I think we can save the MRI argument from these since it's in
GISelKnownBits already, but currently not accessible.
Implementation deferred to avoid dependency on other patches.
This TODO hasn't quite been handled. It's also intended to cover the case where you have target specific instructions (e.g. AArch64::ADDWrr), pseudo-instructions, or target-specific generic opcodes.
To cover this case we also need to call TL.computeNumSignBitsForTargetInstr() for the default: case
FirstAnswer is 1 so this if-statement is already inside the std::max
Why update a variable that and return that rather than just returning the answer as per the other code paths?
Handle the fallback to computeKnownBits, copied directly from the DAG version. This should really get a test, but I don't think there's any operation that is implemented that should hit this path