RE: Java GC descriptor vs. tree_int_cst on 64-bit platforms

["Boehm, Hans" <hans_boehm at hp dot com>]

Here's what I sent to Andrew.  This somewhat fixed the comment that Fergus
Henderson was worried about.  And it includes the change log.

In spite of the subject, this patch is related to the compiler patch only in
that both are necessary for gcj on Itanium. 

Hans

finalize.c:
	- Accomodate finalization requests for static objects.
	  (Will be required by hash synchronization.  May be needed
	  in some configurations now.)

gc_priv.h:
	- Define MIN_WORDS.  All allocation requests are rounded
	  up to at least this size.  Removes a subtle assumption that
	  Java objects have a 2 word header.

gcconfig.h:
	- Adjust Linux/IA64 configuration for non-ancient kernels.
	  (Necessary fix for IA64.)

linux_threads.c:
	- Fix syntax error in currently unused code.  Will be needed
	  for Linux/PA-RISC.

malloc.c:
	- Handle MIN_WORDS.

misc.c:
	- Handle MIN_WORDS.
	- Change stack cleaning code to typically clear about one tenth
	  the memory it used to in the threads configuration.  Occasionally
	  still clear more.  (This is really a fix for a long-standing
	  and fairly significant performance bug with threads.)

os_dep.c:
	- Fix the code for finding the beginning of the data segment under
	  Linux.  I believe this is necessary for some IA64 Linux
	  distributions.  It will also helo other platforms, though those
	  may additionally require a gcconfig.h adjustment.  (This basically
	  works around the absence of a data_start or __data_start
	  definition in glibc.)

test.c:
	- Handle rounding due to MIN_WORDS.



Index: boehm-gc/finalize.c
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/finalize.c,v
retrieving revision 1.5
diff -u -r1.5 finalize.c
--- finalize.c	2000/05/07 00:43:49	1.5
+++ finalize.c	2001/03/30 22:53:28
@@ -324,6 +324,7 @@
     struct finalizable_object * curr_fo, * prev_fo;
     int index;
     struct finalizable_object *new_fo;
+    hdr *hhdr;
     DCL_LOCK_STATE;
 
 #   ifdef THREADS
@@ -402,6 +403,19 @@
 #	endif
         return;
     }
+    GET_HDR(base, hhdr);
+    if (0 == hhdr) {
+      /* We won't collect it, hence finalizer wouldn't be run. */
+      /* This is changed for gcj, but it will be in version 6.0 of the	*/
+      /* standard collector distribution.  It costs virtually nothing	*/
+      /* here, but it's expensive to check in the hash synchronization	*/
+      /* code, where it matters.		-HB			*/
+#     ifdef THREADS
+          UNLOCK();
+    	  ENABLE_SIGNALS();
+#     endif
+      return;
+    }
 #   ifdef THREADS
       new_fo = (struct finalizable_object *)
     	GC_generic_malloc_inner(sizeof(struct finalizable_object),NORMAL);
@@ -413,7 +427,7 @@
         new_fo -> fo_hidden_base = (word)HIDE_POINTER(base);
 	new_fo -> fo_fn = fn;
 	new_fo -> fo_client_data = (ptr_t)cd;
-	new_fo -> fo_object_size = GC_size(base);
+	new_fo -> fo_object_size = hhdr -> hb_sz;
 	new_fo -> fo_mark_proc = mp;
 	fo_set_next(new_fo, fo_head[index]);
 	GC_fo_entries++;
Index: boehm-gc/gc_priv.h
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/gc_priv.h,v
retrieving revision 1.14
diff -u -r1.14 gc_priv.h
--- gc_priv.h	2000/09/04 21:46:26	1.14
+++ gc_priv.h	2001/03/30 22:53:29
@@ -890,6 +890,18 @@
 #   define SMALL_OBJ(bytes) ((bytes) <= WORDS_TO_BYTES(MAXOBJSZ))
 #   define ADD_SLOP(bytes) (bytes)
 # endif
+# ifndef MIN_WORDS
+    /* MIN_WORDS is the size of the smallest allocated object.	*/
+    /* 1 and 2 are the only valid values.			*/
+    /* 2 must be used if:					*/
+    /* - GC_gcj_malloc can be used for objects of requested	*/
+    /*   size smaller than 2 words				*/
+#   if defined(GC_GCJ_SUPPORT)
+#     define MIN_WORDS 2   	/* Smallest allocated object.	*/
+#   else
+#     define MIN_WORDS 1
+#   endif
+# endif
 
