https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78856
--- Comment #3 from Jeffrey A. Law <law at redhat dot com> --- Mostly to record my thoughts... Loop unrolling inserts a __builtin_unreachable as part of unrolling one of the loops. DOM3 looks at the first 10 or so blocks determines they all have a statically determined successor and collapses the mess down to: ;; basic block 2, loop depth 0, count 0, freq 333, maybe hot ;; prev block 0, next block 1, flags: (NEW, REACHABLE, VISITED) ;; pred: ENTRY [100.0%] (FALLTHRU,EXECUTABLE) f[0] = 1; f[1] = 1; f[2] = 1; e.3_3 = e; b.1_1 = b; c.2_2 = c; _31 = c.2_2 != 0; _15 = b.1_1 == 0 ? _31 : 0; iftmp.0_24 = (int) _15; _84 = 2; _4 = 9; _47 = 3; _49 = 9; _56 = 4; _57 = 9; _25 = 5; _86 = 9; __builtin_unreachable (); ;; succ: And of course removes everything else. If I work through the CFG prior to DOM3 by hand, I get the same result -- we always pass control to the block with the __builtin_unreachable. I really wonder if the problem is somewhere in loop unrolling.
