https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80995
Thomas Schwinge <tschwinge at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|Fortran OpenACC kernels |Unparallelized Fortran
|"Unhandled stmt in |OpenACC kernels after
|entry/exit: _14 = |Fortran loop optimizations
|BUILTIN_EXPECT ([...])" |
|after "Introduce fortran |
|loop preheader" |
--- Comment #2 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
With trunk r238114 "Optimize fortran loops with +-1 step" merged in, the two
regressions mentioned before are gone again, but instead, the following fail to
parallelize with "-Os" (only):
* libgomp.oacc-fortran/asyncwait-1.f90
* libgomp.oacc-fortran/deviceptr-1.f90
* libgomp.oacc-fortran/kernels-loop-2.f95
* libgomp.oacc-fortran/kernels-loop-data-2.f95
* libgomp.oacc-fortran/kernels-loop-data-enter-exit-2.f95
* libgomp.oacc-fortran/kernels-loop-data-enter-exit.f95
* libgomp.oacc-fortran/kernels-loop-data-update.f95
* libgomp.oacc-fortran/kernels-loop-data.f95
* libgomp.oacc-fortran/kernels-loop.f95
* libgomp.oacc-fortran/kernels-parallel-loop-data-enter-exit.f95
* libgomp.oacc-fortran/non-scalar-data.f90
With nvptx offloading, this is observed as: "warning: OpenACC kernels construct
will be executed sequentially; will by default avoid offloading to prevent data
copy penalty".
I did not try to understand what it is that parloops doesn't like.
