CPython has been created in 1990. In 1990, it made sense to use C macros. But nowadays, inlined functions can be used instead:"Python versions greater than or equal to 3.6 use C89 with several select C99 features: (...) static inline functions"https://www.python.org/dev/peps/pep-0007/#c-dialectI propose to convert 4 macros to inlined functions:* PyObject_INIT(), PyObject_INIT_VAR()* _Py_NewReference(), _Py_ForgetReference()Advantages:* Functions use regular C syntax* No more corner cases ("traps") of macros* Function arguments have a typeDrawbacks:* Require a specific type can introduce compiler warnings if the caller doesn't pass the proper type (PyObject* or PyVarObject*). _Py_NewReference() and _Py_ForgetReference() seem to be properly used, but not PyObject_INIT() and PyObject_INIT_VAR().The two attached PRs implements these changes.

Context: I was unhappy with _Py_NewReference() macro implementation when I had to modify it to fix a bug,bpo-35053. That's why I would like to convert it to a static inline function.

Does this slow down debug builds at all? It probably will not end will if Py_INCREF is ever not inlined.

I testedPR 10079 using gdb on Fedora 28 with GCC 8.1.1 to check if Py_INCREF/Py_DECREF functions are inlined.I understand that "static inline void Py_INCREF()" is *not* inline by gcc -O0, but it *is* inlined using gcc -Og which is the *default* optimization level of ./configure --with-debug.To develop on Python, I force -O0, because the compilation time matters more than runtime performance for me :-) Compilation on my laptop using MAKEFLAGS=-j9:* -O0: 21s* -0g: 36s (1.7x slower)... But Py_INCREF/DECREF are always inlined, even with -O0, when using __attribute__((always_inline))! I will work on a change to use that.== gcc -O0 ==(gdb) disassemble Py_IncRef Dump of assembler code for function Py_IncRef:   ...   0x000000000047276a <+20>: cmpq   $0x0,-0x8(%rbp)   0x000000000047276f <+25>:je     0x47277d <Py_IncRef+39>   0x0000000000472771 <+27>:mov    -0x8(%rbp),%rax   0x0000000000472775 <+31>:mov    %rax,%rdi   0x0000000000472778 <+34>:callq  0x472523 <_Py_INCREF>   0x000000000047277d <+39>:nop   0x000000000047277e <+40>:leaveq    0x000000000047277f <+41>:retq   (gdb) disassemble Py_DecRef Dump of assembler code for function Py_DecRef:   ...   0x0000000000472794 <+20>:cmpq   $0x0,-0x8(%rbp)   0x0000000000472799 <+25>:je     0x4727b1 <Py_DecRef+49>   0x000000000047279b <+27>:mov    -0x8(%rbp),%rax   0x000000000047279f <+31>:mov    %rax,%rdx   0x00000000004727a2 <+34>:mov    $0xe1,%esi   0x00000000004727a7 <+39>:mov    $0x65c550,%edi   0x00000000004727ac <+44>:callq  0x472554 <_Py_DECREF>   0x00000000004727b1 <+49>:nop   0x00000000004727b2 <+50>:leaveq    0x00000000004727b3 <+51>:retq   == gcc -Og ==(gdb) disassemble Py_IncRef Dump of assembler code for function Py_IncRef:   0x0000000000462de2 <+0>:test   %rdi,%rdi   0x0000000000462de5 <+3>:je     0x462dfb <Py_IncRef+25>   0x0000000000462de7 <+5>:addq   $0x1,0x4bfc09(%rip)        # 0x9229f8 <_Py_RefTotal>   0x0000000000462def <+13>:mov    0x10(%rdi),%rax   0x0000000000462df3 <+17>:add    $0x1,%rax   0x0000000000462df7 <+21>:mov    %rax,0x10(%rdi)   0x0000000000462dfb <+25>:retq   (gdb) disassemble Py_DecRef Dump of assembler code for function Py_DecRef:   0x0000000000463b2e <+0>:test   %rdi,%rdi   0x0000000000463b31 <+3>:je     0x463b6f <Py_DecRef+65>   0x0000000000463b33 <+5>:sub    $0x8,%rsp   0x0000000000463b37 <+9>:subq   $0x1,0x4beeb9(%rip)        # 0x9229f8 <_Py_RefTotal>   0x0000000000463b3f <+17>:mov    0x10(%rdi),%rax   0x0000000000463b43 <+21>:sub    $0x1,%rax   0x0000000000463b47 <+25>:mov    %rax,0x10(%rdi)   0x0000000000463b4b <+29>:je     0x463b68 <Py_DecRef+58>   0x0000000000463b4d <+31>:js     0x463b54 <Py_DecRef+38>   0x0000000000463b4f <+33>:add    $0x8,%rsp   0x0000000000463b53 <+37>:retq      0x0000000000463b54 <+38>:mov    %rdi,%rdx   0x0000000000463b57 <+41>:mov    $0xe1,%esi   0x0000000000463b5c <+46>:mov    $0x5e1120,%edi   0x0000000000463b61 <+51>:callq  0x462da4 <_Py_NegativeRefcount>   0x0000000000463b66 <+56>:jmp    0x463b4f <Py_DecRef+33>   0x0000000000463b68 <+58>:callq  0x463b0c <_Py_Dealloc>   0x0000000000463b6d <+63>:jmp    0x463b4f <Py_DecRef+33>   0x0000000000463b6f <+65>:retq

