]> Cypherpunks repositories - gostls13.git/log
gostls13.git
9 years agocrypto/tls: len(clientRandom) used for serverRandom source
Ralph Corderoy [Sat, 7 Nov 2015 11:45:00 +0000 (11:45 +0000)]
crypto/tls: len(clientRandom) used for serverRandom source

In keysFromMasterSecret(), don't copy from serverRandom into
seed[:len(clientRandom)].  Actually, switch from an array to a slice in
keysFromMasterSecret() and masterFromPreMasterSecret() so the length
need not be given;  that's how it's done elsewhere in the file.

Fixes #13181

Change-Id: I92abaa892d1bba80c2d4f12776341cda7d538837
Reviewed-on: https://go-review.googlesource.com/16697
Run-TryBot: Adam Langley <agl@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Adam Langley <agl@golang.org>
9 years agocrypto/tls: don't send IP literals as SNI values.
Adam Langley [Tue, 10 Nov 2015 19:18:50 +0000 (11:18 -0800)]
crypto/tls: don't send IP literals as SNI values.

(This relands commit a4dcc692011bf1ceca9b1a363fd83f3e59e399ee.)

https://tools.ietf.org/html/rfc6066#section-3 states:

  “Literal IPv4 and IPv6 addresses are not permitted in "HostName".”

However, if an IP literal was set as Config.ServerName (which could
happen as easily as calling Dial with an IP address) then the code would
send the IP literal as the SNI value.

This change filters out IP literals, as recognised by net.ParseIP, from
being sent as the SNI value.

Fixes #13111.

Change-Id: I6e544a78a01388f8fe98150589d073b917087f75
Reviewed-on: https://go-review.googlesource.com/16776
Run-TryBot: Adam Langley <agl@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocrypto/elliptic,crypto/ecdsa: P256 amd64 assembly
Vlad Krasnov [Fri, 17 Apr 2015 13:10:35 +0000 (06:10 -0700)]
crypto/elliptic,crypto/ecdsa: P256 amd64 assembly

This is based on the implementation used in OpenSSL, from a
submission by Shay Gueron and myself. Besides using assembly,
this implementation employs several optimizations described in:

    S.Gueron and V.Krasnov, "Fast prime field elliptic-curve
                             cryptography with 256-bit primes"

In addition a new and improved modular inverse modulo N is
implemented here.

The performance measured on a Haswell based Macbook Pro shows 21X
speedup for the sign and 9X for the verify operations.
The operation BaseMult is 30X faster (and the Diffie-Hellman/ECDSA
key generation that use it are sped up as well).

The adaptation to Go with the help of Filippo Valsorda

Updated the submission for faster verify/ecdh, fixed some asm syntax
and API problems and added benchmarks.

Change-Id: I86a33636747d5c92f15e0c8344caa2e7e07e0028
Reviewed-on: https://go-review.googlesource.com/8968
Run-TryBot: Adam Langley <agl@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Adam Langley <agl@golang.org>
9 years agocmd/compile/internal/gc: avoid potential endless loop in float printing
Robert Griesemer [Tue, 10 Nov 2015 19:32:26 +0000 (11:32 -0800)]
cmd/compile/internal/gc: avoid potential endless loop in float printing

The compiler should not usually call Fconv with an infinity, but if
it does, Fconv will end in an endless loop. Test for infinities early.

Change-Id: I48f366466538b0bd26a851e01258725025babaff
Reviewed-on: https://go-review.googlesource.com/16777
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/go: always use --whole-archive for gccgo packages
Ian Lance Taylor [Tue, 10 Nov 2015 19:25:15 +0000 (11:25 -0800)]
cmd/go: always use --whole-archive for gccgo packages

This is, in effect, what the gc toolchain does.  It fixes cases where Go
code refers to a C global variable; without this, if the global variable
was the only thing visible in the C code, the generated cgo file might
not get pulled in from the archive, leaving the Go variable
uninitialized.

This was reported against gccgo as https://gcc.gnu.org/PR68255 .

Change-Id: I3e769dd174f64050ebbff268fbbf5e6fab1e2a1b
Reviewed-on: https://go-review.googlesource.com/16775
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/dist, cmd/go, misc/cgo/testshared: update testshared and run it on arm
Michael Hudson-Doyle [Wed, 2 Sep 2015 10:26:42 +0000 (22:26 +1200)]
cmd/dist, cmd/go, misc/cgo/testshared: update testshared and run it on arm

And enable PIE in cmd/go because that's all it seems to take.

Change-Id: Ie017f427ace5e91de333a9f7cba9684c4641dfd5
Reviewed-on: https://go-review.googlesource.com/14222
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/internal/obj/arm: access global data via GOT on arm when -dynlink
Michael Hudson-Doyle [Sun, 5 Jul 2015 09:49:11 +0000 (21:49 +1200)]
cmd/internal/obj/arm: access global data via GOT on arm when -dynlink

Change-Id: I88034611f56cc06bb47b0c431075cc78ca8dbb09
Reviewed-on: https://go-review.googlesource.com/14188
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agoruntime: break atomics out into package runtime/internal/atomic
Michael Matloob [Mon, 2 Nov 2015 19:09:24 +0000 (14:09 -0500)]
runtime: break atomics out into package runtime/internal/atomic

This change breaks out most of the atomics functions in the runtime
into package runtime/internal/atomic. It adds some basic support
in the toolchain for runtime packages, and also modifies linux/arm
atomics to remove the dependency on the runtime's mutex. The mutexes
have been replaced with spinlocks.

all trybots are happy!
In addition to the trybots, I've tested on the darwin/arm64 builder,
on the darwin/arm builder, and on a ppc64le machine.

Change-Id: I6698c8e3cf3834f55ce5824059f44d00dc8e3c2f
Reviewed-on: https://go-review.googlesource.com/14204
Run-TryBot: Michael Matloob <matloob@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet/http: Show more data for unsolicited response
Erik Dubbelboer [Thu, 27 Aug 2015 06:09:15 +0000 (14:09 +0800)]
net/http: Show more data for unsolicited response

Show more than one character when we recieve a unsolicited
response on an idle HTTP channel. Showing more than one
byte is really useful when you want to debug your program
when you get this message.

Change-Id: I3caf9f06420e7c2a2de3e4eb302c5dab95428fdb
Reviewed-on: https://go-review.googlesource.com/13959
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agonet/http/internal: ignore chunk-extension when reading chunked encoding bodies
Brad Fitzpatrick [Thu, 5 Nov 2015 06:53:28 +0000 (07:53 +0100)]
net/http/internal: ignore chunk-extension when reading chunked encoding bodies

Fixes #13135

Change-Id: I45666f32cd91102211bf01a306edcb10deb65187
Reviewed-on: https://go-review.googlesource.com/16680
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agomisc/cgo/testsanitizers: fix check for existence of clang
Mohit Agarwal [Mon, 9 Nov 2015 20:22:52 +0000 (01:52 +0530)]
misc/cgo/testsanitizers: fix check for existence of clang

The current mechanism fails if clang cannot be executed by the current
user.  Using the `-x` operator for `test` return TRUE if the file is
executable by the user.

Change-Id: I0f3c8dc3880c1ce5a8a833ff3109eb96853184af
Reviewed-on: https://go-review.googlesource.com/16752
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/dist: run more cgo tests on ppc64x
Michael Hudson-Doyle [Thu, 3 Sep 2015 20:31:43 +0000 (08:31 +1200)]
cmd/dist: run more cgo tests on ppc64x

Change-Id: I992655bb02690ad95122a9e4c45cbd0948b545a0
Reviewed-on: https://go-review.googlesource.com/14238
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/link: enable external linking on ppc64
Michael Hudson-Doyle [Thu, 3 Sep 2015 20:30:15 +0000 (08:30 +1200)]
cmd/link: enable external linking on ppc64

Change-Id: Iffe8ccb55b2c555b2cb8c168cebfbfd5892212df
Reviewed-on: https://go-review.googlesource.com/14236
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/link, runtime: arm64 implementation of addmoduledata
Michael Hudson-Doyle [Wed, 24 Jun 2015 10:31:24 +0000 (22:31 +1200)]
cmd/link, runtime: arm64 implementation of addmoduledata

Change-Id: I62fb5b20d7caa51b77560a4bfb74a39f17089805
Reviewed-on: https://go-review.googlesource.com/13999
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet: fix off by one error while counting interfaces on windows
Alex Brainman [Fri, 6 Nov 2015 06:29:27 +0000 (17:29 +1100)]
net: fix off by one error while counting interfaces on windows

Fixes #12301

Change-Id: I8d01ec9551c6cff7e6129e06a7deb36a3be9de41
Reviewed-on: https://go-review.googlesource.com/16751
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agohtml/template, encoding/asn1: fix test errors
Dominik Honnef [Wed, 28 Oct 2015 07:10:46 +0000 (08:10 +0100)]
html/template, encoding/asn1: fix test errors

Change-Id: I1da1d718609eb6a7b78d29b173ec780bde22c687
Reviewed-on: https://go-review.googlesource.com/16422
Reviewed-by: Ralph Corderoy <ralph@inputplus.co.uk>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agoRevert "crypto/tls: don't send IP literals as SNI values."
Adam Langley [Mon, 9 Nov 2015 23:16:12 +0000 (15:16 -0800)]
Revert "crypto/tls: don't send IP literals as SNI values."

This reverts commit a4dcc692011bf1ceca9b1a363fd83f3e59e399ee.

Change-Id: Ib55fd349a604d6b5220dac20327501e1ce46b962
Reviewed-on: https://go-review.googlesource.com/16770
Reviewed-by: Adam Langley <agl@golang.org>
9 years agocrypto/tls: don't send IP literals as SNI values.
Adam Langley [Mon, 9 Nov 2015 03:10:14 +0000 (19:10 -0800)]
crypto/tls: don't send IP literals as SNI values.

https://tools.ietf.org/html/rfc6066#section-3 states:

  “Literal IPv4 and IPv6 addresses are not permitted in "HostName".”

However, if an IP literal was set as Config.ServerName (which could
happen as easily as calling Dial with an IP address) then the code would
send the IP literal as the SNI value.

This change filters out IP literals, as recognised by net.ParseIP, from
being sent as the SNI value.

Fixes #13111.

Change-Id: Ie9ec7acc767ae172b48c9c6dd8d84fa27b1cf0de
Reviewed-on: https://go-review.googlesource.com/16742
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Adam Langley <agl@golang.org>

9 years agocmd/link: look at the R_AARCH64_RELATIVE relocs to find the gcdata on arm64
Michael Hudson-Doyle [Wed, 1 Jul 2015 23:37:51 +0000 (11:37 +1200)]
cmd/link: look at the R_AARCH64_RELATIVE relocs to find the gcdata on arm64

Change-Id: I5a1864a27ad917aa65c8e65a133f6cc0a980d05f
Reviewed-on: https://go-review.googlesource.com/13998
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/internal/obj/arm64, cmd/link: use two instructions rather than three for loads...
Michael Hudson-Doyle [Mon, 3 Aug 2015 03:45:37 +0000 (15:45 +1200)]
cmd/internal/obj/arm64, cmd/link: use two instructions rather than three for loads from memory

Reduces size of godoc .text section by about 75k (or 1.4%).

