[Mesa-dev] [PATCH 5/6] glsl: teach opt_array_splitting about bindless images

[Ian Romanick]

On 05/31/2017 02:04 AM, Samuel Pitoiset wrote:
> 
> 
> On 05/31/2017 02:46 AM, Ian Romanick wrote:
>> On 05/25/2017 11:07 AM, Samuel Pitoiset wrote:
>>> Memory/format layout qualifiers shouldn't be lost when arrays
>>> of images are splitted by this pass.
>>>
>>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>>> ---
>>>   src/compiler/glsl/opt_array_splitting.cpp | 17 +++++++++++++++--
>>>   1 file changed, 15 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/src/compiler/glsl/opt_array_splitting.cpp
>>> b/src/compiler/glsl/opt_array_splitting.cpp
>>> index e3073b022c..5841454574 100644
>>> --- a/src/compiler/glsl/opt_array_splitting.cpp
>>> +++ b/src/compiler/glsl/opt_array_splitting.cpp
>>> @@ -449,9 +449,22 @@ optimize_split_arrays(exec_list *instructions,
>>> bool linked)
>>>         for (unsigned int i = 0; i < entry->size; i++) {
>>>            const char *name = ralloc_asprintf(mem_ctx, "%s_%d",
>>>                                               entry->var->name, i);
>>> -
>>> -         entry->components[i] =
>>> +         ir_variable *new_var =
>>>               new(entry->mem_ctx) ir_variable(subtype, name,
>>> ir_var_temporary);
>>> +
>>> +         if (entry->var->type->contains_image()) {
>>> +            /* Do not lost memory/format qualifiers when arrays of
>>> images are
>>
>>                           lose
>>
>>> +             * splitted.
>>
>>                    split
>>
>>> +             */
>>> +            new_var->data.memory_read_only =
>>> entry->var->data.memory_read_only;
>>> +            new_var->data.memory_write_only =
>>> entry->var->data.memory_write_only;
>>> +            new_var->data.memory_coherent =
>>> entry->var->data.memory_coherent;
>>> +            new_var->data.memory_volatile =
>>> entry->var->data.memory_volatile;
>>> +            new_var->data.memory_restrict =
>>> entry->var->data.memory_restrict;
>>> +            new_var->data.image_format = entry->var->data.image_format;
>>
>> Is there a reason to only copy these fields for images?  It seems like
>> it should be safe to always do it.
> 
> Which fields you are talking about? I only copy the image fields because
> I'm sure they have to be propagated when arrays are split. Maybe, some
> other fields should be propagated as well but that would require a
> deeper analysis.

I mean the image fields are only copied inside the 'if
(entry->var->type->contains_image())'.  It seems like it should be safe
to just copy them always.

>>> +         }
>>> +
>>> +         entry->components[i] = new_var;
>>>            entry->var->insert_before(entry->components[i]);
>>>         }
>>>  
>>
>