X-Git-Url: http://git.sourceforge.jp/view?a=blobdiff_plain;f=boehm-gc%2Fwin32_threads.c;h=ffb34e0a8b1a9218290b85aa85d4a2b7ea8f064c;hb=5e3b370b4b55fa96c9ed649cd5d8f7bb51960112;hp=ff1d06625f86d3de60aa5f216066974e27607f0f;hpb=341281013a7334da0896c950d9c254e42692c479;p=pf3gnuchains%2Fgcc-fork.git

diff --git a/boehm-gc/win32_threads.c b/boehm-gc/win32_threads.c
index ff1d06625f8..ffb34e0a8b1 100644
--- a/boehm-gc/win32_threads.c
+++ b/boehm-gc/win32_threads.c
@@ -1,6 +1,7 @@
+#include "private/gc_priv.h"
+
 #if defined(GC_WIN32_THREADS) 
 
-#include "private/gc_priv.h"
 #include <windows.h>
 
 #ifdef CYGWIN32
@@ -10,40 +11,210 @@
 # undef pthread_create 
 # undef pthread_sigmask 
 # undef pthread_join 
+# undef pthread_detach
 # undef dlopen 
 
 # define DEBUG_CYGWIN_THREADS 0
 
-  GC_bool GC_thr_initialized = FALSE;
   void * GC_start_routine(void * arg);
   void GC_thread_exit_proc(void *arg);
 
 #endif
 
+/* The type of the first argument to InterlockedExchange.	*/
+/* Documented to be LONG volatile *, but at least gcc likes 	*/
+/* this better.							*/
+typedef LONG * IE_t;
+
 #ifndef MAX_THREADS
-# define MAX_THREADS 64
+# define MAX_THREADS 256
+    /* FIXME:							*/
+    /* Things may get quite slow for large numbers of threads,	*/
+    /* since we look them up with sequential search.		*/
 #endif
 
-struct thread_entry {
-  LONG in_use;
+GC_bool GC_thr_initialized = FALSE;
+
+DWORD GC_main_thread = 0;
+
+struct GC_thread_Rep {
+  LONG in_use; /* Updated without lock.	*/
+  			/* We assert that unused 	*/
+  			/* entries have invalid ids of	*/
+  			/* zero and zero stack fields.  */
   DWORD id;
   HANDLE handle;
-  void *stack;		/* The cold end of the stack.   */
+  ptr_t stack_base;	/* The cold end of the stack.   */
 			/* 0 ==> entry not valid.	*/
-			/* !in_use ==> stack == 0	*/
-  CONTEXT context;
+			/* !in_use ==> stack_base == 0	*/
   GC_bool suspended;
 
 # ifdef CYGWIN32
     void *status; /* hold exit value until join in case it's a pointer */
     pthread_t pthread_id;
+    short flags;		/* Protected by GC lock.	*/
+#	define FINISHED 1   	/* Thread has exited.	*/
+#	define DETACHED 2	/* Thread is intended to be detached.	*/
 # endif
-
 };
 
+typedef volatile struct GC_thread_Rep * GC_thread;
+
+/*
+ * We generally assume that volatile ==> memory ordering, at least among
+ * volatiles.
+ */
+
 volatile GC_bool GC_please_stop = FALSE;
 