Change-Id: I65850aa569aefbddd6cb07c6ae1addcc39cab6a5
Reviewed-on: https://go-review.googlesource.com/13993
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/go: clean the directory path containing the packages
Mohit Agarwal [Mon, 9 Nov 2015 19:26:14 +0000 (00:56 +0530)]
cmd/go: clean the directory path containing the packages

The heuristic for determining if the packages or commands are stale
fails as the mtime comparison happens even though the GOROOT and
current package paths are the same, since the path name isn't
canonicalized before the comparison (GOROOT is).

Fixes: #12690
Change-Id: Ia7d142fbbed8aac2bd2f71d1db4efd1f3ff5aece
Reviewed-on: https://go-review.googlesource.com/16483
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile/internal/gc: fix go.y to match y.go
Robert Griesemer [Mon, 9 Nov 2015 19:26:25 +0000 (11:26 -0800)]
cmd/compile/internal/gc: fix go.y to match y.go

In https://go-review.googlesource.com/#/c/16554/ y.go was modified
manually, but go.y (which is the source for y.go) was not changed.

Change-Id: I1273801bfd1ac65b875f4465033b0d062abff0b7
Reviewed-on: https://go-review.googlesource.com/16745
Reviewed-by: Austin Clements <austin@google.com>
9 years agoruntime: simplify chan ops, take 2
Keith Randall [Sun, 8 Nov 2015 05:28:21 +0000 (21:28 -0800)]
runtime: simplify chan ops, take 2

This change is the same as CL #9345 which was reverted,
except for a small bug fix.

The only change is to the body of sendDirect and its callsite.
Also added a test.

The problem was during a channel send operation.  The target
of the send was a sleeping goroutine waiting to receive.  We
basically do:
1) Read the destination pointer out of the sudog structure
2) Copy the value we're sending to that destination pointer
Unfortunately, the previous change had a goroutine suspend
point between 1 & 2 (the call to sendDirect).  At that point
the destination goroutine's stack could be copied (shrunk).
The pointer we read in step 1 is no longer valid for step 2.

Fixed by not allowing any suspension points between 1 & 2.
I suspect the old code worked correctly basically by accident.

Fixes #13169

The original 9345:

This change removes the retry mechanism we use for buffered channels.
Instead, any sender waking up a receiver or vice versa completes the
full protocol with its counterpart.  This means the counterpart does
not need to relock the channel when it wakes up.  (Currently
buffered channels need to relock on wakeup.)

For sends on a channel with waiting receivers, this change replaces
two copies (sender->queue, queue->receiver) with one (sender->receiver).
For receives on channels with a waiting sender, two copies are still required.

This change unifies to a large degree the algorithm for buffered
and unbuffered channels, simplifying the overall implementation.

Fixes #11506

Change-Id: I57dfa3fc219cffa4d48301ee15fe5479299efa09
Reviewed-on: https://go-review.googlesource.com/16740
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/link, runtime: arm implementation of addmoduledata
Michael Hudson-Doyle [Wed, 2 Sep 2015 10:25:57 +0000 (22:25 +1200)]
cmd/link, runtime: arm implementation of addmoduledata

Change-Id: I3975e10c2445e23c2798a7203a877ff2de3427c7
Reviewed-on: https://go-review.googlesource.com/14189
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/go: send all go build -n output to stderr
Dominik Honnef [Sun, 8 Nov 2015 19:27:14 +0000 (20:27 +0100)]
cmd/go: send all go build -n output to stderr

Also change a -v print, for consistency.

Fixes #12913

Change-Id: I6cc067d9f8dac66b1f9d1a675e0fbe0528371d0d
Reviewed-on: https://go-review.googlesource.com/16737
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/go, cmd/link: allow -buildmode=pie on linux/ppc64le
Michael Hudson-Doyle [Sun, 11 Oct 2015 23:32:03 +0000 (12:32 +1300)]
cmd/go, cmd/link: allow -buildmode=pie on linux/ppc64le

Change-Id: I0d0abbb9503c3a3c35d846fc0379836b2c483dea
Reviewed-on: https://go-review.googlesource.com/15962
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/internal/obj, cmd/link: handle the fact that a few store/loads on ppc64 are DS...
Michael Hudson-Doyle [Tue, 8 Sep 2015 03:21:58 +0000 (15:21 +1200)]
cmd/internal/obj, cmd/link: handle the fact that a few store/loads on ppc64 are DS form

Change-Id: I4fe1af48ec1cd8a23e2f7f2a0257dc989ff7aced
Reviewed-on: https://go-review.googlesource.com/14235
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/internal/ld: skip dwarf output if dsymutil no-ops
Tamir Duberstein [Sun, 8 Nov 2015 00:30:25 +0000 (19:30 -0500)]
cmd/internal/ld: skip dwarf output if dsymutil no-ops

Fixes #11994.

Change-Id: Icee6ffa6e3a9d15b68b4ae9b2716d65ecbdba73a
Reviewed-on: https://go-review.googlesource.com/16702
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/asm/internal/lex: format error correctly
Dominik Honnef [Wed, 28 Oct 2015 06:44:35 +0000 (07:44 +0100)]
cmd/asm/internal/lex: format error correctly

Error doesn't take a format string and appends its own newline. Phrase
the error like the other ones.

Change-Id: Ic3af857e5d4890207c74a6eb59a0d1067b503e1b
Reviewed-on: https://go-review.googlesource.com/16420
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

9 years agonet/http: update bundled http2 revision
Brad Fitzpatrick [Sun, 8 Nov 2015 10:45:04 +0000 (11:45 +0100)]
net/http: update bundled http2 revision

Updates to git rev 042ba42f (https://golang.org/cl/16734)

This moves all the code for glueing the HTTP1 and HTTP2 transports
together out of net/http and into x/net/http2 where others can use it,
and where it has tests.

Change-Id: I143ac8bb61eed36c87fd838b682ebb37b81b8c2c
Reviewed-on: https://go-review.googlesource.com/16735
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Mathieu Lonjaret <mathieu.lonjaret@gmail.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agogo/build: introduce go1.6 build tag
Brad Fitzpatrick [Sun, 8 Nov 2015 09:41:20 +0000 (10:41 +0100)]
go/build: introduce go1.6 build tag

This is needed now for subrepos to be able to conditionally use
API symbols found only after Go 1.5.

Change-Id: I91f8a1154e2a74008e8ca79490e3f12847f9c3b2
Reviewed-on: https://go-review.googlesource.com/16733
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Mathieu Lonjaret <mathieu.lonjaret@gmail.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/compile: mark duffzero as using X0, not AX
Keith Randall [Fri, 6 Nov 2015 16:35:55 +0000 (08:35 -0800)]
cmd/compile: mark duffzero as using X0, not AX

duffzero was changed to use X0 instead of AX in
CL 14408.  This was missed as part of that change.

Change-Id: I72fb0114cfbc035b83bfaa8631d27e6740da2652
Reviewed-on: https://go-review.googlesource.com/16717
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Minux Ma <minux@golang.org>
9 years agonet/http: update http2 bundle
Brad Fitzpatrick [Sat, 7 Nov 2015 15:51:20 +0000 (16:51 +0100)]
net/http: update http2 bundle

To rev a179abb (handle Transport PING frames).

Change-Id: I6e1eef2c9586c23f231803d9364d921248722f12
Reviewed-on: https://go-review.googlesource.com/16732
Reviewed-by: Blake Mizerany <blake.mizerany@gmail.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agonet/http: update http2 bundle to rev d62542
Brad Fitzpatrick [Sat, 7 Nov 2015 15:34:03 +0000 (16:34 +0100)]
net/http: update http2 bundle to rev d62542

Updates to use new client pool abstraction.

Change-Id: I3552018038ee8394d313d3253af337b07be211f6
Reviewed-on: https://go-review.googlesource.com/16730
Reviewed-by: Blake Mizerany <blake.mizerany@gmail.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: use pthread_sigmask, not sigprocmask, on Darwin ARM/ARM64
Ian Lance Taylor [Fri, 6 Nov 2015 19:17:20 +0000 (11:17 -0800)]
runtime: use pthread_sigmask, not sigprocmask, on Darwin ARM/ARM64

Other systems use pthread_sigmask.  It was a mistake to use sigprocmask
here.

Change-Id: Ie045aa3f09cf035fcf807b7543b96fa5b847958a
Reviewed-on: https://go-review.googlesource.com/16720
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agonet/http: enable HTTP/2 support in DefaultTransport
Brad Fitzpatrick [Thu, 5 Nov 2015 09:28:28 +0000 (10:28 +0100)]
net/http: enable HTTP/2 support in DefaultTransport

The GODEBUG option remains, for now, but only for turning it off.
We'll decide what to do with it before release.

This CL includes the dependent http2 change (https://golang.org/cl/16692)
in the http2 bundle (h2_bundle.go).

Updates golang/go#6891

Change-Id: If9723ef627c7ba4f7343dc8cb89ca88ef0fbcb10
Reviewed-on: https://go-review.googlesource.com/16693
Reviewed-by: Blake Mizerany <blake.mizerany@gmail.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: memmove/memclr pointers atomically
Keith Randall [Thu, 5 Nov 2015 20:39:56 +0000 (12:39 -0800)]
runtime: memmove/memclr pointers atomically

Make sure that we're moving or zeroing pointers atomically.
Anything that is a multiple of pointer size and at least
pointer aligned might have pointers in it.  All the code looks
ok except for the 1-pointer-sized moves.

