!8097 Remove redundant phi nodes

From: @ginfung
Reviewed-by: 
Signed-off-by:

mindspore/ccsrc/pipeline/jit/parse/function_block.cc

@@ -49,11 +49,11 @@ AnfNodePtr FunctionBlock::ReadVariable(const std::string &var) {
   if (vars_.count(var)) {
     AnfNodePtr node = vars_[var];
     MS_EXCEPTION_IF_NULL(node);
-    if (node->isa<ValueNode>()) {
-      return NewValueNode(GetValueNode(node));
-    } else {
-      return node;
+    auto iter = resolve_to_removable_phis_.find(node);
+    if (iter != resolve_to_removable_phis_.end()) {
+      return iter->second;
     }
+    return node;
   }
   // get var from predecessor block ,if can't get the make a resolve node to it
   if (matured_) {
@@ -64,7 +64,13 @@ AnfNodePtr FunctionBlock::ReadVariable(const std::string &var) {
       return block->ReadVariable(var);
     } else if (prev_blocks_.empty()) {
       // get namespace and make Reslove
-      return MakeResolveSymbol(var);
+      auto it = var_to_resolve_.find(var);
+      if (it != var_to_resolve_.end()) {
+        return it->second;
+      }
+      auto tmp_node = MakeResolveSymbol(var);
+      var_to_resolve_[var] = tmp_node;
+      return tmp_node;
     }
   }
   // If have more than one predecessor blocks then build a phi node.
@@ -217,6 +223,7 @@ bool FunctionBlock::CollectRemovablePhi(const ParameterPtr &phi) {
     // replace var with new one. This equal to statement in TR "v0 is immediately replaced by v1."
     WriteVariable(var, arg_node);
     removable_phis_[phi] = arg_node;
+    resolve_to_removable_phis_[arg_node] = phi;
     // The following equal to statement "The φ-function defining v1, which now reads φ(v2, v1), is optimized
     // recursively". check if phi1 is assigned with this phi before, then phi1 can be replaced with arg_node.
     for (auto &prev : prev_blocks_) {

mindspore/ccsrc/pipeline/jit/parse/function_block.h

@@ -101,12 +101,21 @@ class FunctionBlock : public std::enable_shared_from_this<FunctionBlock> {
   // keeps all removable phis which will be removed in one pass.
   std::unordered_map<ParameterPtr, AnfNodePtr> removable_phis_;
 
+  // Keeps the map for the resolve node to the removable phi node.
+  // For the case that ReadVariable returns a phi node although this phi node
+  // generated in the prev block is identified as removable. The other blocks
+  // should find this phi node.
+  std::unordered_map<AnfNodePtr, ParameterPtr> resolve_to_removable_phis_;
+
   // hold declared global variables in function
   std::set<std::string> global_vars_;
 
   // other depend need to insert before function return nodes.
   // summary or some other node
   std::vector<AnfNodePtr> auto_depends_;
+
+  // keeps the new made resolve symbol for the variable not found in vars_.
+  std::unordered_map<std::string, AnfNodePtr> var_to_resolve_;
 };
 
 }  // namespace parse

tests/st/control/test_cont_grad.py

@@ -14,6 +14,7 @@
 # ============================================================================
 """ test control ops """
 import numpy as np
+import pytest
 
