Add DPCPP_ENABLE_OPAQUE_POINTERS option

[bader]

This option controls the default behavior of LLVM projects regarding
opaque pointers. The default value is OFF as most of the SYCL compiler
components are not ready to handle opaque pointers.

This change adds a new nightly job testing SYCL compiler with enabled
opaque pointers by default.

[bader]

@dwoodwor-intel, @jcranmer-intel, @chrulski-intel, FYI.

[pvchupin]

CI changes looks ok

[asudarsa]

'Tools' related changes look ok to me. Thanks

[premanandrao]

FE changes look okay to me.

[bader]

@andykaylor, @asudarsa, could you take a look at new commits, please?

[bader]

@andykaylor, @asudarsa, could you take a look at new commits, please?

Ping.

[asudarsa]

Hi @bader
Sorry for delay in response. LGTM except for one issue I have highlighted.
Thanks

[asudarsa]

LGTM. Thanks

[bader]

@andykaylor, this PR holds off follow-up fixes, so I'd like to merge it as soon as possible. Is it okay to commit current version and address further comments (if any) separately?

[bader]

@jchlanda, do you think that this failure on NVIDIA GPU https://github.com/intel/llvm/runs/7391421134?check_suite_focus=true is related to this change? This change is not expected to change the default behavior.

[bader]

@jcranmer-intel, I've added ab05021, which updates a few tests.

• sycl/test/check_device_code/id_queries_fit_int.cpp - now checks are invariant to opaque pointer setting (perfect case).
• sycl/test/check_device_code/kernel_arguments_as.cpp - the purpose of the checks is to validate that address space information is passed via LLVM IR and pointer type is not relevant for this test. I forced opaque pointers use and updated the checks.
• sycl/test/check_device_code/no_offset.cpp - what is interesting in this test is that compiler is able to remove certain kernel parameters if nooffset accessor property is set. Pointer types are not relevant here as well, so I applied the same approach as in sycl/test/check_device_code/kernel_arguments_as.cpp test.
• sycl/test/extensions/sub_group_as.cpp - this is an example of the test where further test updates are required. Type are not important for some of the checks (e.g. spirv_GenericCastToPtrExplicit_ToLocal doesn't change the type, just an address space), but there are examples where LLVM IR must have type information somewhere. I'm hesitant to update more tests like this because not all changes are in place to make the final test fix. We need to figure out what is the plan for ESIMD/matrix types and decide whether we go with [Attr] Extend uses of elementtype attribute #6397 before updating LLVM IR check. Let me know your thoughts on this.

[jchlanda]

@jchlanda, do you think that this failure on NVIDIA GPU https://github.com/intel/llvm/runs/7391421134?check_suite_focus=true is related to this change? This change is not expected to change the default behavior.

It doesn't look related, even if we ignore the assert messages which are a bit flaky, upon hitting the assert PI runtime should generate error 710:

PI CUDA ERROR:
        Value:           710
        Name:            CUDA_ERROR_ASSERT

but this one reports -999 - Unknown PI error (https://github.com/intel/llvm/runs/7391421134?check_suite_focus=true#step:8:1168)

[jcranmer-intel]

We need to figure out what is the plan for ESIMD/matrix types and decide whether we go with #6397 before updating LLVM IR check. Let me know your thoughts on this.

My current thinking for the matrix type situation in the LLVM-SPIRV translator is a short-term fix that has the clang frontend just lower to the correct name of the LLVM struct type in the first place (obviating the need to change in the translator itself), with a longer-term fix of adding the LLVM opaque type support I proposed in the RFC.

[bader]

@jchlanda, do you think that this failure on NVIDIA GPU https://github.com/intel/llvm/runs/7391421134?check_suite_focus=true is related to this change? This change is not expected to change the default behavior.

It doesn't look related, even if we ignore the assert messages which are a bit flaky, upon hitting the assert PI runtime should generate error 710:

PI CUDA ERROR:
        Value:           710
        Name:            CUDA_ERROR_ASSERT

but this one reports -999 - Unknown PI error (https://github.com/intel/llvm/runs/7391421134?check_suite_focus=true#step:8:1168)

I can't reproduce this issue locally, so I restarted CI jobs and it failed again, but with the different error:

error: command failed with exit status: 255

https://github.com/intel/llvm/runs/7450227286?check_suite_focus=true

These look like flaky llvm-test-suite issues not caused by this patch. I'll file a report to investigate them and merge this PR.