[Mesa-dev] [PATCH v2 23/82] glsl: Do not do CSE for expressions involving SSBO loads

[Jordan Justen]

On 2015-06-17 19:33:23, Matt Turner wrote:
> On Wed, Jun 17, 2015 at 5:20 PM, Jordan Justen
> <jordan.l.justen at intel.com> wrote:
> > I wanted to question whether this was required, based on this text
> > from the extension spec:
> >
> > "The ability to write to buffer objects creates the potential for
> >  multiple independent shader invocations to read and write the same
> >  underlying memory. The same issue exists with the
> >  ARB_shader_image_load_store extension provided in OpenGL 4.2, which
> >  can write to texture objects and buffers. In both cases, the
> >  specification makes few guarantees related to the relative order of
> >  memory reads and writes performed by the shader invocations."
> >
> > But I'm not sure if we can reconcile CSE with 'memoryBarrier' and
> > 'barrier'. curro, any thoughts from image load/store?
> 
> I think the point is that if you have two reads of the same location,
> you can't combine them into a single load since in the intervening
> time another shader invocation might have written to that location...?

I don't think the spec requires you to have gotten the newly written
value in that case. So, it might be okay to just read it once.

"In both cases, the specification makes few guarantees related to the
 relative order of memory reads and writes performed by the shader
 invocations."

> The question about the semantics of memory-barriers is a good one though.