[Bug c++/79681] [6/7 Regression] ICE with constexpr and bitfield

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79681

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I was thinking about
--- fold-const.c.jj1    2017-02-17 18:29:24.000000000 +0100
+++ fold-const.c        2017-02-27 14:49:13.816203253 +0100
@@ -3862,6 +3862,39 @@ make_bit_field_ref (location_t loc, tree
 {
   tree result, bftype;

+  /* Attempt to use DECL_BIT_FIELD_REPRESENTATIVE as BIT_FIELD_REF
+     base if possible during FE folding.  */
+  if (in_gimple_form
+      && TREE_CODE (orig_inner) == COMPONENT_REF
+      && (DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (orig_inner, 1))
+         != NULL_TREE))
+    {
+      tree repr = DECL_BIT_FIELD_REPRESENTATIVE (TREE_OPERAND (orig_inner,
1));
+      tree ninner = build3 (COMPONENT_REF, TREE_TYPE (repr),
+                           TREE_OPERAND (orig_inner, 0), repr, NULL_TREE);
+      machine_mode nmode;
+      HOST_WIDE_INT nbitsize, nbitpos;
+      tree noffset;
+      int nunsignedp, nreversep, nvolatilep = 0;
+      tree base = get_inner_reference (ninner, &nbitsize, &nbitpos,
+                                      &noffset, &nmode, &nunsignedp,
+                                      &nreversep, &nvolatilep);
+      if (base == inner
+         && noffset == NULL_TREE
+         && nbitsize >= bitsize
+         && nbitpos <= bitpos
+         && bitpos + bitsize <= nbitpos + nbitsize
+         && !reversep
+         && !nreversep
+         && !nvolatilep)
+       {
+         inner = ninner;
+         bitpos -= nbitpos;
+       }
+      else
+       ggc_free (ninner);
+    }
+
   alias_set_type iset = get_alias_set (orig_inner);
   if (iset == 0 && get_alias_set (inner) != iset)
     inner = fold_build2 (MEM_REF, TREE_TYPE (inner),
which quite often just starts using the representative instead of using
BIT_FIELD_REF.  Unfortunately constexpr.c isn't prepared to handle that either.