[Bug tree-optimization/60510] SLP blocks loop vectorization (with reduction)

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60510

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Blocks|                            |53947

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ah, too late for reassoc to catch it.  Only LIM store motion makes it a
scalar reduction, and:

  # prephitmp_32 = PHI <pretmp_31(5), _18(8)>
  _8 = prephitmp_32;
...
  _13 = _12 + _8;
  _18 = _13 + _16;
  *a_7(D) = _18;

it's already botched after PRE (and before).  So reassoc _could_ catch it
and re-associate

  _8 = *a_7(D);
  _9 = (integer(kind=8)) i_1;
  _10 = _9 + -1;
  _12 = *x_11(D)[_10];
  _15 = *y_14(D)[_10];
  _16 = _15 * _12;
  _13 = _12 + _8;
  _18 = _13 + _16;
  *a_7(D) = _18;

so that the store to *a_7(D) is fed by _8 + ...

But that seems to be a bit too much special-casing.  There isn't a good
pass before vectorization that could handle this so I think this is
finally a reason to make the vectorizer more properly recognize
reductions ...

(and keep the analysis result recorded somewhere, not try to re-match
stuff during transform phase - that's always flaky)