ponder-lab / Hybridize-Functions-Refactoring / #3478

88%

main: 89%

push

github

Read a tensor condition from the tensor-type analysis, not the taint (#1085)

* Read a tensor condition from the tensor-type analysis, not the taint

HAS_UNTRACEABLE_PYTHON_ON_PARAMETERS (35) counted a branch, or a bool()
call, when its condition or argument was derived from a parameter, a
stand-in for "is a tensor" that the tensor-type analysis is responsible
for. It refused a parameter's `is None` test, a Python bool decided
while tracing. Measured on TensorFlow 2.9.3, such a function traces.

A branch now counts only when the tensor-type analysis types its
condition value as a tensor, and bool() only when it types the
argument. The value checked is the condition itself, not the operands
of a comparison that produces it. The None-disagreement test, and the
reach through callees and closures, are unchanged.

Ariadne types no result of a Python comparison operator on a tensor
(wala/ML#1074), so a branch on `tf.reduce_sum(x) > 0.0` is no longer
detected. The five such cases are pinned as not refused, with TODOs.
New cases spelled with tf.greater, which Ariadne types, keep each part
of the rule refused: the region return, the merge, the read after the
join, bool(), and the callee.

Part of #1083.

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

* Pin a branch on ops Ariadne does not model

A branch on tf.reduce_any(tf.math.is_nan(x)) whose paths disagree about
None raises ValueError under tf.function on TensorFlow 2.9.3. The parameter
taint used to carry through the unmodeled calls and refuse it; read from
the tensor-type analysis, the condition has no type, since neither op is
modeled. The new none_return_unmodeled case pins it as not refused, with a
TODO citing wala/ML#1076.

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

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

16 of 16 new or added lines in 2 files covered. (100.0%)

5142 of 5818 relevant lines covered (88.38%)

0.88 hits per line