move parallel factor pattern to canonicalization pass

include/Dialect/TensorExt/IR/TensorExtCanonicalization.td

@@ -82,4 +82,18 @@ def RotatePlusExtractToIndexedExtract : Pat<
     (MakeSingleResultVariadic (Arith_AddIOp $shift, $index, DefOverflow)))
 >;
 
+// Rotating two tensors by the same amount can be converted to a single
+// post-rotation. This can result in eliminating either the rotation (because
+// it can be combined with a later rotation) or the arith op itself, if it is
+// is identical to an existing arith op applied before the rotation.
+foreach ArithOp = [Arith_AddIOp, Arith_SubIOp, Arith_MulIOp] in {
+  def FactorParallelRotationsThroughOp_#ArithOp : Pat<
+    (ArithOp
+      (TensorExt_RotateOp $t1, $i),
+      (TensorExt_RotateOp $t2, $i),
+      $ovf),
+    (TensorExt_RotateOp (ArithOp $t1, $t2, $ovf), $i)
+  >;
+}
+
 #endif  // INCLUDE_DIALECT_TENSOREXT_IR_TENSOREXTCANONICALIZATION_TD_

include/Dialect/TensorExt/Transforms/TensorExtRotationAlignment.td

@@ -80,18 +80,4 @@ foreach InnerOp = [Arith_AddIOp, Arith_SubIOp, Arith_MulIOp] in {
   }
 }
 
-// Rotating two tensors by the same amount can be converted to a single
-// post-rotation. This can result in eliminating either the rotation (because
-// it can be combined with a later rotation) or the arith op itself, if it is
-// is identical to an existing arith op applied before the rotation.
-foreach ArithOp = [Arith_AddIOp, Arith_SubIOp, Arith_MulIOp] in {
-  def FactorParallelRotationsThroughOp_#ArithOp : Pat<
-    (ArithOp
-      (TensorExt_RotateOp $t1, $i),
-      (TensorExt_RotateOp $t2, $i),
-      $ovf),
-    (TensorExt_RotateOp (ArithOp $t1, $t2, $ovf), $i)
-  >;
-}
-
 #endif  // INCLUDE_DIALECT_TENSOREXT_IR_TENSOREXTPATTERNS_TD_