Analyzing code generated for PPC with LRA, I found that sometimes
call-saved registers are overused. It creates long prologs/epilogs.
The following patch fixes it.
The problem was in that BB frequency might be zero.
Committed to trunk as rev. 198792.
The same patch was committed into lra-mike branch.
2013-05-10 Vladimir Makarov <vmaka...@redhat.com>
* lra-assigns.c (find_hard_regno_for): Add 1 to the cost of call
saved registers.
Index: lra-assigns.c
===================================================================
--- lra-assigns.c (revision 198780)
+++ lra-assigns.c (working copy)
@@ -598,7 +598,7 @@ find_hard_regno_for (int regno, int *cos
&& ! df_regs_ever_live_p (hard_regno + j))
/* It needs save restore. */
hard_regno_costs[hard_regno]
- += 2 * ENTRY_BLOCK_PTR->next_bb->frequency;
+ += 2 * ENTRY_BLOCK_PTR->next_bb->frequency + 1;
priority = targetm.register_priority (hard_regno);
if (best_hard_regno < 0 || hard_regno_costs[hard_regno] <
best_cost
|| (hard_regno_costs[hard_regno] == best_cost