Fixes #13160
Update #12552

Change-Id: Ib97d9b918fa9f4cc5c56c67ed90255b7fdfb7b45
Reviewed-on: https://go-review.googlesource.com/16668
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: optimize indexbytebody on amd64
Ilya Tocar [Thu, 29 Oct 2015 15:52:22 +0000 (18:52 +0300)]
runtime: optimize indexbytebody on amd64

Use avx2 to compare 32 bytes per iteration.
Results (haswell):

name                    old time/op    new time/op     delta
IndexByte32-6             15.5ns ± 0%     14.7ns ± 5%   -4.87%        (p=0.000 n=16+20)
IndexByte4K-6              360ns ± 0%      183ns ± 0%  -49.17%        (p=0.000 n=19+20)
IndexByte4M-6              384µs ± 0%      256µs ± 1%  -33.41%        (p=0.000 n=20+20)
IndexByte64M-6            6.20ms ± 0%     4.18ms ± 1%  -32.52%        (p=0.000 n=19+20)
IndexBytePortable32-6     73.4ns ± 5%     75.8ns ± 3%   +3.35%        (p=0.000 n=20+19)
IndexBytePortable4K-6     5.15µs ± 0%     5.15µs ± 0%     ~     (all samples are equal)
IndexBytePortable4M-6     5.26ms ± 0%     5.25ms ± 0%   -0.12%        (p=0.000 n=20+18)
IndexBytePortable64M-6    84.1ms ± 0%     84.1ms ± 0%   -0.08%        (p=0.012 n=18+20)
Index32-6                  352ns ± 0%      352ns ± 0%     ~     (all samples are equal)
Index4K-6                 53.8µs ± 0%     53.8µs ± 0%   -0.03%        (p=0.000 n=16+18)
Index4M-6                 55.4ms ± 0%     55.4ms ± 0%     ~           (p=0.149 n=20+19)
Index64M-6                 886ms ± 0%      886ms ± 0%     ~           (p=0.108 n=20+20)
IndexEasy32-6             80.3ns ± 0%     80.1ns ± 0%   -0.21%        (p=0.000 n=20+20)
IndexEasy4K-6              426ns ± 0%      215ns ± 0%  -49.53%        (p=0.000 n=20+20)
IndexEasy4M-6              388µs ± 0%      262µs ± 1%  -32.42%        (p=0.000 n=18+20)
IndexEasy64M-6            6.20ms ± 0%     4.19ms ± 1%  -32.47%        (p=0.000 n=18+20)

name                    old speed      new speed       delta
IndexByte32-6           2.06GB/s ± 1%   2.17GB/s ± 5%   +5.19%        (p=0.000 n=18+20)
IndexByte4K-6           11.4GB/s ± 0%   22.3GB/s ± 0%  +96.45%        (p=0.000 n=17+20)
IndexByte4M-6           10.9GB/s ± 0%   16.4GB/s ± 1%  +50.17%        (p=0.000 n=20+20)
IndexByte64M-6          10.8GB/s ± 0%   16.0GB/s ± 1%  +48.19%        (p=0.000 n=19+20)
IndexBytePortable32-6    436MB/s ± 5%    422MB/s ± 3%   -3.27%        (p=0.000 n=20+19)
IndexBytePortable4K-6    795MB/s ± 0%    795MB/s ± 0%     ~           (p=0.940 n=17+18)
IndexBytePortable4M-6    798MB/s ± 0%    799MB/s ± 0%   +0.12%        (p=0.000 n=20+18)
IndexBytePortable64M-6   798MB/s ± 0%    798MB/s ± 0%   +0.08%        (p=0.011 n=18+20)
Index32-6               90.9MB/s ± 0%   90.9MB/s ± 0%   -0.00%        (p=0.025 n=20+20)
Index4K-6               76.1MB/s ± 0%   76.1MB/s ± 0%   +0.03%        (p=0.000 n=14+15)
Index4M-6               75.7MB/s ± 0%   75.7MB/s ± 0%     ~           (p=0.076 n=20+19)
Index64M-6              75.7MB/s ± 0%   75.7MB/s ± 0%     ~           (p=0.456 n=20+17)
IndexEasy32-6            399MB/s ± 0%    399MB/s ± 0%   +0.20%        (p=0.000 n=20+19)
IndexEasy4K-6           9.60GB/s ± 0%  19.02GB/s ± 0%  +98.19%        (p=0.000 n=20+20)
IndexEasy4M-6           10.8GB/s ± 0%   16.0GB/s ± 1%  +47.98%        (p=0.000 n=18+20)
IndexEasy64M-6          10.8GB/s ± 0%   16.0GB/s ± 1%  +48.08%        (p=0.000 n=18+20)

Change-Id: I46075921dde9f3580a89544c0b3a2d8c9181ebc4
Reviewed-on: https://go-review.googlesource.com/16484
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
Reviewed-by: Klaus Post <klauspost@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: teach peephole optimizer that duffcopy clobbers X0
Keith Randall [Fri, 6 Nov 2015 08:06:52 +0000 (00:06 -0800)]
runtime: teach peephole optimizer that duffcopy clobbers X0

Duffcopy now uses X0, as of 5cf281a.  Teach the peephole
optimizer that duffcopy clobbers X0 so that it does not
rename registers use X0 across the duffcopy instruction.

Fixes #13171

Change-Id: I389cbf1982cb6eb2f51e6152ac96736a8589f085
Reviewed-on: https://go-review.googlesource.com/16715
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Minux Ma <minux@golang.org>
Reviewed-by: Ilya Tocar <ilya.tocar@intel.com>
9 years agodoc: updated go1.6 with reflect change for unexported embedded structs
Marcel van Lohuizen [Mon, 26 Oct 2015 10:21:21 +0000 (11:21 +0100)]
doc: updated go1.6 with reflect change for unexported embedded structs

Change-Id: I53c196925fb86784b31dea799c27e79574d35fcc
Reviewed-on: https://go-review.googlesource.com/16304
Reviewed-by: Marcel van Lohuizen <mpvl@golang.org>
9 years agoRevert "runtime: simplify buffered channels."
Keith Randall [Fri, 6 Nov 2015 08:29:53 +0000 (08:29 +0000)]
Revert "runtime: simplify buffered channels."

Revert for now until #13169 is understood.

This reverts commit 8e496f1d6923172291658f0a785bdb47cc152325.

Change-Id: Ib3eb2588824ef47a2b6eb9e377a24e5c817fcc81
Reviewed-on: https://go-review.googlesource.com/16716
Reviewed-by: Keith Randall <khr@golang.org>
9 years agocmd/link: don't warn about unnamed symbols in .debug_str section
Ian Lance Taylor [Thu, 5 Nov 2015 02:20:57 +0000 (18:20 -0800)]
cmd/link: don't warn about unnamed symbols in .debug_str section

They reportedly occur with LLVM 3.7 on FreeBSD ARM.

Fixes #13139.

Change-Id: Ia7d053a8662696b1984e81fbd1d908c951c35a98
Reviewed-on: https://go-review.googlesource.com/16667
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Minux Ma <minux@golang.org>
9 years agoarchive/tar: detect truncated files
Joe Tsai [Thu, 1 Oct 2015 09:30:29 +0000 (02:30 -0700)]
archive/tar: detect truncated files

Motivation:
* Reader.skipUnread never reports io.ErrUnexpectedEOF. This is strange
given that io.ErrUnexpectedEOF is given through Reader.Read if the
user manually reads the file.
* Reader.skipUnread fails to detect truncated files since io.Seeker
is lazy about reporting errors. Thus, the behavior of Reader differs
whether the input io.Reader also satisfies io.Seeker or not.

To solve this, we seek to one before the end of the data section and
always rely on at least one call to io.CopyN. If the tr.r satisfies
io.Seeker, this is guarunteed to never read more than blockSize.

Fixes #12557

Change-Id: I0ddddfc6bed0d74465cb7e7a02b26f1de7a7a279
Reviewed-on: https://go-review.googlesource.com/15175
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/go: skip TestBuildOutputToDevNull on Plan 9
David du Colombier [Thu, 5 Nov 2015 09:20:32 +0000 (10:20 +0100)]
cmd/go: skip TestBuildOutputToDevNull on Plan 9

TestBuildOutputToDevNull was added in CL 16585.
However, copying to /dev/null couldn't work on Plan 9,
because /dev/null is a regular file. Since it's not
different from any other file, the logic in copyFile
couldn't distinguish it from another, already existing,
file, that we wouldn't want to overwrite.

Change-Id: Ie8d353f318fedfc7cfb9541fed00a2397e232592
Reviewed-on: https://go-review.googlesource.com/16691
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: David du Colombier <0intro@gmail.com>

9 years agocmd/internal/obj/ppc64: fix assembly of SRADCC with immediate
Michael Hudson-Doyle [Thu, 5 Nov 2015 21:10:07 +0000 (10:10 +1300)]
cmd/internal/obj/ppc64: fix assembly of SRADCC with immediate

sradi and sradi. hide the top bit of their immediate argument apart from the
rest of it, but the code only handled the sradi case.

I'm pretty sure this is the only instruction missing (a couple of the rotate
instructions encode their immediate the same way but their handling looks OK).

This fixes the failure of "GOARCH=amd64 ~/go/bin/go install -v runtime" as
reported in the bug.

Fixes #11987

Change-Id: I0cdefcd7a04e0e8fce45827e7054ffde9a83f589
Reviewed-on: https://go-review.googlesource.com/16710
Reviewed-by: Minux Ma <minux@golang.org>
9 years agocmd/go: skip TestGoGenerateEnv on Plan 9
David du Colombier [Thu, 5 Nov 2015 08:41:55 +0000 (09:41 +0100)]
cmd/go: skip TestGoGenerateEnv on Plan 9

TestGoGenerateEnv was added in CL 16537.
However, Plan 9 doesn't have the env command.

Change-Id: I5f0c937a1b9b456dcea41ceac7865112f2f65c45
Reviewed-on: https://go-review.googlesource.com/16690
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: David du Colombier <0intro@gmail.com>

9 years agoruntime: remove background GC goroutine and mark barriers
Austin Clements [Tue, 27 Oct 2015 21:42:57 +0000 (17:42 -0400)]
runtime: remove background GC goroutine and mark barriers

These are now unused.

Updates #11970.

Change-Id: I43e5c4e5bcda9581bacc63364f96bb4855ab779f
Reviewed-on: https://go-review.googlesource.com/16393
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: remove GC start up/shutdown workaround in mallocgc
Austin Clements [Tue, 27 Oct 2015 21:34:11 +0000 (17:34 -0400)]
runtime: remove GC start up/shutdown workaround in mallocgc

Currently mallocgc detects if the GC is in a state where it can't
assist, but also can't allocate uncontrolled and yields to help out
the GC. This was a workaround for periods when we were trying to
schedule the GC coordinator. It is no longer necessary because there
is no GC coordinator and malloc can always assist with any GC
transitions that are necessary.

Updates #11970.

Change-Id: I4f7beb7013e85e50ae99a3a8b0bb708ba49cbcd4
Reviewed-on: https://go-review.googlesource.com/16392
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: decentralize mark done and mark termination
Austin Clements [Mon, 26 Oct 2015 15:27:37 +0000 (11:27 -0400)]
runtime: decentralize mark done and mark termination

This moves all of the mark 1 to mark 2 transition and mark termination
to the mark done transition function. This means these transitions are
now handled on the goroutine that detected mark completion. This also
means that the GC coordinator and the background completion barriers
are no longer used and various workarounds to yield to the coordinator
are no longer necessary. These will be removed in follow-up commits.

One consequence of this is that mark workers now need to be
preemptible when performing the mark done transition. This allows them
to stop the world and to perform the final clean-up steps of GC after
restarting the world. They are only made preemptible while performing
this transition, so if the worker findRunnableGCWorker would schedule
isn't available, we didn't want to schedule it anyway.

Fixes #11970.

Change-Id: I9203a2d6287eeff62d589ec02ad9cb1e29ddb837
Reviewed-on: https://go-review.googlesource.com/16391
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: account mark worker time before gcMarkDone
Austin Clements [Mon, 26 Oct 2015 20:48:36 +0000 (16:48 -0400)]
runtime: account mark worker time before gcMarkDone

Currently gcMarkDone takes basically no time, so it's okay to account
the worker time after calling it. However, gcMarkDone is about to take
potentially *much* longer because it may perform all of mark
termination. Prepare for this by swapping the order so we account the
time before calling gcMarkDone.

Change-Id: I90c7df68192acfc4fd02a7254dae739dda4e2fcb
Reviewed-on: https://go-review.googlesource.com/16390
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: factor mark done transition
Austin Clements [Sun, 25 Oct 2015 01:30:59 +0000 (21:30 -0400)]
runtime: factor mark done transition

Currently the code for completion of mark 1/mark 2 is duplicated in
background workers and assists. Factor this in to a single function
that will serve as the transition function for concurrent mark.

