This commit replaces some of
for i := len(x) - 1; i >= 0; i-- {...}
style loops, which do not rely on reverse iteration order.
Change-Id: I5542834286562da058200c06e7a173b13760e54d
Reviewed-on: https://go-review.googlesource.com/21044
Reviewed-by: Keith Randall <khr@golang.org>
Escape analysis has a hard time with tree-like
structures (see #13493 and #14858).
This is unlikely to change.
As a result, when invoking a function that accepts
a **Node parameter, we usually allocate a *Node
on the heap. This happens a whole lot.
This CL changes functions from taking a **Node
to acting more like append: It both modifies
the input and returns a replacement for it.
Because of the cascading nature of escape analysis,
in order to get the benefits, I had to modify
almost all such functions. The remaining functions
are in racewalk and the backend. I would be happy
to update them as well in a separate CL.
This CL was created by manually updating the
function signatures and the directly impacted
bits of code. The callsites were then automatically
updated using a bespoke script:
https://gist.github.com/josharian/046b1be7aceae244de39
For ease of reviewing and future understanding,
this CL is also broken down into four CLs,
mailed separately, which show the manual
and the automated changes separately.
They are CLs 20990, 20991, 20992, and 20993.
Passes toolstash -cmp.
name old time/op new time/op delta
Template 335ms ± 5% 324ms ± 5% -3.35% (p=0.000 n=23+24)
Unicode 176ms ± 9% 165ms ± 6% -6.12% (p=0.000 n=23+24)
GoTypes 1.10s ± 4% 1.07s ± 2% -2.77% (p=0.000 n=24+24)
Compiler 5.31s ± 3% 5.15s ± 3% -2.95% (p=0.000 n=24+24)
MakeBash 41.6s ± 1% 41.7s ± 2% ~ (p=0.586 n=23+23)
name old alloc/op new alloc/op delta
Template 63.3MB ± 0% 62.4MB ± 0% -1.36% (p=0.000 n=25+23)
Unicode 42.4MB ± 0% 41.6MB ± 0% -1.99% (p=0.000 n=24+25)
GoTypes 220MB ± 0% 217MB ± 0% -1.11% (p=0.000 n=25+25)
Compiler 994MB ± 0% 973MB ± 0% -2.08% (p=0.000 n=24+25)
name old allocs/op new allocs/op delta
Template 681k ± 0% 574k ± 0% -15.71% (p=0.000 n=24+25)
Unicode 518k ± 0% 413k ± 0% -20.34% (p=0.000 n=25+24)
GoTypes 2.08M ± 0% 1.78M ± 0% -14.62% (p=0.000 n=25+25)
Compiler 9.26M ± 0% 7.64M ± 0% -17.48% (p=0.000 n=25+25)
name old text-bytes new text-bytes delta
HelloSize 578k ± 0% 578k ± 0% ~ (all samples are equal)
CmdGoSize 6.46M ± 0% 6.46M ± 0% ~ (all samples are equal)
name old data-bytes new data-bytes delta
HelloSize 128k ± 0% 128k ± 0% ~ (all samples are equal)
CmdGoSize 281k ± 0% 281k ± 0% ~ (all samples are equal)
name old exe-bytes new exe-bytes delta
HelloSize 921k ± 0% 921k ± 0% ~ (all samples are equal)
CmdGoSize 9.86M ± 0% 9.86M ± 0% ~ (all samples are equal)
Change-Id: I277d95bd56d51c166ef7f560647aeaa092f3f475
Reviewed-on: https://go-review.googlesource.com/20959
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
The Node type ODOT and its variants all represent a selector, with a
simple name to the right of the dot. Before this change this was
represented by using an ONAME Node in the Right field. This ONAME node
served no useful purpose. This CL changes these Node types to store the
symbol in the Sym field instead, thus not requiring allocating a Node
for each selector.
When compiling x/tools/go/types this CL eliminates nearly 5000 calls to
newname and reduces the total number of Nodes allocated by about 6.6%.
It seems to cut compilation time by 1 to 2 percent.
