OpenE2K
/
gcc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Don't cancel loop tree in parloops 

2015-07-31  Tom de Vries  <tom@codesourcery.com>

	PR tree-optimization/66846
	* omp-low.c (expand_omp_taskreg) [ENABLE_CHECKING]: Call
	verify_loop_structure for child_cfun if !LOOPS_NEED_FIXUP.
	(expand_omp_target) [ENABLE_CHECKING]: Same.
	(execute_expand_omp) [ENABLE_CHECKING]: Call verify_loop_structure for
	cfun if !LOOPS_NEED_FIXUP.
	(expand_omp_for_static_nochunk): Handle simple latch bb.  Handle case
	that omp_for already has its own loop struct.
	* tree-parloops.c (create_phi_for_local_result)
	(create_call_for_reduction): Handle simple latch bb.
	(create_parallel_loop): Add simple latch bb to preserve
	LOOPS_HAVE_SIMPLE_LATCHES.  Record new exit.  Handle simple latch bb.
	(gen_parallel_loop): Remove call to cancel_loop_tree.
	(parallelize_loops): Skip loops that are inner loops of parallelized
	loops.
	(pass_parallelize_loops::execute) [ENABLE_CHECKING]: Call
	verify_loop_structure.

From-SVN: r226427
This commit is contained in:
Tom de Vries 2015-07-31 06:26:44 +00:00 committed by Tom de Vries
parent 048e62c34a
commit e67d7a1ea1
3 changed files with 86 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
gcc/ChangeLog
View File

@ -1,3 +1,23 @@
2015-07-31 Tom de Vries <tom@codesourcery.com>
PR tree-optimization/66846
* omp-low.c (expand_omp_taskreg) [ENABLE_CHECKING]: Call
verify_loop_structure for child_cfun if !LOOPS_NEED_FIXUP.
(expand_omp_target) [ENABLE_CHECKING]: Same.
(execute_expand_omp) [ENABLE_CHECKING]: Call verify_loop_structure for
cfun if !LOOPS_NEED_FIXUP.
(expand_omp_for_static_nochunk): Handle simple latch bb. Handle case
that omp_for already has its own loop struct.
* tree-parloops.c (create_phi_for_local_result)
(create_call_for_reduction): Handle simple latch bb.
(create_parallel_loop): Add simple latch bb to preserve
LOOPS_HAVE_SIMPLE_LATCHES. Record new exit. Handle simple latch bb.
(gen_parallel_loop): Remove call to cancel_loop_tree.
(parallelize_loops): Skip loops that are inner loops of parallelized
loops.
(pass_parallelize_loops::execute) [ENABLE_CHECKING]: Call
verify_loop_structure.
2015-07-30 Anatoly Sokolov <aesok@post.ru>
* config/v850/v850.h (LIBCALL_VALUE): Remove macros.

32
gcc/omp-low.c
View File

@ -5604,6 +5604,10 @@ expand_omp_taskreg (struct omp_region *region)
}
if (gimple_in_ssa_p (cfun))
update_ssa (TODO_update_ssa);
#ifdef ENABLE_CHECKING
if (!loops_state_satisfies_p (LOOPS_NEED_FIXUP))
verify_loop_structure ();
#endif
pop_cfun ();
}
@ -6535,7 +6539,8 @@ expand_omp_for_static_nochunk (struct omp_region *region,
body_bb = single_succ (seq_start_bb);
if (!broken_loop)
{
gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb);
gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb
|| single_succ (BRANCH_EDGE (cont_bb)->dest) == body_bb);
gcc_assert (EDGE_COUNT (cont_bb->succs) == 2);
}
exit_bb = region->exit;
@ -6818,6 +6823,11 @@ expand_omp_for_static_nochunk (struct omp_region *region,
if (!broken_loop)
{
ep = find_edge (cont_bb, body_bb);
if (ep == NULL)
{
ep = BRANCH_EDGE (cont_bb);
gcc_assert (single_succ (ep->dest) == body_bb);
}
if (gimple_omp_for_combined_p (fd->for_stmt))
{
remove_edge (ep);
@ -6843,9 +6853,19 @@ expand_omp_for_static_nochunk (struct omp_region *region,
set_immediate_dominator (CDI_DOMINATORS, fin_bb,
recompute_dominator (CDI_DOMINATORS, fin_bb));
struct loop *loop = body_bb->loop_father;
if (loop != entry_bb->loop_father)
{
gcc_assert (loop->header == body_bb);
gcc_assert (broken_loop
|| loop->latch == region->cont
|| single_pred (loop->latch) == region->cont);
return;
}
if (!broken_loop && !gimple_omp_for_combined_p (fd->for_stmt))
{
struct loop *loop = alloc_loop ();
loop = alloc_loop ();
loop->header = body_bb;
if (collapse_bb == NULL)
loop->latch = cont_bb;
@ -8984,6 +9004,10 @@ expand_omp_target (struct omp_region *region)
if (changed)
cleanup_tree_cfg ();
}
#ifdef ENABLE_CHECKING
if (!loops_state_satisfies_p (LOOPS_NEED_FIXUP))
verify_loop_structure ();
#endif
pop_cfun ();
}
@ -9492,6 +9516,10 @@ execute_expand_omp (void)
expand_omp (root_omp_region);
#ifdef ENABLE_CHECKING
if (!loops_state_satisfies_p (LOOPS_NEED_FIXUP))
verify_loop_structure ();
#endif
cleanup_tree_cfg ();
free_omp_regions ();

