[SYCLLowerIR] Fix a bug in hierarchical parallelism implementation

[lbushi25]

When inside a parallel_for_work_group context, calls to functions that contain calls to parallel_for_work_item are being lowered incorrectly into IR in that they are being put under the work-group leader branch in the IR which is semantically incorrect as this function should be called in every work item. This manifests when we have an indirect function call to parallel_for_work_item together with at least one other direct call to parallel_for_work_item in the same parallel_for_work_group context and it leads to a program that hangs. This PR fixes the issue and adds a couple of other tests to check this behavior.

[YuriPlyakhin]

LGTM

[aelovikov-intel]

Why E2E test instead of exercising the transformation directly through opt?

[slawekptak]

SYCL changes LGTM. Edit: I've missed Andrei's comment above - please consider it before merging.

[github-actions]

@intel/llvm-gatekeepers please consider merging

[lbushi25]

Why E2E test instead of exercising the transformation directly through opt?

Thanks, added a test case there as well.

[lbushi25]

@intel/llvm-gatekeepers This should be good to merge!

[sarnex]

I'd prefer if someone reviewed the new code added since last review (@aelovikov-intel ?)

[aelovikov-intel]

I'd prefer if someone reviewed the new code added since last review (@aelovikov-intel ?)

It's outside SYCL RT scope, offload utilities should review it (IIUC).

[lbushi25]

Touching this file has not added any new reviewers, nevertheless I am tagging @againull for his input on the changes made in pfwg_and_pfwi.ll as he seems to have created this file.

[againull]

I wonder what happens if there are additional instructions in foo (in work group scope) except parallel_for_work_item call. Is this handled properly or is it a separate bug?

[lbushi25]

I have added some other instructions to the foo function. No failures introduced.