ponder-lab / Hybridize-Functions-Refactoring / #3568

89%

main: 89%

Pull #1090

github

Model the builtin open as a file-system write in the side-effect analysis

A call whose callee points to Ariadne's object for the builtin `open`
now adds a fake global for the file system to the call's mod set, next
to the output stream that `print` writes. The callee is recognized by
points-to, not by its name, so `opener = open; opener(path)` is caught
and a local function bound to `open` is not. Under tf.function the call
runs once, at trace time, so the functions of #1077 that reach `open`
(`Model.visualize`, `predict_dataset`, `process_image_label`,
`reads_file`, `calls_reader`) now fail with HAS_PYTHON_SIDE_EFFECTS.

The other trace-time effects need engine facts. `random`, `time`, `cv2`
and `tf.io.gfile` have no summary, so their reads point to nothing
(wala/ML#1080), and the `numpy.random` summaries model a draw's result
but not the generator state it advances. Those cases, and
`wild_random_size` in testNumpyOverWildDimension, are pinned as still
converted, with TODOs.

testPythonSideEffects10 and 11 now see a side effect, from their `open`
calls; a write to a file the function does not open stays #273.

Part of #1077.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

9 of 9 new or added lines in 2 files covered. (100.0%)

5220 of 5891 relevant lines covered (88.61%)

0.89 hits per line