Already exposed (on TextBlob), so this makes it available for clients
that may use drawGlyphs directly. (including canvaskit/flutter)
Change-Id: I8b4bd51e13827dc3970d5a6d06f0e0d3031af13c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/408638
Reviewed-by: Julia Lavrova <jlavrova@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Reed <reed@google.com>
Change-Id: Ie1eac40fe678529410f3ae4ab0cc7460dedfa4c2
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/408296
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Herb Derby <herb@google.com>
This is a reland of 0a0f4f5c35
This change makes SkCanvas::quickReject always reject empty draw bounds,
whereas previously scale+translate CTMs allowed bounds with w or h == 0
but otherwise contained in the clip to be drawn. This uncovered some
bugs in Skia where bounds shouldn't be empty, and in Flutter where
bounds were legit empty but not expected by the test.
No code changes needed. The issues that required its revert have been
fixed with:
1. https://github.com/flutter/engine/pull/26053 so that platforms that
use an empty typeface, leading to empty draws are just skipped.
2. https://skia-review.googlesource.com/c/skia/+/406140 so that path
effects update bounds so that Android's 1D dash path effect applied to
a horizontal line is properly not rejected.
Based on the period of time where the original CL was landed, some perf
data was collected.
- There were no significant changes in most SKPs or SVGs, except for a
Flutter page flip skp, which saw a 10% net improvement (perhaps the
flip is drawn with perspective?)
- A 10-20% regression in the motionmark paths skp, but dominated by the MSVC
compiler, so I'm not too concerned about that.
- Perspective microbenchmarks for drawing rectangles are 1.5-2x faster.
- quickReject microbenchmarks are about 2x slower.
The last two microbenchmark results aren't surprising since perspective
was the largest improvement in perf for SkM44::MapRect vs.
SkMatrix::mapRect, and the scale+translate specializations in Skmatrix
were maybe 50% faster than SkM44's. That would account for some of the
slow downs, and the rest could be explained by moving away from the
SIMD rect intersection and nan test.
Since these microreductions don't seem to bleed into more complex
benchmarks, I'm inclined to keep the code simple and not bring back the
custom intrinsics.
Original change's description:
> Simplify quickReject implementation in SkCanvas
>
> - SkCanvas no longer keeps fIsScaleTranslate bool that has to stay in
> sync with the type of the matrix.
> - No more fast or slow path for quickReject, the Sk4f code has been
> completely removed.
> - Uses SkM44::mapRect instead of SkMatrix::mapRect. This is slightly
> slower for S+T, but much faster for other transforms. I'm hopeful we
> won't notice the regression in the grand scheme for S+T, since the
> code is a lot simpler now.
> - The final isFinite() and intersects() check for quickReject uses
> SkRect's functions instead of hand-written SSE/NEON. If we think this
> is optimization is necessary, I'm hoping we can rewrite it in terms
> of skvx instead of specific instructions.
> - Consolidated how the quick-reject bounds outsetting into
> computeDeviceClipBounds, and added an option to skip outsetting for
> the one call site that doesn't want it.
>
> Bug: skia:10987
> Change-Id: I3cf2a73636cdeed06d12cab4548cfb94d1eb074a
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/405198
> Commit-Queue: Mike Reed <reed@google.com>
> Auto-Submit: Michael Ludwig <michaelludwig@google.com>
> Reviewed-by: Mike Reed <reed@google.com>
Bug: skia:10987
Change-Id: Id0d4b4ecebf0b83ae30f7e1a263961ab25de28dd
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/407358
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Change-Id: I674f038600afd6d49316c1ece515941ee5579068
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/406939
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
Makes computeFastBounds not part of the public API, it's only accessible
to subclasses of SkPathEffect, GrStyle, and SkPaint. Subclasses can
invoke it other path effects using SkPathEffectPriv::ComputeFastBounds.
Changes the internal function to
bool computeFastBounds(SkRect* bounds) const;
Subclasses of SkPathEffect must implement this, and can choose to return
false when fast bounds aren't computable.
Provides implementations of computeFastBounds() for path effects
bundled with Skia.
