On Thu, 29 Jul 2021 11:41:39 -0400
Jason Merrill <jason@redhat.com> wrote:

> On 7/22/21 7:15 PM, Sergei Trofimovich wrote:
> > From: Sergei Trofimovich <siarheit@google.com>
> > 
> > r12-1804 ("cp: add support for per-location warning groups.") among other
> > things removed warning suppression from a few places including ptrmemfuncs.
> > 
> > Currently ptrmemfuncs don't have valid BINFO attached which causes ICEs
> > in access checks:
> > 
> >      crash_signal
> >          gcc/toplev.c:328
> >      perform_or_defer_access_check(tree_node*, tree_node*, tree_node*, int, access_failure_info*)
> >          gcc/cp/semantics.c:490
> >      finish_non_static_data_member(tree_node*, tree_node*, tree_node*)
> >          gcc/cp/semantics.c:2208
> >      ...
> > 
> > The change suppresses warnings again until we provide BINFOs for ptrmemfuncs.  
> 
> We don't need BINFOs for PMFs, we need to avoid paths that expect them.
> 
> It looks like the problem is with tsubst_copy_and_build calling 
> finish_non_static_data_member instead of build_ptrmemfunc_access_expr.

Sounds good. I'm not sure what would be the best way to match it. Here is
my attempt seems to survive all regtests:

--- a/gcc/cp/pt.c
+++ b/gcc/cp/pt.c
@@ -20530,7 +20530,13 @@ tsubst_copy_and_build (tree t,
        if (member == error_mark_node)
          RETURN (error_mark_node);

-       if (TREE_CODE (member) == FIELD_DECL)
+       if (object_type && TYPE_PTRMEMFUNC_P(object_type)
+           && TREE_CODE (member) == FIELD_DECL)
+         {
+           r = build_ptrmemfunc_access_expr (object, DECL_NAME(member));
+           RETURN (r);
+         }
+       else if (TREE_CODE (member) == FIELD_DECL)
          {
            r = finish_non_static_data_member (member, object, NULL_TREE);
            if (TREE_CODE (r) == COMPONENT_REF)

> > 	PR c++/101219
> > 
> > gcc/cp/ChangeLog:
> > 
> > 	* typeck.c (build_ptrmemfunc_access_expr): Suppress all warnings
> > 	to avoid ICE.
> > 
> > gcc/testsuite/ChangeLog:
> > 
> > 	* g++.dg/torture/pr101219.C: New test.  
> 
> This doesn't need to be in torture; it has nothing to do with optimization.

Aha, moved to gcc/testsuite/g++.dg/warn/pr101219.C.

--- /dev/null
+++ b/gcc/testsuite/g++.dg/warn/pr101219.C
@@ -0,0 +1,11 @@
+/* PR c++/101219 - ICE on use of uninitialized memfun pointer
+   { dg-do compile }
+   { dg-options "-Wall" } */
+
+struct S { void m(); };
+
+template <int> bool f() {
+  void (S::*mp)();
+
+  return &S::m == mp; // no warning emitted here (no instantiation)
+}

Another question: Is it expected that gcc generates no warnings here?
It's an uninstantiated function (-1 for warn), but from what I
understand it's guaranteed to generate comparison with uninitialized
data if it ever gets instantiated. Given that we used to ICE in
warning code gcc could possibly flag it? (+1 for warn)

Attached full patch as well. Full 'make check' shows no regressions on
x86_64-linux.

-- 

  Sergei