Reorder optimization passes (#2433)

CSE benefits from lifting constants to initializers, and from
initializer deduplication. Hence, it is better to have CSE after the
other two. Furthermore, we want this to apply to all constants,
including `Constant(value_int=1)` etc., so set those options for lifting
constants.

---------

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

Author: gramalingam

onnxscript/optimizer/_function_folding_test.py

@@ -151,7 +151,7 @@ def test_fold_nested_if_function_succeeds(self):
         optimized = optimizer.optimize(model, onnx_shape_inference=False, inline=True)
 
         self.assertEqual(len(optimized.functions), 0)
-        self.assertEqual(len(optimized.graph), 2)
+        self.assertEqual(len(optimized.graph), 1)
         self.assertNotIn("If", {n.op_type for n in optimized.graph})
 
 

onnxscript/optimizer/_optimizer.py

@@ -53,10 +53,10 @@ def optimize_ir(
             early_stop=stop_if_no_change,
         ),
         common_passes.RemoveUnusedNodesPass(),
-        common_passes.CommonSubexpressionEliminationPass(),
-        common_passes.LiftConstantsToInitializersPass(),
+        common_passes.LiftConstantsToInitializersPass(lift_all_constants=True, size_limit=0),
         common_passes.LiftSubgraphInitializersToMainGraphPass(),
         common_passes.DeduplicateInitializersPass(),
+        common_passes.CommonSubexpressionEliminationPass(),
     ]
     if inline:
         # Inline all functions first before optimizing