This also adds back default flush() calls which simply do a flush
without any submit.
Change-Id: Ia8c92bbdecd515d871abfa6364592f502e98656b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/298818
Commit-Queue: Greg Daniel <egdaniel@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
This will not be landed until chrome CL 2269958 lands.
Bug: skia:10425
Change-Id: I2a5081201ca3faed5232e8540086bd4c6f865767
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/299292
Commit-Queue: Adlai Holler <adlai@google.com>
Reviewed-by: Robert Phillips <robertphillips@google.com>
Since gpuSetup can preempt draw's execution it needs to draw the error message too.
This is pulled out of the gpuSetup refactoring.
Change-Id: Iafe06d924fc1b694c59aa3100e9fbe95c4773222
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/299140
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
Change-Id: I09b8f8bb95dac443b64d85340db12f59f8977654
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/299143
Auto-Submit: Florin Malita <fmalita@chromium.org>
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>
Previously, DM destroyed a large number of non-trivial objects at
shutdown time. Because no shutdown order is promised across translation
units by the standard, this can lead to bugs which only reproduce
capriciously, at the whim of the linker.
http://go/totw/110#the-fix-safe-initialization-no-destruction
"Destruction issues are usually solved by defining your static data
in such a way that the destructor never runs. The most common way to do
this is to heap allocate the static object - pointers don't have
destructors."
http://go/cstyle#decision_on_destruction
"Global and static variables that use dynamic initialization or have
non-trivial destructors create complexity that can easily lead to hard-
to-find bugs. Dynamic initialization is not ordered across translation
units, and neither is destruction (except that destruction happens in
reverse order of initialization). When one initialization refers to
another variable with static storage duration, it is possible that this
causes an object to be accessed before its lifetime has begun (or
after its lifetime has ended). Moreover, when a program starts threads
that are not joined at exit, those threads may attempt to access objects
after their lifetime has ended if their destructor has already run."
Change-Id: I54eedcd813295a23923deb925b0ca2adfff69f7d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/297872
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>
This also adds a GPUTEST_FOR_D3D_CONTEXT macro to help with debugging
tests.
Change-Id: I72db01d148755c3bbbbb4d948d441a31dcf9482b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/297717
Reviewed-by: Jim Van Verth <jvanverth@google.com>
Commit-Queue: Greg Daniel <egdaniel@google.com>
The flippity GM doesn't create backend textures but it does perform some custom proxy creation (i.e., using MakeTextureProxyFromData to set the origin) that requires a direct context. This work must be done in onGpuSetup but doesn't entail any fancy lifetime management.
TBR=egdaniel@google.com
Change-Id: I258c0cb66746ca6853a4e228e10407671d7d55d9
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/297697
Commit-Queue: Robert Phillips <robertphillips@google.com>
Reviewed-by: Greg Daniel <egdaniel@google.com>
This reverts commit ddca6ab54a.
Reason for revert: breaking abandoned context bot
Original change's description:
> Fix flippity GM for *ooprddl configs
>
> The flippity GM doesn't create backend textures but it does perform some custom proxy creation (i.e., using MakeTextureProxyFromData to set the origin) that requires a direct context. This work must be done in onGpuSetup but doesn't entail any fancy lifetime management.
>
> Change-Id: Ica4f4f7476778cdf934c3be9ef8c9a28d0d4ba2e
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/297445
> Reviewed-by: Greg Daniel <egdaniel@google.com>
> Commit-Queue: Robert Phillips <robertphillips@google.com>
TBR=egdaniel@google.com,robertphillips@google.com
Change-Id: Ie66a4187a5ff2efee21f60bb9c0d00e2aab3d1e8
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/297696
Reviewed-by: Robert Phillips <robertphillips@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
The flippity GM doesn't create backend textures but it does perform some custom proxy creation (i.e., using MakeTextureProxyFromData to set the origin) that requires a direct context. This work must be done in onGpuSetup but doesn't entail any fancy lifetime management.