I modifiedPR 10079 to add a Py_STATIC_INLINE(TYPE) macro:* Use __attribute__((always_inline)) with GCC and clang* Use __forceinline with MSVCTests on Linux, example:"./configure --with-pydebug CC=clang CFLAGS="-O0" && make clean && make platform"* Linux, gcc -O0: inlined* Linux, clang -O0: inlinedTest done on Fedora 28 with GCC 8.1.1 and clang 6.0.1.

On Windows, a Debug build doesn't inline Py_INCREF/DECREF even if it uses __forceinline. I looked at the Py_IncRef() and Py_DecRef() assembly in Visual Studio using a breakpoint.Using /Ob1, Py_INCREF/DECREF are inlined as expected. I set this option in the pythoncore project.Do you think that I should modify the 38 other projects of the pcbuild solution to enable /Ob1 in debug build?Documentations.Inline Functions (C++):https://docs.microsoft.com/en-us/cpp/cpp/inline-functions-cpp?view=vs-2017-Od: disable optimization ("d" stands for Debug)https://msdn.microsoft.com/en-us/library/aafb762y.aspx/Ob (Inline Function Expansion):https://msdn.microsoft.com/en-us/library/47238hez.aspx

Ok, I confirm that Py_XINCREF() is properly inlined in Py_IncRef() with the latest version of myPR 10079.I tested:* gcc -O0* clang -O0* MSVC: x64 build in Debug mode

Is it guarantied that static inline functions will be inlined and will be not called as external functions from the Python library? The latter would break binary compatibility of extensions compiled with newer Python headers with older binary Python libraries.

> Is it guarantied that static inline functions will be inlined and will be not called as external functions from the Python library? The latter would break binary compatibility of extensions compiled with newer Python headers with older binary Python libraries.First, I'm not sure that the "stable ABI" really works in practice, there are many issues:https://pythoncapi.readthedocs.io/Converting the macro to a static inline function is a minor step in the direction of a more stable API and ABI.To come back to your question: depending on the compiler and compiler options, Py_INCREF/DECREF() can generate a *function call*. I tuned Py_STATIC_INLINE() to always inline Py_INCREF/DECREF() for GCC, clang and MSVC which are the major compilers used by Python on Windows, Linux, macOS and FreeBSD."static inline" is still something new to me. Maybe someone will come with a compiler with which it doesn't work as expected. IMHO we have to go through these issues, and it's only be testing for real that we will see these issues.I mean, we *have to* get ride of these ugly macros used in Python header files. They are causing subtle bugs and are hard to maintain. See my first message for advantages of static inline functions.

> Is it guarantied that static inline functions will be inlined and will be not called as external functions from the Python library? The latter would break binary compatibility of extensions compiled with newer Python headers with older binary Python libraries.Ok, I made more checks. In short,PR 10079 has no impact on the ABI compatibility.I modified Py_STATIC_INLINE() to remove __attribute__((always_inline)).In this case, the ./python binary contains multiple "instances" (what's the correct name for that?) of the "_Py_INCREF" inline function:---vstinner@apu$ readelf -sW ./python | c++filt -t |grep -E '(INC|DEC)REF'    42: 000000000041f908    49 FUNC    LOCAL  DEFAULT   13 _Py_INCREF    43: 000000000041f939   121 FUNC    LOCAL  DEFAULT   13 _Py_DECREF   109: 00000000004234cb    49 FUNC    LOCAL  DEFAULT   13 _Py_INCREF   ...  5639: a486d    49 FUNC    LOCAL  DEFAULT   13 _Py_INCREF  5786: a7abc    49 FUNC    LOCAL  DEFAULT   13 _Py_INCREF  5801: a7f0f    49 FUNC    LOCAL  DEFAULT   13 _Py_INCREF  ...  8126: 00000000006011c5    49 FUNC    LOCAL  DEFAULT   13 _Py_INCREF  8127: 00000000006011f6   121 FUNC    LOCAL  DEFAULT   13 _Py_DECREF  8140: 0000000000601971    49 FUNC    LOCAL  DEFAULT   13 _Py_INCREF  8141: 00000000006019a2   121 FUNC    LOCAL  DEFAULT   13 _Py_DECREF  ...---These functions are *LOCAL*, I understand that they are not exported.I also checked the _struct module:vstinner@apu$ readelf -sW build/lib.linux-x86_64-3.8-pydebug/_struct.cpython-38dm-x86_64-linux-gnu.so | c++filt -t |grep -E '(INC|DEC)REF'    40: 0000000000002e99    55 FUNC    LOCAL  DEFAULT   11 _Py_INCREF    41: 0000000000002ed0   127 FUNC    LOCAL  DEFAULT   11 _Py_DECREFAgain, these functions are "LOCAL", not imported nor exported.I undertand that _struct.so doesn't depend on libpython for INCREF/DECREF: it contains its own "implementation".

