[Bug bootstrap/53300] New: AIX bootstrap related to varpool patch

[Bug bootstrap/53300] New: AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

             Bug #: 53300
           Summary: AIX bootstrap related to varpool patch
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: bootstrap
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: dje@gcc.gnu.org


The recent change to cgraph_remove_unreachable_nodes causes AIX bootstrap
failure because TOC symbols are removed.

ld: 0711-317 ERROR: Undefined symbol: LC..0
ld: 0711-317 ERROR: Undefined symbol: LC..1
ld: 0711-317 ERROR: Undefined symbol: LC..2
ld: 0711-317 ERROR: Undefined symbol: LC..3
ld: 0711-317 ERROR: Undefined symbol: LC..4

These are emitted in output_toc(). I am not sure what dependency is missing to
allow cgraph to believe the symbol is unused. Although both AIX and pLinux use
TARGET_TOC, this does not fail on Linux.

[Bug bootstrap/53300] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-05-09
     Ever Confirmed|0                           |1

--- Comment #1 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-09 19:38:37 UTC ---
Sorry, I referenced the wrong patch.  This is related to
Symbol table 19/many: cleanup varpool/front-end/varasm interactions

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |*-*-aix
   Target Milestone|---                         |4.8.0
            Summary|AIX bootstrap related to    |[4.8 Regression] AIX
                   |varpool patch               |bootstrap related to
                   |                            |varpool patch

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[ro at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

Rainer Orth <ro at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ro at gcc dot gnu.org

--- Comment #2 from Rainer Orth <ro at gcc dot gnu.org> 2012-05-10 12:15:13 UTC ---
I'm seeing a similar failure linking gnatxref on i386-pc-solaris2.10:

Undefined            first referenced
 symbol                  in file
.LC0                                ./xref_lib.o
ld: fatal: symbol referencing errors. No output written to ../../gnatxref
collect2: error: ld returned 1 exit status
gnatlink: error when calling /var/gcc/gcc-4.8.0-20120509/10-gcc/gcc/xgcc
make[3]: *** [common-tools] Error 4
make[3]: Leaving directory `/var/gcc/gcc-4.8.0-20120509/10-gcc/gcc/ada/tools'
make[2]: *** [gnattools-native] Error 2
make[2]: Leaving directory `/var/gcc/gcc-4.8.0-20120509/10-gcc/gnattools'

Reverting your patch allows the build to complete.

  Rainer

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |assemble-failure
             Target|*-*-aix                     |powerpc-ibm-aix*,
                   |                            |i386-pc-solaris2.10

--- Comment #3 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-10 13:12:45 UTC ---
If the solaris problem is the same root cause, hopefully it will be easier to
reproduce.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED

--- Comment #4 from Jan Hubicka <hubicka at gcc dot gnu.org> 2012-05-10 17:17:01 UTC ---
Would be possible to have testcase with an explanation what is missing?

The TOC symbols are output by output_constant_pool_1 that calls
ASM_OUTPUT_SPECIAL_POOL_ENTRY.  I am not sure how varpool variable removal is
involved here, since TOC never lands the varpool.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #5 from Jan Hubicka <hubicka at gcc dot gnu.org> 2012-05-10 17:59:46 UTC ---
i am still somewhat in dark on root of this problem.  What may actually be
problem here is the fact that varpool no longer outputs constant pool decls
relying on constant pool code to manage them on its own.

Does the following hack avoid the problem? Perhaps during the years when
varpool
was outputting constant pool vars something broke in the code tracking when the
var is needed.

Index: varpool.c
===================================================================
--- varpool.c   (revision 187371)
+++ varpool.c   (working copy)
@@ -282,7 +282,8 @@ varpool_assemble_decl (struct varpool_no

   /* Constant pool is output from RTL land when the reference
      survive till this level.  */
-  if (DECL_IN_CONSTANT_POOL (decl))
+  if (DECL_IN_CONSTANT_POOL (decl)
+      && TREE_ASM_WRITTEN (decl))
     return false;

   /* Decls with VALUE_EXPR should not be in the varpool at all.  They

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #6 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-10 18:37:12 UTC ---
The first failure occurs for me when linking gcov, but I am not sure which
object file is missing the symbols.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #7 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-10 20:27:52 UTC ---
Your patch in comment #5 seems to work around the problem. I will try to figure
out which file omits the symbols.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[hubicka at ucw dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #8 from Jan Hubicka <hubicka at ucw dot cz> 2012-05-10 20:48:52 UTC ---
> Your patch in comment #5 seems to work around the problem. I will try to figure
> out which file omits the symbols.

Thank you! If you could test it fully, I will commit it for time being to
unbreak the
bootstrap.  We can solve this problem incrementally.

