https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86889
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
So the issue is that the __builtin_unreachable () elides the g >= 0 test,
effectively making the loop endless. If you disable EVRP and CUNROLLI then
VRP1 will do this transform.
<bb 6> :
# g_2 = PHI <1(2), g_7(5)>
if (g_2 != -1)
goto <bb 3>; [INV]
else
goto <bb 7>; [INV]
It goes like EVRP computing the range of g_2 to [0, 1] (correctly, based on
the __builtin_unreachable) and then CFG cleanup simplifying the conditional
to false, removing the __builtin_unreachable () call.
cunrolli then computes the loop to run twice and inserts that stray [-1] array
reference, exposing it's old latent issue of sometimes emitting stray
iterations as the reporter already analyzed (it can at most handle the _last_
iteration when marking blocks as never executed, but in cases where the IV
increment is at unfortunate places it has to look at the previous peeled
copy as well).
Maybe I can get to improve this. Don't hold your breath.
I think we have dups of this (but maybe closed because they no longer warn).
