[Bug tree-optimization/104466] Inlining functions with restrict parameters can inhibit lim (e.g. in 554.roms_r)

[rguenth at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104466

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Bah, what stupid cut&paste error ...



diff --git a/gcc/tree-ssa-alias.cc b/gcc/tree-ssa-alias.cc
index d434446a997..3e8d2455ba5 100644
--- a/gcc/tree-ssa-alias.cc
+++ b/gcc/tree-ssa-alias.cc
@@ -2420,12 +2420,12 @@ refs_may_alias_p_2 (ao_ref *ref1, ao_ref *ref2, bool
tbaa_p)
          rbase2 = TREE_OPERAND (rbase2, 0);
     }
   if (rbase1 && rbase2
-      && (TREE_CODE (base1) == MEM_REF || TREE_CODE (base1) == TARGET_MEM_REF)
-      && (TREE_CODE (base2) == MEM_REF || TREE_CODE (base2) == TARGET_MEM_REF)
+      && (TREE_CODE (rbase1) == MEM_REF || TREE_CODE (rbase1) ==
TARGET_MEM_REF)
+      && (TREE_CODE (rbase2) == MEM_REF || TREE_CODE (rbase2) ==
TARGET_MEM_REF)
       /* If the accesses are in the same restrict clique... */
-      && MR_DEPENDENCE_CLIQUE (base1) == MR_DEPENDENCE_CLIQUE (base2)
+      && MR_DEPENDENCE_CLIQUE (rbase1) == MR_DEPENDENCE_CLIQUE (rbase2)
       /* But based on different pointers they do not alias.  */
-      && MR_DEPENDENCE_BASE (base1) != MR_DEPENDENCE_BASE (base2))
+      && MR_DEPENDENCE_BASE (rbase1) != MR_DEPENDENCE_BASE (rbase2))
     return false;

   ind1_p = (TREE_CODE (base1) == MEM_REF