[clang] [llvm] [Offload] Provide a kernel library useable by the offload runtime (PR #104168)

[Joseph Huber via cfe-commits]

================
@@ -393,22 +393,17 @@ struct CUDADeviceTy : public GenericDeviceTy {
     return Plugin::success();
   }
 
-  virtual Error callGlobalConstructors(GenericPluginTy &Plugin,
-                                       DeviceImageTy &Image) override {
-    // Check for the presense of global destructors at initialization time. This
-    // is required when the image may be deallocated before destructors are run.
-    GenericGlobalHandlerTy &Handler = Plugin.getGlobalHandler();
-    if (Handler.isSymbolInImage(*this, Image, "nvptx$device$fini"))
-      Image.setPendingGlobalDtors();
-
-    return callGlobalCtorDtorCommon(Plugin, Image, /*IsCtor=*/true);
+  virtual Expected<StringRef>
+  getGlobalConstructorName(DeviceImageTy &Image) override {
+    if (auto Err = prepareGlobalCtorDtorCommon(Image, /*IsCtor=*/true))
+      return Err;
+    return "nvptx$device$init";
   }
-
-  virtual Error callGlobalDestructors(GenericPluginTy &Plugin,
-                                      DeviceImageTy &Image) override {
-    if (Image.hasPendingGlobalDtors())
-      return callGlobalCtorDtorCommon(Plugin, Image, /*IsCtor=*/false);
-    return Plugin::success();
+  virtual Expected<StringRef>
+  getGlobalDestructorName(DeviceImageTy &Image) override {
+    if (auto Err = prepareGlobalCtorDtorCommon(Image, /*IsCtor=*/false))
----------------
jhuber6 wrote:

What happened to the `pendingGlobalDtors` thing? That was required since some configs (i.e. JIT) don't have permanent storage so they could be deallocated during teardown. The better solution would be to make an API that copied in its own buffer (This is what hsa_executable_freeze is AFAIK).

https://github.com/llvm/llvm-project/pull/104168