User's Guide

The Philosophy of XL Fortran Optimizations

XL Fortran optimizations can be characterized according to their aggressiveness, which determines how much risk they carry. Only the very highest optimization level performs aggressive optimizations, and even then the risk is limited to slightly different results in a small subset of possible programs.

The less-aggressive optimizations are intended to produce exactly the same results as an equivalent unoptimized program:

• Code that might cause an exception is not moved unless the exception is certain to occur anyway. In the following example, the program could evaluate the expression N/K before the loop because the result is the same for each iteration of the loop:

        DO 10 J=1,N
          ...
          IF (K .NE. 0) M(J)=N/K
          ...
  10    END
  

  However, it is not moved because K might be 0, and computing N/K results in an exception where none occurs in the unoptimized program.

• The rules for IEEE arithmetic are followed more closely than otherwise. ⁴ For example, X+0.0 is not folded to X, because IEEE rules require that -0.0+0.0 be 0, making X+0 equal to -X in this one case.
• Floating-point calculations are not considered associative. For example, XL Fortran evaluates X*Y*Z left-to-right, even though the program might already have computed Y*Z, because the results might not be identical.

As the optimization level increases, these restrictions are relaxed where there is an opportunity for a performance improvement:

• Calculations like N/K in the previous example and floating-point operations may be moved or rescheduled because they are unlikely to cause exceptions.
• IEEE conformance is not enforced for rules that are unlikely to be needed. The sign of zero might not be correctly preserved, as in the preceding example. However, this might only be a problem in an extreme case, such as multiplying the wrongly signed zero by infinity and ending up with an infinity of the wrong sign. Floating-point operations that might cause an exception might be removed entirely.
• Floating-point expressions might be reassociated, so that results might not be identical.

When you specify the highest level of optimization, XL Fortran assumes that you are requesting speed even at the possibility of some risk, as already explained. If you want as much optimization as possible without the resulting risk, you must add the -qstrict compiler option.

In the past, the XL family of compilers adopted a conservative approach to optimization. This was intended to make an optimized program work exactly the same as an unoptimized one even in extreme cases unlikely to occur in real life. For example, the array reference A(N) might not be optimized, because N might be a huge number so that the program causes a segmentation violation when the address is referenced, and this behavior would be "preserved". With the industry in general favoring a less conservative approach, XL Fortran's highest optimization level now emphasizes performance over identical execution between optimized and unoptimized programs.

The different levels of the -O option incorporate various optimization techniques that are expected to improve performance for many different kinds of programs. The specialized optimization options, such as -qipa, -qhot, and -Q, can improve performance in some kinds of programs but degrade it in others. Therefore, they may require experimentation to determine whether they are appropriate for any given program.

⁴

If IEEE compliance is a concern for you, you should also specify either -qfloat=nomaf or -qfloat=rrm.