I had totally forgotten about it myself. The occasion to take a fresh 
look at it.

Here is a new version considering 1st pubseekoff call returned value to 
find out if it can be used. I removed the check/comparison with 2nd call 
result as it's not usable. In new 4_neg.cc test case if I ask to advance 
istreambuf_iterator by 100000 the returned pos_type is showing 100000 
offset even if it is obviously false. You need to read next byte to see 
that it is eof.

     libstdc++: Implement std::advance for istreambuf_iterator using 
pubseekoff

     If advance increment is smaller than input buffer size just advance 
in this
     buffer thanks to __safe_gbump. If increment is larger check for 
seekoff support
     and use it accordingly. Otherwise fallback on current 
__safe_gbump/underflow loop.

     libstdc++-v3/ChangeLog:

             * include/bits/streambuf_iterator.h (advance): Re-implement
             using streambuf pubseekoff/seekoff when supported.
             * 
testsuite/25_algorithms/advance/istreambuf_iterators/char/4_neg.cc: New
             test case.

Ok to commit when back in stage 1 ?

François


On 31/03/2023 23:03, Jonathan Wakely wrote:
> On Tue, 15 Oct 2019 at 21:20, François Dumont wrote:
>> Here is an update to set _M_sbuf to null if the user advance too much.
>>
>> Note that in this case the streambuf remains un-modified which is
>> different from the current implementation. I think it is another
>> enhancement.
>>
>> I also change the Debug assertion message for something more dedicated
>> to std::advance algo.
>>
>> François
>>
>> On 10/14/19 10:12 PM, François Dumont wrote:
>>> The same way I proposed to review std::copy overload for
>>> istreambuf_iterator we can implement std::advance using pubseekoff.
>>>
>>> It is both a cleaner implementation and avoids yet another friend
>>> declaration.
> Looks like I never sent my review of this one, it's been sitting in my
> draft mails for years, sorry.
>
> It looks like this will fail if the streambuf doesn't support seeking.
> The default behaviour for seekoff is to return -1, in which case
> you'll get -1 for both calls to pubseekoff, and new_pos - cur_pos will
> be zero, which is not equal to n, so you set the istreambuf_iterator
> to end-of-stream. That seems wrong, we could still advance using the
> old code (or just call ++ in a loop!)
>
>
>
>>>      * include/std/streambuf
>>>      (advance(istreambuf_iterator<>&, _Distance)): Remove friend
>>> declaration.
>>>      * include/bits/streambuf_iterator.h (__copy_move_a2): Re-implement
>>> using
>>>      streambuf pubseekoff.
>>>
>>> Tested under Linux x86_64.
>>>
>>> Ok to commit ?
>>>
>>> François
>>>