Change-Id: I4d9f697a15da0d349db3b34d56f3a220dd41d41b
Reviewed-on: https://go-review.googlesource.com/16359
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: eliminate mark completion in scheduler
Austin Clements [Fri, 23 Oct 2015 21:44:11 +0000 (17:44 -0400)]
runtime: eliminate mark completion in scheduler

Currently, findRunnableGCWorker will perform mark completion if there
is no remaining work and no running workers. This used to be necessary
to resolve a race in the transition from mark 1 to mark 2 where we
would enter mark 2 with no mark work (and no dedicated workers), so no
workers would run, so no worker would signal mark completion.

However, we're about to make mark completion also perform the entire
follow-on process, which includes mark termination. We really don't
want to do that in the scheduler if it happens to detect completion.

Conveniently, this hack is no longer necessary because we always
enqueue root scanning work at the beginning of both mark 1 and mark 2,
so a mark worker will always run. Hence, we can simply eliminate it.

Change-Id: I3fc8f27c8da632f0fb732c9f6425e1f457f5652e
Reviewed-on: https://go-review.googlesource.com/16358
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: don't start idle mark workers when barriers are cleared
Austin Clements [Wed, 4 Nov 2015 00:47:07 +0000 (19:47 -0500)]
runtime: don't start idle mark workers when barriers are cleared

Currently, we don't start dedicated or fractional mark workers unless
the mark 1 or mark 2 barriers have been cleared. One intended
consequence of this is that no background workers run between the
forEachP that disposes all gcWork caches and the beginning of mark 2.

However, we (unintentionally) did not apply this restriction to idle
mark workers. As a result, these can start in the interim between mark
1 completion and mark 2 starting. This explains why it was necessary
to reset the root marking jobs using carefully ordered atomic writes
when setting up mark 2. It also means that, even though we definitely
enqueue work before starting mark 2, it may be drained by the time we
reset the mark 2 barrier. If this happens, currently the only thing
preventing the runtime from deadlocking is that the scheduler itself
also checks for mark completion and will signal mark 2 completion.
Were it not for the odd behavior of idle workers, this check in the
scheduler would not be necessary.

Clean all of this up and prepare to remove this check in the scheduler
by applying the same restriction to starting idle mark workers.

Change-Id: Ic1b479e1591bd7773dc27b320ca399a215603b5a
Reviewed-on: https://go-review.googlesource.com/16631
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: decentralize sweep termination and mark transition
Austin Clements [Fri, 23 Oct 2015 19:55:03 +0000 (15:55 -0400)]
runtime: decentralize sweep termination and mark transition

This moves all of GC initialization, sweep termination, and the
transition to concurrent marking in to the off->mark transition
function. This means it's now handled on the goroutine that detected
the state exit condition.

As a result, malloc no longer needs to Gosched() at the beginning of
the GC cycle to prevent over-allocation while the GC is starting up
because it will now *help* the GC to start up. The Gosched hack is
still necessary during GC shutdown (this is easy to test by enabling
gctrace and hitting Ctrl-S to block the gctrace output).

At this point, the GC coordinator still handles later phases. This
requires a small tweak to how we start the GC coordinator. Currently,
starting the GC coordinator is best-effort and may fail if the
coordinator is about to park from the previous cycle but hasn't yet.
We fix this by replacing the park/ready to wake up the coordinator
with a semaphore. This is temporary since the coordinator will be
going away in a few commits.

Updates #11970.

Change-Id: I2c6a11c91e72dfbc59c2d8e7c66146dee9a444fe
Reviewed-on: https://go-review.googlesource.com/16357
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: decentralize concurrent sweep termination
Austin Clements [Fri, 23 Oct 2015 19:04:37 +0000 (15:04 -0400)]
runtime: decentralize concurrent sweep termination

This moves concurrent sweep termination from the coordinator to the
off->mark transition. This allows it to be performed by all Gs
attempting to start the GC.

Updates #11970.

Change-Id: I24428e8599a759398c2ef7ec996ba755a448f947
Reviewed-on: https://go-review.googlesource.com/16356
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: beginning of decentralized off->mark transition
Austin Clements [Fri, 23 Oct 2015 18:15:18 +0000 (14:15 -0400)]
runtime: beginning of decentralized off->mark transition

This begins the conversion of the centralized GC coordinator to a
decentralized state machine by introducing the internal API that
triggers the first state transition from _GCoff to _GCmark (or
_GCmarktermination).

This change introduces the transition lock, the off->mark transition
condition (which is very similar to shouldtriggergc()), and the
general structure of a state transition. Since we're doing this
conversion in stages, it then falls back to the GC coordinator to
actually execute the cycle. We'll start moving logic out of the GC
coordinator and in to transition functions next.

This fixes a minor bug in gcstoptheworld debug mode where passing the
heap trigger once could trigger multiple STW GCs.

Updates #11970.

Change-Id: I964087dd190a639eb5766398f8e1bbf8b352902f
Reviewed-on: https://go-review.googlesource.com/16355
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

9 years agoruntime: move concurrent mark setup off system stack
Austin Clements [Sun, 25 Oct 2015 01:19:52 +0000 (21:19 -0400)]
runtime: move concurrent mark setup off system stack

For historical reasons we currently do a lot of the concurrent mark
setup on the system stack. In fact, at this point the one and only
thing that needs to happen on the system stack is the start-the-world.

Clean up this code by lifting everything other than the
start-the-world off the system stack.

The diff for this change looks large, but the only code change is to
narrow the systemstack call. Everything else is re-indentation.

Change-Id: I1e03b8afc759fad726f2397b05a17d183c2713ce
Reviewed-on: https://go-review.googlesource.com/16354
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: lift state variables from func gc to var work
Austin Clements [Fri, 23 Oct 2015 19:17:04 +0000 (15:17 -0400)]
runtime: lift state variables from func gc to var work

We're about to split func gc across several functions, so lift the
local variables it uses for tracking statistics and state across the
cycle into the global "work" variable.

Change-Id: Ie955f2f1758c7f5a5543ea1f3f33b222bc4b1d37
Reviewed-on: https://go-review.googlesource.com/16353
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: note a minor issue with GODEUG=gcstoptheworld
Austin Clements [Fri, 23 Oct 2015 17:59:08 +0000 (13:59 -0400)]
runtime: note a minor issue with GODEUG=gcstoptheworld

Change-Id: I91cda8d88b0852cd0f868d33c594206bcca0c386
Reviewed-on: https://go-review.googlesource.com/16352
Reviewed-by: Rick Hudson <rlh@golang.org>
9 years agocmd/go: change ar argument to rc
Ian Lance Taylor [Thu, 5 Nov 2015 01:38:53 +0000 (17:38 -0800)]
cmd/go: change ar argument to rc

Put 'r' first because that is the command, and 'c' is the modifier.
Keep 'c' because it means to not warn when creating an archive.
Drop 'u' because it is unnecessary and fails on Arch Linux.

No test because this is only for gccgo (I tested it manually).

Fixes #12310.

Change-Id: Id740257fb1c347dfaa60f7d613af2897dae2c059
Reviewed-on: https://go-review.googlesource.com/16664
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agoruntime: optimize string comparison on amd64
Ilya Tocar [Thu, 29 Oct 2015 14:17:05 +0000 (17:17 +0300)]
runtime: optimize string comparison on amd64

Use AVX2 if possible.
Results below (haswell):

name                            old time/op    new time/op     delta
CompareStringEqual-6              8.77ns ± 0%     8.63ns ± 1%   -1.58%        (p=0.000 n=20+19)
CompareStringIdentical-6          5.02ns ± 0%     5.02ns ± 0%     ~     (all samples are equal)
CompareStringSameLength-6         7.51ns ± 0%     7.51ns ± 0%     ~     (all samples are equal)
CompareStringDifferentLength-6    1.56ns ± 0%     1.56ns ± 0%     ~     (all samples are equal)
CompareStringBigUnaligned-6        124µs ± 1%      105µs ± 5%  -14.99%        (p=0.000 n=20+18)
CompareStringBig-6                 112µs ± 1%      103µs ± 0%   -7.87%        (p=0.000 n=20+17)

name                            old speed      new speed       delta
CompareStringBigUnaligned-6     8.48GB/s ± 1%   9.98GB/s ± 5%  +17.67%        (p=0.000 n=20+18)
CompareStringBig-6              9.37GB/s ± 1%  10.17GB/s ± 0%   +8.54%        (p=0.000 n=20+17)

Change-Id: I1c949626dd2aaf9f633e3c888a9df71c82eed7e1
Reviewed-on: https://go-review.googlesource.com/16481
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Klaus Post <klauspost@gmail.com>
9 years agoruntime: simplify buffered channels.
Keith Randall [Fri, 20 Mar 2015 21:12:05 +0000 (14:12 -0700)]
runtime: simplify buffered channels.

This change removes the retry mechanism we use for buffered channels.
Instead, any sender waking up a receiver or vice versa completes the
full protocol with its counterpart.  This means the counterpart does
not need to relock the channel when it wakes up.  (Currently
buffered channels need to relock on wakeup.)

For sends on a channel with waiting receivers, this change replaces
two copies (sender->queue, queue->receiver) with one (sender->receiver).
For receives on channels with a waiting sender, two copies are still required.

This change unifies to a large degree the algorithm for buffered
and unbuffered channels, simplifying the overall implementation.

Fixes #11506

benchmark                        old ns/op     new ns/op     delta
BenchmarkChanProdCons10          125           110           -12.00%
BenchmarkChanProdCons0           303           284           -6.27%
BenchmarkChanProdCons100         75.5          71.3          -5.56%
BenchmarkChanContended           6452          6125          -5.07%
BenchmarkChanNonblocking         11.5          11.0          -4.35%
BenchmarkChanCreation            149           143           -4.03%
BenchmarkChanSem                 63.6          61.6          -3.14%
BenchmarkChanUncontended         6390          6212          -2.79%
BenchmarkChanSync                282           276           -2.13%
BenchmarkChanProdConsWork10      516           506           -1.94%
BenchmarkChanProdConsWork0       696           685           -1.58%
BenchmarkChanProdConsWork100     470           469           -0.21%
BenchmarkChanPopular             660427        660012        -0.06%

Change-Id: I164113a56432fbc7cace0786e49c5a6e6a708ea4
Reviewed-on: https://go-review.googlesource.com/9345
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
9 years agocmd/dist: remove vestigial -s flag
Matthew Dempsky [Thu, 5 Nov 2015 02:12:30 +0000 (18:12 -0800)]
cmd/dist: remove vestigial -s flag

Fixes #12002.

Change-Id: I7262f4520560ac158fc2ee3ce1d2f7a488d40354
Reviewed-on: https://go-review.googlesource.com/16666
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Minux Ma <minux@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Dave Cheney <dave@cheney.net>
9 years agocrypto/x509: add /etc/ssl/certs to certificate directories
Shenghou Ma [Thu, 5 Nov 2015 02:20:33 +0000 (21:20 -0500)]
crypto/x509: add /etc/ssl/certs to certificate directories

Fixes #12139.

