4. Reduction of Waiting Time for Cache Access

When CPU performance analysis reports show higher ratio of waiting time for L1D cache access on cycle accounting, there might be room to reconsider data access pattern of the source program. Effective techniques for such cases are explained in this section.

• 4.1. Full-Unrolling of Innermost Loop with Non-Contiguous Data Access
• 4.2. Interchange of Array Dimension for AoS Layout
• 4.3. Specifying CONTIGUOUS Attribute to Array Pointer

On cycle accounting, waiting time for calculation and cache access is distinguished for each CPU cycle by execution status of the oldest in-flight CPU instruction in the program order. Therefore, techniques for Reduction of Waiting Time for Calculation might reduce waiting time for cache access together as a result, for example in case of a program where there are many non-contiguous array references within a loop.