-volatile struct thread_entry thread_table[MAX_THREADS];
+volatile struct GC_thread_Rep thread_table[MAX_THREADS];
+
+volatile LONG GC_max_thread_index = 0; /* Largest index in thread_table	*/
+				       /* that was ever used.		*/
+
+extern LONG WINAPI GC_write_fault_handler(struct _EXCEPTION_POINTERS *exc_info);
+
+/*
+ * This may be called from DllMain, and hence operates under unusual
+ * constraints.
+ */
+static GC_thread GC_new_thread(void) {
+  int i;
+  /* It appears to be unsafe to acquire a lock here, since this	*/
+  /* code is apparently not preeemptible on some systems.	*/
+  /* (This is based on complaints, not on Microsoft's official	*/
+  /* documentation, which says this should perform "only simple	*/
+  /* initialization tasks".)					*/
+  /* Hence we make do with nonblocking synchronization.		*/
+
+  /* The following should be a noop according to the win32	*/
+  /* documentation.  There is empirical evidence that it	*/
+  /* isn't.		- HB					*/
+# if defined(MPROTECT_VDB)
+   if (GC_incremental) SetUnhandledExceptionFilter(GC_write_fault_handler);
+# endif
+                /* cast away volatile qualifier */
+  for (i = 0; InterlockedExchange((IE_t)&thread_table[i].in_use,1) != 0; i++) {
+    /* Compare-and-swap would make this cleaner, but that's not 	*/
+    /* supported before Windows 98 and NT 4.0.  In Windows 2000,	*/
+    /* InterlockedExchange is supposed to be replaced by		*/
+    /* InterlockedExchangePointer, but that's not really what I		*/
+    /* want here.							*/
+    if (i == MAX_THREADS - 1)
+      ABORT("too many threads");
+  }
+  /* Update GC_max_thread_index if necessary.  The following is safe,	*/
+  /* and unlike CompareExchange-based solutions seems to work on all	*/
+  /* Windows95 and later platforms.					*/
+  /* Unfortunately, GC_max_thread_index may be temporarily out of 	*/
+  /* bounds, so readers have to compensate.				*/
+  while (i > GC_max_thread_index) {
+    InterlockedIncrement((IE_t)&GC_max_thread_index);
+  }
+  if (GC_max_thread_index >= MAX_THREADS) {
+    /* We overshot due to simultaneous increments.	*/
+    /* Setting it to MAX_THREADS-1 is always safe.	*/
+    GC_max_thread_index = MAX_THREADS - 1;
+  }
+  
+# ifdef CYGWIN32
+    thread_table[i].pthread_id = pthread_self();
+# endif
+  if (!DuplicateHandle(GetCurrentProcess(),
+	               GetCurrentThread(),
+		       GetCurrentProcess(),
+		       (HANDLE*)&thread_table[i].handle,
+		       0,
+		       0,
+		       DUPLICATE_SAME_ACCESS)) {
+	DWORD last_error = GetLastError();
+	GC_printf1("Last error code: %lx\n", last_error);
+	ABORT("DuplicateHandle failed");
+  }
+  thread_table[i].stack_base = GC_get_stack_base();
+  /* Up until this point, GC_push_all_stacks considers this thread	*/
+  /* invalid.								*/
+  if (thread_table[i].stack_base == NULL) 
+    ABORT("Failed to find stack base in GC_new_thread");
+  /* Up until this point, this entry is viewed as reserved but invalid	*/
+  /* by GC_delete_thread.						*/
+  thread_table[i].id = GetCurrentThreadId();
+  /* If this thread is being created while we are trying to stop	*/
+  /* the world, wait here.  Hopefully this can't happen on any	*/
+  /* systems that don't allow us to block here.			*/
+  while (GC_please_stop) Sleep(20);
+  return thread_table + i;
+}
+
+/*
+ * GC_max_thread_index may temporarily be larger than MAX_THREADS.
+ * To avoid subscript errors, we check on access.
+ */
+#ifdef __GNUC__
+__inline__
+#endif
+LONG GC_get_max_thread_index()
+{
+  LONG my_max = GC_max_thread_index;
+
+  if (my_max >= MAX_THREADS) return MAX_THREADS-1;
+  return my_max;
+}
+
+/* This is intended to be lock-free, though that			*/
+/* assumes that the CloseHandle becomes visible before the 		*/
+/* in_use assignment.							*/
+static void GC_delete_gc_thread(GC_thread thr)
+{
+    CloseHandle(thr->handle);
+      /* cast away volatile qualifier */
+    thr->stack_base = 0;
+    thr->id = 0;
+#   ifdef CYGWIN32
+      thr->pthread_id = 0;
+#   endif /* CYGWIN32 */
+    thr->in_use = FALSE;
+}
+
+static void GC_delete_thread(DWORD thread_id) {
+  int i;
+  LONG my_max = GC_get_max_thread_index();
+
+  for (i = 0;
+       i <= my_max &&
+       (!thread_table[i].in_use || thread_table[i].id != thread_id);
+       /* Must still be in_use, since nobody else can store our thread_id. */
+       i++) {}
+  if (i > my_max) {
+    WARN("Removing nonexistent thread %ld\n", (GC_word)thread_id);
+  } else {
+    GC_delete_gc_thread(thread_table+i);
+  }
+}
+
+
+#ifdef CYGWIN32
+
+/* Return a GC_thread corresponding to a given pthread_t.	*/
+/* Returns 0 if it's not there.					*/
+/* We assume that this is only called for pthread ids that	*/
+/* have not yet terminated or are still joinable.		*/
+static GC_thread GC_lookup_thread(pthread_t id)
+{
+  int i;
+  LONG my_max = GC_get_max_thread_index();
+
+  for (i = 0;
+       i <= my_max &&
+       (!thread_table[i].in_use || thread_table[i].pthread_id != id
+	|| !thread_table[i].in_use);
+       /* Must still be in_use, since nobody else can store our thread_id. */
+       i++);
+  if (i > my_max) return 0;
+  return thread_table + i;
+}
+
+#endif /* CYGWIN32 */
 
 void GC_push_thread_structures GC_PROTO((void))
 {
@@ -52,24 +223,32 @@ void GC_push_thread_structures GC_PROTO((void))
     /* no private structures we need to preserve.			*/
 # ifdef CYGWIN32
   { int i; /* pthreads may keep a pointer in the thread exit value */
-    for (i = 0; i < MAX_THREADS; i++)
-      if (thread_table[i].in_use) GC_push_all((ptr_t)&(thread_table[i].status),(ptr_t)(&(thread_table[i].status)+1));
+    LONG my_max = GC_get_max_thread_index();
+
+    for (i = 0; i <= my_max; i++)
+      if (thread_table[i].in_use)
+	GC_push_all((ptr_t)&(thread_table[i].status),
+                    (ptr_t)(&(thread_table[i].status)+1));
   }
 # endif
 }
 
+/* Defined in misc.c */
+extern CRITICAL_SECTION GC_write_cs;
+
 void GC_stop_world()
 {
   DWORD thread_id = GetCurrentThreadId();
   int i;
 
-#ifdef CYGWIN32
   if (!GC_thr_initialized) ABORT("GC_stop_world() called before GC_thr_init()");
-#endif
 
   GC_please_stop = TRUE;
-  for (i = 0; i < MAX_THREADS; i++)
-    if (thread_table[i].stack != 0
+# ifndef CYGWIN32
+    EnterCriticalSection(&GC_write_cs);
+# endif /* !CYGWIN32 */
+  for (i = 0; i <= GC_get_max_thread_index(); i++)
+    if (thread_table[i].stack_base != 0
 	&& thread_table[i].id != thread_id) {
 #     ifdef MSWINCE
         /* SuspendThread will fail if thread is running kernel code */
@@ -84,13 +263,12 @@ void GC_stop_world()
 	DWORD exitCode; 
 	if (GetExitCodeThread(thread_table[i].handle,&exitCode) &&
             exitCode != STILL_ACTIVE) {
-          thread_table[i].stack = 0; /* prevent stack from being pushed */
+          thread_table[i].stack_base = 0; /* prevent stack from being pushed */
 #         ifndef CYGWIN32
             /* this breaks pthread_join on Cygwin, which is guaranteed to  */
 	    /* only see user pthreads 					   */
 	    thread_table[i].in_use = FALSE;
 	    CloseHandle(thread_table[i].handle);
-	    BZERO((void *)(&thread_table[i].context), sizeof(CONTEXT));
 #         endif
 	  continue;
 	}
@@ -99,14 +277,19 @@ void GC_stop_world()
 #     endif
       thread_table[i].suspended = TRUE;
     }
+# ifndef CYGWIN32
+    LeaveCriticalSection(&GC_write_cs);
+# endif /* !CYGWIN32 */
 }
 
