[SYCL] Correctly handle debug information in global offset pass

[jchlanda]

Make sure that cloned functions do not use debug information that points to the original function. In doing so for each cloned function, create a DISubprogram that describe the clone and walk the instructions of cloned functions to correctly fix up the debug information.

[asudarsa]

Thanks for the PR.
Two pointers:

1. I think this will benefit from a review from someone in the debug info team.
2. Looking at the change itself, this issue could arise whenever anyone tries to clone a function. Can we address this issue upstream whenever we clone a function?

Thanks

[jchlanda]

2. Looking at the change itself, this issue could arise whenever anyone tries to clone a function. Can we address this upstream whenever we clone a function?

You are right, I think if we manage to populate the ValueToValueMapTy &VMap of the CloneFunctionInto with all the duplicated functions, then llvm would resolve the debug information for us, and the manual fixups done in the first commit will not be needed.
I've marked this PR as a draft and will explore the VMap solution.

Thank you.

[asudarsa]

Hi @jchlanda

Can you please let me know about the recent changes on this PR? I see only one commit and am unable to comment about the changes after my earlier comment.

Thanks

[jchlanda]

Hi @jchlanda

Can you please let me know about the recent changes on this PR? I see only one commit and am unable to comment about the changes after my earlier comment.

Thanks

HI @asudarsa

I've followed your suggestion and worked on a solution that utilizes llvm's CloneFunction machinery. Essential what this patch does now is that it creates all the functions that would be cloned during the execution of the pass ahead of time - that is before we get to create functions with offset. This is done so that we can populate a value-to-value map that llvm uses to keep track of all the debug information amongst other).

While this PR was in draft state, I rebased it and dropped some of the earlier, now not needed, commits.

Thanks.

[jchlanda]

Ping @intel/dpcpp-tools-reviewers this PR should now be ready for your review.
Thank you.

[jchlanda]

Friendly ping @intel/dpcpp-tools-reviewers

[npmiller]

@intel/dpcpp-tools-reviewers could we get some help with an approval/review on this? It's actually only needed for CUDA and HIP (maybe we should change the codeowners on these files?), and we'd like to get this patch into the next release.

[npmiller]

Updated the codeowners here: #17484

[jchlanda]

With the change to the codeowners structure and @npmiller's approval this should be read to merge @intel/llvm-gatekeepers

[sommerlukas]

@jchlanda Github UI still says that merge is blocked due to missing review from @intel/dpcpp-tools-reviewers, so the change to code owners might only take effect for new PRs.

@intel/dpcpp-tools-reviewers can we get a quick review on this one?

[maarquitos14]

LGTM.

[jchlanda]

@intel/llvm-gatekeepers we should be ready to roll now. Thank you.