DWARF2 PE/COFF port and parsing ?

DWARF2 PE/COFF port and parsing ?

[Brian Ford]

I'm still slowly trying to add DWARF2 support for PE/COFF targets.  I have
the following problem that confuses me and I was hoping for some expert
guidance about how to proceed.  gdb's symbol table code is completely
foreign to me.

If I load an executable into gdb that contains one object file with DWARF2
information (all the rest stabs), I get a segfault here:

1431              if (!do_linear_search
1432                  && (SYMBOL_LANGUAGE (*center) == language_java))
1433                {
1434                  do_linear_search = 1;
1435                }

(gdb) p *center
$1 = (struct partial_symbol *) 0x20

(gdb) bt
#0  0x004479b8 in lookup_partial_symbol (pst=0x1052a0a0, name=0x44bf51
"main",
    linkage_name=0x0, global=1, domain=VAR_DOMAIN)
    at ../../gcc/gdb/symtab.c:1431
#1  0x004480b9 in find_main_psymtab () at ../../gcc/gdb/symtab.c:1608
#2  0x0041c11e in set_initial_language () at ../../gcc/gdb/symfile.c:1278
#3  0x0041bac0 in symbol_file_add_main (args=0x22ff1e
"v9win/tap/vital.sym",
    from_tty=0) at ../../gcc/gdb/symfile.c:1017
#4  0x00417978 in do_captured_command (data=0x22fa90)
    at ../../gcc/gdb/top.c:521
#5  0x004178e1 in do_catch_errors (uiout=0x101240a0, data=0x22fa70)
    at ../../gcc/gdb/top.c:492
#6  0x004177a0 in catcher (func=0x4178d0 <do_catch_errors>,
    func_uiout=0x101240a0, func_args=0x22fa70, func_val=0x22fa6c,
    func_caught=0x22fa68, errstring=0x41798d "", mask=6)
    at ../../gcc/gdb/top.c:424
#7  0x0041793a in catch_errors (func=0x417960 <do_captured_command>,
    func_args=0x22fa90, errstring=0x41798d "", mask=6)
    at ../../gcc/gdb/top.c:504
#8  0x004179ca in catch_command_errors (
    command=0x41baa0 <symbol_file_add_main>,
    arg=0x22ff1e "v9win/tap/vital.sym", from_tty=0, mask=6)
    at ../../gcc/gdb/top.c:541
#9  0x00401eeb in captured_main (data=0x22fda8) at
../../gcc/gdb/main.c:654
#10 0x004178e1 in do_catch_errors (uiout=0x5ef090, data=0x22fd60)
    at ../../gcc/gdb/top.c:492
#11 0x004177a0 in catcher (func=0x4178d0 <do_catch_errors>,
    func_uiout=0x5ef090, func_args=0x22fd60, func_val=0x22fd5c,
    func_caught=0x22fd58, errstring=0x401211 "", mask=6)
    at ../../gcc/gdb/top.c:424
#12 0x0041793a in catch_errors (func=0x401420 <captured_main>,
    func_args=0x22fda8, errstring=0x401211 "", mask=6)
    at ../../gcc/gdb/top.c:504
#13 0x00402331 in gdb_main (args=0x22fda8) at ../../gcc/gdb/main.c:813
#14 0x004010da in main (argc=2, argv=0x10101d08) at ../../gcc/gdb/gdb.c:35
#15 0x610068ea in _libuser32_a_iname ()
#16 0x61006b73 in _libuser32_a_iname ()
#17 0x005eca12 in cygwin_crt0 ()
#18 0x0040103c in mainCRTStartup ()
#19 0x77f1bbb5 in _libuser32_a_iname ()

The strange thing is that if I load the object file by itself, all is
well (ie. info types runs to completion).

Any pointers would be greatly appreciated.  Thanks.

-- 
Brian Ford
Senior Realtime Software Engineer
VITAL - Visual Simulation Systems
FlightSafety International
Phone: 314-551-8460
Fax:   314-551-8444

Re: DWARF2 PE/COFF port and parsing ?

[David Carlton]

On Tue, 20 May 2003 16:14:32 -0500 (CDT), Brian Ford <ford@vss.fsi.com> said:

> I'm still slowly trying to add DWARF2 support for PE/COFF targets.  I have
> the following problem that confuses me and I was hoping for some expert
> guidance about how to proceed.  gdb's symbol table code is completely
> foreign to me.

