bpo-39573: Porting to Python 3.10: Py_SET_SIZE() macro

[vstinner]

In What's New in Python 3.10, propose Py_SET_SIZE(), Py_SET_REFCNT()
and Py_SET_TYPE() macros for backward compatibility with Python 3.9
and older.

https://bugs.python.org/issue39573

[vstinner]

@tacaswell @njsmith @corona10: Would you mind to review this documentation change?

[tacaswell]

Looks good to me, thanks!

[corona10]

lgtm

[vstinner]

@tacaswell: In one of your PR, you wrote "typ" instead of "type". Was it a deliberate typo? Is there an issue in C++? Should I rename the variable to "new_type"?

[vstinner]

Oops, my question is on the second parameter of Py_SET_TYPE() :-)

[vstinner]

I built Py_SET_TYPE() macro with g++ and Python 3.8: I don't have any warning or error. "type" sounds like a valid parameter type.

[tacaswell]

I am not sure, that code come from @eric-wieser via numpy/numpy#16417 (comment)

[eric-wieser]

This macro doesn't behave the same way as the function in 3.9. I think a more accurate macro would be:

Suggested change

	#if PY_VERSION_HEX < 0x030900A4
	# define Py_SET_TYPE(obj, type) do { Py_TYPE(obj) = (type); } while (0)
	#endif
	#if PY_VERSION_HEX < 0x030900A4
	# define Py_SET_TYPE(obj, type) ((Py_TYPE(obj)) = (type), (void)0)
	#endif

which can be used as part of an expression like the function version of Py_SET_TYPE can:

return Py_SET_TYPE(my_obj, some_type), my_obj;

is legal in 3.9 and with my macro, but illegal with this PR.

[vstinner]

Oh wow, the C language is so surprising sometimes :-) I designed Py_SET_SIZE() as a static inline function to ensure that the result is void and to prevent strange usage. But I didn't know that return Py_SET_TYPE(my_obj, some_type), (...); was legit.

I updated macros to use , (void)0. According to https://stackoverflow.com/a/25021889 this syntax is endorsed by the C standard, since NDEBUG uses #define assert(ignore) ((void)0).

[eric-wieser]

Thanks for finding some precedent. I've seen "always use parens and do { } while(0);" given as macro advice forever, but I don't remember ever seeing this case before.

[vstinner]

In the meanwhile, I created the backport compatiblity header file :-) It defines a static inline function which returns void:
https://github.com/pythoncapi/pythoncapi_compat/blob/219cdd858595fb5d478117b0ef4758fedc3d2365/pythoncapi_compat.h#L29

Maybe tomorrow when it will be more mature (than 1 hour old), numpy might want to use it.

[vstinner]

@eric-wieser @tacaswell: Would you mind to review the updated PR?

[eric-wieser]

Looks reasonable to me. Maybe we should try committing this modified macro to numpy first just to prove it works.

[vstinner]

Looks reasonable to me. Maybe we should try committing this modified macro to numpy first just to prove it works.

@eric-wieser @tacaswell: Do you want to try that first?

[eric-wieser]

Just did, CI should be running now

[vstinner]

I see numpy/numpy#16501 :-)

[eric-wieser]

Numpy looks happy with this spelling.

[vstinner]

Thanks for the reviews!