Change-Id: Ied760ac37e2fc21ef951ae872136dc3bfd49bf9f
Reviewed-on: https://go-review.googlesource.com/16671
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/go: put all generate variables in the environment
Ian Lance Taylor [Mon, 2 Nov 2015 17:13:44 +0000 (09:13 -0800)]
cmd/go: put all generate variables in the environment

Fixes #13124.

Change-Id: I8a824156c84016504d29dc2dd2d522149b189be8
Reviewed-on: https://go-review.googlesource.com/16537
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agointernal/syscall/unix: eliminate non-trivial randomTrap initializer
Matthew Dempsky [Wed, 4 Nov 2015 22:51:51 +0000 (14:51 -0800)]
internal/syscall/unix: eliminate non-trivial randomTrap initializer

While here, enable getrandom on arm64 too (using the value found in
include/uapi/asm-generic/unistd.h, which seems to match up with other
GOARCH=arm64 syscall numbers).

Updates #10848.

Change-Id: I5ab36ccf6ee8d5cc6f0e1a61d09f0da7410288b9
Reviewed-on: https://go-review.googlesource.com/16662
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agoruntime: make putfull start mark workers
Austin Clements [Mon, 26 Oct 2015 21:07:02 +0000 (17:07 -0400)]
runtime: make putfull start mark workers

Currently we depend on the good graces and timing of the scheduler to
get opportunities to start dedicated mark workers. In the worst case,
it may take 10ms to get dedicated mark workers going at the beginning
of mark 1 and mark 2 or after the amount of available work has dropped
and gone back up.

Instead of waiting for the regular preemption logic to get around to
us, make putfull enlist a random P if we're not already running enough
dedicated workers. This should improve performance stability of the
garbage collector and is likely to improve the overall performance
somewhat.

No overall effect on the go1 benchmarks. It speeds up the garbage
benchmark by 12%, which more than counters the performance loss from
the previous commit.

name              old time/op  new time/op  delta
XBenchGarbage-12  6.32ms ± 4%  5.58ms ± 2%  -11.68%  (p=0.000 n=20+16)

name                      old time/op    new time/op    delta
BinaryTree17-12              3.18s ± 5%     3.12s ± 4%  -1.83%  (p=0.021 n=20+20)
Fannkuch11-12                2.50s ± 2%     2.46s ± 2%  -1.57%  (p=0.000 n=18+19)
FmtFprintfEmpty-12          50.8ns ± 3%    50.4ns ± 3%    ~     (p=0.184 n=20+20)
FmtFprintfString-12          167ns ± 2%     171ns ± 1%  +2.46%  (p=0.000 n=20+19)
FmtFprintfInt-12             161ns ± 2%     163ns ± 2%  +1.81%  (p=0.000 n=20+20)
FmtFprintfIntInt-12          269ns ± 1%     266ns ± 1%  -0.81%  (p=0.002 n=19+20)
FmtFprintfPrefixedInt-12     237ns ± 2%     231ns ± 2%  -2.86%  (p=0.000 n=20+20)
FmtFprintfFloat-12           313ns ± 2%     313ns ± 1%    ~     (p=0.681 n=20+20)
FmtManyArgs-12              1.05µs ± 2%    1.03µs ± 1%  -2.26%  (p=0.000 n=20+20)
GobDecode-12                8.66ms ± 1%    8.67ms ± 1%    ~     (p=0.380 n=19+20)
GobEncode-12                6.56ms ± 1%    6.56ms ± 2%    ~     (p=0.607 n=19+20)
Gzip-12                      317ms ± 1%     314ms ± 2%  -1.10%  (p=0.000 n=20+19)
Gunzip-12                   42.1ms ± 1%    42.2ms ± 1%  +0.27%  (p=0.044 n=20+19)
HTTPClientServer-12         62.7µs ± 1%    62.0µs ± 1%  -1.04%  (p=0.000 n=19+18)
JSONEncode-12               16.7ms ± 1%    16.8ms ± 2%  +0.59%  (p=0.021 n=20+20)
JSONDecode-12               58.2ms ± 1%    61.4ms ± 2%  +5.43%  (p=0.000 n=18+19)
Mandelbrot200-12            3.84ms ± 1%    3.87ms ± 2%  +0.79%  (p=0.008 n=18+20)
GoParse-12                  3.86ms ± 2%    3.76ms ± 2%  -2.60%  (p=0.000 n=20+20)
RegexpMatchEasy0_32-12       100ns ± 2%     100ns ± 1%  -0.68%  (p=0.005 n=18+15)
RegexpMatchEasy0_1K-12       332ns ± 1%     342ns ± 1%  +3.16%  (p=0.000 n=19+19)
RegexpMatchEasy1_32-12      82.9ns ± 3%    83.0ns ± 2%    ~     (p=0.906 n=19+20)
RegexpMatchEasy1_1K-12       487ns ± 1%     494ns ± 1%  +1.50%  (p=0.000 n=17+20)
RegexpMatchMedium_32-12      131ns ± 2%     130ns ± 1%    ~     (p=0.686 n=19+20)
RegexpMatchMedium_1K-12     39.6µs ± 1%    39.2µs ± 1%  -1.09%  (p=0.000 n=18+19)
RegexpMatchHard_32-12       2.04µs ± 1%    2.04µs ± 2%    ~     (p=0.804 n=20+20)
RegexpMatchHard_1K-12       61.7µs ± 2%    61.3µs ± 2%    ~     (p=0.052 n=18+20)
Revcomp-12                   529ms ± 2%     533ms ± 1%  +0.83%  (p=0.003 n=20+19)
Template-12                 70.7ms ± 2%    71.0ms ± 2%    ~     (p=0.065 n=20+19)
TimeParse-12                 351ns ± 2%     355ns ± 1%  +1.25%  (p=0.000 n=19+20)
TimeFormat-12                362ns ± 2%     373ns ± 1%  +2.83%  (p=0.000 n=18+20)
[Geo mean]                  62.2µs         62.3µs       +0.13%

name                      old speed      new speed      delta
GobDecode-12              88.6MB/s ± 1%  88.5MB/s ± 1%    ~     (p=0.392 n=19+20)
GobEncode-12               117MB/s ± 1%   117MB/s ± 1%    ~     (p=0.622 n=19+20)
Gzip-12                   61.1MB/s ± 1%  61.8MB/s ± 2%  +1.11%  (p=0.000 n=20+19)
Gunzip-12                  461MB/s ± 1%   460MB/s ± 1%  -0.27%  (p=0.044 n=20+19)
JSONEncode-12              116MB/s ± 1%   115MB/s ± 2%  -0.58%  (p=0.022 n=20+20)
JSONDecode-12             33.3MB/s ± 1%  31.6MB/s ± 2%  -5.15%  (p=0.000 n=18+19)
GoParse-12                15.0MB/s ± 2%  15.4MB/s ± 2%  +2.66%  (p=0.000 n=20+20)
RegexpMatchEasy0_32-12     317MB/s ± 2%   319MB/s ± 2%    ~     (p=0.052 n=20+20)
RegexpMatchEasy0_1K-12    3.08GB/s ± 1%  2.99GB/s ± 1%  -3.07%  (p=0.000 n=19+19)
RegexpMatchEasy1_32-12     386MB/s ± 3%   386MB/s ± 2%    ~     (p=0.939 n=19+20)
RegexpMatchEasy1_1K-12    2.10GB/s ± 1%  2.07GB/s ± 1%  -1.46%  (p=0.000 n=17+20)
RegexpMatchMedium_32-12   7.62MB/s ± 2%  7.64MB/s ± 1%    ~     (p=0.702 n=19+20)
RegexpMatchMedium_1K-12   25.9MB/s ± 1%  26.1MB/s ± 2%  +0.99%  (p=0.000 n=18+20)
RegexpMatchHard_32-12     15.7MB/s ± 1%  15.7MB/s ± 2%    ~     (p=0.723 n=20+20)
RegexpMatchHard_1K-12     16.6MB/s ± 2%  16.7MB/s ± 2%    ~     (p=0.052 n=18+20)
Revcomp-12                 481MB/s ± 2%   477MB/s ± 1%  -0.83%  (p=0.003 n=20+19)
Template-12               27.5MB/s ± 2%  27.3MB/s ± 2%    ~     (p=0.062 n=20+19)
[Geo mean]                99.4MB/s       99.1MB/s       -0.35%

Change-Id: I914d8cadded5a230509d118164a4c201601afc06
Reviewed-on: https://go-review.googlesource.com/16298
Reviewed-by: Rick Hudson <rlh@golang.org>
9 years agoruntime: eliminate getfull barrier from concurrent mark
Austin Clements [Mon, 26 Oct 2015 20:29:25 +0000 (16:29 -0400)]
runtime: eliminate getfull barrier from concurrent mark

Currently dedicated mark workers participate in the getfull barrier
during concurrent mark. However, the getfull barrier wasn't designed
for concurrent work and this causes no end of headaches.

In the concurrent setting, participants come and go. This makes mark
completion susceptible to live-lock: since dedicated workers are only
periodically polling for completion, it's possible for the program to
be in some transient worker each time one of the dedicated workers
wakes up to check if it can exit the getfull barrier. It also
complicates reasoning about the system because dedicated workers
participate directly in the getfull barrier, but transient workers
must instead use trygetfull because they have exit conditions that
aren't captured by getfull (e.g., fractional workers exit when
preempted). The complexity of implementing these exit conditions
contributed to #11677. Furthermore, the getfull barrier is inefficient
because we could be running user code instead of spinning on a P. In
effect, we're dedicating 25% of the CPU to marking even if that means
we have to spin to make that 25%. It also causes issues on Windows
because we can't actually sleep for 100µs (#8687).

Fix this by making dedicated workers no longer participate in the
getfull barrier. Instead, dedicated workers simply return to the
scheduler when they fail to get more work, regardless of what others
workers are doing, and the scheduler only starts new dedicated workers
if there's work available. Everything that needs to be handled by this
barrier is already handled by detection of mark completion.

This makes the system much more symmetric because all workers and
assists now use trygetfull during concurrent mark. It also loosens the
25% CPU target so that we can give some of that 25% back to user code
if there isn't enough work to keep the mark worker busy. And it
eliminates the problematic 100µs sleep on Windows during concurrent
mark (though not during mark termination).

The downside of this is that if we hit a bottleneck in the heap graph
that then expands back out, the system may shut down dedicated workers
and take a while to start them back up. We'll address this in the next
commit.

Updates #12041 and #8687.

No effect on the go1 benchmarks. This slows down the garbage benchmark
by 9%, but we'll more than make it up in the next commit.

name              old time/op  new time/op  delta
XBenchGarbage-12  5.80ms ± 2%  6.32ms ± 4%  +9.03%  (p=0.000 n=20+20)

Change-Id: I65100a9ba005a8b5cf97940798918672ea9dd09b
Reviewed-on: https://go-review.googlesource.com/16297
Reviewed-by: Rick Hudson <rlh@golang.org>
9 years agomisc/ios: keep whole buffer in go_darwin_arm_exec
David Crawshaw [Wed, 4 Nov 2015 16:21:55 +0000 (11:21 -0500)]
misc/ios: keep whole buffer in go_darwin_arm_exec

The existing go_darwin_arm_exec.go script does not work with Xcode 7,
not due to any significant changes, but just ordering and timing of
statements from lldb. Unfortunately the current design of
go_darwin_arm_exec.go makes it not obvious what gets stuck where, so
this moves from a moving buffer window to a complete buffer of the
lldb output.