Change-Id: Ica4f4f7476778cdf934c3be9ef8c9a28d0d4ba2e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/297445
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
In order to emulate OOP-R's behavior, GM needs to pass GPU-backed resources to a DDL recorder.
This change allows GMs to create GPU resources first (in onGpuSetup w/ a direct context) and then use them in onDraw (with only a GrRecordingContext).
Change-Id: Ifa3002af73eb9926f653fb4c4bf4542c0749d658
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/294336
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
Clang doesn't like it when operator| is applied to two different enum
types.
Additionally, fixed some nearby line wrapping.
Change-Id: I77190c9bc91b53ebc38d184d73a6a244b8f34ce9
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/295795
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Adlai Holler <adlai@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This CL is not fully comprehensive; for instance, it does not contain
fixes for backends that don't compile on Mac. But it does resolve the
vast majority of cases that trigger -Wimplicit-fallthrough.
A few minor bugs were found and fixed, but none that were likely to
affect normal operation.
Change-Id: I43487602b0d56200ce8b42702e04f66390d82f60
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/295916
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Our "header" reading and writing code didn't agree, so we always failed
to recognize cached program binaries. The asserts in the testing sink
failed to notice, because we did get a 100% cache hit rate, but we
immediately discarded the data we received.
We now also check that we didn't insert anything into the cache, as a
proxy for doing any shader compile work. That change, plus the tweak to
set cached=false when the header fields are invalid (like we do if we
encounter problems further in the blob) detected the problem. Adding the
version tag to the start of the encoded blob fixes the test, and means
that program binary caching is actually working again.
This code still looks (and is) fragile. The next CL is going to rewrite
things to use SkReadBuffer and SkWriteBuffer, make the parsing code less
brittle, and give us a more robust way to detect failure anywhere in the
stream.
Bug: skia:9402
Change-Id: I0329f088e0afce3998494d91ef2206e5eb9cac42
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/294599
Reviewed-by: Brian Salomon <bsalomon@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
OOP-R, on the gpu thread, creates the DDL, pre-compiles its shaders, draws it, flushes and then deletes the DDL. This process triggered a bug (cf. https://skia-review.googlesource.com/c/skia/+/292818 and crbug.com/1056730).
Prior to this CL all the programs were compiled and only at the end was any work flushed - thus it was likely that the bound program would be reset to the correct value when rendering.
With this CL, the addition of the flush right before the DDL deletion, makes it more likely that the wrong program will be bound when rendering begins.
Change-Id: I60479bd429e132d8652bbffde6c8b71094be6225
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/292257
Reviewed-by: Adlai Holler <adlai@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
Parse embedded fonts into SkCustomTypefaces, and pass down the text
animation pipeline. Things seem to mostly work for Latin examples.
Most existing Lottie files come with embedded fonts (the option is
enabled by default), so to minimize disruption only use the new
feature as a fallback for typefaces which cannot be resolved otherwise.
Also introduce a builder flag to prioritize embedded fonts over native
(kPreferEmbeddedFonts), and plumb in existing tools for testing.
Change-Id: Ia2a659f76e354fea6081b0f2e0dce1d8bdf63c52
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/291180
Reviewed-by: Ben Wagner <bungeman@google.com>
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Florin Malita <fmalita@google.com>
Bug: skia:10154
This will make it clear that these files are for Android use and
avoid compiling them for other clients.
Update testing tools to use android::skia::BitmapRegionDecoder, but
only if SK_ENABLE_ANDROID_UTILS is defined.
Take this opportunity to clean up the class:
- The base class, which was originally designed to allow switching
amongst different implementations, is no longer needed. Rename
SkBitmapRegionCodec to android::skia::BitmapRegionDecoder
(following the new convention and matching the Java API name).
Continue to inherit from SkBitmapRegionDecoder temporarily, to
allow Android to switch to the new API.
