allow the loong64 CALL* ops to take variable number of args
Update #40724
Co-authored-by: Xiaolin Zhao <zhaoxiaolin@loongson.cn>
Change-Id: I4706d9651fcbf9a0f201af6820c97b1a924f14e3
Reviewed-on: https://go-review.googlesource.com/c/go/+/521781
Auto-Submit: David Chase <drchase@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Meidan Li <limeidan@loongson.cn>
Reviewed-by: David Chase <drchase@google.com>
Turns out after adding the generic implementation for And/Or we ended up
with duplicated ops that are exactly the same for arm.
Apologies for the oversight, this CL removes the redundant arm code and
adds arm to the generic build flags.
For #61395
Change-Id: Id5e5a5cf113774948f8e772592e898d0810ad1f6
GitHub-Last-Rev: 4d8c857d15
GitHub-Pull-Request: golang/go#64299
Reviewed-on: https://go-review.googlesource.com/c/go/+/544017
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
According to review, buildmode=shared has unfixed shortcomings and
should be considered legacy at this time. So only buildmode=plugin is
going to be added for loong64 which is a relatively new platform.
Change-Id: Iac0b9f57e4ee01755458e180bb24d1b2a146fdf0
Reviewed-on: https://go-review.googlesource.com/c/go/+/480878
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Run-TryBot: WANG Xuerui <git@xen0n.name>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: abner chenc <chenguoqi@loongson.cn>
Reviewed-by: Meidan Li <limeidan@loongson.cn>
This is needed before actual support for buildmode=plugin is added.
Should not affect current behavior.
Change-Id: I86371d7e373fd529cb8710850d7b0fbbf1eb52ca
Reviewed-on: https://go-review.googlesource.com/c/go/+/480877
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Meidan Li <limeidan@loongson.cn>
Reviewed-by: abner chenc <chenguoqi@loongson.cn>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Run-TryBot: WANG Xuerui <git@xen0n.name>
TryBot-Result: Gopher Robot <gobot@golang.org>
Add R21 to the allocatable registers, use R20 and R21 in duff
device. This CL is in preparation for subsequent regABI support.
Updates #40724
Co-authored-by: Xiaolin Zhao <zhaoxiaolin@loongson.cn>
Change-Id: If1661adc0f766925fbe74827a369797f95fa28a9
Reviewed-on: https://go-review.googlesource.com/c/go/+/521775
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Reviewed-by: Meidan Li <limeidan@loongson.cn>
Reviewed-by: Than McIntosh <thanm@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
The function passed to OnceFunc/OnceValue/OnceValues may transitively
keep more allocations alive. As the passed function is guaranteed to be
called at most once, it is safe to drop it after the first call is
complete. This avoids keeping the passed function (and anything it
transitively references) alive until the returned function is GCed.
Change-Id: I2faf397b481d2f693ab3aea8e2981b02adbc7a21
Reviewed-on: https://go-review.googlesource.com/c/go/+/481515
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: qiulaidongfeng <2645477756@qq.com>
Some small changes to help reduce compile time for function flags
computation. The current implementation of panic path detection adds
an entry to a map for every node in the function, which is wasteful
(and shows up in cpu profiles). Switch to only adding entries where
they are useful. This is especially important for functions with large
map literals and other constructs with many non-statement nodes.
Change-Id: I9cfb2cd1cbf480f21298e6102aa99e2d77219f3d
Reviewed-on: https://go-review.googlesource.com/c/go/+/539696
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Fix a bug in the code that analyzes how function parameters are used,
which wasn't properly handling certain types of closures. The code
in question has support for deriving flags for a given parameter based
on whether it is passed to a call. Example:
func foo(f1 func(int)) {
bar(32, f1)
}
func bar(x int, f2 func()) {
f2(x)
}
When analyzing "bar", we can derive the "FeedsIndirectCall" flag for
parameter "f1" by virtue of the fact that it is passed directly to
"bar", and bar's corresponding parameter "f2" has the flag set. For
a more complex example such as
func foo(f1 func(int)) func() int {
return func(q int) int { <<-- HERE
bar(99, f1)
return q
}
}
func bar(x int, f2 func()) {
f2(x)
}
The heuristics code would panic when examining the closure marked above
due to the fact that the call to "bar" was passing an ir.Name with class
PPARAM, but no such param was present in the enclosing function.
Change-Id: I30436ce716b51bfb03e42e7abe76a4514e6b9285
Reviewed-on: https://go-review.googlesource.com/c/go/+/539320
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
These tests are checking the output of test functions that call the
Helper methods. However, they were reaching into package internals
instead of running those test functions as actual tests.
