Peter Bergner wrote:
> On Fri, 2009-06-26 at 20:04 +0400, Maxim Kuvyrkov wrote:
>> Peter Bergner wrote:
>>
>>> In the mean time, I have asked Luis to test your patch on POWER6 to ensure
>>> there are no performance degradations. It doesn't look like there should
>>> be, but...
>> The assembler should be exactly the same. If it isn't, then there's a
>> bug in the patch.
>
> That's exactly what we're testing for. With your patch, we'll always
> decrement the precedence value if we have a pointer. With the old code,
> that occurred only for RTX_OBJ. Maybe that's the only case were we'll
> see pointers, but I personally don't know that for sure.
I tried to avoid any differences from the current behavior ...
+static int
+rs6000_commutative_operand_precedence (const_rtx op, int value)
+{
+ /* Prefer pointer objects over non pointer objects.
+ For rationale see PR28690. */
+ if (GET_RTX_CLASS (GET_CODE (op)) == RTX_OBJ
... and the patch decrements the value only for RTX_OBJs.
+ && ((REG_P (op) && REG_POINTER (op))
+ || (MEM_P (op) && MEM_POINTER (op))))
+ --value;
+
+ return value;
+}
Peter and Luis,
In any case, thanks for testing the patch on PowerPC.
--
Maxim
