git difftool suddenly dies in 2.20

[MarkKharitonov]

I have already downgraded to 2.18, but I am pretty certain the regression comes from 2.20:

C:\DFDeploymentSmokeTests [master ↑3 +0 ~12 -0 !]> git difftool
error: cannot spawn Í?: No such file or directory
fatal: external diff died, stopping at NonUITests/SiteCheck.cs

Everything works once I downgraded to 2.18.

[dscho]

And I am pretty certain that this bug report is incomplete:

$ git difftool                                                                                                                                                    This message is displayed because 'diff.tool' is not configured.
See 'git difftool --tool-help' or 'git help config' for more details.
'git difftool' will now attempt to use one of the following tools:
opendiff kdiff3 tkdiff xxdiff meld kompare gvimdiff diffuse diffmerge ecmerge p4 merge araxis bc codecompare emerge vimdiff
Viewing (1/1): 'README.md'
Launch 'vimdiff' [Y/n]?

When I say y, it works.

This is with a 64-bit Git for Windows installed via the official installer, using Secure Channel and the "check daily for updates" checked, otherwise default options.

[dakotahawkins]

@dscho In case this is the same thing or same class of things, I upgraded this morning and has find kill my console. Will report back or file another issue if it happens again or I get new info.

[dscho]

@dakotahawkins that sounds worrisome. Hopefully you can gather more information on this.

[gunnar22]

I observed the same problem (as reported by MarkKharitonov) in git v2.20 on Windows. (I have not tried other versions after v2.20, I can not say for sure when the regression started).

I did a little bit of research. Try to do like this:
Keep watching the Temp folder where the temporary files are placed before diff starts (repeated for each pair of files in a loop), which is C:/Users/<username>/AppData/Local/Temp
Start diffing between two commits where there are several altered files.
git difftool <commit1> <commit2>
Not necessary to use the -y option. (You may, but it is easier to analyse if you don't).
And answer n to all the Launch 'toolname' [Y/n]? n
so that it does not actually try to do any real diffing of the files.
In my system it loops from file to file and runs fine for the approximately first 5 to 8 file pairs.
Watching the Temp folder, I can see that it places temporary files with decorated names (e.g., q97v0a_MySource.h) there to be diffed.
Then suddenly the error occurs:

When the error occurs, it has NOT placed the temporary files to be diff'ed into the Temp folder.

So, it is not the diffing that is the problem. It seems to be more related to what the error message says: "No such file..."

[dscho]

When the error occurs, it has NOT placed the temporary files to be diff'ed into the Temp folder.

Is it always the same file?

[gunnar22]

Each experiment is repeatable, but if I change the experiment (i.e., specify other commits) it may have a different number of successes and fail on a different file than in the previous experiment.

[vdurante]

I have the same issue here. The issue feels random but it feels like it has more chances of happening on some files.

[morgaalb]

I also have this issue. It seems to occur when using git difftool on a repo with more than 23 submodules. It only seems to happen when running git difftool and not git difftool <path>.

[dscho]

It seems to occur when using git difftool on a repo with more than 23 submodules.

Can you verify that by making a really small repository from scratch, with 23 fake submodules?

[gunnar22]

I guess that submodules is not the reason. My case also has submodules, but a lot fewer than quoted.