[Regression][JitStressRegs=0x2] Assertion failed '!varDsc->lvIsRegCandidate()'

[sandreenko]

The test started to fail 4 days ago : ci example.

cc @dotnet/jit-contrib

"Don't struct-promote opaque vectors" dotnet/coreclr#21314 looks like the most relevant change, @CarolEidt could you please take a look?

[CarolEidt]

The test is JIT\HardwareIntrinsics\General\Vector128_1\Vector128_1_ro\Vector128_1_ro.exe
It doesn't repro on arm64 windows (or on the windows x64 arm64 altjit).

[tannergooding]

The explicit scenario failing is: JIT.HardwareIntrinsics.General.VectorGetAndWithLowerAndUpper__GetAndWithLowerAndUpperByte:RunReflectionScenario(). The RunBasicScenario always runs first, so that succeeded and this may be something specific to the reflection code.

[tannergooding]

Given that this is the VectorGetAndWithLowerAndUpper test, that also means that it is either:

• Testing GetLower/GetUpper, in which case the parameter is a Vector128<T> (the this parameter) and it would return a Vector64<T>
  -or-
• Testing WithLower/WithUpper, in which case the parameter is a Vector128<T> (the this parameter) and a Vector64<T> (the new value to insert) and it would return a new Vector128<T>

[CarolEidt]

@sandreenko - I'm stumped on this. I wasn't able to repro on arm64/linux, even after resetting to dotnet/coreclr@13f8624 which was the last commit in the batch that were being tested.
@dotnet/jit-contrib - any suggestions? I haven't yet tried doing a full run, but it fails to repro with:

export COMPlus_TieredCompilation=0
export COMPlus_JitStressRegs=2

and then either running the .sh file with -coreroot or running the .exe directly with corerun.

[sandreenko]

I saw this in different ci jobs, so it can't be just a noise.
For example job that failed 8 hours ago.

[CarolEidt]

So ... when the test build completes, it teases you with instructions on how to run a single test. However, at least for this test case, the single test ran just fine until I ran it as part of a full test run - then it failed, and continued to fail when run individually, with the same command and environment that was passing before. It's not clear to me what is changing, but clearly the full test run does some additional configuration that wasn't being done before.

In any event, I have a PR with a fix. It wasn't strictly related to dotnet/coreclr#21314, except that that PR caused the lclVar of interest in this case to no longer be promoted, so it was a register candidate - which would have been fine except that due to stress it didn't get a register, and then when it was defined its liveness wasn't being updated.