Move tools to use LKG ILC #107772
Move tools to use LKG ILC #107772
Conversation
|
Are we going to build the non-NativeAOT crossgen2 (the singlefilehost version) against the LKG bits? If so, could we instead just update/use the crossgen2.csproj project and have 2 projects instead of continuing to have the 3 separate projects? |
By that you mean, are we going to pull the single-file host from the LKG as well? I hadn't thought of that, but it makes sense. |
|
Yep that's what I was thinking of. I think however we ship crossgen2, we should ship it on equivalent runtime bits. The alternative would likely be more expensive in infra maintenence for minimal gain, unless it's required by source-build or something. |
| @@ -1,14 +1,7 @@ | |||
| <Project Sdk="Microsoft.NET.Sdk"> | |||
| <PropertyGroup> | |||
| <CrossHostArch Condition="'$(CrossBuild)' == 'true' or '$(TargetArchitecture)' != '$(BuildArchitecture)' or '$(HostOS)' != '$(TargetOS)' or '$(EnableNativeSanitizers)' != ''">$(BuildArchitecture)</CrossHostArch> | |||
There was a problem hiding this comment.
For ilc, this is set unconditionally in the crossarch project
Also, we should make the crossgen and ilc project files use the same naming scheme. I like the naming scheme for ilc project files (ILCompiler.csproj, ILCompiler_crossarch.csproj).
|
I've been experimenting locally, and I don't think we should get rid of the publish project either. The problem is that the SDK is architected such that I can't think of a way to do that without a different project if we also want things to use the non-published copy. If I set |
|
ilc has two projects (native and cross-arch). It does not have any special handling of publishing. Can crossgen be on the same plan? |
|
I think ILC is the incorrect one here -- while not setting Thus far ILC has not depended on any of those behaviors, but I think that's accidental. |
|
Maybe I am missing the bigger picture. Coverage-wise, both approaches are eventually testing what is being shipped:
Live build is better IMO, and it gives us control over enabling nativeaot for new platforms without waiting for the entire update cycle, runtime->sdk->runtime (which takes a month or two). End to end tests in dotnet/sdk repo are covering the nuget feed scenario. |
But it's also testing every version, as you're developing it on a local machine and in CI. So in addition to the complexity in the build system needed to produce the right phase ordering, you also make local development and CI uniquely unstable by changing Native AOT. Add to all of this, debug builds become unusably slow because the debug JIT is incredibly slow. Optimizing for live builds is optimizing for release validation, while using LKG builds is optimizing for infra simplicity and developer experience. |
|
Yup, @jkotas explained it to me about the release vs. debug performance penalty and I agree that live build (for non-testing) should be on release plan. BTW, are you looking to incorporrate the scenario #105004 is covering? I can close that one. Basically it's this workflow https:/am11/CrossRepoCITesting/actions/runs/10847119420/workflow (also covered in runtime CI, see eng/pipelines changes in the PR). |
In theory, yes. In practice, there isn't an LKG apphost for FreeBSD right? I haven't thought of what we should do in that case. Perhaps the answer is "require an LKG apphost just like we require an LKG runtime" |
I think we should introduce a new build option that builds everything that comes from LKG (including LKG apphost if necessary) and require invoking build with that option before invoking regular build if there is no LKG available for the target platform. #105004 (comment) has details. Introducing this build option should be a separate PR from this one. |
Yes, it's about the |
Starts with crossgen2 and moves shared helpers into a central targets file.
|
I believe this is ready for review. Two things this doesn't handle:
I think both of those are best left up to later PRs. |
eng/Subsets.props
Outdated
| <ProjectToBuild Include="$(CoreClrProjectRoot)tools\aot\crossgen2\crossgen2_inbuild.csproj" Category="clr" /> | ||
| <!-- skip the architectures that don't have LKG runtime packs --> | ||
| <ProjectToBuild Include="$(CoreClrProjectRoot)tools\aot\crossgen2\crossgen2_publish.csproj" Condition="'$(NativeAotSupported)' == 'true'" Category="clr" /> |
There was a problem hiding this comment.
A few questions:
- Was there a reason it was moved below?
- This project is also built by
src/installer/pkg/sfx/Microsoft.NETCore.App/Microsoft.NETCore.App.Crossgen2.sfxproj(i.e.packssubset), do we need this condition there? - If answer to 2. is "yes", then this is a regression on linux-riscv64 and illumos-x64 etc. where
NativeAotSupportedis false; i.e. main publishes crossgen2, while PR doesn't.. until the "later PR" is on the horizon, right?
There was a problem hiding this comment.
(1) No, just accidental due to moving things around
(2) I don't think it needs to be
| @@ -9,6 +9,7 @@ | |||
| Condition="'$(DisableVersionCheckImported)' != 'true'" /> | |||
|
|
|||
| <Import Project="$(RepoRoot)eng/liveBuilds.targets" /> | |||
| <Import Project="$(RepositoryEngineeringDir)toolAot.targets" /> | |||
There was a problem hiding this comment.
Do we still need to build crossgen during test build? Would it make more sense to go back to using the one created during the build?
There was a problem hiding this comment.
Looks like there are still reasons to keep this. In particular, I moved all the package overrides here, so if we need to use new ILC packages the overrides are here.
There was a problem hiding this comment.
I am talking about https:/dotnet/runtime/blob/main/src/tests/Common/Directory.Build.targets#L66-L78 and UsePublishedCrossgen2 property. Building tests is reaching out into the product build to publish crossgen2. It should not be needed anymore since the ordering problems that this was working around should not exist after this PR. We should be able to use what we have produced during the product build instead. (My assumption was that the line that I have commented on is necessary to make this work.)
I agree that ability to override ILC LKG as necessary is good. But building tests should not need ILC LKG to publish. If tests need to publish, they should use the live ILC.
There was a problem hiding this comment.
Do you have thoughts about this? Should triggering build of crossgen from the test partition be deleted? (The pain associated with that actually started the conversation that led to this PR.)
|
Will community supported platforms, in my case FreeBSD, still be able to generate a working runtime/ILC/Crossgen when building under Linux? |
|
@Thefrank I don't think there's anything preventing it, but the packages will not be built by default |
|
ugh, turns out we still need the inbuild ILC for the tests |
|
I think this is ready to merge |
|
@jkoritzinsky Mind also reviewing, as Jan is out? |
|
This appears to have broken crossgen outerloop (worse than it was already broken) by making the path for crossgen2.dll incorrect. It looks like #108693 will fix it but in the future it might be good to run crossgen outerloop for changes like this. |
Starts with crossgen2 and moves shared helpers into a central targets file.