 void GC_start_world()
 {
   DWORD thread_id = GetCurrentThreadId();
   int i;
-  for (i = 0; i < MAX_THREADS; i++)
-    if (thread_table[i].stack != 0 && thread_table[i].suspended
+  LONG my_max = GC_get_max_thread_index();
+
+  for (i = 0; i <= my_max; i++)
+    if (thread_table[i].stack_base != 0 && thread_table[i].suspended
 	&& thread_table[i].id != thread_id) {
       if (ResumeThread(thread_table[i].handle) == (DWORD)-1)
 	ABORT("ResumeThread failed");
@@ -122,9 +305,11 @@ ptr_t GC_current_stackbottom()
 {
   DWORD thread_id = GetCurrentThreadId();
   int i;
-  for (i = 0; i < MAX_THREADS; i++)
-    if (thread_table[i].stack && thread_table[i].id == thread_id)
-      return thread_table[i].stack;
+  LONG my_max = GC_get_max_thread_index();
+
+  for (i = 0; i <= my_max; i++)
+    if (thread_table[i].stack_base && thread_table[i].id == thread_id)
+      return thread_table[i].stack_base;
   ABORT("no thread table entry for current thread");
 }
 # ifdef _MSC_VER
@@ -135,10 +320,10 @@ ptr_t GC_current_stackbottom()
     /* The VirtualQuery calls below won't work properly on WinCE, but	*/
     /* since each stack is restricted to an aligned 64K region of	*/
     /* virtual memory we can just take the next lowest multiple of 64K.	*/
-#   define GC_get_lo_stack_addr(s) \
+#   define GC_get_stack_min(s) \
         ((ptr_t)(((DWORD)(s) - 1) & 0xFFFF0000))
 # else
-    static ptr_t GC_get_lo_stack_addr(ptr_t s)
+    static ptr_t GC_get_stack_min(ptr_t s)
     {
 	ptr_t bottom;
 	MEMORY_BASIC_INFORMATION info;
@@ -155,197 +340,76 @@ ptr_t GC_current_stackbottom()
 void GC_push_all_stacks()
 {
   DWORD thread_id = GetCurrentThreadId();
+  GC_bool found_me = FALSE;
   int i;
-  for (i = 0; i < MAX_THREADS; i++)
-    if (thread_table[i].stack) {
-      ptr_t bottom = GC_get_lo_stack_addr(thread_table[i].stack);
-      if (thread_table[i].id == thread_id)
-	GC_push_all_stack((ptr_t)&i, thread_table[i].stack);
-      else {
-	thread_table[i].context.ContextFlags
-			= (CONTEXT_INTEGER|CONTEXT_CONTROL);
-	if (!GetThreadContext(thread_table[i].handle,
-				/* cast away volatile qualifier */
-	    		        (LPCONTEXT)&thread_table[i].context))
+  int dummy;
+  ptr_t sp, stack_min;
+  GC_thread thread;
+  LONG my_max = GC_get_max_thread_index();
+  
+  for (i = 0; i <= my_max; i++) {
+    thread = thread_table + i;
+    if (thread -> in_use && thread -> stack_base) {
+      if (thread -> id == thread_id) {
+	sp = (ptr_t) &dummy;
+	found_me = TRUE;
+      } else {
+        CONTEXT context;
+        context.ContextFlags = CONTEXT_INTEGER|CONTEXT_CONTROL;
+        if (!GetThreadContext(thread_table[i].handle, &context))
 	  ABORT("GetThreadContext failed");
-#	ifdef I386
-	  GC_push_one ((word) thread_table[i].context.Edi);
-    	  GC_push_one ((word) thread_table[i].context.Esi);
-    	  GC_push_one ((word) thread_table[i].context.Ebp);
-    	  GC_push_one ((word) thread_table[i].context.Ebx);
-    	  GC_push_one ((word) thread_table[i].context.Edx);
-    	  GC_push_one ((word) thread_table[i].context.Ecx);
-    	  GC_push_one ((word) thread_table[i].context.Eax);
-	  if (thread_table[i].context.Esp >= (DWORD)thread_table[i].stack
-	      || thread_table[i].context.Esp < (DWORD)bottom) {
-	      WARN("Thread stack pointer 0x%lx out of range, pushing everything",
-		   thread_table[i].context.Esp);
-	      GC_push_all_stack((char *) bottom, thread_table[i].stack);
-	  } else {
-	      GC_push_all_stack((char *) thread_table[i].context.Esp,
-			        thread_table[i].stack);
-	  }
-#       else
-#       ifdef ARM32
-	  if (thread_table[i].context.Sp >= (DWORD)thread_table[i].stack
-	      || thread_table[i].context.Sp < (DWORD)bottom)
-	      ABORT("Thread stack pointer out of range");
-	  GC_push_one ((word) thread_table[i].context.R0);
-	  GC_push_one ((word) thread_table[i].context.R1);
-	  GC_push_one ((word) thread_table[i].context.R2);
-	  GC_push_one ((word) thread_table[i].context.R3);
-	  GC_push_one ((word) thread_table[i].context.R4);
-	  GC_push_one ((word) thread_table[i].context.R5);
-	  GC_push_one ((word) thread_table[i].context.R6);
-	  GC_push_one ((word) thread_table[i].context.R7);
-	  GC_push_one ((word) thread_table[i].context.R8);
-	  GC_push_one ((word) thread_table[i].context.R9);
-	  GC_push_one ((word) thread_table[i].context.R10);
-	  GC_push_one ((word) thread_table[i].context.R11);
-	  GC_push_one ((word) thread_table[i].context.R12);
-	  GC_push_all_stack((char *) thread_table[i].context.Sp,
-			    thread_table[i].stack);
-#       else
-#	ifdef SHx
