Setting COMPlus_EnableHWIntrinsic=0 causes System.PlatformNotSupportedException with optimized libraries code

[echesakov]

While working on #38162 I observed behavior (that I believe is incorrect) when running with .NET 6 SDK RC1.

To reproduce the issue build and run the project using the following sequence of commands:

dotnet build -c Release

set COMPlus_EnableHWIntrinsic=0

dotnet run -c Release

Runtime_60035.csproj

<Project Sdk="Microsoft.NET.Sdk">

  <PropertyGroup>
    <OutputType>Exe</OutputType>
    <TargetFramework>net6.0</TargetFramework>
  </PropertyGroup>

</Project>

Runtime_60035.cs

using System;
using System.Text;
using System.Text.Encodings.Web;

namespace Runtime_60035
{
    class Program
    {
        static void Main(string[] args)
        {
            byte[] inputBytes = Encoding.UTF8.GetBytes("https://github.com/dotnet/runtime");
            Console.WriteLine(UrlEncoder.Default.FindFirstCharacterToEncodeUtf8(inputBytes));
        }
    }
}

The result will be a failure with the following stack trace

Unhandled exception. System.PlatformNotSupportedException: Operation is not supported on this platform.
   at System.Runtime.Intrinsics.X86.Sse2.LoadVector128(Byte* address)
   at System.Text.Encodings.Web.OptimizedInboxTextEncoder.GetIndexOfFirstByteToEncodeSsse3(Byte* pData, UIntPtr lengthInBytes)
   at System.Text.Encodings.Web.OptimizedInboxTextEncoder.GetIndexOfFirstByteToEncode(ReadOnlySpan`1 data)
   at System.Text.Encodings.Web.DefaultUrlEncoder.FindFirstCharacterToEncodeUtf8(ReadOnlySpan`1 utf8Text)
   at Runtime_60035.Program.Main(String[] args) in D:\echesako\GitHub\Runtime_60035\Program.cs:line 12

If in addition to setting COMPlus_EnableHWIntrinsic=0 I set COMPlus_ReadyToRun=0 the issue goes away.

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[echesakov]

@tannergooding I wonder why we take OptimizedInboxTextEncoder.GetIndexOfFirstByteToEncodeSsse3 here with COMPlus_EnableHWIntrinsic=0.

@davidwrighton Do we have special handling for Ssse3.IsSupported and friends in crossgen2 (as we used to have in crossgen) to treat them as regular calls and not as intrinsics? Or this is done via a different mechanism in crossgen2?

[tannergooding]

I wonder why we take OptimizedInboxTextEncoder.GetIndexOfFirstByteToEncodeSsse3 here with COMPlus_EnableHWIntrinsic=0.

@echesakovMSFT There was a regression at one point where EnableHWIntrinsic=0 was no longer impacting all ISAs. It's possible that was reintroduced at some point and missed.

I don't think we have any tests that explicitly check that setting the environment variable correctly controls the IsSupported flags, but perhaps those should get added.

[danmoseley]

I know we discussed this in the past but I don’t know where we ended up. Do we test anything but AVX paths in the libraries (on x64)? And where we have added intrinsic paths for Arm64, do we test the software fallback paths? Just wondering how large the test gap might be.

[tannergooding]

Do we test anything but AVX paths in the libraries (on x64)? And where we have added intrinsic paths for Arm64, do we test the software fallback paths? Just wondering how large the test gap might be.

I know we have tests in the runtime side to test every ISA configuration (just not that the environment variables themselves are doing the right thing).

I don't know where we landed with the libraries jobs alsoo running with those switches....

[danmoseley]

I don't know where we landed with the libraries jobs alsoo running with those switches....

I don't see relevant DOTNET_ or COMPLUS_ in our yml, so I guess we have a test hole: we are shipping code we never test? Should we drive plugging that somehow before 7.0 feature coding starts, do you think? cc @safern

[tannergooding]

so I guess we have a test hole: we are shipping code we never test

It was tested in all the various configurations a couple times, I just don't know/believe we have the infra to do it automatically right now.

[danmoseley]

You’re quite right: I should have said that we do not test in automation. We have other such areas eg DirectoryServices tests need Docker containers. But this one seems feasible as it’s susceptible to environment variables. Could we safely pairwise this into our matrix perhaps eg Server 2022 always has AVX paths disabled?

[tannergooding]

I think we should be able to yes, especially as an outerloop. We just need to define the jobs and ensure the right environment variables are set.

I can sync with Santi or someone else tomorrow and figure this out.

Tagging subscribers to this area: @dotnet/area-system-runtime-intrinsics
See info in area-owners.md if you want to be subscribed.

Issue Details

---

While working on #38162 I observed behavior (that I believe is incorrect) when running with .NET 6 SDK RC1.

To reproduce the issue build and run the project using the following sequence of commands:

dotnet build -c Release

set COMPlus_EnableHWIntrinsic=0

dotnet run -c Release

Runtime_60035.csproj

<Project Sdk="Microsoft.NET.Sdk">

  <PropertyGroup>
    <OutputType>Exe</OutputType>
    <TargetFramework>net6.0</TargetFramework>
  </PropertyGroup>

</Project>

Runtime_60035.cs

using System;
using System.Text;
using System.Text.Encodings.Web;

namespace Runtime_60035
{
    class Program
    {
        static void Main(string[] args)
        {
            byte[] inputBytes = Encoding.UTF8.GetBytes("https://github.com/dotnet/runtime");
            Console.WriteLine(UrlEncoder.Default.FindFirstCharacterToEncodeUtf8(inputBytes));
        }
    }
}

The result will be a failure with the following stack trace

Unhandled exception. System.PlatformNotSupportedException: Operation is not supported on this platform.
   at System.Runtime.Intrinsics.X86.Sse2.LoadVector128(Byte* address)
   at System.Text.Encodings.Web.OptimizedInboxTextEncoder.GetIndexOfFirstByteToEncodeSsse3(Byte* pData, UIntPtr lengthInBytes)
   at System.Text.Encodings.Web.OptimizedInboxTextEncoder.GetIndexOfFirstByteToEncode(ReadOnlySpan`1 data)
   at System.Text.Encodings.Web.DefaultUrlEncoder.FindFirstCharacterToEncodeUtf8(ReadOnlySpan`1 utf8Text)
   at Runtime_60035.Program.Main(String[] args) in D:\echesako\GitHub\Runtime_60035\Program.cs:line 12

If in addition to setting COMPlus_EnableHWIntrinsic=0 I set COMPlus_ReadyToRun=0 the issue goes away.

Author:	echesakovMSFT
Assignees:	-
Labels:	area-System.Runtime.Intrinsics, untriaged
Milestone:	-

[echesakov]

@davidwrighton I am confirming (via the test #60047 I am working on) that the behavior is interaction between a crossgen2-d image and the Enable{ISA} flags.

crossgen2 assumes that Sse2.IsSupported, Ssse3.IsSupported, Sse41.IsSupported and Sse42.IsSupported (i.e. for any baseline ISA) would always return true, so, for example, setting EnableSse42=0 would not prevent the code in the R2R image from taking paths guarded with Sse42.IsSupported. However, for non-baseline ISA (e.g. Avx) setting EnableAvx=0 would lead to discarding the method code during the running time (i.e. the code would be re-jitted).