Getting this right was somewhat subtle, and I added two dubious changes
to produce the exact same output as before. One is to ishairy in
inl.go: the ONAME node increased the cost of ODOT and friends by 1, and
I retained that, although really ODOT is not more expensive than any
other node. The other is to varexpr in walk.go: because the ONAME in
the Right field of an ODOT has no class, varexpr would always return
false for an ODOT, although in fact for some ODOT's it seemingly ought
to return true; I added an && false for now. I will send separate CLs,
that will break toolstash -cmp, to clean these up.
This CL passes toolstash -cmp.
Change-Id: I4af8a10cc59078c436130ce472f25abc3a9b2f80
Reviewed-on: https://go-review.googlesource.com/20890
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Boolean expressions involving t.Thistuple were converted to use
t.Recv(), because it's a bit clearer and will hopefully reveal cases
where we could remove redundant calls to t.Recv() (in followup CLs).
The other cases were all converted to use t.Recvs().NumFields(),
t.Params().NumFields(), or t.Results().NumFields().
Change-Id: I4df91762e7dc4b2ddae35995f8dd604a52c09b09
Reviewed-on: https://go-review.googlesource.com/20796
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
This is an automated rewrite of all the calls of the form:
for f, it := IterFields(t); f != nil; f = it.Next() { ... }
Followup CLs will work on cleaning up the remaining cases.
Change-Id: Ic1005ad45ae0b50c63e815e34e507e2d2644ba1a
Reviewed-on: https://go-review.googlesource.com/20794
Reviewed-by: David Crawshaw <crawshaw@golang.org>
This should probably be considered "experimental" at this stage, but
what it needs is feedback from adventurous adopters. I think the data
structure used for describing escape reasons might be extendable to
allow a cleanup of the underlying algorithms, which suffers from
insufficiently separated concerns (the graph does not deal well with
escape level adjustments, so it is augmented by a second custom-walk
portion of the "flood" phase. It would be better to put it all,
including level adjustments, in a single graph structure, and then
simply flood the graph.
Tweaked to avoid allocations in the no-logging case.
Modified run.go to ignore lines with leading "#" in the output (since
it can never match a line), and in -update_errors to ignore leading
tabs in output lines and to normalize embedded filenames.
Currently requires -m -m because otherwise the noise/update
burden for the other escape tests is considerable.
There is a partial test. Existing escape analysis tests seem to
cover all except the panic case and what looks like it might be
unreachable code in escape analysis.
Fixes#10526.
Change-Id: I2524fdec54facae48b00b2548e25d9e46fcaf832
Reviewed-on: https://go-review.googlesource.com/18041
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
The obj.Fmt* values are only used by gc/fmt.go, so just move them
there. Also, add comments documenting the correspondance between
FmtFoo names and their flag characters to make understanding the
existing documentation slightly less confusing.
While here, add a new FmtFlag named type to represent these values.
Change-Id: I9631214b892557d094823f1ac575d0c43a84007b
Reviewed-on: https://go-review.googlesource.com/20717
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Currently, the only use for this is on the Left side of OKEY nodes
within struct literals. esc and fmt only care so they can recognize
that the ONAME nodes are actually field names, which need special
handling.
sinit additionally needs to know the field's offset within the struct,
which we can provide via Xoffset.
Passes toolstash/buildall.
Change-Id: I362d965e161f4d80fcd9c9bae0dfacc657dc0b29
Reviewed-on: https://go-review.googlesource.com/20676
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Switch TSTRUCT and TINTER to use Fields instead of Type, which wrings
out the remaining few direct uses of the latter.
Preparation for converting fields to use a separate "Field" type.
Passes toolstash/buildall.
Change-Id: I5a2ea7e159d0dde1be2c9afafc10a8f739d95743
Reviewed-on: https://go-review.googlesource.com/20675
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
Change-Id: Ice3aa807169f4fec85745a3991b1084a9f85c1b5
Reviewed-on: https://go-review.googlesource.com/20499
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Accessing the n'th field of a struct is fairly common, and in
particular accessing the 0'th field of the receiver parameter list is
very common. Add helper methods for both of these tasks and update
code to make use of them.
Change-Id: I81f551fecdca306b3800636caebcd0dc106f2ed7
Reviewed-on: https://go-review.googlesource.com/20498
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Dave Cheney <dave@cheney.net>
Mix in several other minor cleanups, including adding some new methods
to Nodes: Index, Addr, SetIndex, SetNodes.
Passes toolstash -cmp.
