SKINT implements all standard features of R7RS `syntax-rules`, including custom ellipsis, non-final ellipsis patterns, non-binding underscore pattern, and `(... tpl)` template escapes. It also supports the following extensions:
Boxes, as defined by SRFI-111 and the future `(scheme box)` library, are supported natively, and can be parts of both patterns and templates. See examples of their use below.
A pattern of the form `(<ellipsis> <pattern>)`, where `<ellipsis>` is the current ellipsis, is interpreted as if it were `<pattern>`, but ellipses and underscores in `<pattern>` lose their special meaning, e.g.:
Note that R7RS prescribes special treatment of keyword identifier at the beginning of the pattern in a `<syntax rule>`: it is matched automatically with the head of the use form, but is not considered a pattern variable. SKINT's pattern escape extension drops this positional restriction, and matches its sub-pattern in a normal way, e.g.:
A pattern of the form `(<ellipsis> <predicate name> <pattern>)` where `<ellipsis>` is the current ellipsis is interpreted as if it were `<pattern>`, with additional constraint that the S-expression it matches should also satisfy the constraint specified by `<predicate name>`. Predicate names are compared to predefined symbols according to `free-identifier=?` rules. The following named pattern escapes are supported:
*`(... number? <pattern>)`
*`(... exact-integer? <pattern>)`
*`(... boolean? <pattern>)`
*`(... char? <pattern>)`
*`(... string? <pattern>)`
*`(... bytevector? <pattern>)`
*`(... id? <pattern>)`
All but the last predicate have the same meaning as the corresponding Scheme procedures. The `id?` predicate checks if the corresponding S-expression is either a symbol or a syntax object representing an identifier.
The rationale for adding these escapes is obvious: while `syntax-rules`-based macros can perform very complex calculations with structured S-expressions, they lack an ability to deal with *atomic* S-expressions (with the exception of identifiers – they can be recognized, but the technique for that is quite complicated).
A template of the form `(<ellipsis> <converter name> <template+>)` where `<ellipsis>` is the current ellipsis is interpreted as follows. First, `<template+>` (which can be any nonempty sequence of `<template>`s), is instantiated recursively, resulting in a list of S-expressions. These S-expressions become arguments to a converter specified by `<converter name>`. It is a syntax error to apply converters to a wrong type or number of arguments. Converter names are compared to predefined symbols according to `free-identifier=?` rules. The following named template escapes are supported:
All but the last two converters have the same meaning as the corresponding Scheme procedures. The `id->string` converter expects either a symbol or a syntax object representing an identifier and produces a string containing a “quote name” of the identifier (the result of applying `symbol->string` to the original name supplied by the user after all substitutions).
The `string->id` converter allows one to produce identifiers having the same syntax properties as identifiers explicitly introduced as part of macro definitions or macro uses. The properies are copied from `<id template>`, which, after all substitutions are performed, should instantiate to an identifier serving as a prototype. The `<template>` argument should instantiate to a string, which is then converted to a symbol via `string->symbol` and then turned into an identifier syntax object *as if* it was introduced side-by-side with the prototype identifier (same expression, same expansion phase).
Please note that identifiers generated with `string->id` are not autorenamed with other “free” template
identifiers; their syntactic identity is defined entirely by that of `<id template>` id, which might have already being renamed by the time `string->id` converter is applied.
Note that in the last example the escaped keyword `ref-id` at the beginning of the pattern was used to bring in the `define-math-constants` from the macro use to serve as a prototype id for introduced `pi` and `e` identifiers, allowing them to capture the corresponding identifier names typed in by the user in `(+ pi e)`. Without simple pattern escape, this keyword would not be treated as a pattern variable.
The above collection of named escapes is selected as *almost* minimal one. Its purpose is not to make `syntax-rules`-based macro programming more convenient, but just to extend its core abilities in dealing with non-structural S-expressions, so it is possible to recognize them and work with them via convertion to/from structural form if a need arises. Arithmetics is limited to what one can do using lists as Peano numbers; also, for numbers and chars, access to ordering is provided, to support simple ranges. One can imitate `string-append` without a dedicated converter, but this unnecessarily complicates generation of identifiers, which is a major use case.