[llvm] [libsycl] Introduce Kernel name based cache (PR #225865)

[Sergey Semenov via llvm-commits]

================
@@ -45,9 +50,31 @@ class DeviceKernelInfo {
   /// \return the device image containing the device code of this kernel.
   DeviceImageManager &getDeviceImage() const { return MDeviceImage; }
 
+  /// Returns a cache entry for pair key \p Context & \p Device if successful,
+  /// null otherwise
+  ol_symbol_handle_t tryGetCachedKernel(ContextImpl *Context,
+                                        ol_device_handle_t Device);
+
+  /// Adds \p Kernel to the cache by using the pair key \p Context & \p Device
+  void addCachedKernel(ContextImpl *Context, ol_device_handle_t Device,
+                       ol_symbol_handle_t Kernel);
+
+  /// Removes every cache entry keyed by \p Context, regardless of device.
+  /// Called by ContextImpl when it is being destroyed.
+  void removeCachedKernelsFor(ContextImpl *Context);
+
 private:
   std::string_view MName;
   DeviceImageManager &MDeviceImage;
+
+  using CacheKeyT = std::pair<ContextImpl *, ol_device_handle_t>;
+  struct CacheKeyHash {
+    std::size_t operator()(const CacheKeyT &Key) const noexcept {
+      return std::hash<ContextImpl *>{}(Key.first);
----------------
sergey-semenov wrote:

Shouldn't we have the device part of the key be represented in the hash too?

https://github.com/llvm/llvm-project/pull/225865