https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84234
--- Comment #6 from Florian Lemaitre ---
I want auto-vectorization, not guided vectorization.
Basically, I try to write a fast RSQRT that will not break auto vectorization,
and this will be integrated in a large scale project where it is easy to
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84234
--- Comment #5 from Jakub Jelinek ---
#pragma omp declare simd or simd attribute say that on the definition compiler
should create alternate entry point(s) for the vectorized variants, and on
declarations that on the definition they were created,
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84234
--- Comment #4 from Florian Lemaitre ---
Ah... ok. I understand.
Indeed, declaring the function const works as expected.
So, I guess we can say it is not a bug, and close the bug report?
However, I really think this should be documented somewhe
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84234
--- Comment #3 from Jakub Jelinek ---
The forward declaration is not ignored, but you either need to use #pragma omp
simd on the loop you want to vectorize, or tell the compiler your function
doesn't have any side-effects that would prevent auto-
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84234
--- Comment #2 from Florian Lemaitre ---
Well, I should mention that using __attribute((simd)) also has the issue.
So I don't think the problem lies in OpenMP front end.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84234
Richard Biener changed:
What|Removed |Added
Keywords||missed-optimization, openmp
