The ARTIFACTS_SUFFIX is being dropped on building scripts because:
1) This will make possible to further simplify Installer scripts in other commit
2) There is no script that uses multi-arch _build/_install at same time on CI
However, there are two use cases: to build Debian and flatpak; and, on Win,
to build on CLANGARM64 and CLANG64 (?). But probably they are pretty niche.
I suppose that people who build on Debian (or other dev-oriented distro)
willn't want to mantain deps in two sys prefixes (pkg and GNOME runtime)
+ two gimp prefixes (out of sandbox and sandbox) due to huge storage use.
Why someone would want to build with emulation on Win ARM64 I don't know.
Anyway, people that know how to do this stuff probably can change the .sh.
3) When building locally, the contributor doesn't need to know the arch at all.
Indeed, without this suffix, the scripts are inline with gimp-web-devel and
prevent some first-sight confusion when reading them that I've seen on IRC.
4) These arch suffixes can be understood as '_install-*' being distributable
which is not true. So, without this suffix we could make more clear
what is a package (when GitLab fix the glob bug in 'expose_as' someday).
---
Despite .gitlab-ci.yml not being a script, it needed to be touched too
because cross scripts depends on Debian jobs due to gimp-data MR.
The numbering is now inline with actual the order of jobs on CI
to make easier to understand what the numbers represent.
To understand why bundling is number 2 see: d09a2a6f and 2dc6f411
The storing of Windows scripts was changed to make easier to
call them, to better convey that they work outside CI and
to be consistent with other build/ subdirectories.
- Dependencies not present in GNOME runtime are now built in deps stage
This makes easier to follow the progress of the overall pipeline and
to know how much time was spent on each stage
(like crossbuilds, the artifact size is brutal but only lasts 2 hours)
- babl and GEGL build now have output in GitLab runner, unlike others deps
This makes clearer to spot if something goes wrong in these crucial deps
- GIMP is still built in its stage but now alone, like Debian and Windows
This makes possible to retrigger only this job when runner errors occours,
without needing to start monolithically the deps+gimp build from scratch
Also:
- babl, GEGL and GIMP now have meson-log.txt artifact like other builds
- dist job now have all commands self-contained on its script (needs to
be merged to be tested according to my tests)
See: https://gitlab.gnome.org/GNOME/gimp/-/merge_requests/1670#note_2152565
This makes the NIGHTLY flatpak use the local GIMP source. This reduces the
internet usage and makes possible to generate .flatpak from every branch out
there, which turns the flatpak finally really useful to code development.
Of course, this makes impossible to run 'flatpak-builder NIGHTLY_manifest.json'
without cloning or downloading GIMP code, but this is not a regression since:
1) the NIGHTLY manifest is in-source to begin with, so same thing as our other
building methods (except macOS), which runs over the code they are stored
2) the chances of someone downloading partial source code (the NIGHTLY manifest
alone) so being unable to build are low since this file isn't too easy to find
3) even if someone do this and don't like, we are not violating flatpak building
pratices since this option exists to be used (and will be in NIGHTLY only)
Despite we not building gimp nightly on aarch64 we need to sync with beta.
So, this updates to LLVM/Clang 18, which fixes 'gexiv2' and 'poppler'
problems with deps on that arch. 'json-c' and 'luajit' were adjusted too
with the use of proper build options to fix errors.
As decided in #10922, 32-bit will be tolerated in 3.0 series because of TWAIN.
So, let's package only the bare minimum files to the 32-bit TWAIN plug-in work.
This reduces the final installation size by approximately 215MB.
It's a common pratice to tell the user what version is being installed
even if this is somewhat obvious (the user opened the installer, he
should know it). Many installers do this and Win MSIX installer too.
However, there is no sign in our Installer about what version is
being installed (the reused splash/intro image doesn't count because
it isn't always different at each micro or revision, and text is small).
---
To accomplish that feature, the Inno langs needed to be patched since
Inno actually forbidden to use that feature when we have a Welcome page.
https://groups.google.com/g/innosetup/c/w0sebw5YAeg
But 'UninstallAppFullTitle' from the following langs couldn't be patched:
- Japanese
- Lativian
There are two approaches regarding the icon for Windows installers .exe:
1) same icon as the app: this give more identity, but creates confusion if
you saved the installer in the same dir of the installed app shortcut.
