R7RS Scheme Progress Report

John Cowan recently gave a talk on the progress of R7RS (slides), the next revision of the Scheme language standard, at LispNYC. After the R6RS debacle, the community stepped back and is now basing the next standard on R5RS; the work has been split into two languages — R7RS-Small and R7RS-Large. The first working group is preparing to issue a final draft of the R7RS-Small language (PDF; clocking in at 73 pages vs. R5RS's 50) within the next few weeks. Read on for a summary of the planned changes to R7RS (more or less in the order of presentation).

The talk details a number of improvements over R5RS and R6RS, and is divided into two portions. The majority of the talk discuses the status of the small language, with the last portion giving a quick update on the future intent of the large language group.

First on the list of major new features is a mandatory library system designed to be easily implemented atop existing module systems. R6RS's library system proved to tackle too much at once and be incompatible with everything already in a use (a persistent concern in the design of the R7RS-Small language). The R7RS system, on the other hand, is static, simple, and just powerful enough to promote implementation of portable libraries.

Exceptions are stripped down from R6RS (which proved too incompatible with existing implementations for practical use). guard (similar to try...catch in other languages) and with-exception-handler are supported: the latter runs the handler in the context from which the error was triggered (permitting recovery from the error like invoke-restart ) Unlike r6rs, exceptions lack a strictly defined hierarchy and can be any object (you could e.g. throw 4 if you really felt like it).

Dynamically scoped variables, in the form of parameters, are now part of the standard. Unlike Common Lisp special variables, parameters are first-class objects bound to expressions rather than symbols that are declared dynamically scoped. For reasons of simplicity it was decided to make parameters immutable (i.e. any "mutation" has to be done by rebinding). This has the (intentional) side-effect of making parameters play more nicely with threads (when mutation is permitted, setting a parameter that has not been rebound in the current thread requires synchronizing all threads and can have unexpected results).

As expected, R7RS includes bytevectors to complement strings. The small language standard only permits accessing bytevectors as ordered sets of unsigned 8-bit values (the large language standard will offer more flexible access). Binary I/O is implemented via a set of parallel procedures (open-binary-input-file vs. open-input-file, etc.) in contrast to the incredibly complicated dual binary/text ports provided by R6RS. Additionally, string and bytevector ports similar to SRFI-6 are provided instead of the incompatible string ports provided by R6RS.

Taylor Campbell noted that integer division in most languages is insufficiently expressive, and so R7RS will provide Euclidean division and centered division in addition to the usual suspects. Mathematicians rejoice!

As with R6RS, Unicode support is mandated. Unlike R6RS, the only characters that must be supported are those present in ASCII. For supported characters, Unicode case mapping and normalization are mandatory. One interesting diversion from Unicode and R6RS is that string order comparison is implementation-dependent: this gives implementers latitude with the internal encoding of strings.

Any user of Scheme knows that the language strives for consistent and obvious names for bindings. R7RS furthers this goal by resolving the long-standing inconsistency with core data structure creation, copying, mutation, mapping, etc. Lists, strings, and vectors now have a consistent set of each: make-TYPE, TYPE-copy, TYPE-set!, etc. Conversion procedures between all three types are also provided.

Finally, number of minor improvements were made. Most notable: record types are compatible with SRFI-6 (widely in use today); case sensitivity is the default (with optional insensitivity via include-ci); s-expression and nested block comments were added; IEEE infinities and NaNs have read syntax; strings may contain C-style escape sequences; Common Lisp circular list notation is supported; and common extensions to syntax-rules were standardized.

The final 20 minutes of the talk were about the status of R7RS-Large.

The R7RS Large language is currently on hold, mostly because all the small language members are on the large language committee as well, so there is a lack of time to work on both simultaneously. Work is planned to resume upon release of the final draft of the small language. Some work, however, has been completed.

The main focus of R7RS-Large is providing Scheme "with batteries included." John Cowan started the process by looking beyond the Lisp and Scheme communities (Python is mentioned) to figure out which libraries modern programmers expect their language to include.

This resulted in a list of around 250 packages that was narrowed down to around 80 packages through an initial voting process. It was decided then than some desirable packages (e.g. a foreign function interface) had to be omitted due to complexity. It is expected that implementers will continue experimenting and gradually come to a consensus on the larger packages using the existing SRFI process, and perhaps another revision of the standard down the road.

Of the packages planned for inclusion, the most prominent are: networking, threads, regular expressions, delimited continuations, URI handling, date and time parsing/arithmetic/formatting, hash tables, ambient environment access, file system directory access, gettext (i18n support), and pattern matching.

Most will be optional; packages will only be made mandatory if a number of the other packages require them. A compliant R7RS-Large implementation will only have to either provide a package fully or not at all (half implementations are forbidden).

Interestingly, R7RS large with all packages will be even larger than Common Lisp.

To avoid getting bogged down in stylistic discussions, a decision was made to focus on functionality above other concerns. The resulting packages may not be aesthetically pleasing to everyone, but will provide useful functionality. Users who disagree with naming, scope of functionality, argument ordering, etc. will be free to use the library system to import only the bindings they want, rename functions, wrap things into the style they want, etc. Basically, a compromise between the MIT approach and "Worse is Better" is being sought.

Here a call for volunteers is made: Since the focus will be on functionality over pure aesthetics, developers outside of the Lisp and Scheme communities are actively encouraged to participate in the R7RS-Large language process. No fixed time commitment would be required; the goal is to get a lot of people involved with a few core members maintaining momentum and guiding the process. The R7RS-Large language is most definitely a language designed by and for developers. So, make your voice heard!

Overall R7RS is shaping up to be the standard R6RS should have been (which, of course, could not have happened without the lessons of R6RS). The split between an elegant core language with each design issue meticulously fretted over and voted upon and a looser library standard should, hopefully, result in a core language that will stand the test of time with a standard library that can be used to get actual work done.