]> Cypherpunks repositories - gostls13.git/log
gostls13.git
10 years agonet: don't miss testing server teardowns when test fails early
Mikio Hara [Sat, 25 Apr 2015 22:52:28 +0000 (07:52 +0900)]
net: don't miss testing server teardowns when test fails early

Change-Id: I9fa678e43b4ae3970323cac474b5f86d4d933997
Reviewed-on: https://go-review.googlesource.com/9382
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agotest: reenable syntax tests
Matthew Dempsky [Mon, 27 Apr 2015 23:06:22 +0000 (16:06 -0700)]
test: reenable syntax tests

These were fixed a little while ago, but overlooked when reenabling
disabled tests.

Update #9968.

Change-Id: I301ef587e580c517a170ad08ff897118b58cedec
Reviewed-on: https://go-review.googlesource.com/9347
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
10 years agodoc/go1.5.txt: go doc
Rob Pike [Mon, 27 Apr 2015 23:23:43 +0000 (16:23 -0700)]
doc/go1.5.txt: go doc

Change-Id: I883017b67e8fa76b6f123e8f9bcec3d6f820bbb3
Reviewed-on: https://go-review.googlesource.com/9348
Reviewed-by: Rob Pike <r@golang.org>
10 years agocmd/go,cmd/doc: add "go doc"
Rob Pike [Fri, 24 Apr 2015 19:28:18 +0000 (12:28 -0700)]
cmd/go,cmd/doc: add "go doc"

Add the new go doc command to the go command, installed in
the tool directory.

(Still to do: tests)

Fix cmd/dist to remove old "package documentation" code that was
stopping it from including cmd/go/doc.go in the build.

Implement the doc command. Here is the help info from "go help doc":

===
usage: go doc [-u] [package|[package.]symbol[.method]]

Doc accepts at most one argument, indicating either a package, a symbol within a
package, or a method of a symbol.

go doc
go doc <pkg>
go doc <sym>[.<method>]
go doc [<pkg>].<sym>[.<method>]

Doc interprets the argument to see what it represents, determined by its syntax
and which packages and symbols are present in the source directories of GOROOT and
GOPATH.

The first item in this list that succeeds is the one whose documentation is printed.
For packages, the order of scanning is determined by the file system, however the
GOROOT tree is always scanned before GOPATH.

If there is no package specified or matched, the package in the current directory
is selected, so "go doc" shows the documentation for the current package and
"go doc Foo" shows the documentation for symbol Foo in the current package.

Doc prints the documentation comments associated with the top-level item the
argument identifies (package, type, method) followed by a one-line summary of each
of the first-level items "under" that item (package-level declarations for a
package, methods for a type, etc.)

The package paths must be either a qualified path or a proper suffix of a path
(see examples below). The go tool's usual package mechanism does not apply: package
path elements like . and ...  are not implemented by go doc.

When matching symbols, lower-case letters match either case but upper-case letters
match exactly.

Examples:
go doc
Show documentation for current package.
go doc Foo
Show documentation for Foo in the current package.
(Foo starts with a capital letter so it cannot match a package path.)
go doc json
Show documentation for the encoding/json package.
go doc json
Shorthand for encoding/json assuming only one json package
is present in the tree.
go doc json.Number (or go doc json.number)
Show documentation and method summary for json.Number.
go doc json.Number.Int64 (or go doc json.number.int64)
Show documentation for the Int64 method of json.Number.

Flags:
-u
Show documentation for unexported as well as exported
symbols and methods.

===

Still to do:

Tests.
Disambiguation when there is both foo and Foo.
Flag for case-sensitive matching.

Change-Id: I83d409a68688a5445f54297a7e7c745f749b9e66
Reviewed-on: https://go-review.googlesource.com/9227
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime/race: fix failing tests
Austin Clements [Mon, 27 Apr 2015 21:15:30 +0000 (17:15 -0400)]
runtime/race: fix failing tests

Some race tests were sensitive to the goroutine scheduling order.
When this changed in commit e870f06, these tests started to fail.

Fix TestRaceHeapParam by ensuring that the racing goroutine has
run before the test exits. Fix TestRaceRWMutexMultipleReaders by
adding a third reader to ensure that two readers wind up on the
same side of the writer (and race with each other) regardless of
the schedule. Fix TestRaceRange by ensuring that the racing
goroutine runs before the main goroutine exits the loop it races
with.

Change-Id: Iaf002f8730ea42227feaf2f3c51b9a1e57ccffdd
Reviewed-on: https://go-review.googlesource.com/9402
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime/race: stop listening to external network addresses
Russ Cox [Mon, 27 Apr 2015 20:18:35 +0000 (16:18 -0400)]
runtime/race: stop listening to external network addresses

This makes the OS X firewall box pop up.
Not run during all.bash so hasn't been noticed before.

Change-Id: I78feb4fd3e1d3c983ae3419085048831c04de3da
Reviewed-on: https://go-review.googlesource.com/9401
Reviewed-by: Austin Clements <austin@google.com>
10 years agoruntime: fix stack use accounting
Austin Clements [Mon, 27 Apr 2015 18:50:42 +0000 (14:50 -0400)]
runtime: fix stack use accounting

ReadMemStats accounts for stacks slightly differently than the runtime
does internally. Internally, only stacks allocated by newosproc0 are
accounted in memstats.stacks_sys and other stacks are accounted in
heap_sys. readmemstats_m shuffles the statistics so all stacks are
accounted in StackSys rather than HeapSys.

However, currently, readmemstats_m assumes StackSys will be zero when
it does this shuffle. This was true until commit 6ad33be. If it isn't
(e.g., if something called newosproc0), StackSys+HeapSys will be
different before and after this shuffle, and the Sys sum that was
computed earlier will no longer agree with the sum of its components.

Fix this by making the shuffle in readmemstats_m not assume that
StackSys is zero.

Fixes #10585.

Change-Id: If13991c8de68bd7b85e1b613d3f12b4fd6fd5813
Reviewed-on: https://go-review.googlesource.com/9366
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: remove unnecessary noescape to fix netbsd
David Crawshaw [Mon, 27 Apr 2015 14:56:38 +0000 (10:56 -0400)]
runtime: remove unnecessary noescape to fix netbsd

I introduced this build failure in golang.org/cl/9302 but failed to
notice due to the other failures on the dashboard.

Change-Id: I84bf00f664ba572c1ca722e0136d8a2cf21613ca
Reviewed-on: https://go-review.googlesource.com/9363
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Minux Ma <minux@golang.org>
10 years agocmd/internal/gc, cmd/internal/ld: add memprofilerate flag
Josh Bleecher Snyder [Thu, 23 Apr 2015 00:53:32 +0000 (17:53 -0700)]
cmd/internal/gc, cmd/internal/ld: add memprofilerate flag

Also call runtime.GC before exit to ensure
that the profiler picks up all allocations.

Fixes #10537.

Change-Id: Ibfbfc88652ac0ce30a6d1ae392f919df6c1e8126
Reviewed-on: https://go-review.googlesource.com/9261
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Minux Ma <minux@golang.org>
Run-TryBot: Minux Ma <minux@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime/race: fix benchmark deadlock
Austin Clements [Mon, 27 Apr 2015 20:31:16 +0000 (16:31 -0400)]
runtime/race: fix benchmark deadlock

Currently TestRaceCrawl fails to wg.Done for every wg.Adds if the
depth ever reaches 0. This causes the test to deadlock. Under the race
detector, this deadlock is not detected, so the test eventually times
out.

This only recently became a problem. Prior to commit e870f06 the depth
would never reach 0 because the strict round-robin goroutine schedule
ensured that all of the URLs were already "seen" by depth 2. Now that
the runtime prefers scheduling the most recently started goroutine,
the test is able to reach depth 0 and trigger this deadlock.

Change-Id: I5176302a89614a344c84d587073b364833af6590
Reviewed-on: https://go-review.googlesource.com/9344
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoregexp: trivial change in comments to update code.google.com link
Dmitry Savintsev [Mon, 27 Apr 2015 09:57:21 +0000 (11:57 +0200)]
regexp: trivial change in comments to update code.google.com link

Replaced code.google.com/p/re2/ with github.com/google/re2/ and
updated the file names (re2-exhaustive.txt.bz2 not re2.txt.gz)
as well as the re2 make command (make log).

Change-Id: I15937b0b8a898d78d45366857ed86421c8d69960
Reviewed-on: https://go-review.googlesource.com/9372
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
10 years agoruntime: fix race in BenchmarkPingPongHog
Russ Cox [Mon, 27 Apr 2015 20:08:11 +0000 (16:08 -0400)]
runtime: fix race in BenchmarkPingPongHog

The master goroutine was returning before
the child goroutine had done its final i < b.N
(the one that fails and causes it to exit the loop)
and then the benchmark harness was updating
b.N, causing a read+write race on b.N.

Change-Id: I2504270a0de30544736f6c32161337a25b505c3e
Reviewed-on: https://go-review.googlesource.com/9368
Reviewed-by: Austin Clements <austin@google.com>
10 years agoruntime: fix some out of date comments and typos
Austin Clements [Mon, 27 Apr 2015 19:42:45 +0000 (15:42 -0400)]
runtime: fix some out of date comments and typos

Change-Id: I061057414c722c5a0f03c709528afc8554114db6
Reviewed-on: https://go-review.googlesource.com/9367
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agoruntime: remove a modulus calculation from pollorder
Josh Bleecher Snyder [Mon, 27 Apr 2015 17:46:02 +0000 (10:46 -0700)]
runtime: remove a modulus calculation from pollorder

This is a follow-up to CL 9269, as suggested
by dvyukov.

There is probably even more that can be done
to speed up this shuffle. It will matter more
once CL 7570 (fine-grained locking in select)
is in and can be revisited then, with benchmarks.

Change-Id: Ic13a27d11cedd1e1f007951214b3bb56b1644f02
Reviewed-on: https://go-review.googlesource.com/9393
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
10 years agoruntime: rename gcController.findRunnable to findRunnableGCWorker
Austin Clements [Fri, 24 Apr 2015 18:17:42 +0000 (14:17 -0400)]
runtime: rename gcController.findRunnable to findRunnableGCWorker

This avoids confusion with the main findrunnable in the scheduler.

