[Openmp-dev] clang: Emit Target Functions when #pragma omp declare target

[Daniel Schürmann via Openmp-dev]

Hello!

I wanted to ask, if this behavior is on purpose or if I am missing 
something:

To me it seems that using "#pragma omp declare target" has no effect on 
the code generation.
The specification states that variables and functions are mapped to the 
device.
In my understanding, this means that device code should be emitted.

However, CGOpenMPRuntime::emitTargetGlobal() does only search for target 
#pragmas and doesn't check for OMPDeclareTargetDeclAttr. This check is 
done in ASTContext::DeclMustBeEmitted(), but it seems, this can never 
happen. The reason is that CodeGenModule::EmitGlobal() returns early 
after calling CGOpenMPRuntime::emitTargetGlobal(). I hope, this is 
understandable.

It might be that this problem never showed up when the device is x86 as 
well. But I don't understand how Nvidia can compile e.g. lulesh like this...

My question is if the intended way would be to change emitTargetGlobal() 
to also emit functions/variables with OMPDeclareTargetDeclAttr.

Thanks in advance,
Daniel