Bug: skia:11974
Change-Id: I545ccf99b4e669d3af9df13acfac28573306fab8
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/406140
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Mike Reed <reed@google.com>
Note: Shouldn't add or change the Skia API, only accepts an optional z value to preScale()
Change-Id: Ic73c723ebc2b75acca1fce5395953434ef1582e6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/404376
Reviewed-by: Mike Reed <reed@google.com>
Reviewed-by: Florin Malita <fmalita@chromium.org>
Commit-Queue: Jorge Betancourt <jmbetancourt@google.com>
This reverts commit 0a0f4f5c35.
Reason for revert: possible cause of flutter roll failure
Original change's description:
> Simplify quickReject implementation in SkCanvas
>
> - SkCanvas no longer keeps fIsScaleTranslate bool that has to stay in
> sync with the type of the matrix.
> - No more fast or slow path for quickReject, the Sk4f code has been
> completely removed.
> - Uses SkM44::mapRect instead of SkMatrix::mapRect. This is slightly
> slower for S+T, but much faster for other transforms. I'm hopeful we
> won't notice the regression in the grand scheme for S+T, since the
> code is a lot simpler now.
> - The final isFinite() and intersects() check for quickReject uses
> SkRect's functions instead of hand-written SSE/NEON. If we think this
> is optimization is necessary, I'm hoping we can rewrite it in terms
> of skvx instead of specific instructions.
> - Consolidated how the quick-reject bounds outsetting into
> computeDeviceClipBounds, and added an option to skip outsetting for
> the one call site that doesn't want it.
>
> Bug: skia:10987
> Change-Id: I3cf2a73636cdeed06d12cab4548cfb94d1eb074a
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/405198
> Commit-Queue: Mike Reed <reed@google.com>
> Auto-Submit: Michael Ludwig <michaelludwig@google.com>
> Reviewed-by: Mike Reed <reed@google.com>
TBR=mtklein@google.com,reed@google.com,michaelludwig@google.com
Change-Id: I1a373740ee167827b9a6a2eee9afb7f814641fb0
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: skia:10987
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/405616
Reviewed-by: Greg Daniel <egdaniel@google.com>
Commit-Queue: Greg Daniel <egdaniel@google.com>
- SkCanvas no longer keeps fIsScaleTranslate bool that has to stay in
sync with the type of the matrix.
- No more fast or slow path for quickReject, the Sk4f code has been
completely removed.
- Uses SkM44::mapRect instead of SkMatrix::mapRect. This is slightly
slower for S+T, but much faster for other transforms. I'm hopeful we
won't notice the regression in the grand scheme for S+T, since the
code is a lot simpler now.
- The final isFinite() and intersects() check for quickReject uses
SkRect's functions instead of hand-written SSE/NEON. If we think this
is optimization is necessary, I'm hoping we can rewrite it in terms
of skvx instead of specific instructions.
- Consolidated how the quick-reject bounds outsetting into
computeDeviceClipBounds, and added an option to skip outsetting for
the one call site that doesn't want it.
Bug: skia:10987
Change-Id: I3cf2a73636cdeed06d12cab4548cfb94d1eb074a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/405198
Commit-Queue: Mike Reed <reed@google.com>
Auto-Submit: Michael Ludwig <michaelludwig@google.com>
Reviewed-by: Mike Reed <reed@google.com>
Chromium has been updated to use makeAnalysisCanvas directly and there are
no more references to SkTextBlobDiffCanvas as a type in its code base.
Since the GlyphTrackingDevice extends SkNoPixelsDevice, any SkCanvas that
uses it is effectively a "no-draw" canvas. However, by returning a base
SkCanvas the text tracking now automatically happens in the context of
the base's AutoLayerForImageFilter handling it applies on every draw. This
means that drawing a text blob with an image filter that modifies the
transform state will now be analyzed in that context automatically
(simplifying code in chrome after this lands).
Another behavioral change is that all non-text draws will still go through
the base SkCanvas' virtuals and invoke the device function. Since it's an
SkNoPixelsDevice, it'll still be a no-op, it just happens a little later.
This won't really impact performance because oop-r already inspects their
operations and only plays back text and transform related ones to the
analysis canvas, so we shouldn't really see non-text draws being invoked
anyways.
