This time it's different. The "architecture" (instruction set) really doesn't matter, I agree with you. What matters is that the new units are dual issue, so the HLSL shaders that get compiled must use a special compilation path that optimizes, merges and combines instructions and data in a way so that they become aligned with how the architecture performs best. This is a difficult task that often requires hand-optimized code.
Random paper that I pulled up to give you an idea:
http://www.ai.mit.edu/projects/aries/Documents/vliw.pdf
I'm quite familiar VLIW. AMD's family of architectures prior to GCN, TeraScale, used several variants of this type of design. Shader compilers have no issue with this, they are tailored to optimize the shader programs from IR to native instructions specific to each GPU architecture. This isn't something new, but PR people can absolutely make it sound fancy.
AMD (and the others) certainly have the ability to override shader programs with their own, but this is rarely done for a whole host of reasons; including the difficulty as most are in IR, the risk of introducing bugs or glitches, the sheer number of shader programs (could be hundreds in a single game), game patches overriding this effort, and the difficulty for developers to debug the game with these "optimizations". I say "optimizations" in quotes, because hand-optimized shaders are usually about removing complexity without sacrificing too much fidelity to notice, unlike the shader compiler which retains the original logic.
AMD certainly have a huge repository of shader programs, which is what they use to develop their shader compilers.
There are games which do compile shaders during rendering. I haven't checked why, but it's certainly conceivable that some "generate" countless permutations of shaders based on object properties to save logic inside the shaders.
And the complexity and amount of shaders in games are going up, so inconceivable that AMD will hand-optimize most of these by hand, it would probably be hundreds of thousands, for each GPU architecture. it would still be a lot if you limited it to only the top games.