That not only produced significant differences in formatting (such as
indentation for subtests), but also caused test flags such as
"-failfast" passed for the overall test run to interfere with the
output formatting.
Now, we run the test functions as real tests in a subprocess,
so that we get the real output and formatting of those tests.
This makes the tests not only more realistic, but also less
sensitive to otherwise-irrelevant implementation details
(such as the names and signatures of unexported types and
functions in the testing package).
Fixes#61016.
Change-Id: I646fbbd7cfeb00382054677f726c05fc9d35d0dc
Reviewed-on: https://go-review.googlesource.com/c/go/+/506955
Auto-Submit: Bryan Mills <bcmills@google.com>
Reviewed-by: Alan Donovan <adonovan@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
When running the code to compute function properties that feed
inlining heuristics, the existing heuristics implementation makes
fairly extensive use of ir.StaticValue and ir.Reassigned to sharpen
the analysis. These calls turn out to cause a significant compile time
increase, due to the fact that each call can potentially walk every
node in the IR for the function. To help with this problem, switch the
heuristics code over to using the new "batch mode" reassignment helper
added in the previous CL.
Change-Id: Ib15a62416134386e34b7cfa1130a4b413a37b225
Reviewed-on: https://go-review.googlesource.com/c/go/+/537977
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
If GOROOT is a symbolic link,
the paths obtained from the
first and second Getwd of TestFileChdir are different,
and this CL fixes the test failure in this situation.
Fixes#64281
Change-Id: I53026b6c54a54be08833396e2c7081ca3ab8c282
GitHub-Last-Rev: 5cc418e625
GitHub-Pull-Request: golang/go#64001
Reviewed-on: https://go-review.googlesource.com/c/go/+/540521
Reviewed-by: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Auto-Submit: Bryan Mills <bcmills@google.com>
Reviewed-by: Bryan Mills <bcmills@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Add a new helper type 'ReassignOracle', useful for doing "batch mode"
reassignment analysis, e.g. deciding whether a given ir.Name or (chain
of connected names) has a single definition and is never reassigned.
The intended usage model is for clients to create/initialize a
ReassignOracle for a given function, then make a series of queries
using it (with the understanding that changing/mutating the func body
IR can invalidate the info cached in the oracle). This oracle is
intended to provide the same sort of analysis that ir.StaticValue and
ir.Reassigned carry out, but at a much reduced cost in compile
time.
Notes:
- the new helper isn't actually used for anything useful in this
patch; it will be hooked into the inline heuristics code as part of
a subsequent CL.
- this is probably not an ideal long-term solution; it would be better
to switch to a scheme based a flag on ir.Name, as opposed to a
side table.
Change-Id: I283e748e440a9f595df495f6aa48ee9c498702d9
Reviewed-on: https://go-review.googlesource.com/c/go/+/539319
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
For 'case Array' in IsZero, check 'v.flag&flagIndir == 0' in the
first place, rename 'array' to 'typ' for consistency, and remove
stale comment.
Add line breaks for long sentence in isZero.
Change-Id: Id06d01fd61eefd205bf4626e6b920ae82b459455
GitHub-Last-Rev: 7225ca3f7b
GitHub-Pull-Request: golang/go#64270
Reviewed-on: https://go-review.googlesource.com/c/go/+/543656
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Jes Cok <xigua67damn@gmail.com>
Reviewed-by: Keith Randall <khr@golang.org>
Auto-Submit: Keith Randall <khr@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Apparently, on Windows, throwing an exception on a non-system-
allocated crash stack causes EXCEPTION_STACK_OVERFLOW and hangs
the process (see issue #63938). Disable crash stack for now, which
gets us back the the behavior of Go 1.21.
Fixes#63938.
Change-Id: I4c090315b93b484e756b242f0de7a9e02f199261
Reviewed-on: https://go-review.googlesource.com/c/go/+/543996
Reviewed-by: Michael Knyszek <mknyszek@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Quim Muntal <quimmuntal@gmail.com>
This is inspired by CL 539915, I'm only submitting now that
CL 456435 has been merged.
This divide the number of objects kept alive by the heap by two
and remove the slice header allocation in New and in the put back.
Change-Id: Ibcd5166bac5a37f365a533e09a28f3b79f81ad58
Reviewed-on: https://go-review.googlesource.com/c/go/+/543515
Reviewed-by: Damien Neil <dneil@google.com>
Auto-Submit: Damien Neil <dneil@google.com>
Reviewed-by: qiulaidongfeng <2645477756@qq.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Change-Id: Iacbcb582e263149fede734822cba2df4b8162968
Reviewed-on: https://go-review.googlesource.com/c/go/+/544015
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Tim King <taking@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
As of Go 1.19, runtime.GOROOT() reports the empty string if the binary
was built with -trimpath. cmd/go/internal/cfg uses the path of the go
command executable to reverse-engineer the correct GOROOT setting,
but that means that cmd/go's "GOPATH set to GOROOT" warning needs to
use cfg.GOROOT instead of runtime.GOROOT().
