(more discussion about relative relative precedences)
Revision as of 13:49, 20 November 2006
This page is dedicated to arguments against syntactic sugar. The request for extended syntactic sugar is present everywhere and the reasons for syntactic sugar are obvious, but there are also serious objections to them. The objections listed here may help to decide when to do without syntactic sugar and which special notations should better be dropped in future versions of Haskell.
Haskell's basic syntax consists of function definition and function application. Though in some cases function application is hard to read and digs into details that are not essential for the situation they describe. For this purpose special syntaxeslike
were introduced for some frequent programming tasks to allow a more pleasant look.
Many people seem to like Haskell only because of its syntactic sugar. But adding syntactic sugar to a language is not a big achievement. Python, Perl, C++ have lots of syntactic sugar, but I wouldn't prefer them to Haskell. Why? Because they lack the transparency of data dependency of functional programming languages, they lack static but easy to use polymorphism, they lack lazy evaluation, they lack reliable modularisation. It's not amazing that Haskell provides a lot of syntactic sugar. It's amazing that every syntactic sugar has pure functional explanations. That proves the power of the functional concept.
1.1 Syntactic heroin
Compiler writers can only lose if they give way to the insistence of users requesting more syntactic sugar. Every user has his own preferred applications, everyone has his taste and everyone wants his special application and his taste to be respected in future language revisions. Who is authorised to decide which application is general and which is too special? Is it more important to have many syntactic alternatives such that all people can write with their individual styles or is it more important that code of several authors have homogenous appearance such that it can be read by all people?
You can bet if new syntactic sugar arises many users will rush at it and forget about the analytic expression the special notation shall replace. To argue against that is like trying to take the most beloved toy from children.
Every special notation leads to the question if it can be extended and generalised. Guards are extended to pattern guards and list comprehension is generalised to parallel list comprehension in current versions of Haskell compilers. Infix notation for alphanumeric functions is already possible in Haskell98but "lacks" the possibility to add arguments like in
The last is not implemented, but was already requested. A solution using only Haskell98 infix operators is already invented (http://www.haskell.org/pipermail/haskell-cafe/2002-July/003215.html). Further on, the more general "MixFix" notation was already proposed (http://www.dcs.gla.ac.uk/mail-www/haskell/msg02005.html), not to forget the silent lifting of map data structures to functions (http://www.haskell.org/pipermail/haskell/2002-October/010629.html). What comes next?
Rodney Bates called the phenomena not only "syntactic sugar" but "syntactic heroin".
People start with a small dosis of syntactic sugar, they quickly want more, because the initial dosis isn't enough for ecstasy any longer. If one drug no longer helps then stronger ones are requested. It is so much tempting because the users requesting syntactic sugar are not responsible for implementing it and for avoiding inferences with other language features.
1.2 Parse errors
Compiler users have contradictory wishes. On the one hand they want more syntactic sugar, on the other hand they want better parser error messages. They don't realize that one is quite the opposite of the other.
E.g. when a parser reads an opening bracket it doesn't know whether it is the start of a list comprehension expressionlike
or the start of a list of comma separated expressionslike
Thus if you accidentally mix bars and commas the parser don't know if you wanted to write a list comprehension or a comma separated list. So it can't tell you precisely what you made wrong.
Type error messages of GHC have already reached a complexity which can't be processed by many Haskell newbies. It is the price to be paid for a type system which tries to cope with as few as possible type hints.
Let's consider another example from the view of a compiler. Internally it transforms the source code
flip (+) 1
then it compiles it like regular functional code.
Though what happens if it encounters an error?
If it reports the error like
type error in
flip (+) 1
(as Hugs November 2002)
you wouldn't understand it,
A compiler which handles this properly must support syntactic sugar at the same level like regular syntax which is obviously more complicated.
1.3 Sugar adds complexity
Syntactic sugar are usually special grammatical constructions. They can interfere badly with other constructions:
But syntactic sugar does not only touch the compilers. Many other tools like those for syntax highlighting (emacs, nedit), source code markup (lhs2TeX), source code formatting (Language.Haskell.Pretty), source code transform (e.g. symbolic differentation), program proofs, debugging, dependency analysis, documentation extraction (haddock) are affected.
Each tool becomes more complicated by more syntactic sugar.
