Not possible to specify IlcInstructionSet < sse4.2 in .NET 10

[MichalStrehovsky]

This is a problem in .NET 10. Benchmark.net is running into it.

ILC is erroring out with error Unrecognized instruction set sse.

ilc --help prints:

The allowable values for the --instruction-set option are described in the table below.
...
x64: base, base, base, sse4.2, sse4.2, sse4.2, sse4.2, sse4.2, avx, avx2, avx2, avx2, avx2, avx2, avx2, avx2, avx512
...

A .NET 9 ILC instead prints:

The allowable values for the --instruction-set option are described in the table below.
...
x64: base, sse, sse2, sse3, ssse3, sse4.1, sse4.2, avx, avx2, aes, bmi, bmi2, fma, lzcnt, pclmul, popcnt, avxvnni
...

Notice base and sse4.2 getting repeated.

I've not tried to narrow it down to a specific PR yet. Maybe @tannergooding could have some idea.

[dotnet-policy-service]

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

[tannergooding]

This would have been #116406 from a couple months back.

We simplified the JIT support to better represent the real world targets rather than any potential target and so (which resulted in significant removal of complex code):

• SSE, SSE2 -> became part of base (they were already the baseline)
• SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT -> became simply sse4.2 (guaranteed even in Arm64 emulated environments and largely needed together for practical use, this was the R2R target and things get weird and often inefficient if not all together)
• AVX2, BMI1, BMI2, FMA, F16C -> become simply avx2 (always bundled in actual hardware)
• AVX512F, AVX512VL, AVX512BW, AVX512CD, AVX512DQ -> became simply avx512 (were already required together)

A breaking change notice was opted against since the configuration knobs and the ILC --instruction-set options were explicitly specified to be undocumented and subject to change without notice (#115983 (comment)).

It is unexpected that the command line lists them over and over however, it should just be one of each like we're getting for avx512 there...

[stephentoub]

I raised this with @MichalStrehovsky because benchmarkdotnet is broken now with .NET 10 when trying to benchmark with Native AOT. When I add --runtimes nativeaot10_0 to the bdn command to target Native AOT in .NET 10, it fails to run benchmarks with:

Generating native code
EXEC : error : Unrecognized instruction set sse [/home/stoub/benchmarks/bin/Release/net10.0/benchmarks-Job-JSQCBN-1/BenchmarkDotNet.Autogenerated.csproj]
  System.CommandLine.CommandLineException: Unrecognized instruction set sse
     at System.CommandLine.Helpers.ConfigureInstructionSetSupport(String, Int32, Boolean, TargetArchitecture, TargetOS, String, String, Logger, Boolean) + 0xc3f
     at ILCompiler.Program.Run() + 0x780
     at ILCompiler.ILCompilerRootCommand.<>c__DisplayClass261_0.<.ctor>b__0(ParseResult) + 0x31a

If this is by design for .NET 10, we need to fix benchmarkdotnet.

[tannergooding]

Looks like this is https://github.com/dotnet/BenchmarkDotNet/blob/e704adb6926f4743a361800676817094ca1787d0/src/BenchmarkDotNet/Toolchains/NativeAot/Generator.cs#L240

I'll put up a PR since it's a simple fix.

@MichalStrehovsky do we want to also fix the repeated instruction set entries for .NET 10? If so, I can get a PR up for that as well

[jkotas]

https://github.com/dotnet/BenchmarkDotNet/blob/e704adb6926f4743a361800676817094ca1787d0/src/BenchmarkDotNet/Toolchains/NativeAot/Generator.cs#L250

Why is this not taking the early out that just returns native?

[tannergooding]

It looks like x86 on x64 and vice versa, as well as x64 on Arm64 and others won't end up matching and allowing native: https://github.com/dotnet/BenchmarkDotNet/blob/e704adb6926f4743a361800676817094ca1787d0/src/BenchmarkDotNet/Portability/RuntimeInformation.cs#L206

Is this a scenario where BDN is doing such a "cross" check?

[tannergooding]

This has been resolved. We'll need to wait for BDN to publish an update