delete duplicated optimize in AddFunction (#56192)

paddle/cinn/backends/codegen_cuda_util.h

@@ -60,14 +60,15 @@ struct CollectHostFunctionVisitor : public ir::IRMutator<> {
  private:
   void Visit(const ir::_LoweredFunc_* op, Expr* expr) override {
     if (op->body.As<ir::Call>()) {
-      host_module_builder.AddFunction(expr->as_lowered_func_ref());
+      host_module_builder.AddFunctionWithoutOptim(expr->as_lowered_func_ref());
     } else {
       if (!op->cuda_axis_info.valid()) {
         expr->as_lowered_func_ref()->cuda_axis_info.set_valid(true);
       }
       auto host_func = CreateHostFunctionGivenDeviceKernel(op);
-      host_module_builder.AddFunction(host_func.as_lowered_func_ref());
-      device_module_builder.AddFunction(
+      host_module_builder.AddFunctionWithoutOptim(
+          host_func.as_lowered_func_ref());
+      device_module_builder.AddFunctionWithoutOptim(
           CreateDeviceFunctionGivenDeviceKernel(*expr).as_lowered_func_ref());
     }
   }

paddle/cinn/ir/module.cc

@@ -30,6 +30,10 @@ void Module::Builder::AddFunction(ir::LoweredFunc func) {
   module_->functions.push_back(func);
 }
 
+void Module::Builder::AddFunctionWithoutOptim(const ir::LoweredFunc &func) {
+  module_->functions.push_back(func);
+}
+
 void Module::Builder::AddBuffer(ir::Buffer buffer) {
   CHECK(buffer->target.defined())
       << "buffer [" << buffer->name << "]'s target is undefined";

paddle/cinn/ir/module.h

@@ -42,6 +42,7 @@ class Module : public ir::IrNodeRef {
     }
 
     void AddFunction(ir::LoweredFunc func);
+    void AddFunctionWithoutOptim(const ir::LoweredFunc& func);
     void AddBuffer(ir::Buffer buffer);
     void Clear();