Interesting article on inline/static inline and symbols:https://gist.github.com/htfy96/50308afc11678d2e3766a36aa60d5f75

New changeset18618e652c56e61a134e596b315a13c7cb997a89 by Victor Stinner in branch 'master':bpo-35059: Add Py_STATIC_INLINE() macro (GH-10093)https://github.com/python/cpython/commit/18618e652c56e61a134e596b315a13c7cb997a89

Why do we need this Py_STATIC_INLINE macro? If you want to have one for enabling always inline, that's fine with me, since it's compiler-specific. But the current Py_STATIC_INLINE macro seems to conflate linkage with always-inline behavior.

Benjamin:> Why do we need this Py_STATIC_INLINE macro? If you want to have one for enabling always inline, that's fine with me, since it's compiler-specific.For about the name, there is already Py_LOCAL_INLINE which also uses "static inline", but has a very different usage: "Py_LOCAL can be used instead of static to get the fastest possible calling convention for functions that are local to a given module." So I chise "Py_STATIC_INLINE" name, it describes the basic implementation ("static inline TYPE").Py_STATIC_INLINE() is designed to replace a preprocessor macro with a function, when you care that the code is "always inlined". (Maybe the name is not perfect ;-))Honestly, I'm not sure that it really matters that the function is "always" inlined. The main issue is that "static" requires to duplicate the function in each file which uses the macro, when the function cannot/is not inlined.Previously, you asked me:> Does this slow down debug builds at all? It probably will not end will if Py_INCREF is ever not inlined.That's why I wrote Py_STATIC_INLINE() to "force" inlining andPR 10094 to enable inlining for Debug build on MSVC.> But the current Py_STATIC_INLINE macro seems to conflate linkage with always-inline behavior.I'm not sure that I get it. Do you talk about "static" inside the macro?bpo-33407 modifies Py_DEPRECATED() to support Visual Stuido, but it requires to modify how the macro is used: it now must be used at the start, rather than at the end:https://github.com/python/cpython/pull/8980/filesI chose to put "static" and "inline" in the same macro. We already have many other similar macros like PyAPI_FUNC(TYPE) and Py_LOCAL(TYPE). I would like to have a common way to "declare a function behaving as a macro".Please see also the discussion on the PR itself, Neil discuss what's the best way to declare an inline function:https://github.com/python/cpython/pull/10079#issuecomment-433230587By the way, was it you who required "static inline" support inPEP 7? :-)

New changesetb4435e20a92af474f117b78b98ddc6f515363af5 by Victor Stinner in branch 'master':bpo-35059: Convert PyObject_INIT() to function (GH-10077)https://github.com/python/cpython/commit/b4435e20a92af474f117b78b98ddc6f515363af5

New changeseta05bef4f5be1bcd0df63ec0eb88b64fdde593a86 by Victor Stinner in branch 'master':bpo-35059, PCbuild: Expand inline funcs in Debug (GH-10094)https://github.com/python/cpython/commit/a05bef4f5be1bcd0df63ec0eb88b64fdde593a86

>bpo-35059, PCbuild: Expand inline funcs in Debug (GH-10094)>https://github.com/python/cpython/commit/a05bef4f5be1bcd0df63ec0eb88b64fdde593a86Too bad: this change broke the compilation on AMD64 Windows7 SP1 3.x, AMD64 Windows8 3.x and AMD64 Windows10 3.x:https://github.com/python/cpython/pull/10094#issuecomment-433405364I don't understand why: it works well on my Windows 10 VM with my up to date Visual Studio Community 2017 (version 15.8.8). The compilation also worked on AppVeyor.For me, the only explanation is that buildbots use older compiler versions.Note: I checked that _decimal is compiled on my VM.

New changeset3b1cba3701fd1321a9bdafa9e683f891369f0cfd by Victor Stinner in branch 'master':bpo-35059, libmpdec: Add missing EXTINLINE in mpdecimal.h (GH-10128)https://github.com/python/cpython/commit/3b1cba3701fd1321a9bdafa9e683f891369f0cfd

> Too bad: this change broke the compilation on AMD64 Windows7 SP1 3.x, AMD64 Windows8 3.x and AMD64 Windows10 3.x: (...)I checked buildbots: myPR 10128 fixed all Windows buildbots, good.

