Signed mismatch for intrinsics?

[beamspease]

Hi,

I noticed that the nvptx repo was using wrapping_add, which was curious. In looking into it, I noticed that the type of the intrinsics seems to differ from what the CUDA guide states. Is this intentional?

http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#built-in-variables

[japaric]

The LLVM definition use i32 so I'm sticking to that. If you are using C/C++ (and CUDA is C/C++ + some language extensions) I suppose it doesn't matter what LLVM (the backend) uses because the compiler will implicitly cast between integer types.

[beamspease]

That makes sense. I'd assume that we're nowhere near 2,147,483,647 blocks so it probably doesn't matter right now for signed/unsigned.