C++ PATCH for c++/85847, ICE with template_id_expr in new()

[Marek Polacek]

The diagnostic code in build_new{,_1} was using maybe_constant_value to fold
the array length, but that breaks while parsing a template, because we might
then leak template codes to the constexpr machinery.

Bootstrapped/regtested on x86_64-linux, ok for trunk/8?

2018-05-23  Marek Polacek  <pola...@redhat.com>

        PR c++/85847
        * init.c (build_new_1): Use fold_non_dependent_expr.
        (build_new): Likewise.

        * g++.dg/cpp0x/new3.C: New test.

diff --git gcc/cp/init.c gcc/cp/init.c
index b558742abf6..d96fec46f65 100644
--- gcc/cp/init.c
+++ gcc/cp/init.c
@@ -2860,7 +2860,7 @@ build_new_1 (vec<tree, va_gc> **placement, tree type, 
tree nelts,
   /* Lots of logic below. depends on whether we have a constant number of
      elements, so go ahead and fold it now.  */
   if (outer_nelts)
-    outer_nelts = maybe_constant_value (outer_nelts);
+    outer_nelts = fold_non_dependent_expr (outer_nelts);
 
   /* If our base type is an array, then make sure we know how many elements
      it has.  */
@@ -3639,7 +3639,7 @@ build_new (vec<tree, va_gc> **placement, tree type, tree 
nelts,
       /* Try to determine the constant value only for the purposes
         of the diagnostic below but continue to use the original
         value and handle const folding later.  */
-      const_tree cst_nelts = maybe_constant_value (nelts);
+      const_tree cst_nelts = fold_non_dependent_expr (nelts);
 
       /* The expression in a noptr-new-declarator is erroneous if it's of
         non-class type and its value before converting to std::size_t is
diff --git gcc/testsuite/g++.dg/cpp0x/new3.C gcc/testsuite/g++.dg/cpp0x/new3.C
index e69de29bb2d..c388acf552e 100644
--- gcc/testsuite/g++.dg/cpp0x/new3.C
+++ gcc/testsuite/g++.dg/cpp0x/new3.C
@@ -0,0 +1,11 @@
+// PR c++/85847
+// { dg-do compile { target c++11 } }
+
+template <class>
+int f(int b) { return b; }
+
+template <class>
+void g()
+{
+  auto a = new int[f<int>(2), 2];
+}