Just a small code rework, no functional change.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

If either the GLX_MESA_swap_control or the GLX_EXT_swap_control
extension is available, use it to set the swap interval to 1.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

When using multiple buffers, no need to set all possible buffers, just
set the right number of buffers actually used per vblank backend.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

If GLX was configured (at build time) to use multiple buffers, we would
still use a single GLX drawable.

Assign a GLX drawable per buffer instead - Note that this doesn't make
any difference in the default case, as GLX uses a single buffer anyway.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

We would not allow transients for group to be minimized as long as a
parent is still visible.

Be more lenient with those, allow those to be minimized independently
of their parents.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes: #349

The window menu entries would check for valid transients or modal dialog
window to enable or disable the above/below entries.

Yet the actual above/below function would simply check for any transient
or modal dialog, hence denying the request for fake dialogs without an
actual parent window.

Fix the logic to remain consistent, and allows fake dialogs to be placed
above or below other windows.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes:#458

When an offscreen window is damaged, there is no need to repaint any
screen region.

Limit damage to the screen region only.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

When using no vblank, with a single buffer, no need to render the entire
screen.

Optimize rendering by limiting the repaint to the damage region.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

When using Xpresent, the compositor maintains multiple pixmap buffers
and therefore needs to accumulate damages for those multiple buffers.

However, when using GLX or no vblank method at all, there is only one
buffer pixmap and there is no need to accumulate damage between buffers.

Currently, the compositor would accumulate damage indefinitely when
using GLX or no vblank, hence damaging way more than actually necessary.

Fix this by accumulating damages only when there is more than one
buffer.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

By definition, the content of the back buffer is undefined, therefore we
need to copy the entire content on flip with GLX.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

XSync can fail, handle the XError gracefully.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes: #448

Typically, the opaque region applies to the client window.

When shaded, the client window is unmapped and the frame shrunk to its
minimum height.

When shaded, the opaque regio nshould simply be ignored as the client
window is no longer visible.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes: #452

The opaque region was clipped against the compositor window extents,
which might be larger than the window as it can also include the drop
shadows of the server-side decorations.

Clip the opaque region against the border size and the client size if
available instead.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

The client might be gone, and trying to kill the X11 client might cause
an unhandled XError.

Make sure to protect XKillClient against XError.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes: #453

If the opaque region is correctly translated and intersect with the
actual window extents, there should be no need to generate additional
damages when the opaque region is updated.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

The opaque region is set by the client, and applies to the client
window.

Therefore the opaque region should be translated to match the actual
client window, not the composited windows which may be the window
manager frame window.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes: #452

Small optimization, the clientSize represents the actual client region,
there is no need to recompute that each time when painting.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

We now store the current tile mode along with the client, that makes it
possible to toggle back and forth between tile modes and back to the
initial position/location.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

No need to preset the size when re-computing the tile size.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

Tiling works by moving and resizing windows to a specific location  on
screen.

GTK clients using client-side decorations (aka CSD) draw the drop shadow
themselves and use the property _GTK_FRAME_EXTENTS to let the window
manager or compositor know about the actual relative frame size and
location.

GTK automatically removes the drop shadow on some specific window
states, such as fullscreen or maximized, and restores the drop shadow
when the window returns to a normal state.

If that update occurs after the tiled size was computed, the drop shadow
is not accounted for and the window will appear smaller than the actual
expected tiled size.

To avoid that issue, the window manager should recompute the tiled size
whenever the _GTK_FRAME_EXTENTS is updated.

Store the tiled mode and recompute the tiled size/location of windows
when the frame extents or the screen size change.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes: #450

Right now, resizing a window would immediately update the compositor
representation of that window.

The compositor, however, is able to catch the changes on configure
notify, so there is no need to force that update early.

Remove the resize of the compositor window on each configure requests.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

The opaque region is set and updated by the client, which may lag when
the window is resized.

That may leave unpainted areas when resizing a window as the opaque
region set by the client could be actually larger than the window itself
and hence prevent areas to be updated while they should, causing
artifacts on screen.

Make sure to limit the clipping of the opaque region by the window
extents to avoid the issue.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

If a client updates the opaque region of its window, regions previously
opaque may become transparent, meaning that they would need to be
repainted to let other windows underneath see through.

The compositor already computes that and adds damage for the relevant
regions.

Make sure to do that only for visible and redirected windows, other
windows shouldn't trigger that though.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

xfwm4 sends a synthetic configure event when required as per the ICCCM
specification, but a force redraw as when maximizing a window while
moving generates a fake move, which defats the logic.

