[Bug bootstrap/30511] False array bound check causes gcc failed to boostrap

["roger at eyesopen dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #2 from roger at eyesopen dot com  2007-01-19 20:55 -------

It looks like the ivopts pass is creating dubious? array references.
The symbol.c.053t.vrp1

  D.12252_12 = (long unsigned int) i_2;
  D.12255_15 = &ns_4->default_type[D.12252_12];
  ts_16 = D.12255_15 + -2328B;

i.e. we now have "&ns->default_type[i] - &ns->default_type['a']".
I wouldn't like to offer an opinion on whether this IV is valid in the
middle-end.  The transformation is certainly unsafe in C, where a pointer
arithmetic is not guaranteed to be valid outside of the declared object.

I think the appropriate fix is that if ivopts is going to create these
suspicious array refs, it should appropriately set TREE_NO_WARNING, which
will then cause the reference to be ignored by the bounds checking in VRP.

Alternatively, we need more context in array bounds checking such that
&x['a'] is not a problem, but x['a'] on it's own is diagnosed.  i.e. we
could perhaps keep an in_indirect_ref_p flag during the tree walking.

I hope this helps.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30511