Optimization of Reciprocals and Square Roots on the i860 Microprocessor

Reciprocal and reciprocal square root operations are partially supported by the i860 floating point unit, whereas square roots are not. We point out the reasons for this, and its consequences for the optimization of code involving many reciprocal square roots, such as many-body simulations involving Coulomb-like potentials. We conclude that code which can be optimized to explicitly combine reciprocals and square roots in the form of reciprocal square roots can attain significantly higher performance, and that assembly language coding of such operations can make the greatest use of the hardware by calculating only to the accuracy required, which may be less than single precision.