Currently, we mix objects with pointers and objects without pointers
("noscan" objects) together in memory. As a result, for every object
we grey, we have to check that object's heap bits to find out if it's
noscan, which adds to the per-object cost of GC. This also hurts the
TLB footprint of the garbage collector because it decreases the
density of scannable objects at the page level.
This commit improves the situation by using separate spans for noscan
objects. This will allow a much simpler noscan check (in a follow up
CL), eliminate the need to clear the bitmap of noscan objects (in a
follow up CL), and improves TLB footprint by increasing the density of
scannable objects.
This is also a step toward eliminating dead bits, since the current
noscan check depends on checking the dead bit of the first word.
This has no effect on the heap size of the garbage benchmark.
We'll measure the performance change of this after the follow-up
optimizations.
This is a cherry-pick from dev.garbage commit d491e550c3. The only
non-trivial merge conflict was in updatememstats in mstats.go, where
we now have to separate the per-spanclass stats from the per-sizeclass
stats.
Change-Id: I13bdc4869538ece5649a8d2a41c6605371618e40
Reviewed-on: https://go-review.googlesource.com/41251
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
In particular, this says that Frames.Function uniquely identifies a
function within a program. We depend on this in various places that
use runtime.Frames in std, but it wasn't actually written down.
Change-Id: Ie7ede348c17673e11ae513a094862b60c506abc5
Reviewed-on: https://go-review.googlesource.com/41610
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This reverts commit 3d86d45dd6.
Reason for revert: Fixing a different way. See #20008.
Change-Id: I6f4382ffdfff83a4504b013ba37100319b2c1aad
Reviewed-on: https://go-review.googlesource.com/42138
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
The code in #20162 contains an embedded interface.
It didn't get dowidth'd by the frontend,
and during DWARF generation, ngotype asked
for a string description of it,
which triggered a request for the number of fields
in the interface, which triggered a dowidth,
which is disallowed in the backend.
The other changes in this CL are to support the test.
Fixes#20162
Change-Id: I4d0be5bd949c361d4cdc89a8ed28b10977e40cf9
Reviewed-on: https://go-review.googlesource.com/42131
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Now that it only takes small values.
Change-Id: I08086d392529d8775b470d65afc2475f8d0e7f4a
Reviewed-on: https://go-review.googlesource.com/42030
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
It used to be simple, and then it got complicated for speed (to reduce
allocations, mostly), but that involved a mutex and hurt multi-core
performance, contending on the mutex.
A change was sent to try to improve that mutex contention in
https://go-review.googlesource.com/c/42110/2/src/net/http/server.go
but that introduced its own allocations (the string->interface{}
boxing for the sync.Map key), which runs counter to the whole point of
that statusLine function: to remove allocations.
Instead, make the code simple again and not have a mutex. It's a bit
slower for the single-core case, but nobody with a single-user HTTP
server cares about 50 nanoseconds:
name old time/op new time/op delta
ResponseStatusLine 37.5ns ± 2% 87.1ns ± 2% +132.42% (p=0.029 n=4+4)
ResponseStatusLine-2 63.1ns ± 1% 43.1ns ±12% -31.67% (p=0.029 n=4+4)
ResponseStatusLine-4 53.8ns ± 8% 40.2ns ± 2% -25.29% (p=0.029 n=4+4)
name old alloc/op new alloc/op delta
ResponseStatusLine 0.00B ±NaN% 0.00B ±NaN% ~ (all samples are equal)
ResponseStatusLine-2 0.00B ±NaN% 0.00B ±NaN% ~ (all samples are equal)
ResponseStatusLine-4 0.00B ±NaN% 0.00B ±NaN% ~ (all samples are equal)
name old allocs/op new allocs/op delta
ResponseStatusLine 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal)
ResponseStatusLine-2 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal)
ResponseStatusLine-4 0.00 ±NaN% 0.00 ±NaN% ~ (all samples are equal)
(Note the code could be even simpler with fmt.Fprintf, but that is
relatively slow and involves a bunch of allocations getting arguments
into interface{} for the call)
Change-Id: I1fa119132dbbf97a8e7204ce3e0707d433060da2
Reviewed-on: https://go-review.googlesource.com/42133
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Bryan Mills <bcmills@google.com>
Using 'go build -tags "foo,bar"' might seem to work when you wanted
-tags "foo bar", since they make up a single tag that doesn't exist and
the build is unaffected.
