Fix GCC 16 LTO build failure with multi-target dispatch #225
Conversation
|
HI @stratakis thanks for reporting the issue and patching it up! Happy to merge this as soon as the CI passes. I am assuming NumPy needs an update after merging this one? |
Yes, numpy bundles it, so it will need an update. |
|
Close/open to trigger CI runs. Not sure why they didnt find a runner. |
|
Also something I'd like to point out is that this file has mixed windows line endings and linux ones (CRLF and LF). Working on Linux it was a bit of a pain, as git by default does conversions of line endings on changes. Not sure if important, but if Linux is the main target you might wanna consider converting everything to LF? |
Yes, linux is the main target and I do want to get that fixed in a separate PR. Feel free to add a commit to this PR, if you would like. Also, please rebase when #226 is merged in a few minutes. |
|
Please rebase to run the CI. |
Remove always_inline from STDSortComparator. This function is passed to std::sort, and GCC 16 rejects forced inlining when the caller and callee have mismatched target attributes during LTO. This occurs when the library is used via the header-only static include with multiple dispatch targets (e.g. AVX-512 and baseline) linked together, as NumPy does. Standalone builds are unaffected since they use -march=skylake-avx512 for the entire translation unit.
stratakis
force-pushed
the
fix_x86_64_gcc_16_build
branch
from
cb42dd6 to
4273453
Compare
|
Thanks @stratakis! |
|
2 participants
Remove always_inline from STDSortComparator. This function is passed to std::sort, and GCC 16 rejects forced inlining when the caller and callee have mismatched target attributes during LTO.
This occurs when the library is used via the header-only static include with multiple dispatch targets (e.g. AVX-512 and baseline) linked together, as NumPy does.
Standalone builds are unaffected since they use -march=skylake-avx512 for the entire translation unit.
Fixes: #224