[release/5.0] Prevent emitting Avx2 instruction for Vector256<T>.AllBitsSet and Vector256.Create(-1) when Avx2 is not supported

[echesakov]

This is backporting of #48383 to 5.0
Closes #48283

Customer Impact

In the two following cases

 var ones = Vector256<int>.AllBitsSet;

and

var minusOnes = Vector256.Create(-1);

on x86 and x64 the JIT emits vpcmpeqd ymmReg, ymmReg, ymmReg instruction (that is supported by Avx2 CPU feature - see pcmpeqb:pcmpeqw:pcmpeqd) even on a machine that doesn't support the feature.

For example, the issue occurs with CPUs that belong to Sandy Bridge or Ivy Bridge microarchitectures.

As you can see in #48283, a customer originally observed the issue on an Ivy Bridge machine. The execution of such code on their machine resulted in crash followed by terminating of the process by OS.

Validation

I don't have access to a machine that doesn't support Avx2. Therefore I used the same methodology we use in CI for ISA-related test scenarios (e.g. runtime-coreclr jitstress-isas-x86) where we set corresponding COMPlus environment variables to simulate that a CPU feature is not supported.

I generated all possible variations of Vector128/256<T>.AllBitsSet and Vector128/256.Create((T)-1) (in the latter case T is a signed integer type) (see Runtime_48283.cs) and manually validated that the JIT doesn't emit the forbidden instructions.

You can find below results of the JIT output for the following three cases - noavx (simulated with COMPlus_EnableAVX=0), avx (simulated with COMPlus_EnableAVX2=0) and avx2 (when none of the ISAs were forbidden) for both x86 and x64.

• noavx-win-x64.asm

• avx-win-x64.asm

• avx2-win-x64.asm

• noavx-win-x86.asm

• avx-win-x86.asm

• avx2-win-x86.asm

After the change, the JIT will emit vcmptrueps ymmReg, ymmReg, ymmReg for such cases. This is similar to what Clang does - https://godbolt.org/z/erd5sf.

Risk

Low, the changes impacts only code generation of hardware intrinsics Vector128/256<int>.AllBitsSet that were introduces in 5.0 and optimization in the JIT of Vector256.Create(-1) that uses the mentioned intrinsics (that was also done in 5.0).

Regression

Yes, the Vector256.Create(-1) on 3.1 wouldn't use Avx2 instruction.
There is no Vector128/256<int>.AllBitsSet API in 3.1.

Note

The change depends on #48613 being merged - where another problem with Vector256.Create(-1) optimization was fixed.
The problem is that, at the moment, the JIT lowers both Vector128.Create(-1) and Vector256.Create(-1) to Vector128<T>.AllBitsSet without taking into account the vector size (see src/coreclr/src/jit/lowerxarch.cpp

[echesakov]

@dotnet/jit-contrib @tannergooding PTAL

[echesakov]

Can I have sign-off from someone on @dotnet/jit-contrib please?

[briansull]

Looks Good

[echesakov]

cc @jeffschwMSFT

[jeffschwMSFT]

Approved. We will take for consideration in .NET 5.0.x