fare-quasiquote — a pattern-matching friendly implementation of Quasiquote

Copyright ⓒ 2002-2020 Fahree Reedaw fare@tunes.org

Purpose

The main purpose of this n+2nd reimplementation of quasiquote, is enable matching of quasiquoted patterns, using trivia.

Now, developing this implementation was also a challenge in understanding the ins and outs of quasiquotation, and in exploring the way it interacts with extended support for constant-building syntactic constructs. And this, at least to me was as much fun as it was an intellectual challenge.

How to use it

The recommended way to use fare-quasiquote is to have your defsystem depend on fare-quasiquote-extras, and at the start your file, use:

(named-readtables:in-readtable :fare-quasiquote)

Then you can match expressions using trivia:match, such as:

(trivia:match '(1 (2 3) 4 5)
  (`(a (b ,c) ,@d) (list a b c d)))

; => (1 2 3 (4 5))

You can also at the SLIME REPL use:

(asdf:load-system "fare-quasiquote-extras")
(named-readtables:in-readtable :fare-quasiquote)

However, beware to not leak fare-quasiquote into systems you load that do not :depends-on ("fare-quasiquote"); and so until ASDF is fixed to do that for you (hopefully some day in the ASDF 3.x series), before you call (asdf:load-system ...), you need to restore the default *readtable* with, e.g.:

(named-readtables:in-readtable :standard)

Discussion

The fare-quasiquote system is a reimplementation of quasiquote with two advantages:

• first it has all the bugs straightened out, including dark corner cases (some implementations still get the simple ``(foo ,@,@bar) wrong);

• second, it expands into a stable format that allows for pattern matching, by privileging list before cons before list* before append before nconc (and by the way, you should never, ever, use nconc).

When using trivia.quasiquote, expressions parsed by fare-quasiquote can be used as pattern matching patterns with trivia. trivia's predecessors optima and fare-matcher used to be supported, but both have long been deprecated.

We recommend you use named-readtables to enable at the beginning and end of your Lisp files, or around their compilation (e.g. using ASDF around-compile hooks). Note however that it is important not to let such readtables leak into the compilation of files that do not depend on fare-quasiquote.

Note that since pattern matchers like trivia do not support backtracking (unlike say an embedded logic programming language), they cannot match append patterns, and those quasiquote templates that expand into something using append can't be used as patterns to match. This means that the use of ,@ or ,. is restricted to the end of a list when used as a pattern.

fare-quasiquote was originally written to work with fare-matcher, and legacy support for fare-matcher is available in fare-quasiquote-matcher, now distributed with fare-matcher itself. But fare-matcher was deprecated in favor of optima, that that was deprecated in favor of trivia. Use trivia.quasiquote to match patterns written using fare-quasiquote.

References

Essential documents consulted while implementing this file:

• Alan Bawden's PEPM 99 paper: Quasiquotation in Lisp
• The CLtL2 Appendix C: Backquote
• The CLHS section 2.4.6: Backquote
• Slate reference manual section 2.6.2 on quoting and unquoting
• Factor handbook on fried quotations
• Common Lisp backquote implementation, written in Common Lisp. (public domain) Author: Guy L. Steele Jr. Date: 27 December 1985. To be used with 2010 patch by Alex Plotnick regarding the simplification pass.
• SBCL backquote implementation (derived from CMUCL, used the October 2010 version).

If you for whatever reason ever feel like reimplementing backquote, you should probably leverage the many tests in my test file quasiquote-test.lisp --- in addition to my own test cases, I notably included the tests from SBCL; they include regression tests for many actual bugs in subtle cases that you might otherwise miss.

Read-time vs Macro-expansion-time

In conformance with the CLHS, fare-quasiquote expands its patterns at read-time, at least unless you enable a feature #+quasiquote-at-macro-expansion-time.

In both cases, fare-quasiquote handles unquoting in #() and #n() syntax by re-defining the syntax for hash-left-paren as well as for backquote.

If you enable feature #+quasiquote-at-macro-expansion-time, fare-quasiquote will expands its patterns at macro-expansion time, using the same convention as Scheme, with symbols quasiquote, unquote, unquote-splicing and unquote-nsplicing defined in package fare-quasiquote (but not exported from it).

Known Issues

• Either at read-time or at macro-expansion-time, the implementation of quasiquote is fragile in case the user explicitly uses its internal syntax marker inside the expression being quasiquoted. Such expressions may lead to confusion between the body of expressions being quasiquoted and the internal quasiquote infrastructure. If you use these kinds of tricks, you're on your own. In fare-quasiquote, the syntax markers are the quasiquote, unquote, unquote-splicing and unquote-nsplicing symbols present in the fare-quasiquote package and not exported. Any objects that would be used as markers instead of these interned symbols would lead to the same "bug" if somehow used inside quasiquoted expressions; at least, non-interned symbols or gensyms would allow to avoid this bug in expressions being read. The "perfect" solution would be to use objects gensymed at the moment a given quasiquote is read, and kept in a lexical variable to prevent introspective cheat by #. syntax. Then, after simplifying expressions, we could pass the read expression through functions that would substitute proper symbols for the markers, from the fare-quasiquote package if pretty-printing is to be supported, or otherwise from the CL package. Note that this would have to be interleaved with support for working inside vectors and other syntax. This is all very tricky and until further notice is left as an exercise to the intrepid reader. Thus, while the behaviour of our implementation isn't strictly correct, we don't go through the hassle of modifying it into something much less readable just for the sake of preventing code that would deliberately confuse the quasiquote engine. Now, if we imagine that some code were dynamically generated based on system introspection, that could contain some of our magic markers, then this code would have to be made safe for interaction with quasiquote; this might (or might not?) require making quasiquote 100% fool-proof.

