[issue45476] [C API] Disallow using PyFloat_AS_DOUBLE() as l-value

[Marc-Andre Lemburg]

Marc-Andre Lemburg <mal at egenix.com> added the comment:

On 15.11.2021 10:54, STINNER Victor wrote:
> 
> I don't understand what you are trying to prove about compilers not inlining when you explicitly ask them... not to inline.

I'm not trying to prove anything, Victor.

I'm only stating the fact that by switching from macros to inline
functions we are giving away control to the compilers and should not
be surprised that Python now suddenly runs a lot slower on systems
which either have inlining optimizations switched off or where the
compiler (wrongly) assumes that creating more assembler would result
in slower code.

I've heard all your arguments against macros, but don't believe the
blanket approach to convert to inline functions is warranted in all
cases, in particular not for code which is private to the interpreter
and where we know that we need the code inlined to not result in
unexpected performance regressions.

I also don't believe that we should assume that Python C extension
authors will unintentionally misuse Python API macros. If they do,
it's their business to sort out any issues, not ours. If we document
that macros may not be used as l-values, that's clear enough. We don't
need to use compiler restrictions to impose such limitations.

IMO, conversion to inline functions should only happen, when

a) the core language implementation has a direct benefit, and

b) it is very unlikely that compilers will not inline the code
   with -O2 settings, e.g. perhaps using a threshold of LOCs
   or by testing with the website Oleg mentioned.

Overall, I think the PEP 670 should get some more attentions from the
SC to have a guideline to use as basis for deciding whether or not
to use the static inline function approach. That way we could avoid
these discussions :-)

BTW: Thanks for the details about -O0 vs. -Og.

----------

_______________________________________
Python tracker <report at bugs.python.org>
<https://bugs.python.org/issue45476>
_______________________________________