[Mesa-dev] [PATCH 7/9] i965/vec4: Don't allow type conversion in opt_vector_float

[Jason Ekstrand]

On Thu, Mar 17, 2016 at 11:17 AM, Matt Turner <mattst88 at gmail.com> wrote:

> On Thu, Mar 17, 2016 at 10:21 AM, Jason Ekstrand <jason at jlekstrand.net>
> wrote:
> > ---
> >  src/mesa/drivers/dri/i965/brw_vec4.cpp | 1 +
> >  1 file changed, 1 insertion(+)
> >
> > diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp
> b/src/mesa/drivers/dri/i965/brw_vec4.cpp
> > index baf72a2..155a550 100644
> > --- a/src/mesa/drivers/dri/i965/brw_vec4.cpp
> > +++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp
> > @@ -375,6 +375,7 @@ vec4_visitor::opt_vector_float()
> >        if (inst->opcode != BRW_OPCODE_MOV ||
> >            inst->dst.writemask == WRITEMASK_XYZW ||
> >            inst->src[0].file != IMM ||
> > +          inst->src[0].type != inst->dst.type ||
>
> Why?
>

That may be the wrong condition.  The reason is that the code below doesn't
look at the source type at all and just assumes that it's a float.  If, for
instance, we had the first case below but instead of 0D we had some
non-zero VF-capable value, it would interpret it as a float, cram it into a
VF, and then copy it to m2 doing a re-interpret instead of doing a
conversion.

Thinking about it harder, I'm convinced that this patch is bogus, but the
bug is real.
--Jasoan


> This breaks legitimate optimizations, like
>
> -mov m2.xy:F, 0.500000F
> -mov m2.zw:F, 0D
> +mov m2:F, [0.5F, 0.5F, 0F, 0F]
>
> and
>
> -mov vgrf6.0.x:D, -1082130432D
> -mov vgrf6.0.y:D, 1056964608D
> -mov vgrf6.0.z:D, 1065353216D
> +mov vgrf6.0.xyz:F, [-1F, 0.5F, 1F, 0F]
>
> and
>
> -mov vgrf7.0.x:D, 1073741824D
> -mov vgrf7.0.yz:D, 0D
> +mov vgrf7.0.xyz:F, [2F, 0F, 0F, 0F]
>
>
> The first one we should just handle in opt_algebraic by fixing the src
> type. The other two look like NIR fail. If we fixed those things, I'd
> be fine with this.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160317/d804ad80/attachment-0001.html>