[Bug tree-optimization/40071] ICE (aliasing assert) in vectorizable_store at tree-vect-stmts.c:3117

[rguenth at gcc dot gnu dot org]

------- Comment #6 from rguenth at gcc dot gnu dot org  2009-05-15 12:11 -------
We are vectorizing

<bb 4>:

<bb 5>:
  # J1_43 = PHI <J1_36(6), J1_3(4)>
  J1_36 = J1_43 + 1;
  D.2354_35 = (long int) J1_36;
  VIEW_CONVERT_EXPR<struct loop_optimization1__r
*[D.2335:D.2339]>(*_init$P_ARRAY_1)[D.2354_35]{lb: D.2335_26 sz: 8} = 0B;
  if (D.2336_6 != J1_36)
    goto <bb 6>;
  else
    goto <bb 7>;

<bb 6>:
  goto <bb 5>;

using a vector pointer of type vector struct loop_optimization1__r **.

The issue here is that for

VIEW_CONVERT_EXPR<struct loop_optimization1__r
*[D.2335:D.2339]>(*_init$P_ARRAY_1)[D.2354_35]

we use the alias set of *_init$P_ARRAY_1 because the array elements are
non-aliased.  As the vectorizer now creates an alias the resulting
stores through the pointer to the vector type do not conflict with the
original stores and we have created wrong code.

How I love Ada ;)

Well, the fix is to teach the vectorizer to never vectorize arrays with
non-aliased components (or any aggregate with such elements).  Or, as
they were introduced to allow more optimization, finally drop this
non-sensical non-aliased component stuff from the middle-end as well.

:P


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40071