cpuset macro /vs/ inline implementation

Daniel Hellstrom daniel at gaisler.com
Thu Oct 31 13:54:29 UTC 2013

On 10/31/2013 02:44 PM, Sebastian Huber wrote:
> On 2013-10-31 14:34, Daniel Hellstrom wrote:
>> On 10/31/2013 01:34 PM, Sebastian Huber wrote:
>>> On 2013-10-31 13:21, Daniel Hellstrom wrote:
>>>> On 10/31/2013 12:38 PM, Sebastian Huber wrote:
>>>>> On 2013-10-31 12:30, Daniel Hellstrom wrote:
>>>>>> I see that the cpu_set_word_t is set to 32-bit which suites most CPUs,
>>>>>> there is
>>>>>> probably a good reason for it but would it make sens to make it dependent on
>>>>>> the architecture? So that 64-bit machines have it sized to 64-bits. I'm not
>>>>>> sure there are 16-bit SMP machines :) .
>>>>> No, this should be a fixed size integer for all architectures to allow static
>>>>> initialization.  Anyway it will be a long way for RTEMS to support 33
>>>>> processors.
>>>> So what is the point of having an array of 32-bits in the set? If there were a
>>>> 16-bit SMP CPU would not performance be better to describe it with a 16-bit
>>>> instead of an 32-bit word? Its probably not realistic with a 16-bit SMP machine
>>>> anyway, more realistic is a 64-bit SMP machine. Would the atomicity of having a
>>>> 64-bit word instead of 32-bit word on a 64-bit machine matter?
>>> Performance of the operations at this API level is completely irrelevant.
>>> What the implementation does with these CPU sets internally is something
>>> different.
>>> Other operating systems don't have a static link-time configuration.
>> Sorry, I'm not sure I understand what you are saying or if we are saying the
>> same thing or that I'm just don't understand as usual :) All I meant with my
>> question was why specify it to 32-bit for all architectures? Why not have it
>> 16-bit for 16-bit CPUs, 32-bit for 32-bit CPUs and 64-bit for 64-bit CPUs. I'm
>> not sure what that has to do with linking or maximum number of CPUs. Linux for
>> example have defined it to a unsigned long, which is a minimum 32-bit and on
>> some 64-bit machines will be 64-bit, I can't see what the problem is with that.
>> For the LEON it is better with a hard 32-bit definition, so I'm satisfied with
>> the current implementation.
> In case this cpu_set_t is architecture specific, then your application configuration has to deal with this and this is bad. We may add CPU sets to the Configuration (rtems_configuration_table) in 
> the future.

Ok, I didn't think that far. Can't see that it would be a big problem though. However that's a reason anyway which I'm not intending to get deeper into.



More information about the devel mailing list