[PATCH] radeon: Deinline indirect register accessor functions

[Ilia Mirkin]

On Mon, May 18, 2015 at 3:04 PM, Christian König
<christian.koenig at amd.com> wrote:
> On 18.05.2015 20:50, Denys Vlasenko wrote:
>>
>> On 05/18/2015 08:06 PM, Christian König wrote:
>>>
>>> I'm actually surprised how often people come along with that. The last
>>> time we tried this it caused a noticeable performance drop.
>>>
>>> Basic problem is that this line:
>>>>
>>>> +    if ((reg < rdev->rmmio_size || reg < RADEON_MIN_MMIO_SIZE) &&
>>>> !always_indirect)
>>>
>>> optimizes away in most of the cases which reduces the call to a readl
>>> which is way faster than the spinlock path.
>>>
>>> So this is a NAK,
>>
>>
>> Fair enough.
>>
>> I'm preparing a v2 where the fast branch of r100_mm_{r,w}reg() will stay
>> inlined.
>>
> Sounds good to be, but IIRC that was suggested the last time this came up as
> well. You might just want to google a bit why it wasn't done like this
> before submitting the patch for review.
>
> BTW: Please CC the dri-devel list as well, cause not everybody is reading on
> linux-kernel.

http://lists.freedesktop.org/archives/dri-devel/2014-April/057349.html
...
http://lists.freedesktop.org/archives/dri-devel/2014-April/057520.html

Actually Lauri was *inlining* the function, not out-of-lining. I made
the suggestion (and you agreed at the time) that the slow-path should
be kept out of line, but apparently there was still high CPU overhead
as a result?

Cheers,

  -ilia