Hi!
cxx_eval_outermost_constant_expr clears TREE_CONSTANT on ADDR_EXPRs that
aren't considered by C++ constant expressions, but that breaks middle-end
which relies on TREE_CONSTANT being set on ADDR_EXPR where the address
is constant.
The following patch just special cases ADDR_EXPR not to clear it.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
As I wrote in the PR, another option would be to restore TREE_CONSTANT
during genericization if it was cleared earlier in the FE.
2018-01-24 Jakub Jelinek <ja...@redhat.com>
PR c++/83993
* constexpr.c (cxx_eval_outermost_constant_expr): Don't clear
TREE_CONSTANT on ADDR_EXPRs.
* g++.dg/init/pr83993-2.C: New test.
--- gcc/cp/constexpr.c.jj 2018-01-24 13:38:40.572913190 +0100
+++ gcc/cp/constexpr.c 2018-01-24 17:03:16.821440047 +0100
@@ -4832,7 +4832,7 @@ cxx_eval_outermost_constant_expr (tree t
if (non_constant_p && !allow_non_constant)
return error_mark_node;
- else if (non_constant_p && TREE_CONSTANT (r))
+ else if (non_constant_p && TREE_CONSTANT (r) && TREE_CODE (r) != ADDR_EXPR)
{
/* This isn't actually constant, so unset TREE_CONSTANT. */
if (EXPR_P (r))
--- gcc/testsuite/g++.dg/init/pr83993-2.C.jj 2018-01-24 17:04:18.823456178
+0100
+++ gcc/testsuite/g++.dg/init/pr83993-2.C 2018-01-24 17:04:39.593454636
+0100
@@ -0,0 +1,14 @@
+// PR c++/83993
+// { dg-do compile }
+// { dg-options "-w" }
+
+int a[5];
+extern int b[];
+int *const c = &a[6];
+int *const d = &b[1];
+
+int
+foo ()
+{
+ return c[-4] + d[-1];
+}
Jakub
