[Bug tree-optimization/63406] New: -Warray-bounds false positive with -O3

[Bug tree-optimization/63406] New: -Warray-bounds false positive with -O3

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63406

            Bug ID: 63406
           Summary: -Warray-bounds false positive with -O3
           Product: gcc
           Version: 4.9.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org

int total_objs = 0;
#ifndef N
#define N 2
#endif
void *objs[N];

void
del_obj (void *obj)
{
  int i = 0;
  while (i < total_objs && objs[i] != obj)
    ++i;
  if (i >= total_objs)
    return;
  while (++i < total_objs)
    objs[i - 1] = objs[i];
  --total_objs;
}

Both loops seem to be unrolled N times (although the second loop clearly could
be
only executed N-1 times, because otherwise it triggers undefined behavior in
the loop), and after unrolling the misdesigned -Warray-bounds warning complains
on the last iteration of the unrolled loop.

[Bug tree-optimization/63406] -Warray-bounds false positive with -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63406

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think we have similar dups elsewhere (warning for last unrolled iteration
which is never executed).

My previous naiive attempts dropped warnings from VRP2 and only warn from
VRP1 (but that regresses some cases in the testsuite).

[Bug tree-optimization/63406] -Warray-bounds false positive with -O3

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63406

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I'd say it would still be worthwhile, if it was just a matter of XFAILing a few
testcases, because the number of false positives is more important, if the
warning is too unreliable, people just will ignore it completely.
And people can combine it with -fsanitize=undefined, Fortran -fbounds-check
etc.