[Bug tree-optimization/57993] [4.9 Regression] ICE: verify_ssa failed (definition in block n does not dominate use in block m)

[wschmidt at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57993

--- Comment #6 from Bill Schmidt <wschmidt at gcc dot gnu.org> ---
Here's the patch I'm currently testing, which corrects the problem for this
test case.  We'll see how it does on regressions.

Index: gcc/gimple-ssa-strength-reduction.c
===================================================================
--- gcc/gimple-ssa-strength-reduction.c    (revision 201267)
+++ gcc/gimple-ssa-strength-reduction.c    (working copy)
@@ -2179,6 +2179,18 @@ phi_add_costs (gimple phi, slsr_cand_t c, int one_
   int cost = 0;
   slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi));

+  /* If we work our way back to a phi that isn't dominated by the hidden
+     basis, this isn't a candidate for replacement.  Indicate this by
+     returning an unreasonably high cost.  It's not easy to detect
+     these situations when determining the basis, so we defer the
+     decision until now.  */
+  basic_block phi_bb = gimple_bb (phi_cand->cand_stmt);
+  slsr_cand_t basis = lookup_cand (c->basis);
+  basic_block basis_bb = gimple_bb (basis->cand_stmt);
+
+  if (phi_bb == basis_bb || !dominated_by_p (CDI_DOMINATORS, phi_bb,
basis_bb))
+    return COST_INFINITE;
+
   for (i = 0; i < gimple_phi_num_args (phi); i++)
     {
       tree arg = gimple_phi_arg_def (phi, i);