Bug: chromium:1187246
Change-Id: I83f86571300751f385b3065dfe889f218fa1edc6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/405196
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
The SkM44::RectToRect function matches the semantics of
SkMatrix::RectToRect(kFill_ScaleToFit). No other ScaleToFit variants are
ported over to SkM44.
skottie uses some instances of kCenter_ScaleToFit so that functionality
may need to be added in the future (in SkM44 or in skottie). There are
no current usages of the kStart and kEnd_ScaleToFit semantics.
The SkM44::mapRect() function is implemented to correspond to the
SkMatrix::mapRect() that returns the mapped rect (instead of modifying a
pointer) and always has ApplyPerspectiveClip::kYes. This was chosen to
keep its behavior simple and because perspective clipping is almost
always the right thing to do. In the new implementation there is no
longer a performance cliff to worry about (see below). For the timebeing
mapRect is hidden behind SkMatrixPriv::MapRect().
Performance:
I added benchmarks for mapRect() on SkM44 and SkMatrix that use the same
matrices to get a fair comparison on their different specializations.
SkMatrix has a very efficient mapRect when it's scale+translate or
simpler, then another impl. for affine matrices, and then falls back to
SkPath clipping when there's perspective. On the other hand, SkM44 only
has 2 modes: affine and perspective.
On my desktop, with a Ryzen 9 3900X, here are the times for 100,000 calls
to mapRect for different types of matrices:
SkMatrix SkM44
scale+translate 0.35 ms 0.42 ms
rotate 1.70 ms 0.42 ms
perspective 63.90 ms 0.66 ms
clipped-perspective 138.0 ms 0.96 ms
To summarize, the SkM44::mapRect is almost as fast as the s+t specialization
in SkMatrix, but for all non-perspective matrices. For perspective matrices
it's only 2x slower than that specialization when no vertices are clipped,
and still almost 2x faster than the affine specialization when vertices are
clipped (and 100x faster than falling back to SkPath).
Given that, there's the open question of whether or not keeping an affine
specialization is worth it for SkM44's code size.
Bug: skia:11720
Change-Id: I6771956729ed64f3b287a9de503513375c9f42a0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/402957
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Reed <reed@google.com>
Auto-Submit: Michael Ludwig <michaelludwig@google.com>
This removes the templated versions of these operators. This should help
the case where clients would get compiler errors in their code with
errors claiming failed matches to these ganesh operators. The errors were
correct, but would be confusing for clients to see. Now with this change
the various operators are defined for specific types so a client shouldn't
get errors for their own enums anymore.
Bug: chromium:1204688
Change-Id: Ie3450834da7734a161af303ca6c8f458dd173513
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/403596
Commit-Queue: Greg Daniel <egdaniel@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
This relands the idea that onMatchFaceStyle is no longer used, but
leaves the baseclass virtual to stage removing it from client
subclasses.
This reverts commit 3c04a65508.
Change-Id: I18570065249c86f7f155c28288dce3ea9d59f619
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/401759
Reviewed-by: Ben Wagner <bungeman@google.com>
Commit-Queue: Mike Reed <reed@google.com>
Add a drawGlyphs to SkCanvas that takes SkRSXform instead of
positions. Update buffer sizing calculations to take
SkRSXform buffers into account.
Change-Id: I14529088199dcd0b1ae78b4605e1ba77fec2000e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/399096
Reviewed-by: Ben Wagner <bungeman@google.com>
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Herb Derby <herb@google.com>
This code appears to be clearing out the SkSurfaceProps in order to
disable LCD text. We ought to be able to only disable the LCD text,
while preserving the SkSurfaceProps flags.
Bug: skia:11396
Change-Id: I7c1f49f59639404535a445f0318ab97b07c20c84
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/397636
Reviewed-by: Mike Reed <reed@google.com>
Reviewed-by: Jim Van Verth <jvanverth@google.com>
Reviewed-by: Michael Ludwig <michaelludwig@google.com>
Commit-Queue: Chris Dalton <csmartdalton@google.com>
This was an experimental feature. It worked (but only the GPU backend).
It was never adopted or used by anyone, to my knowledge. It's a large
amount of code, and a strange corner of SkSL for users to stumble into.
Bug: skia:10680
Change-Id: I0dda0364bce7dbffa58c32de4c7801ec2a6bc42e
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/398222
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
This also connects a lot of the wires needed to use dynamic MSAA in
vulkan. By using the framebuffer object in the render pass we can figure
out the specific framebuffer we want in one place, GrVkGpu::onGetOpsRenderPass,
and then the render pass itself doesn't need any explicit knowledge of
dmsaa stuff.
