ponder-lab / Hybridize-Functions-Refactoring / #2376

81%

main: 83%

push

github

Add a safety precondition for numpy applied to parameter-flowing values (#743)

* Add a parameter-flow numpy safety precondition

A function whose body (transitively) applies a numpy/scipy API to a value flowing from its parameters crashes on first call under `@tf.function` tracing, once the parameters become symbolic tensors. The gate is an SSA def-use taint slice rather than a points-to intersection: a points-to-keyed gate silently passes a crasher under any modeling gap that empties a points-to set, and this precondition exists precisely to hold while upstream modeling is in motion (wala/ML#696's generator fix re-types `compute_loss`'s parameters; this precondition is what makes that safe to absorb).

Taint policy per the design on #740: sources are the parameter value numbers (the receiver excluded for instance methods); propagation is worklist-based over def-use chains; property reads of `shape`/`dtype` launder taint (probe-verified trace-time available on TensorFlow 2.9.3, admitting the `PositionEncoding.call` idiom); sinks are invocations whose callee chain roots at the numpy/scipy module (points-to preferred, import-alias fallback over `numpy`/`np`/`scipy`/`sp` globals—whose over-recognition here over-blocks, the inverse of the TensorFlow chain walk's trade); tainted positional arguments cross into user callees and tainted returns flow back, memoized with a cycle guard.

The failure (`HAS_NUMPY_CALLS_ON_PARAMETERS`, code 16) joins the safety family after `HAS_EAGER_ONLY_CALLS` and before the barren benefit signal. `alwaysCheckNumpyCalls` mirrors the sibling knobs end-to-end (`EvaluationOption`, evaluator handler, `run-headless-evaluator.sh` `CHECK_NUMPY_CALLS`, "numpy calls on parameters"/"numpy calls" CSV columns).

`testNumpyCallsOnParametersBlockHybridization` pins the direct sink, argument-flow into a callee, and the return-flow chain, with clean-callee and no-numpy P1 controls; `testNumpyCallsSanitizedByShape` pins the trace-time carve-ou... (continued)

101 of 112 new or added lines in 4 files covered. (90.18%)

1927 of 2389 relevant lines covered (80.66%)

0.81 hits per line