This approach is also confusing in the task bar (e.g.: running GIMP
stable while installing GIMP unstable);
2) generic icon (e.g. a box, a cd) provided by the tool: more generic if
you downloaded two installers generated by the same tool (ours is Inno)
I choose a middle ground and created a icon with: the app icon and an
"installer" symbol (a package box), which conveys the best of two words.
This also fixes the Inno inborn bug of the unninstaller with install icon.
After running almost all the py plug-ins, I noticed that few py modules and
pkgs are used to justify the need of a slightly faster build time with .pyc.
In the actual bloated status (with all .pyc), lib/python* is about 260 MB big.
With the bare minimum .pyc after the tests above, python*/ is less than 90.
So, let's purge the .pyc at bundling time (I'm not reinventing the wheel,
Krita do this too), which reduces the .zip bundles (so the Inno and MSIX
installs) in 170MB, the Installer .exe in 45MB and MSIX download in 60MB.
---
Let's also disable any .pyc generation, since GIMP installed with Inno in the
system-wide (aka admin) mode or installed with MSIX both doesn't handle well:
py plug-ins work but have CLI errors, unlike Inno user-mode and .zip bundle.
Actually, the Language dialog is always displayed, which is not too good since:
1) This is inconsistent: the install mode dialog (to choose user or system)
appears before using the system lang automatically. Even if Inno fix this:
2) This is tiring: depending on the context, the user can see up to 5 dialogs:
SmartScreen (1) > Install mode (2) > installer lang (3) > dev warning (4)
> SetupMutex (5): https://groups.google.com/g/innosetup/c/bkRtnHz2ygg
3) This is redundant: even if only 2 and 3 are displayed, why 2? Inno will
auto select the sys lang, which the user understand to search GIMP for PC;
4) This is confusing: even the lang dialog message being clear, some users
genuinely conclude that the selection will carry over to GIMP install like
the old Adobe CS installer (#5515). !1628 makes more clear that not, but
reducing confusion even more is a welcome plus considering points above.
So, let's use 'ShowLanguageDialog=auto'. That option ensures that the dialog
will NOT appear if the sys lang have a translation present in the installer,
which matches what every program does and even installers like Inkscape one.
If the lang isn't included in the installer, the dialog WILL appear so the
user can choose one which he is more fluent, being is a nice middle ground.
This solves an old bug that makes Inno report GIMP as being almost 1.8GB,
which is not true (it's 1.3GB at most). That was probably caused by the
overuse of components intersections (e.g.: gs + gimp, py + deps), which
Inno does not handle well (sometimes it not even displayed the cmp size).
TWAIN component is still reporting a wrong size since it's a totally
different beast, it will be fixed later.
Following c6634d66 and 09d6402a
These files are too small and aren't reused so no point keeping them modular.
Also, with them inside .iss we reduce the clutter in the installer/ folder.
Despite that folder mysteriously not being mandatory on Linux
(the contents are different), on Windows (in PCs without MSYS2)
this is required by a warning. So, partially reverts c808d13b
The gimp.org main page was being used as the updates URL but the
downloads page is more direct, avoiding confusion with news or with
the stable big red button (remember: we also have dev installers).
Temporarily drops 'cfitsio' which started weeks ago to change its hash.
In my tests, after "fixing" the hash, sometimes it alternates to another,
which breaks the build again. This is too hard to maintain and unsafe(?).
Following e93e0f49
This defaults 'Segoe UI' for Installer texts and for license's body. This
font is still used as the default in Windows 11 and even in online docs.
We are also bumping to 9 pt as outlined in the final Win32 guidelines:
https://learn.microsoft.com/en-us/windows/win32/uxguide/vis-fonts
I've done microscopic pixel checks and confirmed that this produces a Win32
compliant design, which renders kerning and spacing equal as 'Win + R' and
UWP Notepad (please note that some "legacier" Win32 apps renders different)
---
Also, change from Lucida Console to 'Consolas' (default console font since
at least Win8). Today it is Cascadia Mono but not included with Win 10.
