This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: C++ PATCH for c++/91416 - GC during late parsing collects live data

On 8/12/19 4:37 AM, Richard Biener wrote:
On Sun, Aug 11, 2019 at 7:12 PM Marek Polacek <> wrote:

This is a crash that points to a GC problem.  Consider this test:

   __attribute__ ((unused)) struct S {
     S() { }
   } s;

We're parsing a simple-declaration.  While parsing the decl specs, we parse the
attribute, which means creating a TREE_LIST using ggc_alloc_*.

A function body is a complete-class context so when parsing the
member-specification of this class-specifier, we parse the bodies of the
functions we'd queued in cp_parser_late_parsing_for_member.  This then leads to
this call chain:
cp_parser_function_definition_after_declarator -> expand_or_defer_fn ->
expand_or_defer_fn_1 -> maybe_clone_body -> expand_or_defer_fn ->
cgraph_node::finalize_function -> ggc_collect.

In this test, the ggc_collect call collects the TREE_LIST we had allocated, and
a crash duly ensues.  We can't avoid late parsing of members in this context,
so my fix is to bump function_depth, exactly following cp_parser_lambda_body.
Since we are performing late parsing, we know we have to be nested in a class.
(We still ggc_collect later, in c_parse_final_cleanups.)

So the struct S itself is properly referenced by a GC root?  If so why not
attach the attribute list to the tentative struct instead?  Or do we
fear we have
other non-rooted data live at the point we collect?  If so shouldn't we instead
bump function_depth when parsing a declaration in general?

It's already a significant issue for C++ that we only collect between function declarations, I'm concerned that this change will cause more memory problems.

Perhaps if we start parsing a class-specifier we could push the decl_specifiers onto a vec that is a GC root?


But here's the thing.  This goes back to ancient r104500, at least.  How has
this not broken before?  All you need to trigger it is to enable GC checking
and have a class with a ctor/member function, that has an attribute.

...and is defined in the declaration of something else? Does it happen without declaring the variable 's'?


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]