[Mesa-dev] Proposal to branch off old drivers

[Marek Olšák]

On Fri, May 26, 2017 at 2:34 PM, Timothy Arceri <tarceri at itsqueeze.com> wrote:
> On 26/05/17 21:38, Marek Olšák wrote:
>
>> On May 26, 2017 2:45 AM, "Timothy Arceri" <tarceri at itsqueeze.com
>> <mailto:tarceri at itsqueeze.com>> wrote:
>>
>>     Hi all,
>>
>>     Following on from the discussion here:
>>
>>     https://lists.freedesktop.org/archives/mesa-dev/2017-May/155971.html
>>     <https://lists.freedesktop.org/archives/mesa-dev/2017-May/155971.html>
>>
>>     Back in 2011/12 despite various concerns old hardware would become
>>     useless, dropping support for DRI1 drivers Mesa proved distros were
>>     up to the challenge of packaging up the old driver branch, and since
>>     we maintain compatibility they continue to work without issue.
>>
>>     I'm currently working on uniform packing for gallium drivers which
>>     means updates to struct gl_program_parameter_list and the assumption
>>     that everything is padded to 4 vectors. Rather than updating and
>>     testing i915 to work with this (or even hacking around it), I'd
>>     rather make the proposal to branch off some older drivers.
>>
>>     Why branch them off?
>>
>>     1. IMO there is a bunch of clean-up this would enable such as:
>>
>>     - enabling a bunch of extensions by default and removing on the
>>     runtime checks for these pasted all over the api.
>>     - dropping a bunch of non asm mesa ir code paths
>>     - dropping a bunch of driver function callbacks
>>     - the software tnl code??
>>     - Likely a bunch of other bits and pieces.
>>
>>     2. They are either not in development at all, or being updated
>>     extremely rarely. Testing is often just does this code compile.
>>     Having them in master just opens them to the possibility of breakage.
>>
>>     3. Death by a thousand cuts. While the clean-ups above may not be
>>     huge I would argue a more important outcome is the ability to
>>     preform re-factors, add features, etc without needlessly updating
>>     these drivers.
>>
>>     As someone who re-factored the main gl_* structs last year in the
>>     lead up to shader cache support, I can say my job would have been
>>     much easier if I didn't have to needlessly update the old classic
>>     drivers.
>>     On the gallium side there is are things like adding caps to all the
>>     drivers etc, again not huge but another cut.
>>
>>     4. As the API expands it just adds more overhead for features these
>>     drivers will mostly never support. The drivers likely already run on
>>     systems with much slower cpus.
>>
>>     My specific proposal is:
>>
>>     - Rather than just pointing distros at the last Mesa release as we
>>     did for the DRI1 driver, we create a mesa-pre-dx9-1.0 branch
>>     (branched from 17.1). However unlikely this will at least give us
>>     the possibility to release updates as some dev's have shown interest
>> in.
>>
>>     - Remove the following drivers from master:
>>         Classic:
>>         --------
>>         i915, nouveau, r200, radeon, swrast (classic)
>>
>>         Gallium:
>>         --------
>>         r300, i915g
>>
>>
>> No matter how hard you are trying, there is no real benefit in removing
>> any Gallium drivers. Come on, is the PIPE_CAP thing your only argument? I
>> can add a new PIPE_CAP case to all gallium drivers in 15 seconds. If you
>> can't do that, you need a better editor.
>
>
> No the reason was the same as all the other drivers. To stop
> untested/unmaintained drivers breaking, and to remove various legacy OpenGL
> 2.1 < checks from Mesa core.

r300 is actually a well tested driver and it's still stable and solid.
It's maintained in the sense that I test it before I commit a change
that might break it, so technically it is a maintained driver. r300
supports GL 2.1 and many GL 3.x features and a few from GL 4.x. See
all occurrences of "r300" and "all":
https://cgit.freedesktop.org/mesa/mesa/tree/docs/GL3.txt?id=a984abdad39df2d8ceb4c46e11f4ce1344c36c86



>
>>
>> The only reason for removing a Gallium driver would be if the driver was
>> broken or inferior in some way. That might be i915g, but even that is kinda
>> a shaky ground.
>>
>> I thought you'd only ask about removing pre-i915 classic drivers (i.e.
>> keeping i915). That would have a higher chance of success.
>
>
> The pre-i915 drivers are actually not a big deal since they don't implement
> many features. In my experience its i915 that gets in the way a lot of the
> time.
>
> It seemed like there was more support for this work last week. So I withdraw
> my proposal. I'll just hack around the i915 driver like we always do.

There is also the option of removing i915c and keeping i915g. The
latter needs much less maintenance and gets much less in the way.
Anyway, this is not my decision to make. i915c maintainers can decide
that.

Marek