Shader translator converted single-element vector swizzle to direct index access. It is not wrong, …

Shader translator converted single-element vector swizzle to direct index access. It is not wrong, but unnecessary and moreover it triggers a MESA optimizer bug.
Committed: https://code.google.com/p/angleproject/source/detail?r=1246

[Alok Priyadarshi, 2012-08-10 21:35:00]

The main motivation for this change is to avoid triggering a MESA optimizer bug.

[kbr1, 2012-08-11 00:38:27]

LGTM

I'd like to know why the original code was written the way it was...

Could you make sure a regression test catching the bug is added to
https://github.com/KhronosGroup/WebGL , perhaps under
sdk/tests/conformance/glsl/misc/ ? Thanks.

[Alok Priyadarshi, 2012-08-13 16:45:19]

I am not sure why the original code was written that way. I suspect they were
for folding const vectors. I tested folding with my patch, and it still works.

What kind of test are you looking for? Something that has a single element
access? We certainly do not want to compare translated output.

Daniel: Do you see any issues with this change?

[dgkoch, 2012-08-13 17:33:36]

On 2012/08/13 16:45:19, Alok Priyadarshi wrote:
> I am not sure why the original code was written that way. I suspect they were
> for folding const vectors. I tested folding with my patch, and it still works.
> 
> What kind of test are you looking for? Something that has a single element
> access? We certainly do not want to compare translated output.
> 
> Daniel: Do you see any issues with this change?

Looks good to me.  

I don't know why it was originally written that way either.  I believe that part
of the code came from the original 3DLabs compiler.

[kbr1, 2012-08-13 18:34:58]

On 2012/08/13 16:45:19, Alok Priyadarshi wrote:
> What kind of test are you looking for? Something that has a single element
> access? We certainly do not want to compare translated output.

Agreed, certainly not. I was thinking about a simple rendering test (or shader
compilation test? What triggers the Mesa bug?) that causes the bug to occur.
Simply running the test will ensure that ANGLE -- or Mesa -- doesn't regress
again in this area.