Problems with command line scratch files in zsh

[Dan Nelson]

In the last episode (Jun 12), Mike Meyer said:
> [Format recovered from top posting.]
> In <1118608948.705.6.camel at localhost>, Christopher Black <cblack at securecrossing.com> typed:
> > On Sun, 2005-06-12 at 03:31 -0500, Mike Meyer wrote:
> > > Since going to 5.x with devfd, I've noticed that some of the
> > > shell constructs used by zsh (and other shells - I know zsh
> > > didn't invent this) quit working. To wit:
> > > 
> > > guru% wc <(cat /etc/motd)
> > > wc: /dev/fd/11: open: No such file or directory
> > > 
> > > The <(...) construct runs the pipe in (), and replaces the <(...)
> > > with the name of the /dev/fd/ entry for the output of that pipe.
> > > The file exists for the shell process doing all this. But when
> > > the comm process tries to open the file to read the data, the
> > > file doesn't exist. This is pretty nasty.
> > > 
> > > Anyone got any suggestions on how to fix this? A bug report with
> > > a patch, maybe (I couldn't find any such bug report)?
> > > Workarounds? Maybe this should go to hackers at freebsd.org?
> > Why not just 'cat /etc/motd | wc' ?
> 
> Because I used a trivial example designed to illustrate the problem. A
> less trivial example would be:
> 
> 	comm -12 <(sort file_one) <(sort file_two)
> 
> Of course, this can also be rewritten using temp files instead of
> pipes. But that will be longer, slower, and uglier.
> 
> This worked on 4.X. It ought to work on 5.X.

If you want a tempfile, you should probably use the =() syntax, which
will always use a tempfile.  <() and >() will attempt to use /dev/fd. 
It probably worked on 4.* because 4.* creates 64 /dev/fd/* device nodes
on install.  If for some reason zsh had more than 64 files open
already, it would have failed even on 4.*.

-- 
	Dan Nelson
	dnelson at allantgroup.com