Still getting some crash reports on nightly with a 1MB stack. Let's try one more time
with 2MB just to see if we can eliminate those. If this fails, then we may need to consider
more drastic approaches like disabling OMTP in this case.
Differential Revision: https://phabricator.services.mozilla.com/D47080
--HG--
extra : moz-landing-system : lando
This requires replacing inclusions of it with inclusions of more specific prefs
files.
The exception is that StaticPrefsAll.h, which is equivalent to StaticPrefs.h,
and is used in `Codegen.py` because doing something smarter is tricky and
suitable for a follow-up. As a result, any change to StaticPrefList.yaml will
still trigger recompilation of all the generated DOM bindings files, but that's
still a big improvement over trigger recompilation of every file that uses
static prefs.
Most of the changes in this commit are very boring. The only changes that are
not boring are modules/libpref/*, Codegen.py, and ServoBindings.toml.
Differential Revision: https://phabricator.services.mozilla.com/D39138
--HG--
extra : moz-landing-system : lando
Currently it's completely unclear at use sites that the getters for `once`
static prefs return the pref value from startup, rather than the current pref
value. (Bugs have been caused by this.) This commit improves things by changing
the getter name to make it clear that the pref value obtained is from startup.
This required changing things within libpref so it distinguishes between the
"base id" (`foo_bar`) and the "full id" (`foo_bar` or
`foo_bar_DoNotUseDirectly` or `foo_bar_AtStartup` or
`foo_bar_AtStartup_DoNotUseDirectly`; the name used depends on the `mirror` and
`do_not_use_directly` values in the YAML definition.) The "full id" is used in
most places, while the "base id" is used for the `GetPrefName_*` and
`GetPrefDefault_*` functions.
(This is a nice demonstration of the benefits of the YAML file, BTW. Making
this change with the old code would have involved adding an entry to every
single pref in StaticPrefList.h.)
The patch also rejigs the comment at the top of StaticPrefList.yaml, to clarify
some things.
Differential Revision: https://phabricator.services.mozilla.com/D38604
--HG--
extra : moz-landing-system : lando
The patch also removes the layers.mlgpu.enable-container-resizing pref, because
it's dead.
Differential Revision: https://phabricator.services.mozilla.com/D36159
--HG--
extra : rebase_source : e215d584aed18f865d2e8d00a78e76e9b0323e6e
gfxPrefs Live preferences are almost identical to StaticPrefs.
We leave aside for now those that set a custom change callback as this feature isn't yet supported in StaticPrefs.
Differential Revision: https://phabricator.services.mozilla.com/D31256
--HG--
extra : moz-landing-system : lando
gfxPrefs Live preferences are almost identical to StaticPrefs.
We leave aside for now those that set a custom change callback as this feature isn't yet supported in StaticPrefs.
Differential Revision: https://phabricator.services.mozilla.com/D31256
--HG--
extra : moz-landing-system : lando
gfxPrefs Live preferences are almost identical to StaticPrefs.
We leave aside for now those that set a custom change callback as this feature isn't yet supported in StaticPrefs.
Differential Revision: https://phabricator.services.mozilla.com/D31256
--HG--
extra : moz-landing-system : lando
Now that we have C++14 support we can capture a move only object in a lambda expression.
--HG--
extra : rebase_source : 232639ba334520cf9d38d68190af8fdcd4aa454d
These method names and ordering have gotten out of sync because of
the recent churn.
Differential Revision: https://phabricator.services.mozilla.com/D3288
--HG--
extra : rebase_source : 42bceaeb66a0df4808981b8c9cb0ed70b23f5a30
extra : histedit_source : 228024efe8de4e149f7e7ca66a2bb078bba820ce
This may have been needed at some point, but all the important code for
EndLayerTransaction is in CompositorBridgeChild behind a lock, so this
should be safe.
--HG--
extra : rebase_source : bda4080bc04afa95954732050df7bd25c8177752
extra : histedit_source : 9386583dd923bf0ae44ff783d3ef1c6692944c77
There should only ever be at most four TextureClients here, so
allocated a vector seems wasteful.
--HG--
extra : rebase_source : 6b0f9f7749461eb39cd3c6c6bf7917152ffc9aab
extra : histedit_source : b5539d02c294596a5147dc55b417ef7970f8c0cd
This was needed when there were multiple types of CapturedPaintStates but
is not anymore.
