Hi,
The only thing that graphite modifies is from canonicalize_loop_ivs:
here is the diff between "1" that is the debug_loops (3) before
graphite and "2" that is after graphite.
--- 1 2009-03-02 12:20:03.000000000 -0600
+++ 2 2009-03-02 12:20:18.000000000 -0600
@@ -23,6 +23,8 @@
bb_4 (preds = {bb_3 }, succs = {bb_10 })
{
<bb 4>:
+ D.1655_27 = (unsigned int) width_12(D);
+ D.1656_25 = D.1655_27 + 4294967295;
goto <bb 10>;
}
@@ -90,6 +92,8 @@
bb_11 (preds = {bb_10 }, succs = {bb_6 })
{
<bb 11>:
+ D.1652_24 = (unsigned int) num_comp_14(D);
+ D.1653_6 = D.1652_24 + 4294967295;
goto <bb 6>;
}
@@ -98,16 +102,18 @@
bb_5 (preds = {bb_6 }, succs = {bb_6 })
{
<bb 5>:
+ ivtmp.25_2 = ivtmp.25_4 + 1;
}
bb_6 (preds = {bb_5 bb_11 }, succs = {bb_5 bb_7 })
{
<bb 6>:
- # in_47 = PHI <in_17(5), in_21(11)>
- # out_48 = PHI <out_16(5), out_32(11)>
- # i_49 = PHI <i_18(5), 0(11)>
# SMT.10_50 = PHI <SMT.10_30(5), SMT.10_34(11)>
# SMT.11_52 = PHI <SMT.11_31(5), SMT.11_35(11)>
+ # ivtmp.25_4 = PHI <ivtmp.25_2(5), 0(11)>
+ in_47 = in_21 + ivtmp.25_4;
+ out_48 = out_32 + ivtmp.25_4;
+ i_49 = (int) ivtmp.25_4;
# VUSE <SMT.10_50, SMT.11_52> { SMT.10 SMT.11 }
D.1617_15 = *in_47;
# SMT.10_30 = VDEF <SMT.10_50>
@@ -116,7 +122,7 @@
out_16 = out_48 + 1;
in_17 = in_47 + 1;
i_18 = i_49 + 1;
- if (num_comp_14(D) > i_18)
+ if (ivtmp.25_4 < D.1653_6)
goto <bb 5>;
else
goto <bb 7>;
The fail is in RTL expand in copy_to_mode_reg:
gcc_assert (GET_MODE (x) == mode || GET_MODE (x) == VOIDmode);
(gdb) p x->mode
$16 = SImode
(gdb) p mode
$17 = DImode
It looks like a type problem for the condition to be expanded:
(gdb) p exp
$18 = (tree) 0x7fa9967d2e00
(gdb) pgs
ivtmp.25 < D.1653;
So after figuring out that canonicalize_loop_ivs does compute the
largest precision for all the phi nodes of the loop, such that the new
induction variable can represent all the values of the old IVs, i.e:
for (psi = gsi_start_phis (loop->header);
!gsi_end_p (psi); gsi_next (&psi))
{
phi = gsi_stmt (psi);
res = PHI_RESULT (phi);
if (is_gimple_reg (res) && TYPE_PRECISION (TREE_TYPE (res)) > precision)
precision = TYPE_PRECISION (TREE_TYPE (res));
}
type = lang_hooks.types.type_for_size (precision, 1);
it does not fold_convert the number of iterations to this new type,
and thus we end up building a condition with two different precision
types: 32 for niter and 64 for the new IV.
Attached is a fix for this problem, and the diff between 1 before and
5 after graphite looks like this:
--- 1 2009-03-02 12:20:03.000000000 -0600
+++ 5 2009-03-02 12:54:27.000000000 -0600
@@ -23,6 +23,8 @@
bb_4 (preds = {bb_3 }, succs = {bb_10 })
{
<bb 4>:
+ D.1656_25 = (unsigned int) width_12(D);
+ D.1657_5 = D.1656_25 + 4294967295;
goto <bb 10>;
}
@@ -90,6 +92,9 @@
bb_11 (preds = {bb_10 }, succs = {bb_6 })
{
<bb 11>:
+ D.1652_24 = (unsigned int) num_comp_14(D);
+ D.1653_6 = D.1652_24 + 4294967295;
+ D.1654_4 = (long unsigned int) D.1653_6;
goto <bb 6>;
}
@@ -98,16 +103,18 @@
bb_5 (preds = {bb_6 }, succs = {bb_6 })
{
<bb 5>:
+ ivtmp.25_27 = ivtmp.25_2 + 1;
}
bb_6 (preds = {bb_5 bb_11 }, succs = {bb_5 bb_7 })
{
<bb 6>:
- # in_47 = PHI <in_17(5), in_21(11)>
- # out_48 = PHI <out_16(5), out_32(11)>
- # i_49 = PHI <i_18(5), 0(11)>
# SMT.10_50 = PHI <SMT.10_30(5), SMT.10_34(11)>
# SMT.11_52 = PHI <SMT.11_31(5), SMT.11_35(11)>
+ # ivtmp.25_2 = PHI <ivtmp.25_27(5), 0(11)>
+ in_47 = in_21 + ivtmp.25_2;
+ out_48 = out_32 + ivtmp.25_2;
+ i_49 = (int) ivtmp.25_2;
# VUSE <SMT.10_50, SMT.11_52> { SMT.10 SMT.11 }
D.1617_15 = *in_47;
# SMT.10_30 = VDEF <SMT.10_50>
@@ -116,7 +123,7 @@
out_16 = out_48 + 1;
in_17 = in_47 + 1;
i_18 = i_49 + 1;
- if (num_comp_14(D) > i_18)
+ if (ivtmp.25_2 < D.1654_4)
goto <bb 5>;
else
goto <bb 7>;
Sebastian Pop
--
AMD - GNU Tools
Index: tree-parloops.c
===================================================================
--- tree-parloops.c (revision 144544)
+++ tree-parloops.c (working copy)
@@ -1338,6 +1338,7 @@ canonicalize_loop_ivs (struct loop *loop
affine_iv iv;
edge exit = single_dom_exit (loop);
struct reduction_info *red;
+ gimple_seq stmts;
for (psi = gsi_start_phis (loop->header);
!gsi_end_p (psi); gsi_next (&psi))
@@ -1351,6 +1352,11 @@ canonicalize_loop_ivs (struct loop *loop
type = lang_hooks.types.type_for_size (precision, 1);
+ nit = fold_convert (type, nit);
+ nit = force_gimple_operand (nit, &stmts, true, NULL_TREE);
+ if (stmts)
+ gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
+
gsi = gsi_last_bb (loop->latch);
create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
loop, &gsi, true, &var_before, NULL);
