Empty diagnostics when using Python optimization

[treiher]

We have observed in RecordFlux that the diagnostics of our langkit-based parser is empty when python -O is used (AdaCore/RecordFlux#603). The root cause seems to be an assertion in the Python binding (librflxlang/__init__.py):

1120     @property
1121     def diagnostics(self):
1122         """
1123         Diagnostics for this unit.
1124         """
1125         count = _unit_diagnostic_count(self._c_value)
1126         result = []
1127         diag = Diagnostic._c_type()
1128         for i in range(count):
1129             assert _unit_diagnostic(self._c_value, i, ctypes.byref(diag))
1130             result.append(diag._wrap())
1131         return result

The function _unit_diagnostic seems to have side effects and is not executed when -O is used. If line 1129 is commented out, the issue also occurs without -O.

6313 _unit_diagnostic = _import_func(
6314     "rflx_unit_diagnostic",
6315     [AnalysisUnit._c_type, ctypes.c_uint, ctypes.POINTER(Diagnostic._c_type)],
6316     ctypes.c_int,
6317 )

We use -O for performance reasons, otherwise the assertions and contracts used in RecordFlux would make the execution significantly slower.

Versions

GNAT: Community 2020
Langkit: 21.0.0

[pmderodat]

Hello @treiher,

Thank you for reporting this. A fix is pending review on #486.

[pmderodat]

Hello again,

This is now fixed in the master branch. Thanks again for reporting this bug!