New changeset95cfb818eaffba41333d4bc93253f4e0c6237ca8 by Miss Islington (bot) in branch '3.7':bpo-35059, libmpdec: Add missing EXTINLINE in mpdecimal.h (GH-10128)https://github.com/python/cpython/commit/95cfb818eaffba41333d4bc93253f4e0c6237ca8

New changeset7eac88afd2e39d05a0ed3bc8c0787a2e755a6072 by Miss Islington (bot) in branch '3.6':bpo-35059, libmpdec: Add missing EXTINLINE in mpdecimal.h (GH-10128)https://github.com/python/cpython/commit/7eac88afd2e39d05a0ed3bc8c0787a2e755a6072

See alsobpo-35081: "RenameInclude/internals/ toInclude/pycore/" which is linked to this issue.

On Fri, Oct 26, 2018, at 03:18, STINNER Victor wrote:> > STINNER Victor <vstinner@redhat.com> added the comment:> > Benjamin:> > Why do we need this Py_STATIC_INLINE macro? If you want to have one for enabling always inline, that's fine with me, since it's compiler-specific.> > For about the name, there is already Py_LOCAL_INLINE which also uses > "static inline", but has a very different usage: "Py_LOCAL can be used > instead of static to get the fastest possible calling convention for > functions that are local to a given module." So I chise > "Py_STATIC_INLINE" name, it describes the basic implementation ("static > inline TYPE").I would like to see Py_LOCAL_INLINE removed, too, fwiw.> > Py_STATIC_INLINE() is designed to replace a preprocessor macro with a > function, when you care that the code is "always inlined". (Maybe the > name is not perfect ;-))"always inline" is different from "static inline". So, it's not appropriate to make a macro named the latter that has the former former.> > Honestly, I'm not sure that it really matters that the function is > "always" inlined. The main issue is that "static" requires to duplicate > the function in each file which uses the macro, when the function > cannot/is not inlined.> > > Previously, you asked me:> > > Does this slow down debug builds at all? It probably will not end will if Py_INCREF is ever not inlined.> > That's why I wrote Py_STATIC_INLINE() to "force" inlining andPR 10094 > to enable inlining for Debug build on MSVC.> > > > But the current Py_STATIC_INLINE macro seems to conflate linkage with always-inline behavior.> > I'm not sure that I get it. Do you talk about "static" inside the macro? >bpo-33407 modifies Py_DEPRECATED() to support Visual Stuido, but it > requires to modify how the macro is used: it now must be used at the > start, rather than at the end:>https://github.com/python/cpython/pull/8980/files> > I chose to put "static" and "inline" in the same macro. We already have > many other similar macros like PyAPI_FUNC(TYPE) and Py_LOCAL(TYPE). I > would like to have a common way to "declare a function behaving as a > macro".We don't want functions that behave like macros... otherwise, we would be writing macros. If we want a function to always be inlined, we should explicitly request that from the compiler.> > > Please see also the discussion on the PR itself, Neil discuss what's the > best way to declare an inline function:>https://github.com/python/cpython/pull/10079#issuecomment-433230587> > By the way, was it you who required "static inline" support inPEP 7? :-)Yes, which is why we shouldn't need a macro to write it.

> I would like to see Py_LOCAL_INLINE removed, too, fwiw.Oh. Why? Do you want to directly use "static" and "static inline"? I guess that Py_LOCAL and Py_LOCAL_INLINE have been added to use __fastcall with MSVC.... __fastcall is mostly interesting in x86 (32-bit), but x86-64 calling convention is "fast" by default no? __fastcall pass the first two arguments in registers, but x86-64 already pass the first four arguments in registers...Do you mean that __fastcall is no longer revelant?

>> Py_STATIC_INLINE() is designed to replace a preprocessor macro with a >> function, when you care that the code is "always inlined". (Maybe the >> name is not perfect ;-))>> "always inline" is different from "static inline". So, it's not appropriate to make a macro named the latter that has the former former.Oh ok. So if we decide to keep it, it should be renamed to Py_STATIC_ALMOST_ALWAYS_INLINE() or something like that :-)> We don't want functions that behave like macros... otherwise, we would be writing macros. If we want a function to always be inlined, we should explicitly request that from the compiler.Oh. It seems like I misunderstood you.I understood that you required to have zero impact on performance on debug build.*I* want to use functions because it's the regular C language: regular scope rules, no preprocessor magic, the compiler detects errors if the function is misused, etc.But I'm not sure about the drawbacks of converting a macro to a function. I don't want to be the only one responsible to regressions :-) If you support the change, we can drop "__attribute__((always_inline))" and use "regular" "static inline" functions :-)>> By the way, was it you who required "static inline" support inPEP 7? :-)>> Yes, which is why we shouldn't need a macro to write it.Ok ok.I updated myPR 10079 to simply use "static inline".