Instead, error on any tag that contains a comma.
Fixes#18800.
Change-Id: I6641e03e2ae121c8878d6301c4311aef97026b73
Reviewed-on: https://go-review.googlesource.com/41951
Run-TryBot: Daniel Martí <mvdan@mvdan.cc>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
The comment for Cmd.Stdout and Cmd.Stderr says that it's safe to
set both to the same writer, but it doesn't say that this only
works when both writers are comparable.
This change updates the comment to explain that using a
non-comparable writer may still lead to a race.
Fixes#19804
Change-Id: I63b420034666209a2b6fab48b9047c9d07b825e2
Reviewed-on: https://go-review.googlesource.com/42052
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Profile labels added by the user using pprof.Do, if present will
be in a *labelMap stored in the unsafe.Pointer 'tag' field of
the profile map entry. This change extracts the labels from the tag
field and writes them to the profile proto.
Change-Id: Ic40fdc58b66e993ca91d5d5effe0e04ffbb5bc46
Reviewed-on: https://go-review.googlesource.com/39613
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
If g1 sets its labels and then they are copied into a profile buffer
and then g2 reads the profile buffer and inspects the labels,
the race detector must understand that g1's recording of the labels
happens before g2's use of the labels. Make that so.
Fixes race test failure in CL 39613.
Change-Id: Id7cda1c2aac6f8eef49213b5ca414f7154b4acfa
Reviewed-on: https://go-review.googlesource.com/42111
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Michael Matloob <matloob@golang.org>
More specifically, allow Unicode letters in the directories of GitHub
repositories, which can occur and don't have a valid reason to be
disallowed by go get.
Do so by using a predefined character class, the Unicode character
property class \p{L} that describes the Unicode characters that are
letters:
http://www.regular-expressions.info/unicode.html#category
Since it's not possible to create GitHub usernames or repositories
containing Unicode letters at this time, those parts of the import path
are still restricted to ASCII letters only.
Fix name of tested func in t.Errorf messages.
Fixes#18660.
Change-Id: Ia0ef4742bfd8317d989ef1eb1d7065e382852fe2
Reviewed-on: https://go-review.googlesource.com/41822
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Daniel Martí <mvdan@mvdan.cc>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This is the same technique used in CL 24466. By adding a little bit of
size to the binary, we can remove a function call and gain a lot of
performance.
A raw array ([128]bool) would be faster, but is also be 128 bytes
instead of 16.
Running tip on a Mac:
name old time/op new time/op delta
QuoteMetaAll-4 192ns ±12% 120ns ±11% -37.27% (p=0.000 n=10+10)
QuoteMetaNone-4 186ns ± 6% 64ns ± 6% -65.52% (p=0.000 n=10+10)
name old speed new speed delta
QuoteMetaAll-4 73.2MB/s ±11% 116.6MB/s ±10% +59.21% (p=0.000 n=10+10)
QuoteMetaNone-4 139MB/s ± 6% 405MB/s ± 6% +190.74% (p=0.000 n=10+10)
Change-Id: I68ce9fe2ef1c28e2274157789b35b0dd6ae3efb5
Reviewed-on: https://go-review.googlesource.com/41495
Run-TryBot: Kevin Burke <kev@inburke.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Use it to ensure that dowidth is not called
from the backend on a type whose size
has not yet been calculated.
This is an alternative to CL 42016.
Change-Id: I8c7b4410ee4c2a68573102f6b9b635f4fdcf392e
Reviewed-on: https://go-review.googlesource.com/42018
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Now it's not very different from syscall_dragonfly.go in
golang.org/x/sys/unix repository.
Change-Id: I8dfd22e1ebce9dc2cc71ab9ab7f0c92d93b2b762
Reviewed-on: https://go-review.googlesource.com/41835
Run-TryBot: Mikio Hara <mikioh.mikioh@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
This CL mainly moves some work to the switch on w.Op,
to make a follow-up change simpler and clearer.
