STINNER Victor added the comment: Let me try to explain this issue again.
"f1, f2 = lambda: 1, lambda: 1.0" is compiled to two MAKE_FUNCTION instructions, MAKE_FUNCTION takes a code object as parameter (and a name). The Python compiler merges constants which are seen as "equal", with exceptions to not merge values of different types, or float of different sign. Merging duplicate code objects is a cool micro optimization, I prefer to keep it. My patch keeps this micro optimization, but fix the bug: it ensures that equal constants having different types are not seen as equal. For example, 0 is equal to 0.0, but if when used for code constants, the code objects are seen a different. Patch version 3: * as suggested by Armin Rigo & Serhiy Storchaka: use id(obj) in the constant key for unknown constant types -- in practice, this patch is never taken, it's just to be extra safe (I checked manually by running the whole test suite when an assertion, assertion not in the posted patch) * add a lot of unit tests * add a documentation to _PyCode_ConstantKey() @Serhiy: does it look good to you now? > Would option (1) work if wrap 1 and 1.0 in a tuple? In a list? In a custom > collection? My patch now uses id(obj) in the "constant key" for unknown types. The compiler only emits simple type (int, str, ...), tuple, frozenset and code objects. You *can* other types if you patch manually constants with custom objects, since code_richcomp() now uses the "constant key" function to compare constants. For example, my fat project has a replace_consts() function to inject builtin functions in constants: http://fatoptimizer.readthedocs.org/en/latest/fat.html#replace_consts There is also @asconstants decorator of codetransformer which allow to inject arbitrary types in function constants: https://pypi.python.org/pypi/codetransformer ---------- Added file: http://bugs.python.org/file41671/code_richcompare-3.patch _______________________________________ Python tracker <rep...@bugs.python.org> <http://bugs.python.org/issue25843> _______________________________________ _______________________________________________ Python-bugs-list mailing list Unsubscribe: https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
