Recent regression on (N) after removing array element

Recent regression on (N) after removing array element

[Vin Shelton]

This used to work:

x=( /bin /foo )
x[(r)/foo]=''
x=( $^x(N) )

but the last line now yields an error:
zsh: no matches found: (N)

If I remove this recent change to Src/glob.c, (N) works correctly:

--- /opt/src/zsh-2013-11-17/Src/glob.c 2013-09-23 18:01:42.000000000 -0400
+++ /opt/src/zsh-2013-11-19/Src/glob.c 2013-11-19 10:22:39.000000000 -0500
@@ -1171,7 +1171,7 @@
  break;
     }
  }
- if (*s != Inpar)
+ if (*s != Inpar || s == str)
     break;
  if (isset(EXTENDEDGLOB) && !zpc_disables[ZPC_HASH] && s[1] == Pound) {
     if (s[2] == 'q') {


Sincerely,
  Vin

Re: Recent regression on (N) after removing array element

[Peter Stephenson]

On Tue, 19 Nov 2013 12:08:04 -0500
Vin Shelton <acs@alumni.princeton.edu> wrote:
> This used to work:
> 
> x=( /bin /foo )
> x[(r)/foo]=''
> x=( $^x(N) )
> 
> but the last line now yields an error:
> zsh: no matches found: (N)
> 
> If I remove this recent change to Src/glob.c, (N) works correctly:

What we didn't think of was glob qualifiers that transformed a failed
match in some way.  Although the empty string isn't a valid file name,
it is a valid failed match for a file name.

It's quite hard to see how to improve the test so that we can still
check for something before the (N).  By the time globbing comes along,
it's already been expanded to a bare (N), so we can't see that the user
originally had something in front.

Apart from going back to the original, the only ways round that occur to
me involve reparsing on failure, e.g. reinterpreting as glob qualifiers
if no matches were found.  But that's probably worse than the original
disease.

pws

Re: Recent regression on (N) after removing array element

[Bart Schaefer]

On Nov 19,  5:52pm, Peter Stephenson wrote:
}
} What we didn't think of was glob qualifiers that transformed a failed
} match in some way.  Although the empty string isn't a valid file name,
} it is a valid failed match for a file name.

I was hoping there might be some way we can distinguish, e.g., that a
null token appears there rather than an empty string, but apparently
those are stripped by the time we reach globbing.
 
} Apart from going back to the original, the only ways round that occur to
} me involve reparsing on failure, e.g. reinterpreting as glob qualifiers
} if no matches were found.  But that's probably worse than the original
} disease.

Yeah, I think we have to back this one out.

Re: Recent regression on (N) after removing array element

[Peter Stephenson]

On Tue, 19 Nov 2013 10:02:22 -0800
Bart Schaefer <schaefer@brasslantern.com> wrote:
> } Apart from going back to the original, the only ways round that occur to
> } me involve reparsing on failure, e.g. reinterpreting as glob qualifiers
> } if no matches were found.  But that's probably worse than the original
> } disease.
> 
> Yeah, I think we have to back this one out.

In the absence of immediate smart ideas...

Here's a test so we don't forget.

diff --git a/Src/glob.c b/Src/glob.c
index 385b9e6..e0d0cf6 100644
--- a/Src/glob.c
+++ b/Src/glob.c
@@ -1171,7 +1171,7 @@ zglob(LinkList list, LinkNode np, int nountok)
 		break;
 	    }
 	}
-	if (*s != Inpar || s == str)
+	if (*s != Inpar)
 	    break;
 	if (isset(EXTENDEDGLOB) && !zpc_disables[ZPC_HASH] && s[1] == Pound) {
 	    if (s[2] == 'q') {
diff --git a/Test/D02glob.ztst b/Test/D02glob.ztst
index d6ab733..81b0021 100644
--- a/Test/D02glob.ztst
+++ b/Test/D02glob.ztst
@@ -491,3 +491,11 @@
 >@(scrimf), @(scrimf): 0
 >'(' '*' '[' '^' '@('
 >Nothing should be disabled.
+
+  (
+   setopt nomatch
+   x=( '' )
+   print $^x(N)
+  )
+0:No error with empty null glob with (N).
+>

pws