The following patch adds reload behaviour to LRA for some cases. Some
targets assumes such behaviour.
The patch was successfully bootstrapped on x86/x86-64.
Committed as rev. 184631.
2012-02-28 Vladimir Makarov <vmaka...@redhat.com>
* lra-constraints.c (check_and_process_move): Process source first
for the secondary move.
(process_alt_operands): Increase cost for scratch allocations.
Index: lra-constraints.c
===================================================================
--- lra-constraints.c (revision 184561)
+++ lra-constraints.c (working copy)
@@ -1097,20 +1097,20 @@ check_and_process_move (bool *change_p,
sri.prev_sri = NULL;
sri.icode = CODE_FOR_nothing;
sri.extra_cost = 0;
- if (dclass != NO_REGS)
- {
- in_p = true;
- rclass = dclass;
- x = sreg;
- xclass = sclass;
- }
- else if (sclass != NO_REGS)
+ if (sclass != NO_REGS)
{
in_p = false;
rclass = sclass;
x = dreg;
xclass = dclass;
}
+ else if (dclass != NO_REGS)
+ {
+ in_p = true;
+ rclass = dclass;
+ x = sreg;
+ xclass = sclass;
+ }
else
return false;
temp_assign_p = false;
@@ -2027,6 +2027,14 @@ process_alt_operands (int only_alternati
this_costly_alternative, NULL))
reject++;
}
+ /* We simulate the behaviour of old reload here.
+ Although scratches need hard registers and it
+ might result in spilling other pseudos, no reload
+ insns are generated for the scratches. So it
+ might cost something but probably less than old
+ reload pass believes. */
+ if (lra_former_scratch_p (REGNO (operand_reg[nop])))
+ reject += LOSER_COST_FACTOR;
}
}
else if (did_match)