Change-Id: I8cf40657557a8610a2fe5a2f74598518256ca7f0
Reviewed-on: https://go-review.googlesource.com/9305
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agoruntime: replace STW for enabling write barriers with ragged barrier
Austin Clements [Fri, 27 Mar 2015 21:01:53 +0000 (17:01 -0400)]
runtime: replace STW for enabling write barriers with ragged barrier

Currently, we use a full stop-the-world around enabling write
barriers. This is to ensure that all Gs have enabled write barriers
before any blackening occurs (either in gcBgMarkWorker() or in
gcAssistAlloc()).

However, there's no need to bring the whole world to a synchronous
stop to ensure this. This change replaces the STW with a ragged
barrier that ensures each P has individually observed that write
barriers should be enabled before GC performs any blackening.

Change-Id: If2f129a6a55bd8bdd4308067af2b739f3fb41955
Reviewed-on: https://go-review.googlesource.com/8207
Reviewed-by: Russ Cox <rsc@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agoruntime: add ragged global barrier function
Austin Clements [Fri, 27 Mar 2015 20:49:12 +0000 (16:49 -0400)]
runtime: add ragged global barrier function

This adds forEachP, which performs a general-purpose ragged global
barrier. forEachP takes a callback and invokes it for every P at a GC
safe point.

Ps that are idle or in a syscall are considered to be at a continuous
safe point. forEachP ensures that these Ps do not change state by
forcing all syscall Ps into idle and holding the sched.lock.

To ensure that Ps do not enter syscall or idle without running the
safe-point function, this adds checks for a pending callback every
place there is currently a gcwaiting check.

We'll use forEachP to replace the STW around enabling the write
barrier and to replace the current asynchronous per-M wbuf cache with
a cooperatively managed per-P gcWork cache.

Change-Id: Ie944f8ce1fead7c79bf271d2f42fcd61a41bb3cc
Reviewed-on: https://go-review.googlesource.com/8206
Reviewed-by: Russ Cox <rsc@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agoRevert "cmd/dist: consolidate runtime CPU tests"
Josh Bleecher Snyder [Mon, 27 Apr 2015 17:54:51 +0000 (17:54 +0000)]
Revert "cmd/dist: consolidate runtime CPU tests"

This reverts commit a9e50a6b358df2e725157128c140684264c8ed38.

Change-Id: I3c5e459f1030e36bc249910facdae12303a44151
Reviewed-on: https://go-review.googlesource.com/9394
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
10 years agocmd/dist: consolidate runtime CPU tests
Josh Bleecher Snyder [Sat, 25 Apr 2015 00:06:52 +0000 (17:06 -0700)]
cmd/dist: consolidate runtime CPU tests

Instead of running:

go test -short runtime -cpu=1
go test -short runtime -cpu=2
go test -short runtime -cpu=4

Run just:

go test -short runtime -cpu=1,2,4

This is a return to the Go 1.4.2 behavior.

We lose incremental display of progress and
per-cpu timing information, but we don't have
to recompile and relink the runtime test,
which is slow.

This cuts about 10s off all.bash.

Updates #10571.

Change-Id: I6e8c7149780d47439f8bcfa888e6efc84290c60a
Reviewed-on: https://go-review.googlesource.com/9350
Reviewed-by: Dave Cheney <dave@cheney.net>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
10 years agocmd/internal/ld: remove pointless allocs
Josh Bleecher Snyder [Sat, 25 Apr 2015 03:24:49 +0000 (20:24 -0700)]
cmd/internal/ld: remove pointless allocs

Reduces allocs linking cmd/go and runtime.test
by ~13%. No functional changes.

The most easily addressed sources of allocations
after this are expandpkg, rdstring, and symbuf
string conversion.

These can be reduced by interning strings,
but that increases the overall memory footprint.

Change-Id: Ifedefc9f2a0403bcc75460d6b139e8408374e058
Reviewed-on: https://go-review.googlesource.com/9391
Reviewed-by: David Crawshaw <crawshaw@golang.org>
10 years agoencoding/xml: do not escape newlines
Roger Peppe [Fri, 24 Apr 2015 16:20:21 +0000 (17:20 +0100)]
encoding/xml: do not escape newlines

There is no need to escape newlines in char data -
it makes the XML larger and harder to read.

Change-Id: I1c1fcee1bdffc705c7428f89ca90af8085d6fb73
Reviewed-on: https://go-review.googlesource.com/9310
Reviewed-by: Nigel Tao <nigeltao@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: reset spinning in mspinning if work was ready()ed
Austin Clements [Sat, 25 Apr 2015 02:33:13 +0000 (22:33 -0400)]
runtime: reset spinning in mspinning if work was ready()ed

This fixes a bug where the runtime ready()s a goroutine while setting
up a new M that's initially marked as spinning, causing the scheduler
to later panic when it finds work in the run queue of a P associated
with a spinning M. Specifically, the sequence of events that can lead
to this is:

1) sysmon calls handoffp to hand off a P stolen from a syscall.

2) handoffp sees no pending work on the P, so it calls startm with
   spinning set.

3) startm calls newm, which in turn calls allocm to allocate a new M.

4) allocm "borrows" the P we're handing off in order to do allocation
   and performs this allocation.

5) This allocation may assist the garbage collector, and this assist
   may detect the end of concurrent mark and ready() the main GC
   goroutine to signal this.

6) This ready()ing puts the GC goroutine on the run queue of the
   borrowed P.

7) newm starts the OS thread, which runs mstart and subsequently
   mstart1, which marks the M spinning because startm was called with
   spinning set.

8) mstart1 enters the scheduler, which panics because there's work on
   the run queue, but the M is marked spinning.

To fix this, before marking the M spinning in step 7, add a check to
see if work was been added to the P's run queue. If this is the case,
undo the spinning instead.

Fixes #10573.

Change-Id: I4670495ae00582144a55ce88c45ae71de597cfa5
Reviewed-on: https://go-review.googlesource.com/9332
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

10 years agoruntime: panic when idling a P with runnable Gs
Austin Clements [Fri, 24 Apr 2015 20:37:31 +0000 (16:37 -0400)]
runtime: panic when idling a P with runnable Gs

This adds a check that we never put a P on the idle list when it has
work on its local run queue.

Change-Id: Ifcfab750de60c335148a7f513d4eef17be03b6a7
Reviewed-on: https://go-review.googlesource.com/9324
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
10 years agoruntime: tighten select permutation generation
Josh Bleecher Snyder [Thu, 18 Dec 2014 19:48:39 +0000 (11:48 -0800)]
runtime: tighten select permutation generation

This is the optimization made to math/rand in CL 21030043.

Change-Id: I231b24fa77cac1fe74ba887db76313b5efaab3e8
Reviewed-on: https://go-review.googlesource.com/9269
Reviewed-by: Minux Ma <minux@golang.org>
10 years agodebug/dwarf: update class_string.go to add ClassReferenceSig using stringer.
John Dethridge [Wed, 22 Apr 2015 06:45:42 +0000 (16:45 +1000)]
debug/dwarf: update class_string.go to add ClassReferenceSig using stringer.

Change-Id: I677a5ee273a4d285a8adff71ffcfeac34afc887f
Reviewed-on: https://go-review.googlesource.com/9235
Reviewed-by: Austin Clements <austin@google.com>
10 years agocrypto/tls: call GetCertificate if Certificates is empty.
Adam Langley [Sun, 12 Apr 2015 23:41:31 +0000 (16:41 -0700)]
crypto/tls: call GetCertificate if Certificates is empty.

This change causes the GetCertificate callback to be called if
Certificates is empty. Previously this configuration would result in an
error.

This allows people to have servers that depend entirely on dynamic
certificate selection, even when the client doesn't send SNI.

Fixes #9208.

Change-Id: I2f5a5551215958b88b154c64a114590300dfc461
Reviewed-on: https://go-review.googlesource.com/8792
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Adam Langley <agl@golang.org>

10 years agocrypto/tls: add OCSP response to ConnectionState
Jonathan Rudenberg [Sun, 26 Apr 2015 16:05:37 +0000 (12:05 -0400)]
crypto/tls: add OCSP response to ConnectionState

The OCSP response is currently only exposed via a method on Conn,
which makes it inaccessible when using wrappers like net/http. The
ConnectionState structure is typically available even when using
wrappers and contains many of the other handshake details, so this
change exposes the stapled OCSP response in that structure.

Change-Id: If8dab49292566912c615d816321b4353e711f71f
Reviewed-on: https://go-review.googlesource.com/9361
Reviewed-by: Adam Langley <agl@golang.org>
Run-TryBot: Adam Langley <agl@golang.org>

10 years agocrypto/elliptic: don't unmarshal points that are off the curve
David Leon Gil [Wed, 7 Jan 2015 05:07:24 +0000 (21:07 -0800)]
crypto/elliptic: don't unmarshal points that are off the curve

At present, Unmarshal does not check that the point it unmarshals
is actually *on* the curve. (It may be on the curve's twist.)

This can, as Daniel Bernstein has pointed out at great length,
lead to quite devastating attacks. And 3 out of the 4 curves
supported by crypto/elliptic have twists with cofactor != 1;
P-224, in particular, has a sufficiently large cofactor that it
is likely that conventional dlog attacks might be useful.

This closes #2445, filed by Watson Ladd.

To explain why this was (partially) rejected before being accepted:

In the general case, for curves with cofactor != 1, verifying subgroup
membership is required. (This is expensive and hard-to-implement.)
But, as recent discussion during the CFRG standardization process
has brought out, small-subgroup attacks are much less damaging than
a twist attack.

