https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87926
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-11-21
CC| |hubicka at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed. I _think_ I've seen a duplicate at some point. Given
for (ix = 2 - 1; ix >= 0; ix--)
{
BITMAP_WORD word = elt->bits[ix];
if (word)
goto found_bit;
}
__builtin_unreachable ();
we compute that the loop exit test is never true (ix < 0) and thus elide it.
That causes us to unroll the loop based on the size of elt->bits[] given
that constrains the number of iterations now which is where we hit an
older complete unrolling issue that it doesn't mark all unreachable bits
properly unreachable. That is, remove_exits_and_undefined_stmts isn't
doing it's job properly. Hmm, because a bound is recorded for the IV
decrement (range info!) but not the load. Ah - trailing array. So
the issue is indeed a dup of some PR because the IV already goes out of
bound for the latch of the previous loop copy. Now go and find the dup...
We can avoid the warning by refactoring the code to
for (ix = 2 - 1; ix >= 1; ix--)
{
BITMAP_WORD word = elt->bits[ix];
if (word)
goto found_bit;
}
gcc_assert (elt->bits[ix]);
found_bit:
return ix;
which should be even faster eliding the read of elt->bits[0] for
--disable-checking.
