[Bug rtl-optimization/36419] [4.1/4.3/4.4 Regression] Wrong unwind info with -Os -fasynchronous-unwind-tables

[jakub at gcc dot gnu dot org]

------- Comment #7 from jakub at gcc dot gnu dot org  2008-06-06 08:03 -------
Ok, I'll bootstrap/regtest patch 1).

To answer your question about 2), consider e.g.:

__attribute__((noinline))
void foo (int, int)
{
  static int n = 0;
  void *esp;
  asm volatile ("movl %%esp, %0" : "=r" (esp));
  __builtin_printf ("foo %p\n", esp);
  if (n++ == 0)
    throw 1;
}

__attribute__((noinline))
void baz (int, int, int, int, int, int, int)
{
  static int n = 1;
  void *esp;
  asm volatile ("movl %%esp, %0" : "=r" (esp));
  __builtin_printf ("baz %p\n", esp);
  if (n++ == 0)
    throw 1;
}

struct A { A () { }; ~A (); char c[24]; };

__attribute__((noinline))
A::~A ()
{
  asm volatile ("" : : : "memory");
}

__attribute__((noinline))
void bar ()
{
  A a;
  try
    {
      foo (1, 2);
      baz (3, 4, 5, 6, 7, 8, 9);
    }
  catch (...)
    {
    }
  foo (1, 2);
  baz (3, 4, 5, 6, 7, 8, 9);
}

int
main ()
{
  bar ();
  return 0;
}

with -m32 -Os -fasynchronous-unwind-tables -mpreferred-stack-boundary=4 vs.
-m32 -Os -mpreferred-stack-boundary=4.  The generated .text
is identical between both, yet without -fa-u-t it shows correct stack alignment
(both foo calls have the same %esp, baz 16 bytes below it), but with -fa-u-t
the second foo call has %esp 8 bytes below the first foo call's %esp (so
clearly
one of them is not aligned to the preferred stack boundary), and baz is 16
bytes below it.
The problem is that at the first call _Z3fooii in bar -fa-u-t computed
args_size
8, while without -fa-u-t it is 16.  The first call _Z3baziiiiiii has wrong
args_size too (with -fa-u-t 28, without -fa-u-t 32).  The second call _Z3fooii
already has the same args_size between the two (16) and second call
_Z3baziiiiiii
too (in both cases 32).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36419