[Bug middle-end/58479] [6/7/8 Regression] slow var-tracking on x86_64-linux at -O1 (and above) with -g, but checking disabled

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58479

--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Even with LVUs if I try the #c0 testcase with outer 3 loops commented out (so
only 8x8 unrolling) and with a single dimension array rather than 3 dimensional
one, i.e.:
int d, e, f;

int
main ()
{
  for (d = 0; d < 8; d++)
    for (e = 0; e < 8; e++)
      {
        int t[9] = { f, f, f, f, f, f, f, f, f};
      }
  return 0;
}

we still don't emit any location info for t, even when it is in *.final dump:
(note 735 734 807 2 pr58479-4.c:12 NOTE_INSN_BEGIN_STMT)
(note 807 735 808 2 (var_location t$0 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)
(note 808 807 809 2 (var_location t$1 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)
(note 809 808 810 2 (var_location t$2 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)
(note 810 809 811 2 (var_location t$3 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)
(note 811 810 812 2 (var_location t$4 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)
(note 812 811 813 2 (var_location t$5 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)
(note 813 812 814 2 (var_location t$6 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)
(note 814 813 736 2 (var_location t$7 (mem/c:SI (symbol_ref:DI ("f") [flags
0x2] <var_decl 0x7fffefd9e120 f>) [1 f+0 S4 A32])) NOTE_INSN_VAR_LOCATION)

So unless that changes, not really sure if it is actually worth preserving the
debug stmts at all.  I guess we don't do that because there is no real
instruction that is in the scope where the variable is defined and the
NOTE_INSN_BEGIN_STMTs don't really count (but shouldn't they?).

Alex, can you please have a look?