ponder-lab / Hybridize-Functions-Refactoring / #2501

83%

push

github

Use tensor-origin provenance for tensor-computation detection (#777)

The tensor-computation detector counts any tensor-typed def, and Ariadne
types a numpy array identically to a TensorFlow tensor, so pure-numpy
functions were misreported as computing and hybridized (#774). #775
excluded defs whose callee is a numpy free function, but that signal is
absent for a numpy operator (`m + m`), an interprocedural call returning a
numpy result, and an ndarray method (`m.reshape`), which stayed
misreported.

Replace the callee-type probe with Ariadne's origin record (wala/ML#724):
`computeTensorTypedOrigins` maps each tensor-typed pointer key to its
producing-library origins, and `definesTensor` counts a def unless its
origin set is exactly `{NUMPY}`. A TensorFlow origin, a parameter origin
(wala/ML#726: a tensor parameter is a symbolic tensor under tracing
regardless of its eager feeds, so `f(x): return x + 1` fed a numpy array
still counts), a mixed set, or no evidence all count; a numpy-only def
does not.

Requires Ariadne 0.52.29.


Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

8 of 8 new or added lines in 2 files covered. (100.0%)

2139 of 2616 relevant lines covered (81.77%)

0.82 hits per line