cfe-dev Digest, Vol 55, Issue 4

Hi Anton,

I just sent out a proposal for "module flags metadata". It's similar to what you're proposing here, but more general. If you think it's a benefit to what you would like to do or needs improvements, please send feedback. :slight_smile:

-bw

Hi Bill,
I went over your proposal, and it looks very clear and flexible; however I couldn't find an easy way to use it to represent all the
OpenCL kernel and kernel argument specific metadata.
I think it can store module level information, but for most of the interesting OpenCL metadata (which is function/kernel level) it might
only act as entry point, replacing the "opencl.kernels" named metadata node. Trying to store all the kernel metadata in separate nodes
would make the metadata lookup very inefficient and difficult. I think Anton's proposal below is more suitable for this purpose.

Anton/Bill,
If you have ideas how to use Bill's proposal for OpenCL metadata, I think there should be an OpenCL specific proposal with examples
to show it's feasible.

Thanks
   Guy

Hi Guy,

My initial response is that module flags is a more formalized way to implement the OpenCL data. Whether you add the opencl.kernels metadata as one element or as separate elements to the flags. It's formalized because it will be documented in the LangRef doc. Second, it has a way to merge this data during LTO.

So it's not completely different from Anton's suggestion. Just more formalized. :slight_smile:

-bw