The result is easier code to follow, and it works with Xcode 7.

Updates #12660.

Change-Id: I3b8b890b0bf4474119482e95d84e821a86d1eaed
Reviewed-on: https://go-review.googlesource.com/16634
Reviewed-by: Michael Matloob <matloob@golang.org>
9 years agonet/http: register HTTP/2 before listening in ListenAndServe
Brad Fitzpatrick [Tue, 3 Nov 2015 19:00:37 +0000 (11:00 -0800)]
net/http: register HTTP/2 before listening in ListenAndServe

Change-Id: Icf9b6802945051aa484fb9ebcce71704f5655474
Reviewed-on: https://go-review.googlesource.com/16630
Reviewed-by: Andrew Gerrand <adg@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agosyscall: allow nacl's fake network code to Listen twice on the same address
Brad Fitzpatrick [Wed, 4 Nov 2015 07:17:57 +0000 (02:17 -0500)]
syscall: allow nacl's fake network code to Listen twice on the same address

Noticed from nacl trybot failures on new tests in
https://golang.org/cl/16630

Related earlier fix of mine to nacl's listen code:

  syscall: fix nacl listener to not accept connections once closed
  https://go-review.googlesource.com/15940

Perhaps a better fix (in the future?) would be to remove the listener
from the map at close, but that didn't seem entirely straightforward
last time I looked into it. It's not my code, but it seems that the
map entry continues to have a purpose even after Listener close. (?)

But given that this code is only really used for running tests and the
playground, this seems fine.

Change-Id: I43bfedc57c07f215f4d79c18f588d3650687a48f
Reviewed-on: https://go-review.googlesource.com/16650
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/vet: use testenv
David Crawshaw [Wed, 4 Nov 2015 15:41:05 +0000 (10:41 -0500)]
cmd/vet: use testenv

Fix for iOS builder.

Change-Id: I5b6c977b187446c848182a9294d5bed6b5f9f6e4
Reviewed-on: https://go-review.googlesource.com/16633
Run-TryBot: David Crawshaw <crawshaw@golang.org>
Reviewed-by: Hyang-Ah Hana Kim <hyangah@gmail.com>
9 years agocrypto/md5: uniform Write func
unknown [Wed, 4 Nov 2015 12:55:21 +0000 (15:55 +0300)]
crypto/md5: uniform Write func

Unification of implementation of existing md5.Write function
with other implementations (sha1, sha256, sha512).

Change-Id: I58ae02d165b17fc221953a5b4b986048b46c0508
Reviewed-on: https://go-review.googlesource.com/16621
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/go: check if destination is a regular file
Mohit Agarwal [Mon, 2 Nov 2015 18:34:46 +0000 (00:04 +0530)]
cmd/go: check if destination is a regular file

builder.copyFile ensures that the destination is an object file.  This
wouldn't be true if we are not writing to a regular file and the copy
fails.  Check if the destination is an object file only if we are
writing to a regular file.  While removing the file, ensure that it is a
regular file so that device files and such aren't removed when running
as a user with suggicient privileges.

Fixes #12407

Change-Id: Ie86ce9770fa59aa56fc486a5962287859b69db3d
Reviewed-on: https://go-review.googlesource.com/16585
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: add go:nowritebarrierrec annotation
Austin Clements [Mon, 2 Nov 2015 21:45:07 +0000 (16:45 -0500)]
cmd/compile: add go:nowritebarrierrec annotation

This introduces a recursive variant of the go:nowritebarrier
annotation that prohibits write barriers not only in the annotated
function, but in all functions it calls, recursively. The error
message gives the shortest call stack from the annotated function to
the function containing the prohibited write barrier, including the
names of the functions and the line numbers of the calls.

To demonstrate the annotation, we apply it to gcmarkwb_m, the write
barrier itself.

This is a new annotation rather than a modification of the existing
go:nowritebarrier annotation because, for better or worse, there are
many go:nowritebarrier functions that do call functions with write
barriers. In most of these cases this is benign because the annotation
was conservative, but it prohibits simply coopting the existing
annotation.

Change-Id: I225ca483c8f699e8436373ed96349e80ca2c2479
Reviewed-on: https://go-review.googlesource.com/16554
Reviewed-by: Keith Randall <khr@golang.org>
9 years agocmd/cgo: add a missing newline in writeExports
Ian Lance Taylor [Wed, 4 Nov 2015 01:18:42 +0000 (17:18 -0800)]
cmd/cgo: add a missing newline in writeExports

The code works without the newline, but it looks funny:

func _cgoexp_15afe6549f62_GoFn(a unsafe.Pointer, n int32) { fn := GoFn

This adds a newline after the '{'.

Change-Id: I6c465abe16f47924426d1b22b91004b3a3586ebd
Reviewed-on: https://go-review.googlesource.com/16612
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Minux Ma <minux@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agonet/http: don't panic after request if Handler sets Request.Body to nil
Brad Fitzpatrick [Tue, 3 Nov 2015 20:04:20 +0000 (12:04 -0800)]
net/http: don't panic after request if Handler sets Request.Body to nil

The Server's server goroutine was panicing (but recovering) when
cleaning up after handling a request. It was pretty harmless (it just
closed that one connection and didn't kill the whole process) but it
was distracting.

Updates #13135

Change-Id: I2a0ce9e8b52c8d364e3f4ce245e05c6f8d62df14
Reviewed-on: https://go-review.googlesource.com/16572
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agocmd/compile: make sure instrumented call has type width
Ian Lance Taylor [Mon, 26 Oct 2015 20:58:23 +0000 (13:58 -0700)]
cmd/compile: make sure instrumented call has type width

The width of the type of an external variable defined with a type
literal may not be set when the instrumentation pass is run.  There are
two cases in the standard library that fail without the call to dowidth:

../../../src/encoding/base32/base32.go:322: constant -1000000000 overflows uintptr
../../../src/encoding/base32/base32.go:329: constant -1000000000 overflows uintptr
../../../src/encoding/json/encode.go:385: constant -1000000000 overflows uintptr
../../../src/encoding/json/encode.go:387: constant -1000000000 overflows uintptr

Change-Id: I7c3334f7decdb7488595ffe4090cd262d7334283
Reviewed-on: https://go-review.googlesource.com/16331
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agomisc/cgo/testsanitizers: fix test of whether compiler option works
Ian Lance Taylor [Tue, 3 Nov 2015 15:14:22 +0000 (07:14 -0800)]
misc/cgo/testsanitizers: fix test of whether compiler option works

On older versions of GCC we need to pass a file name before GCC will
report an unrecognized option.

Fixes #13065.

Change-Id: I7ed34c01a006966a446059025f7d10235c649072
Reviewed-on: https://go-review.googlesource.com/16589
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agoruntime: remove dead code
Dmitry Vyukov [Tue, 3 Nov 2015 19:00:21 +0000 (20:00 +0100)]
runtime: remove dead code

runtime.free has long gone.

Change-Id: I058f69e6481b8fa008e1951c29724731a8a3d081
Reviewed-on: https://go-review.googlesource.com/16593
Reviewed-by: Austin Clements <austin@google.com>
Run-TryBot: Austin Clements <austin@google.com>

9 years agoruntime: cache two workbufs to reduce contention
Austin Clements [Fri, 16 Oct 2015 13:53:11 +0000 (09:53 -0400)]
runtime: cache two workbufs to reduce contention

Currently the gcWork abstraction caches a single work buffer. As a
result, if a worker is putting and getting pointers right at the
boundary of a work buffer, it can flap between work buffers and
(potentially significantly) increase contention on the global work
buffer lists.

This change modifies gcWork to instead cache two work buffers and
switch off between them. This introduces one buffers' worth of
hysteresis and eliminates the above performance worst case by
amortizing the cost of getting or putting a work buffer over at least
one buffers' worth of work.

In practice, it's difficult to trigger this worst case with reasonably
large work buffers. On the garbage benchmark, this reduces the max
writes/sec to the global work list from 32K to 25K and the median from
6K to 5K. However, if a workload were to trigger this worst case
behavior, it could significantly drive up this contention.

This has negligible effects on the go1 benchmarks and slightly speeds
up the garbage benchmark.

name              old time/op  new time/op  delta
XBenchGarbage-12  5.90ms ± 3%  5.83ms ± 4%  -1.18%  (p=0.011 n=18+18)

name                      old time/op    new time/op    delta
BinaryTree17-12              3.22s ± 4%     3.17s ± 3%  -1.57%  (p=0.009 n=19+20)
Fannkuch11-12                2.44s ± 1%     2.53s ± 4%  +3.78%  (p=0.000 n=18+19)
FmtFprintfEmpty-12          50.2ns ± 2%    50.5ns ± 5%    ~     (p=0.631 n=19+20)
FmtFprintfString-12          167ns ± 1%     166ns ± 1%    ~     (p=0.141 n=20+20)
FmtFprintfInt-12             162ns ± 1%     159ns ± 1%  -1.80%  (p=0.000 n=20+20)
FmtFprintfIntInt-12          277ns ± 2%     263ns ± 1%  -4.78%  (p=0.000 n=20+18)
FmtFprintfPrefixedInt-12     240ns ± 1%     232ns ± 2%  -3.25%  (p=0.000 n=20+20)
FmtFprintfFloat-12           311ns ± 1%     315ns ± 2%  +1.17%  (p=0.000 n=20+20)
FmtManyArgs-12              1.05µs ± 2%    1.03µs ± 2%  -1.72%  (p=0.000 n=20+20)
GobDecode-12                8.65ms ± 1%    8.71ms ± 2%  +0.68%  (p=0.001 n=19+20)
GobEncode-12                6.51ms ± 1%    6.54ms ± 1%  +0.42%  (p=0.047 n=20+19)
Gzip-12                      318ms ± 2%     315ms ± 2%  -1.20%  (p=0.000 n=19+19)
Gunzip-12                   42.2ms ± 2%    42.1ms ± 1%    ~     (p=0.667 n=20+19)
HTTPClientServer-12         62.5µs ± 1%    62.4µs ± 1%    ~     (p=0.110 n=20+18)
JSONEncode-12               16.8ms ± 1%    16.8ms ± 2%    ~     (p=0.569 n=19+20)
JSONDecode-12               60.8ms ± 2%    59.8ms ± 1%  -1.69%  (p=0.000 n=19+19)
Mandelbrot200-12            3.87ms ± 1%    3.85ms ± 0%  -0.61%  (p=0.001 n=20+17)
GoParse-12                  3.76ms ± 2%    3.76ms ± 1%    ~     (p=0.698 n=20+20)
RegexpMatchEasy0_32-12       100ns ± 2%     101ns ± 2%    ~     (p=0.065 n=19+20)
RegexpMatchEasy0_1K-12       342ns ± 2%     333ns ± 1%  -2.82%  (p=0.000 n=20+19)
RegexpMatchEasy1_32-12      83.3ns ± 2%    83.2ns ± 2%    ~     (p=0.692 n=20+19)
RegexpMatchEasy1_1K-12       498ns ± 2%     490ns ± 1%  -1.52%  (p=0.000 n=18+20)
RegexpMatchMedium_32-12      131ns ± 2%     131ns ± 2%    ~     (p=0.464 n=20+18)
RegexpMatchMedium_1K-12     39.3µs ± 2%    39.6µs ± 1%  +0.77%  (p=0.000 n=18+19)
RegexpMatchHard_32-12       2.04µs ± 2%    2.06µs ± 1%  +0.69%  (p=0.009 n=19+20)
RegexpMatchHard_1K-12       61.4µs ± 2%    62.1µs ± 1%  +1.21%  (p=0.000 n=19+20)
Revcomp-12                   534ms ± 1%     529ms ± 1%  -0.97%  (p=0.000 n=19+16)
Template-12                 70.4ms ± 2%    70.0ms ± 1%    ~     (p=0.070 n=19+19)
TimeParse-12                 359ns ± 3%     344ns ± 1%  -4.15%  (p=0.000 n=19+19)
TimeFormat-12                357ns ± 1%     361ns ± 2%  +1.05%  (p=0.002 n=20+20)
[Geo mean]                  62.4µs         62.0µs       -0.56%

