Re: [tree-ssa] Lowering of VLA's, part 2

[Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>]

Hello,

> > > The stack save should also go inside the BIND_EXPR, but before the
> > > TRY_FINALLY_EXPR.
> > > 
> > > Also, you need to make sure that you don't add these again if the BIND_EXPR
> > > gets gimplified again.
> > 
> > done. Just retesting it, ok if it passes?
> 
> If you're going to get rid of VLAs, perhaps a subsequent patch could get
> rid of the code that deals with VLAs in find_hidden_use_vars_r.

I hoped so, but it turned out to be a bit problematic.  Consider

void zzz ()
{
  int z = 5;
  struct { char a[z]; } x;

  foo (x);
},

translated into

zzz ()
{
  int z.1;
  unsigned int z.2;
  void * saved_stack.3;
  int z;
  struct
  {
    char a[(unsigned int)SAVE_EXPR <z.1>];
  } x;
  extern  foo;

  saved_stack.3 = __builtin_stack_save ();
  try
    {
      z = 5;
      z.1 = z;
      z.2 = (unsigned int)z.1;
      __builtin_stack_alloc (&x, z.2);
      foo (x);
    }
  finally
    {
      __builtin_stack_restore (saved_stack.3)
    };
}

Without hidden use of z.1 that protects it from being optimized out, it
becomes

zzz ()
{
  void * saved_stack.3;
  struct
  {
    char a[(unsigned int)SAVE_EXPR <z.1>];
  } x;
  extern  foo;

  saved_stack.3 = __builtin_stack_save ();
  try
    {
      __builtin_stack_alloc (&x, 5);
      foo (x);
    }
  finally
    {
      __builtin_stack_restore (saved_stack.3)
    };
}

size of x now depends on non-existent variable z.1, which causes a crash
when an attempt to copy it before passing to foo is made.

Zdenek