Patchwork [PR44248] Fix lto/whopr -g link-time -fcompare-debug failure

login
register
mail settings
Submitter Alexandre Oliva
Date June 17, 2010, 7:07 p.m.
Message ID <or4oh1jx4q.fsf@livre.localdomain>
Download mbox | patch
Permalink /patch/56082/
State New
Headers show

Comments

Alexandre Oliva - June 17, 2010, 7:07 p.m.
On Jun 15, 2010, Richard Guenther <richard.guenther@gmail.com> wrote:

> On Tue, Jun 15, 2010 at 5:00 AM, Alexandre Oliva <aoliva@redhat.com> wrote:
>> When VTA is disabled, LTO streams in debug stmts as gimple nops.  This
>> is inadequate, for gimple nops will get in the way of removal of blocks
>> and perhaps other situations in which debug stmts are properly
>> disregarded.  Their presence is enough to bring about codegen
>> differences, causing -fcompare-debug failures when an object file
>> compiled with -g -flto is to be link-time compiled without -g, as is the
>> case when -fcompare-debug is given.
>> 
>> This patch arranges for us to drop the debug stmts entirely, but to do
>> so only after applying fixups.  We could drop them a bit earlier, while
>> filling in the stmts array, but then we'd have to worry about NULL stmt
>> entries in the array (should debug stmts have fixups) and the array
>> would probably be slower, for we'd have to test MAY_HAVE_DEBUG_STMTS
>> multiple times.
>> 
>> Is this ok, if it passes regstrap on x86_64-linux-gnu?

> Ok if you make input_bb return a flag whether a debug stmt was
> streamed in, aggregate that and make the loop over all stmts
> conditional on that to avoid walking over all stmts in a function
> one more time.

How about piggibacking on another existing loop over all stmts, instead?
I tried this with an -flto -fcompare-debug bootstrap and that worked
fine on i686-pc-linux-gnu.  On other platforms, there were
-fcompare-debug failures, but no crashes, so removing the stmts at that
point is fine, we just have to keep them long enough for the uids to be
computed as expected.  Leave the corresponding entries in the stmts
array as NULL is fine.

Ok to install?
Richard Guenther - June 18, 2010, 9:20 a.m.
On Thu, Jun 17, 2010 at 9:07 PM, Alexandre Oliva <aoliva@redhat.com>
wrote:
> On Jun 15, 2010, Richard Guenther <richard.guenther@gmail.com> wrote:
>
>> On Tue, Jun 15, 2010 at 5:00 AM, Alexandre Oliva <aoliva@redhat.com> wrote:
>>> When VTA is disabled, LTO streams in debug stmts as gimple nops.  This
>>> is inadequate, for gimple nops will get in the way of removal of blocks
>>> and perhaps other situations in which debug stmts are properly
>>> disregarded.  Their presence is enough to bring about codegen
>>> differences, causing -fcompare-debug failures when an object file
>>> compiled with -g -flto is to be link-time compiled without -g, as is the
>>> case when -fcompare-debug is given.
>>>
>>> This patch arranges for us to drop the debug stmts entirely, but to do
>>> so only after applying fixups.  We could drop them a bit earlier, while
>>> filling in the stmts array, but then we'd have to worry about NULL stmt
>>> entries in the array (should debug stmts have fixups) and the array
>>> would probably be slower, for we'd have to test MAY_HAVE_DEBUG_STMTS
>>> multiple times.
>>>
>>> Is this ok, if it passes regstrap on x86_64-linux-gnu?
>
>> Ok if you make input_bb return a flag whether a debug stmt was
>> streamed in, aggregate that and make the loop over all stmts
>> conditional on that to avoid walking over all stmts in a function
>> one more time.
>
> How about piggibacking on another existing loop over all stmts, instead?
> I tried this with an -flto -fcompare-debug bootstrap and that worked
> fine on i686-pc-linux-gnu.  On other platforms, there were
> -fcompare-debug failures, but no crashes, so removing the stmts at that
> point is fine, we just have to keep them long enough for the uids to be
> computed as expected.  Leave the corresponding entries in the stmts
> array as NULL is fine.
>
> Ok to install?

Ok if you add a comment like /* Remove debug statements only
here to preserve gimple uid numbering.  */

Thanks,
Richard.

>
>
> --
> Alexandre Oliva, freedom fighter    http://FSFLA.org/~lxoliva/
> You must be the change you wish to see in the world. -- Gandhi
> Be Free! -- http://FSFLA.org/   FSF Latin America board member
> Free Software Evangelist      Red Hat Brazil Compiler Engineer
>
>

Patch

for  gcc/ChangeLog
from  Alexandre Oliva  <aoliva@redhat.com>

	PR debug/44248
	* lto-streamer-in.c (input_bb): Leave debug stmts alone.
	(input_function): Drop them here, if VTA is disabled.
	
Index: gcc/lto-streamer-in.c
===================================================================
--- gcc/lto-streamer-in.c.orig	2010-06-16 17:17:46.000000000 -0300
+++ gcc/lto-streamer-in.c	2010-06-16 17:21:33.000000000 -0300
@@ -1204,13 +1204,6 @@  input_bb (struct lto_input_block *ib, en
     {
       gimple stmt = input_gimple_stmt (ib, data_in, fn, tag);
 
-      /* Change debug stmts to nops on-the-fly if we do not have VTA enabled.
-	 This allows us to build for example static libs with debugging
-	 enabled and do the final link without.  */
-      if (!MAY_HAVE_DEBUG_STMTS
-	  && is_gimple_debug (stmt))
-	stmt = gimple_build_nop ();
-
       find_referenced_vars_in (stmt);
       gsi_insert_after (&bsi, stmt, GSI_NEW_STMT);
 
@@ -1370,11 +1363,21 @@  input_function (tree fn_decl, struct dat
   stmts = (gimple *) xcalloc (gimple_stmt_max_uid (fn), sizeof (gimple));
   FOR_ALL_BB (bb)
     {
-      gimple_stmt_iterator bsi;
-      for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+      gimple_stmt_iterator bsi = gsi_start_bb (bb);
+      while (!gsi_end_p (bsi))
 	{
 	  gimple stmt = gsi_stmt (bsi);
-	  stmts[gimple_uid (stmt)] = stmt;
+	  if (!MAY_HAVE_DEBUG_STMTS && is_gimple_debug (stmt))
+	    {
+	      gimple_stmt_iterator gsi = bsi;
+	      gsi_next (&bsi);
+	      gsi_remove (&gsi, true);
+	    }
+	  else
+	    {
+	      gsi_next (&bsi);
+	      stmts[gimple_uid (stmt)] = stmt;
+	    }
 	}
     }