> If I load an executable into gdb that contains one object file with DWARF2
> information (all the rest stabs), I get a segfault here:

> 1431              if (!do_linear_search
> 1432                  && (SYMBOL_LANGUAGE (*center) == language_java))
> 1433                {
> 1434                  do_linear_search = 1;
> 1435                }

The above code really should be deleted from the function in question,
but never mind that; at least it's being a useful canary in the coal
mine.

> (gdb) p *center
> $1 = (struct partial_symbol *) 0x20

That's bad: 0x20 doesn't look like an address.  So the partial symbol
table in question is screwed up.  Unfortunately, by the time you hit
this bug, we've long since built the partial symbol tables.

For an example of building a partial symbol table, see
dwarf2read.c:dwarf2_build_psymtabs_hard.  It gets created with
start_psymtab_common, then some fields get set, then a function gets
called (scan_partial_symbols, in this case) that eventually translates
into a bunch of calls to add_psymbol_to_list to add the actual partial
symbols, then the number of global and static symbols get set, then
sort_pst_symbols gets called.  From your symptoms, maybe the number of
global and/or static symbols got set wrong (or not at all) somewhere.

What does "p *pst" say?  Unfortunately, the bug may have occurred when
another psymtab was being generated: all the psymtabs within one
objfile share one data structure, so an error when generating one
psymtab may corrupt another psymtab's data.  (Sigh.  Partial symbol
tables are a mess.)

If this doesn't help, maybe you could give some more details about
exactly how you've been tinkering with the readers.

David Carlton
carlton@math.stanford.edu

Re: DWARF2 PE/COFF port and parsing ?

[Brian Ford]

Thanks for the Cc.  I am not on the list.

On 20 May 2003, David Carlton wrote:

> On Tue, 20 May 2003 16:14:32 -0500 (CDT), Brian Ford <ford@vss.fsi.com> said:
>
> > I'm still slowly trying to add DWARF2 support for PE/COFF targets.  I have
> > the following problem that confuses me and I was hoping for some expert
> > guidance about how to proceed.  gdb's symbol table code is completely
> > foreign to me.
>
> > If I load an executable into gdb that contains one object file with DWARF2
> > information (all the rest stabs), I get a segfault here:
>
> > 1431              if (!do_linear_search
> > 1432                  && (SYMBOL_LANGUAGE (*center) == language_java))
> > 1433                {
> > 1434                  do_linear_search = 1;
> > 1435                }
>
> The above code really should be deleted from the function in question,
> but never mind that; at least it's being a useful canary in the coal
> mine.
>
> > (gdb) p *center
> > $1 = (struct partial_symbol *) 0x20
>
> That's bad: 0x20 doesn't look like an address.  So the partial symbol
> table in question is screwed up.  Unfortunately, by the time you hit
> this bug, we've long since built the partial symbol tables.
>
Yup.  I know.

> For an example of building a partial symbol table, see
> dwarf2read.c:dwarf2_build_psymtabs_hard.  It gets created with
> start_psymtab_common, then some fields get set, then a function gets
> called (scan_partial_symbols, in this case) that eventually translates
> into a bunch of calls to add_psymbol_to_list to add the actual partial
> symbols, then the number of global and static symbols get set, then
> sort_pst_symbols gets called.  From your symptoms, maybe the number of
> global and/or static symbols got set wrong (or not at all) somewhere.
>
Thanks for the road map.  That was a good portion of what I was looking
for.  I'll have to trace down the chain.  Your assumption of what the
problem is sounds pretty good to me.

> What does "p *pst" say?  Unfortunately, the bug may have occurred when
> another psymtab was being generated: all the psymtabs within one
> objfile share one data structure, so an error when generating one
> psymtab may corrupt another psymtab's data.  (Sigh.  Partial symbol
> tables are a mess.)
>
(gdb) p *pst
Cannot access memory at address 0x0
(gdb) p pst
$1 = (struct partial_symtab *) 0x0

> If this doesn't help, maybe you could give some more details about
> exactly how you've been tinkering with the readers.
>
Actually, I haven't.  The tinkering I have done is in binutils to get
valid DWARF2 objects and let bfd know that PE/COFF can produce them.  gdb
is current cvs.  I have more tinkering to go, but the objects and
executables should be valid.  libdwarf and dwarfdump have no trouble
parsing them.

Thanks for the help and let me know if you think of anything to add.  I'll
try to start the chase in the psymtab building now that I have a very
fuzzy road map.

-- 
Brian Ford
Senior Realtime Software Engineer
VITAL - Visual Simulation Systems
FlightSafety International
Phone: 314-551-8460
Fax:   314-551-8444

Re: DWARF2 PE/COFF port and parsing ?

