[SYCL] Remove sycl_ext_oneapi_sub_group in preview-breaking mode

[softwaresat]

Guards the deprecated SYCL_INTEL_sub_group feature-test macro from preview-breaking builds.

Documentation is marked for removal or relocation during the next breaking-changes window.

This closes #22581.

[KornevNikita]

This descriptor and all related code should be removed:

llvm/sycl/include/sycl/info/device.hpp

Line 466 in 0f637d2

sub_group_independent_forward_progress

This header should be removed:
https://github.com/intel/llvm/blob/sycl/sycl/include/sycl/ext/oneapi/sub_group.hpp

I guess the rest of code that should be removed will be easy to find if you disable this header and try to compile the project.

[gmlueck]

This deprecated extension also defines a couple of C++ attributes:

• [[intel::sub_group_size(S)]]
• [[intel::named_sub_group_size(NAME)]]

These are implemented in the C++ front-end (not the headers). I think we should deprecate these also. I don't know if there is a way to remove them from the CFE in -fpreview-breaking-changes. If you want to assign this part of the task to the CFE team, that's fine. But, we should create a tracker in that case.

Note that we have functionality similar to [[intel::named_sub_group_size(NAME)]] proposed in sycl_ext_oneapi_named_sub_group_sizes, but that proposal uses properties (not attributes). Therefore, even if we implement that extension, I don't think we would want that attribute.

[KornevNikita]

@softwaresat hi, do you plan to continue working on this?

[softwaresat]

@KornevNikita Sorry, was on a long vacation and just started classes. I'll have it completed before the end of the week. I'd love to start on other issues afterwards as well.

[KornevNikita]

No need to remove this.

[softwaresat]

Updated code

[KornevNikita]

Please mark this and all other removals with #ifndef __INTEL_PREVIEW_BREAKING_CHANGES. I guess it won't break ABI actually, but I'd like to keep all the oneapi::sub_group related stuff until the next major release.

[softwaresat]

Fixed all the removals and kept the other sub_group stuff.

[KornevNikita]

Why do we need this?

[softwaresat]

I added this to emit a deprecation warning for [[intel::named_sub_group_size(NAME)]] while keeping the attribute source-compatible. Should I remove it?

[KornevNikita]

I think we should keep it until the spec is not removed

[softwaresat]

Added it back

[elizabethandrews]

FE change LGTM

[elizabethandrews]

Oh wait I was too quick to approve. I see some of the other deprecated attributes using diagnoseDeprecatedAttribute. I don't recall what this does. Can you clarify why this wasn't used?

[softwaresat]

Oh wait I was too quick to approve. I see some of the other deprecated attributes using diagnoseDeprecatedAttribute. I don't recall what this does. Can you clarify why this wasn't used?

diagnoseDeprecatedAttribute was not used because intel::named_sub_group_size has no replacement attribute spelling. The proposed replacement is a kernel property, so I emit only warn_attribute_spelling_deprecated and intentionally provide no fix-it.

[tahonermann]

I think the diagnostic could be more helpful; see inline comments. The PR title and summary could use updates to clarify what change is being made; neither of those currently indicate this change has anything to do with attributes.

[tahonermann]

The warn_attribute_spelling_deprecated diagnostic doesn't seem appropriate here. That diagnostic is currently used when only some spellings of the attribute are deprecated. For example, all [[cl::XXX]] attribute spellings have been deprecated in favor of their equivalent [[sycl::XXX]] attribute spellings. The actual message produced for the warn_attribute_spelling_deprecated is reasonable for this new deprecation, but I think it would be better to tell users what they should use instead. See the warn_atl_uuid_deprecated, warn_vector_mode_deprecated, and warn_amdgpu_num_reg_attr_deprecated diagnostics as good examples that provide the user more direction.