[OpenCL] __generic address space for CL2.0

[Sahasrabuddhe, Sameer]

On 11/14/2014 3:41 PM, Pekka Jääskeläinen wrote:
> On 11/14/2014 12:01 PM, Anastasia Stulova wrote:
>> One solution to that would be to add __generic to the end of the address
>> space enum list, but then we will have OpenCL items fragmented. I am not
>> sure it's good.
>
> Yep, I don't believe it's important to retain the API compatibility in
> this particular spot if everything else is freely changed.
>
I agree too. The Khronos SPIR generator has a similar patch, and seems 
like the standard thing to do.

Just one question: why not directly use the appropriate number for each 
target, instead of defaulting it to global? The provisional SPIR 2.0 
spec specifies "4" as the generic address space, which is a reasonable 
value to use ... maybe we can use that right away. SPIR is not a "real" 
target anyway, so there's no issue of breaking anything.

For other targets, we could use zero, which seems to mean "undefined 
address space, replaced with default". At least that's what the TCE and 
SPIR target do for CUDA address spaces! Then it is up to each target to 
specify the appropriate value when the OpenCL generic address space gets 
supported on that target.

Sameer.