[XLA:GPU] Short-circuit compilation of no-op IR -> empty PTX.

There's no point constructing/running LLVM pipeline if we know that we have no
kernels in the IR we've generated for the given HLO op. This is often the case
for ops we can optimize away at the HLO level.

PiperOrigin-RevId: 174072540

tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/gpu_backend_lib.cc

@@ -342,6 +342,13 @@ StatusOr<string> CompileModuleToPtx(llvm::Module* module,
                                     std::pair<int, int> compute_capability,
                                     const HloModuleConfig& hlo_module_config,
                                     const string& libdevice_dir_path) {
+  // If the module has no functions or globals, there's nothing to compile. Just
+  // return an empty string.
+  if (module->empty() && module->global_empty()) {
+    VLOG(2) << "Module '" << llvm_ir::AsString(module->getName())
+            << "' is empty. Skipping compilation.";
+    return string();
+  }
   // Link the input module with libdevice, to pull in implementations of some
   // builtins.
   TF_RETURN_IF_ERROR(