[Mesa-dev] [PATCH 00/61] nir: Move to using instructions for derefs

[Rob Clark]

On Wed, Mar 28, 2018 at 9:41 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
>
> On March 28, 2018 17:43:31 Rob Clark <robdclark at gmail.com> wrote:
>
> On Wed, Mar 28, 2018 at 8:16 PM, Jason Ekstrand <jason at jlekstrand.net>
> wrote:
> On March 28, 2018 16:54:33 Rob Clark <robdclark at gmail.com> wrote:
>
> I had noticed the code to remove dead deref's in a few of the passes
> (at least on your wip branch), and had wondered a bit about not just
> requiring all the deref related lowering to happen in ssa and possibly
> require dce after, although admittedly hadn't thought about it *too*
> much yet..
>
> Yeah. Like I said below, it should be ready enough to just have a tiny
> clean-up pass instead of having to run full-on dce.  Maybe just running dce
> is the right choice; I'm not sure.
>
>
> I kinda expected to use the dce clean things up once we are in a
> deref-instruction world.. re: validation passes, could we not just
> allow dead deref instructions to be ok.  That seems like kind of a
> natural thing..
>
> Making validation ignore them is easy. The trickier bit is that they can
> cause problems for any pass which works on all deref instructions as opposed
> to working on texture instructions or intrinsics and tracing the deref chain
> back.  The later are ok because they'll never look at dead derefs.  There
> former (which are likely to be more efficient if we've CSEd derefs) can run
> into trouble as it's not always obvious when a deref is dead.
>
> I'm sure I'll get a better feel for this whole mess as I continue to
> progress.
>
> Defn not trying to second guess you since you are deeper into it that
> I am..  But requiring dce (or a some sort of mini deref-dce) pass in
> various places seems reasonable.. I guess it would be nice (given the
> growing list of nir passes) to have some more formal way to require
> that some pass(es) is run prior to a whatever random pass driver wants
> to run would be nice.  (Not sure if llvm's PassManager provides this..
> if it doesn't, it should.)
>
> Yeah.  We could theoretically use the metadata system for this but it seems
> like a bit of an abuse.  I'll know more once I get done removing deref
> chains.  That process is teaching me about all sorts of things I missed on
> the first pass.
>

mmm, I wouldn't look at it as an abuse, as much as making the metadata
system a bit more generic.  Metadata is really a subset of "state of
the IR".  (Although perhaps that means nir_metadata_* is no longer a
good name ;-))

But this doesn't all have to be done at once.. merging addition of
deref instructions with (nearly) immediate lowering to deref chains
seems relatively safe (piglit was happy) and low impact on drivers.
And would be helpful for clover / vtn pointer (ie. not having to
rebase on top of two large patchsets)

BR,
-R