This chapter introduces Frown by means of example.
Some elementary knowledge of monads is helpful in order to use Frown effectively. For the most basic applications, however, one can possibly do without. This section summarizes the relevant facts.
In Haskell, the concept of a monad is captured by the following class definition.
|
The essential idea of monads is to distinguish between computations and values. This distinction is reflected on the type level: an element of m a represents a computation that yields a value of type a. The trivial or pure computation that immediately returns the value a is denoted return a. The operator (>>=), commonly called ‘bind’, combines two computations: m >>= k applies k to the result of the computation m. The derived operation (>>) provides a handy shortcut if one is not interested in the result of the first computation. Finally, the operation fail is useful for signaling error conditions (a common thing in parsing).
Framing the concept of a monad as a type class is sensible for at least two interrelated reasons. First, we can use the same names (return, ‘>>=’, and fail) for wildly different computational structures.1 Second, by overloading a function with the monad class we effectively parameterize the function by computational structures, that is, we can call the same function with different instances of monads obtaining different computational effects.
The following instance declaration (Result.lhs) defines a simple monad that we will use intensively in the sequel (the monad can be seen as a simplified term implementation of the basic monad operations).
|
In monad speak, this is an exception monad: a computation in Result either succeeds gracefully yielding a value a (represented by the term Return a) or it fails with an error message s (represented by Fail s). That’s all we initially need for Frown: parsing a given input either succeeds producing a semantic value (sometimes called an attribution) or it fails (hopefully, with a clear indication of the syntax error).
Let’s start with a simple example. The following complete Frown source file (Paren1.lg2) defines the language of well-balanced parentheses. The specification of the grammar is enclosed in special curly braces ‘%{ ldots }%’. The remainder contains Haskell source code, that is, a module header and a function declaration.
|
The part enclosed in special curly braces comprises the typical ingredients of a context-free grammar: a declaration of the terminal symbols, a declaration of the nonterminal symbols, and finally the productions or grammar rules.
In general, the terminal symbols are given by Haskell patterns of the same type. Here, we have two character patterns of type Char.
Nonterminals are just identifiers starting with a lower-case letter. By convention, the first nonterminal is also the start symbol of the grammar (this default can be overwritten, see Sec. 3.2.7).
Productions have the general form n : v_1, ldots, v_k; where n is a nonterminal and v_1, …, v_k are symbols. Note that the symbols are separated by commas and terminated by a semicolon. The mandatory trailing semicolon helps to identify so-called є-productions, productions with an empty right-hand side, such as paren : ;.
As a shorthand, we allow to list several alternative right-hand sides separated by a vertical bar. Thus, the above productions could have been written more succinctly as
|
The two styles can be arbitrarily mixed. In fact, it is not even required that productions with the same left-hand side are grouped together (though it is good style to do so).
Now, assuming that the above grammar resides in a file called Paren.g we can generate a Haskell parser by issuing the command
frown Paren.g
This produces a Haskell source file named Paren.hs that contains among other things the function
|
which recognizes the language generated by the start symbol of the same name. Specifically, if inp is a list of characters, then paren inp is a computation that either succeeds indicating that inp is a well-formed parentheses or fails indicating that inp isn’t well-formed. Here is a short interactive session using the Haskell interpreter Hugs (type hugs Paren.hs at the command line).
|
Note that we have to specify the result type of the expressions in order to avoid an unresolved overloading error. Or to put it differently, we have to specify the monad, in which the parsing process takes place. Of course, we are free to assign paren a more constrained type by placing an appropriate type signature in the Haskell section of the grammar file:
|
By the way, since the nonterminal paren carries no semantic value, the type of the computation is simply Result () where the empty tuple type ‘()’ serves as a dummy type. In the next section we will show how to add attributes or semantic values to nonterminals.
Every once in a while parsing fails. In this case, Frown calls a user-supplied function named, well, frown (note that you must supply this function). In our example, frown has type
|
The error function frown is passed the remaining input as an argument, that you can give an indication of the location of the syntax error (more on error reporting in Sec. 3.3). Note that frown must be polymorphic in the result type.
