[Bug debug/105145] New: dropped DWARF location information at -O1/-O2/-O3 upon ftree-dse

[Bug debug/105145] New: dropped DWARF location information at -O1/-O2/-O3 upon ftree-dse

[assaiante at diag dot uniroma1.it]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105145

            Bug ID: 105145
           Summary: dropped DWARF location information at -O1/-O2/-O3 upon
                    ftree-dse
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: debug
          Assignee: unassigned at gcc dot gnu.org
          Reporter: assaiante at diag dot uniroma1.it
  Target Milestone: ---

When debugging this minimized C example, pointer variable l_158 appears as
optimized out at every stepped line, including line 14 where it is used to
compute the value of a function call argument. This happens at -O1/O2/O3 but
not at -Og.

Apparently, -ftree-dse makes the DWARF DIE for l_158 drop a location definition
while the ASM stays unchanged. When adding -fno-tree-dse to -O1/O2/O3, the
variable value appears at lines 13 and 14. However, it stays as optimized out
between line 10 (where it is assigned) to 13. We were not able to pinpoint what
may be behind this latter fact, though.

Below we provide details for -O1 on x64 and a quick assessment of prior gcc
versions.

$ cat a.c
int *a;
int **d = &a;
int *e = &d;
int f;
volatile char g;
short h(int b, int c) { return 0; }
int main() {
  int ***l_134 = e;
  int *l_133 = l_134;
  int l_141 = 0;
  int *l_158 = &l_141;
  short l_185 = 3;
  *d = l_158;
  g = ***l_134 + h(f != &l_158, l_185);
}

GCC and GDB version:
-  gcc (GCC) 12.0.0 20211227 (experimental) - commit id: 500d3f0a302
- GNU gdb (GDB) 11.2

GDB trace:
$ gcc -O1 -g a.c -o opt
$ gdb -q opt
root@1f83e51b6153:/tmp# gdb -q opt
Reading symbols from opt...
(gdb) b 14
Breakpoint 1 at 0x4004aa: file a.c, line 14.
(gdb) r
Starting program: /tmp/opt 

Breakpoint 1, main () at a.c:14
14        g = ***l_134 + h(f != &l_158, l_185);
(gdb) info locals
l_134 = 0x601030 <d>
l_133 = 0x601030 <d>
l_141 = 0
l_158 = <optimized out>
l_185 = 3

ASM at -O1:
000000000040048c <main>:
  40048c:       48 8b 05 95 0b 20 00    mov    0x200b95(%rip),%rax        #
601028 <e>
  400493:       c7 44 24 fc 00 00 00    movl   $0x0,-0x4(%rsp)
  40049a:       00 
  40049b:       48 8d 4c 24 fc          lea    -0x4(%rsp),%rcx
  4004a0:       48 8b 15 89 0b 20 00    mov    0x200b89(%rip),%rdx        #
601030 <d>
  4004a7:       48 89 0a                mov    %rcx,(%rdx)
  4004aa:       48 8b 00                mov    (%rax),%rax
  4004ad:       48 8b 00                mov    (%rax),%rax
  4004b0:       8b 00                   mov    (%rax),%eax
  4004b2:       88 05 88 0b 20 00       mov    %al,0x200b88(%rip)        #
601040 <g>
  4004b8:       b8 00 00 00 00          mov    $0x0,%eax
  4004bd:       c3                      retq

DWARF at -O1:
0x00000101:     DW_TAG_variable
                  DW_AT_name    ("l_158")
                  DW_AT_decl_file       ("/tmp/a.c")
                  DW_AT_decl_line       (11)
                  DW_AT_decl_column     (0x08)
                  DW_AT_type    (0x00000041 "int*")

Through some testing we found out that the optimization flag involved in the
issue is likely -ftree-dse. If we add -fno-tree-dse to the compilation command
line used above, the pointer variable l_158 appears in the current frame with
its value correctly shown at line 14.
The DWARF location info is correctly defined for the variable throughout its
whole lifetime.

ASM at -O1 with -fno-tree-dse:
000000000040048c <main>:
  40048c:       48 8b 05 95 0b 20 00    mov    0x200b95(%rip),%rax        #
601028 <e>
  400493:       c7 44 24 fc 00 00 00    movl   $0x0,-0x4(%rsp)
  40049a:       00 
  40049b:       48 8d 4c 24 fc          lea    -0x4(%rsp),%rcx
  4004a0:       48 8b 15 89 0b 20 00    mov    0x200b89(%rip),%rdx        #
601030 <d>
  4004a7:       48 89 0a                mov    %rcx,(%rdx)
  4004aa:       48 8b 00                mov    (%rax),%rax
  4004ad:       48 8b 00                mov    (%rax),%rax
  4004b0:       8b 00                   mov    (%rax),%eax
  4004b2:       88 05 88 0b 20 00       mov    %al,0x200b88(%rip)        #
601040 <g>
  4004b8:       b8 00 00 00 00          mov    $0x0,%eax
  4004bd:       c3                      retq

DWARF at -O1 with -fno-tree-dse:
0x00000101:     DW_TAG_variable
                  DW_AT_name    ("l_158")
                  DW_AT_decl_line       (11)
                  DW_AT_decl_column     (0x08)
                  DW_AT_type    (0x00000041 "int*")
                  DW_AT_location        (0x00000020: 
                     [0x000000000040049b, 0x00000000004004a0): DW_OP_fbreg -12,
DW_OP_stack_value
                     [0x00000000004004a0, 0x00000000004004b8): DW_OP_reg2 RCX)
                  DW_AT_GNU_locviews    (0x0000001c)

In short, the ASM of the main function stays identical but we have an accurate
location definition that leads to the correct availability of the value of
l_158 in its usage at line 14.

We have also tested older gcc versions (6.4, 7.5, 8.4, 9.3, 10.3, 11.1) and for
the tested optimization levels the behavior is identical to the git version
originally analyzed.

[Bug debug/105145] dropped DWARF location information at -O1/-O2/-O3 upon ftree-dse

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105145

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2022-04-05
             Status|UNCONFIRMED                 |NEW
                 CC|                            |rsandifo at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
There is no infrastructure to retain debug information for memory contents
(which includes address-taken local variables which can eventually become
registers later).  IIRC Richard Sandiford developed some prototype patches to
try improve this a bit a few years ago.

In particular here we have before DSE1

  [t.c:11:8] l_158(address-taken) = [t.c:11:16] &l_141(address-taken);
...
  l_158(address-taken) ={v} {CLOBBER(eol)};

and l_158 is otherwise unused so DSE elides the store.  If we'd call
update_address_taken before DSE we'd rewrite l_158 into SSA form and
then DSE/DCE would retain a # DEBUG stmt for the value.  In this case
it is FRE1 which elides all uses of the variable.

[Bug debug/105145] dropped DWARF location information at -O1/-O2/-O3 upon ftree-dse

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105145

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 105248 has been marked as a duplicate of this bug. ***