Prepend a version number too in source tarballs

[mortenpi]

Currently the files in source tarballs just have julia/ prepended to them (e.g. 1.0.1 tarball), but I propose including the version number too, i.e. prepend julia-x.y.z/.

I switched over to using the --transform option of tar for modifying the filenames. However, this might not be very portable -- it seems to be that --transform/--xform is specific to GNU tar. BSD tar does not appear to have it.. or rather calls a similar option -s if I understand correctly.

How important is it to keep these targets cross-platform? I do not have any other idea how to approach this at the moment, other than doing something similar to the original implementation of cd-ing up and then renaming or symlinking the directory, or copying files over to temporary directory.

[ararslan]

How important is it to keep these targets cross-platform?

It is required. These targets should be able to be run on any system that can build Julia.

[mortenpi]

Different approach now: as we are using an explicit file list, it actually shouldn't be a problem to just temporarily create a symlink in the source tree which points back to the root. I would hope that this would work properly on all UNIX-y platforms. Less certain about Windows though.

[mortenpi]

I checked that this works as intended on Windows in a Cygwin build. And by interpolation, I hope that it works on everything in between.

[staticfloat]

Why || exit 1? I'm not sure I see why that would change any outcome?

[mortenpi]

I always assumed that just exit would just return 0, but it turns out that is not the case -- it returns $?. Removed the 1s.

[staticfloat]

exit 1 explicitly causes bash to exit with return code 1, so I think what you want is exit 0.

[staticfloat]

Actually, I'm not sure that exit 0 is what you want. What are you trying to do? Do you want to ignore ln -s errors? Or do you want ln -s errors to prevent running the tar command? Or do you want ln -s errors to run the tar command but prevent the make build from finishing?

[mortenpi]

If ln fails, I want this target to exit right away with an error. I'd say exit or exit N where N != 0 is what I want.

[staticfloat]

I think the best thing to do then is to not have an exit at all, and replace ln -s . || exit 1; \ with ln -s . && \

[mortenpi]

I also need to connect tar and rm with && then. And could connect sed too -- unlikely to fail, but might as well check.

[staticfloat]

Yes, I think that's a good idea. Best to check all exit codes.

[vtjnash]

It’s a Makefile—it’s already checking the exit code of each bash process (unless prefixed with -). No need to duplicate that.

[vtjnash]

Also, || exit is the same as || true, since the default exit code is success.

Edit: reading above, if it propagates, then it’s equivalent to not being present

[staticfloat]

I'm actually quite surprised this worked on Windows Cygwin; are you on a recent Windows 10? I believe that anything older than a recent Windows 10 might require admin privileges to create symlinks. (I'm not that worried about it since this is an unusual operation, so as long as it works for "some" windows machines I'm okay with it)

[mortenpi]

It should be latest -- it is a fresh install from a few days ago on a VM with all the post-install updates (says version 10.0.17134).

But in any case, doesn't Cygwin implement symlinks as normal files from the Windows perspective? I am basing this on this SO answer, and when I tried it out and looked at the symlink in the Windows file manager, it indeed appears to be just a normal file.

[mortenpi]

AV failure seems unrelated.

[staticfloat]

But in any case, doesn't Cygwin implement symlinks as normal files from the Windows perspective?

I didn't realize this; that's great.

[mortenpi]

As far as I could tell, the only reason why the commands were joined together into a single shell call with \s was so that you could have the DIRNAME variable. But since now it is just julia-${JULIA_COMMIT}, I put this in explicitly, without compromising readability. So instead of &&s I split it up into different shell calls, which should have the same effect (what @vtjnash hinted at).

[mortenpi]

Friendly bump -- anything left to do here?

[StefanKarpinski]

Seems fine to me but I'm not sure I'm qualified to review. Who would you like to review?

[mortenpi]

Not sure. Maybe @staticfloat or @vtjnash could take another look?

[staticfloat]

Looks good to me. Thanks @mortenpi

[staticfloat]

.....ah I clicked "rebase and merge", not "squash and merge". 🙄