[Intel-gfx] [RFC 00/15] Selectable platform support

[Chris Wilson]

Quoting Jani Nikula (2018-02-09 11:49:16)
> On Thu, 08 Feb 2018, Joonas Lahtinen <joonas.lahtinen at linux.intel.com> wrote:
> > Quoting Tvrtko Ursulin (2018-02-08 16:06:41)
> >> 
> >> On 08/02/2018 13:26, Chris Wilson wrote:
> >> > Quoting Tvrtko Ursulin (2018-02-08 13:05:51)
> >> >> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >> >>
> >> >> For Joonas basically. :)
> >> >>
> >> >> Rough goal - add Kconfig options to turn off supported platforms
> >> >> and count on compiler DCE to make the driver smaller.
> >> >>
> >> >> Tested as so much that it boots and renders on Skylake with all
> >> >> platforms/gens older than Gen8 turned off.
> >> >>
> >> >>     text    data     bss     dec     hex filename
> >> >> 1502847   54223    2888 1559958  17cd96 i915.ko.original
> >> >> 1375647   51939    2888 1430474  15d3ca i915.ko.gen8+
> >> >>
> >> >> So only ~124kiB saving. Or ~8.5%. Perhaps once GCC LTO support
> >> >> lands it would be better than this?
> >> > 
> >> > Did you get to the point where the compiler was complaining about
> >> > unused functions?
> >> 
> >> No, but on a random check it seems that it is removing some. For
> >> instance i965_emit_bb_start and i830_emit_bb_start are not in my
> >> build.
> >> 
> >> Doesn't mean I haven't made some other mistake which is preventing
> >> more savings.
> >> 
> >> >> Starts with smaller patches to show the idea step by step on Gen2,
> >> >> then proceeds in larger chunks, to finish with some invasive
> >> >> Coccinelle works to enable the last few kilo-bytes of savings.
> >> > 
> >> > So, if we want to support this, how do we test it?
> >> > 
> >> > Do a per-platform build and check modinfo for pci ids?
> >> > 
> >> > Limit the CI builds to be per-platform and check they work?
> >> 
> >> Extensively. :) Which will probably be a problem. PCI ids is not
> >> enough, I think it would actually need functional testing so a growth
> >> of number of builds we would need to test.
> >
> > My original suggestion was to be able to "pre-select the PCI ID" and
> > let the compiler do the magic with LTO to get rid of the dead
> > code. You would test by comparing the "include all" and "one PCI ID"
> > kernel operation on same system. I'm assuming that the feature would
> > be a useful non-default to opt-in to similarly to a "targeted" initrd.
> 
> So I can imagine someone shipping a specific product with a specific
> platform (or even PCI ID) could use this. Beyond that, who would use
> this? Some rare enthusiasts?
> 
> Is there any way a general purpose distro could use it? I guess only by
> building several drivers with different options (and names) that would
> have different PCI ID matches.

The way I see this effort scaling to general purpose distros is
basically with a flock of modules (or a fat module with disposable
.sections). I'm a few steps sort of $profit.
-Chris