- Use std::unique_ptr instead of passing raw pointers.
Add a test to verify that we only create a BitmapRegionDecoder if
it is one of the supported types.
Change-Id: Ied13fc8acb105fde042553331846d95ae15d6b57
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/287498
Commit-Queue: Leon Scroggins <scroggo@google.com>
Reviewed-by: Derek Sollenberger <djsollen@google.com>
Change-Id: Ic3d9a17674baac8d567ee6229ee2ee0716d9dca9
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/289980
Auto-Submit: Brian Salomon <bsalomon@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Add support for external precomp Skottie layers. This allows embedders
to seamlessly mix custom/Lottie content.
General flow:
* embedders register a PrecompInterceptor callback with
the animation builder
* at build time, Skottie invokes the callback for each pre-composed
layer
- the returned ExternalLayer implementation is used instead of the
Lottie layer payload
- (a nullptr value signals Skottie to use the usual Lottie payload)
* at render time, ExternalLayer::render() is called to defer content
rendering to the embedder
Also implement a sample PrecompInterceptor which attempts to substitute
precmp layers matching a given pattern with external Lottie animations:
precomp_name: "__foo.json" -> Animation("foo.json")
This new mechanism is a generalization of (and supersedes) the old
NestedAnimation hack - so we can remove that.
Change-Id: Id80fe11881c62b8717c2476117c7c03ad5300eef
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/288130
Commit-Queue: Florin Malita <fmalita@chromium.org>
Reviewed-by: Mike Reed <reed@google.com>
The callback lets the caller know when the data uploads to the texture
from the create call are finished. This is important since the caller
cannot delete the backend texture till the gpu is finished on vulkan
and d3d.
This change also removes the hard sync in vulkan during creation.
Change-Id: I660d142219474e22b1337d2b0c81cda66fe18a4b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/286517
Commit-Queue: Greg Daniel <egdaniel@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
Reviewed-by: Robert Phillips <robertphillips@google.com>
Besides better matching Viz's behavior this also reduces a lot of choppiness in the composition RenderTask DAG.
In the previous approach DDL draws and compositing draws would be interleaved resulting in a lot of render target swaps.
This necessitated some reorganization bc I wanted to reuse PromiseImageCallbackContext to manage the tiles' promiseImages.
Change-Id: I513bf060a69ff2bfe0e7b82ae72f149dfede632e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/285056
Reviewed-by: Brian Salomon <bsalomon@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
The GPU thread has privileged access to the GPU so its work can't be easily borrowed.
Change-Id: I1eae4c86ff1c36cc1248f74fc48d76b1c243f0b2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/284764
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
This better matches Chrome's use of DDLs.
With path, image, and text draws stripped out, here is the perf impact of this change:
before CL after CL
w/ DDLs 7.792 1.038
w/o DDLs 0.800 0.876
This perf improvement (in the DDL case) is from backend texture wrapping SkSurfaces being created w/o initialization. The prior method of SkSurface creation was resulting in double clearing of all the surfaces.
This perf improvement won't be seen by Chrome since they've always being using wrapped backend texture SkSurfaces.
TBR=bsalomon@google.com
Bug: 1056730
Change-Id: Ic04d322cad96df845e75437211208495862c6555
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/283866
Reviewed-by: Brian Salomon <bsalomon@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
This reverts commit 7ae9d2fca6.
Reason for revert: Triggering Vulkan Debug layer errors
Original change's description:
> Update DDL test harness to use backendTextures to back tiles
>
> This better matches Chrome's use of DDLs.
>
> With path, image, and text draws stripped out, here is the perf impact of this change:
>
> before CL after CL
> w/ DDLs 7.792 1.038
> w/o DDLs 0.800 0.876
>
> This perf improvement (in the DDL case) is from backend texture wrapping SkSurfaces being created w/o initialization. The prior method of SkSurface creation was resulting in double clearing of all the surfaces.