If myPR 10079 is merged without Py_STATIC_INLINE(), I will remove the macro.I'm not sure if we need an "always inline" macro or not.Note: I'm in favor of moving closer to the C language and not abusing __attribute__(...) :-)

New changeset2aaf0c12041bcaadd7f2cc5a54450eefd7a6ff12 by Victor Stinner in branch 'master':bpo-35059: Convert Py_INCREF() to static inline function (GH-10079)https://github.com/python/cpython/commit/2aaf0c12041bcaadd7f2cc5a54450eefd7a6ff12

New changeset542497aa9f71c664768c3d5b7398c03679d3a7e1 by Victor Stinner in branch 'master':bpo-35059: Remove Py_STATIC_INLINE() macro (GH-10216)https://github.com/python/cpython/commit/542497aa9f71c664768c3d5b7398c03679d3a7e1

>bpo-35059: Remove Py_STATIC_INLINE() macro (GH-10216)Here you have Benjamin, it's gone :-)

TODO: Convert _PyObject_GC_TRACK() and _PyObject_GC_UNTRACK() macros to static inline functions, but there are inter-dependencies issues: seebpo-35081 andhttps://mail.python.org/pipermail/python-dev/2018-October/155592.html

New changeset541497e6197268517b0d492856027774c43e0949 by Victor Stinner in branch 'master':bpo-35059: Convert Py_XINCREF() to static inline function (GH-10224)https://github.com/python/cpython/commit/541497e6197268517b0d492856027774c43e0949

New changeset3c09dca4b5de9fe8c8756251d02f49cf093b88c1 by Victor Stinner in branch 'master':bpo-35059: Convert _Py_Dealloc() to static inline function (GH-10223)https://github.com/python/cpython/commit/3c09dca4b5de9fe8c8756251d02f49cf093b88c1

I am collecting information on how different compilers behave with the `static inline`. I am checking the following compilers:ARM GCCAVR GCCCLANG X86_64CLANG RISC-VELLCCGCC X86_64ICC x86_64MIPS GCCMSP GCCPOWERPC GCCAIX Compiler (xlc)SunPro Compilerswill report back when I have enough information on these.

See alsobpo-35230: "Remove _Py_REF_DEBUG_COMMA".

New changesetf1d002c1e094922b0f17a820f90ff102d68ab253 by Victor Stinner in branch 'master':bpo-35059: Enhance _PyObject_AssertFailed() (GH-10642)https://github.com/python/cpython/commit/f1d002c1e094922b0f17a820f90ff102d68ab253

New changeset271753a27aca2e13275f0827080b915fb438107a by Victor Stinner in branch 'master':bpo-35059: Convert _PyObject_GC_TRACK() to inline function (GH-10643)https://github.com/python/cpython/commit/271753a27aca2e13275f0827080b915fb438107a

New changeset2ff8fb7639a86757c00a7cbbe7da418fffec3870 by Victor Stinner in branch 'master':bpo-35059: Add _PyObject_CAST() macro (GH-10645)https://github.com/python/cpython/commit/2ff8fb7639a86757c00a7cbbe7da418fffec3870

New changesetb37672daf61740fe1ff9d805f6d74bc5ef04012b by Victor Stinner in branch 'master':bpo-35059: Cleanup usage of Python macros (GH-10648)https://github.com/python/cpython/commit/b37672daf61740fe1ff9d805f6d74bc5ef04012b

New changeseta42de742e7c20eeb64699b5785543fea65b2e8d3 by Victor Stinner in branch 'master':bpo-35059: Cast void* to PyObject* (GH-10650)https://github.com/python/cpython/commit/a42de742e7c20eeb64699b5785543fea65b2e8d3

