[SYCL] Fix assertion for sycl keyword using without -fsycl option.

[jyu2-git]

This is to fix assertion when clang compile C++ head file without -fsycl

The problem is when __pipe (one of sycl keyworld) used without
-fsycl
option, the __pipe should not be keyword and keyword status should be
KS_Disable. But it is wrongly set to KS_Future.

The root problem is the KEYSYCL is add to KEYALLCXX:

KEYALLCXX = KEYSYCL | KEYCXX | KEYCXX11 | KEYCXX2A

It seems, we want KEYSYCL to be part of feature of C++. But which C++
should belong: C++11 or C++2a? (for example KEYCX2A... vs concept)

For now we should just use -fsycl option to control this.

Signed-off-by: Yu jennifer.yu@intel.com

[erichkeane]

I looked into this as well, this is correct.

[bader]

__constant is not SYCL keyword after #358.

[erichkeane]

This problem isn't limited to __constant. It'll happen with anything defined with KEYSYCL. __read_only, __write_only, and __pipe have the same problem. @jyu2-git you likely need to just update the test to use one of those other keywords instead.

[jyu2-git]

I don't know if I have updated source. But it seems after 358, we don't defined any KEYSYCL keyword in TokenKinds.def. If that is the case, we need remove KEYSYCL for all use. If not, we still need to remove KEYSYCL form KEYALLCXX

[bader]

I think __pipe should still be a keyword.

[bader]

Could you also add [SYCL] tag to the commit message title, please?

[jyu2-git]

Thanks I will add by end of day. The machine I am using currently used by other. So can not do now.

[keryell]

None of these are actual SYCL keywords. They are just implementation details.
I have the feeling that most of them could be just replaced by some __attribute__((annotate("sycl::pipe"))) or whatever.

[jyu2-git]

None of these are actual SYCL keywords. They are just implementation details.
I have the feeling that most of them could be just replaced by some __attribute__((annotate("sycl::pipe"))) or whatever.

Decision is needed from sycl team.

[agozillon]

None of these are actual SYCL keywords. They are just implementation details.
I have the feeling that most of them could be just replaced by some __attribute__((annotate("sycl::pipe"))) or whatever.

Decision is needed from sycl team.

I think (but could be wrong) @keryell was just giving an opinion for some interesting future direction as I imagine it's outside of the scope of this pull request probably. As I think the annotate attribute might need some tweaking to be fully usable in all contexts; at least from my experience with it anyway. And I imagine some of the code gen/other components relating to the __pipe keyword would have to change.

But having annotate working nicely would be pretty useful for passing arbitrary data down to backends/opt passes and could be useful to make some generic way to express SYCL extensions/attributes. So some discussion on it from the SYCL team may be interesting!

[jyu2-git]

Thanks @agozillon. Should we get this in for now? Thanks. Jennifer

[bader]

LGTM.
Minor comment.

[bader]

Please, update the description. __constant is not a SYCL keyword, so it's a bad example.

Also it might be useful to test that clang emits diagnostic if we try to use __pipe, __read_only, etc. as keywords.

[jyu2-git]

Thanks all, update comment