This makes the output more complete but at the same time less
polluted regarding the bundling scripts (base, deps and sym).
---
Also, packaging scripts (inno and msix) were touched as consequence
of experiments about drop bundling, for saving space (~20MB), of:
- etc/gtk (MSYS2 don't build immodules)
- share/ghostscript (macOS .dmg don't package it)
- share/libthai (macOS .dmg don't package it)
- share/locale (???, decided to kept only gtk and iso-codes to test)
- share/man (installer don't package it, maybe because Windows can't open)
- share/poppler (macOS .dmg don't package it)
- bin/bzip2.exe (???)
- bin/gdk-pixbuf-query-loaders.exe (loaders.cache already bundled)
- bin/gspawn*.exe. (???)
- share/gir-1.0 (installer don't package it)
- share/lua (???)
- share/vala (installer don't package it)
The "???" ones seems to not be used according to my local tests. If
they are, a bug will appear so we will be able to know, comment in
the script (for maximum clarity towards the future) then revert.
In the process, make the Installer localization process less 'masochist'
(in other words: less manual). The custom '.xml' was the only way since
the iso-codes, Dammed Lies and Inno namings don't match.
This reverts commit 085d8a02b5.
For some reason, Windows doesn't allow us to use that feature
(confirmed this after using the version from the Store itself,
not the .msixbundle)
The installer is probably the most complex packaging format so one of the
biggest examples of packaging love over the years into it in our repo.
But the readability of installer scripts... leave a lot to be desired.
Now, the main script is organized following the order of installer pages
with some comments. This will make easier to future contributors work
(but this commit probably isn't perfect since it's a big change to read)
We have counter measures to avoid installing while gimp opened at 'prep time',
but until now none for the installer itself. Let's fix it with SetupMutex.
The final UX is just less worse (than today): the user needs first to select a
language, then only after this the dialog about two installer will be prompt.
Seems that Inno devs willn't fix: https://github.com/jrsoftware/issrc/pull/461
Following 0199faac
We use arm64 runners for the installer job to avoid queues, but...
llvm-objcopy is well know to lack parity and one sympthom is the
'invalid SymbolTableIndex' warning when handling x86 debug symbols.
This commit just makes the installer display a (unchecked) checkbox, like Win
pkg manager (WinGet) so it's doesn't change the actual behavior. It also don't
break string freeze since the .isl string already exists and have translations.
The benefit to some users is having one less click to launch GIMP: [checkbox] >
[close]; instead of: [close] > [start menu icon] > [gimp in 'recommendations'].
This style mimics the color scheme used by Win32 applications since Vista (and
still used in Win11). The all grey scheme ended in XP. Let's move on.
But we will keep the actual window size and without being able to resize since
these two Inno features (WizardSizePercent and WizardResizable) are broken.
(The progress bar was changed too to conform with the defaults and uninstaller)
Now, the installer files match the meson config.h naming:
'gimp_version', 'gimp_app_version' and 'gimp_api_version'.
"devel" was replaced by meson 'gimp_unstable' too.
util_version was dropped since it lost its utility.
1) Right now, MS Partner Center doesn't tell us if the .pdb bunbled as .appxsym
are fine and we only have "Unknow" dumps in the Health page from MS Par. Cen.
My theory, according to my tests with 'SymChk', 'PDBCopy' and 'llvm-pdbutil',
is that this is happening because .pdb from clang or gcc are not "perfect", but
I really have no proof to afirm this, since Partner Center tell us nothing
about them, and we don't even know if the .appxsym were uploaded to begin with.
---
2) The compiler can't generate DWARF (.debug*) symbols when generating .pdb,
which breaks debugging in DrMingw and even lldb according to my tests.
(This is not a fault of the .pdb format but a circumstance: our debuggers
only support DWARF, which is the format already used by MSYS2 packages)
---
So, the .pdb will return probably only in the potential vcpkg transition.
Previous changes broken local building courtesy and probably DirectInput too.
Noticed that the universal variables can not work in some contexts, fixed.
Also, done some little cleanups.