Updates #19838
Change-Id: I86f3181c380dd60960afcc24224f655276b8956c
Reviewed-on: https://go-review.googlesource.com/42010
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Node.Used was written to from the backend
concurrently with reads of Node.Class
for the same ONAME Nodes.
I do not know why it was not failing consistently
under the race detector, but it is a race.
This is likely also a problem with Node.HasVal and Node.HasOpt.
They will be handled in a separate CL.
Fix Used by moving it to gc.Name and making it a separate bool.
There was one non-Name use of Used, marking OLABELs as used.
That is no longer needed, now that goto and label checking
happens early in the front end.
Leave the getters and setters in place,
to ease changing the representation in the future
(or changing to an interface!).
Updates #20144
Change-Id: I9bec7c6d33dcb129a4cfa9d338462ea33087f9f7
Reviewed-on: https://go-review.googlesource.com/42015
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Type.Size and Type.Alignment are for the front end:
They calculate size and alignment if needed.
Type.MustSize and Type.MustAlignment are for the back end:
They call Fatal if size and alignment are not already calculated.
Most uses are of MustSize and MustAlignment,
but that's because the back end is newer,
and this API was added to support it.
This CL was mostly generated with sed and selective reversion.
The only mildly interesting bit is the change of the ssa.Type interface
and the supporting ssa dummy types.
Follow-up to review feedback on CL 41970.
Passes toolstash-check.
Change-Id: I0d9b9505e57453dae8fb6a236a07a7a02abd459e
Reviewed-on: https://go-review.googlesource.com/42016
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
dowidth is fundamentally unsafe to call from the back end;
it will cause data races.
Replace all calls to dowidth in the backend with
assertions that the width has been calculated.
Then fix all the cases in which that was not so,
including the cases from #20145.
Fixes#20145.
Change-Id: Idba3d19d75638851a30ec2ebcdb703c19da3e92b
Reviewed-on: https://go-review.googlesource.com/41970
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Many (most!) of the values of objapi.SymKind are used only in the linker, so
this creates a separate cmd/link/internal/ld.SymKind type, removes most values
from SymKind and maps one to the other when reading object files in the linker.
Two of the remaining objapi.SymKind values are only checked for, never set and
so will never be actually found but I wanted to keep this to the most
mechanical change possible.
Change-Id: I4bbc5aed6713cab3e8de732e6e288eb77be0474c
Reviewed-on: https://go-review.googlesource.com/40985
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Currently, the following two codes generate the identical dwarf info
for type Foo.
prog 1)
type Foo struct {
Bar
}
prog 2)
type Foo struct {
Bar Bar
}
This change adds a go-specific attribute DW_AT_go_embedded_field
to annotate each member entry. Its absence or false value indicates
the corresponding member is not an embedded field.
Update #20037
Change-Id: Ibcbd2714f3e4d97c7b523d7398f29ab2301cc897
Reviewed-on: https://go-review.googlesource.com/41873
Reviewed-by: David Chase <drchase@google.com>
There's been one failure on the race builder so far,
before we started sorting functions by length.
The race detector can only detect actual races,
and ordering functions by length might reduce the odds
of catching some kinds of races. Give it more to chew on.
Updates #20144
Change-Id: I0206ac182cb98b70a729dea9703ecb0fef54d2d0
Reviewed-on: https://go-review.googlesource.com/41973
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Its sole use is in walk.go. 100% code movement.
gsubr.go increasingly contains backend-y things.
With a few more relocations, it could probably be
fruitfully renamed progs.go.
Change-Id: I61ec5c2bc1f8cfdda64c6d6f580952c154ff60e0
Reviewed-on: https://go-review.googlesource.com/41972
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
They were used only in esc.go. 100% code movement.
Also, remove the rather outdated comment at the top of gen.go.
It's not really clear what gen.go is for any more.
Change-Id: Iaedfe7015ef6f5c11c49f3e6721b15d779a00faa
Reviewed-on: https://go-review.googlesource.com/41971
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
When a constant doesn't fit in a single instruction, use two
paired instructions instead of the constant pool. For example
ADD $0xaa00bb, R0, R1
Used to rewrite to:
MOV ?(IP), R11
ADD R11, R0, R1
Instead, do:
ADD $0xaa0000, R0, R1
ADD $0xbb, R1, R1
Same number of instructions.
