On 04/05/2017 11:25 AM, Jakub Jelinek wrote:
On Wed, Apr 05, 2017 at 11:11:54AM -0400, Vladimir Makarov wrote:
--- ira-color.c (revision 246536)
+++ ira-color.c (working copy)
@@ -1367,6 +1367,16 @@ update_costs_from_allocno (ira_allocno_t
|| ALLOCNO_ASSIGNED_P (another_allocno))
continue;
+ if (GET_MODE_SIZE (ALLOCNO_MODE (cp->second)) < GET_MODE_SIZE (mode))
+ /* If we have different modes use the smallest one. It is
+ a sub-register move. It is hard to predict what LRA
+ will reload (the pseudo or its sub-register) but LRA
+ will try to minimize the data movement. Also for some
+ register classes bigger modes might be invalid,
+ e.g. DImode for AREG on x86. For such cases the
+ register move cost will be maximal. */
+ mode = ALLOCNO_MODE (cp->second);
+
cost = (cp->second == allocno
? ira_register_move_cost[mode][rclass][aclass]
: ira_register_move_cost[mode][aclass][rclass]);
@@ -1512,7 +1522,7 @@ update_conflict_hard_regno_costs (int *c
index = ira_class_hard_reg_index[aclass][hard_regno];
if (index < 0)
continue;
- cost = (int) ((unsigned) conflict_costs [i] * mult) / div;
+ cost = (int) (((long) conflict_costs [i] * mult) / div);
If you want something wider than unsigned, wouldn't it be better to
use HOST_WIDE_INT then? Otherwise it will work differently between
32-bit and 64-bit hosts.
Can any of those 3 values be negative? If not, perhaps
unsigned HOST_WIDE_INT?
Thank you for finding this, Jakub. The overflow might happen only when
65K is used for ira_register_move_cost which means an impossible move in
most cases. Still if it happens in rare cases, the behavior might be
different for the native and a cross.
I'll correct the patch.
