https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120428
--- Comment #12 from Shawn Xu ---
Bisecting with -mprefer-vector-width=256 leads to PR112824, which seems to be a
modification on the move_max option. Compiling with -mmove-max=256 reproduces
the issue in 12.1: https://godbolt.org/z/boP6148er
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120428
--- Comment #11 from Shawn Xu ---
Since r15-3078-g6ea25c041964bf63014fcf7bb68fb1f5a0a4e123 mentions
-mprefer-vector-width, I tried setting -mprefer-vector-width=256 for
compilation. This leads to the same regression in 14.2, but not in 13.3.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120428
--- Comment #9 from Shawn Xu ---
(In reply to Jonathan Wakely from comment #8)
> (In reply to Shawn Xu from comment #0)
> > On x86-64 with avx512, PR115444 caused the following code to vectorize
> > sub-optimally:
>
> What made you blame PR1154
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120428
--- Comment #2 from Shawn Xu ---
Shorter reproduction:
// https://godbolt.org/z/z8z4Ye4rq
#include
#include
#include
#include
void permute(std::array& data) {
static constexpr std::array order{0, 1};
std::array buffer{};
for (
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120428
--- Comment #3 from Shawn Xu ---
Seems to only matter when -mavx or -mavx512f is on. Otherwise the generated
assemblies are identical
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120428
Bug ID: 120428
Summary: [15/16 regression] Suboptimal autovec involving
blocked permutation and std::copy
Product: gcc
Version: 15.1.1
Status: UNCONFIRMED
Seve
