Re: noisy gnulib-tool on IRIX

[Eric Blake]

[dropping bug-gnulib]

On 09/07/2010 11:38 PM, Ralf Wildenhues wrote:
> > (hmm, maybe autoconf should use that trick to reduce
> > forking at m4sh startup).
>
>    exec 3>&2 2>/dev/null; command; exec 2>&3 3>&-
>
> has the advantage of not forking, the disadvantage of using another file
> descriptor (which we should probably disallow the user from passing in).

Ah, but autoconf is already tying up other fd's for things like the log, 
so reusing fd 5 at script startup before the log is active won't put any 
additional pressure on the user.

> With modern shells, it should work just as well to
>    { command; } 2>/dev/null
>
> except of course, really old shells will fork a subshell in order to
> execute redirected compound commands, so unless we have some indication
> that this was fixed before some other feature we rely on was fixed  ;-)

In this case, I'm not worried if an old shell forks.  Rather, I'm trying 
to optimize things so that bash can start with fewer forks, making the 
startup experience slightly faster, especially for platforms like cygwin.

> we might need a test for that.

That's for sure.  And I still plan to do a sampling of various shells to 
confirm whether { command; } 2>/dev/null is sufficient at avoiding 
output, regardless of whether it forks on older shells, as it looks 
cleaner than using exec to fiddle around with fds.

> I expect that some shells try to optimize (command) 2>/dev/null by not
> always executing command in a subshell (that would explain some of the
> bugs old dash had), but I don't know how common that is.  builtin vs.
> external command are a potential difference, too, also for old shell
> behavior.

--
Eric Blake   address@hidden    +1-801-349-2682
Libvirt virtualization library http://libvirt.org