In addition, if we fail to find the GOROOT then there is no point in
complaining about GOPATH also being empty: the missing GOROOT will stop
everything right away anyway, so there is no point confusing the user
with an additional warning about GOPATH.
Updates #51461.
Updates #18678.
Updates #3207.
Change-Id: Id7d0f4dc2f229c202dfda4e6e8af5dea909bb16f
Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest,gotip-windows-amd64-longtest
Reviewed-on: https://go-review.googlesource.com/c/go/+/543955
Reviewed-by: Michael Matloob <matloob@golang.org>
Auto-Submit: Bryan Mills <bcmills@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
go get golang.org/x/tools@8966034e # CL 542639
go mod tidy
go mod vendor
Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest
Change-Id: Id2a7322006ff3985e3c97482cd78c8963c4df140
Reviewed-on: https://go-review.googlesource.com/c/go/+/543975
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Tim King <taking@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Robert Findley <rfindley@google.com>
Currently if morestack on g0 happens the wasm runtime prints
"RuntimeError: memory access out of bounds", which is quite misleading.
By switching to a crash stack we can get better stacktraces
for the error.
There is no way to automate tests for this feature on wasm, since
TestG0StackOverflow relies on spawning a subprocess which is not
supported by the wasm port.
The way I got this tested manually is to comment everything in
TestG0StackOverflow, leaving just runtime.G0StackOverflow().
Then it is a matter of invoking the test:
GOOS=js GOARCH=wasm go test runtime -v -run=TestG0StackOverflow
Change-Id: If450f3ee5209bb32efc1abd0a34b1cc4a29d0c46
GitHub-Last-Rev: 0d7c396e4c
GitHub-Pull-Request: golang/go#63956
Reviewed-on: https://go-review.googlesource.com/c/go/+/539995
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
CL 543657 dedup'd the go/defer statement recognition between the
inliner and static devirtualizer. This CL extends that for PGO-based
devirtualization too.
Change-Id: I998753132af1ef17329676f4e17515f16e0acb03
Reviewed-on: https://go-review.googlesource.com/c/go/+/543775
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Michael Pratt <mpratt@google.com>
This reverts commit 6382893890.
Reason for revert: Causes failures on big endian platforms and riscv64.
Possibly a bug in the generic implementation.
For #64284.
For #64285.
Change-Id: Ic1bb8533d9641fae28d0337b36d434b9a575cd7e
Reviewed-on: https://go-review.googlesource.com/c/go/+/543895
Reviewed-by: Heschi Kreinick <heschi@google.com>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
TryBot-Bypass: Michael Knyszek <mknyszek@google.com>
This CL interleaves devirtualization and inlining, so that
devirtualized calls can be inlined.
Fixes#52193.
Change-Id: I681e7c55bdb90ebf6df315d334e7a58f05110d9c
Reviewed-on: https://go-review.googlesource.com/c/go/+/528321
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Bypass: Matthew Dempsky <mdempsky@google.com>
This change introduces new options to set the floating point
mode on ARM targets. The GOARM version number can optionally be
followed by ',hardfloat' or ',softfloat' to select whether to
use hardware instructions or software emulation for floating
point computations, respectively. For example,
GOARM=7,softfloat.
Previously, software floating point support was limited to
GOARM=5. With these options, software floating point is now
extended to all ARM versions, including GOARM=6 and 7. This
change also extends hardware floating point to GOARM=5.
GOARM=5 defaults to softfloat and GOARM=6 and 7 default to
hardfloat.
For #61588
Change-Id: I23dc86fbd0733b262004a2ed001e1032cf371e94
Reviewed-on: https://go-review.googlesource.com/c/go/+/514907
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Michael Knyszek <mknyszek@google.com>
Make it clear that the `block` argument is not always used, and stop
passing it in a case where it didn't really make sense - at the end of a
function.
Change-Id: I2fa86bed6ceb53a1b1cbfda5c3392e7e9da9579d
Reviewed-on: https://go-review.googlesource.com/c/go/+/502115
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: David Chase <drchase@google.com>
os.OpenFile on windows did not use the O_SYNC flag. This meant
that even if the user set O_SYNC, os.OpenFile would ignore it.