[Brian Ford]

On Tue, 20 May 2003, Brian Ford wrote:

> > What does "p *pst" say?  Unfortunately, the bug may have occurred when
> > another psymtab was being generated: all the psymtabs within one
> > objfile share one data structure, so an error when generating one
> > psymtab may corrupt another psymtab's data.  (Sigh.  Partial symbol
> > tables are a mess.)
> >
> (gdb) p *pst
> Cannot access memory at address 0x0
> (gdb) p pst
> $1 = (struct partial_symtab *) 0x0
>
Um, sorry.  That obviously looks wrong because of optimization.  Here is
what it says without -O2:

(gdb) p *pst
$1 = {next = 0x1052a018, filename = 0x1052a0e8 "../../gcc/gcc/libgcc2.c",
  fullname = 0x0, objfile = 0x10148238, section_offsets = 0x1014e440,
  textlow = 4922848, texthigh = 4922848, dependencies = 0x1052a118,
  number_of_dependencies = 4, globals_offset = 1151, n_global_syms = 2,
  statics_offset = 24280, n_static_syms = 21, symtab = 0x0,
  read_symtab = 0x4f1170 <dbx_psymtab_to_symtab>,
  read_symtab_private = 0x1052a100 "D\030\031", readin = 0 '\0'}

-- 
Brian Ford
Senior Realtime Software Engineer
VITAL - Visual Simulation Systems
FlightSafety International
Phone: 314-551-8460
Fax:   314-551-8444

Re: DWARF2 PE/COFF port and parsing ?

[David Carlton]

On Tue, 20 May 2003 16:56:52 -0500 (CDT), Brian Ford <ford@vss.fsi.com> said:

>> What does "p *pst" say?

> (gdb) p *pst
> Cannot access memory at address 0x0
> (gdb) p pst
> $1 = (struct partial_symtab *) 0x0

That's unfortunate; maybe the optimizer has already invalidated pst at
that point?  You could try finding an earlier place in the command
where pst is still valid and have GDB print out the value of pst each
time it hits that location (which will happen a lot), and then, when
GDB crashes, look at the last value for pst and print out 
*((struct partial_sytmab *) <insert last value>)

Maybe that will give a clue as to what's going on; maybe it won't.

>> If this doesn't help, maybe you could give some more details about
>> exactly how you've been tinkering with the readers.

> Actually, I haven't.  The tinkering I have done is in binutils to get
> valid DWARF2 objects and let bfd know that PE/COFF can produce them.  gdb
> is current cvs.  I have more tinkering to go, but the objects and
> executables should be valid.  libdwarf and dwarfdump have no trouble
> parsing them.

Weird.  I'm pretty confused, but maybe following through the
psymtab building will help.

David Carlton
carlton@math.stanford.edu

Re: DWARF2 PE/COFF port and parsing ?

[David Carlton]

On Tue, 20 May 2003 17:05:26 -0500 (CDT), Brian Ford <ford@vss.fsi.com> said:

> On Tue, 20 May 2003, Brian Ford wrote:
>> > What does "p *pst" say?  Unfortunately, the bug may have occurred when
>> > another psymtab was being generated: all the psymtabs within one
>> > objfile share one data structure, so an error when generating one
>> > psymtab may corrupt another psymtab's data.  (Sigh.  Partial symbol
>> > tables are a mess.)
>> >
>> (gdb) p *pst
>> Cannot access memory at address 0x0
>> (gdb) p pst
>> $1 = (struct partial_symtab *) 0x0
>> 
> Um, sorry.  That obviously looks wrong because of optimization.  Here is
> what it says without -O2:

> (gdb) p *pst
> $1 = {next = 0x1052a018, filename = 0x1052a0e8 "../../gcc/gcc/libgcc2.c",
>   fullname = 0x0, objfile = 0x10148238, section_offsets = 0x1014e440,
>   textlow = 4922848, texthigh = 4922848, dependencies = 0x1052a118,
>   number_of_dependencies = 4, globals_offset = 1151, n_global_syms = 2,
>   statics_offset = 24280, n_static_syms = 21, symtab = 0x0,
>   read_symtab = 0x4f1170 <dbx_psymtab_to_symtab>,
>   read_symtab_private = 0x1052a100 "D\030\031", readin = 0 '\0'}

Hmm.  It doesn't look too strange to me.  Are the libraries that GCC
builds being compiled using your DWARF 2 in PE/COFF mechanisms?  If
not, is there any way you can think of that that would cause problems?

