[Mlir-commits] [mlir] [MLIR][LLVM] Import LLVM target triple into MLIR LLVM Dialect (PR #125084)

[Tobias Gysi]

gysit wrote:

> @gysit Re not modeling all fields ... what're we missing? I suspect vector alignments and ... probably some of the address space metadata?

Yes vector and I believe aggregate and function pointer alignment as well. Plus some exotic name mangling flag etc.

> (My long term vision here is that, when going to LLVM, you'd have a, say, #rocdl.target<..., [otherDataLayoutEntries]> or #x86.target<..., [otherDataLayoutEntries]>` attribute sitting in scope so that all the things that're known from the LLVM data layout would be implicitly populated without having to write all of them out.

The main goal would be that the default data layout is chosen based on the target? That sounds like a good option to me.

PS I will land this commit later today should there be no more comments.

https://github.com/llvm/llvm-project/pull/125084