Change-Id: I284042eb9954ff9b7cde80b8b693b1d468c7e1e8
Reviewed-on: https://go-review.googlesource.com/2421
Reviewed-by: Adam Langley <agl@golang.org>
10 years agocrypto/x509: CertificateRequest signature verification
Paul van Brouwershaven [Wed, 11 Mar 2015 07:15:23 +0000 (08:15 +0100)]
crypto/x509: CertificateRequest signature verification

This implements a method for x509.CertificateRequest to prevent
certain attacks and to allow a CA/RA to properly check the validity
of the binding between an end entity and a key pair, to prove that
it has possession of (i.e., is able to use) the private key
corresponding to the public key for which a certificate is requested.

RFC 2986 section 3 states:

"A certification authority fulfills the request by authenticating the
requesting entity and verifying the entity's signature, and, if the
request is valid, constructing an X.509 certificate from the
distinguished name and public key, the issuer name, and the
certification authority's choice of serial number, validity period,
and signature algorithm."

Change-Id: I37795c3b1dfdfdd455d870e499b63885eb9bda4f
Reviewed-on: https://go-review.googlesource.com/7371
Reviewed-by: Adam Langley <agl@golang.org>
10 years agocrypto/tls: add support for session ticket key rotation
Jonathan Rudenberg [Sat, 18 Apr 2015 01:32:11 +0000 (21:32 -0400)]
crypto/tls: add support for session ticket key rotation

This change adds a new method to tls.Config, SetSessionTicketKeys, that
changes the key used to encrypt session tickets while the server is
running. Additional keys may be provided that will be used to maintain
continuity while rotating keys. If a ticket encrypted with an old key is
provided by the client, the server will resume the session and provide
the client with a ticket encrypted using the new key.

Fixes #9994

Change-Id: Idbc16b10ff39616109a51ed39a6fa208faad5b4e
Reviewed-on: https://go-review.googlesource.com/9072
Reviewed-by: Jonathan Rudenberg <jonathan@titanous.com>
Reviewed-by: Adam Langley <agl@golang.org>
10 years agocmd/pprof: handle empty profile gracefully
Håvard Haugen [Thu, 8 Jan 2015 23:11:44 +0000 (00:11 +0100)]
cmd/pprof: handle empty profile gracefully

The command "go tool pprof -top $GOROOT/bin/go /dev/null" now logs that
profile is empty instead of panicking.

Fixes #9207

Change-Id: I3d55c179277cb19ad52c8f24f1aca85db53ee08d
Reviewed-on: https://go-review.googlesource.com/2571
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
10 years agocrypto/tls: add support for Certificate Transparency
Jonathan Rudenberg [Thu, 16 Apr 2015 18:59:22 +0000 (14:59 -0400)]
crypto/tls: add support for Certificate Transparency

This change adds support for serving and receiving Signed Certificate
Timestamps as described in RFC 6962.

The server is now capable of serving SCTs listed in the Certificate
structure. The client now asks for SCTs and, if any are received,
they are exposed in the ConnectionState structure.

Fixes #10201

Change-Id: Ib3adae98cb4f173bc85cec04d2bdd3aa0fec70bb
Reviewed-on: https://go-review.googlesource.com/8988
Reviewed-by: Adam Langley <agl@golang.org>
Run-TryBot: Adam Langley <agl@golang.org>
Reviewed-by: Jonathan Rudenberg <jonathan@titanous.com>
10 years agoencoding/csv: Preallocate records slice
Justin Nuß [Mon, 13 Apr 2015 18:13:14 +0000 (20:13 +0200)]
encoding/csv: Preallocate records slice

Currently parseRecord will always start with a nil
slice and then resize the slice on append. For input
with a fixed number of fields per record we can preallocate
the slice to avoid having to resize the slice.

This change implements this optimization by using
FieldsPerRecord as capacity if it's > 0 and also adds a
benchmark to better show the differences.

benchmark         old ns/op     new ns/op     delta
BenchmarkRead     19741         17909         -9.28%

benchmark         old allocs     new allocs     delta
BenchmarkRead     59             41             -30.51%

benchmark         old bytes     new bytes     delta
BenchmarkRead     6276          5844          -6.88%

Change-Id: I7c2abc9c80a23571369bcfcc99a8ffc474eae7ab
Reviewed-on: https://go-review.googlesource.com/8880
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

10 years agoruntime: signal forwarding for darwin/amd64
David Crawshaw [Fri, 24 Apr 2015 16:47:46 +0000 (12:47 -0400)]
runtime: signal forwarding for darwin/amd64

Follows the linux signal forwarding semantics from
http://golang.org/cl/8712, sharing the implementation of sigfwdgo.
Forwarding for 386, arm, and arm64 will follow.

Change-Id: I6bf30d563d19da39b6aec6900c7fe12d82ed4f62
Reviewed-on: https://go-review.googlesource.com/9302
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agocmd/internal/ld: R_TLS_LE is fine on Darwin too
Michael Hudson-Doyle [Sun, 26 Apr 2015 04:37:31 +0000 (16:37 +1200)]
cmd/internal/ld: R_TLS_LE is fine on Darwin too

Sorry about this.

Fixes #10575

Change-Id: I2de23be68e7d822d182e5a0d6a00c607448d861e
Reviewed-on: https://go-review.googlesource.com/9341
Reviewed-by: Minux Ma <minux@golang.org>
10 years agotesting/quick: align tests with reflect.Kind.
Matt T. Proud [Sun, 12 Apr 2015 17:50:52 +0000 (19:50 +0200)]
testing/quick: align tests with reflect.Kind.

This commit is largely cosmetic in the sense that it is the remnants
of a change proposal I had prepared for testing/quick, until I
discovered that 3e9ed27 already implemented the feature I was looking
for: quick.Value() for reflect.Kind Array.  What you see is a merger
and manual cleanup; the cosmetic cleanups are as follows:

(1.) Keeping the TestCheckEqual and its associated input functions
in the same order as type kinds defined in reflect.Kind.  Since
3e9ed27 was committed, the test case began to diverge from the
constant's ordering.

(2.) The `Intptr` derivatives existed to exercise quick.Value with
reflect.Kind's `Ptr` constant.  All `Intptr` (unrelated to `uintptr`)
in the test have been migrated to ensure the parallelism of the
listings and to convey that `Intptr` is not special.

(3.) Correct a misspelling (transposition) of "alias", whereby it is
named as "Alais".

Change-Id: I441450db16b8bb1272c52b0abcda3794dcd0599d
Reviewed-on: https://go-review.googlesource.com/8804
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agocmd/8l, cmd/internal/ld, cmd/internal/obj/x86: stop incorrectly using the term "inita...
Michael Hudson-Doyle [Thu, 23 Apr 2015 09:53:48 +0000 (21:53 +1200)]
cmd/8l, cmd/internal/ld, cmd/internal/obj/x86: stop incorrectly using the term "inital exec"

The long comment block in obj6.go:progedit talked about the two code sequences
for accessing g as "local exec" and "initial exec", but really they are both forms
of local exec. This stuff is confusing enough without using the wrong words for
things, so rewrite it to talk about 2-instruction and 1-instruction sequences.
Unfortunately the confusion has made it into code, with the R_TLS_IE relocation
now doing double duty as meaning actual initial exec when externally linking and
boring old local exec when linking internally (half of this is my fault). So this
stops using R_TLS_IE in the local exec case. There is a chance this might break
plan9 or windows, but I don't think so. Next step is working out what the heck is
going on on ARM...

Change-Id: I09da4388210cf49dbc99fd25f5172bbe517cee57
Reviewed-on: https://go-review.googlesource.com/9273
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>

10 years agoruntime: Fix bug due to elided return.
Rick Hudson [Thu, 23 Apr 2015 20:28:15 +0000 (16:28 -0400)]
runtime: Fix bug due to elided return.

A previous change to mbitmap.go dropped a return on a
path the seems not to be excersized. This was a mistake that
this CL fixes.

Change-Id: I715ee4ef08f5bf8d9f53cee84e8fb31a237e2d43
Reviewed-on: https://go-review.googlesource.com/9295
Reviewed-by: Austin Clements <austin@google.com>
10 years agocmd/internal/ld: fix R_TLS handling now Xsym is not read from object file
Michael Hudson-Doyle [Thu, 23 Apr 2015 08:04:14 +0000 (20:04 +1200)]
cmd/internal/ld: fix R_TLS handling now Xsym is not read from object file

I think this should fix the arm build. A proper fix involves making the handling
of tlsg less fragile, I'll try that tomorrow.

Update #10557

Change-Id: I9b1b666737fb40aebb6f284748509afa8483cce5
Reviewed-on: https://go-review.googlesource.com/9272
Reviewed-by: Dave Cheney <dave@cheney.net>
Run-TryBot: Dave Cheney <dave@cheney.net>

10 years agoruntime: replace per-M workbuf cache with per-P gcWork cache
Austin Clements [Sun, 19 Apr 2015 19:22:20 +0000 (15:22 -0400)]
runtime: replace per-M workbuf cache with per-P gcWork cache

Currently, each M has a cache of the most recently used *workbuf. This
is used primarily by the write barrier so it doesn't have to access
the global workbuf lists on every write barrier. It's also used by
stack scanning because it's convenient.

This cache is important for write barrier performance, but this
particular approach has several downsides. It's faster than no cache,
but far from optimal (as the benchmarks below show). It's complex:
access to the cache is sprinkled through most of the workbuf list
operations and it requires special care to transform into and back out
of the gcWork cache that's actually used for scanning and marking. It
requires atomic exchanges to take ownership of the cached workbuf and
to return it to the M's cache even though it's almost always used by
only the current M. Since it's per-M, flushing these caches is O(# of
Ms), which may be high. And it has some significant subtleties: for
example, in general the cache shouldn't be used after the
harvestwbufs() in mark termination because it could hide work from
mark termination, but stack scanning can happen after this and *will*
use the cache (but it turns out this is okay because it will always be
followed by a getfull(), which drains the cache).