• This implementation allows for simplifying quasiquoting of literals and other self-evaluating objects into the object itself, instead of a `(quote ,object) expression, if you enable the #+quasiquote-passes-literals feature. This is the behaviour of the simplifier in CMUCL and SBCL, but some people have expressed concerns that it might not be strictly allowed in letter or spirit by the Common Lisp standards, and it can make the pretty-printer trickier.

• This version works inside simple vectors, so as to support unquoting inside the #(...) syntax as the standard mandates. To do that, it replaces the hash-left-paren reader-macro as well as the backquote reader-macro. Note that this does not work in #1A(...) syntax. This phenomenon has been documented before in the following message.

• Interestingly, I haven't seen the following problem stated, to know which is correct of `#5(1 ,@'(2 3)). In other words, does the read argument to #() mean something at read-time or at vector-creation-time? Of course, in the original intended usage, outside of any quasiquoting, read-time and vector-creation-time are one and the same. But what when quasiquote breaks up that identity? Our answer is that it means something at vector-creation-time.

• The CLHS section 2.4.6 says that (x1 x2 ... xn . atom) is same as (append [x1] [x2] [x3] ... [xn] (quote atom)) --- mind that the atom is preserved. This means that you can't conformantly simplify `(a b ,@c) to `(a b . ,c) unless you know that c is a proper list (if it isn't, that's an error). Yet, the CLHS itself suggests the simplification, and all implementations tested agree that a final (quote nil) can be elided:

  for l in sbcl ccl clisp cmucl ecl abcl \
          scl allegro lispworks gcl xcl ; do
      cl -l $l -i \
          '(format t "'$l': ~S~%" `(,@`(a b) ,@`c)))' \
          2>&1 | grep "^$l:" # LW, GCL are verbose
  done
  

  yet at the same time, SBCL still complains about `(,@1). fare-quasiquote follows the consensus, unless #+quasiquote-strict-append is enabled.

• The current implementation of fare-quasiquote tends to simplify away things like `,c to c and, without #+quasiquote-strict-append, also `(,@c) to c. These simplifications probably need to be somehow prevented by default. Maybe with various kinds of (identity) wrappers to indicate quoting-unquoting?

Meta Unquote Protocol (not implemented)

The CLHS specifies that quasiquoting works with simple-vector, but not with other arrays (multi-dimensional or not simple), and not with arbitrary structures. However, a fully general quasiquote facility would support quasiquoting within arbitrary syntax, using a MUP: Meta Unquote Protocol.

The MUP would allow to extend the quasiquote mechanism with support for new constant-building syntactic constructs as such constructs are defined. Maybe we will end up with a full-fledge declarative infrastructure for a Parser-Preprocessor-Pretty-Printer, like camlp4 only more declarative.

The MUP would have an abstract API for arbitrary readers; existing syntax for vectors would implemented using the MUP. For compliance reasons, further MUP extensions would be disabled by default, but could be made available with a suitable function call, e.g. for #A, #S, #P, etc.

The MUP might also implement tagged (and multiple-valued?) quasiquotes, unquotes and quotes.

Note that copying and modifying read-tables is expensive, that dynamically modifying and restoring read-tables might interfere with #. syntax, and that caching modified read-tables will interfere with any subsequent modification of a cached read-table, comparison not being possible. This means that if we wanted the MUP to adapt to existing extensions without modifying existing code, we would have to intercept the definition of syntax reading functions before they are fed to either set-macro-character or set-dispatch-macro-character, or intercept the entire reader protocol. Spooky. Now, this also requires that the current depth of quasiquoting be consulted any time any of the MUP-enabled constructors is read.

The principle of the MUP is that:

• structure readers that don't want to support unquote MUST be wrapped into something that dynamically binds *quasiquote-level* to 0.

• structure readers #c(args) that do want to support unquote MUST accumulate formal arguments to a structure constructor into a list args, then, if *quasiquote-level* is 0, behave like #.(apply `c `args) otherwise, behave like `,(apply `c `args) where c is the name of the constructor for given structure, and args is whichever arguments have been deduced from the syntax, which may include as many levels of unquotations as *quasiquote-level* says. Note that in a strong sense, #. is like , assuming an infinite tower of read-time evaluators à la 3-LISP.

Note that the above is obscured because we're trying to discuss the behaviour of quasiquote-containing programs without having a meta-level quasiquoting facility that could distinguish between what is constant or variable at the meta-level independently from what is constant or variable at the base level: c and args would be better specified through a special meta-level unquotation, the above expressions being in a corresponding special quasiquotation. A feature that would allow for clear separation of levels of meta-language would be a tagged quasiquote feature, as in Slate.

The idea of making circular data-structures work within quasiquotation makes my head ache with overarching pain. You're crazier than I am if you do it and do it right.

PS: If you're able to follow this discussion, you impress me. Come join the TUNES project!