I wrote a pull request to replace static inline functions with C macros:PR 10669. I ran a benchmark on speed.python.org server using the "performance" benchmark suite:http://pyperformance.readthedocs.io/I understand that from the benchmark results that converting macros to static inline functions has no significant impact on performances. Two benchmarks are 1.06x and 1.07x faster but it can be explained by the PGO compilation which is not reliable. One benchmark is way slower, but it seems like the benchmark has an issue. If you look at the 3 latest run on speed.python.org, I see:* 38.2 us (Sept 24)* 43.5 us (Nov 21)* 43.7 us (Nov 22)I don't think that any change in _pickle or pickle explains this significant slowdown. IMHO it's just that the benchmark is not reliable :-/ We have a performance timeline on the last 5 years, and this benchmark doesn't have a straight line, we can see that the result is a little bit random :-/--speed.python.org runs Ubuntu 16.04 with gcc 5.4.0.The result are the two attached (compressed) JSON files:* 2018-11-22_17-38-master-3bb183d7fb83-patch-10669.json.gz: reference, Python using C macros* 2018-11-22_17-38-master-3bb183d7fb83.json.gz: static inline, current master branchComparison ignoring difference smaller than -5% and +5%, macros are the reference:vstinner@apu$ python3 -m perf compare_to --table -G --min-speed=5 macros.json.gz inline.json.gz +--------------+---------+------------------------------+| Benchmark    | macros  | inline                       |+==============+=========+==============================+| regex_dna    | 288 ms  | 269 ms: 1.07x faster (-7%)   |+--------------+---------+------------------------------+| crypto_pyaes | 236 ms  | 223 ms: 1.06x faster (-5%)   |+--------------+---------+------------------------------+| pickle_dict  | 37.8 us | 43.7 us: 1.16x slower (+16%) |+--------------+---------+------------------------------+Not significant (54): (...)Raw comparison, full data, macros are the reference:vstinner@apu$ python3 -m perf compare_to --table -G macros.json.gz inline.json.gz +-------------------------+----------+------------------------------+| Benchmark               | macros   | inline                       |+=========================+==========+==============================+| regex_dna               | 288 ms   | 269 ms: 1.07x faster (-7%)   |+-------------------------+----------+------------------------------+| crypto_pyaes            | 236 ms   | 223 ms: 1.06x faster (-5%)   |+-------------------------+----------+------------------------------+| nqueens                 | 199 ms   | 195 ms: 1.02x faster (-2%)   |+-------------------------+----------+------------------------------+| raytrace                | 1.01 sec | 984 ms: 1.02x faster (-2%)   |+-------------------------+----------+------------------------------+| chaos                   | 224 ms   | 221 ms: 1.02x faster (-2%)   |+-------------------------+----------+------------------------------+| logging_simple          | 21.2 us  | 20.9 us: 1.01x faster (-1%)  |+-------------------------+----------+------------------------------+| unpack_sequence         | 117 ns   | 116 ns: 1.01x faster (-1%)   |+-------------------------+----------+------------------------------+| spectral_norm           | 247 ms   | 244 ms: 1.01x faster (-1%)   |+-------------------------+----------+------------------------------+| regex_v8                | 41.8 ms  | 41.4 ms: 1.01x faster (-1%)  |+-------------------------+----------+------------------------------+| pickle                  | 19.9 us  | 19.7 us: 1.01x faster (-1%)  |+-------------------------+----------+------------------------------+| logging_format          | 24.0 us  | 23.8 us: 1.01x faster (-1%)  |+-------------------------+----------+------------------------------+| scimark_sparse_mat_mult | 9.06 ms  | 8.99 ms: 1.01x faster (-1%)  |+-------------------------+----------+------------------------------+| pathlib                 | 42.1 ms  | 41.8 ms: 1.01x faster (-1%)  |+-------------------------+----------+------------------------------+| meteor_contest          | 188 ms   | 187 ms: 1.01x faster (-1%)   |+-------------------------+----------+------------------------------+| unpickle_pure_python    | 704 us   | 700 us: 1.01x faster (-1%)   |+-------------------------+----------+------------------------------+| python_startup          | 12.0 ms  | 12.0 ms: 1.00x faster (-0%)  |+-------------------------+----------+------------------------------+| python_startup_no_site  | 7.91 ms  | 7.94 ms: 1.00x slower (+0%)  |+-------------------------+----------+------------------------------+| sympy_integrate         | 35.0 ms  | 35.3 ms: 1.01x slower (+1%)  |+-------------------------+----------+------------------------------+| pidigits                | 283 ms   | 285 ms: 1.01x slower (+1%)   |+-------------------------+----------+------------------------------+| hexiom                  | 17.3 ms  | 17.4 ms: 1.01x slower (+1%)  |+-------------------------+----------+------------------------------+| sympy_str               | 403 ms   | 407 ms: 1.01x slower (+1%)   |+-------------------------+----------+------------------------------+| mako                    | 31.9 ms  | 32.2 ms: 1.01x slower (+1%)  |+-------------------------+----------+------------------------------+| django_template         | 293 ms   | 296 ms: 1.01x slower (+1%)   |+-------------------------+----------+------------------------------+| tornado_http            | 368 ms   | 372 ms: 1.01x slower (+1%)   |+-------------------------+----------+------------------------------+| sqlalchemy_declarative  | 287 ms   | 290 ms: 1.01x slower (+1%)   |+-------------------------+----------+------------------------------+| html5lib                | 190 ms   | 192 ms: 1.01x slower (+1%)   |+-------------------------+----------+------------------------------+| xml_etree_iterparse     | 181 ms   | 183 ms: 1.01x slower (+1%)   |+-------------------------+----------+------------------------------+| 2to3                    | 601 ms   | 610 ms: 1.02x slower (+2%)   |+-------------------------+----------+------------------------------+| unpickle_list           | 6.49 us  | 6.60 us: 1.02x slower (+2%)  |+-------------------------+----------+------------------------------+| scimark_monte_carlo     | 210 ms   | 214 ms: 1.02x slower (+2%)   |+-------------------------+----------+------------------------------+| sympy_sum               | 198 ms   | 202 ms: 1.02x slower (+2%)   |+-------------------------+----------+------------------------------+| telco                   | 14.8 ms  | 15.1 ms: 1.02x slower (+2%)  |+-------------------------+----------+------------------------------+| fannkuch                | 855 ms   | 874 ms: 1.02x slower (+2%)   |+-------------------------+----------+------------------------------+| sympy_expand            | 857 ms   | 883 ms: 1.03x slower (+3%)   |+-------------------------+----------+------------------------------+| scimark_lu              | 298 ms   | 307 ms: 1.03x slower (+3%)   |+-------------------------+----------+------------------------------+| xml_etree_generate      | 207 ms   | 213 ms: 1.03x slower (+3%)   |+-------------------------+----------+------------------------------+| float                   | 210 ms   | 216 ms: 1.03x slower (+3%)   |+-------------------------+----------+------------------------------+| xml_etree_process       | 173 ms   | 180 ms: 1.04x slower (+4%)   |+-------------------------+----------+------------------------------+| scimark_sor             | 357 ms   | 373 ms: 1.05x slower (+5%)   |+-------------------------+----------+------------------------------+| json_dumps              | 25.8 ms  | 27.0 ms: 1.05x slower (+5%)  |+-------------------------+----------+------------------------------+| pickle_dict             | 37.8 us  | 43.7 us: 1.16x slower (+16%) |+-------------------------+----------+------------------------------+Not significant (16): regex_effbot; go; json_loads; xml_etree_parse; dulwich_log; nbody; regex_compile; scimark_fft; pickle_pure_python; unpickle; pickle_list; sqlite_synth; richards; logging_silent; deltablue; sqlalchemy_imperative

