SPU --auto-overlay --num-regions

[Alan Modra <amodra at bigpond dot net dot au>]

This extends SPU --auto-overlay to generate linker scripts for more
than one overlay buffer.  It also moves the overlays to before .text
rather than after, for less wasted space on average.

bfd/
	* elf32-spu.h (struct spu_elf_params): Add num_regions.
	* elf32-spu.c (spu_elf_auto_overlay): Handle multiple overlay regions.
ld/
	* emultempl/spuelf.em (params): Init new field.
	(OPTION_SPU_NUM_REGIONS): Define.
	(PARSE_AND_LIST_LONGOPTS, PARSE_AND_LIST_OPTIONS): Add --num-regions.
	(PARSE_AND_LIST_ARGS_CASES): Handle --num-regions.

Note diff -w!

Index: bfd/elf32-spu.c
===================================================================
RCS file: /cvs/src/src/bfd/elf32-spu.c,v
retrieving revision 1.56
diff -u -p -w -r1.56 elf32-spu.c
--- bfd/elf32-spu.c	10 Dec 2008 06:32:52 -0000	1.56
+++ bfd/elf32-spu.c	10 Dec 2008 13:23:06 -0000
@@ -3444,8 +3444,9 @@ spu_elf_auto_overlay (struct bfd_link_in
   unsigned int fixed_size, lo, hi;
   struct spu_link_hash_table *htab;
   unsigned int base, i, count, bfd_count;
-  int ovlynum;
+  unsigned int region, ovlynum;
   asection **ovly_sections, **ovly_p;
+  unsigned int *ovly_map;
   FILE *script;
   unsigned int total_overlay_size, overlay_size;
   struct elf_link_hash_entry *h;
@@ -3625,20 +3626,17 @@ spu_elf_auto_overlay (struct bfd_link_in
   if (!for_each_node (collect_overlays, info, &ovly_p, TRUE))
     goto err_exit;
   count = (size_t) (ovly_p - ovly_sections) / 2;
-
-  script = (*htab->params->spu_elf_open_overlay_script) ();
-
-  if (fprintf (script, "SECTIONS\n{\n OVERLAY :\n {\n") <= 0)
-    goto file_err;
+  ovly_map = bfd_malloc (count * sizeof (*ovly_map));
+  if (ovly_map == NULL)
+    goto err_exit;
 
   memset (&dummy_caller, 0, sizeof (dummy_caller));
-  overlay_size = htab->local_store - fixed_size;
+  overlay_size = (htab->local_store - fixed_size) / htab->params->num_regions;
   base = 0;
   ovlynum = 0;
   while (base < count)
     {
       unsigned int size = 0;
-      unsigned int j;
 
       for (i = base; i < count; i++)
 	{
@@ -3741,9 +3739,44 @@ spu_elf_auto_overlay (struct bfd_link_in
 	  goto err_exit;
 	}
 
-      if (fprintf (script, "  .ovly%d {\n", ++ovlynum) <= 0)
+      while (dummy_caller.call_list != NULL)
+	{
+	  struct call_info *call = dummy_caller.call_list;
+	  dummy_caller.call_list = call->next;
+	  free (call);
+	}
+
+      ++ovlynum;
+      while (base < i)
+	ovly_map[base++] = ovlynum;
+    }
+
+  script = htab->params->spu_elf_open_overlay_script ();
+
+  if (fprintf (script, "SECTIONS\n{\n") <= 0)
 	goto file_err;
-      for (j = base; j < i; j++)
+
+  for (region = 1; region <= htab->params->num_regions; region++)
+    {
+      ovlynum = region;
+      base = 0;
+      while (base < count && ovly_map[base] < ovlynum)
+	base++;
+
+      if (base == count)
+	break;
+
+      if (fprintf (script, " OVERLAY :\n {\n") <= 0)
+	goto file_err;
+
+      while (base < count)
+	{
+	  unsigned int j;
+	  
+	  if (fprintf (script, "  .ovly%u {\n", ovlynum) <= 0)
+	    goto file_err;
+
+	  for (j = base; j < count && ovly_map[j] == ovlynum; j++)
 	{
 	  asection *sec = ovly_sections[2 * j];
 
@@ -3775,7 +3808,7 @@ spu_elf_auto_overlay (struct bfd_link_in
 	    }
 	}
 
-      for (j = base; j < i; j++)
+	  for (j = base; j < count && ovly_map[j] == ovlynum; j++)
 	{
 	  asection *sec = ovly_sections[2 * j + 1];
 	  if (sec != NULL
@@ -3813,18 +3846,20 @@ spu_elf_auto_overlay (struct bfd_link_in
       if (fprintf (script, "  }\n") <= 0)
 	goto file_err;
 
-      while (dummy_caller.call_list != NULL)
-	{
-	  struct call_info *call = dummy_caller.call_list;
-	  dummy_caller.call_list = call->next;
-	  free (call);
+	  base = j;
+	  ovlynum += htab->params->num_regions;
+	  while (base < count && ovly_map[base] < ovlynum)
+	    base++;
 	}
 