This make them easier to understand and to run local, as we do with native ones
(They can't be run local yet, further changes will try to acomplish this a bit)
* Since the appimage is for testing purposes, let's use system 'libc'
* Hardcode python path since the wildcard wasn't expanding misteriously
* Improve system theme recognition making way simpler
Port all plug-ins to retrieve the layers
directly from the image rather than
having them passed in. This resolves some
issues with introspection and sets the
foundation for future API work.
The move to Clang in 85ed2847 didn't take into account caching,
Now, this is fixed or, better saying, reintroduced(?): e545116b
Also, flatpak-builder output was confined to the .log file to make
possible checking the end of the CI output, which was being cut.
The wrapper, like in native artifact, exists just for avoid searching the exe.
But, if the user wanted, wasn't possible to execute gimp-*.exe directly since
the wrapper was needed to generate loaders.cache. Now, loaders.cache are built.
This improves the gimp.cmd wrapper wording a bit to not need the README
anymore. Our dev site was improved some time ago to clarify this:
the Windows builds aren't not incomplete if run correctly (under MSYS2)
Also, add cleanup commands to make the .typelib wizardy less invasive.
Without this, Partner Center refuses the .MSIX by not matching the
entry Identity and DisplayName (which isn't the same as the stable).
Also, rework the naming of the .msix's to be more Microsoft-ish
This commit is mostly a little cleanup:
- Reduce redundancy in Local compatibility layer regarding build commands,
which partily reverts !1171 regarding buggy "ninja && ninja install"
- Reduce redundancy of some variables (MSYS_PREFIX and GIMP_DISTRIB)
- Remove manual QOI install since MSYS2 granted a exception in the win32 drop
This commit also reverts 7cca69cd, a fix from the autotools era that isn't
actually needed according to my tests in CI.
Partially reverts af79bbe0 (regarding .debug "extraction" in cross builds)
Now, we call the split debug script from the main bundling script, which makes
similar to our macOS .app bundling script. This cleans a bit of code in .yml
and make things clearer to the mere mortals in the scripts.
Native arm64 and x64 TWAIN binaries are not built anymore. No need to
discard at dist time what doesn't exist. (32-bit TWAIN not affected)
Also, start to using our own ARM64 runners to do the dist. This is more
reliable since the shared x64 runners can (and indeed) cause long queue.
They were generating a distracting output in CLANG* shells, as noted by
@lillolollo in a comment from MR: Infrastructure/gimp-web-devel!65
In the process, make AppImage and Windows (native) scripts use these
variables, without hardcoding the same variables from .yml anymore.
Creating a temporary config directory for the in-build GIMP (run as a tool or
for unit-testing) is not done as a build target anymore, but in the
in-build-gimp.sh script as a unique temp directory, then cleaned out on exit.
This has a few advantages:
- It is properly cleaned out once the build ends (instead of leaving a full
config dir as trash inside the build dir).
- It is not reused from one build to another (with risk of carrying bugs and
issues over).
- Every use of the in-build GIMP will have its own config directory, and in
particular when they are called in parallel.
As a side update, make sure that all `gimp_exe` runs depend on
`gimp_exe_depends`.
This is recommended by Microsoft for "*nix apps" and adding it to the
.exe manifest will avoid disparity between MSIX and other versions
(e.g. CI crossbuilds and CI native builds/installer/local builds).
Bruno note: This commit was repurposed after ad132ad0.
---
For some reason, on the Windows/Aarch64 runners only, submodule updates fails
with:
> + git submodule update --init
> From https://gitlab.gnome.org/GNOME/gimp-data
> * branch d1041a4d97871c85538938872d0ace69d2fc44c2 -> FETCH_HEAD
> error: Your local changes to the following files would be overwritten by checkout:
> images/README.md
> images/logo/meson.build
> images/meson.build
The big question was: why are there edits to begin with? This is supposed to be
a brand new clone!
After a bit of investigation, it turns out that the whole files have incorrect
EOL endings (a `git diff` returns every possible lines in the whole
repositories, but `git diff --ignore-cr-at-eol` returns nothing). It looks like
this breaks the submodule update.