[Bug tree-optimization/58656] [4.9 Regression] rnflow regressing after r202826

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58656

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
cptrf2 shows

   <bb 32>:
-  pretmp_278 = (integer(kind=8)) ival2_72;
-  pretmp_279 = pretmp_278 + -1;
-  pretmp_280 = *xxtrt_23(D)[pretmp_279];

   <bb 33>:
   # ival2_130 = PHI <ival2_50(32), ival2_138(36)>
   # ival2_134 = PHI <ival2_72(32), ival2_140(36)>
-  # prephitmp_281 = PHI <pretmp_280(32), prephitmp_255(36)>
   _131 = (integer(kind=8)) ival2_130;
   _132 = _131 + -1;
   _133 = *xxtrt_23(D)[_132];
   _135 = (integer(kind=8)) ival2_134;
   _136 = _135 + -1;
-  _137 = prephitmp_281;
+  _137 = *xxtrt_23(D)[_136];
   if (_133 < _137)
     goto <bb 34>;
   else

gentrs has a difference as well.  It's partial-PRE results differing
(we don't have many partial-PRE testcases).

The reason for the difference is likely "failed" translations which are
cached over insert iterations.  I have a patch.