Pointer Analysis Metrics

include/Alias/Infrastructure/Metrics/ and lib/Alias/Infrastructure/Metrics/ provide helpers for measuring precision and soundness-related properties of alias analyses.

Main components:

PointerAnalysisMetrics stores collected statistics.
CollectMetricsOptions configures metric collection.
collectMetricsFromWrapper runs the collection against an alias-analysis wrapper.

This module is primarily for evaluation, comparison, and regression tracking of points-to algorithms.

Collected measurements

The collector records points-to-set size statistics when a wrapper exposes them, including the number of tracked pointers, total, maximum, average, and median set size. It also records direct and resolved indirect call-graph edges, the number of polymorphic indirect call sites, and the average number of targets per indirect call.

Optionally, it samples alias queries over pointer use-site pairs. The result separates NoAlias, MustAlias, MayAlias, and PartialAlias answers. PointerAnalysisMetrics::fractionNoAlias is a compact indication of how many sampled pairs the analysis can prove disjoint; it is most useful when compared across runs using the same module and query limit.

Cost and interpretation

Points-to and call-graph counts are inexpensive summaries of an already-run analysis. Pairwise alias sampling can be much more costly, so CollectMetricsOptions::max_alias_pairs caps it (the default is 50,000). Set this value to zero to omit pair metrics, or use a fixed nonzero value for comparable regression results.

These are comparative metrics, not a proof of soundness. Lower points-to counts or fewer indirect targets commonly indicate greater precision, but the comparison is meaningful only when analyses model the same program, libraries, and external-call assumptions.