c++: Fix constexpr dtors vs invisible ref [PR97388]

For arguments passed by invisible reference, in the IL until genericization
we have the source types on the callee side and while on the caller side
we already pass references to the actual argument slot in the caller, we
undo that in cxx_bind_parameters_in_call's
      if (TREE_ADDRESSABLE (type))
        /* Undo convert_for_arg_passing work here.  */
        x = convert_from_reference (x);
This works fine most of the time, except when the type also has constexpr
destructor; in that case the destructor is invoked in the caller and thus
the unsharing we do to make sure that the callee doesn't modify caller's
values is in that case undesirable, it prevents the changes done in the
callee propagating to the caller which should see them for the constexpr
dtor evaluation.

The following patch fixes that.  While it could be perhaps done for all
TREE_ADDRESSABLE types, I don't see the need to change the behavior
if there is no constexpr non-trivial dtor.

Jason: And we need to avoid memoizing the call, because a later equivalent
call also needs to modify its argument.  And we don't need to unshare
constructors when we aren't memoizing the call, because we already unshared
them when evaluating the TARGET_EXPR representing the copy-initialization of
the argument.

2020-10-20  Jakub Jelinek  <jakub@redhat.com>
	    Jason Merrill  <jason@redhat.com>

	PR c++/97388
	* constexpr.c (cxx_bind_parameters_in_call): Set non_constant_args
	if the parameter type has a non-trivial destructor.
	(cxx_eval_call_expression): Only unshare arguments if we're
	memoizing this evaluation.

	* g++.dg/cpp2a/constexpr-dtor5.C: New test.
	* g++.dg/cpp2a/constexpr-dtor6.C: New test.
	* g++.dg/cpp2a/constexpr-dtor7.C: New test.

1 files changed, 12 insertions, 7 deletions

diff --git a/gcc/cp/constexpr.c b/gcc/cp/constexpr.c
index 37c836a..c959b53 100644
--- a/gcc/cp/constexpr.c
+++ b/gcc/cp/constexpr.c
@@ -1602,6 +1602,11 @@ cxx_bind_parameters_in_call (const constexpr_ctx *ctx, tree t,
 	    arg = adjust_temp_type (type, arg);
 	  if (!TREE_CONSTANT (arg))
 	    *non_constant_args = true;
+	  else if (TYPE_HAS_NONTRIVIAL_DESTRUCTOR (type))
+	    /* The destructor needs to see any modifications the callee makes
+	       to the argument.  */
+	    *non_constant_args = true;
+
 	  /* For virtual calls, adjust the this argument, so that it is
 	     the object on which the method is called, rather than
 	     one of its bases.  */
@@ -2626,14 +2631,14 @@ cxx_eval_call_expression (const constexpr_ctx *ctx, tree t,
 		     problems with verify_gimple.  */
 		  arg = unshare_expr_without_location (arg);
 		  TREE_VEC_ELT (bound, i) = arg;
+
+		  /* And then unshare again so the callee doesn't change the
+		     argument values in the hash table. XXX Could we unshare
+		     lazily in cxx_eval_store_expression?  */
+		  arg = unshare_constructor (arg);
+		  if (TREE_CODE (arg) == CONSTRUCTOR)
+		    vec_safe_push (ctors, arg);
 		}
-	      /* Don't share a CONSTRUCTOR that might be changed.  This is not
-		 redundant with the unshare just above; we also don't want to
-		 change the argument values in the hash table.  XXX Could we
-		 unshare lazily in cxx_eval_store_expression?  */
-	      arg = unshare_constructor (arg);
-	      if (TREE_CODE (arg) == CONSTRUCTOR)
-		vec_safe_push (ctors, arg);
 	      ctx->global->values.put (remapped, arg);
 	      remapped = DECL_CHAIN (remapped);
 	    }