Base class layout in C++

[Eric Botcazou <ebotcazou at libertysurf dot fr>]

Hi,

The following testcase ICEs on mainline on SPARC64:

class v
{
  double x, y;
public:
  v();
};

class w : public v {
public:
  w();
};

void bar(w x);

void foo()
{
  w my_w;
  bar(my_w);
}


Given that the size of struct 'w' is equal to 16 bytes, 'my_w' will be passed 
by value to 'bar'.  The SPARC back-end then needs to know the DECL_SIZE and 
DECL_MODE of the fields 'x' and 'y', as seen through the derived struct, in 
order to generate the call to 'bar'.  Now DECL_SIZE is 0 and DECL_MODE is 
VOIDmode for both fields, which confuses the back-end.

Questions:
(1) Should the C++ front-end have laid out the base struct?
(2) If no, should the middle-end (calls.c for example) lay it out?
(3) If no, should the back-end resort to TYPE_SIZE and TYPE_MODE?


I think (3) may be negatively answered by these lines from tree.h:

/* Holds the machine mode corresponding to the declaration of a variable or
   field.  Always equal to TYPE_MODE (TREE_TYPE (decl)) except for a
   FIELD_DECL.  */
#define DECL_MODE(NODE) (DECL_CHECK (NODE)->decl.mode)

-- 
Eric Botcazou