!49472 Fixed the bug of side effect in fallback.

Merge pull request !49472 from Margaret_wangrui/fallback_side_effect

mindspore/ccsrc/pipeline/jit/fallback.cc

@@ -54,13 +54,14 @@ AnfNodePtr ConvertInterpretedObjectToPyExecute(const FuncGraphPtr &fg, const Val
   // Build new CNode for value node.
   ValuePtrList keys({std::make_shared<StringImm>(value_node_key)});
   ValuePtrList values({std::make_shared<StringImm>(value_node_key)});
-  const auto interpreted_cnode = fg->NewCNodeInOrder({NewValueNode(prim::kPrimPyExecute), NewValueNode(script_str),
-                                                      NewValueNode(std::make_shared<ValueTuple>(keys)),
-                                                      NewValueNode(std::make_shared<ValueTuple>(values))});
+  const auto interpreted_cnode = fg->NewCNode({NewValueNode(prim::kPrimPyExecute), NewValueNode(script_str),
+                                               NewValueNode(std::make_shared<ValueTuple>(keys)),
+                                               NewValueNode(std::make_shared<ValueTuple>(values))});
   constexpr auto debug_recursive_level = 2;
   MS_LOG(DEBUG) << "original node: " << node->DebugString(debug_recursive_level)
                 << ", interpreted_cnode: " << interpreted_cnode->DebugString(debug_recursive_level);
   interpreted_cnode->set_debug_info(node->debug_info());
+  fg->ReplaceInOrder(node, interpreted_cnode);
   return interpreted_cnode;
 }
 }  // namespace mindspore

mindspore/ccsrc/pipeline/jit/static_analysis/static_analysis.cc

@@ -373,7 +373,7 @@ EvalResultPtr AnalysisEngine::InterpretedNodeCall(const CNodePtr &cnode, const A
   auto fg = cnode->func_graph();
   const auto value_tuple_node = fg->NewCNode(value_list);
 
-  const auto getattr_obj_call_node = fg->NewCNodeInOrder(
+  const auto getattr_obj_call_node = fg->NewCNode(
     {NewValueNode(prim::kPrimPyExecute), NewValueNode(script_call_str), NewValueNode(key_tuple), value_tuple_node});
   MS_LOG(DEBUG) << "getattr_obj_call_node: " << getattr_obj_call_node->DebugString();
 
@@ -382,7 +382,7 @@ EvalResultPtr AnalysisEngine::InterpretedNodeCall(const CNodePtr &cnode, const A
   AnalysisEnginePtr eng = conf->engine();
   MS_EXCEPTION_IF_NULL(eng);
   AnfNodeConfigPtr fn_conf = eng->MakeConfig(getattr_obj_call_node, conf->context(), conf->func_graph());
-  return eng->ForwardConfig(conf, fn_conf);
+  return eng->ForwardConfig(conf, fn_conf, false);
 }
 
 AbstractBasePtr AnalysisEngine::GetCNodeOperatorAbstract(const CNodePtr &cnode, const AnalysisContextPtr &context,
@@ -807,7 +807,8 @@ EvaluatorPtr AnalysisEngine::GetEvaluatorFor(const AbstractFunctionPtr &func) {
   MS_LOG(EXCEPTION) << "Cannot GetEvaluator from " << func->type_name();
 }
 
-EvalResultPtr AnalysisEngine::ForwardConfig(const AnfNodeConfigPtr &orig_conf, const AnfNodeConfigPtr new_conf) {
+EvalResultPtr AnalysisEngine::ForwardConfig(const AnfNodeConfigPtr &orig_conf, const AnfNodeConfigPtr new_conf,
+                                            bool need_erase) {
   MS_EXCEPTION_IF_NULL(orig_conf);
   MS_EXCEPTION_IF_NULL(new_conf);
   // If always_eval_flag is true in BaseFuncGraphEvaluaotr, then the CNode with same orig_conf may be forwarded
@@ -816,7 +817,7 @@ EvalResultPtr AnalysisEngine::ForwardConfig(const AnfNodeConfigPtr &orig_conf, c
   MS_LOG(DEBUG) << "Forward orig_conf: " << orig_conf->ToString() << ", to new_conf: " << new_conf->ToString();
   auto old_cnode = orig_conf->node()->cast_ptr<CNode>();
   auto new_cnode = new_conf->node()->cast<CNodePtr>();
-  if (old_cnode != nullptr && new_cnode != nullptr) {
+  if (need_erase && old_cnode != nullptr && new_cnode != nullptr) {
     if (old_cnode->func_graph() == new_cnode->func_graph()) {
       MS_LOG(DEBUG) << "Try to remove forward node from order list, forward node: " << new_cnode->DebugString()
                     << ", as origin node should be in order list, origin_node: " << old_cnode->DebugString();

mindspore/ccsrc/pipeline/jit/static_analysis/static_analysis.h

@@ -1,7 +1,7 @@
 /**
  * This is the C++ adaptation and derivative work of Myia (https://github.com/mila-iqia/myia/).
  *
- * Copyright 2019-2022 Huawei Technologies Co., Ltd
+ * Copyright 2019-2023 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -322,7 +322,8 @@ class AnalysisEngine : public std::enable_shared_from_this<AnalysisEngine> {
 
   // Set the analysis result for orig to the result for new.
   // This sets an entry in anfnode_config_map from orig to new.
-  EvalResultPtr ForwardConfig(const AnfNodeConfigPtr &orig_conf, const AnfNodeConfigPtr new_conf);
+  EvalResultPtr ForwardConfig(const AnfNodeConfigPtr &orig_conf, const AnfNodeConfigPtr new_conf,
+                              bool need_erase = true);
   const PrimEvaluatorMap &PrimConstructors() const { return prim_constructors_; }
 
   FuncGraphPtr root_func_graph() const { return root_func_graph_; }

tests/st/fallback/test_graph_fallback_runtime.py

@@ -902,6 +902,7 @@ class UNet(ms.nn.Cell):
         return out
 
 
+@pytest.mark.skip(reason="No support PyExecute Add.")
 @pytest.mark.level0
 @pytest.mark.platform_x86_gpu_training
 @pytest.mark.platform_arm_ascend_training
@@ -916,7 +917,8 @@ def test_resolve_cust_class():
     net = UNet(UserDefinedNet())
     x = np.array([10], np.float32)
     output = net(ms.Tensor(x))
-    print(output)  # The output should == 200, but failed, check later.
+    print(output)
+    assert output == 200
 
 
 @pytest.mark.level0
@@ -935,3 +937,34 @@ def test_resolve_cust_ms_function_call_class():
     with pytest.raises(RuntimeError) as err:
         net(ms.Tensor(x))
     assert "Nested execution during JIT execution is not supported." in str(err.value)
+
+
+class PrintPyExecuteNet(ms.nn.Cell):
+    def __init__(self, net):
+        super().__init__()
+        self.net = net
+
+    def construct(self, x):
+        out = x * x
+        print("out1:", out)
+        out = self.net(x) + out
+        print("out2:", out)
+        return out
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_print_pyexecute():
+    """
+    Feature: Side effect in Fallback runtime.
+    Description: Side effect in Fallback runtime.
+    Expectation: No error.
+    """
+    net = PrintPyExecuteNet(UserDefinedNet())
+    x = np.array([10], np.float64)
+    output = net(ms.Tensor(x))
+    print(output)
+    assert output == 200