Good:
4 less bytes (no constant pool entry)
One less load.
Bad:
Critical path is one instruction longer.
It's probably worth it to avoid the loads, they are expensive.
Dave Cheney got us some performance numbers: https://perf.golang.org/search?q=upload:20170426.1
TL;DR mean 1.37% improvement.
Change-Id: Ib206836161fdc94a3962db6f9caa635c87d57cf1
Reviewed-on: https://go-review.googlesource.com/41612
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
The "lldb start" phase often times out on the iOS builder. Increase
the timeout and see if that helps.
Change-Id: I92fd67cbfa90659600e713198d6b2c5c78dde20f
Reviewed-on: https://go-review.googlesource.com/41863
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Elias Naur <elias.naur@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
The current code treats condition as special register and write
its raw data directly into instruction.
The fix converts the raw data into correct condition encoding.
Also fix the operand catogery of FCCMP.
Add tests to cover all cases.
Change-Id: Ib194041bd9017dd0edbc241564fe983082ac616b
Reviewed-on: https://go-review.googlesource.com/41511
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Previously, the package did not distinguish between baseline and
extended sequential images. Both are non-progressive images, but the Th
range differs between the two, as per Annex B of
https://www.w3.org/Graphics/JPEG/itu-t81.pdf
Extended sequential images are often emitted by the Guetzli encoder.
Fixes#19913
Change-Id: I3d0f9e16d5d374ee1c65e3a8fb87519de61cff94
Reviewed-on: https://go-review.googlesource.com/41831
Reviewed-by: David Symonds <dsymonds@golang.org>
When using a concurrent backend,
the overall compilation time is bounded
in part by the slowest function to compile.
The number of top-level statements in a function
is an easily calculated and fairly reliable
proxy for compilation time.
Here's a standard compilecmp output for -c=8 with this CL:
name old time/op new time/op delta
Template 127ms ± 4% 125ms ± 6% -1.33% (p=0.000 n=47+50)
Unicode 84.8ms ± 4% 84.5ms ± 4% ~ (p=0.217 n=49+49)
GoTypes 289ms ± 3% 287ms ± 3% -0.78% (p=0.002 n=48+50)
Compiler 1.36s ± 3% 1.34s ± 2% -1.29% (p=0.000 n=49+47)
SSA 2.95s ± 3% 2.77s ± 4% -6.23% (p=0.000 n=50+49)
Flate 70.7ms ± 3% 70.9ms ± 2% ~ (p=0.112 n=50+49)
GoParser 85.0ms ± 3% 83.0ms ± 4% -2.31% (p=0.000 n=48+49)
Reflect 229ms ± 3% 225ms ± 4% -1.83% (p=0.000 n=49+49)
Tar 70.2ms ± 3% 69.4ms ± 3% -1.17% (p=0.000 n=49+49)
XML 115ms ± 7% 114ms ± 6% ~ (p=0.158 n=49+47)
name old user-time/op new user-time/op delta
Template 352ms ± 5% 342ms ± 8% -2.74% (p=0.000 n=49+50)
Unicode 117ms ± 5% 118ms ± 4% +0.88% (p=0.005 n=46+48)
GoTypes 986ms ± 3% 980ms ± 4% ~ (p=0.110 n=46+48)
Compiler 4.39s ± 2% 4.43s ± 4% +0.97% (p=0.002 n=50+50)
SSA 12.0s ± 2% 13.3s ± 3% +11.33% (p=0.000 n=49+49)
Flate 222ms ± 5% 219ms ± 6% -1.56% (p=0.002 n=50+50)
GoParser 271ms ± 5% 268ms ± 4% -0.83% (p=0.036 n=49+48)
Reflect 560ms ± 4% 571ms ± 3% +1.90% (p=0.000 n=50+49)
Tar 183ms ± 3% 183ms ± 3% ~ (p=0.903 n=45+50)
XML 364ms ±13% 391ms ± 4% +7.16% (p=0.000 n=50+40)
A more interesting way of viewing the data is by
looking at the ratio of the time taken to compile
the slowest-to-compile function to the overall
time spent compiling functions.