Update #14473.
Change-Id: I8bd4ae3fde7c5e20ba66e7dd1654fbc70c3ddeb8
Reviewed-on: https://go-review.googlesource.com/20491
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
This CL was automatically generated using a special-purpose AST
rewriting tool, followed by manual editing to put some comments back in
the right places and fix some bad line breaks.
The result is not perfect but it's a big step toward getting back to
sanity, and because it was automatically generated there is a decent
chance that it is correct.
Passes toolstash -cmp.
Update #14473.
Change-Id: I01c09078a6d78e2b008bc304d744b79469a38d3d
Reviewed-on: https://go-review.googlesource.com/20440
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
More idiomatic naming (in particular, matches the naming used for
go/types.Signature).
Also, convert more code to use these methods and/or IterFields.
(Still more to go; only made a quick pass for low hanging fruit.)
Passes toolstash -cmp.
Change-Id: I61831bfb1ec2cd50d4c7efc6062bca4e0dcf267b
Reviewed-on: https://go-review.googlesource.com/20451
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Found by temporarily flipping fields from *NodeList to Nodes and fixing
all the compilation errors. This CL does not actually change any
fields.
Passes toolstash -cmp.
Update #14473.
Change-Id: Ib98fa37e8752f96358224c973a743618a6a0e736
Reviewed-on: https://go-review.googlesource.com/20320
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Eliminates type conversions in a bunch of Oconv(int(n.Op), ...) calls.
Notably, this identified a misuse of Oconv in amd64/gsubr.go to try to
print an assembly instruction op instead of a compiler node op.
Change-Id: I93b5aa49fe14a5eaf868b05426d3b8cd8ab52bc5
Reviewed-on: https://go-review.googlesource.com/20298
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Move a few local fields all the way to []*Node while I'm at it.
Update #14473.
Change-Id: Ib18360879839ac592f778cf1042f111bdf14add3
Reviewed-on: https://go-review.googlesource.com/20197
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
- removed lots of unnecessary int(x) casts
- removed parserline() - was inconsistently used anyway
- minor simplifications in dcl.go
Change-Id: Ibf7de679eea528a31c9692ef1c76a1d9b3239211
Reviewed-on: https://go-review.googlesource.com/20131
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
The tree's pretty inconsistent about single space vs double space
after a period in documentation. Make it consistently a single space,
per earlier decisions. This means contributors won't be confused by
misleading precedence.
This CL doesn't use go/doc to parse. It only addresses // comments.
It was generated with:
$ perl -i -npe 's,^(\s*// .+[a-z]\.) +([A-Z]),$1 $2,' $(git grep -l -E '^\s*//(.+\.) +([A-Z])')
$ go test go/doc -update
Change-Id: Iccdb99c37c797ef1f804a94b22ba5ee4b500c4f7
Reviewed-on: https://go-review.googlesource.com/20022
Reviewed-by: Rob Pike <r@golang.org>
Reviewed-by: Dave Day <djd@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
A slice uses less memory than a NodeList, and has better memory locality
when walking the list.
This uncovered a tricky case involving closures: the escape analysis
pass when run on a closure was appending to the Dcl list of the OCLOSURE
rather than the ODCLFUNC. This happened to work because they shared the
same NodeList. Fixed with a change to addrescapes, and a check to
Tempname to catch any recurrences.
This removes the last use of the listsort function outside of tests.
I'll send a separate CL to remove it.
Unfortunately, while this passes all tests, it does not pass toolstash
-cmp. The problem is that cmpstackvarlt does not fully determine the
sort order, and the change from listsort to sort.Sort, while generally
desirable, produces a different ordering. I could stage this by first
making cmpstackvarlt fully determined, but no matter what toolstash -cmp
is going to break at some point.
In my casual testing the compiler is 2.2% faster.
Update #14473.
Change-Id: I367d66daa4ec73ed95c14c66ccda3a2133ad95d5
Reviewed-on: https://go-review.googlesource.com/19919
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Walking the field name as if it were an expression
caused a called to haspointers with a TFIELD, which panics.
Trigger was a field at a large offset within a large struct,
combined with a struct literal expression mentioning that
field.
