[llvm] [Offload] Introduce offload-tblgen and initial new API implementation (PR #108413)

[Joseph Huber via llvm-commits]

================
@@ -0,0 +1,75 @@
+def : Macro {
+  let name = "OFFLOAD_APICALL";
+  let desc = "Calling convention for all API functions";
+  let condition = "defined(_WIN32)";
+  let value = "__cdecl";
+  let alt_value = "";
+}
+
+def : Macro {
+  let name = "OFFLOAD_APIEXPORT";
+  let desc = "Microsoft-specific dllexport storage-class attribute";
+  let condition = "defined(_WIN32)";
+  let value = "__declspec(dllexport)";
+  let alt_value = "";
+}
+
+def : Macro {
+  let name = "OFFLOAD_DLLEXPORT";
+  let desc = "Microsoft-specific dllexport storage-class attribute";
+  let condition = "defined(_WIN32)";
+  let value = "__declspec(dllexport)";
+}
+
+def : Macro {
+  let name = "OFFLOAD_DLLEXPORT";
+  let desc = "GCC-specific dllexport storage-class attribute";
+  let condition = "__GNUC__ >= 4";
+  let value = "__attribute__ ((visibility (\"default\")))";
+  let alt_value = "";
+}
+
+def : Typedef {
+  let name = "offload_bool_t";
+  let value = "uint8_t";
+  let desc = "compiler-independent type";
+}
+
+def : Handle {
+  let name = "offload_platform_handle_t";
----------------
jhuber6 wrote:

Is `platform` what we call a plugin currently and a device a subset of that? It's tough to think how we should handle this in what's exported, since I would say that ideally the user doesn't need to care about the "plugin" at all. However, that's not true for cases where some queue is only valid with CUDA and not AMDGPU for example. We could probably make it a quality of the device.

https://github.com/llvm/llvm-project/pull/108413