Re: Sibcall arguments passed on the stack

["John David Anglin" <dave at hiauly1 dot hia dot nrc dot ca>]

> On Tue, Sep 30, 2003 at 11:10:58PM -2800, John David Anglin wrote:
> > Yes, that appears to be the best fix.  What had bothered me about it
> > was there appears to have been a concious decision not to add stack
> > slots to the CALL_INSN_FUNCTION_USAGE.
> 
> For non-const functions, there is no need, since "all memory" is
> considered clobberable.  For const functions, what we have now is
> partially a mistake -- iirc Dale found that we add the use when
> we're adding libcall notes, but I think we should do it all the time.

Ok, it looks as if we may have a problem in determining when a function
is non-const (PR 12372).  It would appear that the following function
in gcc.c-torture/execute/20030307-1.c is treated as a const function:

int fcntl_lock(int fd, int op, long long offset, long long count, int type)
{
  return type;
}

This allows stack arguments to be deleted.  However, because of the
number and size of the arguments, the argument `type' has to be passed
on the stack (i.e., in global memory) on the 32-bit pa ports.  Thus,
the function must be non-const in this case as it has to read global
memory.

Dave
-- 
J. David Anglin                                  dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada              (613) 990-0752 (FAX: 952-6602)