Not sure about the policy on target intrinsics in InstructionSimplify
since there don't seem to be any others. However we do in
ConstantFolding and instcombine already. Some of the existing AMDGPU
intrinsic simplifications are in instcombine that really belong here
since they don't introduce new instructions.
I also noticed we seem to now be interpreting strictfp attributes on intrinsic call sites,
so try to handle that.