Re: PR middle-end/28779

[Mark Mitchell <mark at codesourcery dot com>]

Jan Hubicka wrote:

> This patch fixes it by making logic similar to what expand_call does:
> using declaration arguments first, if not avaialble declaration type, if
> not call type and actual arguments otherwise.  All these steps are
> needed as our in-place duplicate_decls causes slight type mismatches
> here, so what should be all the same might be actually slightly
> different. 

I think this logic is sound, but I wonder if there is a way to factor 
that out to common code?  What you're doing is getting a bunch of 
argument types (by one mechanism or another), and then computing a cost 
for each of them.

At last for the first two cases (where you are using either 
DECL_ARGUMENTS or TYPE_ARG_TYPES), you don't need any information about 
the actual call arguments.  It's inconvenient that the information is in 
TREE_TYPE for DECL_ARGUMENTS and TREE_VALUE for TYPE_ARG_TYPES, but we 
could still do something like:

  bool for_each_arg (tree fn,
                     void (*f)(tree type, void *data),
                     void *data)
  {
     if (DECL_ARGUMENTS (fn))
       {
          for (arg = DECL_ARGUMENTS (fn) ...)
            f(TREE_TYPE (arg), data);
          return true;
       }
     else if (TREE_TYPE (fn) && prototype_p (TREE_TYPE (fn))
       {
          ...
          return true;
       }

     return false;
  }

This isn't a big deal, but this is a nasty bit of logic, and it seems 
like we should try to put it in a well-defined place for use and reuse...

--
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713