-	  if (thread_table[i].context.R15 >= (DWORD)thread_table[i].stack
-	      || thread_table[i].context.R15 < (DWORD)bottom)
-	      ABORT("Thread stack pointer out of range");
-	  GC_push_one ((word) thread_table[i].context.R0);
-	  GC_push_one ((word) thread_table[i].context.R1);
-	  GC_push_one ((word) thread_table[i].context.R2);
-	  GC_push_one ((word) thread_table[i].context.R3);
-	  GC_push_one ((word) thread_table[i].context.R4);
-	  GC_push_one ((word) thread_table[i].context.R5);
-	  GC_push_one ((word) thread_table[i].context.R6);
-	  GC_push_one ((word) thread_table[i].context.R7);
-	  GC_push_one ((word) thread_table[i].context.R8);
-	  GC_push_one ((word) thread_table[i].context.R9);
-	  GC_push_one ((word) thread_table[i].context.R10);
-	  GC_push_one ((word) thread_table[i].context.R11);
-	  GC_push_one ((word) thread_table[i].context.R12);
-	  GC_push_one ((word) thread_table[i].context.R13);
-	  GC_push_one ((word) thread_table[i].context.R14);
-	  GC_push_all_stack((char *) thread_table[i].context.R15,
-			    thread_table[i].stack);
+
+        /* Push all registers that might point into the heap.  Frame	*/
+        /* pointer registers are included in case client code was	*/
+        /* compiled with the 'omit frame pointer' optimisation.		*/
+#       define PUSH1(reg) GC_push_one((word)context.reg)
+#       define PUSH2(r1,r2) PUSH1(r1), PUSH1(r2)
+#       define PUSH4(r1,r2,r3,r4) PUSH2(r1,r2), PUSH2(r3,r4)
+#       if defined(I386)
+          PUSH4(Edi,Esi,Ebx,Edx), PUSH2(Ecx,Eax), PUSH1(Ebp);
+	  sp = (ptr_t)context.Esp;
+#       elif defined(ARM32)
+	  PUSH4(R0,R1,R2,R3),PUSH4(R4,R5,R6,R7),PUSH4(R8,R9,R10,R11),PUSH1(R12);
+	  sp = (ptr_t)context.Sp;
+#       elif defined(SHx)
+	  PUSH4(R0,R1,R2,R3), PUSH4(R4,R5,R6,R7), PUSH4(R8,R9,R10,R11);
+	  PUSH2(R12,R13), PUSH1(R14);
+	  sp = (ptr_t)context.R15;
+#       elif defined(MIPS)
+	  PUSH4(IntAt,IntV0,IntV1,IntA0), PUSH4(IntA1,IntA2,IntA3,IntT0);
+	  PUSH4(IntT1,IntT2,IntT3,IntT4), PUSH4(IntT5,IntT6,IntT7,IntS0);
+	  PUSH4(IntS1,IntS2,IntS3,IntS4), PUSH4(IntS5,IntS6,IntS7,IntT8);
+	  PUSH4(IntT9,IntK0,IntK1,IntS8);
+	  sp = (ptr_t)context.IntSp;
+#       elif defined(PPC)
+	  PUSH4(Gpr0, Gpr3, Gpr4, Gpr5),  PUSH4(Gpr6, Gpr7, Gpr8, Gpr9);
+	  PUSH4(Gpr10,Gpr11,Gpr12,Gpr14), PUSH4(Gpr15,Gpr16,Gpr17,Gpr18);
+	  PUSH4(Gpr19,Gpr20,Gpr21,Gpr22), PUSH4(Gpr23,Gpr24,Gpr25,Gpr26);
+	  PUSH4(Gpr27,Gpr28,Gpr29,Gpr30), PUSH1(Gpr31);
+	  sp = (ptr_t)context.Gpr1;
+#       elif defined(ALPHA)
+	  PUSH4(IntV0,IntT0,IntT1,IntT2), PUSH4(IntT3,IntT4,IntT5,IntT6);
+	  PUSH4(IntT7,IntS0,IntS1,IntS2), PUSH4(IntS3,IntS4,IntS5,IntFp);
+	  PUSH4(IntA0,IntA1,IntA2,IntA3), PUSH4(IntA4,IntA5,IntT8,IntT9);
+	  PUSH4(IntT10,IntT11,IntT12,IntAt);
+	  sp = (ptr_t)context.IntSp;
 #       else
-#	ifdef MIPS
-	  if (thread_table[i].context.IntSp >= (DWORD)thread_table[i].stack
-	      || thread_table[i].context.IntSp < (DWORD)bottom)
-	      ABORT("Thread stack pointer out of range");
-	  GC_push_one ((word) thread_table[i].context.IntAt);
-	  GC_push_one ((word) thread_table[i].context.IntV0);
-	  GC_push_one ((word) thread_table[i].context.IntV1);
-	  GC_push_one ((word) thread_table[i].context.IntA0);
-	  GC_push_one ((word) thread_table[i].context.IntA1);
-	  GC_push_one ((word) thread_table[i].context.IntA2);
-	  GC_push_one ((word) thread_table[i].context.IntA3);
-	  GC_push_one ((word) thread_table[i].context.IntT0);
-	  GC_push_one ((word) thread_table[i].context.IntT1);
-	  GC_push_one ((word) thread_table[i].context.IntT2);
-	  GC_push_one ((word) thread_table[i].context.IntT3);
-	  GC_push_one ((word) thread_table[i].context.IntT4);
-	  GC_push_one ((word) thread_table[i].context.IntT5);
-	  GC_push_one ((word) thread_table[i].context.IntT6);
-	  GC_push_one ((word) thread_table[i].context.IntT7);
-	  GC_push_one ((word) thread_table[i].context.IntS0);
-	  GC_push_one ((word) thread_table[i].context.IntS1);
-	  GC_push_one ((word) thread_table[i].context.IntS2);
-	  GC_push_one ((word) thread_table[i].context.IntS3);
-	  GC_push_one ((word) thread_table[i].context.IntS4);
-	  GC_push_one ((word) thread_table[i].context.IntS5);
-	  GC_push_one ((word) thread_table[i].context.IntS6);
-	  GC_push_one ((word) thread_table[i].context.IntS7);
-	  GC_push_one ((word) thread_table[i].context.IntT8);
-	  GC_push_one ((word) thread_table[i].context.IntT9);
-	  GC_push_one ((word) thread_table[i].context.IntK0);
-	  GC_push_one ((word) thread_table[i].context.IntK1);
-	  GC_push_one ((word) thread_table[i].context.IntS8);
-	  GC_push_all_stack((char *) thread_table[i].context.IntSp,
