Well articulated, even if "design an appropriate IR for your target" is a bit of an old hat to more experienced compiler devs, it's good reading to point people

Well articulated, even if "design an appropriate IR for your target" is a bit of an old hat to more experienced compiler devs, it's good reading to point people to. I particularly liked the bit about structured control flow for SIMT architectures. Integrating LLVM even in a compiler targeting CPUs (in my experience) still has plenty of pain points, much less a GPU-oriented one, much less in 2014! There's still plenty of rich design space to explore as well, even if it's ultimately a lot of work.