struct by value doesn't work without runtime translation

[Attila Lendvai]

the issue:

(progn
  (asdf:test-system :cffi-tests)
  (in-package :cffi-tests))

(with-foreign-object (arg '(:struct struct-pair))
  (convert-into-foreign-memory '(40 . 2) '(:struct pair) arg)
  (sumpair arg))

=> a runtime error because the current fsbv assumes that everything goes through runtime translators.

[Attila Lendvai (attila-lendvai)]

and a half-working solution:

defcstruct's define a lisp class to represent the struct type, and this lisp class inherits from both (CFFI::FOREIGN-STRUCT-TYPE CFFI::TRANSLATABLE-FOREIGN-TYPE).

after a long session of staring at trace output and letting my creativity loose in the forest of EXPAND-TO-FOREIGN-DYN, EXPAND-TO-FOREIGN-DYN-INDIRECT, TRANSLATE-TO-FOREIGN, TRANSLATE-INTO-FOREIGN-MEMORY, and friends... i've identified the magic 3 lines that fixes this:

(defmethod expand-to-foreign-dyn-indirect (value var body (type foreign-struct-type))
  `(let ((,var ,value))
     ,@body))

the problem is that this breaks type translation because it kicks in before the method specialized on TRANSLATABLE-FOREIGN-TYPE.

maybe there's someone with more insight into the cffi type internals who can come up with a fix for this?

unfortunately FOREIGN-POINTER is not a class, so we cannot dispatch on the value in CFFI::TRANSLATE-INTO-FOREIGN-MEMORY.

[Attila Lendvai (attila-lendvai)]

FTR, redefining this function achieves the goal and keeps all the tests running:

(defmethod expand-to-foreign-dyn-indirect :around
    (value var body (type translatable-foreign-type))
  (let ((*runtime-translator-form*
          `(if (typep ,value 'foreign-pointer)
               (let ((,var ,value))
                 ,@body)
               (with-foreign-object (,var ',(unparse-type type))
                 (translate-into-foreign-memory ,value ,type ,var)
                 ,@body))))
    (call-next-method)))

but it's kludgy, because it should dispatch on foreign-struct-type, and shouldn't duplicate ,@body.

[Angel Peralta (acort3255)]

Is this truly a bug? It seems like passing pointer to foreign struct should not be passed by value to the corresponding c function:

(defcfun "sumpair" :int
  (p (:struct struct-pair)))

corresponding failing test:

(deftest (fsbv.wfo :expected-to-fail t)
    (with-foreign-object (arg '(:struct struct-pair))
      (convert-into-foreign-memory '(40 . 2) '(:struct struct-pair) arg)
      (sumpair arg))
  42)

Is there something I'm missing here?

[Attila Lendvai (attila-lendvai)]

that's arguable, but i think it is indeed a bug.

rationale:

the CFFI domain does not have a concept of pointers (i.e. there's only one way using with-foreign-object to get hold of a new :struct). and it's an implementation detail how that struct is accounted for. and therefore, it's reasonable to expect CFFI to automatically do the right thing when such a struct is passed to a C fuction that expects a struct by value, or a struct pointer.