[Bug middle-end/68762] link error for inline function decorated with OpenMP declare simd
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68762 Jakub Jelinek changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED --- Comment #7 from Jakub Jelinek --- Fixed.
[Bug middle-end/68762] link error for inline function decorated with OpenMP declare simd
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68762 Martin Liška changed: What|Removed |Added CC||marxin at gcc dot gnu.org --- Comment #6 from Martin Liška --- Jakub: Can the bug be marked as resolved?
[Bug middle-end/68762] link error for inline function decorated with OpenMP declare simd
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68762 --- Comment #5 from Jakub Jelinek --- Author: jakub Date: Mon Aug 8 19:45:53 2016 New Revision: 239248 URL: https://gcc.gnu.org/viewcvs?rev=239248=gcc=rev Log: PR middle-end/68762 * omp-simd-clone.c: Include varasm.h. (simd_clone_create): Copy over DECL_COMDAT, DECL_WEAK, DECL_EXTERNAL, DECL_VISIBILITY, DECL_VISIBILITY_SPECIFIED, DECL_DLLIMPORT_P and for DECL_ONE_ONLY call make_decl_one_only. Fix up spelling in comment and update function name. * g++.dg/vect/pr68762-1.cc: New test. * g++.dg/vect/pr68762-2.cc: New test. * g++.dg/vect/pr68762.h: New file. Added: trunk/gcc/testsuite/g++.dg/vect/pr68762-1.cc trunk/gcc/testsuite/g++.dg/vect/pr68762-2.cc trunk/gcc/testsuite/g++.dg/vect/pr68762.h Modified: trunk/gcc/ChangeLog trunk/gcc/omp-simd-clone.c trunk/gcc/testsuite/ChangeLog
[Bug middle-end/68762] link error for inline function decorated with OpenMP declare simd
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68762 Jakub Jelinek changed: What|Removed |Added Status|UNCONFIRMED |ASSIGNED Last reconfirmed||2016-08-08 Assignee|unassigned at gcc dot gnu.org |jakub at gcc dot gnu.org Ever confirmed|0 |1 --- Comment #4 from Jakub Jelinek --- Created attachment 39072 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=39072=edit gcc7-pr68762.patch Untested fix. That said, using #pragma omp declare simd is usually quite pointless on inline functions, those are usually inlined and then there is no point in the declaration.
[Bug middle-end/68762] link error for inline function decorated with OpenMP declare simd
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68762 Andrew Pinski changed: What|Removed |Added Target||x86_64-*-* --- Comment #3 from Andrew Pinski --- _ZGVbN2v__Z3food should have been in its own comdat section. Note on aarch64-linux-gnu I don't hit this bug but that is because aarch64 does not have a simd version of exp, cos or sin (we should be that a different story).
[Bug middle-end/68762] link error for inline function decorated with OpenMP declare simd
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68762 --- Comment #1 from Richard Biener --- You need to provide a out-of-line copy of foo for example by adding a extern double foo (double d); to foobar.cpp. At least if this were C, not 100% sure about C++. OTOH SIMD support may simply fail to provide proper linkage to the clones.
[Bug middle-end/68762] link error for inline function decorated with OpenMP declare simd
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68762 --- Comment #2 from gilles --- TBH I don't know anything about the C inline keyword, aside from it existence. My issue has definitely to do with the C++ inline keyword. And the important point I (hope I) know about it is that the full definition of the function should be available in every compilation unit where the function is used. This is commonly done by putting the definition of the function into a header file. Another very important point is that the inline keyword is a suggestion given to the compiler, which it is at liberty to follow or to ignore. Therefore, an inline function can (should?) also have an external linkage. But since it is compiled as many times as there are of compilation units where it was defined, all its identical binary definitions have somehow to be "merge" at link time. I must admit that this part is a bit of black magic to me, but I seen that the use of weak symbols into the object files does the trick. Now about the "omp declare simd" attribute: nowhere in the OpenMP standard is it mentioned that an inline function cannot be declared as such. Therefore, considering that some explicit restrictions on the type of functions that can be are listed, it seems fair to assume that inline functions (as not listed as unfit) can be declared simd. And actually, so long as the simd-inline function get inlned by the complier, all goes well even with GCC (it works in any case with the Intel compler BTW). But as soon as, for a reason or another, the function isn't inlined, the compilation fails at the linking stage with a "multiple definition" error. And it looks like the reason is that the symbols for the vectorised versions of the out-of-line inline function are strong symbols instead of weak symbols. So my guess, and please forgive me if what I say is stupid, is that these symbols would only need to be defined weak for the bug to get fixed. I hope this makes (even a little) sense since I feel I'm playing way off my league here.
