Branch prediction for the 32-bit implementation and a new optimized 64-bit implementation.
libm is now somewhat integrated with gcc's -ffinite-math-only option and lots of the wrapper functions have been optimized.