The use of functions and functions of functions (i.e. higher order functions) allows for very flexible usage of program units. This is also true for the function notation, but it is not true for some syntactic sugar.E.g.
which is not possible for list comprehension syntax.Thus
A function can get more arguments as the development goes on.If you are used to write
when the result of a function needs post-processing. Say we have the functions
isLeapYear :: Int -> Bool isLeapYear year = mod year 4 == 0 && (mod year 100 /= 0 || mod year 400 == 0) leapYearText :: Int -> String leapYearText year | isLeapYear year = "A leap year" | otherwise = "Not a leap year"
If you stick to guards you will possibly rewrite it to the clumsy
leapYearText :: Language -> Int -> String leapYearText lang year = translate lang (case () of () | isLeapYear year -> "A leap year" | otherwise -> "Not a leap year")
But what about
leapYearText :: Language -> Int -> String leapYearText lang year = translate lang (if (isLeapYear year) then "A leap year" else "Not a leap year")
leapYearText :: Int -> String leapYearText year = if (isLeapYear year) then "A leap year" else "Not a leap year"
The following section consider several notations and their specific problems.
2.1 Infix notation
Infix notation is problematic for both human readers and source code formatters. The reader doesn't know the precedences of custom infix operators, he has to read the modules which the operators are imported from. This is even more difficult because infix operators are usually imported unqualified, that is you don't know from which module an operator is imported. The same problem arises for source code formatters. You certainly prefer the formatting
a + b * c
a + b * c
A source code formatter can format this properly only if it has access to the imported modules. This is certainly uncommon for a plain source code formatter.
It has been noticed by many people, that the integer numbered precedences are not enough for describing the relations of all the infix operators. http://www.haskell.org/pipermail/haskell-cafe/2005-February/009260.html Fractional and negative fixities were already proposed: http://www.haskell.org/pipermail/haskell-cafe/2006-November/019293.html
You can't pass an argument to a function written in infix notation.
Some library functions are designed for a "reversed" order of arguments, this means that you will most oftenly leave out the first argument on partial application rather than the second one.E.g. the functions
This is because the library designer expect that the user will prefer the infix style,writing
For functions which are not bound to a traditional notation one should avoid this order!A bad example in this respect is the module
Many of the functions of this module alter some bits in a machine word,thus they can be considered as update functions and their type signature should end with
Then you could easily combine several operations by
shiftL 2 . clearBit 7 . setBit 4 . setBit 1
flip shiftL 2 . flip clearBit 7 . flip setBit 4 . flip setBit 1
(`shiftL` 2) . (`clearBit` 7) . (`setBit` 4) . (`setBit` 1)
2.2.1 Special notation for the list typeThe type of a list over type
For beginners it becomes even more complicated to distinguish between the type and the value of a list. Some people try to turn some expression into a list by enclosing it in brackets just like it is done for the list type.I don't see the advantage of
2.2.2 Comma separated list elementsWe are used to the list notation
I think many Haskell users are not aware that it is a special notation.They don't know that it is a replacement for
and because of that they also can't derivethat a function for constructing single element list can be written as
There are two reasons against:
- The theoretical reason: The intuitive list notation using comma separation requires one comma less than the number of elements, an empty list would need -1 commas, which can't be written, obviously.
- The practical reason: The colon is like a terminator. Each list element is followed by the colon, thus it is easier to reorder the elements of a list in an editor. If you have written you can simply cut some elements and the subsequent ':' and then you can insert them whereever you want.(1:2:3:)
Although the list type has so many special support by the Haskell 98 language, there is no need for some syntactic support. The definition
data List a = End | (:) a (List a)
is regular Haskell98 code.The colon should have precedence below
Again, this proves the power of the basic features of Haskell98.
2.2.3 Parallel list comprehensionParallel list comprehension can be replaced by using
3 (n+k) patterns
Therer are some notational ambiguities concerning (n+k) patterns.
See Why I hate n+k
4 If-Then-ElseThe construction
A generalisation of this syntactic exception was already proposed as "MixFix" notation. But it's worth to turn round the question:What is so special about
- Guards can be dropped completely. should be turned into a regular function.ifcould be turned into a function, i.e.case expr ofcould an expression of typecase 0 -> 'a'; 1 -> 'b';. It should be complemented byInt -> Charfunction like that in Case.select
- Infix notation is good for nested application, because reflects the represented structure better than(0:1:2:).((:) 0 ((:) 1 ((:) 2 )))
- Infix usage of functions with alphanumeric names is often just a matter of habit, just for the sake of fanciness, such as which doesn't add anything to readability. If this feature is kept it should remain restricted to function names. It should not be extended to partially applied functions.toLower `map` s
- List comprehension should be used rarely, parallel list comprehension should be dropped completely.
- notation is good for representing imperative and stateful program structures.do
- patterns simulate a number representation which is not used internally and thus it must be emulated with much effort. It should be dropped. Numeric patterns such as(n+k)involve conversions likeand real comparisons (fromIntegerclass!) for matching. It should be thought about dropping them, too.Eq