Finaliser call regression after updating to boehmgc 8.2 #334

Activity

1. tobil4sk commented on Sep 29, 2026

   

   tobil4sk

   on Sep 29, 2026

   MemberAuthor

   More actions

   This minimal project can reproduce the issue, though not consistently

   test.neko

   var runGc = $loader.loadprim("std@run_gc",1);
   var allocAbstract = $loader.loadprim("test@allocAbstract",0);
   {
   	var tmp = allocAbstract();
   	var tmp = allocAbstract();
   }
   runGc(true);

   test.c

   #include <neko.h>
   
   DEFINE_KIND(my_kind);
   
   void freeMyKind(value inAbstract)
   {
   	puts("freeMyKind");
   }
   
   value allocAbstract()
   {
   	puts("allocAbstract");
   
   	value abs = alloc_abstract(my_kind, (void *)11);
   	val_gc(abs, freeMyKind);
   	return abs;
   }
   DEFINE_PRIM(allocAbstract, 0);

   CMakeLists.txt

   cmake_minimum_required(VERSION 3.10)
   
   project(Test C)
   
   add_library(test.ndll MODULE test.c)
   
   find_library(NEKO_LIBRARIES neko HINTS C:\\HaxeToolkit\\neko)
   find_path(NEKO_INCLUDE_DIRS neko.h HINTS C:\\HaxeToolkit\\neko\\include)
   
   target_link_libraries(test.ndll ${NEKO_LIBRARIES})
   target_include_directories(test.ndll PRIVATE ${NEKO_INCLUDE_DIRS})
   
   set_target_properties(test.ndll
   	PROPERTIES
   	PREFIX ""
   	OUTPUT_NAME test
   	SUFFIX .ndll
   )
   
   install(
   	TARGETS test.ndll
   	DESTINATION ${CMAKE_SOURCE_DIR}
   )

   cmake -B build
   cmake --build build --target install
   nekoc test.neko
   neko test.n
2. tobil4sk commented on Oct 9, 2026

   

   tobil4sk

   on Oct 9, 2026

   MemberAuthor

   More actions

   One theory is that the register containing the abstract used to be overwritten by bdwgc code in the runGc call, but now some argument or variable got removed and register containing the abstract pointer remains intact, so the object is treated as live.
3. tobil4sk

   added a commit that references this issue on Oct 10, 2026

   Revert "Update BoehmGC to 8.2 (#307)"

   ...

   75338ce