ponder-lab / Hybridize-Functions-Refactoring / #2855

85%

push

github

Withhold a signature derived only from a declared expected failure (#892)

* Withhold a signature derived only from a declared expected failure

The reduction inferred a specification from every observed call site,
including sites the tests declare must fail. For deep_recommenders'
CIN.call that meant emitting one `TensorSpec` reduced from the site
passing a bare array inside `assertRaisesRegexp`, which admits exactly
the input the body rejects and refuses every conforming caller (#888).

`unittest` enforces `assertRaises`: the test fails if the call succeeds.
So the site is an executable assertion that the callee rejects that
argument, which is the one kind of evidence a specification must not rest
on, and reading the developer's own assertion is what makes this a marker
rather than a heuristic.

`ExpectedFailureContextAnalysis` computes the call-graph nodes reached
only from guarded sites, hopping synthesized trampolines so a Keras
`call` override reaches the frame the guard is written in, and testing
guardedness by dominance, which is what the IR preserves of the `with`
body. Dominance is read directly as unreachability from the entry with
the guard removed, since the bundle does not export WALA's `Dominators`.

The evidence itself is already per-node, so `inferTensorTypes` now keeps
that attribution instead of unioning it away: `getTensorTypes()` still
carries everything observed, so classification is untouched, while
`getConformingTensorTypes()` drops the guarded nodes and is what the
reduction reads. Where only guarded evidence exists, the parameter blocks
with the new `EXPECTED_FAILURE_EVIDENCE_ONLY` absence reason and an INFO
naming the cause, and the function still converts at P1 with a bare
decorator, since hybridizing it preserves behavior whether or not a
specification can be written.

Deliberately not attempted: recovering the two-element nested spec the
conforming site does support. That needs the container evidence collected
where Phase... (continued)

89 of 99 new or added lines in 5 files covered. (89.9%)

3400 of 4016 relevant lines covered (84.66%)

0.85 hits per line