Test failure: System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test

[VincentBu]

Failed in: runtime-coreclr libraries-jitstress-random 20260425.1

Failed tests:

net11.0-linux-Release-x64-jitstress_random_1-AzureLinux.3.Amd64.Open
- System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test
net11.0-linux-Release-x64-jitstress_random_2-AzureLinux.3.Amd64.Open
- System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test

Error message:

Assert.Equal() Failure: Values differ
Expected: <-2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648, -2147483648>
Actual:   <-107374176, -107374176, -107374176, -107374176, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>

Stack trace:

   at System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test() in /_/src/libraries/System.Runtime.Intrinsics/tests/Vectors/Vector512Tests.cs:line 6367
   at System.Reflection.MethodBaseInvoker.InterpretedInvoke_Method(Object obj, IntPtr* args) in /_/src/coreclr/System.Private.CoreLib/src/System/Reflection/MethodBaseInvoker.CoreCLR.cs:line 39
   at System.Reflection.MethodBaseInvoker.InvokeWithNoArgs(Object obj, BindingFlags invokeAttr) in /_/src/libraries/System.Private.CoreLib/src/System/Reflection/MethodBaseInvoker.cs:line 57

Kusto query result:

Queued	Pipeline	Pipeline Configuration	OS	Arch	Test	Outcome	Properties	Build Reason
2026-04-26 10:56:04.460000+00:00	runtime-extra-platforms	Release	osx.15.amd64.open	x64	System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test	Failed	{"operatingSystem":"OSX.15.Amd64.Open","System.PhaseAttempt":"2","BuildNumber":"20260425.1","System.JobAttempt":"2","System.JobId":"440ea7f6-c32f-5952-47d6-56dae7644f8a","configuration":"Release","BuildId":"1396883","Project":"public","DefinitionName":"runtime-extra-platforms","System.StageAttempt":"2","System.JobName":"__default","CollectionUri":"https://dev.azure.com/dnceng-public/","DefinitionId":"154","System.StageName":"Build","runtimeFlavor":"coreclr","AzurePipelinesTestRunId":"38961892","System.PhaseName":"build_iossimulator_x64_Release_AllSubsets_CoreCLR","Reason":"Schedule","architecture":"x64"}	Schedule
2026-04-26 04:30:03.811000+00:00	runtime-extra-platforms	Release	osx.15.amd64.open	x64	System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test	Failed	{"Project":"public","BuildNumber":"20260425.1","System.JobAttempt":"1","DefinitionId":"154","BuildId":"1396883","Reason":"Schedule","configuration":"Release","System.PhaseAttempt":"1","DefinitionName":"runtime-extra-platforms","CollectionUri":"https://dev.azure.com/dnceng-public/","System.PhaseName":"build_iossimulator_x64_Release_AllSubsets_CoreCLR","System.JobId":"d413ef91-6ac1-5648-b23b-ad89eaf81672","System.StageName":"Build","runtimeFlavor":"coreclr","architecture":"x64","System.JobName":"__default","AzurePipelinesTestRunId":"38959244","System.StageAttempt":"1","operatingSystem":"OSX.15.Amd64.Open"}	Schedule
2026-04-25 22:32:01.769000+00:00	runtime-nativeaot-outerloop	Release	azurelinux.3.amd64.open.rt	x64	System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test	Failed	{"System.JobId":"abdbb26b-d534-5e94-28a6-1260271ac680","AzurePipelinesTestRunId":"38955702","DefinitionId":"265","configuration":"Release","System.JobName":"__default","BuildId":"1396680","DefinitionName":"runtime-nativeaot-outerloop","System.StageAttempt":"1","Reason":"Schedule","runtimeFlavor":"coreclr","CollectionUri":"https://dev.azure.com/dnceng-public/","System.PhaseName":"build_linux_x64_Release_NativeAOT_Libs","System.StageName":"Build","architecture":"x64","Project":"public","operatingSystem":"(Debian.13.Amd64.Open)AzureLinux.3.Amd64.Open@mcr.microsoft.com/dotnet-buildtools/prereqs:debian-13-helix-amd64","BuildNumber":"20260425.2","System.JobAttempt":"1","System.PhaseAttempt":"1"}	Schedule
2026-04-25 10:34:35.249000+00:00	runtime-nativeaot-outerloop	Release	azurelinux.3.amd64.open.rt	x64	System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test	Failed	{"System.StageName":"Build","architecture":"x64","System.PhaseName":"build_linux_musl_x64_Release_NativeAOT_Libs","System.StageAttempt":"1","System.JobId":"20e6f673-0b0c-5c85-377c-5d941d4c8ee3","runtimeFlavor":"coreclr","BuildNumber":"20260425.1","operatingSystem":"(Alpine.323.Amd64.Open)AzureLinux.3.Amd64.Open@mcr.microsoft.com/dotnet-buildtools/prereqs:alpine-3.23-helix-amd64","Project":"public","System.PhaseAttempt":"1","BuildId":"1396429","Reason":"Schedule","System.JobAttempt":"1","DefinitionName":"runtime-nativeaot-outerloop","configuration":"Release","System.JobName":"__default","AzurePipelinesTestRunId":"38949484","DefinitionId":"265","CollectionUri":"https://dev.azure.com/dnceng-public/"}	Schedule
2026-04-25 05:16:22.563000+00:00	runtime-extra-platforms	Release	osx.15.amd64.open	x64	System.Runtime.Intrinsics.Tests.Vectors.Vector512Tests.ConvertToInt32Test	Failed	{"BuildId":"1396313","System.StageName":"Build","System.StageAttempt":"1","operatingSystem":"OSX.15.Amd64.Open","configuration":"Release","AzurePipelinesTestRunId":"38946736","architecture":"x64","BuildNumber":"20260424.6","System.JobId":"d413ef91-6ac1-5648-b23b-ad89eaf81672","Project":"public","CollectionUri":"https://dev.azure.com/dnceng-public/","System.JobName":"__default","System.PhaseName":"build_iossimulator_x64_Release_AllSubsets_CoreCLR","DefinitionId":"154","DefinitionName":"runtime-extra-platforms","Reason":"Schedule","System.JobAttempt":"1","System.PhaseAttempt":"1","runtimeFlavor":"coreclr"}	Schedule

