On 03/13/2014 01:15 PM, David Herrmann wrote:
> Hi
>
> On Thu, Mar 13, 2014 at 1:11 PM, Thomas Hellstrom <thellstrom@xxxxxxxxxx> wrote:
>> Hi.
>>
>> Thanks for reviewing. I'll incorporate your suggestions, except this
>> one, and resend.
>>
>>
>> On 03/13/2014 12:19 PM, David Herrmann wrote:
>>> Hi
>>>
>>> On Thu, Mar 13, 2014 at 11:57 AM, Thomas Hellstrom
>>> <thellstrom@xxxxxxxxxx> wrote:
>> ...
>>
>> - if (cmd & IOC_IN) {
>> - if (copy_from_user(kdata, (void __user *)arg,
>> - usize) != 0) {
>> - retcode = -EFAULT;
>> + retcode = drm_ioctl_permit(ioctl->flags, file_priv);
>> + if (unlikely(retcode))
>>
>>> That "unlikely" seems redundant given that all error paths in
>>> drm_ioctl_permit() already are "unlikely".
>> Yes, we know that's true, but I don't think compilers in general can
>> combine branch prediction hints in that way,
>> or even have the information necessary to do it.
>> I mean even if each individual test resulting in an error is unlikely,
>> how could the compiler know that
>> all tests combined would result in an error being unlikely?
> The function is static, so the compiler can see that it returns "!=0"
> only if one of the "unlikely" branches was hit. So I think it's safe
> to assume the whole thing returns "!=0" only in unlikely conditions.
>
But a compiler can't (or shouldn't) make that assumption. Just as an
(adapted) example, imagine that
each test had a 20% probability of returning an error. The probability
of the function returning an error would
then be 68%..
> I'm no big fan of excessive likely/unlikely annotations, but I'm fine
> if you want to keep it.
Fair enough.
Thanks,
Thomas
>
> Thanks
> David
> _______________________________________________
> dri-devel mailing list
> dri-devel@xxxxxxxxxxxxxxxxxxxxx
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/dri-devel
