Eli Zaretskii wrote:
> I'm okay with making 'message' write in one go to stderr, but I don't
> want to pay the price of having stderr buffered globally.

I came up with a fix that does all that. Although the fix does not alter the 
buffering of the stderr stream, it causes 'message' and similar functions to 
write in one go to avoid interleaving output. It also fixes the problem on AIX 
and Solaris where a single call to fprintf is implemented by multiple calls to 
'write' even if the diagnostic is short, botching interleaving. It also fixes 
the INT_MAX overflow and memory-allocation issues of the current master's 
implementation of 'message'.

Proposed patches attached. The first one merely refactors and simplifies the 
C-level code, without changing behavior; this simplifies later patches. The 
second patch fixes buffering of most C-level diagnostics in Emacs, so that they 
interleave well with other processes' diagnostics. The third patch similarly 
fixes buffering of most Lisp-level diagnostics in Emacs. The last patch is one 
more simplification of the C code.

I still far prefer the three-line patch I submitted earlier, as it was much 
simpler and the objections to it were theoretical (i.e., not based on actual 
Emacs code). However, the attached patches work nearly as well, and they do so 
while obeying the constraints you imposed on fixing the problem.