Re: [PATCH] stdio-common: Add the fgetln function

[Florian Weimer <fweimer@redhat.com>]

* Paul Eggert:

> If the stream is not already oriented, FreeBSD getln sets the stream
> to byte-orientation. Should glibc getln do the same?

Our getdelim doesn't do that explicitly.

> On 5/3/22 00:36, Florian Weimer via Libc-alpha wrote:
>> +  /* Discard the old buffer.  This optimizes for a buffered stream,
>> +     with multiple lines in each buffer.  */
>> +  if (fp->_fgetln_buf != NULL)
>> +    {
>> +      free (fp->_fgetln_buf);
>> +      fp->_fgetln_buf = NULL;
>> +    }
>
> Hope you don't mind a bit of bikeshedding here....
>
> Why free the fgetln buffer eagerly? Instead, free it only when
> closing. That would lessen pressure on the memory allocator and would
> save a few insns in fgetln's usual case.

The assumption is that very few lines cross buffer boundaries.

> Come to think of it, how about if we restrict fgetln to streams for
> which either (1) the user has not called setvbuf with a nonnull
> buffer, or (2) the input line fits in the user-supplied setvbuf
> buffer.

It would require a layering violation as far as libio is concerned: a
high-level function such as fgetln cannot reallocate the read buffer.
You mention setvbuf, but there are probably other cases (and of course
GCC 2.95 C++ classes, but we don't need to worry about compatibility
with those, I think).

> Then we wouldn't need to worry about adding a _fgetln_buf
> slot, as fgetln could always return a pointer into the
> already-existing stdio buffer, possibly by enlarging the buffer in
> case (1). (In case (2) fgetln could fail with ENOMEM if the input line
> is longer than the user-supplied buffer.) This would suffice for 99.9%
> of applications and would be more efficient than what FreeBSD does,
> and the whole point of fgetln is low-level efficiency right?

I'm not sure if it's more efficient.  The I/O block granularity would
change depending on where lines end.

Thanks,
Florian