ternary ifs not taken into account by branch coverage

[nedbat]

Originally reported by Antony Lee (Bitbucket: anntzer, GitHub: anntzer)

---

coverage 4.1, python 3.5.2, arch linux

test.py

cond = True

if cond:
    x = 1
else:
    x = 2

x = 1 if cond else 2
x = cond and 1 or 2

Getting branch coverage for this file shows partial coverage of the first if cond:, but not of the ternary (inline) if cond, or of the short-circuiting binary operators.

---

• Bitbucket: https://bitbucket.org/ned/coveragepy/issue/509

[nedbat]

The Python settrace facility only offers feedback about lines executed. Supporting ternary-if and short-circuiting binary operators requires more details than Python offers. There are experiments about getting those details (http://nedbatchelder.com/blog/200804/wicked_hack_python_bytecode_tracing.html), but it's impractical to support in coverage.py for now.

[nedbat]

Original comment by Antony Lee (Bitbucket: anntzer, GitHub: anntzer)

---

Perhaps it may be worth revisiting this now that Py3.7 optionally supports opcode tracing (if I understood https://bugs.python.org/issue31344 correctly)?

[nedbat]

Bytecode tracing is definitely an enabling technology, but there is still a lot to be done, including how to correlate tokens in the source with bytecodes at run-time. There's no information in the compiled code object about that. If someone wants to do some experiments, I'll be very interested in the results.

[graingert]

this also applies to any sort of short circuiting operator eg:

def true():
    return True


def throw_error():
    raise Exception


def error():
    a = true()
    b = throw_error()
    return (a or b)


def no_error():
    return (true() or throw_error())  # True


def test():
    print(no_error())  # True
    print(error())     # Exception()


if __name__ == '__main__':
    test()

of course when expanded to

def no_error():
    if x := true():
        return x
    return throw_error()

it's caught by branch and line coverage

[nedbat]

This would be an amazing feature to implement, but has a number of challenges:

• How do we know what byte codes could be executed? I think it's likely that there are dead op codes that are not reachable.
• How do we report on the results? Python doesn't track a mapping from tokens to byte codes.

[ocehugo]

This thing just happened to me since I was curious about why I was not getting a lack of coverage.

1. Maybe put a warning msg for some ternary operator cases ?

2. Silly solution: why not just expand the source code into ifs for the analysis? Like storing original AST/FST, calculating/piping-in modified AST/FSTs (expanded) and then remapping results back to original AST/FST!? maybe baron package helps!?