Re: `message' not outputting the newline "atomically"

[Paul Eggert <eggert@cs.ucla.edu>]

On 6/24/19 2:33 PM, Lars Ingebrigtsen wrote:
> the code also calls code_convert_string, so I thought one
> more xmalloc didn't make much difference.

If code_convert_string could cause an infloop when reporting 
low-on-memory messages, then we should fix that infloop too. In the 
meantime we shouldn't make things worse by doing yet another heap 
allocation.

One possibility would be to use an auto (C stack) buffer of size (min 
(MAX_ALLOCA, PIPE_BUF)) bytes if that buffer is large enough; otherwise, 
stick with the current code. This would handle the vast majority of 
cases atomically, and the remaining cases would be no worse off than 
they are now.

Still, it would be better if the code didn't need to copy the string 
into a buffer at all. writev would work for that, since stderr is 
unbuffered. There are other possibilities.

> * The patch also needs a FIXME comment saying it fixes only "message"
> output, not the other uses of stderr in Emacs.
> Well...  that's kinda unusual.  The commit message may say so, but
> having that in the code would be odd.
I think it's worth having that comment here. Someone reading this code 
will reasonably wonder why we're going to all this trouble to 
hand-buffer stderr here, when we don't bother doing it elsewhere. The 
commit log is not a good place to document curious code like that.