Honza

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot
                   |                            |gnu.org

--- Comment #9 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2012-05-10 22:36:34 UTC ---
> I'm seeing a similar failure linking gnatxref on i386-pc-solaris2.10:
> 
> Undefined            first referenced
>  symbol                  in file
> .LC0                                ./xref_lib.o

Likewise on x86/Linux.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[hubicka at ucw dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #10 from Jan Hubicka <hubicka at ucw dot cz> 2012-05-10 23:03:44 UTC ---
> Likewise on x86/Linux.
Bootstrap with Ada works for me, but I will test the workaround posted above
and commit if it passes.  We can deal with this incrementally.  Apparently
there is some latent bug in the constant pool handling code...

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #11 from Jan Hubicka <hubicka at gcc dot gnu.org> 2012-05-11 08:40:25 UTC ---
Author: hubicka
Date: Fri May 11 08:40:15 2012
New Revision: 187397

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=187397
Log:

    PR bootstrap/53300
    * varpool.c (varpool_assemble_decl): Also output constat pool entries
    that output_constant_pool missed.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/varpool.c

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[ro at CeBiTec dot Uni-Bielefeld.DE]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #12 from ro at CeBiTec dot Uni-Bielefeld.DE <ro at CeBiTec dot Uni-Bielefeld.DE> 2012-05-11 10:08:11 UTC ---
> --- Comment #5 from Jan Hubicka <hubicka at gcc dot gnu.org> 2012-05-10 17:59:46 UTC ---
[...]
> Does the following hack avoid the problem? Perhaps during the years when
> varpool
> was outputting constant pool vars something broke in the code tracking when the
> var is needed.

It does on i386-pc-solaris2.10.  Ada bootstrap completed with it.

Thanks.
        Rainer

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #15 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-11 15:09:09 UTC ---
Created attachment 27379
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27379
Wrong assembly output for line-map.c

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #13 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-11 15:07:46 UTC ---
Created attachment 27377
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27377
Pre-processed source for libcpp/line-map.c

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #14 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-11 15:08:29 UTC ---
Created attachment 27378
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27378
Correct assembly output for line-map.c

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #16 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-11 15:11:59 UTC ---
The mis-compiled file is libcpp/line-map.c.  I have attached the pre-processed
output for line-map.c on AIX. I also have attached assembly files showing the
correct and incorrect output.  The difference is:

--- line-map.s.good     2012-05-11 11:01:33.000000000 -0400
+++ line-map.s.bad      2012-05-11 11:00:43.000000000 -0400
@@ -4536,26 +4536,6 @@
        .byte 10
        .byte "Macro line maps"
        .byte 10, 0
-       .space 2
-LC..0:
-       .byte "LC_ENTER"
-       .byte 0
-       .space 3
-LC..1:
-       .byte "LC_LEAVE"
-       .byte 0
-       .space 3
-LC..2:
-       .byte "LC_RENAME"
-       .byte 0
-       .space 2
-LC..3:
-       .byte "LC_RENAME_VERBATIM"
-       .byte 0
-       .space 1
-LC..4:
-       .byte "LC_ENTER_MACRO"
-       .byte 0
        .csect _linemap.rw_[RW],4
        .align 2
        .set LANCHOR..1,$ + 0

which corresponds to

enum lc_reason
{
  LC_ENTER = 0,
  LC_LEAVE,
  LC_RENAME,
  LC_RENAME_VERBATIM,
  LC_ENTER_MACRO

};

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #17 from David Edelsohn <dje at gcc dot gnu.org> 2012-05-11 15:44:44 UTC ---
A minimal testcase:

enum lc_reason
{
  LC_ENTER = 0,
  LC_LEAVE,
  LC_RENAME,
  LC_RENAME_VERBATIM,
  LC_ENTER_MACRO

};

const char *
foo (enum lc_reason reason)
{
  const char *lc_reasons_v[LC_ENTER_MACRO + 1]
      = { "LC_ENTER", "LC_LEAVE", "LC_RENAME", "LC_RENAME_VERBATIM",
   "LC_ENTER_MACRO" };

  return lc_reasons_v[reason];
}


Without your TREE_ASM_WRITTEN kludge, the string values of the array are not
emitted in the assembly. The array referencing the strings is emitted, but the
values themselves are not.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P1

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #18 from Andrew Pinski <pinskia at gcc dot gnu.org> 2012-11-15 03:18:18 UTC ---
Does this still happen?

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[dje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #19 from David Edelsohn <dje at gcc dot gnu.org> 2012-11-15 14:04:10 UTC ---
The problem still is present, but when I asked Honza about the status on IRC,
he seemed to say the work-around patch was the correct fix -- that this
behavior is expected.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #20 from Richard Biener <rguenth at gcc dot gnu.org> 2013-01-15 13:54:58 UTC ---
Thus, fixed.

[Bug bootstrap/53300] [4.8 Regression] AIX bootstrap related to varpool patch

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53300

--- Comment #21 from Jan Hubicka <hubicka at gcc dot gnu.org> 2013-01-15 18:23:24 UTC ---
Yep, it is kind of kludgy, but other way to handle it is to rewrite the
constant pool that is tied into target interface so it is rather hard and I do
not see constant pool really being trouble to anything - it works reasonably
well the odd way it is designed.