If this ratio is small (near 0), then increased concurrency might help.
If this ratio is big (near 1), then we're bounded by that single function.
I instrumented the compiler to emit this ratio per-package,
ran 'go build -a -gcflags=-c=C -p=P std cmd' three times,
for varying values of C and P,
and collected the ratios encountered into an ASCII histogram.
Here's c=1 p=1, which is a non-concurrent backend, single process at a time:
90%|
80%|
70%|
60%|
50%|
40%|
30%|
20%|**
10%|***
0%|*********
----+----------
|0123456789
The x-axis is floor(10*ratio), so the first column indicates the percent of
ratios that fell in the 0% to 9.9999% range.
We can see in this histogram that more concurrency will help;
in most cases, the ratio is small.
Here's c=8 p=1, before this CL:
90%|
80%|
70%|
60%|
50%|
40%|
30%| *
20%| *
10%|* * *
0%|**********
----+----------
|0123456789
In 30-40% of cases, we're mostly bound by the compilation time
of a single function.
Here's c=8 p=1, after this CL:
90%|
80%|
70%|
60%|
50%| *
40%| *
30%| *
20%| *
10%| *
0%|**********
----+----------
|0123456789
The sorting pays off; we are bound by the
compilation time of a single function in over half of packages.
The single * in the histogram indicates 0-10%.
The actual values for this chart are:
0: 5%, 1: 1%, 2: 1%, 3: 4%, 4: 5%, 5: 7%, 6: 7%, 7: 7%, 8: 9%, 9: 55%
This indicates that efforts to increase or enable more concurrency,
e.g. by optimizing mutexes or increasing the value of c,
will probably not yield fruit.
That matches what compilecmp tells us.
Further optimization efforts should thus focus instead on one of:
(1) making more functions compile concurrently
(2) improving the compilation time of the slowest functions
(3) speeding up the remaining serial parts of the compiler
(4) automatically splitting up some large autogenerated functions
into small ones, as discussed in #19751
I hope to spend more time on (1) before the freeze.
Adding process parallelism doesn't change the story much.
For example, here's c=8 p=8, after this CL:
90%|
80%|
70%|
60%|
50%|
40%| *
30%| *
20%| *
10%| ***
0%|**********
----+----------
|0123456789
Since we don't need to worry much about p,
these histograms can help us select a good
general value of c to use as a default,
assuming we're not bounded by GOMAXPROCS.
Here are some charts after this CL, for c from 1 to 8:
c=1 p=1
90%|
80%|
70%|
60%|
50%|
40%|
30%|
20%|**
10%|***
0%|*********
----+----------
|0123456789
c=2 p=1
90%|
80%|
70%|
60%|
50%|
40%|
30%|
20%|
10%| **** *
0%|**********
----+----------
|0123456789
c=3 p=1
90%|
80%|
70%|
60%|
50%|
40%|
30%|
20%| *
10%| ** * *
0%|**********
----+----------
|0123456789
c=4 p=1
90%|
80%|
70%|
60%|
50%|
40%|
30%| *
20%| *
10%| * *
0%|**********
----+----------
|0123456789
c=5 p=1
90%|
80%|
70%|
60%|
50%|
40%|
30%| *
20%| *
10%| * *
0%|**********
----+----------
|0123456789
c=6 p=1
90%|
80%|
70%|
60%|
50%|
40%| *
30%| *
20%| *
10%| *
0%|**********
----+----------
|0123456789
c=7 p=1
90%|
80%|
70%|
60%|
50%| *
40%| *
30%| *
20%| *
10%| **
0%|**********
----+----------
|0123456789
c=8 p=1
90%|
80%|
70%|
60%|
50%| *
40%| *
30%| *
20%| *
10%| *
0%|**********
----+----------
|0123456789
Given the increased user-CPU costs as
c increases, it looks like c=4 is probably
the sweet spot, at least for now.
Pleasingly, this matches (and explains)
the results of the standard benchmarking
that I have done.
Updates #15756
Change-Id: I82b606c06efd34a5dbd1afdbcf66a605905b2aeb
Reviewed-on: https://go-review.googlesource.com/41192
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
