In Defense of NIR

For context:

>>NIR (pronounced “ner”) is a new IR (internal representation) for the Mesa shader compiler that will sit between the old IR (GLSL IR) and back-end compilers.

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.

Near infrared??

NIR is lean C89 with bits of c99/c11 and that alone is enough to favor it instead of anything else requiring a massive compiler able to understand without a flinch the insanely (usage of this word is fair) complex c++ syntax (c++11 c++14 c++17 c++20, etc). As an AMD hardware user, I am sad that Lisa L. did not get the memo about c++ toxicity, neither Gabe from Valve, because even if ACO (the glue between AMD GPU isa and NIR) is far less worse than llvm, it still requires such c++ compiler. At least it is moving in the "right"(TM) direction. I wish ACO was like NIR, lean C89 with bits of c99/c11, that to keep the door open of reasonably working alternative C compilers.