[clang] [AMDGPU][Clang] Support for type inferring extended image builtins for AMDGPU (PR #164358)

[Juan Manuel Martinez Caamaño via cfe-commits]

================
@@ -3,7 +3,7 @@
 
 typedef int int4 __attribute__((ext_vector_type(4)));
 typedef float float4 __attribute__((ext_vector_type(4)));
-typedef half half4 __attribute__((ext_vector_type(4)));
+typedef _Float16 half4 __attribute__((ext_vector_type(4)));
----------------
jmmartinez wrote:

> adding a new case requires new builtin def. one for HIP and another for OpenCL.

Not necessarly. Instead of describing the builtin signature as `"V4xiffQtV4ibii"` we add a new case, let's say `X` which becomes `__fp16` on OpenCL and `_Float16` on HIP. The previous signature becomes `"V4XiffQtV4ibii"`.

And in the switch in `ASTContext.cpp:12405` we add a new case to handle it:

```cpp
  case 'X':
    assert(HowLong == 0 && !Signed && !Unsigned &&
           "Bad modifiers used with 'X'!");
    Type = Type = Context.getLangOpts().OpenCL ? Context.HalfTy : Context.Float16Ty;;
    break;
```

At first sight, it looks strange to map one builtin to different types depending on the language. But it's what we already do for pointer address spaces (`v*3`, maps to `void __shared__*` on CUDA/HIP, `void __local__*` on OpenCL and `void __attribute__((addrspace(3)))*` for the rest).

https://github.com/llvm/llvm-project/pull/164358