name                      old speed      new speed      delta
GobDecode-12              88.7MB/s ± 1%  88.1MB/s ± 2%  -0.68%  (p=0.001 n=19+20)
GobEncode-12               118MB/s ± 1%   117MB/s ± 1%  -0.42%  (p=0.046 n=20+19)
Gzip-12                   60.9MB/s ± 2%  61.7MB/s ± 2%  +1.21%  (p=0.000 n=19+19)
Gunzip-12                  460MB/s ± 2%   461MB/s ± 1%    ~     (p=0.661 n=20+19)
JSONEncode-12              116MB/s ± 1%   115MB/s ± 2%    ~     (p=0.555 n=19+20)
JSONDecode-12             31.9MB/s ± 2%  32.5MB/s ± 1%  +1.72%  (p=0.000 n=19+19)
GoParse-12                15.4MB/s ± 2%  15.4MB/s ± 1%    ~     (p=0.653 n=20+20)
RegexpMatchEasy0_32-12     317MB/s ± 2%   315MB/s ± 2%    ~     (p=0.141 n=19+20)
RegexpMatchEasy0_1K-12    2.99GB/s ± 2%  3.07GB/s ± 1%  +2.86%  (p=0.000 n=20+19)
RegexpMatchEasy1_32-12     384MB/s ± 2%   385MB/s ± 2%    ~     (p=0.672 n=20+19)
RegexpMatchEasy1_1K-12    2.06GB/s ± 2%  2.09GB/s ± 1%  +1.54%  (p=0.000 n=18+20)
RegexpMatchMedium_32-12   7.62MB/s ± 2%  7.63MB/s ± 2%    ~     (p=0.800 n=20+18)
RegexpMatchMedium_1K-12   26.0MB/s ± 1%  25.8MB/s ± 1%  -0.77%  (p=0.000 n=18+19)
RegexpMatchHard_32-12     15.7MB/s ± 2%  15.6MB/s ± 1%  -0.69%  (p=0.010 n=19+20)
RegexpMatchHard_1K-12     16.7MB/s ± 2%  16.5MB/s ± 1%  -1.19%  (p=0.000 n=19+20)
Revcomp-12                 476MB/s ± 1%   481MB/s ± 1%  +0.97%  (p=0.000 n=19+16)
Template-12               27.6MB/s ± 2%  27.7MB/s ± 1%    ~     (p=0.071 n=19+19)
[Geo mean]                99.1MB/s       99.3MB/s       +0.27%

Change-Id: I68bcbf74ccb716cd5e844a554f67b679135105e6
Reviewed-on: https://go-review.googlesource.com/16042
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: fix finalization and profiling of tiny allocations
Dmitry Vyukov [Tue, 3 Nov 2015 11:19:15 +0000 (12:19 +0100)]
runtime: fix finalization and profiling of tiny allocations

Handling of special records for tiny allocations has two problems:
1. Once we queue a finalizer we mark the object. As the result any
   subsequent finalizers for the same object will not be queued
   during this GC cycle. If we have 16 finalizers setup (the worst case),
   finalization will take 16 GC cycles. This is what caused misbehave
   of tinyfin.go. The actual flakiness was caused by the fact that fing
   is asynchronous and don't always run before the check.
2. If a tiny block has both finalizer and profile specials,
   it is possible that we both queue finalizer, preserve the object live
   and free the profile record. As the result heap profile can be skewed.

Fix both issues by analyzing all special records for a single object at once.

Also, make tinyfin test stricter and remove reliance on real time.

Also, add a test for the problem 2. Currently heap profile missed about
a half of live memory.

Fixes #13100

Change-Id: I9ae4dc1c44893724138a4565ca5cae29f2e97544
Reviewed-on: https://go-review.googlesource.com/16591
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Dmitry Vyukov <dvyukov@google.com>

9 years agostrings: add asm version of Index() for short strings on amd64
Ilya Tocar [Wed, 28 Oct 2015 15:05:05 +0000 (18:05 +0300)]
strings: add asm version of Index() for short strings on amd64

Currently we have special case for 1-byte strings,
This extends this to strings shorter than 32 bytes on amd64.
Results (broadwell):

name                 old time/op  new time/op  delta
IndexRune-4          57.4ns ± 0%  57.5ns ± 0%   +0.10%        (p=0.000 n=20+19)
IndexRuneFastPath-4  20.4ns ± 0%  20.4ns ± 0%     ~     (all samples are equal)
Index-4              21.0ns ± 0%  21.8ns ± 0%   +3.81%        (p=0.000 n=20+20)
LastIndex-4          7.07ns ± 1%  6.98ns ± 0%   -1.21%        (p=0.000 n=20+16)
IndexByte-4          18.3ns ± 0%  18.3ns ± 0%     ~     (all samples are equal)
IndexHard1-4         1.46ms ± 0%  0.39ms ± 0%  -73.06%        (p=0.000 n=16+16)
IndexHard2-4         1.46ms ± 0%  0.30ms ± 0%  -79.55%        (p=0.000 n=18+18)
IndexHard3-4         1.46ms ± 0%  0.66ms ± 0%  -54.68%        (p=0.000 n=19+19)
LastIndexHard1-4     1.46ms ± 0%  1.46ms ± 0%   -0.01%        (p=0.036 n=18+20)
LastIndexHard2-4     1.46ms ± 0%  1.46ms ± 0%     ~           (p=0.588 n=19+19)
LastIndexHard3-4     1.46ms ± 0%  1.46ms ± 0%     ~           (p=0.283 n=17+20)
IndexTorture-4       11.1µs ± 0%  11.1µs ± 0%   +0.01%        (p=0.000 n=18+17)

Change-Id: I892781549f558f698be4e41f9f568e3d0611efb5
Reviewed-on: https://go-review.googlesource.com/16430
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>

9 years agoruntime: enlarge GC work buffer size
Austin Clements [Fri, 2 Oct 2015 13:51:12 +0000 (09:51 -0400)]
runtime: enlarge GC work buffer size

Currently the GC work buffers are only 256 bytes and hence can record
only 24 64-bit pointer. They were reduced from 4K in commits db7fd1c
and a15818f as a way to minimize the amount of work the per-P workbuf
caches could "hide" from the mark phase and carry in to the mark
termination phase. However, this approach wasn't very robust and we
later added a "mark 2" phase to address this problem head-on.

Because of mark 2, there's now no benefit to having very small work
buffers. But there are plenty of downsides: small work buffers
increase contention on the work lists, increase the frequency and
hence net overhead of acquiring and releasing work buffers, and
somewhat increase memory overhead of the GC.

This commit expands work buffers back to 4K (504 64-bit pointers).
This reduces the rate of writes to work.full in the garbage benchmark
from a peak of ~780,000 writes/sec to a peak of ~32,000 writes/sec.

This has negligible effect on the go1 benchmarks. It slightly slows
down the garbage benchmark.

name              old time/op  new time/op  delta
XBenchGarbage-12  5.37ms ± 5%  5.60ms ± 2%  +4.37%  (p=0.000 n=20+20)

Change-Id: Ic9cc28e7a125d23d9faf4f5e690fb8aa9bcdfb28
Reviewed-on: https://go-review.googlesource.com/15893
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: make assists preemptible
Austin Clements [Thu, 15 Oct 2015 21:58:17 +0000 (17:58 -0400)]
runtime: make assists preemptible

Currently, assists are non-preemptible, which means a heavily
assisting G can block other Gs from running. At the beginning of a GC
cycle, it can also delay scang, which will spin until the assist is
done. Since scanning is currently done sequentially, this can
seriously extend the length of the scan phase.

Fix this by making assists preemptible. Since the assist holds work
buffers and runs on the system stack, this must be done cooperatively:
we make gcDrainN return on preemption, and make the assist return from
the system stack and voluntarily Gosched.

This is prerequisite to enlarging the work buffers. Without this
change, the delays and spinning in scang increase significantly.

This has no effect on the go1 benchmarks.

name              old time/op  new time/op  delta
XBenchGarbage-12  5.72ms ± 4%  5.37ms ± 5%  -6.11%  (p=0.000 n=20+20)

Change-Id: I829e732a0f23b126da633516a1a9ec1a508fdbf1
Reviewed-on: https://go-review.googlesource.com/15894
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: replace assist sleep loop with park/ready
Austin Clements [Thu, 15 Oct 2015 01:31:33 +0000 (21:31 -0400)]
runtime: replace assist sleep loop with park/ready

GC assists must block until the assist can be satisfied (either
through stealing credit or doing work) or the GC cycle ends.
Currently, this is implemented as a retry loop with a 100 µs delay.
This obviously isn't ideal, as it wastes CPU and delays mutator
execution. It also has the somewhat peculiar downside that sleeping a
G requires allocation, and this requires working around recursive
allocation.

Replace this timed delay with a proper scheduling queue. When an
assist can't be satisfied immediately, it adds the allocating G to a
queue and parks it. Any time background scan credit is flushed, it
consults this queue, directly satisfies the debt of queued assists,
and wakes up satisfied assists before flushing any remaining credit to
the background credit pool.

No effect on the go1 benchmarks. Slightly speeds up the garbage
benchmark.

name              old time/op  new time/op  delta
XBenchGarbage-12  5.81ms ± 1%  5.72ms ± 4%  -1.65%  (p=0.011 n=20+20)

Updates #12041.

Change-Id: I8ee3b6274dd097b12b10a8030796a958a4b0e7b7
Reviewed-on: https://go-review.googlesource.com/15890
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: change p.runq from []*g to []guintptr
Austin Clements [Mon, 2 Nov 2015 21:59:39 +0000 (16:59 -0500)]
runtime: change p.runq from []*g to []guintptr

This eliminates many write barriers in the scheduler code that are
unnecessary and will interfere with upcoming changes where the garbage
collector will have to invoke run queue functions in contexts that
must not have write barriers.

