Funny that you should say this... Every 686+ chip produced by Intel and AMD today has a compiler in hardware. It takes for its input x86 instructions, compiles them to RISC operations, and peephole-optimizes the hell out of that. It is very unlikely that the shader compiler in this gadget is within one order of magnitude as complex as what is in every i5. Maybe not even two.
I don't know if those x86->RISC compilers are in rewritable microcode, but that's the way I would bet. That compiler's output has total access to all memory and devices. It might even run when you install new microcode, and doctor that, Ken Thompson-style, to have the same back door as the old.
There are probably good reasons for China to have their own, independent CPU designs, even at a cost of reduced performance. And for the rest of us, too, maybe.