Question

The test(...) helper in TestTensorflow2Model counts tensor variables by iterating over all (CGNode, vn) pairs in the analysis results that match the target function name. Under k-CFA context sensitivity, the same IR variable (same value number in the same method) can appear under multiple calling contexts (e.g., distinct trampolines), inflating the count relative to context-insensitive analysis.

This is a design question, not clearly a bug: tests could reasonably reflect either context-sensitive counts (current behavior) or IR-level source counts (one entry per (CGNode, vn) regardless of context).

Design Question

Should the test helper count:

1. By (CGNode, vn) — current behavior. Faithfully reflects what the analysis computed. Count changes when context-sensitivity depth changes.
2. By vn only — deduplicate across contexts. Count reflects IR-level source variables. Stable under context-sensitivity changes.

Option 1 is the current implementation; option 2 is what older tests assumed before higher-k context sensitivity was added.

Impact of the Current Behavior

• Test expectations become coupled to context-sensitivity implementation details.
• Changes to calling-context structure (e.g., adding/removing trampolines, changing k in k-CFA) silently shift the expected count.

Possible Resolution

• Decide on the counting semantics (option 1 or option 2 above) and document it in the test helper's Javadoc.
• If option 2: deduplicate by vn in the test helper; existing tests that expected the context-insensitive count continue to work.
• If option 1: accept that test expectations may need updating when context-sensitivity changes; document this as a known maintenance cost.