data
type is factored out by the introduction of a separate type which explicitly represents the different variants within the original type.
data Expr = Lit Float 
Add Expr Expr 
Mul Expr Expr 
Sub Expr Expr
eval (Lit r) = r
eval (Add e1 e2)
= eval e1 + eval e2
eval (Mul e1 e2)
= eval e1 * eval e2
eval (Sub e1 e2)
= eval e1  eval e2

data Expr = Lit Float 
Bin BinOp Expr Expr
data BinOp = Add  Mul  Sub
eval (Lit r) = r
eval (Binop op e1 e2)
= evalOp op (eval e1) (eval e2)
evalOp Add = (+)
evalOp Mul = (*)
evalOp Sub = ()

General comment:
It is possible to see this refactoring first being used left to right, to allow the addition of constructors in a localised way, and then from right to left, to permit the introduction of binary functions over the data type.
Note that the application of the refactoring in either direction requires the underlying data
type to be modified, and so all clients of the type, and indeed all constructors of values of the type, need to be modified.
Left to right comment: The code on the LHS repeats the functionality of "apply the appropriate binary operator to the evaluated arguments' for the three different kinds of expression. This commonaility is reflected in the RHS code, once the differences are factored out in the interpretation of the new data type, 
Right to left comment: The RHS complicates the definition by introducing the layer of binary operators. It also forces all 
Left to right conditions: The functions over the data type must exhibit a degree of common behaviour. 
Right to left conditions: There is no restriction on the right to left transition, which can be seen as the unfolding of the partical application of the 