[Bug middle-end/61409] [5/6/7 regression] -Wmaybe-uninitialized false-positive with -O2

[aldyh at gcc dot gnu.org]

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

--- Comment #19 from Aldy Hernandez <aldyh at gcc dot gnu.org> ---
Also, unless I'm missing something, in Jeff's analysis, I see no reference to
j, which plays a pivotal role.

In the testcase in comment 14, we can see that the guard for ptr_14 is actually
[i && j==0]:

    if (i)
      {
        if (j)
          return; /* bar(); */
        ptr = init();
      }

...because on j != 0, we exit, which is what the .uninit dump is suggesting
with:

[AFTER NORMALIZATION -- [DEF]:
ptr_14 = PHI <ptr_18(D)(9), ptr_22(5)>
is guarded by :

 (.NOT.) j_20(D) != 0 (.AND.) i_17(D) != 0

Meanwhile, the use we are warning on is [i != 0]:

[AFTER NORMALIZATION -- [USE]:
rw = ptr_14;
is guarded by :

i_17(D) != 0

So the problem is that while the guard for the DEF is [i != 0 && j == 0], the
guard for the USE is only [i != 0].  Uninit should somehow be aware that if PTR
was set, then  j==0, because otherwise we would've exited the function.

You can see that .uninit is expecting the use to be guarded by [i != 0 && j ==
0], by adding "&& !j" to the guard and silencing the warning:

   if (i && !j)
    {
      rw=ptr;
    }