-      base = i;
+      if (fprintf (script, " }\n") <= 0)
+	goto file_err;
     }
+
+  free (ovly_map);
   free (ovly_sections);
 
-  if (fprintf (script, " }\n}\nINSERT AFTER .text;\n") <= 0)
+  if (fprintf (script, "}\nINSERT BEFORE .text;\n") <= 0)
     goto file_err;
   if (fclose (script) != 0)
     goto file_err;
Index: bfd/elf32-spu.h
===================================================================
RCS file: /cvs/src/src/bfd/elf32-spu.h,v
retrieving revision 1.9
diff -u -p -w -r1.9 elf32-spu.h
--- bfd/elf32-spu.h	10 Dec 2008 06:32:52 -0000	1.9
+++ bfd/elf32-spu.h	10 Dec 2008 13:23:06 -0000
@@ -55,6 +55,7 @@ struct spu_elf_params
   bfd_vma local_store_hi;
 
   /* Control --auto-overlay feature.  */
+  unsigned int num_regions;
   unsigned int auto_overlay_fixed;
   unsigned int auto_overlay_reserved;
   int extra_stack_space;
Index: ld/emultempl/spuelf.em
===================================================================
RCS file: /cvs/src/src/ld/emultempl/spuelf.em,v
retrieving revision 1.29
diff -u -p -w -r1.29 spuelf.em
--- ld/emultempl/spuelf.em	10 Dec 2008 06:32:52 -0000	1.29
+++ ld/emultempl/spuelf.em	10 Dec 2008 13:23:06 -0000
@@ -39,7 +39,7 @@ static struct spu_elf_params params =
   &spu_elf_relink,
   0, ovly_normal, 0, 0, 0, 0,
   0, 0x3ffff,
-  0, 0, 2000
+  1, 0, 0, 2000
 };
   
 static char *auto_overlay_file = 0;
@@ -520,7 +520,8 @@ PARSE_AND_LIST_PROLOGUE='
 #define OPTION_SPU_AUTO_OVERLAY		(OPTION_SPU_STACK_SYMS + 1)
 #define OPTION_SPU_AUTO_RELINK		(OPTION_SPU_AUTO_OVERLAY + 1)
 #define OPTION_SPU_OVERLAY_RODATA	(OPTION_SPU_AUTO_RELINK + 1)
-#define OPTION_SPU_FIXED_SPACE		(OPTION_SPU_OVERLAY_RODATA + 1)
+#define OPTION_SPU_NUM_REGIONS		(OPTION_SPU_OVERLAY_RODATA + 1)
+#define OPTION_SPU_FIXED_SPACE		(OPTION_SPU_NUM_REGIONS + 1)
 #define OPTION_SPU_RESERVED_SPACE	(OPTION_SPU_FIXED_SPACE + 1)
 #define OPTION_SPU_EXTRA_STACK		(OPTION_SPU_RESERVED_SPACE + 1)
 #define OPTION_SPU_NO_AUTO_OVERLAY	(OPTION_SPU_EXTRA_STACK + 1)
@@ -537,6 +538,7 @@ PARSE_AND_LIST_LONGOPTS='
   { "auto-overlay", optional_argument, NULL, OPTION_SPU_AUTO_OVERLAY },
   { "auto-relink", no_argument, NULL, OPTION_SPU_AUTO_RELINK },
   { "overlay-rodata", no_argument, NULL, OPTION_SPU_OVERLAY_RODATA },
+  { "num-regions", required_argument, NULL, OPTION_SPU_NUM_REGIONS },
   { "fixed-space", required_argument, NULL, OPTION_SPU_FIXED_SPACE },
   { "reserved-space", required_argument, NULL, OPTION_SPU_RESERVED_SPACE },
   { "extra-stack-space", required_argument, NULL, OPTION_SPU_EXTRA_STACK },
@@ -557,6 +559,7 @@ PARSE_AND_LIST_OPTIONS='
   --auto-relink               Rerun linker using auto-overlay script.\n\
   --overlay-rodata            Place read-only data with associated function\n\
                               code in overlays.\n\
+  --num-regions               Number of overlay buffers (default 1).\n\
   --fixed-space=bytes         Local store for non-overlay code and data.\n\
   --reserved-space=bytes      Local store for stack and heap.  If not specified\n\
                               ld will estimate stack size and assume no heap.\n\
@@ -621,6 +624,16 @@ PARSE_AND_LIST_ARGS_CASES='
       params.auto_overlay |= 4;
       break;
 
+    case OPTION_SPU_NUM_REGIONS:
+      {
+	char *end;
+	params.num_regions = strtoul (optarg, &end, 0);
+	if (*end == 0)
+	  break;
+	einfo (_("%P%F: invalid --num-regions `%s'\''\n"), optarg);
+      }
+      break;
+
     case OPTION_SPU_FIXED_SPACE:
       {
 	char *end;

-- 
Alan Modra
Australia Development Lab, IBM