Coverage regressions for "with" statement on CPython 3.10

[skirpichev]

Per request I've opened a separate issue.

There are regressions in the diofant/integrals/rde.py on lines with the "with" statement (e.g. L309 and L546). On the CPython 3.9 & the latest coveragepy release - these lines are marked as covered.

To reproduce issue, install the diofant e.g. with pip install -e .[develop] in the git tree, install pytest and coveragepy from git (I use coveragepy at 57a691f) and run:

pytest -n0 -m 'not slow and not xfail and not regression' --cov diofant \
    diofant/tests/integrals/test_rde.py diofant/tests/integrals/test_risch.py
coverage html

Here are two examples (there are four such "with" statements). The function bound_degree():

The function cancel_primitive():

The first one you may view as a new "feature" of the coveragepy. But not too helpful, as it looks for me, since L320 and L342 (i.e. ValueError exceptions) were marked as uncovered before. Maybe it's a documentation issue: how works the branch coverage here - seems unclear for me (wrt "if" statements, for example). For instance, why there are (and on L546) 3 branches:

<line number="309" hits="1" branch="true" condition-coverage="66% (2/3)" missing-branches="exit"/>
<line number="546" hits="1" branch="true" condition-coverage="66% (2/3)" missing-branches="exit"/>

But the second example looks definitely wrong. The NotImplementedError on L552 is ignored by the coveragepy configuration. I think, then this "branch" of the "with" statement on L546 - should be ignored too (as it happens for if/elif/else statements).

[nedbat]

Thanks. This was coverage.py being over-zealous about warning about missed branches. I've fixed it in commit f6d3e88.