Starting from:
$30

$24

Homework 5 solution

Files you should submit: Calc.hs, containing a module of the same name.



As we saw in class, Haskell’s type classes provide ad-hoc polymor-phism, that is, the ability to decide what to do based on the type of an input. This homework explores one interesting use of type classes in constructing domain-specific languages.







Expressions




On day one of your new job as a software engineer, you’ve been asked to program the brains of the company’s new blockbuster prod-uct: a calculator. But this isn’t just any calculator! Extensive focus group analysis has revealed that what people really want out of their calculator is something that can add and multiply integers. Anything more just clutters the interface.




Your boss has already started by modeling the domain with the following data type of arithmetic expressions:




data ExprT = Lit Integer




Add ExprT ExprT



Mul ExprT ExprT



deriving (Show, Eq)




This type is capable of representing expressions involving integer constants, addition, and multiplication. For example, the expression (2 + 3) 4 would be represented by the value




Mul (Add (Lit 2) (Lit 3)) (Lit 4).




Your boss has already provided the definition of ExprT in ExprT.hs, so as usual you just need to add import ExprT to the top of your file. However, this is where your boss got stuck.




Exercise 1




Write Version 1 of the calculator: an evaluator for ExprT, with the signature




eval :: ExprT - Integer




For example, eval (Mul (Add (Lit 2) (Lit 3)) (Lit 4)) == 20.







Exercise 2




The UI department has internalized the focus group data and is ready to synergize with you. They have developed the front-facing user-interface: a parser that handles the textual representation of the selected language. They have sent you the module Parser.hs, which exports parseExp, a parser for arithmetic expressions. If you pass the constructors of ExprT to it as arguments, it will convert Strings representing arithmetic expressions into values of type ExprT. For example:




*Calc parseExp Lit Add Mul "(2+3)*4"




Just (Mul (Add (Lit 2) (Lit 3)) (Lit 4))

*Calc parseExp Lit Add Mul "2+3*4"




Just (Add (Lit 2) (Mul (Lit 3) (Lit 4)))

*Calc parseExp Lit Add Mul "2+3*"




Nothing




Leverage the assets of the UI team to implement the value-added function




evalStr :: String - Maybe Integer




which evaluates arithmetic expressions given as a String, produc-ing Nothing for inputs which are not well-formed expressions, and Just n for well-formed inputs that evaluate to n.




Exercise 3




Good news! Early customer feedback indicates that people really do love the interface! Unfortunately, there seems to be some disagree-ment over exactly how the calculator should go about its calculating business. The problem the software department (i.e. you) has is that while ExprT is nice, it is also rather inflexible, which makes catering to diverse demographics a bit clumsy. You decide to abstract away the properties of ExprT with a type class.




Create a type class called Expr with three methods called lit, add, and mul which parallel the constructors of ExprT. Make an instance of Expr for the ExprT type, in such a way that







mul (add (lit 2) (lit 3)) (lit 4) :: ExprT




== Mul (Add (Lit 2) (Lit 3)) (Lit 4)




Think carefully about what types lit, add, and mul should have. It may be helpful to consider the types of the ExprT constructors, which you can find out by typing (for example)

*Calc :t Lit




at the ghci prompt.




Remark. Take a look at the type of the foregoing example expression:




*Calc :t mul (add (lit 2) (lit 3)) (lit 4)




Expr a = a




What does this mean? The expression mul (add (lit 2) (lit 3)) (lit 4)




has any type which is an instance of the Expr type class. So writing it by itself is ambiguous: GHC doesn’t know what concrete type you want to use, so it doesn’t know which implementations of mul, add, and lit to pick.

One way to resolve the ambiguity is by giving an explicit type signature, as in the above example. Another way is by using such an expression as part of some larger expression so that the context in which it is used determines the type. For example, we may write a function reify as follows:




reify :: ExprT - ExprT




reify = id




To the untrained eye it may look like reify does no actual work!




But its real purpose is to constrain the type of its argument to ExprT.




Now we can write things like




reify $ mul (add (lit 2) (lit 3)) (lit 4)




at the ghci prompt.







Exercise 4




The marketing department has gotten wind of just how flexible the calculator project is and has promised custom calculators to some big clients. As you noticed after the initial roll-out, everyone loves the interface, but everyone seems to have their own opinion on what the semantics should be. Remember when we wrote ExprT and thought that addition and multiplication of integers was pretty cut and dried? Well, it turns out that some big clients want customized calculators with behaviors that they have decided are right for them.




The point of our Expr type class is that we can now write down arithmetic expressions once and have them interpreted in various ways just by using them at various types.







Make instances of Expr for each of the following types:




Integer — works like the original calculator



Bool— every literal value less than or equal to 0 is in-terpreted as False, and all positive Integers are interpreted as True; “addition” is logical or, “multiplication” is logical and








MinMax




— “addition” is taken to be the max function, while “multiplication” is the min function



Mod7— all values should be in the ranage 0 . . . 6, and all arithmetic is done modulo 7; for example, 5 + 3 = 1.