Bug: skia:11809
Change-Id: I3e4e71fa6f9536fdaf915d5369a2f8a24bf48c9b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/397156
Reviewed-by: Brian Salomon <bsalomon@google.com>
Reviewed-by: Jim Van Verth <jvanverth@google.com>
Commit-Queue: Greg Daniel <egdaniel@google.com>
Also, disable quick-reject for now until bounds are fixed.
Change-Id: I08b9fa037d279a99fc393364a71cb171e3d16d4a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/397458
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Reed <reed@google.com>
With this change, there is no need to produce a SkTextBlob when
using the drawGlyph, drawSimpleText, or drawString apis. These
calls just produce a light weight wrapper sending the wrapper
to onDrawGlyphRunList for rendering.
For recording, recording canvas converts the SkGlyphRunList into
a blob, and calls SkRecorder::onDrawTextBlob.
Remove unused call: drawPosTextCommon.
Change-Id: I173ba2793f74b521b33a6fb3dbd8d98945216a3b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/388719
Reviewed-by: Herb Derby <herb@google.com>
Reviewed-by: Ben Wagner <bungeman@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Sk3Perspective -> SkM44::Perspective
Sk3LookAt -> SkM44::LookAt
Also adds some SK_API tags to the SkV[2,3,4] structs. Also fixes
linkage issues around Sk3Perspective/LookAt by moving them into the
exported SkM44 (if we don't like them as SkM44 factories, will just need
to add SK_API tags to old Sk3Perspective/Lookat directly).
Change-Id: I3f125211b76899f216e63cc8d587776004516e36
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/388476
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Michael Ludwig <michaelludwig@google.com>
Now that these methods are used in Chromium to produce PDFs with
embedded text, properly document how to use them.
Change-Id: I68fb477b65ec41af9fcc46429275bda03680bff0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/387976
Commit-Queue: Ben Wagner <bungeman@google.com>
Reviewed-by: Mike Reed <reed@google.com>
Reviewed-by: Dominik Röttsches <drott@chromium.org>
The global fPointSizeLimit which backs the above calls is protected
by a mutex because it can be called from multiple threads. This
created contention with multi-threaded strike cache use.
A search of the Google, Android, and Chrome source show no use of
SetFontCachePointSizeLimit, and the only uses of
GetFontCachePointSizeLimit are internal. In effect, the mutex
protected a constant.
Remove all uses, and replace with constants.
Bug: skia:11777
Change-Id: I9a2c3f3ee849cd07d04efa7113cb3ea9c600927f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/387676
Reviewed-by: Ben Wagner <bungeman@google.com>
Commit-Queue: Herb Derby <herb@google.com>
Skia does not call set or get filter-quality any more
(except for legacy picture deserialization)
Change-Id: I504caf407ca68392481b771040e5d3280bf7da7f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/387439
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Reed <reed@google.com>
A follow-up CL can remove the filter-quality from onProgram.
Change-Id: I770e3b1fd0907bf3824ed402502fa67325a433d5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/381799
Reviewed-by: Florin Malita <fmalita@chromium.org>
Commit-Queue: Mike Reed <reed@google.com>
Expose text and cluster information in SkTextBlob::Iter when
SK_UNTIL_CRBUG_1187654_IS_FIXED is defined. The names are postfixed
with _forTest to indicate that these values are only being exposed for
testing purposes. This will allow blink tests to better verify the
output SkTextBlobs from the ShapeResultBloberizer. The long term goal
is to only store these results when necessary in a blink side type.
Bug: chromium:1187654,chromium:738643
Change-Id: I8db20a8423e5b0652429ddc16cf8fd14940217cb
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384336
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Ben Wagner <bungeman@google.com>
PictureShader = picture + tiling + depth/colorspace + filtering [+ scale]
Today we cache the imageshader that is used to rendering. However, the
key for that cache is the pictureshader's ID itself... which means if
we have several, all using the same picture (but maybe diff tiling) we
would create dup cache entries.