This change replaces this cache with a per-P gcWork object. This
gcWork cache can be used directly by scanning and marking (as long as
preemption is disabled, which is a general requirement of gcWork).
Since it's per-P, it doesn't require synchronization, which simplifies
things and means the only atomic operations in the write barrier are
occasionally fetching new work buffers and setting a mark bit if the
object isn't already marked. This cache can be flushed in O(# of Ps),
which is generally small. It follows a simple flushing rule: the cache
can be used during any phase, but during mark termination it must be
flushed before allowing preemption. This also makes the dispose during
mutator assist no longer necessary, which eliminates the vast majority
of gcWork dispose calls and reduces contention on the global workbuf
lists. And it's a lot faster on some benchmarks:

benchmark                          old ns/op       new ns/op       delta
BenchmarkBinaryTree17              11963668673     11206112763     -6.33%
BenchmarkFannkuch11                2643217136      2649182499      +0.23%
BenchmarkFmtFprintfEmpty           70.4            70.2            -0.28%
BenchmarkFmtFprintfString          364             307             -15.66%
BenchmarkFmtFprintfInt             317             282             -11.04%
BenchmarkFmtFprintfIntInt          512             483             -5.66%
BenchmarkFmtFprintfPrefixedInt     404             380             -5.94%
BenchmarkFmtFprintfFloat           521             479             -8.06%
BenchmarkFmtManyArgs               2164            1894            -12.48%
BenchmarkGobDecode                 30366146        22429593        -26.14%
BenchmarkGobEncode                 29867472        26663152        -10.73%
BenchmarkGzip                      391236616       396779490       +1.42%
BenchmarkGunzip                    96639491        96297024        -0.35%
BenchmarkHTTPClientServer          100110          70763           -29.31%
BenchmarkJSONEncode                51866051        52511382        +1.24%
BenchmarkJSONDecode                103813138       86094963        -17.07%
BenchmarkMandelbrot200             4121834         4120886         -0.02%
BenchmarkGoParse                   16472789        5879949         -64.31%
BenchmarkRegexpMatchEasy0_32       140             140             +0.00%
BenchmarkRegexpMatchEasy0_1K       394             394             +0.00%
BenchmarkRegexpMatchEasy1_32       120             120             +0.00%
BenchmarkRegexpMatchEasy1_1K       621             614             -1.13%
BenchmarkRegexpMatchMedium_32      209             202             -3.35%
BenchmarkRegexpMatchMedium_1K      54889           55175           +0.52%
BenchmarkRegexpMatchHard_32        2682            2675            -0.26%
BenchmarkRegexpMatchHard_1K        79383           79524           +0.18%
BenchmarkRevcomp                   584116718       584595320       +0.08%
BenchmarkTemplate                  125400565       109620196       -12.58%
BenchmarkTimeParse                 386             387             +0.26%
BenchmarkTimeFormat                580             447             -22.93%

(Best out of 10 runs. The delta of averages is similar.)

This also puts us in a good position to flush these caches when
nearing the end of concurrent marking, which will let us increase the
size of the work buffers while still controlling mark termination
pause time.

Change-Id: I2dd94c8517a19297a98ec280203cccaa58792522
Reviewed-on: https://go-review.googlesource.com/9178
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: fix check for pending GC work
Austin Clements [Fri, 24 Apr 2015 00:00:10 +0000 (20:00 -0400)]
runtime: fix check for pending GC work

When findRunnable considers running a fractional mark worker, it first
checks if there's any work to be done; if there isn't there's no point
in running the worker because it will just reschedule immediately.
However, currently findRunnable just checks work.full and
work.partial, whereas getfull can *also* draw work from m.currentwbuf.
As a result, findRunnable may not start a worker even though there
actually is work.

This problem manifests itself in occasional failures of the
test/init1.go test. This test is unusual because it performs a large
amount of allocation without executing any write barriers, which means
there's nothing to force the pointers in currentwbuf out to the
work.partial/full lists where findRunnable can see them.

This change fixes this problem by making findRunnable also check for a
currentwbuf. This aligns findRunnable with trygetfull's notion of
whether or not there's work.

Change-Id: Ic76d22b7b5d040bc4f58a6b5975e9217650e66c4
Reviewed-on: https://go-review.googlesource.com/9299
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: start dedicated mark workers even if there's no work
Austin Clements [Thu, 23 Apr 2015 23:51:03 +0000 (19:51 -0400)]
runtime: start dedicated mark workers even if there's no work

Currently, findRunnable only considers running a mark worker if
there's work in the work queue. In principle, this can delay the start
of the desired number of dedicated mark workers if there's no work
pending. This is unlikely to occur in practice, since there should be
work queued from the scan phase, but if it were to come up, a CPU hog
mutator could slow down or delay garbage collection.

This check makes sense for fractional mark workers, since they'll just
return to the scheduler immediately if there's no work, but we want
the scheduler to start all of the dedicated mark workers promptly,
even if there's currently no queued work. Hence, this change moves the
pending work check after the check for starting a dedicated worker.

Change-Id: I52b851cc9e41f508a0955b3f905ca80f109ea101
Reviewed-on: https://go-review.googlesource.com/9298
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agoruntime: fix some out-of-date comments
Austin Clements [Thu, 23 Apr 2015 22:55:23 +0000 (18:55 -0400)]
runtime: fix some out-of-date comments

bgMarkCount no longer exists.

Change-Id: I3aa406fdccfca659814da311229afbae55af8304
Reviewed-on: https://go-review.googlesource.com/9297
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agomisc/cgo/testcshared: make test.bash resilient against noise.
Hyang-Ah Hana Kim [Fri, 24 Apr 2015 18:11:29 +0000 (14:11 -0400)]
misc/cgo/testcshared: make test.bash resilient against noise.

Instead of comparing against the entire output that may include
verbose warning messages, use the last line of the output and check
it includes the expected success message (PASS).

Change-Id: Iafd583ee5529a8aef5439b9f1f6ce0185e4b1331
Reviewed-on: https://go-review.googlesource.com/9304
Reviewed-by: David Crawshaw <crawshaw@golang.org>
10 years agocmd/go: rename doc.go to alldocs.go in preparation for "go doc"
Rob Pike [Fri, 24 Apr 2015 16:10:44 +0000 (09:10 -0700)]
cmd/go: rename doc.go to alldocs.go in preparation for "go doc"

Also rename and update mkdoc.sh to mkalldocs.sh

Change-Id: Ief3673c22d45624e173fc65ee279cea324da03b5
Reviewed-on: https://go-review.googlesource.com/9226
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: implement xadduintptr and update system mstats using it
Srdjan Petrovic [Thu, 16 Apr 2015 21:32:18 +0000 (14:32 -0700)]
runtime: implement xadduintptr and update system mstats using it

The motivation is that sysAlloc/Free() currently aren't safe to be
called without a valid G, because arm's xadd64() uses locks that require
a valid G.

The solution here was proposed by Dmitry Vyukov: use xadduintptr()
instead of xadd64(), until arm can support xadd64 on all of its
architectures (not a trivial task for arm).

Change-Id: I250252079357ea2e4360e1235958b1c22051498f
Reviewed-on: https://go-review.googlesource.com/9002
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
10 years agomisc/cgo/testcshared: add a c-shared test for android/arm.
Hyang-Ah Hana Kim [Thu, 23 Apr 2015 21:27:38 +0000 (17:27 -0400)]
misc/cgo/testcshared: add a c-shared test for android/arm.

- main3.c tests main.main is exported when compiled for GOOS=android.
- wait longer for main2.c (it's slow on android/arm)
- rearranged test.bash

Fixes #10070.

Change-Id: I6e5a98d1c5fae776afa54ecb5da633b59b269316
Reviewed-on: https://go-review.googlesource.com/9296
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: Hyang-Ah Hana Kim <hyangah@gmail.com>

10 years agocmd/internal/gc, cmd/internal/ld, cmd/internal/obj: teach compiler about local symbols
Michael Hudson-Doyle [Fri, 17 Apr 2015 20:14:08 +0000 (08:14 +1200)]
cmd/internal/gc, cmd/internal/ld, cmd/internal/obj: teach compiler about local symbols

This lets us avoid loading string constants via the GOT and (together with
http://golang.org/cl/9102) results in the fannkuch benchmark having very similar
register usage with -dynlink as without.

Change-Id: Ic3892b399074982b76773c3e547cfbba5dabb6f9
Reviewed-on: https://go-review.googlesource.com/9103
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>

10 years agoruntime: simplify process for starting GC goroutine
Austin Clements [Wed, 22 Apr 2015 19:05:08 +0000 (15:05 -0400)]
runtime: simplify process for starting GC goroutine

Currently, when allocation reaches the GC trigger, the runtime uses
readyExecute to start the GC goroutine immediately rather than wait
for the scheduler to get around to the GC goroutine while the mutator
continues to grow the heap.

Now that the scheduler runs the most recently readied goroutine when a
goroutine yields its time slice, this rigmarole is no longer
necessary. The runtime can simply ready the GC goroutine and yield
from the readying goroutine.

Change-Id: I3b4ebadd2a72a923b1389f7598f82973dd5c8710
Reviewed-on: https://go-review.googlesource.com/9292
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

10 years agoruntime: use park/ready to wake up GC at end of concurrent mark
Austin Clements [Wed, 22 Apr 2015 21:44:36 +0000 (17:44 -0400)]
runtime: use park/ready to wake up GC at end of concurrent mark

Currently, the main GC goroutine sleeps on a note during concurrent
mark and the first background mark worker or assist to finish marking
use wakes up that note to let the main goroutine proceed into mark
termination. Unfortunately, the latency of this wakeup can be quite
high, since the GC goroutine will typically have lost its P while in
the futex sleep, meaning it will be placed on the global run queue and
will wait there until some P is kind enough to pick it up. This delay
gives the mutator more time to allocate and create floating garbage,
growing the heap unnecessarily. Worse, it's likely that background
marking has stopped at this point (unless GOMAXPROCS>4), so anything
that's allocated and published to the heap during this window will
have to be scanned during mark termination while the world is stopped.

This change replaces the note sleep/wakeup with a gopark/ready
scheme. This keeps the wakeup inside the Go scheduler and lets the
garbage collector take advantage of the new scheduler semantics that
run the ready()d goroutine immediately when the ready()ing goroutine
sleeps.