Change-Id: I702d0ac99cfd00ffff406e7362917db6a43e7e55
Reviewed-on: https://go-review.googlesource.com/16556
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

9 years agomisc/ios: fix an error when getenv encounters unset variable
Michael Matloob [Mon, 2 Nov 2015 22:37:31 +0000 (17:37 -0500)]
misc/ios: fix an error when getenv encounters unset variable

The error message should indicate the name of the unset variable,
rather than the value. The value will alwayse be empty.

Change-Id: I6f6c165074dfce857b6523703a890d205423cd28
Reviewed-on: https://go-review.googlesource.com/16555
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agotest: use go:noinline consistently
Todd Neal [Fri, 30 Oct 2015 02:45:19 +0000 (21:45 -0500)]
test: use go:noinline consistently

Replace various implementations of inlining prevention with
"go:noinline"

Change-Id: Iac90895c3a62d6f4b7a6c72e11e165d15a0abfa4
Reviewed-on: https://go-review.googlesource.com/16510
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Todd Neal <todd@tneal.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/link: remove duplicate symtab entry for global functions
Michael Hudson-Doyle [Mon, 2 Nov 2015 20:36:53 +0000 (09:36 +1300)]
cmd/link: remove duplicate symtab entry for global functions

golang.org/cl/16436 added a local symbol for every global function, but also
added a duplicate entry for the global symbol. Surprisingly this hasn't caused
any noticeable problems, but it's still wrong.

Change-Id: Icd3906760f8aaf7bef31ffd4f2d866d73d36dc2c
Reviewed-on: https://go-review.googlesource.com/16581
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agobytes: speed up Compare() on amd64
Ilya Tocar [Wed, 28 Oct 2015 20:20:26 +0000 (23:20 +0300)]
bytes: speed up Compare() on amd64

Use AVX2 if available.
Results (haswell), below:

name                           old time/op    new time/op     delta
BytesCompare1-6                  11.4ns ± 0%     11.4ns ± 0%     ~     (all samples are equal)
BytesCompare2-6                  11.4ns ± 0%     11.4ns ± 0%     ~     (all samples are equal)
BytesCompare4-6                  11.4ns ± 0%     11.4ns ± 0%     ~     (all samples are equal)
BytesCompare8-6                  9.29ns ± 2%     8.76ns ± 0%   -5.72%        (p=0.000 n=16+17)
BytesCompare16-6                 9.29ns ± 2%     9.20ns ± 0%   -1.02%        (p=0.000 n=20+16)
BytesCompare32-6                 11.4ns ± 1%     11.4ns ± 0%     ~           (p=0.191 n=20+20)
BytesCompare64-6                 14.4ns ± 0%     13.1ns ± 0%   -8.68%        (p=0.000 n=20+20)
BytesCompare128-6                20.2ns ± 0%     18.5ns ± 0%   -8.27%        (p=0.000 n=16+20)
BytesCompare256-6                29.3ns ± 0%     24.5ns ± 0%  -16.38%        (p=0.000 n=16+16)
BytesCompare512-6                46.8ns ± 0%     37.1ns ± 0%  -20.78%        (p=0.000 n=18+16)
BytesCompare1024-6               82.9ns ± 0%     62.3ns ± 0%  -24.86%        (p=0.000 n=20+14)
BytesCompare2048-6                155ns ± 0%      112ns ± 0%  -27.74%        (p=0.000 n=20+20)
CompareBytesEqual-6              10.1ns ± 1%     10.0ns ± 1%     ~           (p=0.527 n=20+20)
CompareBytesToNil-6              10.0ns ± 2%      9.4ns ± 0%   -6.57%        (p=0.000 n=20+17)
CompareBytesEmpty-6              8.76ns ± 0%     8.76ns ± 0%     ~     (all samples are equal)
CompareBytesIdentical-6          8.76ns ± 0%     8.76ns ± 0%     ~     (all samples are equal)
CompareBytesSameLength-6         10.6ns ± 1%     10.6ns ± 1%     ~           (p=0.240 n=20+20)
CompareBytesDifferentLength-6    10.6ns ± 0%     10.6ns ± 1%     ~           (p=1.000 n=20+20)
CompareBytesBigUnaligned-6        132±s ± 1%      105±s ± 1%  -20.61%        (p=0.000 n=20+18)
CompareBytesBig-6                 125±s ± 1%      105±s ± 1%  -16.31%        (p=0.000 n=20+20)
CompareBytesBigIdentical-6       8.13ns ± 0%     8.13ns ± 0%     ~     (all samples are equal)

name                           old speed      new speed       delta
CompareBytesBigUnaligned-6     7.94GB/s ± 1%  10.01GB/s ± 1%  +25.96%        (p=0.000 n=20+18)
CompareBytesBig-6              8.38GB/s ± 1%  10.01GB/s ± 1%  +19.48%        (p=0.000 n=20+20)
CompareBytesBigIdentical-6      129TB/s ± 0%    129TB/s ± 0%   +0.01%        (p=0.003 n=17+19)

Change-Id: I820f31bab4582dd4204b146bb077c0d2f24cd8f5
Reviewed-on: https://go-review.googlesource.com/16434
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
Reviewed-by: Klaus Post <klauspost@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
9 years agocmd/link: concurrent obj copy for external linking
David Crawshaw [Wed, 28 Oct 2015 18:41:58 +0000 (14:41 -0400)]
cmd/link: concurrent obj copy for external linking

Change-Id: I630ae29ecb39252642883398cc51d49133c6f3d7
Reviewed-on: https://go-review.googlesource.com/16451
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agotest: update tinyfin test
Brad Fitzpatrick [Mon, 2 Nov 2015 15:46:44 +0000 (07:46 -0800)]
test: update tinyfin test

* use new(int32) to be pedantic about documented SetFinalizer rules:
  "The argument x must be a pointer to an object allocated by calling
  new or by taking the address of a composite literal"

* remove the amd64-only restriction. The GC is fully precise everywhere
  now, even on 32-bit. (keep the gccgo restriction, though)

* remove a data race (perhaps the actual bug) and use atomic.LoadInt32
  for the final check. The race detector is now happy, too.

Updates #13100

Change-Id: I8d05c0ac4f046af9ba05701ad709c57984b34893
Reviewed-on: https://go-review.googlesource.com/16535
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agotest: add gcc6798, valid code that caused gccgo to crash
Ian Lance Taylor [Sun, 1 Nov 2015 20:30:54 +0000 (12:30 -0800)]
test: add gcc6798, valid code that caused gccgo to crash

Change-Id: I288bd3091ea81db7b616747cbec8958a31d98b7e
Reviewed-on: https://go-review.googlesource.com/16532
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/go, runtime: define GOBUILDMODE_shared rather than shared when dynamically linking
Michael Hudson-Doyle [Thu, 29 Oct 2015 22:48:43 +0000 (11:48 +1300)]
cmd/go, runtime: define GOBUILDMODE_shared rather than shared when dynamically linking

To avoid collisions with what existing code may already be doing.

Change-Id: Ice639440aafc0724714c25333d90a49954372230
Reviewed-on: https://go-review.googlesource.com/16503
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agonet: make Dial, Listen{,Packet} for TCP/UDP with invalid port fail
Mikio Hara [Tue, 21 Jul 2015 03:51:01 +0000 (12:51 +0900)]
net: make Dial, Listen{,Packet} for TCP/UDP with invalid port fail

This change makes Dial, Listen and ListenPacket with invalid port fail
whatever GODEBUG=netdns is.

Please be informed that cgoLookupPort with an out of range literal
number may return either the lower or upper bound value, 0 or 65535,
with no error on some platform.

Fixes #11715.

Change-Id: I43f9c4fb5526d1bf50b97698e0eb39d29fd74c35
Reviewed-on: https://go-review.googlesource.com/12447
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agopath/filepath: fix EvalSymlinks(".") on windows
Alex Brainman [Mon, 21 Sep 2015 06:12:24 +0000 (16:12 +1000)]
path/filepath: fix EvalSymlinks(".") on windows

Also new tests added. So, perhaps, this CL corrects
even more broken EvalSymlinks behaviour.

Fixes #12451

Change-Id: I81b9d92bab74bcb8eca6db6633546982fe5cec87
Reviewed-on: https://go-review.googlesource.com/16192
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Alex Brainman <alex.brainman@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/link: support new 386/amd64 relocations
Ian Lance Taylor [Sat, 31 Oct 2015 21:44:48 +0000 (14:44 -0700)]
cmd/link: support new 386/amd64 relocations

The GNU binutils recently picked up support for new 386/amd64
relocations.  Add support for them in the Go linker when doing an
internal link.

The 386 relocation R_386_GOT32X was proposed in
https://groups.google.com/forum/#!topic/ia32-abi/GbJJskkid4I .  It can
be treated as identical to the R_386_GOT32 relocation.

The amd64 relocations R_X86_64_GOTPCRELX and R_X86_64_REX_GOTPCRELX were
proposed in
https://groups.google.com/forum/#!topic/x86-64-abi/n9AWHogmVY0 .  They
can both be treated as identical to the R_X86_64_GOTPCREL relocation.

The purpose of the new relocations is to permit additional linker
relaxations in some cases.  We do not attempt to support those cases.

While we're at it, remove the unused and in some cases out of date
_COUNT names from ld/elf.go.

Fixes #13114.

Change-Id: I34ef07f6fcd00cdd2996038ecf46bb77a49e968b
Reviewed-on: https://go-review.googlesource.com/16529
Reviewed-by: Minux Ma <minux@golang.org>
9 years agogo/types: fix TypeString(nil, nil)
Matthew Dempsky [Thu, 29 Oct 2015 01:55:30 +0000 (18:55 -0700)]
go/types: fix TypeString(nil, nil)

The code is meant to return "<nil>", but because of a make([]Type, 8)
call that should be make([]Type, 0, 8), the nil Type happens to
already appear in the array.

Change-Id: I2db140046e52f27db1b0ac84bde2b6680677dd95
Reviewed-on: https://go-review.googlesource.com/16464
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
9 years agomath: fix bad shift in Expm1
Matthew Dempsky [Fri, 30 Oct 2015 02:16:20 +0000 (19:16 -0700)]
math: fix bad shift in Expm1

Noticed by cmd/vet.

Expected values array produced by Python instead of Keisan because:

1) Keisan's website calculator is painfully difficult to copy/paste
values into and out of, and

2) after tediously computing e^(vf[i] * 10) - 1 via Keisan I
discovered that Keisan computing vf[i]*10 in a higher precision was
giving substantially different output values.

Also, testing uses "close" instead of "veryclose" because 386's
assembly implementation produces values for some of the test cases
that fail "veryclose".  Curiously, Expm1(vf[i]*10) is identical to
Exp(vf[i]*10)-1 on 386, whereas with the portable implementation
they're only "veryclose".

Investigating these questions is left to someone else.  I just wanted
to fix the cmd/vet warning.

Fixes #13101.

Change-Id: Ica8f6c267d01aa4cc31f53593e95812746942fbc
Reviewed-on: https://go-review.googlesource.com/16505
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Klaus Post <klauspost@gmail.com>
Reviewed-by: Robert Griesemer <gri@golang.org>