Debug info support

[bcardosolopes]

LLVM lowering doesn't yet include debug information, nor CIRGen currently emits it in any form. Most of the missing parts are tracked by functions in UnimplementedFeatures::*.

This issue can be used to track the overall work needed, but specific tasks should be tracked in their own issues. There are also multiple ways to go about implementing this, so it's probably good to start some discussion here before putting up a PR.

[PikachuHyA]

I plan to support debug info in ClangIR.
and I have found relevant materials.

• [CIR] Enabling debug loc to be progatated into LLVM. #96
• [CIR][Lowering] Attaching more debug information to CIR module #99
• [CIR][Lowering][debuginfo] Disable debug info if -g is not specified #1145
• Consider capturing more location information #1062
• add names to generated IR #969

Here’s a summary

• Detach debug information generation from CIRGen
• Incrementally add attributes as needed, e.g. #tu_name
• Only changes in clang/lib/CIR will be accepted.
• Consider utilizing OpaqueLoc.

Could you please provide me with more suggestions?

[bcardosolopes]

Detach debug information generation from CIRGen

Yeah, we would love if debug info emission can be done as separate pass, but take it with a grain of salt because we don't know all the restrictions necessary to make it happen, I guess we'll learn with your PRs too :)

Only changes in clang/lib/CIR will be accepted.

As a matter of principle, but if we need to change things upstream to make our lives easier we should consider.

Could you please provide me with more suggestions?

Seems like there are currently limitations of trying to represent type recursion with debug info, it's possible it could affect your work - @gysit might have more pointer for you.

[gysit]

Seems like there are currently limitations of trying to represent type recursion with debug info, it's possible it could affect your work - @gysit might have more pointer for you.

The recursive debug info support is quite complex but it exists. There is a lengthy thread where different solutions have been discussed:
https://discourse.llvm.org/t/handling-cyclic-dependencies-in-debug-info/67526/18?u=gysit

The idea today's implementation is to translate a cyclic graph into a directed acyclic graph by representing recursive nodes either by the translated node + id or by the id only if the node has been traversed before.

The translation process from cyclic graph to directed acyclic graph is quite complex. Luckily there is a CyclicReplacer cache that should simplify things considerably. An example usage is the DebugImporter that is used to import LLVM IR debug info into LLVM dialect:
https://github.com/llvm/llvm-project/blob/84efad0b471543003c0724c85f158f66fccfdc0f/mlir/lib/Target/LLVMIR/DebugImporter.cpp#L419
I believe this is a good starting point if you plan to approach this problem.

Otherwise, LLVM dialect covers a lot of LLVM IR's debug metadata but definitely not everything. It may very well be that you find missing fields or even missing node types if you start working with it. Features have been added on a need basis.

The following file contains the available DINode types (the link points to the composite type node but there are more):
https://github.com/llvm/llvm-project/blob/84efad0b471543003c0724c85f158f66fccfdc0f/mlir/include/mlir/Dialect/LLVMIR/LLVMAttrDefs.td#L374C5-L374C29