[PATCH v3] c++/modules: Fix dangling pointer with imported_temploid_friends

[Nathaniel Shead <nathanieloshead@gmail.com>]

On Thu, May 02, 2024 at 02:05:38PM -0400, Jason Merrill wrote:
> On 5/1/24 21:34, Nathaniel Shead wrote:
> > On Thu, May 02, 2024 at 12:15:44AM +1000, Nathaniel Shead wrote:
> > > On Wed, May 01, 2024 at 09:57:38AM -0400, Patrick Palka wrote:
> > > > 
> > > > On Wed, 1 May 2024, Nathaniel Shead wrote:
> > > > 
> > > > > Bootstrapped and regtested on x86_64-pc-linux-gnu, OK for trunk (and
> > > > > later 14.2)?  I don't think making it a GTY root is necessary but I felt
> > > > > perhaps better to be safe than sorry.
> > > > > 
> > > > > Potentially another approach would be to use DECL_UID instead like how
> > > > > entity_map does; would that be preferable?
> > > > > 
> > > > > -- >8 --
> > > > > 
> > > > > I got notified by Linaro CI and by checking testresults that there seems
> > > > > to be some occasional failures in tpl-friend-4_b.C on some architectures
> > > > > and standards modes since r15-59-gb5f6a56940e708.  I haven't been able
> > > > > to reproduce but looking at the backtrace I suspect the issue is that
> > > > > we're adding to the 'imported_temploid_friend' map a decl that is
> > > > > ultimately discarded, which then has its address reused by a later decl
> > > > > causing a failure in the assert in 'set_originating_module'.
> > > > > 
> > > > > This patch attempts to fix the issue in two ways: by ensuring that we
> > > > > only store the decl if we know it's a new decl (and hence won't be
> > > > > discarded), and by making the imported_temploid_friends map a GTY root
> > > > > so that even if the decl does get discarded later the address isn't
> > > > > reused.
> > > > > 
> > > > > gcc/cp/ChangeLog:
> > > > > 
> > > > > 	* module.cc (imported_temploid_friends): Mark GTY, and...
> > > > > 	(init_modules): ...allocate from GGC.
> > > > > 	(trees_in::decl_value): Only write to imported_temploid_friends
> > > > > 	for new decls.
> > > > > 
> > > > > Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
> > > > > ---
> > > > >   gcc/cp/module.cc | 7 ++++---
> > > > >   1 file changed, 4 insertions(+), 3 deletions(-)
> > > > > 
> > > > > diff --git a/gcc/cp/module.cc b/gcc/cp/module.cc
> > > > > index 5b8ff5bc483..37d38bb9654 100644
> > > > > --- a/gcc/cp/module.cc
> > > > > +++ b/gcc/cp/module.cc
> > > > > @@ -2731,7 +2731,7 @@ static keyed_map_t *keyed_table;
> > > > >      need to be attached to the same module as the temploid.  This maps
> > > > >      these decls to the temploid they are instantiated them, as there is
> > > > >      no other easy way to get this information.  */
> > > > > -static hash_map<tree, tree> *imported_temploid_friends;
> > > > > +static GTY(()) hash_map<tree, tree> *imported_temploid_friends;
> > > > >   /********************************************************************/
> > > > >   /* Tree streaming.   The tree streaming is very specific to the tree
> > > > > @@ -8327,7 +8327,8 @@ trees_in::decl_value ()
> > > > >     if (TREE_CODE (inner) == FUNCTION_DECL
> > > > >         || TREE_CODE (inner) == TYPE_DECL)
> > > > >       if (tree owner = tree_node ())
> > > > > -      imported_temploid_friends->put (decl, owner);
> > > > > +      if (is_new)
> > > > > +	imported_temploid_friends->put (decl, owner);
> > > > 
> > > > Hmm, I'm not seeing this code path getting reached for tpl-friend-4_b.C.
> > > > It seems we're instead adding to imported_temploid_friends from
> > > > propagate_defining_module, during tsubst_friend_function.
> > > > 
> > > > What seems to be happening is that we we first tsubst_friend_function
> > > > 'foo' from TPL<int>, and then we tsubst_friend_function 'foo' from DEF<int>,
> > > > which ends up calling duplicate_decls, which ggc_frees this 'foo'
> > > > redeclaration that is still present in the imported_temploid_friends map.
> > > > 
> > > > So I don't think marking imported_temploid_friends as a GC root would
> > > > help with this situation.  If we want to keep imported_temploid_friends
> > > > as a tree -> tree map, I think we just need to ensure that a decl
> > > > is removed from the map upon getting ggc_free'd from e.g.  duplicate_decls.
> 
> Could we instead move the call to propagate_defining_module down a few
> lines, after the pushdecl?
> 

Doing that for tsubst_friend_class seems to work OK with my current test
cases, but for tsubst_friend_function doing so causes ICEs in
'module_may_redeclare' within duplicate_decls because the function is
already marked as attached but the originating module information hasn't
been setup yet.

I suppose with tsubst_friend_class it works though because we can't ever
take the pushdecl branch if an existing type exists that we would call
duplicate_decls on.