-			    thread_table[i].stack);
-#	else
-#	ifdef PPC
-	  if (thread_table[i].context.Gpr1 >= (DWORD)thread_table[i].stack
-	      || thread_table[i].context.Gpr1 < (DWORD)bottom)
-	      ABORT("Thread stack pointer out of range");
-	  GC_push_one ((word) thread_table[i].context.Gpr0);
-	  /* Gpr1 is stack pointer */
-	  /* Gpr2 is global pointer */
-	  GC_push_one ((word) thread_table[i].context.Gpr3);
-	  GC_push_one ((word) thread_table[i].context.Gpr4);
-	  GC_push_one ((word) thread_table[i].context.Gpr5);
-	  GC_push_one ((word) thread_table[i].context.Gpr6);
-	  GC_push_one ((word) thread_table[i].context.Gpr7);
-	  GC_push_one ((word) thread_table[i].context.Gpr8);
-	  GC_push_one ((word) thread_table[i].context.Gpr9);
-	  GC_push_one ((word) thread_table[i].context.Gpr10);
-	  GC_push_one ((word) thread_table[i].context.Gpr11);
-	  GC_push_one ((word) thread_table[i].context.Gpr12);
-	  /* Gpr13 is reserved for the kernel */
-	  GC_push_one ((word) thread_table[i].context.Gpr14);
-	  GC_push_one ((word) thread_table[i].context.Gpr15);
-	  GC_push_one ((word) thread_table[i].context.Gpr16);
-	  GC_push_one ((word) thread_table[i].context.Gpr17);
-	  GC_push_one ((word) thread_table[i].context.Gpr18);
-	  GC_push_one ((word) thread_table[i].context.Gpr19);
-	  GC_push_one ((word) thread_table[i].context.Gpr20);
-	  GC_push_one ((word) thread_table[i].context.Gpr21);
-	  GC_push_one ((word) thread_table[i].context.Gpr22);
-	  GC_push_one ((word) thread_table[i].context.Gpr23);
-	  GC_push_one ((word) thread_table[i].context.Gpr24);
-	  GC_push_one ((word) thread_table[i].context.Gpr25);
-	  GC_push_one ((word) thread_table[i].context.Gpr26);
-	  GC_push_one ((word) thread_table[i].context.Gpr27);
-	  GC_push_one ((word) thread_table[i].context.Gpr28);
-	  GC_push_one ((word) thread_table[i].context.Gpr29);
-	  GC_push_one ((word) thread_table[i].context.Gpr30);
-	  GC_push_one ((word) thread_table[i].context.Gpr31);
-	  GC_push_all_stack((char *) thread_table[i].context.Gpr1,
-			    thread_table[i].stack);
-#	else
-#	ifdef ALPHA
-	  if (thread_table[i].context.IntSp >= (DWORD)thread_table[i].stack
-	      || thread_table[i].context.IntSp < (DWORD)bottom)
-	      ABORT("Thread stack pointer out of range");
-	  GC_push_one ((word) thread_table[i].context.IntV0);
-	  GC_push_one ((word) thread_table[i].context.IntT0);
-	  GC_push_one ((word) thread_table[i].context.IntT1);
-	  GC_push_one ((word) thread_table[i].context.IntT2);
-	  GC_push_one ((word) thread_table[i].context.IntT3);
-	  GC_push_one ((word) thread_table[i].context.IntT4);
-	  GC_push_one ((word) thread_table[i].context.IntT5);
-	  GC_push_one ((word) thread_table[i].context.IntT6);
-	  GC_push_one ((word) thread_table[i].context.IntT7);
-	  GC_push_one ((word) thread_table[i].context.IntS0);
-	  GC_push_one ((word) thread_table[i].context.IntS1);
-	  GC_push_one ((word) thread_table[i].context.IntS2);
-	  GC_push_one ((word) thread_table[i].context.IntS3);
-	  GC_push_one ((word) thread_table[i].context.IntS4);
-	  GC_push_one ((word) thread_table[i].context.IntS5);
-	  GC_push_one ((word) thread_table[i].context.IntFp);
-	  GC_push_one ((word) thread_table[i].context.IntA0);
-	  GC_push_one ((word) thread_table[i].context.IntA1);
-	  GC_push_one ((word) thread_table[i].context.IntA2);
-	  GC_push_one ((word) thread_table[i].context.IntA3);
-	  GC_push_one ((word) thread_table[i].context.IntA4);
-	  GC_push_one ((word) thread_table[i].context.IntA5);
-	  GC_push_one ((word) thread_table[i].context.IntT8);
-	  GC_push_one ((word) thread_table[i].context.IntT9);
-	  GC_push_one ((word) thread_table[i].context.IntT10);
-	  GC_push_one ((word) thread_table[i].context.IntT11);
-	  GC_push_one ((word) thread_table[i].context.IntT12);
-	  GC_push_one ((word) thread_table[i].context.IntAt);
-	  GC_push_all_stack((char *) thread_table[i].context.IntSp,
-			    thread_table[i].stack);
-#	else
-	      --> architecture not supported
-#	endif /* !ALPHA */
-#	endif /* !PPC */
-#	endif /* !MIPS */
-#	endif /* !SHx */
-#	endif /* !ARM32 */
-#	endif /* !I386 */
+#         error "architecture is not supported"
+#       endif
+      }
+
+      stack_min = GC_get_stack_min(thread->stack_base);
+
+      if (sp >= stack_min && sp < thread->stack_base)
+        GC_push_all_stack(sp, thread->stack_base);
+      else {
+        WARN("Thread stack pointer 0x%lx out of range, pushing everything\n",
+	     (unsigned long)sp);
+        GC_push_all_stack(stack_min, thread->stack_base);
       }
     }
+  }
+  if (!found_me) ABORT("Collecting from unknown thread.");
 }
 
