Convert JIT\Directed tests to merged test groups

[markples]

(based on #81969)

[ghost]

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

Issue Details

---

(based on #81969)

Author:	markples
Assignees:	markples
Labels:	area-System.Reflection.Metadata
Milestone:	-

[markples]

/azp run runtime-coreclr outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[BrianBohe]

LGTM! Let see if that typo was the only problem

[BrianBohe]

/azp run runtime-coreclr outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[markples]

/azp run runtime-coreclr outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[BrianBohe]

Does it make sense to split HardwareIntrinsics_ro? Would that avoid these timeouts? I don't see issues related to this pr

[BrianBohe]

Does it make sense to split HardwareIntrinsics_ro? Would that avoid these timeouts? I don't see issues related to this pr

What do you think about this @trylek ?

[trylek]

What is the number of tests in the group and its duration in checked mode (without any stress modes)?

[markples]

It looks like it has 5834. I see a successful x64 run that took about 5 minutes. Even the timeouts are only showing 5-10 minutes in the test group. This includes timestamps that appear to be after the test script is terminated, so I don't think it's getting stuck. (These tests are generated, etc., so there are a lot of them... @davidwrighton added various things like the test striping mechanism just to handle them at all.. so maybe there's more to do there.)

There's a comment:

<!-- For Vector512, we only have a very small pool of machines with acceleration support, so they are always outerloop -->

Could this be a case where queue time is counting against a timeout value?

[markples]

/azp run runtime-coreclr outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[trylek]

Hmm, both these values seem far beyond what I'd consider reasonable. Please remember that in GC / JIT stress modes the tests are typically two to three orders of magnitude slower and Helix items taking 5000 minutes are less fun than they seem. Under JIT/Methodical, I tried to keep the group size around 400 tests, we can certainly be more aggressive for tests that are really tiny but running the entire Pri1 test set as a single Helix item is obviously not desirable for multiple reasons.

[davidwrighton]

@trylek, the tests are setup so that under striping, we run only about 600 or so tests , which results in them finishing in a vaguely reasonable timeframe under stress. If we're seeing intermittent failures here, then yep, we've got a real product bug to look into, and what I saw here looked like an intermittent failure that should be investigated.

[markples]

Thanks @davidwrighton for pointing out the details on striping - I now see the msbuild variable is called NumberOfStripesToUseInStress so my reporting of a normal run was inaccurate.

I am curious if you can share your reasoning for suspecting a product issue here. Looking at one of the failures (AzDO build https://dev.azure.com/dnceng-public/public/_build/results?buildId=200759&view=results), I see

top of log:

BEGIN EXECUTION
"C:\h\w\9D430947\p\watchdog.exe" 300 "C:\h\w\9D430947\p\corerun.exe" -p "System.Reflection.Metadata.MetadataUpdater.IsSupported=false" HardwareIntrinsics_ro.dll 22:21:46.449 Running test: global::JIT.HardwareIntrinsics.Arm._AdvSimd.Arm64.Program.Abs_Vector128_Double()
Supported ISAs:

bottom of log:

22:26:48.580 Passed test: global::JIT.HardwareIntrinsics.General._Vector128_1.Program.op_AdditionByte()
22:26:48.581 Running test: global::JIT.HardwareIntrinsics.General._Vector128_1.Program.op_AdditionDouble()
Beginning scenario: RunBasicScenario_UnsafeRead
SUCCESS: The process "corerun.exe" with PID 5812 has been terminated.
2023-03-10T22:26:49.815Z	INFO run.py	run(48)	main	Beginning reading of test results.

Start @ 22:21:46.449
Start failing test @ 22:26:48.581
Results message @ 22:26:49.815

The failing test is only running for just over a second before the overall 5 minute timeout hits.

Recent changes have greatly bumped up this value, so if you are seeing a product issue, then it may be hidden now.

[markples]

The failures we've seen so far have been seen elsewhere, so I will resolve the conflicts and update this for merging. @trylek, Brian and I have both made changes/reviewed this, so please let me know if there's anything that you'd like to look at more closely. Thanks!

[trylek]

Looks great to me, thanks Mark!

[SamMonoRT]

@markples@JulieLeeMSFT@BrianBohe - Looking at https://dev.azure.com/dnceng-public/public/_build/results?buildId=226494&view=logs&j=e45de9b4-b3b3-54f9-2ea5-8e56201c788d&t=92db4d17-27b4-5f87-01bd-8cef7a796420 failure,

1. For such large extensive changes to test infrastructure, given same tests are run for both runtimes, please kick off "/azp run runtime-extra-platforms" which should trigger some more lanes, one of which is linux-arm64 Release AllSubsets_Mono_LLVMFullAot_RuntimeTests llvmfullaot' extra-platforms lane. During that, the LLVM AOT cross-compile CoreCLR tests job runs, fails on a couple tests, hence the Send to Helix job doesn't collect any logs. Mono doesn't seem to have any aot-llvm arm64 runtime test coverage due to this.

2. Example of that is https://dev.azure.com/dnceng-public/public/_build/results?buildId=226494&view=logs&j=e45de9b4-b3b3-54f9-2ea5-8e56201c788d&t=92db4d17-27b4-5f87-01bd-8cef7a796420

3. The two test failing in above are : Regression_1.dll:

• Assertion at /__w/1/s/src/mono/mono/mini/aot-compiler.c:3956, condition `image_index < MAX_IMAGE_INDEX' not met
  GitHub_27678.dll:
• Assertion at /__w/1/s/src/mono/mono/mini/memory-access.c:120, condition `size < MAX_INLINE_COPY_SIZE' not met

4. A PR was merged to fix one of those : [mono][aot] Fix an assert in the aot compiler. #84343

5. For the second one, GitHub_27678.dll, it seems it is ignoring the condition in the issues.target file to exclude for Mono runtime. But on deeper look in the failure logs, it shows file at .../coreclr/linux.arm64.Release/JIT/Regression/Regression_3/GitHub_27678.dll, while as in the issues.target file, it seems to be ignoring from the JitBlue directory as in: ....JIT/Regression/JitBlue/GitHub_27678

You should work to disable the test correctly for Mono. Also @trylek I want to understand motivation behind the coreclr cross-compiler tests running under a Mono AOT-LLVM arm64 lane -- shouldn't these be two different lanes ?

[markples]

@SamMonoRT Thank you for the info and investigation. I'll add runtime-extra-platforms.

The naming in issues.targets will work the same before and after merging unless certain internal test changes are made (I believe that splitting a single Main into multiple [Fact]s can cause this). However, merged tests cause this AOT step to happen during Regression_3 for all of the tests inside the group before we check for specific tests. I opened #84380, which effectively pulls GitHub_27678 into a separate executable so that the exclusion happens in time.

[markples]

@SamMonoRT I just wanted to double-check that you weren't waiting for me on anything else. I think these two were the full impact (and both builds fixed so I abandoned my PR - one still fails at runtime but issue.targets handles that).