On 12/05/11 10:18, Jakub Jelinek wrote:
On Thu, May 12, 2011 at 10:09:42AM +0100, Hari Sandanagobalane wrote:
The REGNO_REG_CLASS is generally an array of size
FIRST_PSEUDO_REGISTER. So, the indexes go from 0 to
FIRST_PSEUDO_REGISTER-1.
That is true.
I think the ">=" condition is fine in that
case. Do you agree?
That is wrong. It is perfectly fine to handle hard reg with
regno (FIRST_PSEUDO_REGISTER - 1) if it has nregs 1, or
hard reg with regno (FIRST_PSEUDO_REGISTER - 2) if it has nregs 1 or 2.
The following loop starts with nregs--, therefore for
hard_regno + nregs == FIRST_PSEUDO_REGISTER at the place of your test
REGNO_REG_CLASS will be used in the loop for hard_regno + nregs - 1
down to hard_regno + 0.
Ooh yes. You are right. Thanks for that!
I have now committed the revised patch to mainline as revision 173699.
Cheers
Hari
With this change, the patch is ok. Thanks for the patch.
+ {
+ SET_HARD_REG_BIT (ira_prohibited_class_mode_regs[cl][j],
+ hard_regno);
+ continue;
+ }
pclass = ira_pressure_class_translate[REGNO_REG_CLASS
(hard_regno)];
for (nregs-- ;nregs>= 0; nregs--)
if (((enum reg_class) pclass
Jakub