 
 /*
Index: boehm-gc/gcconfig.h
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/gcconfig.h,v
retrieving revision 1.15.4.1
diff -u -r1.15.4.1 gcconfig.h
--- gcconfig.h	2001/03/17 01:00:18	1.15.4.1
+++ gcconfig.h	2001/03/30 22:53:29
@@ -1116,11 +1116,20 @@
 #       define CPP_WORDSZ 64
 	/* This should really be done through /proc, but that	*/
 	/* requires we run on an IA64 kernel.			*/
-#       define STACKBOTTOM ((ptr_t) 0xa000000000000000l)
+	/* The following works on NUE and older kernels:	*/
+/* #       define STACKBOTTOM ((ptr_t) 0xa000000000000000l)	*/
+	/* This does not work on NUE:				*/
+#       define LINUX_STACKBOTTOM
 	/* We also need the base address of the register stack	*/
-	/* backing store.  There is probably a better way to	*/
-	/* get that, too ...					*/
-#	define BACKING_STORE_BASE ((ptr_t) 0x9fffffff80000000l)
+	/* backing store.  There should be a better way to get	*/
+	/* this:						*/
+#	define APPROX_BS_BASE ((word)GC_stackbottom-0x80000000)
+	/* We round to the next multiple of 1 MB, to compensate	*/
+	/* for the fact that the stack base is displaced by	*/
+	/* the environment, etc.				*/
+#	define BACKING_STORE_BASE \
+		(ptr_t)((APPROX_BS_BASE + 0xfffff) & ~0xfffff)
+
 #	if 1
 #	    define SEARCH_FOR_DATA_START
 #	    define DATASTART GC_data_start
Index: boehm-gc/gcj_mlc.c
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/gcj_mlc.c,v
retrieving revision 1.2
diff -u -r1.2 gcj_mlc.c
--- gcj_mlc.c	2000/12/23 02:23:27	1.2
+++ gcj_mlc.c	2001/03/30 22:53:29
@@ -267,7 +267,6 @@
         } else {
             *opp = obj_link(op);
             GC_words_allocd += lw;
-            FASTUNLOCK();
         }
 	*(void **)op = ptr_to_struct_containing_descr;
 	UNLOCK();
Index: boehm-gc/linux_threads.c
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/linux_threads.c,v
retrieving revision 1.8
diff -u -r1.8 linux_threads.c
--- linux_threads.c	2000/12/30 12:18:38	1.8
+++ linux_threads.c	2001/03/30 22:53:29
@@ -611,7 +611,7 @@
 	/* Needs to be plausible, since an asynchronous stack mark	*/
 	/* should not crash.						*/
 #   else
-      me -> stack_end = (ptr_t)(((word)(&dummy) & ~(GC_page_size - 1));
+      me -> stack_end = (ptr_t)((word)(&dummy) & ~(GC_page_size - 1));
       me -> stack_ptr = me -> stack_end + 0x10;
 #   endif
     /* This is dubious, since we may be more than a page into the stack, */
Index: boehm-gc/malloc.c
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/malloc.c,v
retrieving revision 1.4
diff -u -r1.4 malloc.c
--- malloc.c	2000/04/19 10:09:58	1.4
+++ malloc.c	2001/03/30 22:53:30
@@ -52,7 +52,7 @@
 	  lw = GC_size_map[lb];
 #	else
 	  lw = ALIGNED_WORDS(lb);
-	  if (lw == 0) lw = 1;
+	  if (lw == 0) lw = MIN_WORDS;
 #       endif
 	opp = &(kind -> ok_freelist[lw]);
         if( (op = *opp) == 0 ) {
Index: boehm-gc/misc.c
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/misc.c,v
retrieving revision 1.12
diff -u -r1.12 misc.c
--- misc.c	2000/04/19 10:09:58	1.12
+++ misc.c	2001/03/30 22:53:30
@@ -107,13 +107,17 @@
     {
 	register unsigned i;
 
-	/* Map size 0 to 1.  This avoids problems at lower levels. */
-	  GC_size_map[0] = 1;
+	/* Map size 0 to something bigger.			   */
+        /* This avoids problems at lower levels. 		   */
 	/* One word objects don't have to be 2 word aligned.	   */
-	  for (i = 1; i < sizeof(word); i++) {
-	      GC_size_map[i] = 1;
+	  for (i = 0; i < sizeof(word); i++) {
+	      GC_size_map[i] = MIN_WORDS;
 	  }
-	  GC_size_map[sizeof(word)] = ROUNDED_UP_WORDS(sizeof(word));
+#	  if MIN_WORDS > 1
+	    GC_size_map[sizeof(word)] = MIN_WORDS;
+#	  else
+	    GC_size_map[sizeof(word)] = ROUNDED_UP_WORDS(sizeof(word));
+#	  endif
 	for (i = sizeof(word) + 1; i <= 8 * sizeof(word); i++) {
 #           ifdef ALIGN_DOUBLE
 	      GC_size_map[i] = (ROUNDED_UP_WORDS(i) + 1) & (~1);
@@ -202,10 +206,10 @@
  */
 word GC_stack_last_cleared = 0;	/* GC_no when we last did this */
 # ifdef THREADS
-#   define CLEAR_SIZE 2048
-# else
-#   define CLEAR_SIZE 213
+#   define BIG_CLEAR_SIZE 2048	/* Clear this much now and then.	*/
+#   define SMALL_CLEAR_SIZE 256 /* Clear this much every time.		*/
 # endif
+# define CLEAR_SIZE 213  /* Granularity for GC_clear_stack_inner */
 # define DEGRADE_RATE 50
 