 void GC_get_next_stack(char *start, char **lo, char **hi)
@@ -353,9 +417,10 @@ void GC_get_next_stack(char *start, char **lo, char **hi)
     int i;
 #   define ADDR_LIMIT (char *)(-1L)
     char * current_min = ADDR_LIMIT;
-
-    for (i = 0; i < MAX_THREADS; i++) {
-    	char * s = (char *)thread_table[i].stack;
+    LONG my_max = GC_get_max_thread_index();
+  
+    for (i = 0; i <= my_max; i++) {
+    	char * s = (char *)thread_table[i].stack_base;
 
 	if (0 != s && s > start && s < current_min) {
 	    current_min = s;
@@ -366,7 +431,7 @@ void GC_get_next_stack(char *start, char **lo, char **hi)
     	*lo = ADDR_LIMIT;
 	return;
     }
-    *lo = GC_get_lo_stack_addr(current_min);
+    *lo = GC_get_stack_min(current_min);
     if (*lo < start) *lo = start;
 }
 
@@ -376,7 +441,7 @@ void GC_get_next_stack(char *start, char **lo, char **hi)
 
 /* We register threads from DllMain */
 
-GC_API HANDLE GC_CreateThread(
+GC_API HANDLE WINAPI GC_CreateThread(
     LPSECURITY_ATTRIBUTES lpThreadAttributes, 
     DWORD dwStackSize, LPTHREAD_START_ROUTINE lpStartAddress, 
     LPVOID lpParameter, DWORD dwCreationFlags, LPDWORD lpThreadId )
@@ -391,94 +456,39 @@ GC_API HANDLE GC_CreateThread(
 /* must properly intercept thread creation.			*/
 
 typedef struct {
-    HANDLE child_ready_h, parent_ready_h;
-    volatile struct thread_entry * entry;
     LPTHREAD_START_ROUTINE start;
     LPVOID param;
 } thread_args;
 
-DWORD WINAPI thread_start(LPVOID arg);
+static DWORD WINAPI thread_start(LPVOID arg);
 
-HANDLE WINAPI GC_CreateThread(
+GC_API HANDLE WINAPI GC_CreateThread(
     LPSECURITY_ATTRIBUTES lpThreadAttributes, 
     DWORD dwStackSize, LPTHREAD_START_ROUTINE lpStartAddress, 
     LPVOID lpParameter, DWORD dwCreationFlags, LPDWORD lpThreadId )
 {
     HANDLE thread_h = NULL;
-    HANDLE child_ready_h, parent_ready_h;
 
-    int i;
-    thread_args args;
+    thread_args *args;
 
-    /* allocate thread slot */
-    LOCK();
-    for (i = 0; i != MAX_THREADS && thread_table[i].in_use; i++)
-	;
-    if (i != MAX_THREADS) {
-	thread_table[i].in_use = TRUE;
+    if (!GC_is_initialized) GC_init();
+    		/* make sure GC is initialized (i.e. main thread is attached) */
+    
+    args = GC_malloc_uncollectable(sizeof(thread_args)); 
+	/* Handed off to and deallocated by child thread.	*/
+    if (0 == args) {
+	SetLastError(ERROR_NOT_ENOUGH_MEMORY);
+        return NULL;
     }
-    UNLOCK();
-
-    if (i != MAX_THREADS) {
-
-	/* create unnamed unsignalled events */
-	if (child_ready_h = CreateEvent(NULL, FALSE, FALSE, NULL)) {
-	    if (parent_ready_h = CreateEvent(NULL, FALSE, FALSE, NULL)) {
-
-		/* set up thread arguments */
-		args.child_ready_h = child_ready_h;
-		args.parent_ready_h = parent_ready_h;
-		args.entry = &thread_table[i];
-		args.start = lpStartAddress;
-		args.param = lpParameter;
-
-		thread_h = CreateThread(lpThreadAttributes,
-					dwStackSize, thread_start,
-					&args,
-					dwCreationFlags & ~CREATE_SUSPENDED,
-					lpThreadId);
-
-		if (thread_h) {
-
-		    /* fill in ID and handle; tell child this is done */
-		    thread_table[i].id = *lpThreadId;
-		    if (!DuplicateHandle(GetCurrentProcess(),
-				 	 thread_h,
-					 GetCurrentProcess(),
-			 		 &thread_table[i].handle,
-			  		 0,
-					 0,
-					 DUPLICATE_SAME_ACCESS)) {
-			DWORD last_error = GetLastError();
-			GC_printf1("Last error code: %lx\n", last_error);
-			ABORT("DuplicateHandle failed");
-		    }
-		    SetEvent (parent_ready_h);
-
-		    /* wait for child to fill in stack and copy args */
-		    WaitForSingleObject (child_ready_h, INFINITE);
-
-		    /* suspend the child if requested */
-		    if (dwCreationFlags & CREATE_SUSPENDED)
-			SuspendThread (thread_h);
-
-		    /* let child call given function now (or when resumed) */
-		    SetEvent (parent_ready_h);
-
-		} else {
-		    CloseHandle (parent_ready_h);
-		}
-	    }
-	}
 
-	CloseHandle (child_ready_h);
+    /* set up thread arguments */
+    	args -> start = lpStartAddress;
+    	args -> param = lpParameter;
 
-	if (thread_h == NULL)
-	    thread_table[i].in_use = FALSE;
-
-    } else { /* no thread slot found */
-	SetLastError (ERROR_TOO_MANY_TCBS);
-    }
+    thread_h = CreateThread(lpThreadAttributes,
+    			    dwStackSize, thread_start,
+    			    args, dwCreationFlags,
+    			    lpThreadId);
 
     return thread_h;
 }
@@ -486,19 +496,9 @@ HANDLE WINAPI GC_CreateThread(
 static DWORD WINAPI thread_start(LPVOID arg)
 {
     DWORD ret = 0;
-    thread_args args = *(thread_args *)arg;
+    thread_args *args = (thread_args *)arg;
 
-    /* wait for parent to fill in ID and handle */
-    WaitForSingleObject (args.parent_ready_h, INFINITE);
-    ResetEvent (args.parent_ready_h);
-
-    /* fill in stack; tell parent this is done */
-    args.entry->stack = GC_get_stack_base();
-    SetEvent (args.child_ready_h);
-
-    /* wait for parent to tell us to go (in case it needs to suspend us) */
-    WaitForSingleObject (args.parent_ready_h, INFINITE);
-    CloseHandle (args.parent_ready_h);
+    GC_new_thread();
 