> module Paren where > import Result > > {- frown :-( -} > > data Stack = Empty > | T_1_2 Stack > | T_2_3 Stack > | T_2_5 Stack > | T_4_5 Stack > | T_4_6 Stack > | T_5_4 Stack > > data Nonterminal = Paren > > paren tr = parse_1 tr Empty >>= (\ Paren -> return ()) > > parse_1 ts st = parse_2 ts (T_1_2 st) > > parse_2 tr@[] st = parse_3 tr (T_2_3 st) > parse_2 (’(’ : tr) st = parse_5 tr (T_2_5 st) > parse_2 ts st = frown ts > > parse_3 ts (T_2_3 (T_1_2 st)) = return Paren > > parse_4 (’(’ : tr) st = parse_5 tr (T_4_5 st) > parse_4 (’)’ : tr) st = parse_6 tr (T_4_6 st) > parse_4 ts st = frown ts > > parse_5 ts st = parse_4 ts (T_5_4 st) > > parse_6 ts (T_4_6 (T_5_4 (T_2_5 (T_1_2 st)))) = > = parse_2 ts (T_1_2 st) > parse_6 ts (T_4_6 (T_5_4 (T_4_5 (T_5_4 st)))) > = parse_4 ts (T_5_4 st) > > {- )-: frown -} > > frown _ = fail "syntax error"
![]()
Figure 3.2: The LR(0) automaton underlying the parser of Fig. 3.1.
Now, let’s augment the grammar of Sec. 3.2.1 by semantic values (Paren2.lg). Often, the parser converts a given input into some kind of tree representation (the so-called abstract syntax tree). To represent nested parentheses we simply use binary trees (an alternative employing n-ary trees is given in Sec. 4.1).
|
Attributes are always given in curly braces. When we declare a nonterminal, we have to specify the types of its attributes as in paren {Tree}. The rules of the grammar can be seen as functions from the right-hand side to the left-hand side. On the right-hand side, Haskell variables are used to name the values of attributes. The values of the attributes on the left-hand side are then given by Haskell expressions, in which the variables of the right-hand side may occur free. The Haskell expressions can be arbitrary, except that they must not be layout-sensitive.
In general, a nonterminal may have an arbitrary number of attributes (see Sec. 4.4 for an example). Note that Frown only supports so-called synthesized attributes (inherited attributes can be simulated, however, with the help of a reader monad, see Sec. 3.2.8, or with functional attributes, see Sec. 4.2).
The parser generated by Frown now has type
|
The following interactive session illustrates its use.
|
The parsers of the two previous sections take a list of characters as input. In practice, a parser usually does not work on character streams directly. Rather, it is prefaced by a lexer that first converts the characters into a list of so-called tokens. The separation of the lexical analysis from the syntax analysis usually leads to a clearer design and as a benevolent side-effect it also improves efficiency (Sec. 3.4.2 shows how to combine lexing and parsing in Frown, though).
A simple token type is shown in Fig 3.3 (Terminal1.lhs). (Note that the type comprises more constructors than initially needed.)
> module Terminal where > import Maybe > > data Op = Plus | Minus | Times | Divide > deriving (Show) > > name :: Op -> String > name Plus = "+" > name Minus = "-" > name Times = "*" > name Divide = "/" > > app :: Op -> (Int -> Int -> Int) > app Plus = (+) > app Minus = (-) > app Times = (*) > app Divide = div > > data Terminal = Numeral Int > | Ident String > | Addop Op > | Mulop Op > | KWLet > | KWIn > | Equal > | LParen > | RParen > | EOF > deriving (Show) > > ident, numeral :: String -> Terminal > ident s = fromMaybe (Ident s) (lookup s keywords) > numeral s = Numeral (read s) > > keywords :: [(String, Terminal)] > keywords = [ ("let", KWLet), ("in", KWIn) ]
Fig. 3.4 (Lexer.lhs) displays a simple lexer for arithmetic expressions, which are built from numerals using the arithmetic operators ‘+’, ‘-’, ‘*’, and ‘/’.
