[Bug debug/46724] New: Wrong debug info: Invalid variable location

[Bug debug/46724] New: Wrong debug info: Invalid variable location

[krebbel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

           Summary: Wrong debug info: Invalid variable location
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: major
          Priority: P3
         Component: debug
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: krebbel@gcc.gnu.org


The guality/nrv-1.c testcase currently fails on s390x when compiled with -O2:

Breakpoint 1, f () at /build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c:20
20        a2.i[4] = 7;  /* { dg-final { gdb-test 20 "a2.i\[0\]" "42" } } */
Missing separate debuginfos, use: debuginfo-install glibc-2.12.1-2.s390x
(gdb) p a2.i[0]
$1 = -336596880

The memory location of 'a2' is given as being based on r1:
readelf --debug-dump=info
...
 <2><85>: Abbrev Number: 9 (DW_TAG_variable)
    <86>   DW_AT_name        : a2
    <89>   DW_AT_decl_file   : 1
    <8a>   DW_AT_decl_line   : 16
    <8b>   DW_AT_type        : <0x2d>
    <8f>   DW_AT_location    : 2 byte block: 71 0       (DW_OP_breg1: 0)
...

But in fact r1 is not used at all. Instead the location is in r2:

f:
.LFB0:
        .file 1 "/build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c"
        .loc 1 15 0
        .cfi_startproc
        .loc 1 17 0
        lhi     %r3,42
        st      %r3,0(%r2)    <-- a2.i[0] = 42;
        ...

a2 is returned by f. The address of the location where the return value should
reside is passed to f as implicit first parameter in r2.

Right before cprop_hardreg the code looks like this. r2 is copied into r1 and
r1 is used to reference a2. DECL_RTL (<retval>) is (mem/s:SI (reg/f:DI 1 %r1
[orig:45 D.1986 ] [45]).

nrv-1.c.201r.ce3:

(insn 2 8 47 2 (set (reg/f:DI 1 %r1 [orig:45 D.1986 ] [45])
        (reg:DI 2 %r2 [ D.1986 ]))
/build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c:15 62 {*movdi_64}
     (nil))

(insn 47 2 6 2 (set (reg:SI 3 %r3)
        (const_int 42 [0x2a]))
/build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c:17 66 {*movsi_zarch}
     (nil))

(insn 6 47 45 2 (set (mem/s:SI (reg/f:DI 1 %r1 [orig:45 D.1986 ] [45]) [2
<retval>.i+0 S4 A32])
        (reg:SI 3 %r3))
/build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c:17 66 {*movsi_zarch}
     (expr_list:REG_DEAD (reg:SI 3 %r3)
        (nil)))

Then r1 is propagated into insn 6 and insn 2 is removed in DCE while DECL_RTL
(<retval>) still references r1.

nrv-1.c.203r.cprop_hardreg:

(insn 2 8 47 2 (set (reg/f:DI 1 %r1 [orig:45 D.1986 ] [45])
        (reg:DI 2 %r2 [ D.1986 ]))
/build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c:15 62 {*movdi_64}
     (nil))

(insn 47 2 6 2 (set (reg:SI 3 %r3)
        (const_int 42 [0x2a]))
/build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c:17 66 {*movsi_zarch}
     (nil))

(insn 6 47 45 2 (set (mem/s:SI (reg/f:DI 2 %r2 [orig:45 D.1986 ] [45]) [2
<retval>.i+0 S4 A32])
        (reg:SI 3 %r3))
/build3/gcc-head/gcc/testsuite/gcc.dg/guality/nrv-1.c:17 66 {*movsi_zarch}
     (expr_list:REG_DEAD (reg:SI 3 %r3)
        (nil)))

When generating the debug info for a2 add_location_or_const_value_attribute
first tries to find an RTX for a2 directly by invoking rtl_for_decl_location p
on the VAR_DECL of a2 what returns NULL. It then invokes loc_list_from_tree
which uses the DECL_VALUE_EXPR of the a2 decl - this is <retval>.
rtl_for_decl_location then returns (mem/s:SI (reg/f:DI 1 %r1 [orig:45 D.1986 ]
[45]) for <retval>.

