[Bug middle-end/61409] [4.9/5/6 regression] -Wmaybe-uninitialized false-positive with -O2

[law at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61409

--- Comment #16 from Jeffrey A. Law <law at redhat dot com> ---
AFAICT tree-ssa-uninit won't look at the predicate associated with an undefined
PHI argument and test it against the predicate for the actual use.

ie, given this PHI from the testcase:

;;   basic block 6, loop depth 0
;;    pred:       9
;;                5
  # ptr_1 = PHI <ptr_5(D)(9), ptr_9(5)>

We want to look at the control dependent path that leads to the edge (9,5). 
For this test, that edge is control dependent on bb2:

;;   basic block 2, loop depth 0
;;    pred:       ENTRY
  if (i_4(D) != 0)
    goto <bb 3>;
  else
    goto <bb 9>;


ie, we know that for ptr_1 to take the value ptr_5 that i_4 != 0 must be false.
 So the guard for the edge (9,5) is NOT i_4 != 0.

And in this testcase, the actual use of ptr_1 is guarded by:

  if (i_4(D) != 0)
    goto <bb 7>;
  else
    goto <bb 11>;


When i_4 != 0 is true, then we'll get to the use. So the guard for the use is
i_4 != 0

Those two guards can never both be true.  So there's no way at runtime for the
value ptr_5 to flow into ptr_1 and then into the use of ptr_1 in bb7.

And AFAICT, tree-ssa-uninit.c doesn't have the code to do that kind of
analysis.