Re: [bug-mes] MesCC neglects alignment requirements of the System V ABI

[Danny Milosavljevic]

Hi Mark,
Hi Janneke,

> Suppose I have the following struct:
> 
>   struct { char a; int b; } st;
> 
> If I understand correctly the code correctly, the offset of 'b' will be
> 1 byte into the structure.

@Mark: I think you are right, it does that O_o.

@Janneke: Hmm, we might want to change it (to make it pad each field to an 
entire word) eventually (for ARMv4 and many other platforms).

Currently, it is the following:

(define (struct->type-entry info name fields)
  (let ((size (apply + (map (compose (cut ->size <> info) cdr) fields))))
    (cons `(tag ,name) (make-type 'struct size fields))))

(define (->size o info)
  (cond ((and (type? o) (eq? (type:type o) 'union))
         (apply max (map (compose (cut ->size <> info) cdr) (struct->fields 
o))))
        ((type? o) (type:size o))
        ((pointer? o) (->size (get-type "*" info) info))
        ((c-array? o) (* (c-array:count o) ((compose (cut ->size <> info) 
c-array:type) o)))
        ((local? o) ((compose (cut ->size <> info) local:type) o))
        ((global? o) ((compose (cut ->size <> info) global:type) o))
        ((bit-field? o) ((compose (cut ->size <> info) bit-field:type) o))
        ((and (pair? o) (pair? (car o)) (bit-field? (cdar o))) ((compose (cut 
->size <> info) cdar) o))
        ((string? o) (->size (get-type o info) info))
        (else (error "->size>: not a <type>:" o))))

[...]
 `(("char" . ,(make-type 'signed 1 #f))

(the new padded version could be called "allocation" instead of "size")

For completeness: The current version is allowed in C, so that is a valid way 
to implement C - though annoying to have a backend for.  If we wanted to 
support the System V ABI then the current version wouldn't be allowed.

pgpqEFhqjnpsb.pgp
Description: OpenPGP digital signature