[Python-Dev] Add Py_SETREF and Py_XSETREF to the stable C API

[Serhiy Storchaka]

09.11.17 04:08, Raymond Hettinger пише:
> 
>> On Nov 8, 2017, at 8:30 AM, Serhiy Storchaka <storchaka at gmail.com> wrote:
>>
>> Macros Py_SETREF and Py_XSETREF were introduced in 3.6 and backported to all maintained versions ([1] and [2]). Despite their names they are private. I think that they are enough stable now and would be helpful in third-party code. Are there any objections against adding them to the stable C API? [3]
> 
> I have mixed feeling about this.  You and Victor seem to really like these macros, but they have been problematic for me.  I'm not sure whether it is a conceptual issue or a naming issue, but the presence of these macros impairs my ability to read code and determine whether the refcounts are correct.  I usually end-up replacing the code with the unrolled macro so that I can count the refs across all the code paths.

If the problem is with naming, what names do you prefer? This already 
was bikeshedded (I insisted on discussing names before introducing the 
macros), but may now you have better ideas?

The current code contains 212 usages of Py_SETREF and Py_XSETREF. Maybe 
10% of them correctly used temporary variables before introducing these 
macros, and these macros just made the code shorter. But in the rest of 
cases the refcount was decremented before setting new value. Not always 
this caused a problem, but it is too hard to prove that using such code 
is safe in every concrete case. Unrolling all these invocations will 
make the code larger and more cumbersome, and it is hard to do 
automatically.

> The other issue is that when there are multiple occurrences of these macros for multiple variables, it interferes with my best practice of deferring all decrefs until the data structures are in a fully consistent state.  Any one of these can cause arbitrary code to run.  I greatly prefer putting all the decrefs at the end to increase my confidence that it is okay to run other code that might reenter the current code.  Pure python functions effectively have this built-in because the locals all get decreffed at the end of the function when a return-statement is encountered.  That practice helps me avoid hard to spot re-entrancy issues.

I agree with you. If you need to set two or more attributes 
synchronously, Py_SETREF will not help you. This should be clearly 
explained in the documentation.

Several subsequent Py_SETREFs may be an error. When I created my patches 
for using Py_SETREF I encountered several such cases and used different 
code for them. Maybe still there is not completely correct code, but in 
any case it is better now than before introducing Py_SETREF.

But in many case you need to set only one attribute or different 
attributes are not tightly related.

> Lastly, I think we should have a preference to not grow the stable C API.  Bigger APIs are harder to learn and remember, not so much for you and Victor who use these frequently, but for everyone else who has to lookup all the macros whose function isn't immediately self-evident.

I agree with you again. But these macros are pretty helpful. They allow 
to write the safer code easy. And they are more used than any other C 
API addition in 3.5, 3.6, and 3.7.

3.7:
Py_X?SETREF -- 216
Py_UNREACHABLE -- 65
Py_RETURN_RICHCOMPARE -- 15
PyImport_GetModule -- 28
PyTraceMalloc_(T|Unt)rack -- 9
PyOS_(BeforeFork|AfterFork_(Parent|Child)) -- 24
Py_tss_NEEDS_INIT -- 6
PyInterpreterState_GetID -- 3

3.6:
PySlice_Unpack -- 22
PySlice_AdjustIndices -- 22
PyErr_SetImportErrorSubclass -- 3
PyErr_ResourceWarning -- 6
PyOS_FSPath -- 9
Py_FinalizeEx -- 21
Py_MEMBER_SIZE -- 4

3.5:
PyCodec_NameReplaceErrors -- 3
PyErr_FormatV -- 6
PyCoro_New -- 3
PyCoro_CheckExact -- 14
PyModuleDef_Init -- 30
PyModule_FromDefAndSpec2? -- 10
PyModule_ExecDef -- 3
PyModule_SetDocString -- 4
PyModule_AddFunctions -- 3
PyNumber_MatrixMultiply -- 4
PyNumber_InPlaceMatrixMultiply -- 4
Py_DecodeLocale -- 25
Py_EncodeLocale -- 11

Some older macros:
Py_STRINGIFY -- 15
Py_ABS -- 54
Py_MIN -- 66
Py_MAX - 56

The above number include declaration and definition, hence remove 2 or 3 
per name for getting the number of usages.

Only added in 3.4 Py_UNUSED beats them (291) because it is used in 
generated Argument Clinic code.