[Bug tree-optimization/67470] [5/6 Regression] ICE at -O3 on x86_64-linux-gnu in compute_live_loop_exits, at tree-ssa-loop-manip.c:235

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67470

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
g_40 = PHI <0B(9), &a(6)>
  invariant up to level 1, cost 21.

pretmp_42 = g_40 == 0B;
  invariant up to level 1, cost 22.

but we're only moving pretmp_42, not g_40.  The PHI is controlled by
if (f_16(D) != 0).  But when applying stmt movement we are faced with
g_40 = PHI <0B(9), &a(6)> turned into

<bb 11>:
# g_40 = PHI <0B(26), &a(21), &a(22), 0B(25)>

whoops.  That's because predicated store-motion already messed up the
CFG without ensuring we have forwarder blocks that at least preserve
the existing PHI node structure.

This is a latent issue.