Is any information on this public? I've believed for a while that domain specific optimisations are the right way to go but haven't found many examples of it in

Is any information on this public? I've believed for a while that domain specific optimisations are the right way to go but haven't found many examples of it in practice. LLVM is largely (at least originally) tuned for clang's output which tends towards simple IR that LLVM will clean up later, with a fair bias towards making numerical benchmarks run faster.