Style warning when matching structs on SBCL.

[marcoheisig]

The following snippet of code produces a weird style warning on SBCL. (Both on 1.4.5 and 1.4.15):

(in-package #:cl-user)

(defstruct (foo (:predicate foop)))

(trivia:defpattern foo ()
  (let ((it (gensym)))
    `(trivia:guard1 ,it (foop ,it))))

(defun foo-finder (x)
  (trivia:match x
    ((foo) (format t "Found a foo!"))))

The style warning is

; in: defun foo-finder ; (TRIVIA.LEVEL2:MATCH X ; ((FOO) (FORMAT T "Found a foo!"))) ; --> TRIVIA.LEVEL2:MATCH2 TRIVIA.LEVEL2:MATCH2+ ; ==> ; (TRIVIA.LEVEL2:MATCH2*+ (X) ; (T) ; (((FOO)) (FORMAT T "Found a foo!")) ; ((TRIVIA.LEVEL2.IMPL::_) NIL)) ; ; caught style-warning: ; undefined type: (foo) ; ; compilation unit finished ; Undefined type: ; (foo) ; caught 1 STYLE-WARNING condition

I only get this style warning with the latest Trivia release on Quicklisp, so it must have been due to a recent change.

[guicho271828]

• I can reproducible this on SBCL 1.5.5
• You can find the source of this condition on the debugger by (setf *break-on-signals* t) and continue (i.e. ignore))) the first signal.
• The ~error~ (Edit: style-warning) comes from type-i:all-compound-types, which is a function that tries to enumerate all possible type specifier for an atomic type specifier. For example foo is a valid typespec then (foo) is also a valid typespec, while (foo *) is not a valid typespec. In constrast array, (array), (array *) (array * *) are all valid typespecs. The function heuristically tries to catch when to stop adding *s to the type specifier. https://github.com/guicho271828/type-i/blob/master/src/infer-typep.lisp#L39 See comments in the source code.
• It seems during the compile time the error is converted into a style warning.
• Trivia switched to balland2006 optimizer recently, which makes it depend on type-i. So yes, your analysis is correct.

[marcoheisig]

Thank you for the quick reply and the excellent explanation!

I am trying to think what a sensible workaround would be. I see the following options:

1. Simply muffle style warnings on SBCL where necessary.
2. Improve type-i:all-compound-types such that when the atomic type specifier in question is a subtype of structure-class, it only emits foo and not (foo).

I think I prefer option 2. If you want, I can implement that and open a pull request.

[guicho271828]

I also prefer 2. thanks

[marcoheisig]

Here we go: https://github.com/guicho271828/type-i/pull/2.

[guicho271828]

merged. thanks!