[crossgen2] Making sure Vector64/128/256 are not generated unless CoreLib is part of the version bubble

[cshung]

No description provided.

[4creators]

Is it possible to provide a description of the problem and software architecture associated with it? It would make the issue understandable to all coreclr contributors.

[cshung]

Is it possible to provide a description of the problem and software architecture associated with it? It would make the issue understandable to all coreclr contributors.

This issue is created to track the work required to address one of these work items mentioned as a comment in my pull request.

dotnet/coreclr#26772 (review)

The pull request was created as a port of these 3 pull requests (#24689, dotnet/coreclr#24917, dotnet/coreclr#25365). As I understand it, the original pull request allow crossgen to use hardware intrinsics. The port was intended to allow crossgen2 to also use hardware intrinsics.

To be brutally honest, I don't actually understand what the problem is, at least not in a clearly reproducible manner. The comment must have written to make sure certain scenarios will not fail. I have yet to figure out what those scenarios are.

@MichalStrehovsky who authored the original pull requests and the original comment should be able to explain in detail what the problem is.

[MichalStrehovsky]

To be brutally honest, I don't actually understand what the problem is, at least not in a clearly reproducible manner

@cshung does this comment help? https://github.com/dotnet/coreclr/pull/24689/files#diff-f2c1beb35334bda2c2e3c3cbbc0ccd54R9559-R9566

I wasn't part of the discussion that lead to the initial disabling, but reading through the linked issue gives a good summary of why we need to avoid AOT compiling them for now (until the issue is resolved) and why it would be okay to let precompilation happen in CoreLib (or in things that version with CoreLib).

[mangod9]

@janvorli is there anything actionable here?

[janvorli]

I believe @davidwrighton has fixed that in one of his changes a long time ago, but I am not 100% sure.

[davidwrighton]

I'll look into this in depth tomorrow.

[davidwrighton]

This didn't get implemented, but I've just put together a PR that identifies where the ABI may be unstable and surgically disables code gen for those cases. It preserves the ability to use Vector intrinsics even in the presence of unstable abi.