[dotnet-policy-service]

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

[mangod9]

Some analysis is below, @EgorBo and @tannergooding to validate:

Symptoms

Vector512Tests.ConvertToInt32Test produces wrong results on all non-AVX512 x64 platforms:

Expected: <-2147483648 × 16>   (Int32.MinValue for all 16 elements)
Actual:   <-107374176, -107374176, -107374176, -107374176, 0, 0, 0, 0, ...>

• Only 4 of 16 elements populated (128-bit operation instead of 512-bit)
• The 4 computed values are wrong (0xF99999A0 instead of 0x80000000)
• Remaining elements are zero/uninitialized — vary between runs

Failing platforms (all Release, all non-AVX512):

• runtime-coreclr libraries-jitstress-random — Linux x64
• runtime-extra-platforms — macOS x64 iossimulator
• runtime-nativeaot-outerloop — Linux x64 (Alpine + AzureLinux)

First failure: April 25, 2026 05:16 UTC

---

Root Cause: Missing ISA Guard + Recent Trigger

The missing guard (since April 2023)

PR: #84932 — "Accelerate Vector512 ConvertToSingle and ConvertToInt32"

When NI_Vector512_ConvertToInt32 was added to impSpecialIntrinsic in hwintrinsicxarch.cpp, the Vector512 case was grouped with Vector128/Vector256 without an AVX-512 ISA check:

case NI_Vector128_ConvertToInt32:
case NI_Vector256_ConvertToInt32:
case NI_Vector512_ConvertToInt32:   // ← no ISA gate for Vector512!
{
    op1     = impSIMDPopStack();
    retNode = gtNewSimdCvtNode(retType, op1, TYP_INT, simdBaseType, simdSize);
    break;
}

Vector128 and Vector256 work fine without AVX-512 (using SSE/AVX instructions). But when the Vector512 case enters gtNewSimdCvtNode → gtNewSimdCvtNativeNode, it unconditionally selects NI_AVX512_ConvertToVector512Int32WithTruncation — an AVX-512 instruction — for simdSize=64.

Compare with ConvertToInt64 which correctly gates on AVX-512:

case NI_Vector512_ConvertToInt64:
{
    if (compOpportunisticallyDependsOn(InstructionSet_AVX512))  // ← proper gate
    {
        op1     = impSIMDPopStack();
        retNode = gtNewSimdCvtNode(retType, op1, TYP_LONG, simdBaseType, simdSize);
    }
    break;  // falls back to managed when AVX512 unavailable
}

Why it was harmless for 3 years: On non-AVX512 machines, the JIT never reached this code path. The managed fallback in Vector512.ConvertToInt32 (which splits into two Vector256 operations) always ran instead — the JIT had no reason to intrinsify a Vector512 call that would be expanded from the managed code.