New changesete89607c0fc8d6edbf19c06ba42ff0f00e6c4273f by Victor Stinner in branch 'master':bpo-35059: NEWS entry for macros converted to inline funcs (GH-10671)https://github.com/python/cpython/commit/e89607c0fc8d6edbf19c06ba42ff0f00e6c4273f

I ran a coarse benchmark on the debug mode usingPR 10669: I ran the full Python test suite.=> I don't see any significant impact on performance.* C macros:PR 10669* static inline functions: commit3bb183d7fb83ad6a84ec13dea90f95d67be35c69 (PR 10669 parent)git clean -fdx # remove *all* untracked files./configure --with-pydebug && make./python -m test -j0 -r --randseed=2411717Result:* C macros: 3 min 22 sec* static inline functions: 3 min 26 sec (+4 sec, +2%)I used my laptop to run the benchmark: Fedora 29 with GCC 8.2.1. I was still using the laptop to run other tasks.

Unexpected cool effect of static inline functions (copy of my email):https://mail.python.org/pipermail/python-dev/2018-November/155747.htmlLe jeu. 15 nov. 2018 à 01:06, Gregory P. Smith <greg at krypto.org> a écrit :> I expect the largest visible impact may be that a profiler may now attribute CPU cycles takes by these code snippets to the function from the .h file rather than directly to the functions the macro expanded in in the past due to additional debug symbol info attribution. Just more data. Consider that a win.Oh. That's very interesting.I just tried gdb and I confirm that gdb understands well inlinedfunction. When I debug Python, gdb moves into Py_INCREF() orPy_DECREF() when I use "next".I also tried perf record/perf report: if I annotate a function(assembler code of the function), perf shows me the C code of inlinedPy_INCREF and Py_DECREF!That's nice!Victor

Stefan Behnel wrote:https://mail.python.org/pipermail/python-dev/2018-November/155759.html"""It's also slower to compile, given that function inlining happens at a muchlater point in the compiler pipeline than macro expansion. The C compilerwon't even get to see macros in fact, whereas whether to inline a functionor not is a dedicated decision during the optimisation phase based onmetrics collected in earlier stages. For something as ubiquitous asPy_INCREF/Py_DECREF, it might even be visible in the compilation times."""I ran a benchmark on the compilation time usingPR 10669: there is no significant slowdown (+4 seconds, 6% slower, in the worst case).I ran a benchmark on my laptop:* Fedora 29* Intel(R) Core(TM) i7-6820HQ CPU @ 2.70GHz: 4 physical cores (8 threads)* GCC 8.2.1* MAKEFLAGS=-j9Result in release mode:* git clean -fdx; ./configure; time make # -03* C macros: 1m12,158s* static inline functions: 1m16,294s (+4.136s, +6%)Result in debug mode:* git clean -fdx; ./configure --with-pydebug; time make # -Og* C macros: 0m39,727s* static inline functions: 0m40,423s (+0.696s, +2%)I only used "real" time (I ignored user and sys times).

"PyObject* PyObject_INIT(PyObject *op, PyTypeObject *typeobj)"and"PyVarObject* PyObject_INIT_VAR(PyVarObject *op, PyTypeObject *typeobj, Py_ssize_t size)"Don't cast their argument to PyObject*/PyVarObject* which can introduce new warnings when upgrading from Python 3.7 to Python 3.8.

