[Mesa-dev] [PATCH 3/3] glsl: Generate smaller values for uniform locations

[Ian Romanick]

On 06/10/2013 03:54 PM, Fredrik Höglund wrote:
> On Monday 10 June 2013, Brian Paul wrote:
>> On 06/10/2013 11:52 AM, Ian Romanick wrote:
>>> From: Ian Romanick <ian.d.romanick at intel.com>
>>>
>>> Previously we would generate uniform locations as (slot << 16) +
>>> array_index.  We do this two handle applications that assume the
>
> to handle
>
>>> location of a[2] will be +1 from the location of a[1].  This resulted in
>>> every uniform location being at least 0x10000.  We've now encountered
>>> two applications that assume uniform values will be smaller.
>>>
>>> NEITHER BEHAVIOR IS GUARNATEED
>>
>> GUARANTEED
>>
>>    OR IMPLIED BY ANY VERSION OF OpenGL.
>
> I wonder if OpenGL doesn't in fact imply this.
>
> Section 4.4.3 of the GLSL 4.30 spec says this about the location layout-
> qualifier:
>
> 	"Individual elements of a uniform array are assigned consecutive locations
> 	 with the first element taking location /location/."
>
> and
>
> 	"The first inner-most scalar, vector or matrix member or element takes
> 	 the specified /location/ and the compiler assigns the next inner-most
> 	 member or element the next incremental location value. Each subsequent
> 	 inner-most member or element gets incremental locations for the entire
> 	 structure or array. This rule applies to nested structures and arrays and
> 	 gives each inner-most scalar, vector, or matrix type a unique location.
> 	 For arrays without an explicit size, the size is calculated based on its
> 	 static usage. When the linker generates locations for uniforms without an
> 	 explicit location, it assumes for all uniforms with an explicit location all
> 	 their array elements and structure members are used and the linker will
> 	 not generate a conflicting location, even if that element of member is
> 	 deemed unused."
>
> It seems unlikely to me that the intent is for the compiler to only apply
> this rule to uniforms with an explicit location, even though the spec can
> be read that way.

Good catch. :)  I'll amend the commit message.  I forgot that in 4.3 we 
changed the spec to make the guarantee that a[2] has location +1 from 
a[1].  We should add some piglit tests for that.  However, there is 
still not guarantee that the location values will be "small."

We'll have to make some additional changes in this area to support 
GL_ARB_explicit_uniform_location, but I wanted to make a first patch 
that would be suitable for the stable branches.

> Fredrik