On Thu, Dec 15, 2016 at 6:29 PM, Jason Ekstrand <[email protected]>
wrote:

> On Dec 15, 2016 3:08 PM, "Eric Anholt" <[email protected]> wrote:
>
> Jason Ekstrand <[email protected]> writes:
>
> > This keeps some of Connor's original code.  However, while I was at it,
> > I updated this very old pass to a bit more modern NIR.
>
> I love how much smaller this code is.  Assuming that the phi builder
> does that job correctly, this is pretty trivial to read and understand.
>
>
> The bigger vars_to_ssa has been using the phi builder for some time now
> and it's handled everything we've ever thrown at it.
>
> The only things I saw were:
>
> - It doesn't call nir_metadata_preserve() like the old pass did, while
>   it adds new SSA values.  Seems like it should.
>
>
> Yup.   I'll get that fixed.
>
> - It asserts that reg->num_array_elems == 0, while the old pass would
>   just skip those.  It looks like the only source of arrays would be
>   nir_lower_locals_to_regs(), which only i965 is calling.  Is that
>   something we should be using?  Should we be handling the array reg
>   case in this pass and in our drivers?
>
>
> I've got a patch for that.  I needed it for my hack testing.  I'll as it
> to the pass.  Incidentally, I don't think the old pass handles either of
> those correctly, it just assumed everything was a "normal" register.
>

I just sent a v3 with all of the comments addressed.  It's currently
running on Jenkins with a hack patch that, instead of just going out of
SSA, goes out of SSA completely, calls regs_to_ssa to go back in, and then
goes out of SSA as normal.  I think it's pretty well tested now. :)

https://cgit.freedesktop.org/~jekstrand/mesa/commit/?h=jenkins_gl
_______________________________________________
mesa-dev mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

Reply via email to