I ran my benchmarks, I close thePR 10669 (replace static inline functions with macros). You should still be able to use it as patch:https://github.com/python/cpython/pull/10669.patch

New changesetb509d52083e156f97d6bd36f2f894a052e960f03 by Victor Stinner in branch 'master':bpo-35059: PyObject_INIT() casts to PyObject* (GH-10674)https://github.com/python/cpython/commit/b509d52083e156f97d6bd36f2f894a052e960f03

> Drawbacks: Require a specific type can introduce compiler warnings if the caller doesn't pass the proper type (PyObject* or PyVarObject*). _Py_NewReference() and _Py_ForgetReference() seem to be properly used, but not PyObject_INIT() and PyObject_INIT_VAR().I worked around this issue by adding a macro to cast the argument and declare the static inline function with a different name. Example:static inline void _Py_INCREF(PyObject *op){    _Py_INC_REFTOTAL;    op->ob_refcnt++;}#define Py_INCREF(op) _Py_INCREF(_PyObject_CAST(op))

I close the issue, it seems like all subtasks have been completed.Summary of the issue:* The following macros have been converted to static inline functions:  - Py_INCREF(), Py_DECREF()  - Py_XINCREF(), Py_XDECREF()  - PyObject_INIT(), PyObject_INIT_VAR()  - _Py_NewReference(), _Py_ForgetReference()  - _Py_Dealloc()  - _PyObject_GC_TRACK(), _PyObject_GC_UNTRACK()* There is no significant impact on performance:  * I ran performance benchmark on Python compiled in release mode  * I ran the Python test suite on Python compiled in debug mode  * I measured the compilation time in release an debug mode* It has been decided to use "static inline" to declare inline function (write directly "static inline", no macro).* It has been decided to no use __attribute__((always_inline)) nor __forceinline (ask the compiler to always inline).--Benjamin Peterson would "like to see Py_LOCAL_INLINE removed, too, fwiw", but it's part of the public C API. It would require a deprecation period. I'm not interested to touch the public C API.Benjamin: please open a new issue if you still want to remove it :-)

Advantages of inline functions over C macros:https://mail.python.org/pipermail/python-dev/2018-November/155805.htmlThere are multiple advantages:* Better development and debugging experience: tools understandinlined functions much better than C macros: gdb, Linux perf, etc.* Better API: arguments now have a type and the function has a returntype. In practice, some macros still cast their argument to PyObject*to not introduce new compiler warnings in Python 3.8. For example,even if Py_INCREF() is documented (*) as a function expectingPyObject*, it accepts any pointer type (PyTupleObject*,PyUnicodeObject*, etc.). Technically, it also accepts PyObject** whichis a bug, but that's a different story ;-)* Much better code, just plain regular C. C macros are ugly: "do { ...} while (0)" workaround, additional parenthesis around each argument,strange "expr1, expr2" syntax of "macro expression" which returns avalue (inline function just uses regular "return" and ";" at the endof instructions), strange indentation, etc.* No more "macro pitfals":https://gcc.gnu.org/onlinedocs/cpp/Macro-Pitfalls.html* Local variables no longer need a magic name to avoid risk of nameconflict, and have a clearly defined scope. Py_DECREF() and_Py_XINCREF() no longer need a local variable since it's argumentalready has a clearly defined type: PyObject*. I introduced a newvariable in _Py_Dealloc() to fix a possible race condition.Previously, the variable was probably avoided because it's tricky usevariables in macros.* #ifdef can now be used inside the inline function: it makes the codeeasier to understand.* etc.Are you aware that Python had macros like:#define _Py_REF_DEBUG_COMMA ,#define _Py_CHECK_REFCNT(OP) /* a semicolon */;I let you judge the quality of this macro:#define _Py_NewReference(op) (                          \    _Py_INC_TPALLOCS(op) _Py_COUNT_ALLOCS_COMMA         \    _Py_INC_REFTOTAL  _Py_REF_DEBUG_COMMA               \    Py_REFCNT(op) = 1)Is it an expression? Can it be used in "if (test)_Py_NewReference(op);"? It doesn't use the "do { ... } while (0)"protection against macro pitfals.(*) Py_INCREF doc:https://docs.python.org/dev/c-api/refcounting.html#c.Py_INCREF

Could we have used function overloading to handle the different types? Rather than reintroducing the macro for the sake of the cast?

> Could we have used function overloading to handle the different types?Rather than reintroducing the macro for the sake of the cast?Sorry, I don't know what is function overloading. Is it a C++ thing?Py_INCREF() must accept any type based on PyObject.At least, this issue shouldn't make the situation worse :-)Please open a new issue if you have a solution for this problem. I am nowcurious since I tried many things and I failed to find anything working forall cases.

Ah, you're right, it's only C++. My bad.

New changeset07923f32b16ba39165a58a5f47e807ca04ae17aa by Victor Stinner in branch 'master':bpo-35059: Enhance _PyObject_GC_TRACK() macros (GH-20931)https://github.com/python/cpython/commit/07923f32b16ba39165a58a5f47e807ca04ae17aa

Seebpo-45094: "Consider using __forceinline and __attribute__((always_inline)) on static inline functions (Py_INCREF, Py_TYPE) for debug builds".