[llvm] [AMDGPU] prevent shrinking udiv/urem if operands exceed smax_bitwidth (PR #116733)

[Jay Foad via llvm-commits]

https://github.com/jayfoad commented:

I think it would make more sense for this function to have separate paths for signed (based on computeNumSignBits) and unsigned (based on leading zeros, from computeKnownBits). The way it is written now, the unsigned case will potentially call _both_ computeKnownBits and computeNumSignBits on both arguments, which seems like overkill.

https://github.com/llvm/llvm-project/pull/116733