------- Comment #5 from jakub at gcc dot gnu dot org 2008-01-04 17:09 -------
That's only part of the problem. The other is that ompexpand is very upset if
new basic blocks are inserted in unexpected places, e.g. on the BRANCH_EDGE
from OMP_FOR bb or on the FALLTHRU_EDGE from OMP_CONTINUE. expand_omp_for_*
changes the CFG a lot and so adding profile insns on the edges it will revamp
is premature.
The following patch lets me make -C gcc -k check
RUNTESTFLAGS='--target_board=unix/-fprofile-arcs gomp.exp'
and
make -C */libgomp -k check RUNTESTFLAGS='--target_board=unix/-fprofile-arcs'
(without it there are many ICEs):
--- tree-profile.c.jj16 2007-12-07 12:21:06.000000000 +0100
+++ tree-profile.c 2008-01-04 17:57:28.000000000 +0100
@@ -171,6 +171,19 @@ tree_gen_edge_profiler (int edgeno, edge
{
tree ref, one, stmt1, stmt2, stmt3;
+ if (flag_openmp)
+ {
+ tree stmt = last_stmt (e->src);
+ if (stmt)
+ {
+ /* Avoid inserting profiling instructions on edges
+ which OpenMP expand will replace. */
+ if (TREE_CODE (stmt) == OMP_CONTINUE
+ || TREE_CODE (stmt) == OMP_FOR)
+ return;
+ }
+ }
+
/* We share one temporary variable declaration per function. This
gets re-set in tree_profiling. */
if (gcov_type_tmp_var == NULL_TREE)
In addition to this some bit in cfun? to avoid pass_tree_profile would be
needed
and expand_omp_parallel should set that for the child fns it creates.
Is this ok for you?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34610
