[Bug middle-end/86505] [6/7/8/9 Regression] __builtin_va_arg_pack_len() computes the number of arguments wrongly

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86505

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2
             Status|UNCONFIRMED                 |ASSIGNED
      Known to work|                            |4.4.7
            Version|unknown                     |8.1.0
           Keywords|                            |wrong-code
   Last reconfirmed|                            |2018-07-12
          Component|c                           |middle-end
     Ever confirmed|0                           |1
            Summary|__builtin_va_arg_pack_len() |[6/7/8/9 Regression]
                   |computes the number of      |__builtin_va_arg_pack_len()
                   |arguments wrongly           |computes the number of
                   |                            |arguments wrongly
   Target Milestone|---                         |6.5
      Known to fail|                            |4.5.2, 6.4.0, 7.3.0, 8.1.0,
                   |                            |9.0

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  This is because funA is inlined as follows

__attribute__((always_inline))
funB (unsigned int param)
{
  int D.1995;
  int D.1994;
  unsigned int param;
  int D.1981;
  int _4;
  int _6;

  <bb 2> :
  param_5 = param_2(D);
  _6 = __builtin_va_arg_pack_len ();

  <bb 3> :
<L1>:
  _7 = _6;
  _4 = _7;

  <bb 4> :
<L0>:
  return _4;

}

here obviously while va_arg_pack_len is correctly not resolved it needs
to be adjusted by the additional variadic args that got passed to funA.

The testcase is somewhat artificial because we refuse to inline once
variadic args are accessed in the callee but still it shows a bug.

I will try to fix that.