[flang] [compiler-rt] [clang] [llvm] [mlir] [clang-tools-extra] [lldb] [libc] [libcxx] [AArch64] add intrinsic to generate a bfi instruction (PR #79672)

[David Green via cfe-commits]

davemgreen wrote:

I see. The issue is that the opposite is often true as well - if we add a target specific intrinsic for this then, whilst we get a single instruction being emitted, we don't see all the other optimizations that the compiler can and should be performing.

Things like constant folding, combining into other instructions, known-bits analysis or any form of vectorization will all be blocked by the intrinsic. It can take quite some work to add all those features in (if they are possible), and without them can potentially lead to worse results. Plus more things to maintain.

BFI isn't a trivial instructions to match as it involves certain masks and shifts. There might certainly be advantages to having an intrinsic. I would like to try and see what the problems would be with generated code using normal operations first though, if we can. If there are optimizations we can make based on the existing code then that would help in all cases (c, mlir, rust, etc), not just frontends that are producing the intrinsics.

https://github.com/llvm/llvm-project/pull/79672