8.3 Loop Table Messages

The Loop Table (/listoptions=l ) includes an entry for each loop indicating whether it was optimized, or why it was not. This section lists the possible messages and gives a brief explanation for each. The two most common reasons for a loop to be left serial are that the iterations were not independent and that the loop contained I/O statements.

Messsages referring to DO loops apply to C for loops:

• bad intrinsic call - Loop optimization may give unpredictable results.

• bad operator - Loop optimization may give unpredictable results.

• branches out of-into loop - Code may be iteration- order dependent.

• cannot allocate temporary array - Usually, a data dependence could have been lifted by expanding a scalar into an array, but there was a name conflict.

• concurrentized - Should not appear. If it does, indicates miscellaneous optimization.

• distributed - Loop divided into smaller loops, some of which may be further optimized.

• DO statement not optimizable - For example, the iteration count of a while loop is not known.

• ELSE IF - Some loops inside nested IFs require more information to optimize than KAP can deduce.

• empty loop - Loop with no executable statements inside (nothing to optimize), or loop whose body will never be executed.

• fused with other loops - Adjacent loops joined into a larger loop to reduce overhead.

• GOTO in loop - Possible flow patterns are too complicated for KAP.

• I/O statement in loop - Input/Output is assumed to be order-dependent.

• left as DO loop - Usually, a data dependence prevented concurrentization.

• left as DOALL - Loop was already explicitly concurrent in the input dialect of C.

• left as a serial loop - Usually, too much synchronization required.

• nested IFs - The branching pattern was too complicated.

• no optimizable statements - For example, no arithmetic in loop.

• not enough work - Optimization overhead greater than the expected gain.

• optimization disabled - By a directive or assertion.

• replaced by calls - Some loop nests can be replaced by a call to a library routine.

• scalar directive - A directive or assertion turned off optimization.

• scalar inner loop - Inside a parallel loop already.

• scalar mode preferable - Usually, too much synchronization needed.

• too many inner loops - LIMIT > integer VERY complicated loop nest; KAP will try to optimize the inner loops. Try increasing /limit .

• too many stmts/DD arcs - VERY complicated loop nest; try increasing /arclimit.

• too much synchronization - Loop iterations depend on each other in a complicated way; the overhead from untangling them would be greater than expected gain.

• unoptimizable call - Function with unknown side effects.

• unoptimizable operation - Character string manipulation.

• unoptimizable statements in loop - Usually I/O.

• unoptimizable symbol use - Variable used in a complicated way.

• unoptimizable variable or operation type - Only numerical work is easily concurrentized.

• unrolled - Loop could not be concurrentized, so KAP unrolled it (see explanations of unrolling in Chapter 4 and Chapter 7).

• unrolled completely, fused or removed - Loop has been turned into straight-line code, joined to another loop, or dropped entirely.

• unrolled; already in a parallel loop - Loop could be concurrentized, but target machine does not support nested parallelism; KAP unrolled it.

• vectorized - Can be used for optimizations which enhance the vectorization performed by the C compiler.