>
> This perf improvement won't be seen by Chrome since they've always being using wrapped backend texture SkSurfaces.
>
> TBR=bsalomon@google.com
> Change-Id: Ice3993ca125fce37804e58c353c265cf659dbe2f
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/283456
> Reviewed-by: Brian Salomon <bsalomon@google.com>
> Commit-Queue: Robert Phillips <robertphillips@google.com>
TBR=egdaniel@google.com,bsalomon@google.com,robertphillips@google.com
Change-Id: Ife023ede0774ec2cce4c0d6e7708c036347ebf54
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/283648
Reviewed-by: Robert Phillips <robertphillips@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
This better matches Chrome's use of DDLs.
With path, image, and text draws stripped out, here is the perf impact of this change:
before CL after CL
w/ DDLs 7.792 1.038
w/o DDLs 0.800 0.876
This perf improvement (in the DDL case) is from backend texture wrapping SkSurfaces being created w/o initialization. The prior method of SkSurface creation was resulting in double clearing of all the surfaces.
This perf improvement won't be seen by Chrome since they've always being using wrapped backend texture SkSurfaces.
TBR=bsalomon@google.com
Change-Id: Ice3993ca125fce37804e58c353c265cf659dbe2f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/283456
Reviewed-by: Brian Salomon <bsalomon@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
There apparently were other tests that had similar issues and this fix
is much better and is more future proof.
Change-Id: I4835c7e5772b9e70249a69255aae8808be172eef
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/282681
Commit-Queue: Greg Daniel <egdaniel@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Auto-Submit: Greg Daniel <egdaniel@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Some may not want to always compile XPS on Windows
Change-Id: Icd4cc993667fdce740216b9c52a0a649dcf79645
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/278782
Reviewed-by: Ben Wagner <bungeman@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
ANGLE no longer supports NVPR, so this isn't necessary.
Change-Id: I68e720f17238c960e6f6ca07fe108f5fefed7705
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/278467
Commit-Queue: Greg Daniel <egdaniel@google.com>
Auto-Submit: Brian Osman <brianosman@google.com>
Reviewed-by: Greg Daniel <egdaniel@google.com>
This CL:
1) Fixes a GrTexture access in GrTextureEffect that was blocking pre-compilation
2) Adds program pre-compilation to the DDL Via - which would've caught the GrTextureEffect problem on the bots
3) Adds some #if'ed out code for collecting program pre-compilation stats
Bug: skia:9455
Change-Id: Ibcb07ae855b7a644e1f22c3427a928f116ab300d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/275336
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
This helper method can be used by the MeshDrawOps to (pre-)create their GrProgramInfos.
Bug: skia:9455
Change-Id: I41b7c2aefc0f633a1d32996c7f0cce3d11f8fcb1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/273815
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
With the addition of the DDL program pre-compilation we need to know how it is working.
This CL also fixes some threading bugs.
Bug: skia:9455
Change-Id: I20da58a7f1b19685687fae1d159d4e0db8a4964d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/273001
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
This also allows us to remove all the one off Fence code that we
implemented in all the backend TestContexts
Change-Id: I9ff7ba4690cf3f19a180f51fc510991a112bb62c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/272456
Reviewed-by: Chris Dalton <csmartdalton@google.com>
Commit-Queue: Greg Daniel <egdaniel@google.com>
This is through a private API but it, at least, connects things end-to-end.
Bug: skia:9455
Change-Id: Ib34d49c5c4e4cfa5fa599afc5c967fcadc3de10e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/268627
Commit-Queue: Robert Phillips <robertphillips@google.com>
Reviewed-by: Greg Daniel <egdaniel@google.com>
Having this as a sink rather than a Via allows us to do more aggressive things with threads and shared contexts.
