Some targets require library functions to implement floating-point operations. Some of those targets don't always round exactly correctly. Only round-to-nearest mode is supported by the compiler and the RTS library. When the rounding isn't done correctly, the floating-point operation may round in the wrong direction, causing a 1 ULP error in the result. This bug does not apply to targets that use hardware to implement floating-point fundamental operations (add, subtract, multiply) in hardware.