[Dwarf-Discuss] Register kinds
Richard.Relph at amd.com
Mon Nov 1 11:43:45 PDT 2010
> Relph, Richard wrote:
> > I appreciate the warning. And I wish we could debug at 'machine
> level'. But that decision isn't really up for review.
> > What areas do you think might be particularly problematic?
> Offhand, I can think that there might
> be problems in CIE/FDE generation, which
> provides info to walk the stack.
Our current hope is that the compiler will do this completely.
GPUs (at least AMD GPUs) prefer fully inlined programs. OpenCL programs (kernels) are usually the inner most computational elements of larger programs. The larger programs are run on the host, not on the GPU. That said, OpenCL does support standard C function call syntax, but not recursion. Therefore we have our compilation stack configured to in-line everything. So there's no "calling convention" to speak of, never mind frame pointers and the like. There's no interrupts or context switching. It's a very constrained execution environment.
So we're expecting the compiler to track which IL instructions are replicated as a result of inlining (and unrolling, where that occurs) and pass that information to the debugger via DWARF.
> If anything else occurs to me, I'll let you
Thanks. Our environment is 'different' in many respects. We think DWARF has the flexibility to do what we want, but none of us are 100% certain it will all work in the end.
More information about the Dwarf-Discuss