[SYCL] Add support for __registered_kernels__

[premanandrao]

This change adds support for a new attribute
__sycl_detail::__registered_kernels__, which appears at translation
unit scope. The parameter for this attribute is a list of pairs like:

[[__sycl_detail__::__registered_kernels__(
  {"foo", foo},
  {"(void(*)(int, int*))iota", (void(*)(int, int*))iota},
  {"kernel<float>", kernel<float>}
)]];

The first element in each pair is a string, and the second element is a
constant expressiton for a pointer to a SYCL free function kernel.

The change creates the kernel's wrapper function and generates
module-level metadata of the form:

!sycl_registered_kernels = !{!0, !1}
!0 = !{!"foo", !"mangled-name-of-wrapper-for-foo"}
!1 = !{!"kernel<float>", !"mangled-name-of-wrapper-for-kernel<float>"}

where the first element in the pair of strings, is the first element
of the pair in __registered_kernels__ and the second element is the
mangled named of the wrapper corresponding to the free function.

[jopperm]

Not a real review, but I played with this locally and it LGTM for the SYCL-RTC use-case, thanks!

[jopperm]

Minor documentation nit: The PR description hints at a named MD !sycl_registered_kernels being emitted, whereas the implementation actually uses a module flag.

[premanandrao]

Thank you for pointing this out. It was a thinko on my part.
Please let me know if I should change it - I am not really sure which has advantages over the other.

[jopperm]

Please let me know if I should change it - I am not really sure which has advantages over the other.

Will you be tackling the transformation of this metadata to a property set as well? Maybe this can inform the decision. Anecdotally, the ComputeModuleRuntimeInfo utility (in SYCLLowerIR) doesn't look at module flags right now, only named MD.

@gmlueck Do you have a strong(er) opinion here?

[premanandrao]

Will you be tackling the transformation of this metadata to a property set as well?

No, that won't be me.

I will wait for Greg to confirm, but it looks like it should be a named metadata then.

[gmlueck]

From what I can tell, module flags are a special kind of named MD which has clear semantics for how the named MD is merged together when IR modules are merged. I think the merging semantics are not important for sycl_registered_kernels, though, because the kernel compiler will never merge together multiple IR modules. We will always compile the IR straight to SPIR-V (or PTX, etc.)

Therefore, I think using named MD would be OK. On the other hand, I can't think of any reason why it would be problematic to use module flags either. If it's easier to consume named MD from sycl-post-link, I think we could go with named MD, but I don't have a strong preference.

@cperkinsintel I think you are coordinating work for the kernel compiler. Who will be responsible for converting sycl_registered_kernels into a property set?

[premanandrao]

Therefore, I think using named MD would be OK.

I changed it to use named MD now. It really was a mistake on my part initially to use module flags, but then I got curious as to how it was different. Your explanation that it is a special case of named MD makes perfect sense. Thanks!

[jopperm]

LGT-not-super-familiar-with-the-frontend-me, thanks!

[premanandrao]

@intel/dpcpp-cfe-reviewers, could you please review?

[elizabethandrews]

@Fznamznon can you take a look at this? I was unable to today and will be out for part of the day tomorrow. I don't want to hold this up.