[cfe-dev] OpenCL support

[Peter Collingbourne]

On Fri, Dec 10, 2010 at 11:18:22PM -0000, Mike Gist wrote:
> -----Original Message-----
> From: cfe-dev-bounces at cs.uiuc.edu on behalf of Anton Lokhmotov
> Sent: Fri 10/12/2010 18:21
> To: cfe-dev at cs.uiuc.edu
> Subject: [cfe-dev] OpenCL support
>  
> > 1) For kernel functions (with the __kernel qualifier), we insert named
> > metadata nodes into LLVM-IR.  Other options we have evaluated included:
> > - Shipping metadata separately from bitcode.  This would be hard to
> > standardise on and maintain.
> > - Using imaginative name mangling.  This would be pretty horrible too and
> > easy to misuse.
> >
> > The drawback of course is that metadata can be "silently dropped".  However,
> > we are not aware of any transforms that would do this.  Perhaps the
> > community should consider defining "sticky metadata"?
> 
> You could also consider placing all kernel functions in a 'kernel' section, or
> adding a function attribute for kernels.
> 
> Inserting metadata feels a little dirty to me, it shouldn't be used if it would change
> the semantics of the program (which it would in this case). Some kind of 'sticky' metadata
> would indeed get around 'dropping', however I don't believe (I may be wrong) that metadata was
> designed for this purpose. I'm not sure that trying to fit it to this purpose is a good idea.

The current PTX backend uses a target-specific calling convention for
kernel functions (CallingConv::PTX_Kernel).   Another approach we may
consider is to standardise a kernel calling convention across backends.

Thanks,
-- 
Peter