> module Lexer (module Terminal, module Lexer) where > import Char > import Terminal > > lexer :: String -> [Terminal] > lexer [] = [] > lexer (’+’ : cs) = Addop Plus : lexer cs > lexer (’-’ : cs) = Addop Minus : lexer cs > lexer (’*’ : cs) = Mulop Times : lexer cs > lexer (’/’ : cs) = Mulop Divide : lexer cs > lexer (’=’ : cs) = Equal : lexer cs > lexer (’(’ : cs) = LParen : lexer cs > lexer (’)’ : cs) = RParen : lexer cs > lexer (c : cs) > | isAlpha c = let (s, cs’) = span isAlphaNum cs in ident (c : s) : lexer cs’ > | isDigit c = let (s, cs’) = span isDigit cs in numeral (c : s) : lexer cs’ > | otherwise = lexer cs
The following grammar, which builds upon the lexer, implements a simple evaluator for arithmetic expressions (Calc.lg).
|
The terminal declaration now lists patterns of type Terminal. Note that terminals may also carry semantic values. The single argument of Numeral, for instance, records the numerical value of the numeral.
When declaring a terminal we can optionally define a shortcut using an as-clause as, for example, in LParen as "(". The shortcut can be used in the productions possibly improving their readability.
Here is an example session demonstrating the evaluator.
|
The expression that determines the value of an attribute is usually a pure one. It is, however, also possible to provide a monadic action that computes the value of the attribute. The computation lives in the underlying parsing monad. Monadic actions are enclosed in ‘{% ldots }’ braces and have type m t where m is the type of the underlying monad and t is the type of attributes.
As an example, the following variant of the desktop calculator (MCalc.lg) prints all intermediate results (note that we only list the changes to the preceeding example).
|
|
The following session illustrates its working.
|
In general, monadic actions are useful for performing ‘side-effects’ (for example, in order to parse %include directives) and for interaction with a monadic lexer (see Sec. 3.3.1).
In the previous examples we have encoded the precedences of the operators (‘*’ binds more tightly than ‘+’) into the productions of the grammar. However, this technique soon becomes unwieldy for a larger expression language. So let’s start afresh. The grammar file shown in Fig. 3.5 (Let1.lg) uses only a single nonterminal for expressions (we have also extended expressions by local definitions).
> module Let where > import Lexer > import Monad > > data Expr = Const Int | Var String | Bin Expr Op Expr | Let Decl Expr > deriving (Show) > > data Decl = String :=: Expr > deriving (Show) > > %{ > > Terminal = Numeral {Int} > | Ident {String} > | Addop {Op} > | Mulop {Op} > | KWLet as "let" > | KWIn as "in" > | Equal as "=" > | LParen as "(" > | RParen as ")"; > > expr {Expr}; > expr {Const n} : Numeral {n}; > {Var s} | Ident {s}; > {Bin e1 op e2} | expr {e1}, Addop {op}, expr {e2}; > {Bin e1 op e2} | expr {e1}, Mulop {op}, expr {e2}; > {Let d e} | "let", decl {d}, "in", expr {e}; > {e} | "(", expr {e}, ")"; > > decl {Decl}; > decl {s :=: e} : Ident {s}, "=", expr {e}; > > }% > > frown _ = fail "syntax error"
Also note that the grammar has no Nonterminal declaration. Rather, the terminal symbols are declared by supplying type signatures before the respective rules. Generally, type signatures are preferable to a Nonterminal declaration if the grammar is long.
