On Wed, May 21, 2008 at 9:39 AM, Philip Balister <philip@xxxxxxxxxxxx> wrote:
> Paul Walmsley wrote:
>>
>> Hello Kyungmin,
>>
>> On Wed, 21 May 2008, Kyungmin Park wrote:
>>
>>> On Wed, May 21, 2008 at 3:21 AM, Paul Walmsley <paul@xxxxxxxxx> wrote:
>>>>
>>>> static void omap_mask_irq(unsigned int irq)
>>>> {
>>>> - int offset = (irq >> 5) << 5;
>>>> + int offset = irq & (~(IRQ_BITS_PER_REG - 1));
>>>>
>>>> - if (irq >= 64)
>>>> - irq %= 64;
>>>> - else if (irq >= 32)
>>>> - irq %= 32;
>>>> + irq %= IRQ_BITS_PER_REG;
>>>
>>> Is it the right conversion?
>>> If the irq is greater then 32 and less then or equal to 64 it's
>>> result is different.
>>> E.g, If irq is 63 then original irq is 63, but new code is 31
>>
>> Hmm, in that condition, the result looks the same to me: irq % 32, either
>> way?
>>
>> More practically, if you look at what it does with that irq variable
>> afterwards, it seems to be a bug if irq is ever greater than 31:
>>
>> intc_bank_write_reg(1 << irq, &irq_banks[0], INTC_MIR_CLEAR0 +
>> offset);
>>
>> I think the only case where the new code would work differently than the
>> previous code is if irq > 95. But that would be a bug, since the shift
>> value would then be > 32, for a 32-bit register.
>>
>>> And if this code is right, how about to use mask instead of modulo op?
>>> irq &= (IRQ_BITS_PER_REG - 1);
>>
>> Hehe, very good point, that would probably save even more cycles! If you
>> agree with the above, perhaps I can convert the code to use that also, and
>> add your Signed-off-by also?
>
> On some code where I used % to detect a counter passing multiples of a
> certain number, oprofile revealed that the % operator burned a lot of CPU
> cycles. I suspect this had to do with the counter increasing to very large
> numbers, but ever since, I've tried to avoid the % operator in critical
> paths.
>
Yes, In embedded environment, we should avoid the multiple, divide,
and modulo op as much as possible.
It's best to use the bit operations.
Signed-off-by: Kyungmin Park <kyungmin.park@xxxxxxxxxxx>
BR,
Kyungmin Park
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
