Currently, for AMDGPU, when compiling for OpenCL, we unconditionally use
`private` as the default address space. This is wrong for cases where
the `generic` address space is available, and is corrected via this
patch. In general, this AS map abuse is a bad hack and we should re-work
it altogether, but at least after this patch we will stop being
incorrect for e.g. OpenCL 2.0.
The __builtin_alloca was returning a flat pointer with no address space
when compiled using openCL1.2 or below but worked fine with openCL2.0
and above. This accounts to the fact that later uses the concept of
generic address space which supports cast to other address space(i.e to
private address space which is used for stack allocation) .
But, in actuality, as it returns pointer to the stack, it should be
pointing to private address space irrespective of openCL version becuase
builtin_alloca allocates stack memory used for current function in which
it is called. Thus,it requires redefintion of the builtin function with
appropraite return pointer to private address space.
