrc32,armcrc32arm64,armdp,armrdm,armrdmarm64,armsha1,armsha256,armsve,armsve2
// Reduced from 29.2 KiB to 0.7 KiB in 00:00:16
// Debug: Outputs -3
// Release: Outputs -2
using System.Numerics;
using System.Runtime.Intrinsics;
using System.Runtime.Intrinsics.Arm;

public class Program
{
    public static void Main()
    {
        var vr8 = Vector.Create<uint>(1);
        var vr9 = Vector128.CreateScalar(4701966028735610216UL).AsVector();
        var vr10 = Sve.ShiftLeftLogical(vr8, vr9);
        var vr11 = Vector128.CreateScalar(2U).AsVector();
        var vr12 = Sve.Subtract(vr10, vr11);
        var vr13 = Vector.Create<uint>(1);
        var vr14 = Vector128.CreateScalar(0U).AsVector();
        var vr15 = Sve.CreateBreakBeforePropagateMask(vr12, vr13, vr14);
        sbyte vr16 = (sbyte)Sve.SaturatingDecrementByActiveElementCount(0, vr15);
        System.Console.WriteLine(vr16);
    }
}

cc @dotnet/jit-contrib @dotnet/arm64-contrib

[dotnet-policy-service]

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

[jakobbotsch]

FWIW @a74nh there are plenty of SVE-looking failures in the Fuzzlyn runs, see e.g. https://dev.azure.com/dnceng-public/public/_build/results?buildId=1295334&view=ms.vss-build-web.run-extensions-tab from this Sunday.
Feel free to investigate these proactively (Fuzzlyn pipeline runs every Saturday/Sunday). I only tend to open a single issue every week because I am unsure how many of these have the same root cause. Also, they may or may not be SVE related at all.

[a74nh]

Will take a look

[a74nh]

This is due to the code that converts between constant masks and patterns. It's just a little too optimistic in pattern matching (it is ignoring bits it shouldn't).

This whole area has been quite buggy. And we'll be ripping it out again to support agnostic SVE lengths. I'll look at getting this fixed, but I wonder if the better solution is to just to start on reworking the constant mask support.