--HG--
extra : rebase_source : 3648edccca7c73e3e3aa7a5c3e0d48d12d6324c5
extra : histedit_source : 021814c10b6578970c3a6d234c1e6641ad26b095
This commit moves ContentClient from creating a CapturedBufferState for
buffer operations, to performing all of those operations on the
DrawTarget(Capture). Creating a DrawTargetCapture is now performed
by the RotatedBuffer when we BeginPaint, all operations are performed
on this capture, and then it's returned to the ClientPaintedLayer
as a PaintTask.
This commit is an involved refactoring of ContentClient and RotatedBuffer
to get this all to work. Here are the major parts:
1. RotatedBuffer is refactored to always perform operations on a single
DrawTarget, which may be a single DT, dual DT, or capture.
2. RotatedBuffer adds BeginCapture and EndCapture methods to switch
which DT is used in operations
3. ContentClient uses the RB capture methods when we are async painting
4. CC::BeginPaint is refactored to only perform capturing on a single
RotatedBuffer. This is because we can't have the output of one
PaintTask be the input of a different PaintTask due to the design
of the Snapshot API.
a. This can occur, today, by doing a FinalizeFrame only to later
fail to Unrotate the buffer, causing a new RB to be created
and painted into
b. The previous PaintThread code worked because it used the
buffer operations which didn't use Snapshot's
c. This is fixed by not doing FinalizeFrame on a buffer if we
realize we cannot unrotate it, and switching to initializing
a buffer using the front buffer which should be up to date.
d. I don't like touching this code, but it passes reftests,
might be a performance improvement, and I've tested it on
known regressions from the last time I messed up this code.
5. CC::PrepareForPaint is inlined into BeginPaint because dual draw
targets can be cleared correctly from a previous commit
6. The code paths in ClientPaintedLayer are unified because we no
longer need to special case this beyond setting the correct
ContentClient flag.
7. CapturedPaintState and CapturedBufferState are removed in favor
of PaintTask. Additionally EndLayer is no longer needed as all
quadrants of a rotated buffer are in the same capture, so we
don't need special case flushing code.
MozReview-Commit-ID: 9UI40dwran
--HG--
extra : rebase_source : 809d9816970648468de972c30b0c230c2f21e27b
extra : source : 405ad351821813333c0e989b93e2aeb49ba8552c
This commit renames CapturedTiledPaintState to PaintTask as in a future
commit I will fold CapturedPaintState into it.
MozReview-Commit-ID: 8py7SrK4s29
--HG--
extra : rebase_source : 7abdf127351cdc82ee4c40112dce7150bdb67243
extra : source : 01110727f2e9e0846fc06997653e04860efb23dc
This commit refactors TiledContentClient to not create PaintThread
buffer operations, but to instead perform all of these operations
on the DrawTarget(Capture). This simplifies the code dramatically
and allows us to add flushing behavior to DrawTargetCapture in a
future commit.
With this change, CapturedTiledPaintState is simply a container
for a DrawTarget, DrawTargetCapture, and keep-alive TextureClients.
Part of this commit is moving the logic of locking the texture
clients, constructing a dual draw target, and constructing a capture
into TiledContentClient so it can be shared.
MozReview-Commit-ID: 2rwz9aDI737
--HG--
extra : rebase_source : 4ac317f632c0a2c21480bc88e6246f4dc0daf0be
extra : source : 56d967e03ee225e032034ffd193b6f42b343226b
This commit adds an operation to perform 'edge padding' on a draw
target. By default this is performed using LockBits, but it's
overriden in DrawTargetTiled and DrawTargetCapture to propagate
the call so it functions correctly.
This helps TiledContentClient move from applying this operation
on a per texture client basis, to being able to do it on the
DrawTargetTiled after painting. This in turn helps move all
paint thread operations into DrawTargetCapture.
MozReview-Commit-ID: 2ncOTxGXQfk
--HG--
rename : gfx/layers/BufferEdgePad.cpp => gfx/2d/BufferEdgePad.cpp
rename : gfx/layers/BufferEdgePad.h => gfx/2d/BufferEdgePad.h
extra : rebase_source : a3315644fe31f2a432935dcbfdb9969c58b691e1
extra : source : 699c954992f87db7fc792f5562090de42a8162cb
system-info is a stub on Tier3 platforms while physical vs. logical
difference only matters for hyper-threading. As hyper-threading
is usually available on CPUs with more than 2 physical cores this
change has no impact there as the default is clamped to [1, 4].
