https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80453
--- Comment #9 from Richard Biener ---
So the issue is that vn_phi_eq depends on cond_stmts_equal_p which may change
if we value-number the controlling condition after the PHI itself which can
trivially happen with SCC based value-numbering as controlling conditions are
not part of the SCC.
Index: gcc/tree-ssa-sccvn.c
===
--- gcc/tree-ssa-sccvn.c(revision 246964)
+++ gcc/tree-ssa-sccvn.c(working copy)
@@ -2941,10 +2941,12 @@ cond_stmts_equal_p (gcond *cond1, gcond
else
return false;
+#if 0
lhs1 = vn_valueize (lhs1);
rhs1 = vn_valueize (rhs1);
lhs2 = vn_valueize (lhs2);
rhs2 = vn_valueize (rhs2);
+#endif
return ((expressions_equal_p (lhs1, lhs2)
&& expressions_equal_p (rhs1, rhs2))
|| (commutative_tree_code (code1)
would fix this (and pessimize PHI VN of course).