Re: [Qemu-devel] [PATCH] configure: save git working tree information in "pkgversion"

[Laszlo Ersek]

On 05/31/16 19:45, Eric Blake wrote:
> On 05/31/2016 11:01 AM, Laszlo Ersek wrote:
> 
>>>> Grepping git's Documentation/RelNotes/ directory, I find:
>>>> - in "1.6.6.txt": the introduction of --dirty
>>>> - in "1.7.6.4.txt": an apparently important bugfix for --dirty
>>>>
>>>> Version 1.7.6.4 of git was tagged on Sep 23 2011.
>>>>
>>>> Does this information help in deciding if we can use --dirty?
>>>
>>> 5 years old sounds new enough for my liking :-)
>>>
>>> I guess we could use --dirty and catch the non-zero exit code and just
>>> re-try without --dirty.
>>
>> But, if we can't use --dirty, I should probably use the plus-sign
>> fallback (we need *something* to mark a dirty state).
>>
>> In which case however, shouldn't we just go with the current patch,
>> which doesn't care about --dirty at all? Otherwise, some build hosts
>> will append "-dirty", and others will append "+".
>>
>> IMO we should either require --dirty, or go with the current patch.
> 
> Gnulib's build-aux/git-version-gen script doesn't yet use --dirty, but
> may be an inspiration for how to generate the same suffix:
> 
> # Test whether to append the "-dirty" suffix only if the version
> # string we're using came from git.  I.e., skip the test if it's "UNKNOWN"
> # or if it came from .tarball-version.
> if test "x$v_from_git" != x; then
>   # Don't declare a version "dirty" merely because a time stamp has changed.
>   git update-index --refresh > /dev/null 2>&1
> 
>   dirty=`exec 2>/dev/null;git diff-index --name-only HEAD` || dirty=
>   case "$dirty" in
>       '') ;;
>       *) # Append the suffix only if there isn't one already.
>           case $v in
>             *-dirty) ;;
>             *) v="$v-dirty" ;;
>           esac ;;
>   esac
> fi
> 

BTW, my patch has a functionality bug. Consider the case when you change
some of the tracked files, then stage all those changes with "git add",
then *undo* the changes in the working tree only. In this case, my patch
will report "dirty" ("+"), because there will be both staged changes
(relative to the HEAD commit) and working tree changes (relative to the
index). But that's incorrect -- the working tree actually matches the
HEAD commit, so the build qualifies as "clean".

On the other hand, git-diff-index will do the right thing, namely:

       git-diff-index <tree-ish>
           compares the <tree-ish> and the files on the filesystem.

which is exactly right. The index (= the staged changes) are irrelevant
for a build; only the working tree matters.

(Anyway, this is moot now; I'll happily leave it to Fam! :))

Thanks
Laszlo