[PATCH] Expose more parallelism+ cleanups to autopar
Razya Ladelsky
RAZYA@il.ibm.com
Tue Dec 22 12:47:00 GMT 2009
Hi,
This patch includes the change in
http://gcc.gnu.org/ml/gcc-patches/2009-11/msg00995.html,
which releases a constraint, that results in exposing a lot more
parallelization opportunities, as more
loops are being tested for parallelism.
As a result, I experienced a few new SPEC and testsuite failures when
autopar is enabled.
I solved most of them, the fixes are included in this patch.
(part of the fixes resolve also older failures such as PR38168)
The main changes are:
- Avoid parallelization when the function has_nonlocal_label.
- Avoid parallelization when the preheader is irreducible.
- Try to optimize when estimated_loop_iterations_int is unresolved.
- I added the duplication of the original latch in
gimple_duplicate_sese_tail,
and updated the cfg accordingly.
The current status of failing tests (with autopar enabled):
cactus (I already reduced to a testcase and will open a PR)
perlbench (runtime failure, will open a PR)
comp-goto-1.c (an old one, PR38153)
gfortran.dg/graphite/block-2.f
gfortran.dg/graphite/id-1.f90
gfortran.dg/graphite/id-4.f90
gfortran.dg/graphite/pr38459.f90
I will open a PR for the graphite testcases failures as well.
Bootstrap and testsuite pass successfully (as autopar is not enabled by
default).
OK for trunk?
Thanks,
Razya
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: expose_parallel.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20091222/20d2a288/attachment.txt>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: ChangeLog.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20091222/20d2a288/attachment-0001.txt>
More information about the Gcc-patches
mailing list