https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123207
--- Comment #20 from Victor Do Nascimento <victorldn at gcc dot gnu.org> ---
It appears that hammersley_create(double *out, int n) is the only function
within rayshade.c that needs the -mprefer-vector-width=128 for the blender
benchmark to succeed.
I initially applied the `#pragma GCC push_options; #pragma GCC optimize("O0");
#pragma GCC pop_options` sequence to the different functions in rayshade.c to
isolate the function.
Then, having moved `hammersley_create` to its own file, I had gcc compile it to
its own object file with -mprefer-vector-width=128 and the miscompare indeed
does go away.
I've looked at the array populated by the loop in question for both 128-bit and
the "native" vector widths and found no difference between them in GDB...
Investigation ongoing
