Bug 93977

The handling of arguments seems to not account for the required alignment of pointer kernel arguments. The alignment is tracked for the pointer value itself, but nothing is recording the alignment of the pointer element type needed for the runtime to provide the correctly aligned pointer.

I believe this is why the conformance test kernel_memory_alignment_local fails for all of the parameter tests. The global and constant versions seem to work, I'm assuming because every allocated buffer happens to have an alignment of 16 or higher anyway.

Also here:
         const unsigned arg_store_size = TD.getTypeStoreSize(arg_type);

         // OpenCL 1.2 specification, Ch. 6.1.5: "A built-in data
         // type that is not a power of two bytes in size must be
         // aligned to the next larger power of two".  We need this
         // alignment for three element vectors, which have
         // non-power-of-2 store size.
         const unsigned arg_api_size = util_next_power_of_two(arg_store_size);

TD.getTypeAllocSize(arg_type) should be used instead of manually rounding it.