8.1 Introduction
This section provides an introduction to writing robust macros with
syntax-parse and syntax classes.
As a running example we use the following task: write a macro named
mylet that has the same syntax and behavior as Racket’s
let form. The macro should produce good error messages when
used incorrectly.
Here is the specification of mylet’s syntax:
(mylet ([var-id rhs-expr] ...) body ...+) | (mylet loop-id ([var-id rhs-expr] ...) body ...+) |
|
For simplicity, we handle only the first case for now. We return to
the second case later in the introduction.
First, we import syntax-parse into the transformer environment,
since we will use it to implement a macro transformer.
We get the first version of mylet by essentially
transliterating the syntax specification above. The result is similar
to what one would write using syntax-rules or perhaps
syntax-case.
Note the use of ... and ...+ in the pattern;
... means match zero or more repetitions of the preceeding
pattern; ...+ means match one or more. Only ... may
be used in the template, however.
> (mylet ([a 1] [b 2]) (+ a b)) |
3 |
> (mylet (b 2) (sub1 b)) |
mylet: bad syntax at: (mylet (b 2) (sub1 b)) |
> (mylet ([1 a]) (add1 a)) |
lambda: not an identifier, identifier with default, or |
keyword at: 1 |
> (mylet ([#:x 1] [y 2]) (* x y)) |
struct eval:125:0: expects 0 arguments plus an argument |
with keyword #:x, given 2: 1 2 |
When used correctly, the macro works, but it behaves very badly in the
presence of errors. In some cases, mylet blithely accepts
illegal syntax and passes it along to lambda, with strange
consequences.
These examples of illegal syntax are not to suggest that a typical
programmer would make such mistakes attempting to use
mylet. At least, not often. After an initial learning
curve. But macros are also used by inexpert programmers and as targets
of other macros (or code generators), and many macros are far more
complex than mylet. Macros must validate their syntax and
report appropriate errors. Furthermore, the macro writer benefits from
the machine-checked specification of syntax in the form of more
readable, maintainable code.
The first step toward validation and high-quality error reporting is
annotating each of the macro’s pattern variables with the syntax
class that describes its acceptable syntax. In mylet, each
variable must be an identifier (id for short) and
each right-hand side must be an expr (expression). An
annotated pattern variable is written by concatenating the
pattern variable name, a colon character, and the syntax class
name.For an alternative to the “colon” syntax, see the
~var pattern form.
Note that the syntax class annotations do not appear in the template
(i.e., var, not var:id).
The syntax class annotations are checked when we use the macro.
> (mylet ([a 1] [b 2]) (+ a b)) |
3 |
> (mylet (["a" 1]) (add1 a)) |
mylet: expected identifier at: "a" |
The
expr syntax class does not actually check that the term
it matches is a valid expression—
that would require calling that
macro expander. Instead,
expr just means not a keyword.
> (mylet ([a #:whoops]) 1) |
mylet: expected expression at: #:whoops |
Also,
syntax-parse knows how to report a few kinds of errors
without any help:
> (mylet ([a 1 2]) (* a a)) |
mylet: unexpected term at: 2 |
There are other kinds of errors, however, that this macro does not
handle gracefully:
> (mylet (a 1) (+ a 2)) |
mylet: bad syntax at: (mylet (a 1) (+ a 2)) |
It’s too much to ask for the macro to respond, “This expression is
missing a pair of parentheses around (a 1).” The pattern
matcher is not that smart. But it can pinpoint the source of the
error: when it encountered a it was expecting what we might
call a “binding pair,” but that term is not in its vocabulary yet.
To allow syntax-parse to synthesize better errors, we must
attach descriptions to the patterns we recognize as discrete
syntactic categories. One way of doing that is by defining new syntax
classes:Another way is the ~describe pattern
form.
Note that we write b.var and b.rhs now. They are the
nested attributes formed from the annotated pattern variable
b and the attributes var and rhs of the
syntax class binding.
Now the error messages can talk about “binding pairs.”
> (mylet (a 1) (+ a 2)) |
mylet: expected binding pair at: a |
Errors are still reported in more specific terms when possible:
> (mylet (["a" 1]) (+ a 2)) |
mylet: expected identifier at: "a" |
There is one other constraint on the legal syntax of
mylet. The variables bound by the different binding pairs
must be distinct. Otherwise the macro creates an illegal
lambda form:
> (mylet ([a 1] [a 2]) (+ a a)) |
lambda: duplicate argument name at: a |
Constraints such as the distinctness requirement are expressed as side
conditions, thus:
> (mylet ([a 1] [a 2]) (+ a a)) |
mylet: duplicate variable name at: a |
The #:fail-when keyword is followed by two expressions: the
condition and the error message. When the condition evaluates to
anything but #f, the pattern fails. Additionally, if the
condition evaluates to a syntax object, that syntax object is used to
pinpoint the cause of the failure.