 from mindspore import dtype as ms
 from mindspore import Tensor
@@ -1150,3 +1151,147 @@ def test_if_by_if_forward_all_const_branch():
     end = Tensor(np.array(3), dtype=ms.float32)
     x = Tensor(np.array(0), dtype=ms.float32)
     net(idx, end, x)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_if_const_grad():
+    class MyNet(nn.Cell):
+        def __init__(self):
+            super().__init__()
+            self.add = P.TensorAdd()
+
+        def construct(self, *inputs):
+            out = self.add(*inputs)
+            return out
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+            self.weights = ParameterTuple(net.trainable_params())
+
+        def construct(self, *inputs):
+            a = 1
+            b = 2
+            if a > 0:
+                b = 1
+            a += b
+            return grad_by_list(self.net, self.weights)(*inputs)
+
+    context.set_context(mode=context.GRAPH_MODE)
+    my_net = MyNet()
+    net = GradNet(my_net)
+    a = Tensor(np.array(0), dtype=ms.int32)
+    b = Tensor(np.array(1), dtype=ms.int32)
+    net(a, b)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_if_by_if_const_grad():
+    class MyNet(nn.Cell):
+        def __init__(self):
+            super().__init__()
+            self.add = P.TensorAdd()
+
+        def construct(self, *inputs):
+            out = self.add(*inputs)
+            return out
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+            self.weights = ParameterTuple(net.trainable_params())
+
+        def construct(self, *inputs):
+            a = 1
+            b = 2
+            if a > 0:
+                b = 1
+            if a < 0:
+                b = 0
+            if a == 0:
+                b = 3
+            a += b
+            return grad_by_list(self.net, self.weights)(*inputs)
+
+    context.set_context(mode=context.GRAPH_MODE)
+    my_net = MyNet()
+    net = GradNet(my_net)
+    a = Tensor(np.array(0), dtype=ms.int32)
+    b = Tensor(np.array(1), dtype=ms.int32)
+    net(a, b)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_while_const_grad():
+    class MyNet(nn.Cell):
+        def __init__(self):
+            super().__init__()
+            self.add = P.TensorAdd()
+
+        def construct(self, *inputs):
+            out = self.add(*inputs)
+            return out
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+            self.weights = ParameterTuple(net.trainable_params())
+
+        def construct(self, *inputs):
+            a = 1
+            while a > 1:
+                a = a - 1
+            return grad_by_list(self.net, self.weights)(*inputs)
+
+    context.set_context(mode=context.GRAPH_MODE)
+    my_net = MyNet()
+    net = GradNet(my_net)
+    a = Tensor(np.array(0), dtype=ms.int32)
+    b = Tensor(np.array(1), dtype=ms.int32)
+    net(a, b)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_if_by_while_const_grad():
+    class MyNet(nn.Cell):
+        def __init__(self):
+            super().__init__()
+            self.add = P.TensorAdd()
+
+        def construct(self, *inputs):
+            out = self.add(*inputs)
+            return out
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+            self.weights = ParameterTuple(net.trainable_params())
+
+        def construct(self, *inputs):
+            a = 1
+            b = 2
+            if a > 0:
+                b = 0
+            while a > 1:
+                a = a - 1
+            a += b
+            return grad_by_list(self.net, self.weights)(*inputs)
+
+    context.set_context(mode=context.GRAPH_MODE)
+    my_net = MyNet()
+    net = GradNet(my_net)
+    a = Tensor(np.array(0), dtype=ms.int32)
+    b = Tensor(np.array(1), dtype=ms.int32)
+    net(a, b)

tests/ut/cpp/optimizer/opt_test.cc

@@ -187,7 +187,7 @@ TEST_F(TestOptOpt, CSE) {
   FuncGraphManagerPtr manager1 = Manage(test_graph1);
   draw::Draw("opt_cse_before_1.dot", test_graph1);
 
-  ASSERT_EQ(manager1->all_nodes().size(), 10);
+  ASSERT_EQ(manager1->all_nodes().size(), 9);
 
   auto cse = std::make_shared<CSE>();
   ASSERT_TRUE(cse != nullptr);
@@ -205,7 +205,7 @@ TEST_F(TestOptOpt, CSE) {
 
   FuncGraphManagerPtr manager2 = Manage(test_graph2);
   draw::Draw("opt_cse_before_2.dot", test_graph2);
-  ASSERT_EQ(manager2->all_nodes().size(), 22);
+  ASSERT_EQ(manager2->all_nodes().size(), 16);
   is_changed = cse->Cse(test_graph2, manager2);
   ASSERT_TRUE(is_changed);
   ASSERT_EQ(manager2->all_nodes().size(), 12);