The last two variants work with Integers internally, but in order to provide different instances, we wrap those Integers in newtype wrappers. These are used just like the data constructors we’ve seen before.



newtype MinMax newtype Mod7




= MinMax Integer deriving (Eq, Show)




= Mod7 Integer deriving (Eq, Show)



Once done, the following code should demonstrate our family of calculators:




testExp :: Expr a = Maybe a

testExp = parseExp lit
add mul "(3 * -4) + 5"
testInteger
= testExp
:: Maybe Integer
testBool
= testExp
:: Maybe Bool
testMM
=
testExp
:: Maybe
MinMax
testSat
=
testExp
:: Maybe
Mod7



Try printing out each of those tests in ghci to see if things are working. It’s great how easy it is for us to swap in new semantics for the same syntactic expression!




You must complete at least one of the following two exercises:







Exercise 5 (do this OR exercise 6)




The folks down in hardware have finished our new custom CPU, so we’d like to target that from now on. The catch is that a stack-based architecture was chosen to save money. You need to write a version of your calculator that will emit assembly language for the new processor.
The hardware group has provided you with StackVM.hs, which is a software simulation of the custom CPU. The CPU supports six operations, as embodied in the StackExp data type:




data StackExp = PushI Integer




PushB Bool



Add



Mul



And



Or



deriving Show




type Program = [StackExp]




PushI and PushB push values onto the top of the stack, which can store both Integer and Bool values. Add, Mul, And, and Or each pop the top two items off the top of the stack, perform the appropriate operation, and push the result back onto the top of the stack. For example, executing the program




[PushB True, PushI 3, PushI 6, Mul]




will result in a stack holding True on the bottom, and 18 on top of that.

If there are not enough operands on top of the stack, or if an op-eration is performed on operands of the wrong type, the processor will melt into a puddle of silicon goo. For a more precise specifica-tion of the capabilities and behavior of the custom CPU, consult the reference implementation provided in StackVM.hs.




Your task is to implement a compiler for arithmetic expressions. Simply create an instance of the Expr type class for Program, so that arithmetic expressions can be interpreted as compiled programs. For any arithmetic expression exp :: Expr a = a it should be the case that







stackVM exp == Right [IVal exp]




Note that in order to make an instance for Program (which is a type synonym) you will need to enable the TypeSynonymInstances language extension, which you can do by adding




{-# LANGUAGE TypeSynonymInstances #-}




as the first line in your file.




Finally, put together the pieces you have to create a function

compile :: String - Maybe Program




which takes Strings representing arithmetic expressions and com-piles them into programs that can be run on the custom CPU.




Exercise 6 (do this OR exercise 5)




Some users of your calculator have requested the ability to give names to intermediate values and then reuse these stored values later.

To enable this, you first need to give arithmetic expressions the ability to contain variables. Create a new type class HasVars a which contains a single method var :: String - a. Thus, types which are instances of HasVars have some notion of named variables.

Start out by creating a new data type VarExprT which is the same as ExprT but with an extra constructor for variables. Make VarExprT an instance of both Expr and HasVars. You should now be able to write things like




*Calc add (lit 3) (var "x") :: VarExprT




But we can’t stop there: we want to be able to interpret expres-sions containing variables, given a suitable mapping from variables to values. For storing mappings from variables to values, you should use the Data.Map module. Add




import qualified Data.Map as M




at the top of your file. The qualified import means that you must prefix M. whenever you refer to things from Data.Map. This is stan-dard practice, since Data.Map exports quite a few functions with names that overlap with names from the Prelude. Consult the Data.Map documentation to read about the operations that are sup-ported on Maps.




Implement the following instances:




instance HasVars (M.Map String Integer - Maybe Integer)




instance Expr (M.Map String Integer - Maybe Integer)




































































































http://hackage.haskell.org/ packages/archive/containers/latest/ doc/html/Data-Map.html



The first instance says that variables can be interpreted as func-tions from a mapping of variables to Integer values to (possibly) Integer values. It should work by looking up the variable in the mapping.




The second instance says that these same functions can be inter-preted as expressions (by passing along the mapping to subexpres-sions and combining results appropriately).

Note: to write these instances you will need to enable the FlexibleInstances language extension by putting






{-# LANGUAGE FlexibleInstances #-}




as the first line in your file.




Once you have created these instances, you should be able to test them as follows:




withVars :: [(String, Integer)]




- (M.Map String Integer - Maybe Integer)




- Maybe Integer




withVars vs exp = exp $ M.fromList vs




*Calc :t add (lit 3) (var "x")




add (lit 3) (var "x") :: (Expr a, HasVars a) = a *Calc withVars [("x", 6)] $ add (lit 3) (var "x") Just 9




*Expr withVars [("x", 6)] $ add (lit 3) (var "y") Nothing




*Calc withVars [("x", 6), ("y", 3)]




$ mul (var "x") (add (var "y") (var "x"))




Just 54

More products