Of course, the rewritten grammar is no longer LALR(k) simply because it is ambiguous. For instance, ‘1+2*3’ can be parsed as Bin (Const 1) Plus (Bin (Const 2) Times (Const 3)) or as Bin (Bin (Const 1) Plus (Const 2)) Times (Const 3). Frown is also unhappy with the grammar: it reports six shift/reduce conflicts:
* warning: 6 shift/reduce conflicts |
This means that Frown wasn’t able to produce a deterministic parser. Or rather, it produced a deterministic parser by making some arbitrary choices to avoid non-determinism (shifts are preferred to reductions, see Sec. 3.2.6). However, we can also instruct Frown to produce a non-deterministic parser, that is, one that generates all possible parses of a given input. We do so by supplying the option --backtrack:
frown --backtrack Let.g
The generated parser expr now has type
|
Note that the underlying monad must be an instance of MonadPlus (defined in the standard library Monad). The list monad and the Maybe monad are both instances of MonadPlus. The following session shows them in action.
|
The list monad supports ‘deep backtracking’: all possible parses are returned (beware: the number grows exponentionally). The Maybe monad implements ‘shallow backtracking’: it commits to the first solution (yielding the same results as the parser generated without the option --backtrack).
Instead of resorting to a backtracking parser we may also help Frown to generate the ‘right’ deterministic parser by assigning precedences to terminal symbols. The understand the working of precedences it is necessary to provide some background of the underlying parsing technique.
LR parsers work by repeatedly performing two operations: shifts and reductions. A shift moves a terminal from the input onto the stack, the auxiliary data structure maintained by the parser. A reduction replaces a top segment of the stack matching the right-hand side of a production by its left-hand side. Parsing succeeds if the input is empty and the stack consists of a start symbol only. As an example, consider parsing ‘N*N+N’.
|
At this point, there are two possibilities: we can either perform a reduction (using the production e : e, *, e;) or shift the next input symbol. Both choices are viable.
|
|
Alas, the two choices also result in different parse trees. By default, Frown prefers shifts to reductions. As a consequence, N*N+N is parsed as N*(N+N), that is, ‘+’ binds more tightly than ‘*’.
Now, we can direct the resolution of conflicts by assigning precedences and associativity to terminal symbols. The following declarations will do in our example (Let2.g).
|
Thus, ‘*’ takes precedence over ‘+’, which in turn binds more tightly than ‘in’. For instance, let a = 4 in a + 2 is parsed as let a = 4 in (a + 2). A conflict between two symbols of equal precedence is resolved using associativity: the succession 1+2+3 of left-associative operators is grouped as (1+2)+3; likewise for right-associative operators; sequences of non-associative operators are not well-formed.
Given the fixity declarations above Frown now produces the ‘right’ deterministic parser, which can be seen in action below.
|
In general, a conflict between the actions ‘reduce by rule r’ and ‘shift terminal t’ is resolved as follows (the precedence of a rule is given by the precedence of the rightmost terminal on the right-hand side):
condition | action | example | ||
prec r < prec t | shift | reduce by e:e,+,e; versus shift * | ||
left t | reduce | reduce by e:e,*,e; versus shift * | ||
prec r = prec t | right t | shift | reduce by e:e,++,e; versus shift ++ | |
nonassoc t | fail | reduce by e:e,==,e; versus shift == | ||
prec r > prec t | reduce | reduce by e:e,*,e; versus shift + | ||
Just in case you may wonder: there are no shift/shift conflicts by construction; reduce/reduce conflicts cannot be cured using precedences and associativity.
A grammar may have several start symbols. In this case, Frown generates multiple parsers, one for each start symbol (actually, these are merely different entry points into the LR(0) automaton4). We mark a symbol as a start symbol simply by putting an asterix before its declaration (either in a Nonterminal declaration or in a separate type signature). Consider our previous example: most likely we want parsers both for expressions and declarations. Thus, we write
|
and get parsers of type.
|
This section does not introduce any new features of Frown and can be safely skipped on first reading. Its purpose is to show how to simulate inherited attributes using a reader monad (see also Sec. 4.2). Generally, inherited attributes are used to pass context information down the parse tree. As an example, consider implementing an evaluator for arithmetic expressions that include variables and let-bindings (Let3.lg). To determine the value of a variable we need to pass down an environment that records the values of bound variables. The reader monad displayed in Fig. 3.6 (Reader.lhs) serves exactly this purpose.
