[Driver][SYCL] Use existing option for device-only and some cleanup

[mdtoguchi]

Update -fsycl-device-only usage to be an alias to --offload-device-only. This requires an additional unique check for SYCL enabling due to the expectation that -fsycl-device-only also implies -fsycl.

Update associated tests that use -fsycl-device-only -emit-llvm. Our default is already bitcode output, fixing this up aligns better with community expectations when using --offload-device-only.

Also do some variable cleanup to better be in sync with community variable names. And use a more common diagnostic for invalid options with -fsycl.

[bader]

Update associated tests that use -fsycl-device-only -emit-llvm. Our default is already bitcode output, fixing this up aligns better with community expectations when using --offload-device-only.

I find this behavior a bit confusing. My understanding is that by --offload-device-only should produce the device code in "executable" format as standard compiler invocation does for the host compilation.
SPIR target has a limitation. We don't have a native back-end for that target, but we can use SPIR-V translator as SPIR backend compiler to create a SPIR-V format. In that flow, using --offload-device-only --emit-llvm is reasonable.
This could be discussed and addressed separately.

[mdtoguchi]

Update associated tests that use -fsycl-device-only -emit-llvm. Our default is already bitcode output, fixing this up aligns better with community expectations when using --offload-device-only.

I find this behavior a bit confusing. My understanding is that by --offload-device-only should produce the device code in "executable" format as standard compiler invocation does for the host compilation. SPIR target has a limitation. We don't have a native back-end for that target, but we can use SPIR-V translator as SPIR backend compiler to create a SPIR-V format. In that flow, using --offload-device-only --emit-llvm is reasonable. This could be discussed and addressed separately.

To support the ability to emit SPIR-V using the translator, we have -fsycl-device-only -fsycl-device-obj=spirv. Our default is -fsycl-device-obj=llvmir

[bader]

To support the ability to emit SPIR-V using the translator, we have -fsycl-device-only -fsycl-device-obj=spirv. Our default is -fsycl-device-obj=llvmir

Right. The primary reason for that is to support Relocatable Device Code (-fgpu-rdc) by default. We keep device code in LLVM format to be able to link translation units with device code.

My point is that community compiler uses no-gpu-rdc by default (see https://clang.llvm.org/docs/HIPSupport.html#device-code-compilation). Potentially SPIR-V is also "linkable" file format, but we don't have a production quality SPIR-V linker tool.

We can also consider switching to no-gpu-rdc for SYCL mode by the default as well. This should be a separate discussion, so, please, do not consider my comments as blocking.

[mdtoguchi]

@intel/llvm-gatekeepers, this should be ready for merge, thanks!