The PowerISA does not have a roundeven instruction, like x86-64-v2
or aarch64, however it does have support for round (frin, ISA 2.03)
and clz (cnttzd).  Thus the __builtin_roundeven inline version is
faster for recent POWER chips.

Since the implementations do not include math.h, use __builtin_round
instead of round (so compiler can lower to the instruction if
available).

Also optimize the result adjustment to use copysign instead of
floating point comparison (it shows slight better generated code).