[SYCL] Fix function attribute merging on template declarations

[0x12CC]

#20419 implements attribute merging for add_ir_attributes_* attributes. This commit fixes an issue that occurs when merging attributes on function template declarations.

Attribute merging is deferred until template instantiation if an attribute contains dependent expressions. Previously, we only checked if the existing attribute was dependent but not if the new attribute was dependent. This would cause the merge to fail but the new attribute would still be dropped from the AST. This commit ensures both attributes are checked and new dependent attributes do not disappear.

[0x12CC]

@intel/dpcpp-cfe-reviewers, friendly ping.

[Fznamznon]

Makes sense, just a couple of questions.

https://github.com/intel/llvm/pull/20419/changes also brings changes for two other attributes - SYCLAddIRAttributesKernelParameterAttr and SYCLAddIRAttributesGlobalVariableAttr . Do they have the same problem?

[0x12CC]

Thank you, @Fznamznon,

https://github.com/intel/llvm/pull/20419/changes also brings changes for two other attributes - SYCLAddIRAttributesKernelParameterAttr and SYCLAddIRAttributesGlobalVariableAttr . Do they have the same problem?

I think so. The two other functions with similar logic are SemaSYCL::addSYCLAddIRAttributesKernelParameterAttr and SemaSYCL::addSYCLAddIRAttributesGlobalVariableAttr.

I fixed SemaSYCL::addSYCLAddIRAttributesFunctionAttr since these attributes are used for adding properties to free function kernels. I didn't change the other two since I couldn't find any way to write SYCL code that caused the merging issue. For the add_ir_attributes_kernel_parameter, the entire compile-time property list goes into one attribute. Similarly, for add_ir_attributes_global_variable, a single attribute is used to annotate all the device global metadata. Free functions are unique since the user can directly control the number of attributes in the AST using the SYCL_EXT_ONEAPI_FUNCTION_PROPERTY macro.

[0x12CC]

The failing end-to-end tests are unrelated to this change:

1. Regression/queue_submit.cpp, reported in Regression/queue_submit.cpp sporadic fail on PVC #21095.
2. Basic/out_of_order_queue_status_khr_empty.cpp, reported in queue::submit doesn't Enqueue Kernels before It Returns/after Host Tasks are Complete under OpenCL Backend sometimes #20837.

@intel/llvm-gatekeepers, this PR should be ready to merge.