> module Reader where > > newtype Reader env a = Reader { apply :: env -> a } > > instance Monad (Reader env) where > return a = Reader (\ env -> a) > m >>= k = Reader (\ env -> apply (k (apply m env)) env) > fail s = Reader (error s) > > getenv :: Reader env env > getenv = Reader (\ env -> env) > > withenv :: env -> Reader env a -> Reader env’ a > withenv env m = Reader (\ env’ -> apply m env)
We need some additional helper functions for accessing and extending environments
|
The following grammar implements the desired evaluator.
|
Note that there are two monads around: the parsing monad (in fact, expr is parametric in this monad) and the reader monad, which is embedded in the attributes. The parser returns a value of type Reader Int, to which we pass an empty initial environment.
|
Let’s see the evaluator in action.
|
The chances that parsing succeeds are probably smaller than the chances that it fails. Good error messages are indispensable to turn the latter into the former case. Up to now we only produced the rather uninformative message "syntax error". Fortunately, we are in a good position to do better. LR parsing has the nice property that it detects a syntax error at the earliest possible moment: parsing fails as soon as the input cannot be extended to a legal sentence of the grammar. For instance, the syntax error in let a = 4 * (7 + 1 − 1 in a * a is detected after reading the keyword ‘in’.
Now, all we have to do is to keep track of context information: the current line and column number and possibly the filename. This section prepares the ground for maintaining state information; the parser that actually keeps track of line numbers etc is only shown in the next section.
Unsurprisingly, to maintain state information we employ monads again. This time, we require a state monad. The natural place for maintaining information about line numbers etc is, of course, the lexer. Consequently, we turn the stream-based lexer of type String -> [Terminal] into a monadic one of type
|
where M is the state monad. The idea is that each time get is called it returns the next token and updates its internal state.
The first version of the monadic lexer shown in Fig. 3.7 (MLexer1.lhs) has no internal state apart from the input stream, that is, it provides no additional functionality compared to the stream-based lexer.
> module MLexer ( module Terminal, module MLexer ) where > import Terminal > import Char > > type CPS a answer = (a -> answer) -> answer > > newtype Lex m a = Lex { unLex :: forall ans . CPS a (String -> m ans) } > > instance (Monad m) => Monad (Lex m) where > return a = Lex (\ cont -> cont a) > m >>= k = Lex (\ cont -> unLex m (\ a -> unLex (k a) cont)) > fail s = lift (fail s) > > lift :: (Monad m) => m a -> Lex m a > lift m = Lex (\ cont inp -> m >>= \ a -> cont a inp) > > run :: (Monad m) => Lex m a -> (String -> m a) > run parser inp = unLex parser (\ a rest -> return a) inp > > get :: (Monad m) => Lex m Terminal > get = > Lex (\ cont inp -> > let lexer [] = cont (EOF) [] > lexer (’+’ : cs) = cont (Addop Plus) cs > lexer (’-’ : cs) = cont (Addop Minus) cs > lexer (’*’ : cs) = cont (Mulop Times) cs > lexer (’/’ : cs) = cont (Mulop Divide) cs > lexer (’=’ : cs) = cont (Equal) cs > lexer (’(’ : cs) = cont (LParen) cs > lexer (’)’ : cs) = cont (RParen) cs > lexer (c : cs) > | isSpace c = lexer cs > | isAlpha c = let (s, cs’) = span isAlphaNum cs in cont (ident (c : s)) cs’ > | isDigit c = let (s, cs’) = span isDigit cs in cont (numeral (c : s)) cs’ > | otherwise = lexer cs > in lexer inp) > > frown :: (Monad m) => Terminal -> Lex m a > frown t = Lex (\ cont inp -> > fail ("\n*** syntax error:\n" ++ context 4 inp)) > > context :: Int -> String -> String > context n inp = unlines (take n (lines inp ++ ["<end of input>"]))
Note that we use a continuation-based state monad, Lex m, which requires local universal quantification (a non-Haskell 98 feature). Actually, Lex is even a monad transformer so that we can freely choose a base monad (such as Result or IO). Of course, an ‘ordinary’ state monad would do, as well. The monadic lexer get incorporates more or less the stream-based lexer. We only changed the recursive calls to lexer (ie t : lexer cs) into invocations of the continuation (ie cont t cs). The error routine frown now has type
|
that is, frown is no longer passed the remaining input but only the look-ahead token.
