Tests which depend on the host don't use a live built version

[ViktorHofer]

Both the linker tests and the singlefile tests use a prebuilt version of the host instead of the live built one. In general we want to use live built assets where possible. In addition to that, when the repository revs to the next major version and the SDK doesn't yet understand the net7.0 tfm, the prebuilt host isn't discovered without help (KnownAppHostPack AppHostPackVersion update).

TL;DR: Tests which depend on the apphost should have a dependency on the host subset being built first.

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

Issue Details

---

Both the linker tests and the singlefile tests use a prebuilt version of the host instead of the live built one. In general we want to use live built assets where possible. In addition to that, when the repository revs to the next major version and the SDK doesn't yet understand the net7.0 tfm, the prebuilt host isn't discovered without help (KnownAppHostPack AppHostPackVersion update).

TL;DR: Tests which depend on the apphost should have a dependency on the host subset being built first.

Author:	ViktorHofer
Assignees:	-
Labels:	area-Infrastructure
Milestone:	7.0.0

[jkoritzinsky]

cc: @agocke

[agocke]

TL;DR: Tests which depend on the apphost should have a dependency on the host subset being built first.

Note: singlefilehost is part of the CLR partition. I think the single file tests all use the just-built single file host, but I may be forgetting some

[am11]

Tests which depend on the apphost should have a dependency on the host subset being built first.

Since cab3a63, corehost is built before libs subset, so I guess this objective is completed more broadly.

Can't wait to see live corehost being dogfooded to those tests (and other usages of dotnet(1) in the repo). :)

[ViktorHofer]

@agocke now that the live built host is part of the testhost folder this item tracks removing the remaining dependencies on a prebuilt host. It would be great if we could tackle this before we rev to .NET 8.

The change would need to be done directly in targetingpack.targets by supplying the path to the host

[am11]

Host packages were also removed from Versions.props in #71725.
#72371 was closed so bot can open a new PR without host dependencies updates (it hasn't opened a new PR so far).

ps- I think we can also use live ilasm and remove its package as well?

[ViktorHofer]

ps- I think we can also use live ilasm and remove its package as well?

In theory yes, but the tricky part is the Microsoft.NET.IL.Sdk which today is used by all the .ilproj project files. We would need to convert those to not use the Sdk attribute and instead import the .targets file manually.

[hoyosjs]

ps- I think we can also use live ilasm and remove its package as well?

In theory yes, but the tricky part is the Microsoft.NET.IL.Sdk which today is used by all the .ilproj project files. We would need to convert those to not use the Sdk attribute and instead import the .targets file manually.

There's an extension point in the SDK, where setting ILAsmToolPath is enough to use a custom ILAsm. We use it here:

runtime/Directory.Build.props

Line 118 in 178553b

<ILAsmToolPath Condition="'$(DotNetBuildFromSource)' == 'true' or '$(BuildArchitecture)' == 's390x'">$([MSBuild]::NormalizeDirectory('$(ArtifactsBinDir)', 'coreclr', '$(TargetOS).$(TargetArchitecture).$(Configuration)'))</ILAsmToolPath>

for s390x and sourcebuild. Don't see why not extend except for having to build the IL tools component.

[ViktorHofer]

Right, we definitely want to use such an extension point, but I was referring to that even then we would still use the Microsoft.NET.IL.Sdk prebuilt package which we should also get rid off to avoid the self dependency in the source build graph.

See #71326 for more details.