This allows the LibCallSimplifier to consistently produce only intrinsics
when optimizing an existing exp2 intrinsic.
For targets that do not support FLDEXP natively (i.e. most of them),
legalization can Expand it to the open coded SINT_TO_FP + FEXP2 + FMUL
or convert it to the corresponding LibCall. This has been turned on
This fixes a bug where instcombine incorrectly generated a library call
for the AMDGPU target.