diff --git a/autostatic.c b/autostatic.c
index 092aff2..21c1748 100755
--- a/autostatic.c
+++ b/autostatic.c
@@ -34,36 +34,73 @@
  *      59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
  */
 
+#include "pthread.h"
+#include "implement.h"
+
 #if defined(PTW32_STATIC_LIB)
 
 #if defined(__MINGW64__) || defined(__MINGW32__) || defined(_MSC_VER)
 
-#include "pthread.h"
-#include "implement.h"
+/* For an explanation of this code (at least the MSVC parts), refer to
+ *
+ * http://www.codeguru.com/cpp/misc/misc/threadsprocesses/article.php/c6945/
+ * ("Running Code Before and After Main")
+ *
+ * Compatibility with MSVC8 was cribbed from Boost:
+ *
+ * http://svn.boost.org/svn/boost/trunk/libs/thread/src/win32/tss_pe.cpp
+ *
+ * In addition to that, because we are in a static library, and the linker
+ * can't tell that the constructor/destructor functions are actually
+ * needed, we need a way to prevent the linker from optimizing away this
+ * module. The pthread_win32_autostatic_anchor() hack below (and in
+ * implement.h) does the job in a portable manner.
+ */
 
-static void on_process_init(void)
+static int on_process_init(void)
 {
     pthread_win32_process_attach_np ();
+    return 0;
 }
 
-static void on_process_exit(void)
+static int on_process_exit(void)
 {
     pthread_win32_thread_detach_np  ();
     pthread_win32_process_detach_np ();
+    return 0;
 }
 
 #if defined(__MINGW64__) || defined(__MINGW32__)
-# define attribute_section(a) __attribute__((section(a)))
+__attribute__((section(".ctors"), used)) static int (*gcc_ctor)(void) = on_process_init;
+__attribute__((section(".dtors"), used)) static int (*gcc_dtor)(void) = on_process_exit;
 #elif defined(_MSC_VER)
-# define attribute_section(a) __pragma(section(a,long,read)); __declspec(allocate(a))
+#  if _MSC_VER >= 1400	/* MSVC8 */
+#    pragma section(".CRT$XCU", long, read)
+#    pragma section(".CRT$XPU", long, read)
+__declspec(allocate(".CRT$XCU")) static int (*msc_ctor)(void) = on_process_init;
+__declspec(allocate(".CRT$XPU")) static int (*msc_dtor)(void) = on_process_exit;
+#  else
+#    pragma data_seg(".CRT$XCU")
+static int (*msc_ctor)(void) = on_process_init;
+#    pragma data_seg(".CRT$XPU")
+static int (*msc_dtor)(void) = on_process_exit;
+#    pragma data_seg()	/* reset data segment */
+#  endif
 #endif
 
-attribute_section(".ctors") void *gcc_ctor = on_process_init;
-attribute_section(".dtors") void *gcc_dtor = on_process_exit;
-
-attribute_section(".CRT$XCU") void *msc_ctor = on_process_init;
-attribute_section(".CRT$XPU") void *msc_dtor = on_process_exit;
-
 #endif /* defined(__MINGW64__) || defined(__MINGW32__) || defined(_MSC_VER) */
 
+/* This dummy function exists solely to be referenced by other modules
+ * (specifically, in implement.h), so that the linker can't optimize away
+ * this module. Don't call it.
+ */
+void pthread_win32_autostatic_anchor(void) { abort(); }
+
 #endif /* PTW32_STATIC_LIB */
+
+#if ! defined(PTW32_BUILD_INLINED)
+/*
+ * Avoid "translation unit is empty" warnings
+ */
+typedef int foo;
+#endif
diff --git a/implement.h b/implement.h
index 92f5740..7529ad2 100644
--- a/implement.h
+++ b/implement.h
@@ -109,6 +109,16 @@ typedef VOID (APIENTRY *PAPCFUNC)(DWORD dwParam);
 #  if defined(_MSC_VER) && _MSC_VER < 1300
      typedef long intptr_t;
 #  endif
+
+/*
+ * Don't allow the linker to optimize away autostatic.obj in static builds.
+ */
+#if defined(PTW32_STATIC_LIB)
+void pthread_win32_autostatic_anchor(void);
+#  if defined(__MINGW64__) || defined(__MINGW32__)
+__attribute__((unused, used))
+#  endif
+static void (*local_autostatic_anchor)(void) = pthread_win32_autostatic_anchor;
 #endif
 
 typedef enum