Analysis of libstdc++/2071 test case

[Loren James Rittle <rittle at latour dot rsch dot comm dot mot dot com>]

Thanks to Benjamin for telling me where I might look, I now understand
the failure mode of this test case (simplified from libstdc++/1620
and/or slightly extended from libstdc++/2071):

#include <iostream>

using namespace std;

int main()
{
        int i;
        cin >> i;
        cout << "i == " << i << endl;
}

and why it is not being seen on all platforms.

By repeatedly single-stepping through the code when stdin is hooked to
interactive and non-interactive streams, I have discovered a much
simpler test case (with about as much return value checking as the
libstdc++ code ;-) that reveals a known portability issue (and indeed
mirrors the exact sequence of relevant IO calls made by libstdc++ code
in the above test case).

#include <errno.h>
#include <stdio.h>

int main ()
{
  char buf;
  char buf2;
  int error = 0;

  fread (&buf, 1, 1, stdin);
  fseek (stdin, 0, SEEK_CUR);
  fread (&buf2, 1, 1, stdin);

  printf ("%d %d %d\n", buf, buf2, error);
}

When run on Red Hat Linux 6.1 (Cartman) (with whatever glibc was the default):

; a.out
1234<CR>
1 2

When run on Solaris 5.7 (with Sun libc):

; a.out 
1234<CR>
<CR>
1 

When run on FreeBSD 4.2 (with BSD libc):

; a.out 
1234<CR>
<CR>
1 

When run on Digital UNIX V4.0A (with OSF1 libc):

; a.out
1234<CR>
1 2


When run on Red Hat Linux 6.1 (Cartman), FreeBSD 4.2, Digital UNIX V4.0A:

; echo 1234 | a.out
1 2

When run on Solaris 5.7 (with Sun libc):

; echo 1234|a.out
1 (unprintable character)

All mentioned platforms:

; cat >B
1234<CR><EOF>
; a.out <B
1 2

My reading of ISO-C-1999 was of no help in resolving which set of
implementations are wrong, if any.  Not that it would matter since
Solaris is considered a primary platform and failure when cin is an
interactive stream breaks almost all non-trivial C++ programs.

No patch yet.

Regards,
Loren