Title

SRFI 26: Notation for Specializing Parameters without Currying

Author

Sebastian Egner

Status

This SRFI is currently in ``final'' status. To see an explanation of each status that a SRFI can hold, see here. You can access the discussion via the archive of the mailing list.

Abstract

When programming in functional style, it is frequently necessary to specialize some of the parameters of a multi-parameter procedure. For example, from the binary operation cons one might want to obtain the unary operation (lambda (x) (cons 1 x)). This specialization of parameters is also known as "partial application", "operator section" or "projection".

The mechanism proposed here allows to write this sort of specialization in a simple and compact way. The mechanism is best explained by a few examples:

(cut cons (+ a 1) <>) is the same as (lambda (x2) (cons (+ a 1) x2))
(cut list 1 <> 3 <> 5) is the same as (lambda (x2 x4) (list 1 x2 3 x4 5))
(cut list) is the same as (lambda () (list))
(cut list 1 <> 3 <...>) is the same as (lambda (x2 . xs) (apply list 1 x2 3 xs))
(cut <> a b) is the same as (lambda (f) (f a b))

As you see, the macro cut specializes some of the parameters of its first argument. The parameters that are to show up as formal variables of the result are indicated by the symbol <>, pronouced as "slot". In addition, the symbol <...>, pronounced as "rest-slot", matches all residual arguments of a variable argument procedure. As you can see from the last example above, the first argument can also be a slot, as one should expect in Scheme.

In addition to cut, there is a variant called cute (a mnemonic for "cut with evaluated non-slots") which evaluates the non-slot expressions at the time the procedure is specialized, not at the time the specialized procedure is called. For example,

(cute cons (+ a 1) <>) is the same as (let ((a1 (+ a 1))) (lambda (x2) (cons a1 x2)))

As you see from comparing this example with the first example above, the cute-variant will evaluate (+ a 1) once, while the cut-variant will evaluate it during every invokation of the resulting procedure.

The mechanism proposed in this SRFI allows specializing any subset of the variables of a procedure. The result can be of fixed arity or of variable arity. The mechanism does not allow permutation, omission, duplication or any other processing of the arguments; for this it is necessary to write to use a different mechanism such as lambda.

Rationale

A particularly elegant way to deal with specialization is known as currying (Schönfinkel 1924, Curry 1958). The idea of currying is to reduce multi-argument functions to single-argument functions by regarding an n-ary function as a unary function mapping its first argument into an (n-1)-ary function (which is curried in turn). This point of view, apart from its theoretical elegance, allows an extremely compact notation for specializing the first argument of a function. In the first example, one could simply write (cons 1).

Yet, Scheme is not a curried language---the number of arguments passed to a procedure must match the number of its parameters at all times. This allows zero- and variable-arity procedures but in order to specialize parameters one usually has to write down a lambda-expression and invent some irrelevant identifiers for its formal variables (x in the example). For this reason, the mechanism proposed in this SRFI provides a simple and compact notation for specializing any subset of the parameters of a procedure.

Note: The mechanism proposed here is not currying!

The purpose of the mechanism proposed here is to make the benefits of currying available within the programming language Scheme. There are two primary benefits of currying in practice: Higher-order types are substantially simplified and there is a simple notation for specializing parameters. The type aspect is irrelevant as Scheme has latent typing. The specialization aspect is largly covered with this SRFI.

Here are a few more examples for illustration:
(map (cut * 2 <>) '(1 2 3 4))
(map (cut vector-set! x <> 0) indices)
(for-each (cut write <> port) exprs)
(map (cut <> x y z) (list min max))
(for-each (cut <>) thunks)

Specification

The formal syntax of a specialized expression, in the style of the Revised^5 Report on the Algorithmic Language Scheme:

<cut-expression> --> (cut <slot-or-expr> <slot-or-expr>*)
|(cut <slot-or-expr> <slot-or-expr>* <...>) ; with "rest-slot"
|(cute <slot-or-expr> <slot-or-expr>*) ; evaluate non-slots at specialization time
|(cute <slot-or-expr> <slot-or-expr>* <...>) ; with "rest-slot"
<slot-or-expr> --> <>; a "slot"
| <expression>; a "non-slot expression"

The macro cut transforms a <cut-expression> into a <lambda expression> with as many formal variables as there are slots in the list <slot-or-expr>*. The body of the resulting <lambda expression> calls the first <slot-or-expr> with arguments from <slot-or-expr>* in the order they appear. In case there is a rest-slot symbol, the resulting procedure is also of variable arity, and the body calls the first <slot-or-expr> with all arguments provided to the actual call of the specialized procedure.

The macro cute is similar to the macro cut, except that it first binds new variables to the result of evaluating the non-slot expressions (in an unspecific order) and then substituting the variables for the non-slot expressions. In effect, cut evaluates non-slot expressions at the time the resulting procedure is called, whereas cute evaluates the non-slot expressions at the time the procedure is constructed.

