[llvm] [Offload] Use `amd_signal_async_handler` for host function calls (PR #154131)

[Matt Arsenault via llvm-commits]

================
@@ -1525,13 +1530,22 @@ struct AMDGPUStreamTy {
       InputSignal = consume(OutputSignal).second;
     }
 
-    // "Leaking" the thread here is consistent with other work added to the
-    // queue. The input and output signals will remain valid until the output is
-    // signaled.
-    std::thread(CallbackWrapper, InputSignal, OutputSignal, Callback, UserData)
-        .detach();
+    auto *CallbackData = new CallbackDataType{Callback, UserData, OutputSignal};
+    if (InputSignal && InputSignal->load()) {
+      hsa_status_t Status = hsa_amd_signal_async_handler(
+          InputSignal->get(), HSA_SIGNAL_CONDITION_EQ, 0, callbackWrapper,
+          CallbackData);
 
-    return Plugin::success();
+      return Plugin::check(Status, "error in hsa_amd_signal_async_handler: %s");
+    } else {
----------------
arsenm wrote:

No else after return

https://github.com/llvm/llvm-project/pull/154131