[RFC][patch] Work around for gold/10400 bug.

[RFC][patch] Work around for gold/10400 bug.

[Paul Pluzhnikov]

Greetings,

Perhaps I am jumping the gun here, but I'd like some comments on this
patch as well.

The patch below is on top of my earlier 'speedup find_fde' patch:

  http://sourceware.org/ml/gdb-patches/2009-07/msg00528.html

and works around the gold/10400 issue:

  http://sourceware.org/bugzilla/show_bug.cgi?id=10400

by using the first FDE entry when there are incompatible FDEs "covering"
the same code region. The first entry exactly corresponds to the COMDAT
section gold actually selected, and is the right one to use.

[Gold issue has been fixed, but it will be quite a while before I can pick
up that fix, and it would be nice to have a GDB workaround.]

Tested on Linux/x86_64 with no new failures.

Comments?

Thanks,
--
Paul Pluzhnikov

2009-07-21  Paul Pluzhnikov  <ppluzhnikov@google.com>

	gold/10400
	* dwarf2-frame.c (qsort_fde_cmp): Use stable sort.


Index: src/gdb/dwarf2-frame.c
===================================================================
--- src.orig/gdb/dwarf2-frame.c	2009-07-21 16:33:41.000000000 -0700
+++ src/gdb/dwarf2-frame.c	2009-07-21 16:43:46.000000000 -0700
@@ -1955,8 +1955,16 @@
   struct dwarf2_fde *aa = *(struct dwarf2_fde **)a;
   struct dwarf2_fde *bb = *(struct dwarf2_fde **)b;
   if (aa->initial_location == bb->initial_location)
-    /* Put eh_frame entries after debug_frame ones.  */
-    return aa->eh_frame_p - bb->eh_frame_p;
+    {
+      if (aa->address_range != bb->address_range
+          && aa->eh_frame_p == 0 && bb->eh_frame_p == 0)
+        /* Linker bug, e.g. gold/10400.
+           Work around it by keeping stable sort order.  */
+        return (char *) a - (char *) b;
+      else
+        /* Put eh_frame entries after debug_frame ones.  */
+        return aa->eh_frame_p - bb->eh_frame_p;
+    }
 
   return aa->initial_location - bb->initial_location;
 }

Re: [RFC][patch] Work around for gold/10400 bug.

[Andreas Schwab]

ppluzhnikov@google.com (Paul Pluzhnikov) writes:

> +      if (aa->address_range != bb->address_range
> +          && aa->eh_frame_p == 0 && bb->eh_frame_p == 0)
> +        /* Linker bug, e.g. gold/10400.
> +           Work around it by keeping stable sort order.  */
> +        return (char *) a - (char *) b;

This can overflow.

Andreas.

-- 
Andreas Schwab, schwab@redhat.com
GPG Key fingerprint = D4E8 DBE3 3813 BB5D FA84  5EC7 45C6 250E 6F00 984E
"And now for something completely different."

Re: [RFC][patch] Work around for gold/10400 bug.

[Paul Pluzhnikov]

[-- Attachment #1: Type: text/plain, Size: 330 bytes --]

On Wed, Jul 22, 2009 at 1:55 AM, Andreas Schwab<schwab@redhat.com> wrote:

>> +       return (char *) a - (char *) b;
>
> This can overflow.

Fixed in the updated patch.

Thanks,
-- 
Paul Pluzhnikov

009-07-21  Paul Pluzhnikov  <ppluzhnikov@google.com>

       gold/10400
       * dwarf2-frame.c (qsort_fde_cmp): Use stable sort.

[-- Attachment #2: gdb-gold-workaround-20090722.txt --]
[-- Type: text/plain, Size: 969 bytes --]

Index: src/gdb/dwarf2-frame.c
===================================================================
--- src.orig/gdb/dwarf2-frame.c	2009-07-22 13:26:17.000000000 -0700
+++ src/gdb/dwarf2-frame.c	2009-07-22 13:29:46.000000000 -0700
@@ -1958,9 +1958,18 @@
 {
   struct dwarf2_fde *aa = *(struct dwarf2_fde **)a;
   struct dwarf2_fde *bb = *(struct dwarf2_fde **)b;
+
   if (aa->initial_location == bb->initial_location)
-    /* Put eh_frame entries after debug_frame ones.  */
-    return aa->eh_frame_p - bb->eh_frame_p;
+    {
+      if (aa->address_range != bb->address_range
+          && aa->eh_frame_p == 0 && bb->eh_frame_p == 0)
+        /* Linker bug, e.g. gold/10400.
+           Work around it by keeping stable sort order.  */
+        return (a < b) ? -1 : 1;
+      else
+        /* Put eh_frame entries after debug_frame ones.  */
+        return aa->eh_frame_p - bb->eh_frame_p;
+    }
 
   return (aa->initial_location < bb->initial_location) ? -1 : 1;
 }

Re: [RFC][patch] Work around for gold/10400 bug.

[Paul Pluzhnikov]

On Wed, Jul 22, 2009 at 1:32 PM, Paul Pluzhnikov<ppluzhnikov@google.com> wrote:

> Fixed in the updated patch.

http://sourceware.org/ml/gdb-patches/2009-07/msg00569.html

Ping?
-- 
Paul Pluzhnikov

Re: [RFC][patch] Work around for gold/10400 bug.

[Daniel Jacobowitz]

On Wed, Jul 29, 2009 at 05:42:44PM -0700, Paul Pluzhnikov wrote:
> On Wed, Jul 22, 2009 at 1:32 PM, Paul Pluzhnikov<ppluzhnikov@google.com> wrote:
> 
> > Fixed in the updated patch.
> 
> http://sourceware.org/ml/gdb-patches/2009-07/msg00569.html
> 
> Ping?

I guess this is OK.  It's a bit shorter-lived than most things we have
added workarounds for, but it at least has an explanatory comment :-)

-- 
Daniel Jacobowitz
CodeSourcery