[Bug debug/46724] Wrong debug info: Invalid variable location

[krebbel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

Andreas Krebbel <krebbel at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |s390x-ibm-linux
           Priority|P3                          |P2
           Keywords|                            |wrong-debug
               Host|                            |s390x-ibm-linux
   Target Milestone|---                         |4.6.0
      Known to fail|                            |4.6.0
              Build|                            |s390x-ibm-linux

[Bug debug/46724] Wrong debug info: Invalid variable location

[krebbel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #1 from Andreas Krebbel <krebbel at gcc dot gnu.org> 2010-11-30 14:37:39 UTC ---
Compiling with
GCC 4.5.2 (gcc version 4.5.2 20101108 (prerelease) [gcc-4_5-branch revision
166433] (GCC))

GDB returns "<value optimized out>" for 'a2' everywhere in function 'f'.

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[krebbel at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #2 from Andreas Krebbel <krebbel at gcc dot gnu.org> 2010-12-02 09:51:17 UTC ---
Marked as 4.6 regression. The behavior of 4.5.2 isn't helpful either but it
doesn't return a wrong value as with 4.6.

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.12.02 12:11:48
                 CC|                            |aoliva at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org
     Ever Confirmed|0                           |1

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-02 12:11:48 UTC ---
I guess one option would be:
1) allow DECL_DEBUG_EXPR even on RESULT_DECL (i.e. not use VAR_DECL_CHECK
there)
2) when creating function_result_decl in assign_parms_augmented_arg_list,
   either don't mark that artificial PARM_DECL with DECL_IGNORED_P but
   DECL_NAMELESS and put it into DECL_ARGUMENTS (but then it might confuse
   debuggers), or create a separate artificial DECL_NAMELESS DECL_BY_REFERENCE
   variable which will be put into outermost BLOCK.
3) in assign_parms, after assign_parm_setup_reg initializes it, we could
   emit a DEBUG_INSN for this artificial variable, equating it to the
artificial
   PARM_DECL's DECL_RTL
4) and set DECL_DEBUG_EXPR of the RESULT_DECL to this artificial VAR_DECL
The DECL_BY_REFERENCE thingie is there because var-tracking doesn't VALUE track
aggregates, so we just want to track its address.

Alex, what do you think about this?

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #4 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-12-16 05:38:44 UTC ---
Considering that we create a decl for the implicit argument that holds the
address of <result>, perhaps the best approach would be to emit debug info for
that artificial decl.  The initial implicit binding of argument variables to
their incoming locations should take care of the debug info for the artificial
parameter throughout the function, and the location of <result> could then be
given in terms of the artificial variable (dwarf calls).  Does that sound
sensible?

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

Alexandre Oliva <aoliva at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot       |aoliva at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #5 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-12-17 06:13:57 UTC ---
Mine

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #6 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-12-17 06:19:30 UTC ---
Created attachment 22792
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22792
Patch I'm testing that fixes the bug

Here's what I meant.  It does fix the bug at hand, and (surprisingly) we don't
end up emitting debug info for the artificial parameter.  Tracking it is enough
to fix the location info for the variable that maps to the result.  It looks
like this:

        .byte   0x8     # uleb128 0x8; (DIE (0x15b) DW_TAG_variable)
        .ascii "a2\0"   # DW_AT_name
        .byte   0x1     # DW_AT_decl_file
(../../gcc/testsuite/gcc.dg/guality/nrv-1.c)
        .byte   0x10    # DW_AT_decl_line
        .4byte  0xf5    # DW_AT_type
        .4byte  .LLST1  # DW_AT_location

.LLST1:
        .8byte  .LVL0   # Location list begin address (*.LLST1)
        .8byte  .LVL1-1 # Location list end address (*.LLST1)
        .2byte  0x2     # Location expression size
        .byte   0x72    # DW_OP_breg2
        .byte   0       # sleb128 0
        .8byte  0       # Location list terminator begin (*.LLST1)
        .8byte  0       # Location list terminator end (*.LLST1)

The range covers the entire function.  I haven't looked into why we don't
reduce the single-entry entire-function location list to a simpler location.