However, on Intel i3-* CPUs with 2 physical and 4 logical cores this
bumps the default from 1 to 3.
MozReview-Commit-ID: 1Yh8rJL2JcN
--HG--
extra : rebase_source : 5c563ec8e388a3fd05a0650e8d4c330d48675332
This commit moves ContentClient from creating a CapturedBufferState for
buffer operations, to performing all of those operations on the
DrawTarget(Capture). Creating a DrawTargetCapture is now performed
by the RotatedBuffer when we BeginPaint, all operations are performed
on this capture, and then it's returned to the ClientPaintedLayer
as a PaintTask.
This commit is an involved refactoring of ContentClient and RotatedBuffer
to get this all to work. Here are the major parts:
1. RotatedBuffer is refactored to always perform operations on a single
DrawTarget, which may be a single DT, dual DT, or capture.
2. RotatedBuffer adds BeginCapture and EndCapture methods to switch
which DT is used in operations
3. ContentClient uses the RB capture methods when we are async painting
4. CC::BeginPaint is refactored to only perform capturing on a single
RotatedBuffer. This is because we can't have the output of one
PaintTask be the input of a different PaintTask due to the design
of the Snapshot API.
a. This can occur, today, by doing a FinalizeFrame only to later
fail to Unrotate the buffer, causing a new RB to be created
and painted into
b. The previous PaintThread code worked because it used the
buffer operations which didn't use Snapshot's
c. This is fixed by not doing FinalizeFrame on a buffer if we
realize we cannot unrotate it, and switching to initializing
a buffer using the front buffer which should be up to date.
d. I don't like touching this code, but it passes reftests,
might be a performance improvement, and I've tested it on
known regressions from the last time I messed up this code.
5. CC::PrepareForPaint is inlined into BeginPaint because dual draw
targets can be cleared correctly from a previous commit
6. The code paths in ClientPaintedLayer are unified because we no
longer need to special case this beyond setting the correct
ContentClient flag.
7. CapturedPaintState and CapturedBufferState are removed in favor
of PaintTask. Additionally EndLayer is no longer needed as all
quadrants of a rotated buffer are in the same capture, so we
don't need special case flushing code.
MozReview-Commit-ID: 9UI40dwran
--HG--
extra : rebase_source : 2f63464c1f8ca03992700b33838c4aa56608f872
This commit renames CapturedTiledPaintState to PaintTask as in a future
commit I will fold CapturedPaintState into it.
MozReview-Commit-ID: 8py7SrK4s29
--HG--
extra : rebase_source : 1b5259cca6520761ae99e64157d047441b90b563
This commit refactors TiledContentClient to not create PaintThread
buffer operations, but to instead perform all of these operations
on the DrawTarget(Capture). This simplifies the code dramatically
and allows us to add flushing behavior to DrawTargetCapture in a
future commit.
With this change, CapturedTiledPaintState is simply a container
for a DrawTarget, DrawTargetCapture, and keep-alive TextureClients.
Part of this commit is moving the logic of locking the texture
clients, constructing a dual draw target, and constructing a capture
into TiledContentClient so it can be shared.
MozReview-Commit-ID: 2rwz9aDI737
--HG--
extra : rebase_source : 16a4b87263f28b32f5bcb5fd6d9756548f137e11
This commit adds an operation to perform 'edge padding' on a draw
target. By default this is performed using LockBits, but it's
overriden in DrawTargetTiled and DrawTargetCapture to propagate
the call so it functions correctly.
This helps TiledContentClient move from applying this operation
on a per texture client basis, to being able to do it on the
DrawTargetTiled after painting. This in turn helps move all
paint thread operations into DrawTargetCapture.
MozReview-Commit-ID: 2ncOTxGXQfk
--HG--
rename : gfx/layers/BufferEdgePad.cpp => gfx/2d/BufferEdgePad.cpp
rename : gfx/layers/BufferEdgePad.h => gfx/2d/BufferEdgePad.h
extra : rebase_source : ab850358a763853d50d1f374f28e67a197740443
system-info is a stub on Tier3 platforms while physical vs. logical
difference only matters for hyper-threading. As hyper-threading
is usually available on CPUs with more than 2 physical cores this
change has no impact there as the default is clamped to [1, 4].
