Floating-Point Modes

Switching the floating-point precision mode on x86 inserts an additional instruction into the code and also causes stalls in the processor pipeline. For Java methods using 32-bit and 64-bit floating-point operations, switching precision is inevitable. However, our experiments have shown that while running the benchmarks, only one method in the Java class libraries and one method in the benchmark programs (mpegaudio) use two precision modes within a single method. Even in these cases the less frequent precision is used at most two times. To take advantage of this fact, we choose a default precision mode for each method and set the precision mode to the default at the beginning of each method that uses floating-point operations. Thus, default precision operations require no additional instructions, while the precision mode needs to be set and reset around the occasional non-default precision operation. Our calling convention considers the precision mode to be a preserved value, so it must be reset at the end of the method.

Because non-default precision operations are rare, this strategy significantly decreases the number of switches required and thus increases performance. Precisely analyzing the control-flow and optimizing the number of switches in a given method would help, but may be too expensive for just-in-time compilation.