Re: getndelim2 improvements

[Bruno Haible]

Eric Blake wrote:
> >   4. EOF when reading the first byte, no error, and offset > 0.
> >      Return value was 0 and is now -1.
> 
> I still think this case is like getdelim failing at EOF, particularly since...
> 
> > 
> >   5. EOF when reading the first byte, no error, and offset == 0.
> >      Return value is -1.
> 
> offset should not need to be consulted if the stream is already at EOF.

One can argue that getndelim2 should behave like this, and if I had designed
it, maybe I would have specified it like you say.

But when you look at the history ("gitk lib/getndelim2.c",
"gitk lib/getnline.c") you see that since the function getstr() was
introduced in coreutils in 1995, with an offset argument, it returned 0
in case 4.

+      if (c == EOF ...)
+       {
+         /* Return partial line, if any.  */
+         if (read_pos == *lineptr)
+           return -1;
+         else
+           break;
+       }

To treat case 4 differently, the line in the middle would have to read

          if (read_pos == *lineptr + offset)

I'm hesitating to change an interface that exists for 13 years, saying
that it is a "bug fix". Even if the number of users of this interface
(with offset > 0) appears to be zero (according to Google codesearch).

> Yes, we can revert this part of the patch, and improve the documentation.  It 
> also means we need to beef up the testcase to stress all 5 conditions, and 
> ensure we don't regress from the more-heavily documented interface.

I'll do that on Sunday, if you are not quicker than me.

Bruno