When a value number cannot be traced (empty points-to set and no recoverable property chain), TensorGenerator throws IllegalArgumentException: Empty points-to set and could not trace properties for value number ... rather than flooring to ⊤:
TensorGenerator.java:931 (in the shape path).TensorGenerator.java:1967 (in the dtype path).
This is the same crash-floor anti-pattern as #611 and #612, one level deeper: any caller that probes the shape/dtype of an unmodeled or content-dependent value aborts the whole analysis. Surfaced while attempting #619, where probing a tf.eye batch_shape tensor (a tf.constant(json.loads(...)), and tf.shape(x)) threw here and collapsed the otherwise-floored tf.eye result from ⊤ to ⊥.
Fix
Replace each throw with a ⊤ floor (null shapes / EnumSet.of(DType.UNKNOWN) dtypes), per the lattice conventions in CONTRIBUTING.md, with an unresolvable-value regression test. This unblocks safe shape/dtype probing (e.g. the #619 batch-rank recovery).
When a value number cannot be traced (empty points-to set and no recoverable property chain),
TensorGeneratorthrowsIllegalArgumentException: Empty points-to set and could not trace properties for value number ...rather than flooring to ⊤:TensorGenerator.java:931(in the shape path).TensorGenerator.java:1967(in the dtype path).This is the same crash-floor anti-pattern as #611 and #612, one level deeper: any caller that probes the shape/dtype of an unmodeled or content-dependent value aborts the whole analysis. Surfaced while attempting #619, where probing a
tf.eyebatch_shapetensor (atf.constant(json.loads(...)), andtf.shape(x)) threw here and collapsed the otherwise-flooredtf.eyeresult from ⊤ to ⊥.Fix
Replace each throw with a ⊤ floor (
nullshapes /EnumSet.of(DType.UNKNOWN)dtypes), per the lattice conventions inCONTRIBUTING.md, with an unresolvable-value regression test. This unblocks safe shape/dtype probing (e.g. the #619 batch-rank recovery).