Make sure to send a synthetic configure event on force redraw, that
fixes an issue with VTE terminals (like xfce4-terminal) which do not
resize correctly if maximized while pushing toward the top of the
screen.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

If the window has no opaque region, no point in trying to clip it out
and return immediately.

No functional change.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

When using multi buffers (Present, GLX), if the client updates its
opaque region in between (e.g. when resizing the window) we may cull out
regions on screen that won't get updated, leaving trails of unpainted
areas.

To avoid that, make sure we add damage for the delta in opaque region.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

Opaque region is a hint set by the clients via the _NET_WM_OPAQUE_REGION
property to indicate the compositor which regions of an ARGB window are
considered opaque and therefore can be culled out during repaint.

While this is of little use with regular RGB windows, with the planned
move to client-side decorations in xfce, we'll have more and more ARGB
windows, making such an optimization a beneficial improvement.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

Xfwm4 4.14 made the cursor zoom along with the desktop. While this
sounds like a sensible thing to do, it's not necessarily an improvement
for accessibility, because an enlarged cursor also hides what's being
zoomed underneath.

Add a new option to control whether the pointer is zoomed along with the
rest of the desktop.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
Closes: xfce/xfwm4#439

Make Xi2 support selectable at configure time and disable it by default.

Xi2 in the window manager causes some very subtile bugs with legacy
applications using the core X11 protocol.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
xfce/xfwm4#445
xfce/xfwm4#427

Signed-off-by Olivier Fourdan <fourdan@xfce.org>
xfce/xfwm4#444
xfce/xfwm4#436
xfce/xfwm4#433
xfce/xfwm4#405

When the compositor is disabled on a given screen, there is no point in
looking for XID in the hash table.

That fixes errors like:

 g_hash_table_lookup: assertion 'hash_table != NULL' failed

When the compositor is disabled.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

Previously we would leak error traps in free_win_data() causing
slowdowns overtime as the list of requests grows.

While fixing the leak is crucial, it also unveiled several places where
actual error traps are missing, causing xfwm4 to abort more often.

Based on XErrors reported, the issue is triggered when freeing pixmaps
or render pictures.

Add more XErrors traps to catch those errors.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>
xfce/xfwm4#433
xfce/xfwm4#435
xfce/xfwm4#436

bob5972 authored 4 years ago and bob5972 committed 4 years ago

On my system when the window manager is mostly idling (0-12 tiny
updates per second), about half the CPU time in xfwm4 is spent in
find_cwindow_in_screen, iterating the linked list looking for a
matching Window id.  Instead, this patch adds a GHashTable to map
Window ids to CWindows, completely eliminating that cost in my
profiles, resulting in a 0.3-1.0% of a core total CPU reduction for
this workload.

Signed-off-by: Michael Banack <bob5972@banack.net>

bob5972 authored 4 years ago and bob5972 committed 4 years ago

The call to g_timeout_add in add_repair seems to be mixing up priority
and callback interval.  Instead, let's call g_timeout_add_full to make
it clear there are two separate fields, and add an appropriate define
for the callback interval, setting it to the same 1ms value it was
currently using.

Signed-off-by: Michael Banack <bob5972@banack.net>

bob5972 authored 4 years ago and Olivier Fourdan committed 4 years ago

On my system, 2-4% of the CPU time in xfwm4 is being spend in
find_cwindow_in_screeni() pulling in entries from the linked list on
a lightly updating desktop.

After this change, 25-50% of that CPU is gone, presumably because the
slice allocators have better locality of the list-entries.

Closes: #416

bob5972 authored 4 years ago and Olivier Fourdan committed 4 years ago

The function free_win_data() may generate X11 errors as the client
window may already be gone.

To avoid those, the entire function code is supposed to be enclosed
within an error trap/push.

A typo in the code would however push twice instead of a balanced
push/pop which causes the list of event sequences in GDK to grow
indefinitely, leading to slowdowns as the list grows over time.

Fix the leak by balancing the error push with an error pop at the end
of the function.

xfce/xfwm4#351

The function compositorCheckCMSelection() would use the atom index in
our table of atoms, instead of the actual atom, so this is unlikely to
work as expected.

Fix the code to use the appropriate atom as it should.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

Prefer using memory allocation functions which initialize the memory
with 0's to be on the safe side.

No functional change.

Signed-off-by: Olivier Fourdan <fourdan@xfce.org>

A pid is not necessarily a long on all systems, so casting the value of
NET_WM_PID directly into a GPid may cause a stack overflow.

Fixes: 49c1a33a
Closes: #411


Signed-off-by: Olivier Fourdan <fourdan@xfce.org>