Assertion failed 'value != 0' in 'System.Tests.DoubleTests_GenericMath:GetExponentShortestBitLengthTest()' during 'Do value numbering'

[MichalStrehovsky]

ILC: Assertion failed 'value != 0' in 'System.Tests.DoubleTests_GenericMath:GetExponentShortestBitLengthTest()' during 'Do value numbering' (IL size 314; hash 0x60919730; FullOpts)
  
D:\a\_work\1\s\artifacts\bin\coreclr\windows.x64.Checked\build\Microsoft.NETCore.Native.targets(262,5): error MSB3073: The command ""D:\a\_work\1\s\artifacts\bin\coreclr\windows.x64.Checked\ilc-published\\ilc" @"D:\a\_work\1\s\artifacts\obj\System.Runtime.Tests\Release\net8.0-windows\native\System.Runtime.Tests.ilc.rsp"" exited with code 57005. [D:\a\_work\1\s\src\libraries\System.Runtime\tests\System.Runtime.Tests.csproj::TargetFramework=net8.0-windows]
##[error]artifacts\bin\coreclr\windows.x64.Checked\build\Microsoft.NETCore.Native.targets(262,5): error MSB3073: (NETCORE_ENGINEERING_TELEMETRY=Build) The command ""D:\a\_work\1\s\artifacts\bin\coreclr\windows.x64.Checked\ilc-published\\ilc" @"D:\a\_work\1\s\artifacts\obj\System.Runtime.Tests\Release\net8.0-windows\native\System.Runtime.Tests.ilc.rsp"" exited with code 57005.

This is failing the NativeAOT System.Runtime tests in runtime-extra-platform runs with checked runtime/release framework combo.

https://dev.azure.com/dnceng-public/public/_build/results?buildId=155217&view=logs&jobId=62c84a0c-1188-56fe-6938-e058a633beb1&j=62c84a0c-1188-56fe-6938-e058a633beb1&t=e508e8f5-7a40-5591-4aae-9625248f10c0

Looks to be related to #80789 cc @tannergooding

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch, @kunalspathak
See info in area-owners.md if you want to be subscribed.

Issue Details

---

ILC: Assertion failed 'value != 0' in 'System.Tests.DoubleTests_GenericMath:GetExponentShortestBitLengthTest()' during 'Do value numbering' (IL size 314; hash 0x60919730; FullOpts)
  
D:\a\_work\1\s\artifacts\bin\coreclr\windows.x64.Checked\build\Microsoft.NETCore.Native.targets(262,5): error MSB3073: The command ""D:\a\_work\1\s\artifacts\bin\coreclr\windows.x64.Checked\ilc-published\\ilc" @"D:\a\_work\1\s\artifacts\obj\System.Runtime.Tests\Release\net8.0-windows\native\System.Runtime.Tests.ilc.rsp"" exited with code 57005. [D:\a\_work\1\s\src\libraries\System.Runtime\tests\System.Runtime.Tests.csproj::TargetFramework=net8.0-windows]
##[error]artifacts\bin\coreclr\windows.x64.Checked\build\Microsoft.NETCore.Native.targets(262,5): error MSB3073: (NETCORE_ENGINEERING_TELEMETRY=Build) The command ""D:\a\_work\1\s\artifacts\bin\coreclr\windows.x64.Checked\ilc-published\\ilc" @"D:\a\_work\1\s\artifacts\obj\System.Runtime.Tests\Release\net8.0-windows\native\System.Runtime.Tests.ilc.rsp"" exited with code 57005.

This is failing the NativeAOT System.Runtime tests in runtime-extra-platform runs with checked runtime/release framework combo.

https://dev.azure.com/dnceng-public/public/_build/results?buildId=155217&view=logs&jobId=62c84a0c-1188-56fe-6938-e058a633beb1&j=62c84a0c-1188-56fe-6938-e058a633beb1&t=e508e8f5-7a40-5591-4aae-9625248f10c0

Looks to be related to #80789 cc @tannergooding

Author:	MichalStrehovsky
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[EgorBo]

Presumably assert should just be removed from those helpers (or special cased to return value for 0)

[tannergooding]

@EgorBo, the issue for BSR and BSF is that 0 as an input is undefined and there is an actual functional difference between them for Intel vs AMD and is documented as being "undefined".

Which is why all the usages in managed have an explicit path handling 0 (or in the case of Log2, doing value |= 1 first).

Likewise in importercalls we always introduce a (op1 == 0) ? ... : BSF (or (op1 == 0) ? ... : BSR) QMARK.

So it's unclear what is occurring that could lead to this and why it is only showing up in NativeAOT

[tannergooding]

@MichalStrehovsky, what's the easiest way to repro this and see get a JIT dump?

[MichalStrehovsky]

So it's unclear what is occurring that could lead to this and why it is only showing up in NativeAOT

Could this be because NativeAOT targets the minimum x64 we support and this is untested in libs tests outside NativeAOT? We had several of those.

@MichalStrehovsky, what's the easiest way to repro this and see get a JIT dump?

Follow this workflow:

runtime/docs/workflow/building/coreclr/nativeaot.md

Line 102 in 58176f5

### Running library tests

. Then you'll get a rsp file to invoke the compiler under obj. Invoke the compiler again with the rsp file and pass --codegenopt:JitDump=blah. Alternatively, you can add <ItemGroup><IlcArg Include="--codegenopt:JitDump=blah" /></ItemGroup> to the test project file.

[tannergooding]

Could this be because NativeAOT targets the minimum x64 we support and this is untested in libs tests outside NativeAOT? We had several of those.

We also hit this for crossgen since lzcnt is "opportunistic" and we validate this path in the outerloop with the various Enable*=0 paths that all hit it currently passing.

It's possible its getting masked via some other env variable being set (maybe one of the JitStress ones?), but I don't initially see what could trigger it.

Follow this workflow:

Thanks! Will give this a try.

[MichalStrehovsky]

We also hit this for crossgen since lzcnt is "opportunistic" and we validate this path in the outerloop with the various Enable*=0 paths that all hit it currently passing.

I think we don't run libs tests with crossgen - this is asserting while compiling a test in src/libraries. We'll do some limited crossgenning of libs tests after #80946, but likely not with extra COMPlus_ matrix. We're only equipped to find these bugs in src/tests.