/MATH_LIBRARY

  Controls selection of math library routines to provide accurate or
  fast results for VSI Fortran intrinsic functions, such as SQRT.

  The default is /MATH_LIBRARY=ACCURATE.

  This qualifier applies only to IEEE data types (when you specify
  /FLOAT=IEEE_FLOAT).

  You can only specify one of the qualifier options.

  ACCURATE
    Causes the compiler to produce the very accurate results and
    error checking expected of quality compiler products.  It uses
    the standard set of math library routines for the applicable
    intrinsics.

    The standard math library routines are designed to obtain very
    accurate "near correctly rounded" results and provide the
    robustness needed to check for IEEE exceptional argument values,
    rather than achieve the fastest possible run-time execution
    speed.  Using /MATH_LIBRARY=ACCURATE allows user control of
    arithmetic exception handling with the /IEEE_MODE qualifier.

  FAST
    Causes the compiler to use versions of certain math library
    routines that perform faster computations than the standard, more
    accurate math library routines, but with slightly less fractional
    accuracy and less reliable arithmetic exception handling.  Using
    /MATH_LIBRARY=FAST allows certain math library functions to get
    significant performance improvements when the applicable
    intrinsic function is used.

    If you specify /MATH_LIBRARY=FAST, the math library routines do
    not necessarily check for IEEE exceptional values and the
    /IEEE_MODE qualifier is ignored.

    When you use MATH_LIBRARY=FAST, you should carefully check the
    calculated output from your program to verify it is not relying
    on the full fractional accuracy of the floating-point data type
    to produce correct results.  You should also verify that the
    calculated output is not producing unexpected exceptional values
    (exception handling is indeterminate).

    Programs that do not produce acceptable results with
    /MATH_LIBRARY=FAST and single-precision data might produce
    acceptable results if they are modified (or compiled) to use
    double-precision data.

    The specific intrinsic routines that have special fast math
    routines depend on the version of the OpenVMS Alpha operating
    system in use.  Allowed error bounds vary with each routine.