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-17 08:55:06 UTC ---
Ah, the trick is that you keep the artificial PARM_DECL out of DECL_ARGUMENTS
as it was and manually add it in var-tracking.
To avoid dropping the DECL_NAME test you could perhaps give some artificial
name to this PARM_DECL and make it DECL_NAMELESS.

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-12-17 09:44:03 UTC ---
Created attachment 22794
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22794
gcc46-pr46724-incremental.patch

Incremental patch for that.

BTW, the reason why a location list is used is because it is needed - it
doesn't cover the whole function, when inside of the abort call r2 might be
clobbered within the abort call.  Therefore the range ends at .LVL1 - 1 instead
of .LVL1 (end of function).

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

Alexandre Oliva <aoliva at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #22792|0                           |1
        is obsolete|                            |
  Attachment #22794|0                           |1
        is obsolete|                            |

--- Comment #9 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-12-18 08:02:42 UTC ---
Created attachment 22806
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22806
Revised patch

This patch integrates Jakub's proposed changes and fixes a regression in
pr46815.C on x86_64-linux-gnu.  A difference in DECL_BY_REFERENCE in the
RESULT_DECL required different handling in the DEBUG_VALUE_EXPR and in
var-tracking for proper handling.  Regstrapping again after the changes...

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

Alexandre Oliva <aoliva at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #22806|0                           |1
        is obsolete|                            |

--- Comment #10 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-12-18 08:05:20 UTC ---
Created attachment 22807
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22807
The actual revised patch

Doh, failed to refresh the patch file.

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #11 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-12-22 03:49:06 UTC ---
Author: aoliva
Date: Wed Dec 22 03:49:00 2010
New Revision: 168160

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=168160
Log:
PR debug/46724
* function.c (assign_parms_augmented_arg_list): Name and mark
DECL of result address as NAMELESS rather than IGNORED.
(assign_parms): Set DECL_VALUE_EXPR for indirect result.
* tree.h (tree_decl_common::decl_flag_2): Document RESULT_DECL.
(DECL_HAS_VALUE_EXPR_P): Accept RESULT_DECL.
* dwarf2out.c (loc_list_from_tree) <RESULT_DECL>: Use
DECL_VALUE_EXPR.
* dbxout.c (dbxout_expand_expr) <RESULT_DECL>: Likewise.
* var-tracking.c (vt_add_function_parameter): New, split out of...
(vt_add_function_parameters): ... this.  Handle incoming
pointer to hold result.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/dbxout.c
    trunk/gcc/dwarf2out.c
    trunk/gcc/function.c
    trunk/gcc/tree.h
    trunk/gcc/var-tracking.c

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

Alexandre Oliva <aoliva at gcc dot gnu.org> changed:

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

--- Comment #12 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2010-12-22 03:55:54 UTC ---
Fixed

[Bug debug/46724] [4.6 Regression] Wrong debug info: Invalid variable location

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46724

--- Comment #13 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-01-19 22:00:09 UTC ---
Author: aoliva
Date: Wed Jan 19 22:00:00 2011
New Revision: 169034

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=169034
Log:
PR debug/47079
PR debug/46724
* function.c (instantiate_expr): Instantiate incoming rtl of
implicit arguments, and recurse on VALUE_EXPRs.
(instantiate_decls): Instantiate rtl and VALUE_EXPR of result.
* var-tracking.c (adjust_mems): Reject virtual_incoming_args_rtx.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/function.c
    trunk/gcc/var-tracking.c