Following 4cbb9360
Most of the distros provides the dictionary pre-installed but
some like Gentoo not. So, let's bundle "share/libthai" for
maximum portability.
The only trace of reason for this dir being bundled was a pango crash: #2496.
Since the crash was gone, no other reason left, so it was discarded: c808d13b.
However, according to my tests, "word breaking" does not work without this.
(This is what thbrk.c#L22 source says but was never clear in our scripts).
So, let's bundle it with proper comments to make clear why this is needed.
Resolves#12623
Compose and Recompose are in the same plug-in.
When we converted compose.c to use GimpChoice in 7b29ac24,
we did so without checking if it was being called by
Compose or Recompose. As a result, we always overwrote
compose_type with 0 (RGB mode) on Recompose since it
did not have a "compose-type" GimpChoice property.
This patch adds a check to make sure we're not in a Recompose
call before trying to set compose_type from the GimpChoice
parameter.
This option is run by the CI and makes so that the test-file-plug-ins
plug-in is always installed in the CI, even on stable releases. But by
default, they are not installed on stable releases.
It was changed to "GIMP-continuous-ARCH.AppImage" because
I had hope of linking GitLab artifacts on gimp-web (which failed)
so let's use "GIMP-GIMP_VERSION-ARCH.AppImage" again.
Also, uppercase AppId 'Continuous' suffix to be consistent with Flatpak.
In particular, we could end up with mask of wrong bytes per pixel, which
was what was happening in the report.
This commit adds a new GimpDrawableFilterMask class because we needed to
implement a specific is_attached() method for effect masks.
Note: the file is added to POTFILES but the only localized string
already existed elsewhere. So this doesn't break string freeze.
Without this change, we could be using 5GB per job, which
makes the ccache too old so kills its purpose. Also, if
we consider the pipeline it could reach 10GB per runner.
So, let's limit the 'deps' jobs to 256MB, which should be
enough considering my tests and babl and GEGL coding pacing.
GIMP, however, needed way more in my tests, so I set 768MB.
Partially reverts e01973b9
This makes the AppImage .sh script multiarch aware and
make Debian pipeline a GL 'matrix' for easier maintenance.
As consequence, making an arm64 .appimage is pretty easy now,
so let's make one since this arch is not that rare in Linux.
Adding the mode and opacity arguments makes it very similar to
gimp_drawable_merge_new_filter() signature, but also to its own
Script-fu wrapper (gimp-drawable-append-new-filter).
It's just better to be consistent. Also if either of these args need to
be changed immediately afterwards, then we spare one additional
re-render for nothing (for instance in the PSD load plug-in!).
This patch loads legacy 'sofi' data, which represents
the original Color Overlay layer style. It also sets up
the ability to add the other legacy layer styles once
we've matched the right GEGL operations and properties.
Note that this can not read the modern descriptor format
for layer styles.
See: https://learn.microsoft.com/en-us/windows/win32/shell/app-registration#using-the-app-paths-subkey
Starting with Windows 7, Microsoft "strongly" recommends to install per user.
This still applies today for Windows 10: per-user is the default for our
MSIX installed from MS Store and for our .exe installer obtained by winget.
Note: this commit just change the "(recommended)" option in an Inno dialog.
Users can install GIMP in system-wide (admin mode) as always, nothing changed.
This is not a fix for #12359 yet, but an improvement to XCF loading.
When filters fail to load for any reason, we should still load the
layer. Only drop the broken filter.
I've tested lightly this removal. I don't think this code was really
needed or logical anyway, though maybe I am missing something because I
barely discover the existence of this "Image Selection" feature of dock
windows!
When loading a GimpConfig property for a filter, we assumed that
the string is always valid. This patch adds code to check if we read
in the string correctly - if we didn't (or if it was tampered with), we
don't try to deserialize an empty string.
The code to retrieve the alpha-variant of the indexed palette was wrong,
and therefore babl was creating a brand-new palette fornat. Not sure if
the palette was empty, or was random or anything. It's even weird it
even produced somewhat resembling colors.
This is useful right before releases according our release procedure:
https://testing.developer.gimp.org/core/maintainer/release/
Also, checking the variable content make the script more safe,
considering that in the future we will publish automatically from CI.
Filters with custom dialogues like Curves and Levels did not
have the existing filter set before being initialized. This meant
that if you have a different layer selected, editing an existing
filter would add a new filter to your selected layer rather than
editing the filter on its own layer.
This patch sets the NDE filter in the tool before initialization,
so that it can get the correct layer to edit and update the
filter settings on that one.