The changes to the grammar are minor: we have to declare an ‘end of file’ token marked by a star (Let4.lg)
|
and we have to provide a type signature for the generated parser (in the Haskell section).
|
The type signature is necessary to avoid an ‘unresolved top-level overloading’ error (the monomorphism restriction strikes again).
When we generate the Haskell parser we must supply the option --lexer to inform Frown that we use a monadic lexer.
frown --lexer Let.g
For completeness, here is an interactive session (note that in the case of error the look-ahead token is not displayed).
|
The monadic lexer shown in Fig. 3.8 (MLexer2.lhs) builds upon the one given in the previous section. The state monad Lex m has been extended to keep track of the current line number and the current line itself. The current line is displayed in case of a lexical or syntax error. As an aside, note that the column number can be recreated from the rest of the input and the length of the current line.
> module MLexer ( module Terminal, module MLexer ) where > import Terminal > import Char > > type CPS a answer = (a -> answer) -> answer > > newtype Lex m a = Lex { unLex :: forall ans . CPS a (String -> Int -> String -> m ans) } > > instance (Monad m) => Monad (Lex m) where > return a = Lex (\ cont -> cont a) > m >>= k = Lex (\ cont -> unLex m (\ a -> unLex (k a) cont)) > fail s = lift (fail s) > > lift :: (Monad m) => m a -> Lex m a > lift m = Lex (\ cont inp line cur -> m >>= \ a -> cont a inp line cur) > > run :: (Monad m) => Lex m a -> (String -> m a) > run parser inp = unLex parser (\ a rest line cur -> return a) inp 1 (current inp) > > current :: String -> String > current s = takeWhile (/= ’\n’) s > > get :: (Monad m) => Lex m Terminal > get = > Lex (\ cont inp line cur -> > let lexer [] n x = cont (EOF) [] n x > lexer (’\n’ : cs) n x = lexer cs (n + 1) (current cs) > lexer (’+’ : cs) n x = cont (Addop Plus) cs n x > lexer (’-’ : cs) n x = cont (Addop Minus) cs n x > lexer (’*’ : cs) n x = cont (Mulop Times) cs n x > lexer (’/’ : cs) n x = cont (Mulop Divide) cs n x > lexer (’=’ : cs) n x = cont (Equal) cs n x > lexer (’(’ : cs) n x = cont (LParen) cs n x > lexer (’)’ : cs) n x = cont (RParen) cs n x > lexer (c : cs) n x > | isSpace c = lexer cs n x > | isAlpha c = let (s, cs’) = span isAlphaNum cs in cont (ident (c : s)) cs’ n x > | isDigit c = let (s, cs’) = span isDigit cs in cont (numeral (c : s)) cs’ n x > | otherwise = fail ("\n*** lexical error at " > ++ position cs n x ++ ":\n" > ++ context 4 cs x) > in lexer inp line cur) > > frown :: (Monad m) => Terminal -> Lex m a > frown t = Lex (\ cont inp line cur -> > fail ("\n*** syntax error at " > ++ position inp line cur ++ ":\n" > ++ context 4 inp cur)) > > position :: String -> Int -> String -> String > position inp line cur = "(line " ++ show line ++ ", column " ++ show col ++ ")" > where col = length cur - length (current inp) > > context :: Int -> String -> String -> String > context n inp cur = unlines ([cur, replicate col’ ’ ’ ++ "^"] > ++ take n (lines (drop 1 (dropWhile (/= ’\n’) inp)) > ++ ["<end of input>"])) > where col’ = length cur - length (current inp) - 1
The following session shows the new lexer in action.
|
In the case of a lexical error the cursor ‘^’ points at the offending character. In the case of a syntax error the cursor points at the last character of the offending token (recall that the part of the input up to and including this token is the shortest prefix of the input that cannot be extended to a legal sentence of the grammar).
