Re: [PATCH] Fix call functions command bug in 64-bit programs for AIX and PC read in psymtab-symtab warning

[Aditya Kamath1 <Aditya.Kamath1@ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 5115 bytes --]

Hi Ulrich, Tom and community,

>My understanding, and that is apparently confirmed by the
>text in the GNU stabs documentation here:
>https://sourceware.org/gdb/download/onlinedocs/stabs/Transformations-On-Global-Variables.html
>is that the "address" va lue in the debug entry is not used,
>but the actual address should be taken from the .text entry.

>Now I'm not sure why this doesn't work correctly, but this
>may be a place to investigate further.

>(One interesting thing to note is that the names do not match
>exactly, as the .text entry has the leading '.' that is used
>on ppc64 ... maybe this throws off some of the stabs logic?)

Thank you so much for this explanation. So here is our investigation further.

This __new_exitfn is coming from libgcc. Libgcc is -g compiled with stabs debugging format.

bash-5.1$ dump -tov /opt/freeware/lib/libgcc_s.a | grep __new_exitfn
[32]    m   0x200007e8     .data     1  unamex                    __new_exitfn_called
[8978]  m   0x100002cc     .text     2  extern                    .__new_exitfn
[8981]  m   0x00000000     debug     0     fun                    __new_exitfn:F1=*2=xsexit_function:
[9039]  m   0x00000000     debug     0    gsym                    __new_exitfn_called:G13=14=@s64;r14;0;01777777777777777777777;
[9053]  m   0x20000060     .data     1  extern                    __new_exitfn_called

We see that gcc 10 has this debug symbol which is copied in every binary compiled with gcc 10.

But if I use gcc 11 whose debug format  is dwarf and -g compiled this symbol is as below. And I do not see that bug anymore.

dump -tov /opt/freeware/lib/gcc/powerpc-ibm-aix7.2.0.0/11/libgcc_s.a | grep __new_exitfn
[32]    m   0x200007d8     .data     1  unamex                    __new_exitfn_called
[2813]  m   0x100003e4     .text     2  extern                    .__new_exitfn
[2834]  m   0x20000050     .data     1  extern                    __new_exitfn_called


So folks who will use this version of gcc will not face this problem.

So this looks like a gcc problem..

From: Ulrich Weigand <Ulrich.Weigand@de.ibm.com>
Date: Monday, 24 April 2023 at 9:14 PM
To: gdb-patches@sourceware.org <gdb-patches@sourceware.org>, Aditya Kamath1 <Aditya.Kamath1@ibm.com>, tom@tromey.com <tom@tromey.com>
Cc: Sangamesh Mallayya <sangamesh.swamy@in.ibm.com>, simon.marchi@efficios.com <simon.marchi@efficios.com>
Subject: Re: [PATCH] Fix call functions command bug in 64-bit programs for AIX and PC read in psymtab-symtab warning
Aditya Kamath1 <Aditya.Kamath1@ibm.com> wrote:

>So here is the thing, there are 4 values for which the psymtab entries are created. Their addresses are 0, 1a0, 0 and 8.
>
>In the object dump outputs I see they are like this..
>[235]   m   0x00000000     debug     0    gsym                    __new_exitfn_called:G13=14=@s64;r14;0;01777777777777777777777;
>[236]   m   0x00000008     debug     0    gsym                    __exit_funcs:G4
>[188]   m   0x000001a0     debug     0     fun                    __internal_atexit:F8=r8;-2147483648;2147483647;
>[177]   m   0x00000000     debug     0     fun                    __new_exitfn:F1=*2=xsexit_function:

It seems to be these numbers are not actual addresses,
which looks like the root cause of the problem.

In fact, looking e.g. at "__new_exitfn" in the detailed dump below:

>[174]   m   0x100007a8     .text     2  extern                    .__new_exitfn
>[175]   a2           0             416    7670     185
>[176]   a4  0x000000ac       0    0     LD       PR    0    0
>[177]   m   0x00000000     debug     0     fun                    __new_exitfn:F1=*2=xsexit_function:
>[178]   m   0x100007a8     .text     1     fcn                    .bf
>[179]   a1           0        75     0       0       0
>[180]   m   0x0000001f     debug     0   rpsym                    listp:R3=*4=*5=xsexit_function_list:
>[181]   m   0x00000003     debug     0    rsym                    r:r1
>[182]   m   0x00000009     debug     0    rsym                    i:r6=7=r7;0;037777777777;
>[183]   m   0x10000924     .text     1     fcn                    .ef
>[184]   a1           0       131     0       0       0

We see that there is an actual symbol (in the .text section)
at address 0x100007a8, covering 416 byte (until 0x10000948).

*In addition*, there is a "debug" (stabs) symbol giving the
type of that function.  This also has an address field, but
this is simply zero.

My understanding, and that is apparently confirmed by the
text in the GNU stabs documentation here:
https://sourceware.org/gdb/download/onlinedocs/stabs/Transformations-On-Global-Variables.html
is that the "address" value in the debug entry is not used,
but the actual address should be taken from the .text entry.

Now I'm not sure why this doesn't work correctly, but this
may be a place to investigate further.

(One interesting thing to note is that the names do not match
exactly, as the .text entry has the leading '.' that is used
on ppc64 ... maybe this throws off some of the stabs logic?)

Bye,
Ulrich