     /* Clear the thread entry even if we exit with an exception.	*/
     /* This is probably pointless, since an uncaught exception is	*/
@@ -506,16 +506,12 @@ static DWORD WINAPI thread_start(LPVOID arg)
 #ifndef __GNUC__
     __try {
 #endif /* __GNUC__ */
-	ret = args.start (args.param);
+	ret = args->start (args->param);
 #ifndef __GNUC__
     } __finally {
 #endif /* __GNUC__ */
-	LOCK();
-	args.entry->stack = 0;
-	args.entry->in_use = FALSE;
-	      /* cast away volatile qualifier */
-	BZERO((void *) &args.entry->context, sizeof(CONTEXT));
-	UNLOCK();
+	GC_free(args);
+	GC_delete_thread(GetCurrentThreadId());
 #ifndef __GNUC__
     }
 #endif /* __GNUC__ */
@@ -549,7 +545,6 @@ int WINAPI WinMain(HINSTANCE hInstance, HINSTANCE hPrevInstance,
     DWORD thread_id;
 
     /* initialize everything */
-    InitializeCriticalSection(&GC_allocate_ml);
     GC_init();
 
     /* start the main thread */
@@ -579,126 +574,44 @@ DWORD WINAPI main_thread_start(LPVOID arg)
 
 # else /* !MSWINCE */
 
-LONG WINAPI GC_write_fault_handler(struct _EXCEPTION_POINTERS *exc_info);
-
-/* threadAttach/threadDetach routines used by both CYGWIN and DLL
- * implementation, since both recieve explicit notification on thread
- * creation/destruction.
- */
-static void threadAttach() {
-  int i;
-  /* It appears to be unsafe to acquire a lock here, since this	*/
-  /* code is apparently not preeemptible on some systems.	*/
-  /* (This is based on complaints, not on Microsoft's official	*/
-  /* documentation, which says this should perform "only simple	*/
-  /* inititalization tasks".)					*/
-  /* Hence we make do with nonblocking synchronization.		*/
-
-  /* The following should be a noop according to the win32	*/
-  /* documentation.  There is empirical evidence that it	*/
-  /* isn't.		- HB					*/
-# if defined(MPROTECT_VDB)
-   if (GC_incremental) SetUnhandledExceptionFilter(GC_write_fault_handler);
-# endif
-                /* cast away volatile qualifier */
-  for (i = 0; InterlockedExchange((LONG*)&thread_table[i].in_use,1) != 0; i++) {
-    /* Compare-and-swap would make this cleaner, but that's not 	*/
-    /* supported before Windows 98 and NT 4.0.  In Windows 2000,	*/
-    /* InterlockedExchange is supposed to be replaced by		*/
-    /* InterlockedExchangePointer, but that's not really what I	*/
-    /* want here.							*/
-    if (i == MAX_THREADS - 1)
-      ABORT("too many threads");
-  }
-  thread_table[i].id = GetCurrentThreadId();
-# ifdef CYGWIN32
-    thread_table[i].pthread_id = pthread_self();
-# endif
-  if (!DuplicateHandle(GetCurrentProcess(),
-	               GetCurrentThread(),
-		       GetCurrentProcess(),
-		       (HANDLE*)&thread_table[i].handle,
-		       0,
-		       0,
-		       DUPLICATE_SAME_ACCESS)) {
-	DWORD last_error = GetLastError();
-	GC_printf1("Last error code: %lx\n", last_error);
-	ABORT("DuplicateHandle failed");
-  }
-  thread_table[i].stack = GC_get_stack_base();
-  if (thread_table[i].stack == NULL) 
-    ABORT("Failed to find stack base in threadAttach");
-  /* If this thread is being created while we are trying to stop	*/
-  /* the world, wait here.  Hopefully this can't happen on any	*/
-  /* systems that don't allow us to block here.			*/
-  while (GC_please_stop) Sleep(20);
-}
-
-static void threadDetach(DWORD thread_id) {
-  int i;
-
-  LOCK();
-  for (i = 0;
-       i < MAX_THREADS &&
-       (!thread_table[i].in_use || thread_table[i].id != thread_id);
-       i++) {}
-  if (i >= MAX_THREADS ) {
-    WARN("thread %ld not found on detach", (GC_word)thread_id);
-  } else {
-    thread_table[i].stack = 0;
-    thread_table[i].in_use = FALSE;
-    CloseHandle(thread_table[i].handle);
-      /* cast away volatile qualifier */
-    BZERO((void *)&thread_table[i].context, sizeof(CONTEXT));
-  }
-  UNLOCK();
-}
-
-#ifdef CYGWIN32
-
 /* Called by GC_init() - we hold the allocation lock.	*/
 void GC_thr_init() {
     if (GC_thr_initialized) return;
+    GC_main_thread = GetCurrentThreadId();
     GC_thr_initialized = TRUE;
 
-#if 0
-    /* this might already be handled in GC_init... */
-    InitializeCriticalSection(&GC_allocate_ml);
-#endif
-
     /* Add the initial thread, so we can stop it.	*/
-    threadAttach();
+    GC_new_thread();
 }
 
+#ifdef CYGWIN32
+
 struct start_info {
     void *(*start_routine)(void *);
     void *arg;
+    GC_bool detached;
 };
 
 int GC_pthread_join(pthread_t pthread_id, void **retval) {
     int result;
     int i;
+    GC_thread me;
 
 #   if DEBUG_CYGWIN_THREADS
-      GC_printf3("thread 0x%x(0x%x) is joining thread 0x%x.\n",(int)pthread_self(),
-		 GetCurrentThreadId(), (int)pthread_id);
+      GC_printf3("thread 0x%x(0x%x) is joining thread 0x%x.\n",
+		 (int)pthread_self(), GetCurrentThreadId(), (int)pthread_id);
 #   endif
 
-    /* Can't do any table lookups here, because thread being joined 
-       might not have registered itself yet */
+    /* Thread being joined might not have registered itself yet. */
+    /* After the join,thread id may have been recycled.		 */
+    /* FIXME: It would be better if this worked more like	 */
+    /* pthread_support.c.					 */
+
+    while ((me = GC_lookup_thread(pthread_id)) == 0) Sleep(10);
 
     result = pthread_join(pthread_id, retval);
 
-    LOCK();
-    for (i = 0; !thread_table[i].in_use || thread_table[i].pthread_id != pthread_id;
-         i++) {
-      if (i == MAX_THREADS - 1) {
-        GC_printf1("Failed to find thread 0x%x in pthread_join()\n", pthread_id);
-        ABORT("thread not found on detach");
-      }
-    }
-    UNLOCK();
-    threadDetach(thread_table[i].id);
+    GC_delete_gc_thread(me);
 
