Bug 1092144

Description of problem:

When GCC does a single unrolling of a loop in which a function call has been inlined, an inlined_subroutine DIE is emitted only for the first inlining, not for the second one in the loop. As a result, DWARF consumers such as SystemTap/GDB only fire once at startup when probing the inlined function, rather than at every iteration.

Version-Release number of selected component (if applicable):

$ uname -r
3.10.0-121.el7.x86_64
$ rpm -q gcc
gcc-4.8.2-16.el7.x86_64

How reproducible:

All the time

Steps to Reproduce:

$ cat tmp.c
#include <stdio.h>
#include <unistd.h>

inline int ibar (void) {
  usleep(250000);
  return printf("printing\n");
}

int main (int argc, char *argv[]) {
  int j = 0;
  for (;;) {
    j += ibar();
    if (argc > 1)
      return;
  }
  return j;
}
$ gcc tmp.c -g -O -o tmp
$

Actual results:

$ gdb tmp
(gdb) break ibar
Breakpoint 1 at 0x4005c0: ibar. (2 locations)
(gdb) run
Starting program: /notnfs/jlebon/codebase/systemtap/systemtap/testsuite/systemtap.unprivileged/tmp

Breakpoint 1, main (argc=1, argv=0x7fffffffe298) at tmp.c:12
12          j += ibar();
(gdb) c
Continuing.
printing
printing
printing
^C
Program received signal SIGINT, Interrupt.
(gdb) q
$

Expected results:

GDB breaks at every iteration rather than only the first one.

Additional info:

The loop is unrolled once:

$ objdump -d tmp | grep '<main>' -A 18
00000000004005e2 <main>:
  4005e2:       53                      push   %rbx
  4005e3:       89 fb                   mov    %edi,%ebx
  4005e5:       bf 90 d0 03 00          mov    $0x3d090,%edi
  4005ea:       e8 d1 fe ff ff          callq  4004c0 <usleep@plt>
  4005ef:       bf ba 06 40 00          mov    $0x4006ba,%edi
  4005f4:       e8 87 fe ff ff          callq  400480 <puts@plt>
  4005f9:       83 fb 01                cmp    $0x1,%ebx
  4005fc:       7f 16                   jg     400614 <main+0x32>
  4005fe:       bf 90 d0 03 00          mov    $0x3d090,%edi
  400603:       e8 b8 fe ff ff          callq  4004c0 <usleep@plt>
  400608:       bf ba 06 40 00          mov    $0x4006ba,%edi
  40060d:       e8 6e fe ff ff          callq  400480 <puts@plt>
  400612:       eb ea                   jmp    4005fe <main+0x1c>
  400614:       b8 00 00 00 00          mov    $0x0,%eax
  400619:       5b                      pop    %rbx
  40061a:       c3                      retq
  40061b:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)

We can see ibar is inlined once outside the loop at 4005e5, and once in the loop at 4005fe. However, the DWARF info only has a single inlined_subroutine DIE with entry_pc pointing to 4005e5:

$ readelf -w tmp
...
 <1><29d>: Abbrev Number: 14 (DW_TAG_subprogram)
    <29e>   DW_AT_external    : 1
    <29e>   DW_AT_name        : (indirect string, offset: 0x36): ibar
    <2a2>   DW_AT_decl_file   : 1
    <2a3>   DW_AT_decl_line   : 4
    <2a4>   DW_AT_prototyped  : 1
    <2a4>   DW_AT_type        : <0x62>
    <2a8>   DW_AT_inline      : 3       (declared as inline and inlined)
...
 <2><343>: Abbrev Number: 22 (DW_TAG_inlined_subroutine)
    <344>   DW_AT_abstract_origin: <0x29d>
    <348>   DW_AT_entry_pc    : 0x4005e5
    <350>   DW_AT_ranges      : 0x0
    <354>   DW_AT_call_file   : 1
    <355>   DW_AT_call_line   : 12

Note however that the associated ranges do indicate the two locations:

$ readelf -w tmp
...
Contents of the .debug_ranges section:

    Offset   Begin    End
    00000000 00000000004005e5 00000000004005f9
    00000000 00000000004005fe 0000000000400614
    00000000 <End of list>

Shouldn't these be separate DW_TAG_inlined_subroutine DIEs instead?