Add unions for ICD2 tags in cl_icd_dispatch

[karolherbst]

This requires C11/C++11 as support for anonymous unions got added there.

I need this, because otherwise rustc doesn't let me use a constant in place of a function pointer. But I think being more explicit here would make the code a bit cleaner for loader implementors.

[karolherbst]

rust bug for reference: rust-lang/rust#141660

[Kerilk]

I love the suggestion, but there is another instance of anonymous union where guards exist:

OpenCL-Headers/CL/cl.h

Lines 137 to 167 in 1ce4f1c

	#if defined(__GNUC__)
	__extension__ /* Prevents warnings about anonymous union in -pedantic builds */
	#endif
	#if defined(_MSC_VER) && !defined(__STDC__)
	#pragma warning( push )
	#pragma warning( disable : 4201 ) /* Prevents warning about nameless struct/union in /W4 builds */
	#endif
	#ifdef __clang__
	#pragma clang diagnostic push
	#pragma clang diagnostic ignored "-Wc11-extensions" /* Prevents warning about nameless union being C11 extension*/
	#endif
	#if defined(_MSC_VER) && defined(__STDC__)
	/* Anonymous unions are not supported in /Za builds */
	#else
	union {
	#endif
	#endif
	cl_mem buffer;
	#ifdef CL_VERSION_2_0
	#if defined(_MSC_VER) && defined(__STDC__)
	/* Anonymous unions are not supported in /Za builds */
	#else
	cl_mem mem_object;
	};
	#endif
	#if defined(_MSC_VER) && !defined(__STDC__)
	#pragma warning( pop )
	#endif
	#ifdef __clang__
	#pragma clang diagnostic pop
	#endif

I think we would need something similar to enable this.

[karolherbst]

mhhh.. not a big fan of replicating the same code block, but maybe this can be moved into a macro or some sorts. Will play around with it later.

[karolherbst]

Turns out there were already a bit of helper macros in cl_platform for anonymous structs we can also reuse for unions. Cleaned up code in cl.h and used the more streamlined stuff in cl_icd.h as well

[Kerilk]

Looking good, and generally an improvement I think.
The definition of the khr icd extension will have to be amended as well to reflect the change:
https://github.com/KhronosGroup/OpenCL-Docs/blob/bc6abc46166ff046ee17a2d7a42bb28e7e7fb7b7/api/cl_khr_icd.asciidoc?plain=1#L75-L78
https://github.com/KhronosGroup/OpenCL-Docs/blob/bc6abc46166ff046ee17a2d7a42bb28e7e7fb7b7/api/cl_khr_icd.asciidoc?plain=1#L244-L250

[bashbaug]

LGTM

I'm a little nervous about losing some compiler-specific warning suppression, but if the macros are working well for the other anonymous structs, they should work fine here, too.

[karolherbst]

LGTM

I'm a little nervous about losing some compiler-specific warning suppression, but if the macros are working well for the other anonymous structs, they should work fine here, too.

yeah.. I've checked clang docs and it uses __extension__ (__CL_ANON_STRUCT__ expands to it) similar to gcc, so I don't think those changes are changing anything in this regards.

The only part I'm a bit more concerned about is MSVC and the pop/push of the warning is a bit global. Maybe they have a similar mechanism like __extension__ we could use instead?

[Kerilk]

We can rerun the CI in KhronosGroup/OpenCL-ICD-Loader#252 once this one is merged as an additional check.
(Works fine on my machine.™)

[bashbaug]

Merging as discussed in the June 17th teleconference.