[Intel-gfx] [PATCH v5 17/19] drm/i915: Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset

[Michel Thierry]

On 7/28/2015 3:43 PM, Chris Wilson wrote:
> On Tue, Jul 28, 2015 at 12:12:11PM +0100, Michel Thierry wrote:
>> On 7/27/2015 10:11 PM, Chris Wilson wrote:
>>> On Thu, Jul 16, 2015 at 10:33:29AM +0100, Michel Thierry wrote:
>>>> +	if (!(entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
>>>> +	    (vma->node.start + vma->node.size) >= (1ULL << 32))
>>>> +		return true;
>>>
>>> gcc completely screwed this up here and used 0 for 1ULL<<32.
>>>
>>> Note that we can allow state + size == 4G (since the end is exclusive),
>>> so I went with
>>>
>>>    if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
>>>        (vma->node.start + vma->node.size - 1) >> 32)
>>>           return true;
>>>
>>> instead.
>>> -Chris
>>>
>>
>> Thanks, I'll include this change in the next patch version.
>
> I've also got a couple of other stylistic changes, plus an earlier
> request:
> ...
I updated my patch with these changes, thanks.

> diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> index 209e8e2b07be..78fc8810d6e0 100644
> --- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> +++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
> @@ -680,8 +680,8 @@ eb_vma_misplaced(struct i915_vma *vma)
>          if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
>                  return !only_mappable_for_reloc(entry->flags);
>
> -       if (!(entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) &&
> -           (vma->node.start + vma->node.size) >= (1ULL << 32))
> +       if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
> +           (vma->node.start + vma->node.size - 1) >> 32)
>                  return true;
>
>          return false;
>
Also updated this part to follow your suggestion.

I also spent a bit more time trying to figure what gcc was doing here. 
But I can't reproduce it locally, I get sizeof(1ULL<<32) = 8 and 
1ULL<<32 doesn't seem to be zero (in 32 and 64 bit kernels).

Could it be related to the gcc version? I'm using 4.8.4.

-Michel