The trigger — PR #127124 + PR #127402 (April 24, 2026)

PR: #127124 — "JIT: propagate vector constants through assertion prop"
PR: #127402 — "Allow O2K_CONST_VEC to handle all TYP_SIMD32/64"
Author: @EgorBo

These PRs enable the JIT to constant-propagate Vector512 values. The test does:

Vector512.ConvertToInt32(Vector512.Create(float.MinValue))

With the new constant propagation, Vector512.Create(float.MinValue) becomes a compile-time constant. The JIT now attempts to fold the ConvertToInt32 call through the intrinsic path (gtNewSimdCvtNode) instead of falling back to managed code — even on non-AVX512 machines. This hits the missing guard for the first time.

Timeline

Date	Event
Apr 2023	PR #84932 adds NI_Vector512_ConvertToInt32 without AVX-512 gate
Apr 2024	PR #97529 adds gtNewSimdCvtNode saturation logic (correctly gated by assert, but only in Debug)
Apr 2024	PR #100993 adds ConvertToInt32Test (test has existed for 2 years)
Apr 24, 2026	PR #127124 + #127402 enable Vector512 constant prop (bug exposed)
Apr 25, 2026	First test failures on iossimulator, NativeAOT

---

The Bug in Detail

In hwintrinsicxarch.cpp (line 1849-1858), when the JIT recognizes NI_Vector512_ConvertToInt32, it calls gtNewSimdCvtNode without checking for AVX-512 support:

case NI_Vector128_ConvertToInt32:
case NI_Vector256_ConvertToInt32:
case NI_Vector512_ConvertToInt32:   // ← no ISA gate for Vector512!
{
    op1     = impSIMDPopStack();
    retNode = gtNewSimdCvtNode(retType, op1, TYP_INT, simdBaseType, simdSize);
    break;
}

Compare with ConvertToInt64 which correctly gates:

case NI_Vector512_ConvertToInt64:
{
    if (compOpportunisticallyDependsOn(InstructionSet_AVX512))  // ← proper gate
    {
        op1     = impSIMDPopStack();
        retNode = gtNewSimdCvtNode(retType, op1, TYP_LONG, simdBaseType, simdSize);
    }
    break;  // falls back to managed when AVX512 unavailable
}

For ConvertToInt32, Vector128 and Vector256 work fine without AVX-512 (using SSE/AVX instructions). But Vector512 requires AVX-512 for the native truncation instruction, and there's no fallback in gtNewSimdCvtNativeNode for simdSize=64 without it.

---

Fix

File: src/coreclr/jit/hwintrinsicxarch.cpp — Gate Vector512 ConvertToInt32 on AVX-512, matching the ConvertToInt64 pattern:

case NI_Vector128_ConvertToInt32:
case NI_Vector256_ConvertToInt32:
case NI_Vector512_ConvertToInt32:
{
    assert(sig->numArgs == 1);
    assert(simdBaseType == TYP_FLOAT);

    if ((simdSize == 64) && !compOpportunisticallyDependsOn(InstructionSet_AVX512))
    {
        // Vector512 float->int conversion requires AVX-512 for the native
        // truncation instruction. Fall back to managed implementation.
        break;
    }

    op1     = impSIMDPopStack();
    retNode = gtNewSimdCvtNode(retType, op1, TYP_INT, simdBaseType, simdSize);
    break;
}

Same pattern applied to ConvertToInt32Native.

When AVX-512 is unavailable and simdSize == 64, retNode stays null and the JIT falls back to the managed Vector512.ConvertToInt32 implementation, which correctly splits into two Vector256.ConvertToInt32 calls using SSE/AVX instructions.

Optionally, the assertions in gtNewSimdCvtNode and gtNewSimdCvtNativeNode in gentree.cpp can also be tightened to add (simdSize != 64) to the non-AVX512 condition — this would catch the issue in Debug/Checked builds if any other caller bypasses the gate in the future.

[EgorBo]

@mangod9 in case if you're interested in feedback for the analysis, it seems like it's not correct, my local copilot identified a use-before-def bug elsewhere https://github.com/dotnet/runtime/pull/127524/changes

[mangod9]

ok interesting, so the determination that it was a pre-existing issue exposed by the latest PR was incorrect, or the suggested fix was not right? Maybe it over pivoted on the fact that it was failing on all non-avx512 legs?

[EgorBo]

ok interesting, so the determination that it was a pre-existing issue exposed by the latest PR was incorrect, or the suggested fix was not right? Maybe it over pivoted on the fact that it was failing on all non-avx512 legs?

The bug was pre-existing, but the suggested fix doesn't fix the problem and effectively incorrect. But I am not sure the analysis could catch that easily, my local one had dive into JitDump