Re: [PATCH] Add option to build with -O3

[Austin S Hemmelgarn <ahferroin7@gmail.com>]

On 2014-03-07 07:42, Richard Weinberger wrote:
> 
> 
> Am 07.03.2014 13:39, schrieb Austin S Hemmelgarn:
>> On 2014-03-06 08:28, Richard Weinberger wrote:
>>> On Wed, Mar 5, 2014 at 6:37 AM, Jon Ringle <jringle@gridpoint.com>
>>> wrote:
>>>>
>>>>
>>>> On Wed, 5 Mar 2014, Greg KH wrote:
>>>>
>>>>> On Tue, Mar 04, 2014 at 07:01:49PM -0500, Jon Ringle wrote:
>>>>>> +config CC_OPTIMIZE_FOR_SPEED +    bool "Optimze for speed
>>>>>> (-O3)" +    help +      Enabling this option will pass "-O3"
>>>>>> to gcc +      resulting in a larger kernel (but possibly
>>>>>> faster)
>>>>>
>>>>> Are you sure about that?  Have you measured it?
>>>>
>>>> I do know that there is an improvement performance-wise for my
>>>> particular use-case.
>>>>
>>>> My target is an ARM board being built with gcc-4.8.2. My board
>>>> has on it a sc16is740 that is used as an RS-485 port. The
>>>> sc16is740 is on the i2c bus, so when an interrupt comes in to
>>>> indicate that there is data available to be read, I need to get
>>>> the data over the i2c bus. I do this on a kthread to do this
>>>> work. The i2c transactions (using i2c-davinci driver) are also 
>>>> interrupt driven. I was seeing a lot of lost packets when
>>>> receiving data at only 19200. Adding the -O3 compile option
>>>> helped in this regard in that I am now rarely seeing packet
>>>> loss.
>>>
>>> Please also see: 
>>> http://www.gossamer-threads.com/lists/linux/kernel/1835761?do=post_view_threaded#1835761
>>>
>>>
>> Ironically, combining these might achieve a significant performance
>> improvement over CONFIG_GENERIC_CPU and -O2.
> 
> *might*
> We still need a sane proof.
> 
> Thanks,
> //richard
> 
I'm not arguing that, I just don't have the time to test it at the
moment; and, "performance improvement" also can mean different things to
different people (a lot of RT folks for example don't care if things run
faster if it makes them much less deterministic).