Idea:
1. only cache the image (rastered picture), not an imageShader
2. key the cache on the picture's ID, not the shader's
Several implications of this:
1. Should get more cache reuse, since we don't care about the
shader's ID (which is just wrapping a picture+tiling, etc.)
2. We also eliminate the indirection of creating a PictureImage. Instead
we're creating real (pixel) images, and caching those. This removes one
extra layer of "cache".
Idea: when we cache something for pict
Change-Id: I51cf4e9bff3c91ce1872876597d3d565039d8c7a
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/377844
Commit-Queue: Mike Reed <reed@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
SkFilterQuality is recorded in the chrome UMA metrics, the enum should
not be reordered or changed; otherwise, the result may produce false
alarm. Adding the comment to reflect it t avoid future misunderstanding.
Bug:1176091
Change-Id: I6cdf0d8c4f59a5548e456a2c641b2b6158abde48
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/374417
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Reed <reed@google.com>
This will allow users to create PDFs with the actual text embedded. This
will allow for correct search and copy operations on the generated PDF.
Since these are now public, SkTextBlobBuilderPriv is no longer needed
and is removed. For consistency, the allocRunRSXform overload is renamed
to allocRunTextRSXform.
Change-Id: I44be82d9038a433e1221d5cbfd8ed113ecb6d4fa
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/375017
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Ben Wagner <bungeman@google.com>
The 'lang' was never stored or used, so this makes no current practical
difference. The original intent was to be able to specify a language so
that it could be emitted as a 'Lang' override in the 'ActualText' when
generating a PDF. However, due to the way 'ActualText' is generally used
this would be impractical. If there is ever a desire to mark up sections
of the PDF with a specific language it would be better handled in a
different way.
Change-Id: Id63596190235fc45ce17249b9b578b6f9b838b2b
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/375060
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Ben Wagner <bungeman@google.com>
This paves the way for promise image sharing among direct & recording
contexts, and untethers promise images from DDL recorder.
Followup CLs will migrate us to actually use this entry point,
and then migrate Chrome to do same.
Bug: skia:10286
Change-Id: I0ad46e8e4b91d8bc03cb039b304d2ea6d8a65c35
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/373716
Commit-Queue: Brian Salomon <bsalomon@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>
Reviewed-by: Robert Phillips <robertphillips@google.com>
For clients like chrome that want to rely on the numeric values
Change-Id: Ib8ecf2e404b159ff26e44d41bd60f98609ff3ad0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/372976
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Reed <reed@google.com>
This reverts commit af68258c2e.
Reason for revert: SkFilterQuality is not a scoped enum, so this defines `kMaxValue` without any scoping to Skia by naming convention or namespace, breaking compatibility with other systems that want to use that. (E.g. Android)
Original change's description:
> Add number ordering to SkFilterQuality
>
> As explained in the bug, I would like to add a histogram to study how
> users use the image smoothing quality in Chrome. In order to do that,
> we need to ensure that the enums are not reordered or changed, so I
> added kNone_SkFilterQuality = 0, etc. I also added kMaxValue for
> histogram to work.
>
> Bug: 1176091
> Change-Id: I3af3213f699016a525caad5b0b51e0cc7aab2c52
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/368157
> Commit-Queue: Mike Reed <reed@google.com>
> Auto-Submit: Yi Xu <yiyix@chromium.org>
> Reviewed-by: Mike Reed <reed@google.com>
TBR=reed@google.com,yiyix@chromium.org
Change-Id: Iec24ba3864af69e34f2721a72de1c0a271d328f6
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: 1176091
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/371936
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>
As explained in the bug, I would like to add a histogram to study how
users use the image smoothing quality in Chrome. In order to do that,
we need to ensure that the enums are not reordered or changed, so I
added kNone_SkFilterQuality = 0, etc. I also added kMaxValue for
histogram to work.
Bug: 1176091
Change-Id: I3af3213f699016a525caad5b0b51e0cc7aab2c52
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/368157
Commit-Queue: Mike Reed <reed@google.com>
Auto-Submit: Yi Xu <yiyix@chromium.org>
Reviewed-by: Mike Reed <reed@google.com>
Change-Id: Id28ed827c4a896805c6d4eead339146fdd49e35f
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/359560
Reviewed-by: Mike Klein <mtklein@google.com>
Commit-Queue: Brian Salomon <bsalomon@google.com>
