Loop Optimizations

The compiler can do several types of loop optimization. Some of the more common ones are as follows:

• Loop unrolling

• Loop peeling

• Loop interchange

• Loop fission

• Loop fusion

Loop unrolling consists of repeating several iterations of a loop within the loop body, and adjusting the loop index accordingly. As the body of the loop becomes larger, the compiler can schedule the instructions more efficiently. Also reduced is the overhead caused by the loop index increment and conditional check operations. The remainder of the loop is handled using loop peeling.

Loop peeling consists of removing a number of loop iterations from the loop, and moving them in front of or after the loop, as appropriate.

Loop interchange changes the ordering of nested loops to minimize memory stride, to maximize cache-line hit rates.

Loop fusion consists of combining adjacent or closely located loops into a single loop. The benefits of loop fusion are similar to loop unrolling. In addition, if common data is accessed in the two pre-optimized loops, cache locality is improved by loop fusion, providing the compiler with more opportunities to exploit instruction-level parallelism.

Loop fission is the opposite of loop fusion: a loop is split into two or more loops. This optimization is appropriate if the number of computations in a loop becomes excessive, leading to register spills that degrade performance. Loop fission can also come into play if a loop contains conditional statements. Sometimes it is possible to split the loops into two: one with the conditional statement and one without. This can increase opportunities for software pipelining in the loop without the conditional statement.

Sometimes, with nested loops, the compiler applies loop fission to split a loop apart, and then performs loop fusion to recombine the loop in a different way to increase performance. In this case, you see compiler commentary similar to the following:

    Loop below fissioned into 2 loops
    Loop below fused with loop on line 116
    [116]    for (i=0;i<nvtxs;i++) {