We can do even better! We can instruct Frown to pass a list of expected tokens to the error routine frown (by supplying the option --expected).
frown --lexer --expected Let.g
Frown uses the shortcuts given in the terminal declaration for generating lists of expected tokens. This means, in particular, that a token is not included in such a list if it does not have a shortcut. In our running example, we want every token to be listed. Therefore, we add shortcuts for every terminal symbol (Let6.lg).
|
The error routine frown now takes an additional argument of type [String] (MLexer3.lhs).
|
The interactive session listed in Fig. 3.9 is a bit longer than usual to illustrate the quality of the error messages.
> Let>> run expr "let\n a = 4 * [7 + 1 - 1)\n in a * a" :: IO Expr > > Program error: user error ( > *** lexical error at (line 2, column 13): > a = 4 * [7 + 1 - 1) > ^ > in a * a > <end of input> > ) > Let>> run expr "let\n a = 4 * (7 + 1 - 1\n in a * a" :: IO Expr > > Program error: user error ( > *** syntax error at (line 3, column 3): > in a * a > ^ > <end of input> > * expected: + or -, * or /, )) > Let>> run expr "let\n a = 4 * (7 + 1 - 1)\n a * a" :: IO Expr > > Program error: user error ( > *** syntax error at (line 3, column 2): > a * a > ^ > <end of input> > * expected: + or -, * or /, in) > Let>> run expr "\n a = 4 * (7 + 1 - 1)\n in a * a" :: IO Expr > > Program error: user error ( > *** syntax error at (line 2, column 7): > a = 4 * (7 + 1 - 1) > ^ > in a * a > <end of input> > * expected: + or -, * or /, <end of input>) > Let>> run expr "let\n a = 4 * (7 + - 1)\n in a * a" :: IO Expr > > Program error: user error ( > *** syntax error at (line 2, column 18): > a = 4 * (7 + - 1) > ^ > in a * a > <end of input> > * expected: <numeral>, <identifier>, let, () > Let>> run expr "let\n a = 4 (7 + 1 - 1)\n in a * a" :: IO Expr > > Program error: user error ( > *** syntax error at (line 2, column 12): > a = 4 (7 + 1 - 1) > ^ > in a * a > <end of input> > * expected: + or -, * or /, in)
So far we have content ourselves with reporting syntax errors. To a limited extent it is also possible to correct errors. Consider the last rule of the following grammar (Let7.lg).
|
The symbol insert ")" instructs Frown to automatically insert a ")" token if parsing would otherwise fail. The special symbol insert ")" can be seen as being defined by the є-production insert ")" : ;. The difference to an ‘ordinary’ user-defined є-production is that the rule is only applied if every other action would fail.
The following session shows the error correction in action.
|
In the last query the missing parenthesis ‘)’ is inserted just before the keyword ‘in’. This may or may not what the user intended!
It is generally a good idea to notify the user if a token is inserted. This is relatively easy to accomplish using monadic actions (Let8.lg). The parsing monad is now Lex IO; the monad transformer Lex proves its worth.
|
|
|
Let’s repeat the last query of the previous session.
|
The reader is invited to extend the code so that the current source location is additionally printed (informing the user where the token has been inserted).
When we define grammars we often find ourselves repeatedly writing similar rules. A common pattern is the repetition of symbols. As an example, the following rules define a repetition of t symbols.
|
As an aside, note that the second rule is intentionally left-recursive. LR parsers prefer left to right recursion: the rules above use constant stack space whereas the right-recursive variant requires space linear in the length of the input.