Fixes#14405
Change-Id: I4589badae27cf3d7cf365f3a66c13447512f41f9
Reviewed-on: https://go-review.googlesource.com/19699
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Consider this code:
func f(*int)
func g() {
p := new(int)
f(p)
}
where f is an assembly function.
In general liveness analysis assumes that during the call to f, p is dead
in this frame. If f has retained p, p will be found alive in f's frame and keep
the new(int) from being garbage collected. This is all correct and works.
We use the Go func declaration for f to give the assembly function
liveness information (the arguments are assumed live for the entire call).
Now consider this code:
func h1() {
p := new(int)
syscall.Syscall(1, 2, 3, uintptr(unsafe.Pointer(p)))
}
Here syscall.Syscall is taking the place of f, but because its arguments
are uintptr, the liveness analysis and the garbage collector ignore them.
Since p is no longer live in h once the call starts, if the garbage collector
scans the stack while the system call is blocked, it will find no reference
to the new(int) and reclaim it. If the kernel is going to write to *p once
the call finishes, reclaiming the memory is a mistake.
We can't change the arguments or the liveness information for
syscall.Syscall itself, both for compatibility and because sometimes the
arguments really are integers, and the garbage collector will get quite upset
if it finds an integer where it expects a pointer. The problem is that
these arguments are fundamentally untyped.
The solution we have taken in the syscall package's wrappers in past
releases is to insert a call to a dummy function named "use", to make
it look like the argument is live during the call to syscall.Syscall:
func h2() {
p := new(int)
syscall.Syscall(1, 2, 3, uintptr(unsafe.Pointer(p)))
use(unsafe.Pointer(p))
}
Keeping p alive during the call means that if the garbage collector
scans the stack during the system call now, it will find the reference to p.
Unfortunately, this approach is not available to users outside syscall,
because 'use' is unexported, and people also have to realize they need
to use it and do so. There is much existing code using syscall.Syscall
without a 'use'-like function. That code will fail very occasionally in
mysterious ways (see #13372).
This CL fixes all that existing code by making the compiler do the right
thing automatically, without any code modifications. That is, it takes h1
above, which is incorrect code today, and makes it correct code.
Specifically, if the compiler sees a foreign func definition (one
without a body) that has uintptr arguments, it marks those arguments
as "unsafe uintptrs". If it later sees the function being called
with uintptr(unsafe.Pointer(x)) as an argument, it arranges to mark x
as having escaped, and it makes sure to hold x in a live temporary
variable until the call returns, so that the garbage collector cannot
reclaim whatever heap memory x points to.
For now I am leaving the explicit calls to use in package syscall,
but they can be removed early in a future cycle (likely Go 1.7).
The rule has no effect on escape analysis, only on liveness analysis.
Fixes#13372.
Change-Id: I2addb83f70d08db08c64d394f9d06ff0a063c500
Reviewed-on: https://go-review.googlesource.com/18584
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Brief background on "why heap allocate". Things can be
forced to the heap for the following reasons:
1) address published, hence lifetime unknown.
2) size unknown/too large, cannot be stack allocated
3) multiplicity unknown/too large, cannot be stack allocated
4) reachable from heap (not necessarily published)
The bug here is a case of failing to enforce 4) when an
object Y was reachable from a heap allocation X forced
because of 3). It was found in the case of a closure
allocated within a loop (X) and assigned to a variable
outside the loop (multiplicity unknown) where the closure
also captured a map (Y) declared outside the loop (reachable
from heap). Note the variable declared outside the loop (Y)
is not published, has known size, and known multiplicity
(one). The only reason for heap allocation is that it was
reached from a heap allocated item (X), but because that was
not forced by publication, it has to be tracked by loop
level, but escape-loop level was not tracked and thus a bug
results.
The fix is that when a heap allocation is newly discovered,
use its looplevel as the minimum loop level for downstream
escape flooding.
Every attempt to generalize this bug to X-in-loop-
references-Y-outside loop succeeded, so the fix was aimed
to be general. Anywhere that loop level forces heap
allocation, the loop level is tracked. This is not yet
tested for all possible X and Y, but it is correctness-
conservative and because it caused only one trivial
regression in the escape tests, it is probably also
performance-conservative.
The new test checks the following:
1) in the map case, that if fn escapes, so does the map.
2) in the map case, if fn does not escape, neither does the map.
