[Mesa-dev] [PATCH 1/2] i965: remove cache_aux_free_func array

[Emil Velikov]

On 21 October 2015 at 22:44, Matt Turner <mattst88 at gmail.com> wrote:
> On Wed, Oct 21, 2015 at 2:16 PM, Emil Velikov <emil.l.velikov at gmail.com> wrote:
>> On 21 October 2015 at 21:33, Kenneth Graunke <kenneth at whitecape.org> wrote:
>>> On Monday, October 19, 2015 02:54:56 PM Emil Velikov wrote:
>>>> Ping on these two trivial patches ?
>>>>
>>>> -Emil
>>>
>>> Oh, sorry, I thought I'd sent R-bs for these...
>>>
>>> Both are
>>> Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
>> Thanks Ken. I was wondering if people looked at them and went "meh ...
>> too small, we need something beefier" :-P
>>
>> And now ... some C++ questions. I realise that templates (or is it
>> only STL?) are out of the question, but how about
>>  - Initialization upon object declaration, rather than copy
>> constructors ? Rather trivial yet we have thousands of
>
> You're talking about things like this?
>
>    fs_reg reg = fs_reg(...)
>
> I seem to remember trying at one point to replace those with just
>
>    fs_reg reg(...)
>
> and found that it made absolutely no difference in the compiled code.
>
> The second's probably preferable if for no other reason than it's
> shorter, but I don't know that there's anything to be gained...
>
That's precisely what I meant.

>> duplicated/wasted CPU cycles due to it. One example is the memset()
>> from {fs,src,dst}_reg. Does the compiler squash/optimize those for us
>> ?
>
> Not sure. Experiments and data are welcome.
>
> I think people preferred the memset because there wasn't an
> alternative safe way of ensuring everything was initialized? Not sure.
>
Ack. I'll give it a few quick tests.

>>  - Where is the line about "big enough to pass as reference" rather
>> than a copy for i965 ? It seems that older code(?) and extremely
>> common things such as the *_reg are passed around as copies.
>
> We want to use const references for dst_reg/src_reg/fs_reg where
> possible (see commit e58992aed and the three immediately before it for
> data). I don't think there should be many more instances of this.
>
I think I've saw a few more cases where we can do the same. I'll send
patches as I come across them.

> brw_reg is 8 bytes, so there's no reason to pass it as anything but by value.
Indeed.

Thanks
Emil