Fix pdumper-related GC bug

* src/alloc.c (mark_maybe_pointer): Also mark pointers
to pdumper objects, even when the pointers are tagged.
Add a FIXME saying why this isn’t enough.

1 files changed, 11 insertions, 0 deletions

diff --git a/src/alloc.c b/src/alloc.c
index 76d8c7ddd11..752eaec1352 100644
--- a/src/alloc.c
+++ b/src/alloc.c
@@ -4755,6 +4755,17 @@ mark_maybe_pointer (void *p)
      definitely _don't_ have an object.  */
   if (pdumper_object_p (p))
     {
+      /* FIXME: This code assumes that every reachable pdumper object
+         is addressed either by a pointer to the object start, or by
+         the same pointer with an LSB-style tag.  This assumption
+         fails if a pdumper object is reachable only via machine
+         addresses of non-initial object components.  Although such
+         addressing is rare in machine code generated by C compilers
+         from Emacs source code, it can occur in some cases.  To fix
+         this problem, the pdumper code should grok non-initial
+         addresses, as the non-pdumper code does.  */
+      uintptr_t mask = VALMASK;
+      p = (void *) ((uintptr_t) p & mask);
       /* Don't use pdumper_object_p_precise here! It doesn't check the
          tag bits. OBJ here might be complete garbage, so we need to
          verify both the pointer and the tag.  */