procyberian/gcc
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Always insert out of SSA copies on edges except for loop->latch.

Browse source 
2010-06-23  Sebastian Pop  <sebastian.pop@amd.com>

	* graphite-sese-to-poly.c (rewrite_phi_out_of_ssa): Always insert out
	of SSA copies on edges except for loop->latch.

From-SVN: r163123
This commit is contained in:
Sebastian Pop 2010-08-11 20:25:11 +00:00 • committed by Sebastian Pop
parent ee646fc62b
commit 4aa9a16727
3 changed files with 15 additions and 49 deletions
Download patch file Download diff file Expand all files Collapse all files

5
gcc/ChangeLog
View file

@ -1,3 +1,8 @@
2010-08-02 Sebastian Pop <sebastian.pop@amd.com>
* graphite-sese-to-poly.c (rewrite_phi_out_of_ssa): Always insert out
of SSA copies on edges except for loop->latch.
2010-08-02 Sebastian Pop <sebastian.pop@amd.com>
* graphite-sese-to-poly.c (rewrite_cross_bb_scalar_deps_out_of_ssa):

5
gcc/ChangeLog.graphite
View file

@ -1,3 +1,8 @@
2010-06-23 Sebastian Pop <sebastian.pop@amd.com>
* graphite-sese-to-poly.c (rewrite_phi_out_of_ssa): Always insert out
of SSA copies on edges except for loop->latch.
2010-06-23 Sebastian Pop <sebastian.pop@amd.com>
* graphite-sese-to-poly.c (rewrite_cross_bb_scalar_deps_out_of_ssa):

54
gcc/graphite-sese-to-poly.c
View file

@ -2246,58 +2246,14 @@ rewrite_phi_out_of_ssa (gimple_stmt_iterator *psi)
for (i = 0; i < gimple_phi_num_args (phi); i++)
{
tree arg = gimple_phi_arg_def (phi, i);
edge e = gimple_phi_arg_edge (phi, i);
/* Try to avoid the insertion on edges as much as possible: this
would avoid the insertion of code on loop latch edges, making
the pattern matching of the vectorizer happy, or it would
avoid the insertion of useless basic blocks. Note that it is
incorrect to insert out of SSA copies close by their
definition when they are more than two loop levels apart:
for example, starting from a double nested loop
| a = ...
| loop_1
| loop_2
| b = phi (a, c)
| c = ...
| end_2
| end_1
the following transform is incorrect
| a = ...
| Red[0] = a
| loop_1
| loop_2
| b = Red[0]
| c = ...
| Red[0] = c
| end_2
| end_1
whereas inserting the copy on the incoming edge is correct
| a = ...
| loop_1
| Red[0] = a
| loop_2
| b = Red[0]
| c = ...
| Red[0] = c
| end_2
| end_1
*/
if (TREE_CODE (arg) == SSA_NAME
&& is_gimple_reg (arg)
&& gimple_bb (SSA_NAME_DEF_STMT (arg))
&& (flow_bb_inside_loop_p (bb->loop_father,
gimple_bb (SSA_NAME_DEF_STMT (arg)))
|| flow_bb_inside_loop_p (loop_outer (bb->loop_father),
gimple_bb (SSA_NAME_DEF_STMT (arg)))))
/* Avoid the insertion of code in the loop latch to please the
pattern matching of the vectorizer. */
if (e->src == bb->loop_father->latch)
insert_out_of_ssa_copy (zero_dim_array, arg, SSA_NAME_DEF_STMT (arg));
else
insert_out_of_ssa_copy_on_edge (gimple_phi_arg_edge (phi, i),
zero_dim_array, arg);
insert_out_of_ssa_copy_on_edge (e, zero_dim_array, arg);
}
var = force_gimple_operand (zero_dim_array, &stmts, true, NULL_TREE);