Now, Frown allows to capture recurring patterns using so-called rule schemes. Here is the scheme for a repetition of symbols (of arity 0).
|
The first line contains many’s type signature: it simply says that neither many nor many’s argument x possess attributes. Given this scheme we can simply write many t for a repetition of t symbols.
The rule for repetition becomes more interesting if the argument possesses an attribute (is of arity 1). In this case, many returns a list of semantic values.
|
(The use of list concatenation ‘++’ in the second rule incurs a runtime penalty which we will cure later.) The first line contains again the type signature, which we may read as a conditional clause: if x has one attribute of type a, then many x has one attribute of type [a]. This schemes comes in handy if we extend our little expression language by applications and abstractions (we assume that the abstract syntax has been extended suitably; aexpr denotes atomic expressions).
|
Note that if we pass terminal symbols as arguments to rule schemes they must be written with (empty) curly braces—Frown can only identify terminal symbols, ie patterns, if they have exactly the same syntactic form as in the terminal declaration. Think of ‘{}’ as a placeholder.
In the above definition of many we have used list concatenation to append an element to a list. The following improved definition does away with this linear-time operation employing Hughes’ efficient sequence type [3].
|
These schemata are predefined in Frown. There is a caveat, however: the singleton production many x : many’ x may introduce a shift/reduce conflict, see Sec. 4.3.
Actually, both the many scheme with no attributes and the scheme above with one attribute are predefined. In general, it is possible to use the same name for schemes and nonterminals of different arity. The only restriction is that the arity of the scheme must determine the arity of its arguments.
Another useful variation of many is sepBy x sep which denotes a list of x symbols separated by sep symbols (sepBy and sepBy1 are predefined, as well).
|
This scheme is useful for adding tuples to our expression language.
|
For a complete list of predefined schemes see Sec. 5.3.
The terminal symbols of a grammar are given by Haskell patterns. Up to now we have seen only simple patterns. Patterns, however, may also be nested or even overlapping. In the latter case, one should be careful to list specific patterns before general ones in a Terminal declaration (Frown preserves the relative ordering of patterns when generating case expressions). Here is a simple example.
|
Note that keywords are declared just by listing them before the general pattern for identifiers.
Alternatively, terminal symbols can be specifed using so-called guards, Boolean functions of type Terminal -> Bool. Guards are most useful for defining character classes as in the following example.
|
A guard is introduced by the keyword guard, followed by its Haskell definition, followed by the mandatory shortcut. The shortcut can then be used as a terminal symbol of arity 1: its attribute of type Terminal is the very input symbol that matched the guard.
|
Using guards one can quite easily define character-based grammars that include lexical syntax (that is, whose parsers combine lexing and parsing). Fig. 3.10 lists a variant of the desktop calculator that works without a separate lexer.
> module Calc where > import Result > import Char > > type Terminal = Char > > %{ > > Terminal = guard {isDigit} as "digit" > | ’+’ > | ’*’ > | ’(’ > | ’)’; > > Nonterminal = expr {Integer} > | term {Integer} > | factor {Integer} > | numeral {Integer}; > > expr {v1 + v2} : expr {v1}, ’+’, term {v2}; > {e} | term {e}; > term {v1 * v2} : term {v1}, ’*’, factor {v2}; > {e} | factor {e}; > factor {n} : numeral {n}; > {e} | ’(’, expr {e}, ’)’; > numeral {encode c} : "digit" {c}; > {n * 10 + encode c} | numeral {n}, "digit" {c}; > > }% > > encode c = toInteger (fromEnum c - fromEnum ’0’) > > frown _ = fail "syntax error"
Note that the type Terminal must be defined in the Haskell section. The reader may wish to extend the grammar so that two tokens can be separated by white space.
⟨To do: type grammar.⟩
⟨To do: --prefix und --suffix.⟩
⟨To do: --debug und --pagewidth.⟩
|
⟨To do: optimizations (--optimize).⟩
⟨To do: which format benefits from GHC extensions (--ghc)?⟩
⟨To do: NOINLINE pragmas (--noinline).⟩
⟨To do: --signature.⟩