For the json benchmark from x/benchmarks with GOMAXPROCS=4, this
reduces the delay in waking up the GC goroutine and entering mark
termination once concurrent marking is done from ~100ms to typically
<100µs.

Change-Id: Ib11f8b581b8914f2d68e0094f121e49bac3bb384
Reviewed-on: https://go-review.googlesource.com/9291
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: use timer for GC control revise rather than timeout
Austin Clements [Wed, 22 Apr 2015 20:35:45 +0000 (16:35 -0400)]
runtime: use timer for GC control revise rather than timeout

Currently, we use a note sleep with a timeout in a loop in func gc to
periodically revise the GC control variables. Replace this with a
fully blocking note sleep and use a periodic timer to trigger the
revise instead. This is a step toward replacing the note sleep in func
gc.

Change-Id: I2d562f6b9b2e5f0c28e9a54227e2c0f8a2603f63
Reviewed-on: https://go-review.googlesource.com/9290
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: yield time slice to most recently readied G
Austin Clements [Wed, 22 Apr 2015 18:42:26 +0000 (14:42 -0400)]
runtime: yield time slice to most recently readied G

Currently, when the runtime ready()s a G, it adds it to the end of the
current P's run queue and continues running. If there are many other
things in the run queue, this can result in a significant delay before
the ready()d G actually runs and can hurt fairness when other Gs in
the run queue are CPU hogs. For example, if there are three Gs sharing
a P, one of which is a CPU hog that never voluntarily gives up the P
and the other two of which are doing small amounts of work and
communicating back and forth on an unbuffered channel, the two
communicating Gs will get very little CPU time.

Change this so that when G1 ready()s G2 and then blocks, the scheduler
immediately hands off the remainder of G1's time slice to G2. In the
above example, the two communicating Gs will now act as a unit and
together get half of the CPU time, while the CPU hog gets the other
half of the CPU time.

This fixes the problem demonstrated by the ping-pong benchmark added
in the previous commit:

benchmark                old ns/op     new ns/op     delta
BenchmarkPingPongHog     684287        825           -99.88%

On the x/benchmarks suite, this change improves the performance of
garbage by ~6% (for GOMAXPROCS=1 and 4), and json by 28% and 36% for
GOMAXPROCS=1 and 4. It has negligible effect on heap size.

This has no effect on the go1 benchmark suite since those benchmarks
are mostly single-threaded.

Change-Id: I858a08eaa78f702ea98a5fac99d28a4ac91d339f
Reviewed-on: https://go-review.googlesource.com/9289
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: benchmark for ping-pong in the presence of a CPU hog
Austin Clements [Wed, 22 Apr 2015 20:38:04 +0000 (16:38 -0400)]
runtime: benchmark for ping-pong in the presence of a CPU hog

This benchmark demonstrates a current problem with the scheduler where
a set of frequently communicating goroutines get very little CPU time
in the presence of another goroutine that hogs that CPU, even if one
of those communicating goroutines is always runnable.

Currently it takes about 0.5 milliseconds to switch between
ping-ponging goroutines in the presence of a CPU hog:

BenchmarkPingPongHog     2000     684287 ns/op

Change-Id: I278848c84f778de32344921ae8a4a8056e4898b0
Reviewed-on: https://go-review.googlesource.com/9288
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: factor checking if P run queue is empty
Austin Clements [Wed, 22 Apr 2015 16:18:01 +0000 (12:18 -0400)]
runtime: factor checking if P run queue is empty

There are a variety of places where we check if a P's run queue is
empty. This test is about to get slightly more complicated, so factor
it out into a new function, runqempty. This function is inlinable, so
this has no effect on performance.

Change-Id: If4a0b01ffbd004937de90d8d686f6ded4aad2c6b
Reviewed-on: https://go-review.googlesource.com/9287
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agocmd/internal/gc: add and test write barrier debug output
Russ Cox [Fri, 17 Apr 2015 04:34:18 +0000 (00:34 -0400)]
cmd/internal/gc: add and test write barrier debug output

We can expand the test cases as we discover problems.
This is some basic tests plus all the things I got wrong
in some recent work.

Change-Id: Id875fcfaf74eb087ae42b441fe47a34c5b8ccb39
Reviewed-on: https://go-review.googlesource.com/9158
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
10 years agohash/crc32: clarify documentation
Aamir Khan [Wed, 22 Apr 2015 10:24:48 +0000 (19:24 +0900)]
hash/crc32: clarify documentation

Explicitly specify that we represent polynomial in reversed notation

Fixes #8229

Change-Id: Idf094c01fd82f133cd0c1b50fa967d12c577bdb5
Reviewed-on: https://go-review.googlesource.com/9237
Reviewed-by: David Chase <drchase@google.com>
10 years agocmd/dist: allow $GO_TEST_TIMEOUT_SCALE to override timeoutScale
Shenghou Ma [Thu, 23 Apr 2015 06:16:31 +0000 (02:16 -0400)]
cmd/dist: allow $GO_TEST_TIMEOUT_SCALE to override timeoutScale

Some machines are so slow that even with the default timeoutScale,
they still timeout some tests. For example, currently some linux/arm
builders and the openbsd/arm builder are timing out the runtime
test and CL 8397 was proposed to skip some tests on openbsd/arm
to fix the build.

Instead of increasing timeoutScale or skipping tests, this CL
introduces an environment variable $GO_TEST_TIMEOUT_SCALE that
could be set to manually set a larger timeoutScale for those
machines/builders.

Fixes #10314.

Change-Id: I16c9a9eb980d6a63309e4cacd79eee2fe05769ee
Reviewed-on: https://go-review.googlesource.com/9223
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
10 years agoruntime: signal forwarding
Srdjan Petrovic [Thu, 9 Apr 2015 18:12:12 +0000 (11:12 -0700)]
runtime: signal forwarding

Forward signals to signal handlers installed before Go installs its own,
under certain circumstances.  In particular, as iant@ suggests, signals are
forwarded iff:
   (1) a non-SIG_DFL signal handler existed before Go, and
   (2) signal is synchronous (i.e., one of SIGSEGV, SIGBUS, SIGFPE), and
    (3a) signal occured on a non-Go thread, or
    (3b) signal occurred on a Go thread but in CGo code.

Supported only on Linux, for now.

Change-Id: I403219ee47b26cf65da819fb86cf1ec04d3e25f5
Reviewed-on: https://go-review.googlesource.com/8712
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agoencoding/base64: Optimize EncodeToString and DecodeString.
Egon Elbre [Mon, 13 Apr 2015 10:21:00 +0000 (13:21 +0300)]
encoding/base64: Optimize EncodeToString and DecodeString.

benchmark                   old ns/op     new ns/op     delta
BenchmarkEncodeToString     31281         23821         -23.85%
BenchmarkDecodeString       156508        82254         -47.44%

benchmark                   old MB/s     new MB/s     speedup
BenchmarkEncodeToString     261.88       343.89       1.31x
BenchmarkDecodeString       69.80        132.81       1.90x

Change-Id: I115e0b18c3a6d5ef6bfdcb3f637644f02f290907
Reviewed-on: https://go-review.googlesource.com/8808
Reviewed-by: Nigel Tao <nigeltao@golang.org>
10 years agocmd/9g, etc: remove // fallthrough comments
Josh Bleecher Snyder [Thu, 23 Apr 2015 23:36:52 +0000 (16:36 -0700)]
cmd/9g, etc: remove // fallthrough comments

They are vestiges of the c2go transition.

Change-Id: I22672e40373ef77d7a0bf69cfff8017e46353055
Reviewed-on: https://go-review.googlesource.com/9265
Reviewed-by: Minux Ma <minux@golang.org>
10 years agomath/big: add partial arm64 assembly support
Josh Bleecher Snyder [Tue, 21 Apr 2015 17:33:30 +0000 (10:33 -0700)]
math/big: add partial arm64 assembly support

