Re: Y03arguments.ztst fails

[Peter Stephenson <pws@xxxxxxxxxxxxxxxxxxxxxxxx>]

Peter Stephenson wrote:
> I'm seeing these failures in Y03arguments.ztst too, they don't seem to
> be there in 4.1.1-test-1.  I would think it's more likely to be in the
> functions than the C code, since no-one dares touch the latter.

It's the change below from 18458.  Oliver, what's the right fix?  I presume
the test needs updating to add some `+'s explicitly, or to set the
style.  The latter is probably easiest --- adding 

zstyle ":completion:*:options" prefix-needed no

to comptest makes Y03arguments.ztst pass.

Index: _describe
===================================================================
RCS file: /cvsroot/zsh/zsh/Completion/Base/Utility/_describe,v
retrieving revision 1.13
retrieving revision 1.14
diff -u -p -r1.13 -r1.14
--- _describe	23 May 2002 12:16:42 -0000	1.13
+++ _describe	22 Apr 2003 12:33:08 -0000	1.14
@@ -20,6 +20,10 @@ elif [[ "$1" = -t* ]]; then
   shift
 fi
 
+[[ "$_type" = options && ! -prefix [-+]* ]] && \
+    zstyle -T ":completion:${curcontext}:options" prefix-needed &&
+        return 1
+
 # Do the tests. `showd' is set if the descriptions should be shown.
 
 zstyle -T ":completion:${curcontext}:$_type" verbose && _showd=yes


> I'm also getting random failures with multios tests which I haven't seen
> before.  I don't think the underlying problem is new, I think it's the
> same issue as this comment:
> 
> # Following two tests have to be separated since in
> #   print bar >foo >bar && print "$(<foo) $(<bar)"
> # the multios aren't flushed until after the substitutions take
> # place.  This can't be right.

Grud.  I can partially fix this by storing a list of auxiliary processes
for which the job should wait.  This applies for both multios and >(...)
subsituttions.  Unfortunately there are two problems.

1. It only works for builtins.  This is because in e.g. `cat > >(...)'
the cat is forked before the >(...), so that doesn't get into the job
table of the shell which is waiting for the cat.  The only two obvious
fixes are (i) reorder the fork of >(...) to before the fork of the cat
--- messy and inefficient, requiring extra management of file
descriptors in the parent shell (as is currently done for builtins) (ii)
make the cat process fork again and wait for both the cat and the >(...)
(since it's not feasible to assume one must exit before the other) ---
even messier and more inefficient.  [Waiting for cats to do anything is
always frustrating.]

2. It doesn't always work even then.  Possibly this is because I've got
my wires crossed and haven't implemented properly, but also possibly
it's to do with the additional complexities for handling processes in
the current shell, a.k.a. `list_pipe', see exec.c passim.  Anyway, it
seems to work from the command line, and for the multios tests, but it
doesn't work for this test in D03procsubst.ztst:

  paste <(cut -f2 FILE1) <(cut -f4 FILE2) > >(sed 's/e/E/g' >OUTFILE)
  sleep 1       # since the sed is asynchronous
  cat OUTFILE
0:>(...) substitution
>SEcond ViErtE

even in the modified form:

# slightly desperate hack to force >(...) to be synchronous
  { paste <(cut -f2 FILE1) <(cut -f4 FILE2) } > >(sed 's/e/E/g' >OUTFILE)
  cat OUTFILE
0:>(...) substitution
>SEcond	ViErtE

In fact, I'm fairly clueless about which tests in the list_pipe code
should be modified to look at the auxiliary process list as well as the
main process list.  So I've given up on this for now.

-- 
Peter Stephenson <pws@xxxxxxxxxxxxxxxxxxxxxxxx>
Work: pws@xxxxxxx
Web: http://www.pwstephenson.fsnet.co.uk