RE: libstdc++/10094: Regression: wfilebuf::showmanyc return value too high.

[Pétur Runólfsson <peturr02 at ru dot is>]

> Synopsis: Regression: wfilebuf::showmanyc return value too high.
> 
> State-Changed-From-To: open->analyzed
> State-Changed-By: paolo
> State-Changed-When: Sat Mar 15 13:32:25 2003
> State-Changed-Why:
>     Ok, however the (interesting) analysis is science fiction
>     ;-) : the new showmanyc it's still in review!!

Indeed :-) seems there is a bug in the old implementation...

>     Anyway, what to do here? Which kind of estimate is not
>     affected by the problem?!?

Obviously, if codecvt::always_noconv() == true, this isn't a
problem. However, it seems unreasonable to require showmanyc to
read and convert the characters before returning.

I think the real problem here is that underflow() shouldn't
return eof() when codecvt::in() fails. Consider this example:

wifstream stream;
stream.imbue(locale("se_NO.UTF-8"));
// foo begins with the bytes \xc0\xff followed by lots of
// stuff.
stream.open("foo");

int n;
stream >> n;
wifstream::iostate s = stream.rdstate();

What is the value of s?

wistream::operator>>(int&) calls rdbuf()->sbumpc() or sgetc(),
which call underflow(). underflow() reads bytes from the file,
attempts to convert with codecvt::in(), which fails.

Currently, this causes underflow() to return eof(), which
causes sgetc() to return eof(), which means that s is set to
eofbit | failbit. I see nothing in the standard that requires
this behaviour (*), however, eofbit is described so:

  eofbit: indicates that an input operation reached the end
          of an input sequence;  

This appears to mean that eofbit() should *not* be set when
underflow() has a conversion error, however, badbit seems
just right:

  badbit: indicates a loss of integrity in an input or output
          sequence (such as an irrecoverable read error from
          a file);

So it seems that !(s & eofbit) && (s & badbit) should hold.
The only reasonable way to achieve this is for underflow() to
throw an exception (probably of type ios_base::failure).

Back to showmanyc:

If underflow() throws an exception on conversion errors, the
proposed implementation of showmanyc is OK, since showmanyc
only guarantees that underflow() doesn't return eof() until
the specified number of characters has been read, there is
a footnote that indicates that it is OK for underflow() to
throw an exception anytime.

Petur

(*) Indeed, I don't see anything about what underflow()
should do when codecvt::in() fails. If you know of anything
please let me know :-)