[Bug tree-optimization/69224] [4.9/5/6 Regression] -Warray-bounds false positive with -O3 and struct pointer parameter

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69224

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so we have

                for (int j = 0; j < s->d - 1; j++) {
                    if ((c[i] >= s->x[j]) && (c[i] <= s->x[j + 1])) {
                        b[2*j][i] = s->x[j];
                        break;
                    }
                }

and we can conclude j is in [0, 4] because of the c[i] >= s->x[j] check.
But we can _not_ conclude it is in [0, 3] because of the short-circuited
c[i] <= s->x[j+1] check.  So we're left with

 if (c[4] >= s->x[4])
  {
    if (c[5] <= s->x[j + 1])
      {
 ...
      }
  }

in the last iteration.  The only issue is that we warn "is above array bounds"
rather than "may be" (well, if the access is executed then it _is_ above
array bounds, we never warn "maybe", all array-bound warnigns are "maybe"
in some way, unless in main() and always executable).

So I think we can't do anything here but not optimize (completely peel).

The issue is that usually VRP doesn't warn if it gets a range for the
access that is [0, 5] because its ranges are conservative and that would
result in way too many diagnostics.  Now the complete peeling makes the
ranges very precise (single element) and we get those false positive
warnings.

Fact is that at runtime unless s->d is "proper" (I guess it's 5) the
access may be out of bounds.