benchmark                       old ns/op      new ns/op      delta
BenchmarkAddVV_1                18.7           14.8           -20.86%
BenchmarkAddVV_2                21.8           16.6           -23.85%
BenchmarkAddVV_3                26.1           17.1           -34.48%
BenchmarkAddVV_4                30.4           21.9           -27.96%
BenchmarkAddVV_5                35.5           19.8           -44.23%
BenchmarkAddVV_1e1              63.0           28.3           -55.08%
BenchmarkAddVV_1e2              593            178            -69.98%
BenchmarkAddVV_1e3              5691           1490           -73.82%
BenchmarkAddVV_1e4              56868          20761          -63.49%
BenchmarkAddVV_1e5              569062         207679         -63.51%
BenchmarkAddVW_1                15.8           12.6           -20.25%
BenchmarkAddVW_2                17.8           13.1           -26.40%
BenchmarkAddVW_3                21.2           13.9           -34.43%
BenchmarkAddVW_4                23.6           14.7           -37.71%
BenchmarkAddVW_5                26.0           15.8           -39.23%
BenchmarkAddVW_1e1              41.3           21.6           -47.70%
BenchmarkAddVW_1e2              383            145            -62.14%
BenchmarkAddVW_1e3              3703           1264           -65.87%
BenchmarkAddVW_1e4              36920          14359          -61.11%
BenchmarkAddVW_1e5              370345         143046         -61.37%
BenchmarkAddMulVVW_1            33.2           32.5           -2.11%
BenchmarkAddMulVVW_2            58.0           57.2           -1.38%
BenchmarkAddMulVVW_3            95.2           93.9           -1.37%
BenchmarkAddMulVVW_4            108            106            -1.85%
BenchmarkAddMulVVW_5            159            156            -1.89%
BenchmarkAddMulVVW_1e1          344            340            -1.16%
BenchmarkAddMulVVW_1e2          3644           3624           -0.55%
BenchmarkAddMulVVW_1e3          37344          37208          -0.36%
BenchmarkAddMulVVW_1e4          373295         372170         -0.30%
BenchmarkAddMulVVW_1e5          3438116        3425606        -0.36%
BenchmarkBitLen0                7.21           4.32           -40.08%
BenchmarkBitLen1                6.49           4.32           -33.44%
BenchmarkBitLen2                7.23           4.32           -40.25%
BenchmarkBitLen3                6.49           4.32           -33.44%
BenchmarkBitLen4                7.22           4.32           -40.17%
BenchmarkBitLen5                6.52           4.33           -33.59%
BenchmarkBitLen8                7.22           4.32           -40.17%
BenchmarkBitLen9                6.49           4.32           -33.44%
BenchmarkBitLen16               8.66           4.32           -50.12%
BenchmarkBitLen17               7.95           4.32           -45.66%
BenchmarkBitLen31               8.69           4.32           -50.29%
BenchmarkGCD10x10               5021           5033           +0.24%
BenchmarkGCD10x100              5571           5572           +0.02%
BenchmarkGCD10x1000             6707           6729           +0.33%
BenchmarkGCD10x10000            13526          13419          -0.79%
BenchmarkGCD10x100000           85668          83242          -2.83%
BenchmarkGCD100x100             24196          23936          -1.07%
BenchmarkGCD100x1000            28802          27309          -5.18%
BenchmarkGCD100x10000           64111          51704          -19.35%
BenchmarkGCD100x100000          385840         274385         -28.89%
BenchmarkGCD1000x1000           262892         236269         -10.13%
BenchmarkGCD1000x10000          371393         277883         -25.18%
BenchmarkGCD1000x100000         1311795        589055         -55.10%
BenchmarkGCD10000x10000         9596740        6123930        -36.19%
BenchmarkGCD10000x100000        16404000       7269610        -55.68%
BenchmarkGCD100000x100000       776660000      419270000      -46.02%
BenchmarkHilbert                13478980       13402270       -0.57%
BenchmarkBinomial               9802           9440           -3.69%
BenchmarkBitset                 142            142            +0.00%
BenchmarkBitsetNeg              328            279            -14.94%
BenchmarkBitsetOrig             853            861            +0.94%
BenchmarkBitsetNegOrig          1489           1444           -3.02%
BenchmarkMul                    420949000      410481000      -2.49%
BenchmarkExp3Power0x10          1148           1229           +7.06%
BenchmarkExp3Power0x40          1322           1376           +4.08%
BenchmarkExp3Power0x100         2437           2486           +2.01%
BenchmarkExp3Power0x400         9456           9346           -1.16%
BenchmarkExp3Power0x1000        113623         108701         -4.33%
BenchmarkExp3Power0x4000        1134933        1101481        -2.95%
BenchmarkExp3Power0x10000       10773570       10396160       -3.50%
BenchmarkExp3Power0x40000       101362100      97788300       -3.53%
BenchmarkExp3Power0x100000      921114000      885249000      -3.89%
BenchmarkExp3Power0x400000      8323094000     7969020000     -4.25%
BenchmarkFibo                   322021600      92554450       -71.26%
BenchmarkScanPi                 1264583        321065         -74.61%
BenchmarkStringPiParallel       1644661        554216         -66.30%
BenchmarkScan10Base2            1111           1080           -2.79%
BenchmarkScan100Base2           6645           6345           -4.51%
BenchmarkScan1000Base2          84084          62405          -25.78%
BenchmarkScan10000Base2         3105998        932551         -69.98%
BenchmarkScan100000Base2        257234800      40113333       -84.41%
BenchmarkScan10Base8            571            573            +0.35%
BenchmarkScan100Base8           2810           2543           -9.50%
BenchmarkScan1000Base8          47383          25834          -45.48%
BenchmarkScan10000Base8         2739518        567203         -79.30%
BenchmarkScan100000Base8        253952400      36495680       -85.63%
BenchmarkScan10Base10           553            556            +0.54%
BenchmarkScan100Base10          2640           2385           -9.66%
BenchmarkScan1000Base10         50865          24049          -52.72%
BenchmarkScan10000Base10        3279916        549313         -83.25%
BenchmarkScan100000Base10       309121000      36213140       -88.29%
BenchmarkScan10Base16           478            483            +1.05%
BenchmarkScan100Base16          2353           2144           -8.88%
BenchmarkScan1000Base16         48091          24246          -49.58%
BenchmarkScan10000Base16        2858886        586475         -79.49%
BenchmarkScan100000Base16       266320000      38190500       -85.66%
BenchmarkString10Base2          736            730            -0.82%
BenchmarkString100Base2         2695           2707           +0.45%
BenchmarkString1000Base2        20549          20388          -0.78%
BenchmarkString10000Base2       212638         210782         -0.87%
BenchmarkString100000Base2      1944963        1938033        -0.36%
BenchmarkString10Base8          524            517            -1.34%
BenchmarkString100Base8         1326           1320           -0.45%
BenchmarkString1000Base8        8213           8249           +0.44%
BenchmarkString10000Base8       72204          72092          -0.16%
BenchmarkString100000Base8      769068         765993         -0.40%
BenchmarkString10Base10         1018           982            -3.54%
BenchmarkString100Base10        3485           3206           -8.01%
BenchmarkString1000Base10       37102          18935          -48.97%
BenchmarkString10000Base10      188633         88637          -53.01%
BenchmarkString100000Base10     124490300      19700940       -84.17%
BenchmarkString10Base16         509            502            -1.38%
BenchmarkString100Base16        1084           1098           +1.29%
BenchmarkString1000Base16       5641           5650           +0.16%
BenchmarkString10000Base16      46900          46745          -0.33%
BenchmarkString100000Base16     508957         505840         -0.61%
BenchmarkLeafSize0              8934320        8149465        -8.78%
BenchmarkLeafSize1              237666         118381         -50.19%
BenchmarkLeafSize2              237807         117854         -50.44%
BenchmarkLeafSize3              1688640        353494         -79.07%
BenchmarkLeafSize4              235676         116196         -50.70%
BenchmarkLeafSize5              2121896        430325         -79.72%
BenchmarkLeafSize6              1682306        351775         -79.09%
BenchmarkLeafSize7              1051847        251436         -76.10%
BenchmarkLeafSize8              232697         115674         -50.29%
BenchmarkLeafSize9              2403616        488443         -79.68%
BenchmarkLeafSize10             2120975        429545         -79.75%
BenchmarkLeafSize11             2023789        426525         -78.92%
BenchmarkLeafSize12             1684830        351985         -79.11%
BenchmarkLeafSize13             1465529        337906         -76.94%
BenchmarkLeafSize14             1050498        253872         -75.83%
BenchmarkLeafSize15             683228         197384         -71.11%
BenchmarkLeafSize16             232496         116026         -50.10%
BenchmarkLeafSize32             245841         126671         -48.47%
BenchmarkLeafSize64             301728         190285         -36.93%

Change-Id: I63e63297896d96b89c9a275b893c2b405a7e105d
Reviewed-on: https://go-review.googlesource.com/9260
Reviewed-by: David Crawshaw <crawshaw@golang.org>
10 years agoruntime: deflake TestNewOSProc0, fix _rt0_amd64_linux_lib stack alignment
Srdjan Petrovic [Wed, 22 Apr 2015 18:44:46 +0000 (11:44 -0700)]
runtime: deflake TestNewOSProc0, fix _rt0_amd64_linux_lib stack alignment

This addresses iant's comments from CL 9164.

Change-Id: I7b5b282f61b11aab587402c2d302697e76666376
Reviewed-on: https://go-review.googlesource.com/9222
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agoruntime: fix underflow in next_gc calculation
Austin Clements [Thu, 23 Apr 2015 17:02:31 +0000 (13:02 -0400)]
runtime: fix underflow in next_gc calculation

Currently, it's possible for the next_gc calculation to underflow.
Since next_gc is unsigned, this wraps around and effectively disables
GC for the rest of the program's execution. Besides being obviously
wrong, this is causing test failures on 32-bit because some tests are
running out of heap.

This underflow happens for two reasons, both having to do with how we
estimate the reachable heap size at the end of the GC cycle.

One reason is that this calculation depends on the value of heap_live
at the beginning of the GC cycle, but we currently only record that
value during a concurrent GC and not during a forced STW GC. Fix this
by moving the recorded value from gcController to work and recording
it on a common code path.

The other reason is that we use the amount of allocation during the GC
cycle as an approximation of the amount of floating garbage and
subtract it from the marked heap to estimate the reachable heap.
However, since this is only an approximation, it's possible for the
amount of allocation during the cycle to be *larger* than the marked
heap size (since the runtime allocates white and it's possible for
these allocations to never be made reachable from the heap). Currently
this causes wrap-around in our estimate of the reachable heap size,
which in turn causes wrap-around in next_gc. Fix this by bottoming out
the reachable heap estimate at 0, in which case we just fall back to
triggering GC at heapminimum (which is okay since this only happens on
small heaps).

Fixes #10555, fixes #10556, and fixes #10559.

Change-Id: Iad07b529c03772356fede2ae557732f13ebfdb63
Reviewed-on: https://go-review.googlesource.com/9286
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agoruntime: Improve scanning performance
Rick Hudson [Wed, 22 Apr 2015 19:06:35 +0000 (15:06 -0400)]
runtime: Improve scanning performance

To achieve a 2% improvement in the garbage benchmark this CL removes
an unneeded assert and avoids one hbits.next() call per object
being scanned.

Change-Id: Ibd542d01e9c23eace42228886f9edc488354df0d
Reviewed-on: https://go-review.googlesource.com/9244
Reviewed-by: Austin Clements <austin@google.com>
10 years agoruntime: disable TestNewOSProc0 on android/arm.
Hyang-Ah Hana Kim [Thu, 23 Apr 2015 17:55:56 +0000 (13:55 -0400)]
runtime: disable TestNewOSProc0 on android/arm.

newosproc0 does not work on android/arm.
See issue #10548.

Change-Id: Ieaf6f5d0b77cddf5bf0b6c89fd12b1c1b8723f9b
Reviewed-on: https://go-review.googlesource.com/9293
Reviewed-by: David Crawshaw <crawshaw@golang.org>
10 years agoimage/png: don't silently swallow io.ReadFull's io.EOF error when it
Nigel Tao [Thu, 23 Apr 2015 04:14:31 +0000 (14:14 +1000)]
image/png: don't silently swallow io.ReadFull's io.EOF error when it
lands exactly on an IDAT row boundary.

