[llvm] [Offload] Implement the remaining initial Offload API (PR #122106)

[Joseph Huber via llvm-commits]

================
@@ -1327,6 +1327,34 @@ class CUDAGlobalHandlerTy final : public GenericGlobalHandlerTy {
     DeviceGlobal.setPtr(reinterpret_cast<void *>(CUPtr));
     return Plugin::success();
   }
+
+  Error getGlobalMetadataFromImage(GenericDeviceTy &Device,
+                                   DeviceImageTy &Image,
+                                   GlobalTy &ImageGlobal) override {
+    // If the image is an ELF we can use the generic path, otherwise fall back
+    // and use cuModuleGetGlobal to query the image.
+    if (utils::elf::isELF(Image.getMemoryBuffer().getBuffer())) {
+      return GenericGlobalHandlerTy::getGlobalMetadataFromImage(Device, Image,
+                                                                ImageGlobal);
+    }
+
+    CUDADeviceImageTy &CUDAImage = static_cast<CUDADeviceImageTy &>(Image);
----------------
jhuber6 wrote:

It's very distinct, `Image` means "look up up from an ELF I have right here" while `Device` means "Look it up using the vendor API after it's been loaded". The former is way faster.

https://github.com/llvm/llvm-project/pull/122106