[C++ Patch] PR 90915 [9/10 Regression] ICE in has_attribute, at c-family/c-attribs.c:4221

[Marek Polacek]

On Wed, Jan 29, 2020 at 10:06:51PM +0100, Paolo Carlini wrote:
> Hi,
> 
> On 29/01/20 19:00, Jason Merrill wrote:
> > On 1/29/20 4:31 AM, Paolo Carlini wrote:
> > > Hi,
> > > 
> > > in this regression we issue a diagnostic about an incomplete type
> > > (only a warning by default) and then we crash when we try to query
> > > has_attribute on a member of the type because in such cases the
> > > member remains an IDENTIFIER_NODE which of course doesn't have a
> > > TREE_TYPE neither a DECL_ATTRIBUTES... Simply recognizing
> > > IDENTIFIER_NODEs and returning false works fine, not sure if we want
> > > to do something more sophisticated. Tested x86_64-linux.
> > 
> > Why are we getting to has_attribute at all for a type-dependent argument?
> 
> Because the implementation of __builtin_has_attribute, largely shared with
> the C front-end, doesn't know about templates at all? :-/
> 
> Not sure it's the best time to complete it, but shouldn't be too difficult.

This ICEs even with a more reasonable test like

template<typename T>
void foo ()
{
  static_assert(!__builtin_has_attribute(T::a, aligned));
}

The problem here is that __builtin_has_attribute doesn't handle type-dependent
arguments at all.  To handle type-dependent arguments we'd have to introduce
a new template code, like STATIC_ASSERT or ADDRESSOF_EXPR (or a new generic
template code for built-ins?), but that's always a pain.

Or, meanwhile, we could just sorry.  Martin, what do you think?

--- a/gcc/cp/parser.c
+++ b/gcc/cp/parser.c
@@ -8682,7 +8682,12 @@ cp_parser_has_attribute_expression (cp_parser *parser)
   location_t atloc = cp_lexer_peek_token (parser->lexer)->location;
   if (tree attr = cp_parser_gnu_attribute_list (parser, /*exactly_one=*/true))
     {
-      if (oper != error_mark_node)
+      if (oper == error_mark_node)
+   /* Nothing.  */;
+      else if (type_dependent_expression_p (oper))
+   sorry_at (atloc, "%<__builtin_has_attribute%> with dependent argument "
+         "not supported yet");
+      else
    {
      /* Fold constant expressions used in attributes first.  */
      cp_check_const_attributes (attr);


Marek