Re: [patch] Fix occasional 2MB gap in the linked binary

[Jan Kratochvil <jan dot kratochvil at redhat dot com>]

On Wed, 13 Aug 2008 07:13:41 +0200, Alan Modra wrote:
> I was hoping Jakub would review this change since it is his code..
> My opinion is that this would be cleaner if you just saved dot to
> min_base here, and moved all the aligning and tweaking to one place,
> in ldlang.c.

My former patch nullified/regressed the excessive VMA skip optimization:
  http://www.sourceware.org/ml/binutils/2004-12/msg00152.html
The MIN_BASE value should be used in its original way.  I just moved its
computing to ldlang.c.

The value I formerly computed was already stored in BASE so I reused it for
not overlapping the file offsets inside a MAXPAGE.


Regards,
Jan

2008-08-20  Jan Kratochvil  <jan.kratochvil@redhat.com>

	PR ld/6833
	* ldexp.c (fold_binary <DATA_SEGMENT_ALIGN>): Do not align
	EXPLD.DATASEG.MIN_BASE.
	* ldlang.c (lang_size_sections): New variable OLD_BASE.  Rename
	OLD_MIN_BASE to MIN_BASE with the former alignment from `ldexp.c'.
	Use OLD_BASE now for the minimal base check after the base decrease by
	the maximum alignment found.

--- ld/ldexp.c	7 May 2008 14:30:41 -0000	1.74
+++ ld/ldexp.c	20 Aug 2008 13:10:57 -0000
@@ -415,7 +415,7 @@ fold_binary (etree_type *tree)
 		      if (expld.phase == lang_allocating_phase_enum)
 			{
 			  expld.dataseg.phase = exp_dataseg_align_seen;
-			  expld.dataseg.min_base = align_n (expld.dot, maxpage);
+			  expld.dataseg.min_base = expld.dot;
 			  expld.dataseg.base = expld.result.value;
 			  expld.dataseg.pagesize = commonpage;
 			  expld.dataseg.maxpagesize = maxpage;
--- ld/ldlang.c	17 Aug 2008 03:12:50 -0000	1.290
+++ ld/ldlang.c	20 Aug 2008 13:11:02 -0000
@@ -5056,17 +5056,23 @@ lang_size_sections (bfd_boolean *relax, 
     {
       /* If DATA_SEGMENT_ALIGN DATA_SEGMENT_RELRO_END pair was seen, try
 	 to put expld.dataseg.relro on a (common) page boundary.  */
-      bfd_vma old_min_base, relro_end, maxpage;
+      bfd_vma min_base, old_base, relro_end, maxpage;
 
       expld.dataseg.phase = exp_dataseg_relro_adjust;
-      old_min_base = expld.dataseg.min_base;
       maxpage = expld.dataseg.maxpagesize;
+      /* MIN_BASE is the absolute minimum address we are allowed to start the
+	 read-write segment (byte before will be mapped read-only).  */
+      min_base = (expld.dataseg.min_base + maxpage - 1) & ~(maxpage - 1);
+      /* OLD_BASE is the address for a feasible minimum address which will
+	 still not cause a data overlap inside MAXPAGE causing file offset skip
+	 by MAXPAGE.  */
+      old_base = expld.dataseg.base;
       expld.dataseg.base += (-expld.dataseg.relro_end
 			     & (expld.dataseg.pagesize - 1));
       /* Compute the expected PT_GNU_RELRO segment end.  */
       relro_end = ((expld.dataseg.relro_end + expld.dataseg.pagesize - 1)
 		   & ~(expld.dataseg.pagesize - 1));
-      if (old_min_base + maxpage < expld.dataseg.base)
+      if (min_base + maxpage < expld.dataseg.base)
 	{
 	  expld.dataseg.base -= maxpage;
 	  relro_end -= maxpage;
@@ -5077,7 +5083,8 @@ lang_size_sections (bfd_boolean *relax, 
 	{
 	  /* The alignment of sections between DATA_SEGMENT_ALIGN
 	     and DATA_SEGMENT_RELRO_END caused huge padding to be
-	     inserted at DATA_SEGMENT_RELRO_END.  Try some other base.  */
+	     inserted at DATA_SEGMENT_RELRO_END.  Try to start a bit lower so
+	     that the section alignments will fit in.  */
 	  asection *sec;
 	  unsigned int max_alignment_power = 0;
 
@@ -5091,8 +5098,7 @@ lang_size_sections (bfd_boolean *relax, 
 
 	  if (((bfd_vma) 1 << max_alignment_power) < expld.dataseg.pagesize)
 	    {
-	      if (expld.dataseg.base - (1 << max_alignment_power)
-		  < old_min_base)
+	      if (expld.dataseg.base - (1 << max_alignment_power) < old_base)
 		expld.dataseg.base += expld.dataseg.pagesize;
 	      expld.dataseg.base -= (1 << max_alignment_power);
 	      lang_reset_memory_regions ();