[Libomptarget] Statically link all plugin runtimes

Summary:
This patch overhauls the `libomptarget` and plugin interface. Currently,
we define a C API and compile each plugin as a separate shared library.
Then, `libomptarget` loads these API functions and forwards its internal
calls to them. This was originally designed to allow multiple
implementations of a library to be live. However, since then no one has
used this functionality and it prevents us from using much nicer
interfaces. If the old behavior is desired it should instead be
implemented as a separate plugin.

This patch replaces the `PluginAdaptorTy` interface with the
`GenericPluginTy` that is used by the plugins. Each plugin exports a
`createPlugin_<name>` function that is used to get the specific
implementation. This code is now shared with `libomptarget`.

There are some notable improvements to this.
1. Massively improved lifetimes of life runtime objects
2. The plugins can use a C++ interface
3. Global state does not need to be duplicated for each plugin +
   libomptarget
4. Easier to use and add features and improve error handling
5. Less function call overhead / Improved LTO performance.

Additional changes in this plugin are related to contending with the
fact that state is now shared. Initialization and deinitialization is
now handled correctly and in phase with the underlying runtime, allowing
us to actually know when something is getting deallocated.

Depends on llvm#86971 llvm#86875 llvm#86868

Author: jhuber6

clang/test/Driver/linker-wrapper-image.c

@@ -30,8 +30,8 @@
 
 //      OPENMP: define internal void @.omp_offloading.descriptor_reg() section ".text.startup" {
 // OPENMP-NEXT: entry:
-// OPENMP-NEXT:   %0 = call i32 @atexit(ptr @.omp_offloading.descriptor_unreg)
 // OPENMP-NEXT:   call void @__tgt_register_lib(ptr @.omp_offloading.descriptor)
+// OPENMP-NEXT:   %0 = call i32 @atexit(ptr @.omp_offloading.descriptor_unreg)
 // OPENMP-NEXT:   ret void
 // OPENMP-NEXT: }
 

llvm/lib/Frontend/Offloading/OffloadWrapper.cpp

@@ -232,12 +232,13 @@ void createRegisterFunction(Module &M, GlobalVariable *BinDesc,
   // Construct function body
   IRBuilder<> Builder(BasicBlock::Create(C, "entry", Func));
 
+  Builder.CreateCall(RegFuncC, BinDesc);
+
   // Register the destructors with 'atexit'. This is expected by the CUDA
   // runtime and ensures that we clean up before dynamic objects are destroyed.
-  // This needs to be done before the runtime is called and registers its own.
+  // This needs to be done after plugin initialization to ensure that it is
+  // called before the plugin runtime is destroyed.
   Builder.CreateCall(AtExit, UnregFunc);
-
-  Builder.CreateCall(RegFuncC, BinDesc);
   Builder.CreateRetVoid();
 
   // Add this function to constructors.

openmp/libomptarget/include/PluginManager.h

@@ -16,7 +16,6 @@
 #include "DeviceImage.h"
 #include "ExclusiveAccess.h"
 #include "Shared/APITypes.h"
-#include "Shared/PluginAPI.h"
 #include "Shared/Requirements.h"
 
 #include "device.h"
@@ -34,38 +33,10 @@
 #include <mutex>
 #include <string>
 
-struct PluginManager;
-
-/// Plugin adaptors should be created via `PluginAdaptorTy::create` which will
-/// invoke the constructor and call `PluginAdaptorTy::init`. Eventual errors are
-/// reported back to the caller, otherwise a valid and initialized adaptor is
-/// returned.
-struct PluginAdaptorTy {
-  /// Try to create a plugin adaptor from a filename.
-  static llvm::Expected<std::unique_ptr<PluginAdaptorTy>>
-  create(const std::string &Name);
-
-  /// Name of the shared object file representing the plugin.
-  std::string Name;
-
-  /// Access to the shared object file representing the plugin.
-  std::unique_ptr<llvm::sys::DynamicLibrary> LibraryHandler;
-
-#define PLUGIN_API_HANDLE(NAME)                                                \
-  using NAME##_ty = decltype(__tgt_rtl_##NAME);                                \
-  NAME##_ty *NAME = nullptr;
+#include "PluginInterface.h"
+using GenericPluginTy = llvm::omp::target::plugin::GenericPluginTy;
 
-#include "Shared/PluginAPI.inc"
-#undef PLUGIN_API_HANDLE
-
-  /// Create a plugin adaptor for filename \p Name with a dynamic library \p DL.
-  PluginAdaptorTy(const std::string &Name,
-                  std::unique_ptr<llvm::sys::DynamicLibrary> DL);
-
-  /// Initialize the plugin adaptor, this can fail in which case the adaptor is
-  /// useless.
-  llvm::Error init();
-};
+struct PluginManager;
 
 /// Struct for the data required to handle plugins
 struct PluginManager {
@@ -80,6 +51,8 @@ struct PluginManager {
 
   void init();
 
+  void deinit();
+
   // Register a shared library with all (compatible) RTLs.
   void registerLib(__tgt_bin_desc *Desc);
 
@@ -92,10 +65,10 @@ struct PluginManager {
         std::make_unique<DeviceImageTy>(TgtBinDesc, TgtDeviceImage));
   }
 
-  /// Initialize as many devices as possible for this plugin adaptor. Devices
-  /// that fail to initialize are ignored. Returns the offset the devices were
-  /// registered at.
-  void initDevices(PluginAdaptorTy &RTL);
+  /// Initialize as many devices as possible for this plugin. Devices  that fail
+  /// to initialize are ignored. Returns the offset the devices were registered
+  /// at.
+  void initDevices(GenericPluginTy &RTL);
 
   /// Return the device presented to the user as device \p DeviceNo if it is
   /// initialized and ready. Otherwise return an error explaining the problem.
@@ -151,8 +124,8 @@ struct PluginManager {
   // Initialize all plugins.
   void initAllPlugins();
 
-  /// Iterator range for all plugin adaptors (in use or not, but always valid).
-  auto pluginAdaptors() { return llvm::make_pointee_range(PluginAdaptors); }
+  /// Iterator range for all plugins (in use or not, but always valid).
+  auto plugins() { return llvm::make_pointee_range(Plugins); }
 
   /// Return the user provided requirements.
   int64_t getRequirements() const { return Requirements.getRequirements(); }
@@ -164,14 +137,14 @@ struct PluginManager {
   bool RTLsLoaded = false;
   llvm::SmallVector<__tgt_bin_desc *> DelayedBinDesc;
 
-  // List of all plugin adaptors, in use or not.
-  llvm::SmallVector<std::unique_ptr<PluginAdaptorTy>> PluginAdaptors;
+  // List of all plugins, in use or not.
+  llvm::SmallVector<std::unique_ptr<GenericPluginTy>> Plugins;
 
-  // Mapping of plugin adaptors to offsets in the device table.
-  llvm::DenseMap<const PluginAdaptorTy *, int32_t> DeviceOffsets;
+  // Mapping of plugins to offsets in the device table.
+  llvm::DenseMap<const GenericPluginTy *, int32_t> DeviceOffsets;
 
-  // Mapping of plugin adaptors to the number of used devices.
-  llvm::DenseMap<const PluginAdaptorTy *, int32_t> DeviceUsed;
+  // Mapping of plugins to the number of used devices.
+  llvm::DenseMap<const GenericPluginTy *, int32_t> DeviceUsed;
 
   // Set of all device images currently in use.
   llvm::DenseSet<const __tgt_device_image *> UsedImages;