 word GC_min_sp;		/* Coolest stack pointer value from which
we've */
@@ -262,10 +266,12 @@
 {
     register word sp = (word)GC_approx_sp();  /* Hotter than actual sp */
 #   ifdef THREADS
-        word dummy[CLEAR_SIZE];
-#   else
-    	register word limit;
+        word dummy[SMALL_CLEAR_SIZE];
+	unsigned random_no = 0;  /* Should be more random than it is ... */
+				 /* Used to occasionally clear a bigger	 */
+				 /* chunk.				 */
 #   endif
+    register word limit;
     
 #   define SLOP 400
 	/* Extra bytes we clear every time.  This clears our own	*/
@@ -283,7 +289,14 @@
 	/* thus more junk remains accessible, thus the heap gets	*/
 	/* larger ...							*/
 # ifdef THREADS
-    BZERO(dummy, CLEAR_SIZE*sizeof(word));
+    if (++random_no % 13 == 0) {
+	limit = sp;
+	MAKE_HOTTER(limit, BIG_CLEAR_SIZE*sizeof(word));
+	return GC_lear_stack_inner(arg, limit);
+    } else {
+	BZERO(dummy, SMALL_CLEAR_SIZE*sizeof(word));
+	return arg;
+    }
 # else
     if (GC_gc_no > GC_stack_last_cleared) {
         /* Start things over, so we clear the entire stack again */
Index: boehm-gc/os_dep.c
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/os_dep.c,v
retrieving revision 1.14
diff -u -r1.14 os_dep.c
--- os_dep.c	2000/06/19 18:46:45	1.14
+++ os_dep.c	2001/03/30 22:53:30
@@ -144,20 +144,37 @@
 #endif
 
 #if defined(SEARCH_FOR_DATA_START)
-  /* The following doesn't work if the GC is in a dynamic library.	*/
   /* The I386 case can be handled without a search.  The Alpha case	*/
   /* used to be handled differently as well, but the rules changed	*/
   /* for recent Linux versions.  This seems to be the easiest way to	*/
   /* cover all versions.						*/
-  ptr_t GC_data_start;
+
+# ifdef LINUX
+#   pragma weak __data_start
+    extern int __data_start;
+#   pragma weak data_start
+    extern int data_start;
+# endif /* LINUX */
+  extern int _end;
 
-  extern char * GC_copyright[];  /* Any data symbol would do. */
+  ptr_t GC_data_start;
 
   void GC_init_linux_data_start()
   {
     extern ptr_t GC_find_limit();
 
-    GC_data_start = GC_find_limit((ptr_t)GC_copyright, FALSE);
+#   ifdef LINUX
+      /* Try the easy approaches first:	*/
+      if (&__data_start != 0) {
+	  GC_data_start = (ptr_t)(&__data_start);
+	  return;
+      }
+      if (&data_start != 0) {
+	  GC_data_start = (ptr_t)(&data_start);
+	  return;
+      }
+#   endif /* LINUX */
+    GC_data_start = GC_find_limit((ptr_t)(&_end), FALSE);
   }
 #endif
 
Index: boehm-gc/test.c
===================================================================
RCS file: /cvs/gcc/gcc/boehm-gc/test.c,v
retrieving revision 1.4
diff -u -r1.4 test.c
--- test.c	2000/04/19 10:09:58	1.4
+++ test.c	2001/03/30 22:53:30
@@ -958,17 +958,17 @@
 		"This test program is not designed for leak detection
mode\n");
 	(void)GC_printf0("Expect lots of problems.\n");
 #   endif
-    if (GC_size(GC_malloc(7)) != 8
+    if (GC_size(GC_malloc(7)) != 8 &&
+	GC_size(GC_malloc(7)) != MIN_WORDS * sizeof(GC_word)
 	|| GC_size(GC_malloc(15)) != 16) {
 	    (void)GC_printf0("GC_size produced unexpected results\n");
 	    FAIL;
     }
-    if (GC_size(GC_malloc(0)) != 4 && GC_size(GC_malloc(0)) != 8) {
+    if (GC_size(GC_malloc(0)) != MIN_WORDS * sizeof(GC_word)) {
     	(void)GC_printf0("GC_malloc(0) failed\n");
 	    FAIL;
     }
-    if (GC_size(GC_malloc_uncollectable(0)) != 4
-        && GC_size(GC_malloc_uncollectable(0)) != 8) {
+    if (GC_size(GC_malloc_uncollectable(0)) != MIN_WORDS * sizeof(GC_word))
{
     	(void)GC_printf0("GC_malloc_uncollectable(0) failed\n");
 	    FAIL;
     }