3) in the &x case, that if fn escapes, so does &x.
4) in the &x case, if fn does not escape, neither does &x.
Fixes#13799.
Change-Id: Ie280bef2bb86ec869c7c206789d0b68f080c3fdb
Reviewed-on: https://go-review.googlesource.com/18234
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Update old c-style comments to look like Go comments. Also replace some
lingering references to old .c files that don't exist anymore.
Change-Id: I72b2407a40fc76c23e9048643e0622fd70b4cf90
Reviewed-on: https://go-review.googlesource.com/16190
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Turns out the summary information for the ... args was
already correctly computed, all that lacked was to make
use of it and correct tests that documented our prior
deficiencies.
Fixes#12006
Change-Id: Ie8adfab7547f179391d470679598f0904aabf9f7
Reviewed-on: https://go-review.googlesource.com/15200
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
The existing test did not take into account the implicit
dereference of &fixedArray and thus heap-escaped when it
was not necessary.
Also added a detailed test for this and related cases.
Fixes#12588
Change-Id: I951e9684a093082ccdca47710f69f4366bd6b3cf
Reviewed-on: https://go-review.googlesource.com/15130
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Found with https://github.com/remyoudompheng/go-misc/deadcode:
deadcode: walk.go:2228:1: applywritebarrier_bv is unused
deadcode: subr.go:355:1: gethunk is unused
deadcode: subr.go:1991:1: localexpr is unused
deadcode: dcl.go:82:1: poptodcl is unused
deadcode: swt.go:810:1: dumpcase is unused
deadcode: esc.go:251:1: satAdd8 is unused
deadcode: esc.go:387:1: outputsPerTag is unused
deadcode: obj.go:190:1: duint64 is unused
deadcode: obj.go:287:1: dstringptr is unused
deadcode: plive.go:95:1: xmalloc is unused
deadcode: plive.go:119:1: freeblock is unused
followed by
deadcode: go.go:633:1: hunk is unused
deadcode: go.go:635:1: nhunk is unused
deadcode: go.go:637:1: thunk is unused
after 'gethunk' was removed.
Some dead code in bv.go, mparith3.go, and dcl.go was left as is.
Passes go build -a -toolexec 'toolstash -cmp' std cmd.
Change-Id: Ia63519adedc8650d7095572ddd454fd923d3204d
Reviewed-on: https://go-review.googlesource.com/14610
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Convert some fields of struct Type in go.go from uint8 to bool.
This change passes go build -toolexec 'toolstash -cmp' -a std.
Change-Id: I0a6c53f8ee686839b5234010ee2de7ae3940d499
Reviewed-on: https://go-review.googlesource.com/14370
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This one of a set of changes to make the transition away from NodeList
easier by removing cases in which NodeList doesn't act semi-trivially like a
[]*Node.
This CL was originally prepared by Josh Bleecher Snyder <josharian@gmail.com>.
This change passes go build -toolexec 'toolstash -cmp' -a std.
Change-Id: I582ff8b077eb384b84721a1edb0c1efbc0c40059
Reviewed-on: https://go-review.googlesource.com/14304
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Dave Cheney <dave@cheney.net>
Run-TryBot: Dave Cheney <dave@cheney.net>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This helps vet see a real issue:
cmd/internal/gc$ go vet
gen.go:1223: unreachable code
Fixes#12106.
Change-Id: I720868b07ae6b6d5a4dc6b238baa8c9c889da6d8
Reviewed-on: https://go-review.googlesource.com/14083
Reviewed-by: Minux Ma <minux@golang.org>
Run-TryBot: Minux Ma <minux@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
ODOTTYPE should be treated a whole lot like ODOT,
but it was missing completely from the switch in
escwalk and thus escape status did not propagate
to fields.
Since interfaces are required to trigger this bug,
the test was added to escape_iface.go.
Fixes#11931.
Change-Id: Id0383981cc4b1a160f6ad447192a112eed084538
Reviewed-on: https://go-review.googlesource.com/12921
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Reviewed-by: Russ Cox <rsc@golang.org>
Suggested during code reviews of last 15 CLs (or so).
Change-Id: If780f6eb47a7a31df133c64d5dcf0eaf04d8447b
Reviewed-on: https://go-review.googlesource.com/10675
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
