Re: [PATCH v2] completion: reduce overhead of clearing cached --options

Hi,

SZEDER Gábor wrote:

[...]

As discussed at [1], this fails catastrophically when sourced by git-completion.zsh, which explicitly clears ZSH_VERSION. By catastrophically, I mean that Rick van Hattem and Dave Borowitz report that it segfaults.

I can't reproduce it since I am having trouble following the instructions at the top of git-completion.zsh to get the recommended zsh completion script loading mechanism working at all. (I've succeeded in using git's bash completion on zsh before, but that was many years ago.) First attempt:

 1. rm -fr ~/.zsh ~/.zshrc
    # you can move them out of the way instead for a less destructive
    # reproduction recipe
 2. zsh
 3. hit "q"
 4. zstyle ':completion:*:*:git:*' script \
      $(pwd)/contrib/completion/git-completion.zsh
 5. git checkout <TAB>

Expected result: segfault

Actual result: succeeds, using zsh's standard completion script (not git's).

I also tried

 1. rm -fr ~/.zsh ~/.zshrc
    # you can move them out of the way instead for a less destructive
    # reproduction recipe
 2. zsh
 3. hit "2"
 4. mkdir ~/.zsh; cp contrib/completion/git-completion.zsh ~/.zsh/_git
 5. echo 'fpath=(~/.zsh $fpath)' >>~/.zshrc
 6. ^D; zsh
 7. git checkout <TAB>

and a similar process with fpath earlier in the zshrc file. Whatever I try, I end up with one of two results: either it uses zsh's standard completion, or it spews a stream of errors about missing functions like compgen. What am I doing wrong?

Related question: what would it take to add a zsh completion sanity check to t/t9902-completion.sh?

When running with "set -x", here is what Dave Borowitz gets before the segfault. I don't have a stacktrace because, as mentioned above, I haven't been able to reproduce it.

	str=git
	[[ -n git ]]
	service=git
	i=
	_compskip=default
	eval ''
	ret=0
	[[ default = *patterns* ]]
	[[ default = all ]]
	str=/usr/bin/git
	[[ -n /usr/bin/git ]]
	service=_dispatch:70: bad math expression: operand expected at `/usr/bin/g...'

	zsh: segmentation fault  zsh

Here's a minimal fix, untested. As a followup, as Gábor suggests at [2], it would presumably make sense to stop overriding ZSH_VERSION, using this GIT_ scoped variable everywhere instead.

Thoughts?

Reported-by: Rick van Hattem <wolph@wol.ph>
Reported-by: Dave Borowitz <dborowitz@google.com>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>

[1] https://public-inbox.org/git/01020163c683e753-04629405-15f8-4a30-9dc3-e4e3f2a5aa26-000000@eu-west-1.amazonses.com/ [2] https://public-inbox.org/git/20180606114147.7753-1-szeder.dev@gmail.com/