Fixes #10493

Change-Id: I12be7c5bdcde7032e17ed1d4400db5f17c72bc87
Reviewed-on: https://go-review.googlesource.com/9270
Reviewed-by: Rob Pike <r@golang.org>
10 years agodoc/faq: replace reference to goven with gomvpkg
Dmitry Savintsev [Fri, 17 Apr 2015 19:58:22 +0000 (21:58 +0200)]
doc/faq: replace reference to goven with gomvpkg

github.com/kr/goven says it's deprecated and anyway
it would be preferable to point users to a standard Go tool.

Change-Id: Iac4a0d13233604a36538748d498f5770b2afce19
Reviewed-on: https://go-review.googlesource.com/8969
Reviewed-by: Minux Ma <minux@golang.org>
10 years agonet: use Go's DNS resolver when system configuration permits
Brad Fitzpatrick [Thu, 16 Apr 2015 21:33:25 +0000 (14:33 -0700)]
net: use Go's DNS resolver when system configuration permits

If the machine's network configuration files (resolv.conf,
nsswitch.conf) don't have any unsupported options, prefer Go's DNS
resolver, which doesn't have the cgo & thread over.

It means users can have more than 500 DNS requests outstanding (our
current limit for cgo lookups) and not have one blocked thread per
outstanding request.

Discussed in thread https://groups.google.com/d/msg/golang-dev/2ZUi792oztM/Q0rg_DkF5HMJ

Change-Id: I3f685d70aff6b47bec30b63e9fba674b20507f95
Reviewed-on: https://go-review.googlesource.com/8945
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

10 years agocmd/internal/gc: remove /*untyped*/ comments
Josh Bleecher Snyder [Thu, 23 Apr 2015 01:04:23 +0000 (18:04 -0700)]
cmd/internal/gc: remove /*untyped*/ comments

They are vestiges of the c2go translation.

Change-Id: I9a10536f5986b751a35cc7d84b5ba69ae0c2ede7
Reviewed-on: https://go-review.googlesource.com/9262
Reviewed-by: Minux Ma <minux@golang.org>
10 years agoimage/jpeg: have the LargeImageWithShortData test only allocate 64 MiB, not 604
Nigel Tao [Wed, 22 Apr 2015 23:13:32 +0000 (09:13 +1000)]
image/jpeg: have the LargeImageWithShortData test only allocate 64 MiB, not 604
MiB.

Fixes #10531

Change-Id: I9eece86837c3df2b1f7df315d5ec94bd3ede3eec
Reviewed-on: https://go-review.googlesource.com/9238
Run-TryBot: Nigel Tao <nigeltao@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
10 years agoruntime: fix build after CL 9164 on Linux
Shenghou Ma [Wed, 22 Apr 2015 22:03:18 +0000 (18:03 -0400)]
runtime: fix build after CL 9164 on Linux

There is an assumption that the function executed in child thread
created by runtime.close should not return. And different systems
enforce that differently: some exit that thread, some exit the
whole process.

The test TestNewOSProc0 introduced in CL 9161 breaks that assumption,
so we need to adjust the code to only exit the thread should the
called function return.

Change-Id: Id631cb2f02ec6fbd765508377a79f3f96c6a2ed6
Reviewed-on: https://go-review.googlesource.com/9246
Reviewed-by: Dave Cheney <dave@cheney.net>
10 years agolog/syslog: make the BUG notes visible on golang.org
Shenghou Ma [Wed, 22 Apr 2015 20:27:59 +0000 (16:27 -0400)]
log/syslog: make the BUG notes visible on golang.org

It was only visible when you run godoc with explicit GOOS=windows,
which is less useful for people developing portable application on
non-windows platforms.

Also added a note that log/syslog is not supported on NaCl.

Change-Id: I81650445fb2a5ee161da7e0608c3d3547d5ac2a6
Reviewed-on: https://go-review.googlesource.com/9245
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agocmd/link, cmd/internal/goobj: update constants, regenerate testdata
Michael Hudson-Doyle [Sun, 19 Apr 2015 23:53:20 +0000 (11:53 +1200)]
cmd/link, cmd/internal/goobj: update constants, regenerate testdata

The constants in cmd/internal/goobj had gone stale (we had three copies of
these constants, working on reducing that was what got me to noticing this).

Some of the changes to link.hello.darwin.amd64 are the change from absolute
to %rip-relative addressing, a change which happened quite a while ago...

Depends on http://golang.org/cl/9113.

Fixes #10501.

Change-Id: Iaa1511f458a32228c2df2ccd0076bb9ae212a035
Reviewed-on: https://go-review.googlesource.com/9105
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

10 years agoruntime: use reachable heap estimate to set trigger/goal
Austin Clements [Tue, 21 Apr 2015 18:24:25 +0000 (14:24 -0400)]
runtime: use reachable heap estimate to set trigger/goal

Currently, we set the heap goal for the next GC cycle using the size
of the marked heap at the end of the current cycle. This can lead to a
bad feedback loop if the mutator is rapidly allocating and releasing
pointers that can significantly bloat heap size.

If the GC were STW, the marked heap size would be exactly the
reachable heap size (call it stwLive). However, in concurrent GC,
marked=stwLive+floatLive, where floatLive is the amount of "floating
garbage": objects that were reachable at some point during the cycle
and were marked, but which are no longer reachable by the end of the
cycle. If the GC cycle is short, then the mutator doesn't have much
time to create floating garbage, so marked≈stwLive. However, if the GC
cycle is long and the mutator is allocating and creating floating
garbage very rapidly, then it's possible that marked≫stwLive. Since
the runtime currently sets the heap goal based on marked, this will
cause it to set a high heap goal. This means that 1) the next GC cycle
will take longer because of the larger heap and 2) the assist ratio
will be low because of the large distance between the trigger and the
goal. The combination of these lets the mutator produce even more
floating garbage in the next cycle, which further exacerbates the
problem.

For example, on the garbage benchmark with GOMAXPROCS=1, this causes
the heap to grow to ~500MB and the garbage collector to retain upwards
of ~300MB of heap, while the true reachable heap size is ~32MB. This,
in turn, causes the GC cycle to take upwards of ~3 seconds.

Fix this bad feedback loop by estimating the true reachable heap size
(stwLive) and using this rather than the marked heap size
(stwLive+floatLive) as the basis for the GC trigger and heap goal.
This breaks the bad feedback loop and causes the mutator to assist
more, which decreases the rate at which it can create floating
garbage. On the same garbage benchmark, this reduces the maximum heap
size to ~73MB, the retained heap to ~40MB, and the duration of the GC
cycle to ~200ms.

Change-Id: I7712244c94240743b266f9eb720c03802799cdd1
Reviewed-on: https://go-review.googlesource.com/9177
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agocmd/go: refactor creation of top-level actions for -buildmode=shared
Michael Hudson-Doyle [Tue, 14 Apr 2015 10:03:21 +0000 (12:03 +0200)]
cmd/go: refactor creation of top-level actions for -buildmode=shared

Change-Id: I429402dd91243cd9415b054ee17bfebccc68ed57
Reviewed-on: https://go-review.googlesource.com/9197
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>

10 years agoruntime: include heap goal in gctrace line
Austin Clements [Tue, 21 Apr 2015 20:38:38 +0000 (16:38 -0400)]
runtime: include heap goal in gctrace line

This may or may not be useful to the end user, but it's incredibly
useful for us to understand the behavior of the pacer. Currently this
is fairly easy (though not trivial) to derive from the other heap
stats we print, but we're about to change how we compute the goal,
which will make it much harder to derive.

Change-Id: I796ef233d470c01f606bd9929820c01ece1f585a
Reviewed-on: https://go-review.googlesource.com/9176
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agoruntime: avoid divide-by-zero in GC trigger controller
Austin Clements [Tue, 21 Apr 2015 17:46:54 +0000 (13:46 -0400)]
runtime: avoid divide-by-zero in GC trigger controller

The trigger controller computes GC CPU utilization by dividing by the
wall-clock time that's passed since concurrent mark began. Since this
delta is nanoseconds it's borderline impossible for it to be zero, but
if it is zero we'll currently divide by zero. Be robust to this
possibility by ignoring the utilization in the error term if no time
has elapsed.

Change-Id: I93dfc9e84735682af3e637f6538d1e7602634f09
Reviewed-on: https://go-review.googlesource.com/9175
Reviewed-by: Rick Hudson <rlh@golang.org>
10 years agocmd/internal/gc, cmd/internal/ld: fixes for global vars of types from other modules
Michael Hudson-Doyle [Wed, 1 Apr 2015 03:20:44 +0000 (16:20 +1300)]
cmd/internal/gc, cmd/internal/ld: fixes for global vars of types from other modules

To make the gcprog for global data containing variables of types defined in other shared
libraries, we need to know a lot about those types. So read the value of any symbol with
a name starting with "type.". If a type uses a mask, the name of the symbol defining the
mask unfortunately cannot be predicted from the type name so I have to keep track of the
addresses of every such symbol and associate them with the type symbols after the fact.

I'm not very happy about this change, but something like this is needed and this is as
pleasant as I know how to make it.

Change-Id: I408d831b08b3b31e0610688c41367b23998e975c
Reviewed-on: https://go-review.googlesource.com/8334
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>

10 years agocmd/5g, etc, cmd/internal/gc, cmd/internal/obj, etc: coalesce bool2int implementations
Michael Hudson-Doyle [Wed, 22 Apr 2015 00:41:14 +0000 (12:41 +1200)]
cmd/5g, etc, cmd/internal/gc, cmd/internal/obj, etc: coalesce bool2int implementations

There were 10 implementations of the trivial bool2int function, 9 of which
were the only thing in their file.  Remove all of them in favor of one in
cmd/internal/obj.