Change-Id: I3ca1076686fa4f53387c12a9506e01910c1bc3e4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/272016
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
unused declaration
passing std::move() to const&
These are triggering warnings in google3
Change-Id: I12cebd0a8fd218e7755718fed7acec7908d386a1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/271138
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Reed <reed@google.com>
Split apart creation of the callback contexts and the backend textures
This allows the texture upload to be done separately on the
gpu-thread
Add a backendFormat member to the promise image callback context
This allows the promise images to be created in CreatePromiseImages
before the backend textures have been created (i.e., the backend
textures can now be created on the gpu-thread so we have no
guarantee they will be available when the SKP is being reinflated
w/ promise images)
Change-Id: I1e21385e450a5ef27dd6950d9d6aee737aa7515d
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/270939
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
TileData now gets a pointer to the final surface
This allows the tile to, once rendered, compose itself into
the final surface
DDLTileHelper now stores the TileData in a dumb array
SkTArray is overkill and, since TileData*s are being doled
out to threads, we never want reallocation
Added DDLTileHelper::kickOffThreadedWork
The old code only performed DDL creation in parallel. This
entry point also replays the DDLs and composes them into the
final surface in parallel
Change-Id: I66e02ef7f8291b4d402e22bee0ad3546e930609e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/270796
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
I'm adding a new DDL Sink and would like to reuse this functionality.
TBR=egdaniel@google.com
Change-Id: I17ae929557400be4edd8df78ab6580872024bc15
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/270799
Reviewed-by: Robert Phillips <robertphillips@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
This reverts commit 7b2fcfbc5a.
Reason for revert: Need to loosen up error handling
Original change's description:
> Carve some helper functions off of GPUSink
>
> I'm adding a new DDL Sink and would like to reuse this functionality.
>
> Change-Id: I9f4535ca5e0f36925bf896cb0076eab73fe60fd1
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/270639
> Commit-Queue: Robert Phillips <robertphillips@google.com>
> Reviewed-by: Greg Daniel <egdaniel@google.com>
TBR=egdaniel@google.com,robertphillips@google.com
Change-Id: I97968834b5719c2061d53caff0dcf08a80917beb
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/270798
Reviewed-by: Robert Phillips <robertphillips@google.com>
Commit-Queue: Robert Phillips <robertphillips@google.com>
I'm adding a new DDL Sink and would like to reuse this functionality.
Change-Id: I9f4535ca5e0f36925bf896cb0076eab73fe60fd1
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/270639
Commit-Queue: Robert Phillips <robertphillips@google.com>
Reviewed-by: Greg Daniel <egdaniel@google.com>
Initially, Error was written with the intent that an empty string meant
Ok and anything else meant Fatal. This made things simple with implicit
constructors from strings. With the introduction of Nonfatal the state
was now tied up with an additional boolean. Now the empty string meant
Ok and causes the new boolean to be ignored, or at least that is the way
it was used since Error didn't actually enforce that itself. This leads
to GMs which return kSkip but don't set the message to not be skipped.
This could be fixed in several ways.
The first would be for the GMSrc to notice that a GM had returned kSkip
with an empty message and create the Error::Nonfatal with a non-empty
message. This has the downside of being some extra unexpected complexity
and doesn't prevent similar issues from arising in the future.
The second would be to change how DM interprets the Error, and if the
non-fatal bit is set treat that as a sign to skip, even if the message
is empty. This fixes the stated issue, but doesn't fix the issue where a
GM can return kFail but also leave the message empty. This could again
be fixed by either modifying GMSrc::draw or GM::drawContent, but this
also seems a bit brittle in not preventing this from happening again in
the future.
So this replaces Error with Result, which makes the status orthogonal to
the message. It does lose the automatic conversion from string, but by
being able to wrap the many uses of SkStringPrintf the explicit nature
doesn't add much additional noise to the more complex failure reports.
Change-Id: Ibf48b67faa09a91a4a9d792d204bd9810b441c6c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/270362
Commit-Queue: Ben Wagner <bungeman@google.com>
Reviewed-by: Mike Klein <mtklein@google.com>
