I've done a good bit of CUDA and OpenCL. I haven't dug into this GCC implementation, but my guess is that you're banned from using any features that require so

I've done a good bit of CUDA and OpenCL. I haven't dug into this GCC implementation, but my guess is that you're banned from using any features that require sophisticated runtime support (like exceptions or RTTI). Things like templates, which are a compile-time mechanism, should work fine. Also, you still have the usual nasty issues of the GPU having a separate memory space from the CPU. This is just a code generation backend; not some transparent GPU system model for C/C++.