[auto-monad] Refactor ascend_auto_monad

1. Remove output parameter pool for ascend control flow;
2. Remove duplicate code for Switch and SwitchLayer;
3. Add 'return' attribute to label_goto that used for return;
4. Disable tail call optimize for graphs with 'recursive' flag.

mindspore/ccsrc/utils/utils.h

@@ -398,6 +398,8 @@ constexpr auto kAttrNeedCseAfterRecompute = "need_cse_after_recompute";
 constexpr auto kAttrParallelDimInfo = "parallel_dim_info";
 constexpr auto kAttrStitch = "stitch";
 constexpr auto kAttrTopoSortRhsFirst = "topo_sort_rhs_first";
+constexpr auto kAttrSwitchLayer = "switch_layer";
+constexpr auto kAttrReturn = "return";
 
 // attr value
 constexpr auto kValueTargetSwitch = "target_switch";

mindspore/core/ir/func_graph.h

@@ -86,6 +86,7 @@ const char FUNC_GRAPH_FLAG_SPECIALIZE_PARAMETER[] = "spec_param";
 const char kFuncGraphFlagUndetermined[] = "Undeterminate";
 const char kFuncGraphFlagBackPropEntry[] = "BackPropEntry";
 const char kFuncGraphFlagReAutoMonad[] = "ReAutoMonad";
+const char kFuncGraphFlagRecursive[] = "Recursive";
 
 namespace abstract {
 class AbstractKeywordArg;

tests/st/control/test_switch_layer.py

@@ -24,11 +24,12 @@ from mindspore.common import dtype as mstype
 class CaseNet(nn.Cell):
     def __init__(self):
         super(CaseNet, self).__init__()
-        self.conv = nn.Conv2d(1, 3, 3)
+        self.conv = nn.Conv2d(1, 1, 3)
         self.relu = nn.ReLU()
+        self.relu1 = nn.ReLU()
         self.softmax = nn.Softmax()
         self.layers1 = (self.relu, self.softmax)
-        self.layers2 = (self.conv, self.relu)
+        self.layers2 = (self.conv, self.relu1)
 
     def construct(self, x, index1, index2):
         x = self.layers1[index1](x)
@@ -50,7 +51,3 @@ def test_switch_layer():
     true_value = relu(data)
     ret = np.allclose(value.asnumpy(), true_value.asnumpy())
     assert ret
-
-    idx3 = Tensor(3, mstype.int32)
-    with pytest.raises(IndexError):
-        value = net(data, idx3, idx2)