Re: aliasing warnings [patch]

[Zack Weinberg <zack at bitmover dot com>]

Jason Merrill wrote:
> >>>>> Zack Weinberg <zack@bitmover.com> writes:
> 
>  > The warnings would be much more useful if they named the types, but I
>  > don't see a way of extracting a type name from the tree that defines
>  > it.  c-aux-info.c has a lengthy private function to do it.
> 
> The C++ frontend has convenient error-reporting functions that handle
> printing out all sorts of trees.  It might be nice to move that into
> the language-independent code.

I shall look at this.  The C front end has plenty of warnings/errors
that could be improved.

>  > I'm not sure how many false negatives this will generate.  Casting
>  > through char * or void * suppresses the warning, so a genuine aliasing
>  > bug can be masked - in particular, a memcpy() that takes char * args
>  > and casts them to long * may have problems but get no warning.
> 
> Would it make sense to warn about casts from char*?  The standard only says
> that anything can be accessed through a char*, not vice versa.

Consider this fragment:

{
   char buf[sizeof(struct foo)];
   struct foo *x = (struct foo *)buf;
   x->bar = value1;
   x->baz = value2;
   ...

   fwrite(buf, sizeof(struct foo), 1, file);
}

I believe this is well-defined.

> How about adding this to the C++ frontend, too?

It will have to be rewritten for C++'s type checking code, which is
completely different.  I fear I don't understand the C++ new-style
casts well enough to do it right.

zw