Do not diagnose unsupported __float128 type if it's not used in device code

[Fznamznon]

Now LLORG clang emits the following diagnostics from standard headers:

/usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0/bits/std_abs.h:102:7: error: __float128 is not supported on this target
  abs(__float128 __x)
      ^
/usr/lib/gcc/x86_64-linux-gnu/7.3.0/../../../../include/c++/7.3.0/bits/std_abs.h:101:3: error: __float128 is not supported on this target
  __float128

and so on.
We disabled such diagnostics for device by the following change:
#181 (comment)
Experiments show that we don't emit any diagnostics if __float128 is used in device code. IMO, that is not fully correct and this diagnostic should be delayed.

[Fznamznon]

Okay, we can start with simple disabling of diagnostic, like is was done for CUDA for _Float16 (see https://reviews.llvm.org/D57369). Then improve it once we implement deferred diagnostics for SYCL.
@bader , do you have any objections?

[bader]

No.

Okay, we can start with simple disabling of diagnostic, like is was done for CUDA for _Float16 (see https://reviews.llvm.org/D57369). Then improve it once we implement deferred diagnostics for SYCL.

How difficult is to "implement deferred diagnostics for SYCL"?

[Fznamznon]

No.

Okay, we can start with simple disabling of diagnostic, like is was done for CUDA for _Float16 (see https://reviews.llvm.org/D57369). Then improve it once we implement deferred diagnostics for SYCL.

How difficult is to "implement deferred diagnostics for SYCL"?

I'll take a look. I need some time to understand how it's supposed to work.

[Fznamznon]

Okay, I've learned something new about deferred diagnostics. Actually, we already have almost working implementation of this feature. But we still emit diagnostics in SemaSYCL from MarkDeviceFunction Visitor's run, because deferred diagnostics aren't fully working now.
We still do it, because we don't build device call graph for diagnostics properly. FunctionDecl which represents operator () of lambda which represents SYCL kernel object just doesn't get into the DeviceCallGraph and deferred diagnostics aren't emitted.
The problem is already brilliantly described in details by @erichkeane in https://github.com/intel/llvm/pull/654/files#r327656088 .
All we need is fix this problem in a proper way and we get working deferred diagnostics (I just tried one of Erich's suggestions and changed using of getCurFunctionDecl() to getCurLexicalContext() , but as far as I understand that works not for all cases).
I think I can try to do more experiments, fix our deferred diagnostics system and upstream it. Then we won't need to upstream patches which silent any diagnostics.

[Fznamznon]

Fix and cleanups for deferred diagnostics is here #971 . Deferring of diagnostics about unsupported __float128 type added as well.

[Fznamznon]

Submitted a change to disable diagnostic https://reviews.llvm.org/D74387

[Fznamznon]

https://reviews.llvm.org/D74387 is submitted.