Should not use __attribute__((optimize(…)))

[jwilk]

GCC documentation says this about the optimize attribute:

This attribute should be used for debugging purposes only. It is not suitable in production code.

In practice, __attribute__((optimize("Os"))) seems to re-enable strict aliasing, despite -fno-strict-aliasing specified on command-line:

i686-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -Wstrict-prototypes -fno-strict-aliasing -Wdate-time -D_FORTIFY_SOURCE=2 -g -fdebug-prefix-map=/build/python2.7-5qX40H/python2.7-2.7.15=. -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I/usr/include/python2.7 -c build/temp.linux-i386-2.7/src/decode.c -o build/temp.linux-i386-2.7/build/temp.linux-i386-2.7/src/decode.o -pthread
build/temp.linux-i386-2.7/src/decode.c: In function 'initdecode':
build/temp.linux-i386-2.7/src/decode.c:51244:3: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
   __Pyx_INCREF(((PyObject *)(&PyString_Type)));
   ^~~~~~~~~~~~

^{(full build log)}

I was going to file a bug against GCC, but then I realized the attribute shouldn't be used in the first place, and there's apparently plenty of bugs about it in GCC Bugzilla already.

[scoder]

Thanks for finding that out. The cold attribute seems to have a similar effect, though:
https://gcc.gnu.org/onlinedocs/gcc-8.1.0/gcc/Common-Function-Attributes.html#index-cold-function-attribute

[charris]

@scoder Will Cython be changing the use of the attribute?

[charris]

I ask because if it is true that __attribute__((optimize("Os"))) re-enables strict aliasing, that would be a problem for NumPy.

[scoder]

Yes, this will change in 0.29. See 9ddac71.

[charris]

Do you have an estimate of the 0.29 timeline?

[scoder]

You can always define the C macro CYTHON_SMALL_CODE to nothing to disable this unconditionally.

I'd like to release it rather earlier than later, but there is no clear release timeframe. #2499 is the most important open ticket. Once that's solved, and it'll take a couple of person days to get right, I think I'd look into a release. So, expect a couple of weeks.

I'll also consider a 0.28.5 release in the meantime that would have this fix.

[mattip]

Just a FYI that numpy/numpy#11665 adopted defining CYTHON_SMALL_CODE= was merged for an upcoming NumPy 1.15.1, thanks for helping work through this.

[scoder]

Seeing that this causes actual trouble to projects, I decided to release 0.28.5 with (almost) only this fix. It's uploaded to PyPI.

[charris]

Thanks @scoder .