GitLab

This will make Python plug-in available for the main build step.
Also it should make the distcheck step shorter, hence hopefully will
prevent it from timing out so often.

Looking further at this widget, many things are not right. Here are the
changes:
- Use binary prefixes (i.e. kibibyte, mebibyte and gibibyte) instead of
  decimal ones. We are making binary shifts so we were actually showing
  the wrong units.
- Round the value to the closest integer when showing it, not towards 0.
  Otherwise I had cases where it was showing 7GiB for an actual value of
  7.69GiB (default as computed by GIMP from my actual physical memory).
  Note that I am actually unsure even rounding makes sense. Shouldn't we
  rather show double values with a few digits after the decimal points?
  For such values, I think it would make sense.
- Do not edit the internally saved accurate value when the entry is
  edited to the same less accurate value as our saved value would be
  shown too. In particular when changing the display unit to a bigger
  one, we don't want to lose accuracy. This is especially true for low
  values. Say you don't have a lot of memory and you set the Tile cache
  size to 1.5GiB (1536MiB), you certainly don't want it to become either
  1 or 2GiB when switching display unit to GiB. Now even if the number
  will still display with less accuracy, the internal value will stay
  accurate.

(cherry picked from commit d886bb1b)

This bug doesn't happen when setting value through the GUI as in such
case, the unit never changed. It happened when setting a value which
could not be properly displayed by current unit (typically smaller than
1 in this unit, or with remainder).
In such a case, we should not manually set private->shift before
gimp_int_combo_box_set_active(), or the callback was failing to
reconfigure the GtkAdjustement, in particular min and max values.

As a consequence, hitting a Preferences reset, with a GimpMemSizeEntry
in Gigabytes, it got reset to Kilobytes with the max values capped at
4096. So I realized a Reset ended me with a Tile cache size of 4096 KB
in particular, which is of course ridiculously small and would be a
problem if one doesn't notice the issue immediately.

(cherry picked from commit 0be4e5c1)

Even without Python enabled, these are used for the `dist` step, hence
also for distcheck. These checks are not blocking normal builds though.

Unfortunately the dist target requires pygtk.
Note that even adding --disable-python to DISTCHECK_CONFIGURE_FLAGS
doesn't help as pygtk-codegen-2.0 is needed to generate various .defs
file added to the DIST.
Yet pygtk is absent from Debian testing nowadays because it is too old
and unmaintained. So let's build pygtk ourselves.

Among the few tricks to make pygtk build fine:
- force Python2 by setting PYTHON env variable (detect Python 3
  otherwise).
- patch to allow building against latest Pango.

Undoing the apparent mistake (revert 27e78849) and changing the 2nd
occurrence instead.

(cherry picked from commit be0e63aa)

(cherry picked from commit d7de0781)

Add a new "Messages" boolean parameter to performance logs, which,
when set, records GLIB log messages in the performance log as
markers, with an accompanying sample capturing their backtrace.
This option is enabled by default.

(cherry picked from commit 9e0fdc8e)

... which set/remove a GLIB log handler for all domains used by
GIMP.

Use the new functions in errors.c, instead of using
g_log_set_handler() directly.

(cherry picked from commit a143bfdf)

(cherry picked from commit 0fae4613)

(cherry picked from commit ca6f5774)

(cherry picked from commit 1d70fee1)

(cherry picked from commit d72e6b3a)

(cherry picked from commit 9b29d178)

(cherry picked from commit d11cbbbb)

Add a new performance-log-progressive-coalesce.py tool, which
coalesces partial address maps in progressive performacne logs into
a single global address map, suitable for processing by the rest of
the tools.

Use the new tool as part of the pipeline in performance-log-viewer.

(cherry picked from commit 76f9e5fe)

Add a new performance-log-close-tags.py tool, which closes unclosed
tags in incomplete performance logs, allowing the rest of the
perofmance-log tools to process them.  This is necessary for
unfinished progressive logs.

Use the new tool as part of the pipeline in performance-log-viewer.

(cherry picked from commit e5fe1ef1)

Add an option to record progressive performance logs.  Progressive
logs contain complete information after each recorded sample, by
writing partial address maps at each sample, containing all new
addresses introduced by the sample.  Furthermore, when recording a
progressive log, the output stream is flushed after each sample.

This allows recording complete logs even in cases where they can't
be properly terminated, such as when GIMP crashes or freezes in the
middle of the log.

Progressive logs are disabled by default, since they potentially
increase the sampling cost.  They can be enabled through a toggle
in the log file-dialog, or through the
GIMP_PERFORMANCE_LOG_PROGRESSIVE environment varaible.

(cherry picked from commit 146c2343)

When recording a performance log, allow setting the log parametrs
through the file dialog.  Currently, this includes the sample
frequency, and the option to include backtraces.

These options are still controllable through the
GIMP_PERFORMANCE_LOG_SAMPLE_FREQUENCY and
GIMP_PERFORMANCE_LOG_BACKTRACE environment variables.  When set,
the variables override the values entered through the UI.

(cherry picked from commit 126002c5)

In the Windows implementation of gimp_backtrace_new(), return NULL
if the backend is not fully initialized.

(cherry picked from commit 95b69dd7)

In gimp_edit_selection_tool_translate(), perform the active-item
checks even when the effective translation offset is 0, and only
bail afterwards.  This avoids erroneously calling
gimp_tool_message_literal() with a NULL message in this case.

(cherry picked from commit 3a462a5e)

Runs in the gimp stage in parallel to the make && make check build.
May catch dist errors.

Distcheck is allowed to fail. The idea is that being able to build GIMP
is more important and should get fast feedback, while make distcheck is
usually only used when a release is prepared.

Reviewer note: not sure about the "allowed to fail" part because we also
want to fix dist bugs as soon as we can (it's easier to find a bug
source when we get notified as soon as a guilty commit breaks the CI).
But anyway this is a step forward from current CI. So let's merge this
for now.

In file_open_image(), mount non-native files *before* looking up a
file-proc.  Previously, we'd only mount the file after the initial
lookup, and fail to perform a second lookup if the mount succeeded,
leaving us with a NULL file-proc and a subsequent segfault.

Additionally, simplify the rest of the remote-file code-path.

(cherry picked from commit 0829dba9)

(cherry picked from commit 81a3370e)