6.8.10 Hygiene and the Top-Level

Consider the following macro.

(define-syntax-rule (defconst name val)
    (define t val)
    (define-syntax-rule (name) t)))

If we use it to make a couple of bindings:

(defconst foo 42)
(defconst bar 37)

The expansion would look something like this:

  (define t 42)
  (define-syntax-rule (foo) t))
  (define t 37)
  (define-syntax-rule (bar) t))

As the two t bindings were introduced by the macro, they should be introduced hygienically – and indeed they are, inside a lexical contour (a let or some other lexical scope). The t reference in foo is distinct to the reference in bar.

At the top-level things are more complicated. Before Guile 2.2, a use of defconst at the top-level would not introduce a fresh binding for t. This was consistent with a weaselly interpretation of the Scheme standard, in which all possible bindings may be assumed to exist, at the top-level, and in which we merely take advantage of toplevel define of an existing binding being equivalent to set!. But it’s not a good reason.

The solution is to create fresh names for all bindings introduced by macros – not just bindings in lexical contours, but also bindings introduced at the top-level.

However, the obvious strategy of just giving random names to introduced toplevel identifiers poses a problem for separate compilation. Consider without loss of generality a defconst of foo in module a that introduces the fresh top-level name t-1. If we then compile a module b that uses foo, there is now a reference to t-1 in module b. If module a is then expanded again, for whatever reason, for example in a simple recompilation, the introduced t gets a fresh name; say, t-2. Now module b has broken because module a no longer has a binding for t-1.

If introduced top-level identifiers “escape” a module, in whatever way, they then form part of the binary interface (ABI) of a module. It is unacceptable from an engineering point of view to allow the ABI to change randomly. (It also poses practical problems in meeting the recompilation conditions of the Lesser GPL license, for such modules.) For this reason many people prefer to never use identifier-introducing macros at the top-level, instead making those macros receive the names for their introduced identifiers as part of their arguments, or to construct them programmatically and use datum->syntax. But this approach requires omniscience as to the implementation of all macros one might use, and also limits the expressive power of Scheme macros.

There is no perfect solution to this issue. Guile does a terrible thing here. When it goes to introduce a top-level identifier, Guile gives the identifier a pseudo-fresh name: a name that depends on the hash of the source expression in which the name occurs. The result in this case is that the introduced definitions expand as:

  (define t-1dc5e42de7c1050c 42)
  (define-syntax-rule (foo) t-1dc5e42de7c1050c))
  (define t-10cb8ce9fdddd6e9 37)
  (define-syntax-rule (bar) t-10cb8ce9fdddd6e9))

However, note that as the hash depends solely on the expression introducing the definition, we also have:

(defconst baz 42)
⇒ (begin
    (define t-1dc5e42de7c1050c 42)
    (define-syntax-rule (baz) t-1dc5e42de7c1050c))

Note that the introduced binding has the same name! This is because the source expression, (define t 42), was the same. Probably you will never see an error in this area, but it is important to understand the components of the interface of a module, and that interface may include macro-introduced identifiers.