[Bug tree-optimization/52406] [4.7 Regression] likely wrong code bug

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52406

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Blocks|                            |50067

--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-02-28 14:32:58 UTC ---
The issue is that we want to disambiguate a[i].f1 and a[i].f2, so for the
"base object" we zero out all known indices, resulting in a[0].f1 and a[0].f2
which we then disambiguate (and conclude that for all i there cannot be
a dependence).

Now, when we mix pointer accesses with array accesses most of the index
analysis falls apart (which is what the fix for PR50067 tries to make
work more reliably - see its comment #13 on the dr_may_alias_p issue ...)

So, it's really wrong to try to fixup DR_BASE_OBJECT to make dr_may_alias_p
work, and it is equally wrong to use DR_BASE_OBJECT in dr_may_alias_p.
Using DR_REF (a safe bet) falls foul of failing a load of testcases, for
example gcc.dg/vect/pr37027.c which is no longer vectorized because

(compute_affine_dependence
  stmt_a: D.1722_7 = a[i_24].f1;
  stmt_b: D.1725_11 = a[i_24].f2;
) -> dependence analysis failed

previously we'd have used a[0].f1 and a[0].f2 in the disambiguation in
dr_may_alias_p and disambiguated the accesses.

We can try a similar trick as with REALPART/IMAGPART_EXPR to recover this.
Add a constant access function for outer COMPONENT_REFs (those we can strip
off the base object).