49
gcc/tree-parloops.c
View File

@ -1032,21 +1032,22 @@ create_phi_for_local_result (reduction_info **slot, struct loop *loop)
struct reduction_info *const reduc = *slot;
edge e;
gphi *new_phi;
basic_block store_bb;
basic_block store_bb, continue_bb;
tree local_res;
source_location locus;
/* STORE_BB is the block where the phi
should be stored. It is the destination of the loop exit.
(Find the fallthru edge from GIMPLE_OMP_CONTINUE). */
store_bb = FALLTHRU_EDGE (loop->latch)->dest;
continue_bb = single_pred (loop->latch);
store_bb = FALLTHRU_EDGE (continue_bb)->dest;
/* STORE_BB has two predecessors. One coming from the loop
(the reduction's result is computed at the loop),
and another coming from a block preceding the loop,
when no iterations
are executed (the initial value should be taken). */
if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (loop->latch))
if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (continue_bb))
e = EDGE_PRED (store_bb, 1);
else
e = EDGE_PRED (store_bb, 0);
@ -1055,7 +1056,7 @@ create_phi_for_local_result (reduction_info **slot, struct loop *loop)
locus = gimple_location (reduc->reduc_stmt);
new_phi = create_phi_node (local_res, store_bb);
add_phi_arg (new_phi, reduc->init, e, locus);
add_phi_arg (new_phi, lhs, FALLTHRU_EDGE (loop->latch), locus);
add_phi_arg (new_phi, lhs, FALLTHRU_EDGE (continue_bb), locus);
reduc->new_phi = new_phi;
return 1;
@ -1134,7 +1135,8 @@ create_call_for_reduction (struct loop *loop,
{
reduction_list->traverse <struct loop *, create_phi_for_local_result> (loop);
/* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */
ld_st_data->load_bb = FALLTHRU_EDGE (loop->latch)->dest;
basic_block continue_bb = single_pred (loop->latch);
ld_st_data->load_bb = FALLTHRU_EDGE (continue_bb)->dest;
reduction_list
->traverse <struct clsn_data *, create_call_for_reduction_1> (ld_st_data);
}
@ -1981,7 +1983,7 @@ create_parallel_loop (struct loop *loop, tree loop_fn, tree data,
tree new_data, unsigned n_threads, location_t loc)
{
gimple_stmt_iterator gsi;
basic_block bb, paral_bb, for_bb, ex_bb;
basic_block bb, paral_bb, for_bb, ex_bb, continue_bb;
tree t, param;
gomp_parallel *omp_par_stmt;
gimple omp_return_stmt1, omp_return_stmt2;
@ -2052,8 +2054,12 @@ create_parallel_loop (struct loop *loop, tree loop_fn, tree data,
gcc_assert (exit == single_dom_exit (loop));
guard = make_edge (for_bb, ex_bb, 0);
single_succ_edge (loop->latch)->flags = 0;
end = make_edge (loop->latch, ex_bb, EDGE_FALLTHRU);
/* Split the latch edge, so LOOPS_HAVE_SIMPLE_LATCHES is still valid. */
loop->latch = split_edge (single_succ_edge (loop->latch));
single_pred_edge (loop->latch)->flags = 0;
end = make_edge (single_pred (loop->latch), ex_bb, EDGE_FALLTHRU);
rescan_loop_exit (end, true, false);
for (gphi_iterator gpi = gsi_start_phis (ex_bb);
!gsi_end_p (gpi); gsi_next (&gpi))
{
@ -2102,7 +2108,8 @@ create_parallel_loop (struct loop *loop, tree loop_fn, tree data,
SSA_NAME_DEF_STMT (initvar) = for_stmt;
/* Emit GIMPLE_OMP_CONTINUE. */
gsi = gsi_last_bb (loop->latch);
continue_bb = single_pred (loop->latch);
gsi = gsi_last_bb (continue_bb);
omp_cont_stmt = gimple_build_omp_continue (cvar_next, cvar);
gimple_set_location (omp_cont_stmt, loc);
gsi_insert_after (&gsi, omp_cont_stmt, GSI_NEW_STMT);
@ -2298,10 +2305,6 @@ gen_parallel_loop (struct loop *loop,
scev_reset ();
/* Cancel the loop (it is simpler to do it here rather than to teach the
expander to do it). */
cancel_loop_tree (loop);
/* Free loop bound estimations that could contain references to
removed statements. */
FOR_EACH_LOOP (loop, 0)
@ -2587,6 +2590,7 @@ parallelize_loops (void)
unsigned n_threads = flag_tree_parallelize_loops;
bool changed = false;
struct loop *loop;
struct loop *skip_loop = NULL;
struct tree_niter_desc niter_desc;
struct obstack parloop_obstack;
HOST_WIDE_INT estimated;
@ -2604,6 +2608,19 @@ parallelize_loops (void)
FOR_EACH_LOOP (loop, 0)
{
if (loop == skip_loop)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"Skipping loop %d as inner loop of parallelized loop\n",
loop->num);
skip_loop = loop->inner;
continue;
}
else
skip_loop = NULL;
reduction_list.empty ();
if (dump_file && (dump_flags & TDF_DETAILS))
{
@ -2663,6 +2680,7 @@ parallelize_loops (void)
continue;
changed = true;
skip_loop = loop->inner;
if (dump_file && (dump_flags & TDF_DETAILS))
{
if (loop->inner)
@ -2729,6 +2747,11 @@ pass_parallelize_loops::execute (function *fun)
if (parallelize_loops ())
{
fun->curr_properties &= ~(PROP_gimple_eomp);
#ifdef ENABLE_CHECKING
verify_loop_structure ();
#endif
return TODO_update_ssa;
}