Implementation

The reference implementation defines the two macros cut and cute using macro mechanism of R5RS. It does not use any other SRFI or any library. The implementation makes use of internal macros to run through the list of <slot-or-expr;>s. As macros in R5RS are hygienic and referentially transparent, the macro mechanism makes sure the names of the newly introduced formal variables are unique and do not clash. The template (param ... slot), see Sect. 7.1.5. of R5RS, allows to preserve the order of arguments---which would get reversed otherwise. The reference implementation has been written by Al Petrofsky. It can be found here.

Finally, there is a small collection of confidence tests. It checks some special cases of the mechanism defined in this SRFI and signals an error in case something is wrong. Passing the tests does not mean a correct implementation.

Design Rationale

Why not real currying/uncurrying?

It is possible in Scheme to implement a macro turning a multi-argument procedure into a nesting of single-argument procedures and back. These operations are usually called "curry" and "uncurry" in other programming languages. Yet, Scheme remains an inherently uncurried language and is not prepared to deal with curried procedures in a convenient way. Hence, a "by the book" implementation of currying would only be useful if you apply it in the sequence "curry, specialize some arguments, and uncurry again"---which is exactly the purpose of the macro cut specified in this document. The primary relevance of currying/uncurrying in Scheme is to teach concepts of combinatory logic.

Why not a more general mechanism, also allowing permutation, omission and duplication of arguments?

The reason is that I, the author of this SRFI, consider more general mechanisms too dangerous to mix them with the mechanism proposed here. In particular, as soon as parameters are being rearranged it is usually necessary to be aware of the meaning of the parameters; unnamed variables can be quite harmful then. The mechanism proposed here is designed to prevent this. Please refer to the discussion threads "OK, how about...," (Alan Bawden), "is that useful?" (Walter C. Pelissero), and "l, the ultimate curry that is not curry" (Al Petrofsky).

Why are the macro called cut/cute and not [enter your favourite here]?

Well, the original name proposed for this SRFI was curry which immediately stirred some emotions as it does not what is commonly known as currying. Some alternatives have been discussed, such as section, specialise, specialize, partial-apply, partial-call, partial-lambda, _j, _i, $, &, srfi-26, foobar, xyz, schoenfinkelize, curry-which-isnt-curry, tandoori, and it has also been suggested to pick a five letter symbol uniformly at random and fix this as a name. To be fair, not all of these name have been put forward as serious proposals, some of them were merely to illustrate a point in the discussion. In addition, I have played with the game of the name quite a bit and considered other candidates not listed here. Despite the fact that the discussion list only represents a highly biased random sample of people's opinion (motivation to post a message is higher if you disagree, for example) it told me that the SRFI could potentially benefit from a different name---however impractical it may be to go for unanimous popularity. The name cut refers to "operator section", as the concept is often called in other programming languages, but I tend to remember it as the acronym for "Curry Upon This" ;-). The names for the evaluating version of cut that have been proposed were cut!, cutlet, cut*, and cute.

Is it possible to implement the SRFI without macros?

Not really. As Stephan Houben has pointed out during the discussion (refer to "Implementing it as a procedure") it is possible to implement the cute-mechanism as a procedure. Refer also to Al Petrofsky's posting "Problems with "curry"'s formal specification" for details. However, the procedural implementation comes with a slight performance penalty and it is not possible the implement the cut-mechanism as a procedure, too. As both are needed, we rely on macros to implement the SRFI.

Why is there another symbol for the rest-slot when lambda-expressions use the dotted notation for variable length argument lists?

There are two reasons. The first one is the existence of a procedural implementation of a related mechanism (refer to the previous paragraph). For a procedure, however, it is not possible to have dotted notation. The second reason is the way the hygienic macro mechanism in R5RS is defined to deal with dotted notation, as Felix Winkelmann has pointed out. Refer to the discussion threads "Improper lists in macros [WAS: none]".

Why is it impossible to specify when a non-slot is evaluated individually per non-slot?

Cut evaluates all non-slots at the time the specialized procedure is called and cute evaluates all non-slots at the time the procedure is being specialized. These are only the two extremes and it is possible to define a syntax that allows to choose per non-slot. However, I am convinced that the benefit of the greater flexibility is not worth the risk of confusion. If a piece of code really depends on the distinction, it might be better to make this explicit through let and lambda.

Why is (cut if <> 0 1) etc. illegal?

It is specified that a <slot-or-expr> must be either the slot symbol or an <expression> in the sense of R5RS, Section 7.1.3. As if is no <expression>, the above case is illegal. The reason why cut and cute are restricted in this sense is the difficulty of defining the meaning of such generalized expressions. Please refer to the discussion archive for details.

Acknowledgements

An important part of this SRFI is based on the contribution of other people, mostly through the discussion archive. In particular, the semantics and the design rationale have been greatly improved in the course of the discussion. I would like to thank all who have contributed.

Copyright

Copyright (C) Sebastian Egner (2002). All Rights Reserved.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.


Editor: Mike Sperber
Author: Sebastian Egner
Last modified: Wed Jun 19 10:54:36 MST 2002