However, on Intel i3-* CPUs with 2 physical and 4 logical cores this
bumps the default from 1 to 3.
MozReview-Commit-ID: 1Yh8rJL2JcN
--HG--
extra : rebase_source : 77613cbb99c14f19217592080bfd51ea2194422b
This commit ports over the last remaining operation for tiling that doesn't work
on the paint thread.
The difficult part for edge padding is that it is done outside of ValidateTile
so it doesn't have an associated CapturedTilePaintState to be added to as an
operation. We need it to be in the same paint state so that it's guaranteed
to be run after painting has finished.
This commit changes edge padding to instead be decided inside of ValidateTile
and either sent to the paint thread if there is OMTP or executed right away.
MozReview-Commit-ID: JDD4rH1fVwW
--HG--
extra : source : 9b0a54842d3169960a606fa1dd335acf6aa70dbe
extra : intermediate-source : bcbab66c16c5cc2b917f12b4481bbbb8fe3eb097
This was done automatically replacing:
s/mozilla::Move/std::move/
s/ Move(/ std::move(/
s/(Move(/(std::move(/
Removing the 'using mozilla::Move;' lines.
And then with a few manual fixups, see the bug for the split series..
MozReview-Commit-ID: Jxze3adipUh
On windows it's possible for us to fallback from D2D to Skia rendering at any time due to a device reset.
If this happens with `enable-tiles-if-skia-pomtp` enabled we could begin to use tiling. This can cause a
crash if we never created the worker threads because at initialization time we weren't using tiling.
Another way to fix this would be to dynamically create the worker threads in UpdateRenderMode if we
have switched to skia. That's a larger change and more might be required, so I'd rather just fix the
crash for now.
This commit also fixes a pref that should be `Once` instead of `Live`.
MozReview-Commit-ID: JQidXPjI7ER
--HG--
extra : amend_source : ba9a3746faea3a7355251b7e97e3f7c5dc1ba06b
There's a circular dependency between `UsesTiling` and `InitOMTPConfig` because
we try to disable OMTP if we will be using tiling and edge padding is enabled.
Now that edge padding is disabled everywhere except android it should be safe
to assume that if edge padding is enabled then we'll be using tiling as well
and should disable OMTP.
This commit also removes a check on `UsesTiling` from CalculateWorkerCount
as it is redundant.
MozReview-Commit-ID: 1ruWPwXfLwO
--HG--
extra : rebase_source : d489c52c728d2ff356d2413b9b1044dfa3f63135
This commit adds a paint worker thread pool to PaintThread, and dispatches
tiled paints to it. The thread pool is only created if tiling is enabled,
and its size is set by 'layers.omtp.paint-workers' and defaults to 1. If
-1 is specified, it will be sized to 'max((cpu_cores * 3) / 4, 1)'.
The one tricky part of dispatching tiled paints to a thread pool is the
AsyncEndLayerTransaction message that must execute once all paints are
finished. Previously, this runnable would be queued after all the paints
had been queued, ensuring it would be run after they had all completed.
With a thread pool, there is no guarantee. Instead this commit, uses
a flag on CompositorBridgeChild to signify whether all of the paints
have been queued ('mOutstandingAsyncEndLayerTransaction'), and after
every tiled paint it is examined to see if that paint was the last
paint, and if it is to run AsyncEndLayerTransaction. In addition,
if the async paints complete before we even mark the end of the
layer transaction, we queue it like normal.
The profiler markers are also complicated by using a thread pool.
I don't know of a great way to keep them working as they are per
thread, so for now I've removed them. I may have been the only
one using them anyway.
MozReview-Commit-ID: 5LIJ9GWSfCn
--HG--
extra : rebase_source : 0c26806f337a1b4b1511945f9c72e787b426c5ba
Sylvestre Ledru
Ryan VanderMeulen
Lee Salzman
Ryan Hunt
Nicholas Nethercote
Jean-Yves Avenard
Gurzau Raul
Bas Schouten
Jan Beich
Brindusan Cristian
Cosmin Sabou
Csoregi Natalia
Emilio Cobos Álvarez
Markus Stange
Noemi Erli