David Carlton
carlton@math.stanford.edu

Re: DWARF2 PE/COFF port and parsing ?

[Brian Ford]

On 21 May 2003, Jim Blandy wrote:

> You may find the 'maint info psymtab' and 'maint info symtab'
> functions helpful in debugging this kind of thing.  The output
> includes addresses for the objects it describes.
>
Thanks.  The problem is, I crash as soon as I load the symbol file while
looking up main.  So, the command interface isn't very helpful.

-- 
Brian Ford
Senior Realtime Software Engineer
VITAL - Visual Simulation Systems
FlightSafety International
Phone: 314-551-8460
Fax:   314-551-8444

Re: DWARF2 PE/COFF port and parsing ?

[Brian Ford]

Sorry.  Somehow this missed the list and went straight to David.

On Tue, 20 May 2003, Brian Ford wrote:

> On 20 May 2003, David Carlton wrote:
>
> > On Tue, 20 May 2003 17:05:26 -0500 (CDT), Brian Ford <ford@vss.fsi.com> said:
> >
> > > (gdb) p *pst
> > > $1 = {next = 0x1052a018, filename = 0x1052a0e8 "../../gcc/gcc/libgcc2.c",
> > >   fullname = 0x0, objfile = 0x10148238, section_offsets = 0x1014e440,
> > >   textlow = 4922848, texthigh = 4922848, dependencies = 0x1052a118,
> > >   number_of_dependencies = 4, globals_offset = 1151, n_global_syms = 2,
> > >   statics_offset = 24280, n_static_syms = 21, symtab = 0x0,
> > >   read_symtab = 0x4f1170 <dbx_psymtab_to_symtab>,
> > >   read_symtab_private = 0x1052a100 "D\030\031", readin = 0 '\0'}
> >
> > Hmm.  It doesn't look too strange to me.  Are the libraries that GCC
> > builds being compiled using your DWARF 2 in PE/COFF mechanisms?  If
> > not, is there any way you can think of that that would cause problems?
> >
> No, it must be a case of corruption as they are still in stabs.
>
> Here is my object's pst at dwarf2read.c:1354:
>
> (top-gdb) p *pst
> $1 = {next = 0x1052a4e0,
>   filename = 0x105039d8 "../../../visual/vital/buf_init/bufs.c",
>   fullname = 0x0, objfile = 0x10148540, section_offsets = 0x1014e748,
>   textlow = 0, texthigh = 0, dependencies = 0x0, number_of_dependencies =
> 0,
>   globals_offset = 0, n_global_syms = 42, statics_offset = 0,
>   n_static_syms = 420, symtab = 0x0,
>   read_symtab = 0x542d40 <dwarf2_psymtab_to_symtab>,
>   read_symtab_private = 0x10503a00 " ¸R\020", readin = 0 '\0'}
>
> The one that appears corrupt is *pst->next->next:
>
> (top-gdb) p *pst->next
> $4 = {next = 0x1052a460, filename = 0x1052a528
> "../../gcc/gcc/gbl-ctors.h",
>   fullname = 0x0, objfile = 0x10148540, section_offsets = 0x1014e748,
>   textlow = 0, texthigh = 0, dependencies = 0x1052a560,
>   number_of_dependencies = 1, globals_offset = 0, n_global_syms = 0,
>   statics_offset = 0, n_static_syms = 0, symtab = 0x0,
>   read_symtab = 0x4f1170 <dbx_psymtab_to_symtab>,
>   read_symtab_private = 0x1052a548 "", readin = 0 '\0'}
> (top-gdb) p *pst->next->next
> $5 = {next = 0x1052a3d8, filename = 0x1052a4a8 "../../gcc/gcc/libgcc2.c",
>   fullname = 0x0, objfile = 0x10148540, section_offsets = 0x1014e748,
>   textlow = 0, texthigh = 0, dependencies = 0x1052a4d8,
>   number_of_dependencies = 1, globals_offset = 0, n_global_syms = 0,
>   statics_offset = 0, n_static_syms = 0, symtab = 0x0,
>   read_symtab = 0x4f1170 <dbx_psymtab_to_symtab>,
>   read_symtab_private = 0x1052a4c0 "", readin = 0 '\0'}
>

-- 
Brian Ford
Senior Realtime Software Engineer
VITAL - Visual Simulation Systems
FlightSafety International
Phone: 314-551-8460
Fax:   314-551-8444