 #   if DEBUG_CYGWIN_THREADS
       GC_printf3("thread 0x%x(0x%x) completed join with thread 0x%x.\n",
@@ -729,10 +642,15 @@ GC_pthread_create(pthread_t *new_thread,
 
     si -> start_routine = start_routine;
     si -> arg = arg;
+    if (attr != 0 &&
+        pthread_attr_getdetachstate(attr, &si->detached)
+	== PTHREAD_CREATE_DETACHED) {
+      si->detached = TRUE;
+    }
 
 #   if DEBUG_CYGWIN_THREADS
-      GC_printf2("About to create a thread from 0x%x(0x%x)\n",(int)pthread_self(),
-		      					      GetCurrentThreadId);
+      GC_printf2("About to create a thread from 0x%x(0x%x)\n",
+		 (int)pthread_self(), GetCurrentThreadId);
 #   endif
     result = pthread_create(new_thread, attr, GC_start_routine, si); 
 
@@ -750,6 +668,8 @@ void * GC_start_routine(void * arg)
     void *(*start)(void *);
     void *start_arg;
     pthread_t pthread_id;
+    GC_thread me;
+    GC_bool detached;
     int i;
 
 #   if DEBUG_CYGWIN_THREADS
@@ -764,17 +684,19 @@ void * GC_start_routine(void * arg)
     LOCK();
     /* We register the thread here instead of in the parent, so that	*/
     /* we don't need to hold the allocation lock during pthread_create. */
-    threadAttach();
+    me = GC_new_thread();
     UNLOCK();
 
     start = si -> start_routine;
     start_arg = si -> arg;
-    pthread_id = pthread_self();
+    if (si-> detached) me -> flags |= DETACHED;
+    me -> pthread_id = pthread_id = pthread_self();
 
     GC_free(si); /* was allocated uncollectable */
 
-    pthread_cleanup_push(GC_thread_exit_proc, pthread_id);
+    pthread_cleanup_push(GC_thread_exit_proc, (void *)me);
     result = (*start)(start_arg);
+    me -> status = result;
     pthread_cleanup_pop(0);
 
 #   if DEBUG_CYGWIN_THREADS
@@ -782,20 +704,12 @@ void * GC_start_routine(void * arg)
 		 (int)pthread_self(),GetCurrentThreadId());
 #   endif
 
-    LOCK();
-    for (i = 0; thread_table[i].pthread_id != pthread_id; i++) {
-      if (i == MAX_THREADS - 1)
-        ABORT("thread not found on exit");
-    }
-    thread_table[i].status = result;
-    UNLOCK();
-
     return(result);
 }
 
 void GC_thread_exit_proc(void *arg)
 {
-    pthread_t pthread_id = (pthread_t)arg;
+    GC_thread me = (GC_thread)arg;
     int i;
 
 #   if DEBUG_CYGWIN_THREADS
@@ -804,25 +718,47 @@ void GC_thread_exit_proc(void *arg)
 #   endif
 
     LOCK();
-    for (i = 0; thread_table[i].pthread_id != pthread_id; i++) {
-      if (i == MAX_THREADS - 1)
-        ABORT("thread not found on exit");
+    if (me -> flags & DETACHED) {
+      GC_delete_thread(GetCurrentThreadId());
+    } else {
+      /* deallocate it as part of join */
+      me -> flags |= FINISHED;
     }
     UNLOCK();
-
-#if 0
-    /* TODO: we need a way to get the exit value after a pthread_exit so we can stash it safely away */
-    thread_table[i].status = ???
-#endif
 }
 
 /* nothing required here... */
 int GC_pthread_sigmask(int how, const sigset_t *set, sigset_t *oset) {
   return pthread_sigmask(how, set, oset);
 }
-int GC_pthread_detach(pthread_t thread) {
-  return pthread_detach(thread);
+
+int GC_pthread_detach(pthread_t thread)
+{
+    int result;
+    GC_thread thread_gc_id;
+    
+    LOCK();
+    thread_gc_id = GC_lookup_thread(thread);
+    UNLOCK();
+    result = pthread_detach(thread);
+    if (result == 0) {
+      LOCK();
+      thread_gc_id -> flags |= DETACHED;
+      /* Here the pthread thread id may have been recycled. */
+      if (thread_gc_id -> flags & FINISHED) {
+        GC_delete_gc_thread(thread_gc_id);
+      }
+      UNLOCK();
+    }
+    return result;
+}
+
+GC_PTR GC_get_thread_stack_base()
+{
+  extern GC_PTR _tlsbase __asm__ ("%fs:4");
+  return _tlsbase;
 }
+
 #else /* !CYGWIN32 */
 
 /*
@@ -834,15 +770,17 @@ BOOL WINAPI DllMain(HINSTANCE inst, ULONG reason, LPVOID reserved)
 {
   switch (reason) {
   case DLL_PROCESS_ATTACH:
-    InitializeCriticalSection(&GC_allocate_ml);
     GC_init();	/* Force initialization before thread attach.	*/
     /* fall through */
   case DLL_THREAD_ATTACH:
-    threadAttach();
+    GC_ASSERT(GC_thr_initialized);
+    if (GC_main_thread != GetCurrentThreadId()) {
+        GC_new_thread();
+    } /* o.w. we already did it during GC_thr_init(), called by GC_init() */
     break;
 
   case DLL_THREAD_DETACH:
-    threadDetach(GetCurrentThreadId());
+    GC_delete_thread(GetCurrentThreadId());
     break;
 
   case DLL_PROCESS_DETACH:
@@ -850,15 +788,10 @@ BOOL WINAPI DllMain(HINSTANCE inst, ULONG reason, LPVOID reserved)
       int i;
 
       LOCK();
-      for (i = 0; i < MAX_THREADS; ++i)
+      for (i = 0; i <= GC_get_max_thread_index(); ++i)
       {
           if (thread_table[i].in_use)
-          {
-              thread_table[i].stack = 0;
-              thread_table[i].in_use = FALSE;
-              CloseHandle(thread_table[i].handle);
-              BZERO((void *) &thread_table[i].context, sizeof(CONTEXT));
-          }
+	    GC_delete_gc_thread(thread_table + i);
       }
       UNLOCK();