Removing last remnants of pragma block at the top of pybind11.h #3186
Conversation
rwgk
changed the title
rwgk
force-pushed
the
wattributes_compromise
branch
from
42db2f7
to
162f9b6
Compare
…ulting CUDA, GCC7, GCC8 to PYBIND11_NOINLINE_DISABLED, with the option to define PYBIND11_NOINLINE_FORCED.
for more information, see https://pre-commit.ci
…is not expected to make a difference, but who knows what I am overlooking).
rwgk
force-pushed
the
wattributes_compromise
branch
from
5f79963
to
fa53ba5
Compare
rwgk
changed the title
| // the default under pybind11/tests). | ||
| #if !defined(PYBIND11_NOINLINE_FORCED) && \ | ||
| (defined(__CUDACC__) || (defined(__GNUC__) && (__GNUC__ == 7 || __GNUC__ == 8))) | ||
| # define PYBIND11_NOINLINE_DISABLED |
There was a problem hiding this comment.
I am guessing there is no clean way to detect if the warning is enabled is there? If there is seems like it might make sense to add it to the conditional.
There was a problem hiding this comment.
I am guessing there is no clean way to detect if the warning is enabled is there?
I don't think so, not even noinline is portable, even less so introspection of compiler-specific warnings related to it.
If there is seems like it might make sense to add it to the conditional.
If it was easy, yes.
But how likely is it that anyone cares enough about a 1.7% binary size reduction when building with CUDA? If someone does, they can add the suggested command line options. If they don't like that, they could send us a PR. — The group of affected people and the binary size saving both seem so marginal to me, I cannot imagine it's worth anyone's time, cost vs benefit.
…he comment and empty line that was added in PR pybind#3087; those were made obsolete by the pragma cleanup that concluded with PR pybind#3186.
…nt when underlying type is bool or of char type) (#3232) * Minor tweaks. * Restoring tests/pybind11_tests.h version from master, removing just the comment and empty line that was added in PR #3087; those were made obsolete by the pragma cleanup that concluded with PR #3186. * More-to-the-point test for Python 3.
Defaulting CUDA, GCC7, GCC8 to
PYBIND11_NOINLINE_DISABLED, with the option to definePYBIND11_NOINLINE_FORCED. Relevantinline/noinlineshared-library size differences are tabulated here. The measured size saving when usingnoinlineare only 1.7% for CUDA, -0.2% for GCC7, and 0.0% for GCC8 (using-DCMAKE_BUILD_TYPE=MinSizeRel, the default under pybind11/tests).Follow-on to PR #3127, based on results obtained under PR #3125.
Worksheet
This PR changes 2 files, removes 9 lines, adds 11 lines (6 of which are for a comment).
Suggested changelog entry:
CUDA, GCC7, GCC8 builds that prioritize
noinlineover-Wattributesdiagnostics need to add-DPYBIND11_NOINLINE_FORCED -Wno-attributesto theg++command lines. However, the measured shared-library size saving when usingnoinlineare only 1.7% for CUDA, -0.2% for GCC7, and 0.0% for GCC8 (using-DCMAKE_BUILD_TYPE=MinSizeRel, the default under pybind11/tests).