Syntax classes can have side conditions, too. Here is the macro
rewritten to include another syntax class representing a “sequence of
distinct binding pairs.”
Here we’ve introduced the #:with clause. A #:with
clause matches a pattern with a computed term. Here we use it to bind
var and rhs as attributes of
distinct-bindings. By default, a syntax class only exports
its patterns’ pattern variables as attributes, not their nested
attributes.The alternative would be to explicitly declare
the attributes of distinct-bindings to include the nested
attributes b.var and b.rhs, using the
#:attribute option. Then the macro would refer to
bs.b.var and bs.b.rhs.
Alas, so far the macro only implements half of the functionality
offered by Racket’s let. We must add the
“named-let” form. That turns out to be as simple as adding
a new clause:
We are able to reuse the
distinct-bindings syntax class, so
the addition of the “named-
let” syntax requires only three
lines.
But does adding this new case affect
syntax-parse’s ability
to pinpoint and report errors?
> (mylet ([a 1] [b 2]) (+ a b)) |
3 |
> (mylet (["a" 1]) (add1 a)) |
mylet: expected identifier at: "a" |
> (mylet ([a #:whoops]) 1) |
mylet: expected expression at: #:whoops |
> (mylet ([a 1 2]) (* a a)) |
mylet: unexpected term at: 2 |
> (mylet (a 1) (+ a 2)) |
mylet: expected binding pair at: a |
> (mylet ([a 1] [a 2]) (+ a a)) |
mylet: duplicate variable name at: a |
The error reporting for the original syntax seems intact. We should
verify that the named-
let syntax is working, that
syntax-parse is not simply ignoring that clause.
> (mylet loop ([a 1] [b 2]) (+ a b)) |
3 |
> (mylet loop (["a" 1]) (add1 a)) |
mylet: expected identifier at: "a" |
> (mylet loop ([a #:whoops]) 1) |
mylet: expected expression at: #:whoops |
> (mylet loop ([a 1 2]) (* a a)) |
mylet: unexpected term at: 2 |
> (mylet loop (a 1) (+ a 2)) |
mylet: expected binding pair at: a |
> (mylet loop ([a 1] [a 2]) (+ a a)) |
mylet: duplicate variable name at: a |
How does syntax-parse decide which clause the programmer was
attempting, so it can use it as a basis for error reporting? After
all, each of the bad uses of the named-let syntax are also
bad uses of the normal syntax, and vice versa. And yet the macro doen
not produce errors like “mylet: expected sequence of
distinct binding pairs at: loop.”
The answer is that syntax-parse records a list of all the
potential errors (including ones like loop not matching
distinct-binding) along with the progress made before
each error. Only the error with the most progress is reported.
For example, in this bad use of the macro,
> (mylet loop (["a" 1]) (add1 a)) |
mylet: expected identifier at: "a" |
there are two potential errors: expected
distinct-bindings at
loop and expected
identifier at
"a". The
second error occurs further in the term than the first, so it is
reported.
For another example, consider this term:
> (mylet (["a" 1]) (add1 a)) |
mylet: expected identifier at: "a" |
Again, there are two potential errors: expected
identifier at
(["a" 1]) and expected
identifier at
"a". They both occur at the second term (or first argument,
if you prefer), but the second error occurs deeper in the
term. Progress is based on a left-to-right traversal of the syntax.
A final example: consider the following:
> (mylet ([a 1] [a 2]) (+ a a)) |
mylet: duplicate variable name at: a |
There are two errors again: duplicate variable name at
([a 1] [a 2]) and expected
identifier at
([a 1] [a 2]). Note that as far as
syntax-parse is concerned, the
progress associated with the duplicate error message is the second
term (first argument), not the second occurrence of
a. That’s
because the check is associated with the entire
distinct-bindings pattern. It would seem that both errors
have the same progress, and yet only the first one is reported. The
difference between the two is that the first error is from a
post-traversal check, whereas the second is from a normal
(i.e., pre-traversal) check. A post-traveral check is considered to
have made more progress than a pre-traversal check of the same term;
indeed, it also has greater progress than any failure
within
the term.
It is, however, possible for multiple potential errors to occur with
the same progress. Here’s one example:
> (mylet "not-even-close") |
mylet: expected identifier or expected sequence of distinct |
binding pairs at: "not-even-close" |
Even with all of the annotations we have added to our macro, there are
still some misuses that defy
syntax-parse’s error reporting
capabilities, such as this example:
> (mylet) |
mylet: bad syntax at: (mylet) |
The philosophy behind
syntax-parse is that in these
situations, a generic error such as “bad syntax” is justified. The
use of
mylet here is so far off that the only informative
error message would include a complete recapitulation of the syntax of
mylet. That is not the role of error messages, however; it is
the role of documentation.
This section has provided an introduction to syntax classes, side
conditions, and progress-ordered error reporting. But
syntax-parse has many more features. Continue to the
Examples section for samples of other features in
working code, or skip to the subsequent sections for the complete
reference documentation.