c++: Fix handling of no-linkage decls for modules

When testing the changes for PR c++/112631 we discovered that currently
we don't emit definitions of block-scope function declarations if
they're not used in the module interface TU, which causes issues if they
are used by importers.

This patch fixes the handling of no-linkage declarations for C++20. In
particular, a type declared in a function with vague linkage or declared
in a module CMI could potentially be accessible outside its defining TU,
and as such we can't assume that function declarations using that type
can never be defined in another TU.

A complication with handling this is that we're only strictly interested
in declarations with a module CMI, but when parsing the global module
fragment we don't yet know whether or not this module will have a CMI
until we reach the "export module" line (or not). Since this case is
IFNDR anyway (by [basic.def.odr] p11) we just tentatively assume while
parsing the GMF that this module will have a CMI; once we see (or don't
see) an 'export module' declaration we can commit to that knowledge for
future declarations.

gcc/cp/ChangeLog:

	* cp-tree.h (module_maybe_has_cmi_p): New function.
	* decl.cc (grokfndecl): Mark block-scope functions as public if
	they could be visible in other TUs.
	* decl2.cc (no_linkage_error): Don't error for declarations that
	could be defined in other TUs since C++20. Suppress duplicate
	errors from 'check_global_declaration'.
	* tree.cc (no_linkage_check): In relaxed mode, don't consider
	types in a module CMI to have no linkage.

gcc/testsuite/ChangeLog:

	* g++.dg/cpp2a/linkage-1.C: New test.
	* g++.dg/modules/block-decl-3.h: New test.
	* g++.dg/modules/block-decl-3_a.C: New test.
	* g++.dg/modules/block-decl-3_b.C: New test.
	* g++.dg/modules/block-decl-3_c.C: New test.
	* g++.dg/modules/linkage-1_a.C: New test.
	* g++.dg/modules/linkage-1_b.C: New test.
	* g++.dg/modules/linkage-1_c.C: New test.
	* g++.dg/modules/linkage-2.C: New test.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
Reviewed-by: Jason Merrill <jason@redhat.com>

1 files changed, 16 insertions, 5 deletions

diff --git a/gcc/cp/tree.cc b/gcc/cp/tree.cc
index e75be9a..f1a23ff 100644
--- a/gcc/cp/tree.cc
+++ b/gcc/cp/tree.cc
@@ -2971,7 +2971,8 @@ verify_stmt_tree (tree t)
 
 /* Check if the type T depends on a type with no linkage and if so,
    return it.  If RELAXED_P then do not consider a class type declared
-   within a vague-linkage function to have no linkage.  Remember:
+   within a vague-linkage function or in a module CMI to have no linkage,
+   since it can still be accessed within a different TU.  Remember:
    no-linkage is not the same as internal-linkage.  */
 
 tree
@@ -3012,7 +3013,15 @@ no_linkage_check (tree t, bool relaxed_p)
       /* Only treat unnamed types as having no linkage if they're at
 	 namespace scope.  This is core issue 966.  */
       if (TYPE_UNNAMED_P (t) && TYPE_NAMESPACE_SCOPE_P (t))
-	return t;
+	{
+	  if (relaxed_p
+	      && TREE_PUBLIC (CP_TYPE_CONTEXT (t))
+	      && module_maybe_has_cmi_p ())
+	    /* This type could possibly be accessed outside this TU.  */
+	    return NULL_TREE;
+	  else
+	    return t;
+	}
 
       for (r = CP_TYPE_CONTEXT (t); ; )
 	{
@@ -3023,10 +3032,12 @@ no_linkage_check (tree t, bool relaxed_p)
 	    return no_linkage_check (TYPE_CONTEXT (t), relaxed_p);
 	  else if (TREE_CODE (r) == FUNCTION_DECL)
 	    {
-	      if (!relaxed_p || !vague_linkage_p (r))
-		return t;
-	      else
+	      if (relaxed_p
+		  && (vague_linkage_p (r)
+		      || (TREE_PUBLIC (r) && module_maybe_has_cmi_p ())))
 		r = CP_DECL_CONTEXT (r);
+	      else
+		return t;
 	    }
 	  else
 	    break;