Internal table access performance varies with subscript versus index declaration

Performance Resolution

In COBOL, INDEXED BY typically outperforms subscripts because indices store the displacement (the actual byte offset from the start of the table), whereas subscripts store a logical occurrence number. For every subscript access, the runtime must calculate (Index * Element Size) + Base Address. Indices bypass this multiplication, allowing the CPU to use direct register offsets or simple addition.

Measurement and Isolation Strategy

To isolate the cost of address arithmetic from I/O and business logic, use a synthetic harness that iterates over a large internal table without performing external calls. To ensure the measurement reflects production behavior, execute the harness in a representative LPAR with production-level buffer pools.

• Profiling Tools: Use vendor-specific performance monitors (such as IBM z/OS RMF or specialized COBOL profiling options) to capture CPU time at the paragraph level. If available, enable COMPILER OPT levels to see if the compiler hoists subscript calculations.
• Harness Design: Create a loop that performs a simple read-and-sum operation on a table large enough to exceed the L1 cache. This forces the system to demonstrate the overhead of repeated displacement calculations across a wide memory range.

Assembly Verification Patterns

Inspect the LIST or MAP output to confirm the generated machine code. The difference is visible in the instructions used to resolve the table element address:

Access Method	Assembly Pattern (Typical)	Operation
Subscript	MLT or SLL (Shift Left)	Multiplies the subscript by the element length before adding to base.
INDEXED BY	LA (Load Address) or ADD	Adds the pre-calculated displacement directly to the base address.

Assumptions and Uncertainty

This analysis assumes a standard mainframe-style COBOL implementation. The reported 2–10× performance delta is highly variable. Modern optimizing compilers may automatically convert subscripts to indices if the loop bounds are constant, or hardware decimal assist units may reduce the cost of the multiplication to a negligible level. Verification via assembly is the only way to confirm if the optimization is actually occurring.

Required Diagnostic: Please specify the current compiler version and optimization level (e.g., OPT(1), OPT(2)), as this determines whether subscript hoisting is active.