> > > > But it seems simpler to use DECL_UID as the key instead, since those
> > > > never get reused even after the decl gets ggc_free'd IIUC.
> 
> It still means garbage entries in the hash_map, which is undesirable even if
> it doesn't cause the same kind of breakage.
> 
> Incidentally, decl_tree_map is preferable to hash_map<tree,tree> when the
> key is always a decl.
> 
> Jason
> 

Ah thanks, didn't know about decl_tree_map.  I feel that I prefer using
DECL_UIDs explicitly here though; it's also consistent with the existing
usage in entity_map_t, and it looks like decl_tree_map is still perhaps
vulnerable to the original issue here (since DECL_UID is only used for
hashing and not for equality, it looks like?).

Though that said, 'decl_constraints' in constraints.cc seems to be using
it fine (well, with a GTY marking) by using 'remove_constraints' within
duplicate_decls to clear this:  Here's a version of the patch that
implements in the same way, is something like this preferable?

Bootstrapped and regtested on x86_64-pc-linux-gnu.

-- >8 --

I got notified by Linaro CI and by checking testresults that there seems
to be some occasional failures in tpl-friend-4_b.C on some architectures
and standards modes since r15-59-gb5f6a56940e708.  I haven't been able
to reproduce but looking at the backtrace I suspect the issue is that
we're adding to the 'imported_temploid_friend' map a decl that is
ultimately discarded, which then has its address reused by a later decl
causing a failure in the assert in 'set_originating_module'.

This patch fixes the problem by ensuring 'imported_temploid_friends' is
correctly marked as a GTY root, and that 'duplicate_decls' properly
removes entries from the map for declarations that it frees.

gcc/cp/ChangeLog:

	* cp-tree.h (remove_defining_module): Declare.
	* decl.cc (duplicate_decls): Call remove_defining_module on
	to-be-freed newdecl.
	* module.cc (imported_temploid_friends): Mark as GTY root...
	(init_modules): ...and allocate from ggc.
	(trees_in::decl_value): Only track for declarations that won't
	be discarded.
	(remove_defining_module): New function.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
---
 gcc/cp/cp-tree.h |  1 +
 gcc/cp/decl.cc   |  4 ++++
 gcc/cp/module.cc | 19 ++++++++++++++++---
 3 files changed, 21 insertions(+), 3 deletions(-)

diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h
index 933504b4821..b7c09bf78c4 100644
--- a/gcc/cp/cp-tree.h
+++ b/gcc/cp/cp-tree.h
@@ -7420,6 +7420,7 @@ extern void set_instantiating_module (tree);
 extern void set_defining_module (tree);
 extern void maybe_key_decl (tree ctx, tree decl);
 extern void propagate_defining_module (tree decl, tree orig);
+extern void remove_defining_module (tree decl);
 
 extern void mangle_module (int m, bool include_partition);
 extern void mangle_module_fini ();
diff --git a/gcc/cp/decl.cc b/gcc/cp/decl.cc
index 378311c0f04..4af3ec48b42 100644
--- a/gcc/cp/decl.cc
+++ b/gcc/cp/decl.cc
@@ -3340,6 +3340,10 @@ duplicate_decls (tree newdecl, tree olddecl, bool hiding, bool was_hidden)
   if (flag_concepts)
     remove_constraints (newdecl);
 
+  /* And similarly for any module tracking data.  */
+  if (modules_p ())
+    remove_defining_module (newdecl);
+
   ggc_free (newdecl);
 
   return olddecl;
diff --git a/gcc/cp/module.cc b/gcc/cp/module.cc
index 44dc81eed3e..520dd710549 100644
--- a/gcc/cp/module.cc
+++ b/gcc/cp/module.cc
@@ -2731,7 +2731,7 @@ static keyed_map_t *keyed_table;
    need to be attached to the same module as the temploid.  This maps
    these decls to the temploid they are instantiated them, as there is
    no other easy way to get this information.  */
-static hash_map<tree, tree> *imported_temploid_friends;
+static GTY((cache)) decl_tree_cache_map *imported_temploid_friends;
 
 /********************************************************************/
 /* Tree streaming.   The tree streaming is very specific to the tree
@@ -8327,7 +8327,8 @@ trees_in::decl_value ()
   if (TREE_CODE (inner) == FUNCTION_DECL
       || TREE_CODE (inner) == TYPE_DECL)
     if (tree owner = tree_node ())
-      imported_temploid_friends->put (decl, owner);
+      if (is_new)
+	imported_temploid_friends->put (decl, owner);
 
   /* Regular typedefs will have a NULL TREE_TYPE at this point.  */
   unsigned tdef_flags = 0;
@@ -19336,6 +19337,18 @@ propagate_defining_module (tree decl, tree orig)
     }
 }
 
+/* DECL is being freed, clear data we don't need anymore.  */
+
+void
+remove_defining_module (tree decl)
+{
+  if (!modules_p ())
+    return;
+
+  if (imported_temploid_friends)
+    imported_temploid_friends->remove (decl);
+}
+
 /* Create the flat name string.  It is simplest to have it handy.  */
 
 void
@@ -20550,7 +20563,7 @@ init_modules (cpp_reader *reader)
       entity_map = new entity_map_t (EXPERIMENT (1, 400));
       vec_safe_reserve (entity_ary, EXPERIMENT (1, 400));
       imported_temploid_friends
-	= new hash_map<tree,tree> (EXPERIMENT (1, 400));
+	= decl_tree_cache_map::create_ggc (EXPERIMENT (1, 400));
     }
 
 #if CHECKING_P
-- 
2.43.2