This change adds a new flag FILE_FLAG_WRITE_THROUGH, which is
the equivalent of O_SYNC flag on Linux and is documented in
https://docs.microsoft.com/en-us/windows/win32/api/fileapi/nf-fileapi-createfileaFixes#35358
Change-Id: Ib338caed5bb2f215723bfe30a2551a83998d92c9
GitHub-Last-Rev: 82c6275cb4
GitHub-Pull-Request: golang/go#64027
Reviewed-on: https://go-review.googlesource.com/c/go/+/541015
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Emmanuel Odeke <emmanuel@orijtech.com>
Run-TryBot: Jes Cok <xigua67damn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Quim Muntal <quimmuntal@gmail.com>
Following CL 543035, this CL updates race syso for Linux/PPC64LE.
Now we have update all of them (except OpenBSD).
For #61395.
Fixes#62624.
Change-Id: I9e1d758355114a50ff206e5d78dc4ea8a06367d8
Reviewed-on: https://go-review.googlesource.com/c/go/+/543397
Run-TryBot: Cherry Mui <cherryyz@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Mauri de Souza Meneguzzo <mauri870@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
There doesn't appear to be any need for this code. EditChildren won't
recurse into the closure body anyway.
Split out into a separate commit in case I'm overlooking something.
Change-Id: I004d1aa04865896de972bf3323b1622cc08a0d18
Reviewed-on: https://go-review.googlesource.com/c/go/+/543659
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Than McIntosh <thanm@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
The early return here is meant to suppress inlining of the function
call itself. However, it also suppresses recursing to visit the call
arguments, which are safe to inline.
Change-Id: I75887574c00931cb622277d04a822bc84c29bfa2
Reviewed-on: https://go-review.googlesource.com/c/go/+/543658
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Than McIntosh <thanm@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
The devirtualizer and inliner both want to recognize call expressions
that are part of a go or defer statement. This CL refactors them to
use a single CallExpr.GoDefer flag, which gets set during
normalization of go/defer statements during typecheck.
While here, drop some OCALLMETH assertions. Typecheck has been
responsible for desugaring them into OCALLFUNC for a while now, and
ssagen will check this again for us later anyway.
Change-Id: I3fc370f4417431aae97239313da6fe523f512a2e
Reviewed-on: https://go-review.googlesource.com/c/go/+/543657
Reviewed-by: Than McIntosh <thanm@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
This reverts commit go.dev/cl/528657.
Reason for revert: broke a lot of builders.
Change-Id: I70c33062020e997c4df67b3eaa2e886cf0da961e
Reviewed-on: https://go-review.googlesource.com/c/go/+/543660
Reviewed-by: Than McIntosh <thanm@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Without having all the architectures implementing the And/Or operators
merged I can't proceed with the public sync/atomic apis. This CL adds a
generic implementation that should work for all the remaining arches,
while waiting for the native assembly implementations in CL 531835,
CL 531678, CL 531895.
I regret the oversight of not pushing this earlier.
For #61395
Change-Id: Ib2d67f359fe324b4743eb79e9c8e52e8f6f5476c
GitHub-Last-Rev: d350927ba1
GitHub-Pull-Request: golang/go#64214
Reviewed-on: https://go-review.googlesource.com/c/go/+/543175
Reviewed-by: Cherry Mui <cherryyz@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
Wasmtime 14.0.0 introduced new CLI flags and removed the existing
flags, in particular the --max-wasm-stack flag we were using to avoid
errors in some tests.
This introduces a regular expression based switch that uses the old
flags for wasmtime versions < 14 and the new flags otherwise.
Fixes#63718
Change-Id: I44673e7d9f8729065757abdbf8c41e8a61897d6a
Reviewed-on: https://go-review.googlesource.com/c/go/+/541219
Reviewed-by: David Chase <drchase@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Dmitri Shuralyov <dmitshur@golang.org>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
Reviewed-by: Mauri de Souza Meneguzzo <mauri870@gmail.com>
For some types where the zero value is a value where all bits of this type are 0 optimize it.
goos: windows
goarch: amd64
pkg: reflect
cpu: AMD Ryzen 7 7840HS w/ Radeon 780M Graphics
│ old.txt │ new.txt │
│ sec/op │ sec/op vs base │
IsZero/StructInt_512-16 109.75n ± 0% 72.61n ± 1% -33.84% (p=0.000 n=12)
Change-Id: I56de8b95f4d4482068960d6f38938763fa1caa90
GitHub-Last-Rev: c143f0cd76
GitHub-Pull-Request: golang/go#64220
Reviewed-on: https://go-review.googlesource.com/c/go/+/543355
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Auto-Submit: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
