https://gcc.gnu.org/bugzilla/show_bug.cgi?id=60997
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |RESOLVED
Resolution|--- |FIXED
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
On trunk -floop-interchange is done by default now, but using
-fno-loop-interchange generates faster code. Also loop-interchange is still
performed
with -fopenmp:
> /space/rguenther/install/gcc-9.2/bin/gfortran t.f90 -Ofast -fopenmp
> -fopt-info-loop
t.f90:37:0: optimized: loop vectorized using 16 byte vectors
t.f90:37:0: optimized: loop versioned for vectorization because of possible
aliasing
t.f90:68:0: optimized: loops interchanged in loop nest
t.f90:69:0: optimized: loop vectorized using 16 byte vectors
t.f90:51:0: optimized: basic block part vectorized using 16 byte vectors
t.f90:34:0: optimized: basic block part vectorized using 16 byte vectors
t.f90:17:0: optimized: basic block part vectorized using 16 byte vectors
t.f90:68:0: optimized: loops interchanged in loop nest
...
and curiously with -fopenmp loop interchange pays back.
So I'd say "fixed". But the bugreport is about a non longer existing state
of the compiler (GRAPHITE doing interchange).
Oh, and -Ofast -fopenmp -floop-nest-optimize is "fastest".
