GitLab

This was broken with the unsafe_set speed ups. Because the "unsafe" bit
here is the part that avoids the GObject.Value boxing, Python integers
end up getting mapped to int, which then breaks with large values.

NFS gives a different error code to other network mounts for some
reason. Annoyingly, it's a generic error code so while I'm hesitant to
blanket handle it like this... really we're putting up a confirmation
dialog here so it's not the worst thing.

This isn't currently used, but it was when we were using
check_save_modified() for drag-n-drop and file selector save checks.
The problem was that when actually trying to save we prompted using the
passed buffer list, but didn't refer to that list when trying to save.

The benefit here is that while state in check-save is async (because we
use the async saving of GtkSource.FileSaver), confirm-discard has no
state and is a straight up modal confirm with no complications.

The places we're now doing this is when the user changes the file
selector, and when they drag-n-drop a new file on to the pane. In both
cases, I think these aren't situations where the user is likely to
actually have accidentally forgotten to save (since they're extremely
deliberate manipulations of the current comparison), so the simple
confirm seems like a reasonable complexity trade-off.

The intention here was for this to essentially be a documentation
change, i.e., so that users can see that the shortcut exists. In
reality, this is a behaviour change when hitting enter on a folder row.

Previously, this would expand/collapse the row because it was a treeview
row activation signal and that seemed like a reasonable thing to do.
This is still what will happen on actual row activation (i.e., double-
clicking a folder row), but now hitting enter will activate the Compare
action and will create a new folder comparison when activating a folder
row.

I don't think either of these options are obviously better than the
other, so I'm going to go with this new behaviour for now.

The actual reason I went to change this was that if you tried to use the
menu bar items, the main actions (i.e., compare, copy, delete) were all
insensitive. This happens because we unset their sensitivity on treeview
focus loss. I suspect that the intention here was to keep consistent
action state, but in reality we can just invoke the action based on the
last focus pane, which is what the user expects.

When doing this, I found that the focus handler block/unblock around
popup handling was doing nothing useful, and then that the popup menu
handling was actually slightly confused with the custom filter menu
code, so... now it's mostly deleted.

Currently we write the window state a *lot*, because we constantly get
spammed with `size-allocate` signals when nothing has changed (e.g.,
we've scrolled a comparison). The window size is typically unchanged,
so here we add the simplest possible check before assigning to our
bound properties. In my testing, this completely eliminates the spurious
writes.

There's also the dconf part of this where it's constantly re-writing
values that haven't actually changed; see dconf#40 for the
discussion about that.

See c7bbe78d for the removal reasoning.

Here we're just giving the user a somewhat better error message. Ideally
we'd replicate the previous offer of "hey why don't you save as UTF-8
instead?", but the save lifecycle makes this difficult here. If this was
an often-hit check then we'd do it, but... this code hasn't worked in
over three years and people have only ever complained about bugs in the
encoding check, not the save-as-UTF-8 bit.

The existing handling was Python 2 era. In the current code, we'll
always have a `str`-type root for our `os.listdir()` call, so the
entries will always be `str`s. This patch handles the Python 3 path
handling situation of getting surrogate escaped paths (in the case of a
bad file name vs. filesystem encoding) by just checking for a valid
re-encode and treating any failure as an encoding error.