FRE - SCCVN problem with initialized global variables

[Daniel Berlin]

On 7/18/07, Andreas Krebbel <Andreas.Krebbel@de.ibm.com> wrote:
> Hi,
>
> GCC currently doesn't bootstrap on s390x.  The problem is that
> gengtype is miscompiled and I suspect the fre - sccvn changes to be
> the culprit.
>
> When sccvn performs the depth-first search for uses it might reach
> global variable definitions

No, global variable definitions are *never* the SSA_NAME_DEF_STMT of
anything, so it won't reach them.

We explicitly pull out initializers from variables that are:
  /* Pull out any truly constant values.  */
          if (TREE_READONLY (rhs)
              && TREE_STATIC (rhs)
              && DECL_INITIAL (rhs)
              && valid_gimple_expression_p (DECL_INITIAL (rhs)))
            return DECL_INITIAL (rhs);


This is exacfly the same set of constraints CCP uses, and my guess is
if you disable SCCVN from doing this, you will discover the same code
is generated anyway.


>. If the global variable is initialized
> with a value that value is taken as constant and used to replace uses
> of that variable.
>
> e.g.:
>
> int global = 1;
>
> int
> foo (int a)
> {
>   return a != global;
> }
>
> int
> main ()
> {
>   global = 3;
>   if (foo (3))
>     abort ();
> }
>
> becomes (compiled with -O1):
>
> foo (a)
> {
>   int global.0;
>   int D.1626;
>
> <bb 2>:
>   global.0_1 = 1;
>   D.1626_3 = a_2(D) != global.0_1;
>   return D.1626_3;
>
> }
> ...
>

This only happens where ipa-reference proves the global is  readonly
and marks it TREE_READONLY

(in this case, because there are no external function calls)
HTH,
Dan