[Bug fortran/67170] PRE can't hoist out a readonly argument

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67170

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-08-11
     Ever confirmed|0                           |1

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ah.

  # USE = nonlocal escaped { D.3438 }
  # CLB = nonlocal escaped
  foo (&D.3438);

As it clobbers nonlocal it clobbers the incoming argument :/  Because of course
as the argument is pointing to possibly global memory it can access it via
other means than through the argument - and the function attribute in use
(".r") just means accesses through the first argument do not modify the
contents pointed to.

I think Fortran guarantees that things passed as argument are not modified
in other ways (or it makes that undefined).  Not sure about called functions
though.

Now.  In principle I think C says that with

void foo (int * __restrict p)
{
 ...
 bar ();
 ...
}

bar () can only access *p through the use of p or a derived pointer.  So


int x;
bar()
{
  x = 1;
}
main()
{
  foo (&x); 
}

would invoke undefined behavior.

Of course our restrict implementation isn't able to capture that (calls
don't have a clique/base which they would get from the "memory" passed
as reference arguments).

So - confirmed.  But no easy way out :(

Old idea about special-casing recursive calls also exists but without clear
idea of what would be legal special alias answers here.