[NewIR] support c_broadcast (#56284)

* [NewIR] support c_broadcast

* add legacyOpList

paddle/fluid/ir/dialect/utils.cc

@@ -22,7 +22,7 @@ const std::unordered_set<std::string> LegacyOpList = {
     "pd.fused_softmax_mask_upper_triangle_grad",
     "pd.load_combine",
     "pd.c_concat",
-    "pd.load_combine"};
+    "pd.c_broadcast_"};
 
 enum class AttrType {
   UNDEFINED = 0,

paddle/phi/api/yaml/legacy_ops.yaml

@@ -123,6 +123,16 @@
   backward : batch_norm_grad
   optional : reserve_space
 
+- op : c_broadcast
+  args : (Tensor x, int ring_id=0, int root=0, bool use_calc_stream=false)
+  output : Tensor(out)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : c_broadcast
+  inplace : (x -> out)
+
 - op : c_concat
   args : (Tensor x, int rank, int nranks, int ring_id, bool use_calc_stream, bool use_model_parallel)
   output : Tensor(out)

paddle/phi/api/yaml/op_compat.yaml

@@ -2990,6 +2990,12 @@
     yolo_loss : GetYoloLossExpectedKernelType
     yolo_loss_grad : GetYoloLossExpectedKernelType
 
+- op: c_broadcast
+  inputs :
+    x : X
+  outputs :
+    out : Out
+
 - op: channel_shuffle
   inputs:
     {x: X}