[clang] [llvm] [DebugInfo] Place local ODR-uniqued types in decl DISubprograms (PR #142166)

[Vladislav Dzhidzhoev via cfe-commits]

dzhidzhoev wrote:

> > So I was thinking in the opposite direction: should we force unique distinct definition DISubporgrams containing ODR-uniqued types, as well as it's done for declaration subprograms contained by ODR-uniqued types?
> 
> I'm suggesting more broadly: we should force-unique all DISubprograms (even if they don't contain odr-uniqued types), by making them not distinct anymore & generalizing the unquing infrastructure you quoted in langref.
> 
> This should address a broader class of problems - including, for instance, LTO where a subprogram is inlined/out of line in multiple files.

In order to implement that, I'd suggest the following changes:
* Allow different llvm::Functions to refer to the same (distinct) DISubprogram (such case can pop up during LTO if subprograms are uniquied). I have sketched the implementation of that in https://github.com/llvm/llvm-project/pull/152680, https://github.com/llvm/llvm-project/pull/162852, https://github.com/llvm/llvm-project/pull/162854.
* Modify LLVM passes/DIBuilder interface so that (still distinct) subprograms are not modified, new subprograms are created when necessary.
* Modify CloneFunctionInto so that DISubprogram attached to a function is cloned only when explicitly requested, to reuse DISubprograms for function clones, and to establish one-to-one correspondence between abstract DISubprograms and abstract DIEs on the level of one CU/Module (when LTO is not enabled). It will also help to field-test https://github.com/llvm/llvm-project/pull/162854.
* Switch to uniqued DISubprograms. Merge retainedNodes of DISubprograms that are about to be uniqued with each other during LTO.
* Possibly, implement ODR-uniquing of DISubprograms based on linkageName during LTO.
* Attempt mergin https://github.com/llvm/llvm-project/pull/75385. Modify it in a way that local types are filtered out from DISubprogram's retainedNodes whenever DISubprogram is cloned (but the original DISubprogram is still used) to avoid ODR violations.

Does that sound reasonable? And will the performance regression caused by extra uniquing during LTO be acceptable?

https://github.com/llvm/llvm-project/pull/142166