Change-Id: I9c51d30716239df51186860b9842a5e9b27264d3
Reviewed-on: https://go-review.googlesource.com/9230
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agogo/constants: rename go/exact to go/constants
Alan Donovan [Wed, 22 Apr 2015 15:25:29 +0000 (11:25 -0400)]
go/constants: rename go/exact to go/constants

since the "precision" parameter means constant arithmetic is not
necessarily exact.

As requested by gri, within go/types, the local import name 'exact'
has been kept, to reduce the diff with the x/tools branch.  This may
be changed later.

Since the go/types.bash script was already obsolete, I added a comment
to this effect.

Tested with all.bash.

Change-Id: I45153688d9d8afa8384fb15229b0124c686059b4
Reviewed-on: https://go-review.googlesource.com/9242
Reviewed-by: Rob Pike <r@golang.org>
10 years agoruntime: merge clone0 and clone
Srdjan Petrovic [Sat, 18 Apr 2015 00:27:07 +0000 (17:27 -0700)]
runtime: merge clone0 and clone

We initially added clone0 to handle the case when G or M don't exist, but
it turns out that we could have just modified clone.  (It also helps that
the function we're invoking in clone0 no longer needs arguments.)

As a side-effect, newosproc0 is now supported on all linux archs.

Change-Id: Ie603af75d8f164310fc16446052d83743961f3ca
Reviewed-on: https://go-review.googlesource.com/9164
Reviewed-by: David Crawshaw <crawshaw@golang.org>
10 years agogo/exact: future-proof API: permit setting precision limit
Robert Griesemer [Tue, 21 Apr 2015 19:47:11 +0000 (12:47 -0700)]
go/exact: future-proof API: permit setting precision limit

Added a prec parameter to MakeFromLiteral (which currently must
always be 0). This will permit go/types to provide an upper limit
for the precision of constant values, eventually. Overflows can be
returned with a special Overflow value (very much like the current
Unknown values).

This is a minimal change that should prevent the need for future
backward-incompatible API changes.

Change-Id: I6c9390d7cc4810375e26c53ed3bde5a383392330
Reviewed-on: https://go-review.googlesource.com/9168
Run-TryBot: Robert Griesemer <gri@golang.org>
Reviewed-by: Alan Donovan <adonovan@google.com>
10 years agonet/http: fix race between dialing and canceling
Daniel Morsing [Mon, 20 Apr 2015 22:02:07 +0000 (23:02 +0100)]
net/http: fix race between dialing and canceling

In the brief window between getConn and persistConn.roundTrip,
a cancel could end up going missing.

Fix by making it possible to inspect if a cancel function was cleared
and checking if we were canceled before entering roundTrip.

Fixes #10511

Change-Id: If6513e63fbc2edb703e36d6356ccc95a1dc33144
Reviewed-on: https://go-review.googlesource.com/9181
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
10 years agonet/http: make ServeContent errors return more specific HTTP status codes
Brad Fitzpatrick [Tue, 21 Apr 2015 20:36:44 +0000 (13:36 -0700)]
net/http: make ServeContent errors return more specific HTTP status codes

Previously all errors were 404 errors, even if the real error had
nothing to do with a file being non-existent.

Fixes #10283

Change-Id: I5b08b471a9064c347510cfcf8557373704eef7c0
Reviewed-on: https://go-review.googlesource.com/9200
Reviewed-by: Daniel Morsing <daniel.morsing@gmail.com>
10 years agonet/http: fix rare Transport readLoop goroutine leak
Brad Fitzpatrick [Tue, 21 Apr 2015 20:13:16 +0000 (13:13 -0700)]
net/http: fix rare Transport readLoop goroutine leak

There used to be a small window where if a server declared it would do
a keep-alive connection but then actually closed the connection before
the roundTrip goroutine scheduled after being sent a response from the
readLoop goroutine, then the readLoop goroutine would loop around and
block forever reading from a channel because the numExpectedResponses
accounting was done too late.

Fixes #10457

Change-Id: Icbae937ffe83c792c295b7f4fb929c6a24a4f759
Reviewed-on: https://go-review.googlesource.com/9169
Reviewed-by: Daniel Morsing <daniel.morsing@gmail.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

10 years agoruntime: fix more vet reported issues
Shenghou Ma [Wed, 22 Apr 2015 01:43:25 +0000 (21:43 -0400)]
runtime: fix more vet reported issues

Change-Id: Ie8dfdb592ee0bfc736d08c92c3d8413a37b6ac03
Reviewed-on: https://go-review.googlesource.com/9241
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agoruntime: check error codes for arm64 system calls
Keith Randall [Tue, 21 Apr 2015 22:47:58 +0000 (15:47 -0700)]
runtime: check error codes for arm64 system calls

Unlike linux arm32, linux arm64 does not set the condition codes to indicate
whether a system call failed or not.  We must check if the return value
is in the error code range (the same as amd64 does).

Fixes runtime.TestBadOpen test.

Change-Id: I97a8b0a17b5f002a3215c535efa91d199cee3309
Reviewed-on: https://go-review.googlesource.com/9220
Reviewed-by: Russ Cox <rsc@golang.org>
10 years agoruntime: fix arm64 asm vet issues
Josh Bleecher Snyder [Tue, 21 Apr 2015 23:15:05 +0000 (16:15 -0700)]
runtime: fix arm64 asm vet issues

Several naming changes and a real issue in asmcgocall_errno.

Change-Id: Ieb0a328a168819fe233d74e0397358384d7e71b3
Reviewed-on: https://go-review.googlesource.com/9212
Reviewed-by: Minux Ma <minux@golang.org>
10 years agonet: replace server tests
Mikio Hara [Tue, 21 Apr 2015 13:10:09 +0000 (22:10 +0900)]
net: replace server tests

This change replaces server tests with new ones that require features
introduced after go1 release, such as runtime-integrated network poller,
Dialer, etc.

Change-Id: Icf1f94f08f33caacd499cfccbe74cda8d05eed30
Reviewed-on: https://go-review.googlesource.com/9195
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agoimage/jpeg: ensure that we can't unread a byte if we didn't read a byte.
Nigel Tao [Fri, 17 Apr 2015 06:48:55 +0000 (16:48 +1000)]
image/jpeg: ensure that we can't unread a byte if we didn't read a byte.

Fixes #10413

Change-Id: I7a4ecd042c40f786ea7406c670d561b1c1179bf0
Reviewed-on: https://go-review.googlesource.com/8998
Reviewed-by: Rob Pike <r@golang.org>
10 years agonet: deflake zero byte IO tests on datagram
Mikio Hara [Mon, 20 Apr 2015 14:15:00 +0000 (23:15 +0900)]
net: deflake zero byte IO tests on datagram

This change deflakes zero byte read/write tests on datagram sockets, and
enables them by default.

Change-Id: I52f1a76f8ff379d90f40a07bb352fae9343ea41a
Reviewed-on: https://go-review.googlesource.com/9194
Reviewed-by: Ian Lance Taylor <iant@golang.org>
10 years agonet: fix WriteTo on Plan 9
Mikio Hara [Tue, 21 Apr 2015 14:15:12 +0000 (23:15 +0900)]
net: fix WriteTo on Plan 9

This change excludes internal UDP header size from a result of number of
bytes written on WriteTo.

Change-Id: I847d57f7f195657b6f14efdf1b4cfab13d4490dd
Reviewed-on: https://go-review.googlesource.com/9196
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: David du Colombier <0intro@gmail.com>
10 years agomath/big: test that subVW and addVW work with arbitrary y
Josh Bleecher Snyder [Tue, 21 Apr 2015 22:43:44 +0000 (15:43 -0700)]
math/big: test that subVW and addVW work with arbitrary y

Fixes #10525.

Change-Id: I92dc87f5d6db396d8dde2220fc37b7093b772d81
Reviewed-on: https://go-review.googlesource.com/9210
Reviewed-by: Robert Griesemer <gri@golang.org>
10 years agomisc/cgo/testcshared: add c-shared test with no exports
Ian Lance Taylor [Tue, 21 Apr 2015 17:46:29 +0000 (10:46 -0700)]
misc/cgo/testcshared: add c-shared test with no exports

The purpose of this test is to make sure that -buildmode=c-shared
works even when the shared library can be built without invoking cgo.

Change-Id: Id6f95af755992b209aff770440ca9819b74113ab
Reviewed-on: https://go-review.googlesource.com/9166
Reviewed-by: David Crawshaw <crawshaw@golang.org>
10 years agoRevert "go/internal/gcimporter: populate (*types.Package).Imports"
Alan Donovan [Tue, 21 Apr 2015 19:43:18 +0000 (19:43 +0000)]
Revert "go/internal/gcimporter: populate (*types.Package).Imports"

This reverts commit 8d7d02f14525874eafb6b0a00916bdb0bc24bc03.

Reverted because it breaks go/build's "deps" test.

Change-Id: I61db6b2431b3ba0d2b3ece5bab7a04194239c34b
Reviewed-on: https://go-review.googlesource.com/9174
Reviewed-by: Alan Donovan <adonovan@google.com>
10 years agogo/internal/gcimporter: populate (*types.Package).Imports
Alan Donovan [Wed, 15 Apr 2015 18:43:29 +0000 (14:43 -0400)]
go/internal/gcimporter: populate (*types.Package).Imports

This is an upstream change to the tools repo:
https://go-review.googlesource.com/#/c/8924/

Change-Id: I01fb1b2e9ec834354994c544f65c8ec8267c9626
Reviewed-on: https://go-review.googlesource.com/8954
Run-TryBot: Robert Griesemer <gri@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
10 years agocmd/go: depend on runtime/cgo if external linking mode is forced
Ian Lance Taylor [Mon, 20 Apr 2015 18:47:08 +0000 (11:47 -0700)]
cmd/go: depend on runtime/cgo if external linking mode is forced

In external linking mode, the linker automatically imports
runtime/cgo.  When the user uses non-standard compilation options,
they have to know to run go install runtime/cgo.  When the go tool
adds non-standard compilation options itself, we can't force the user
to do that.  So add the dependency ourselves.

Bad news: we don't currently have a clean way to know whether we are
going to use external linking mode.  This CL duplicates logic split
between cmd/6l and cmd/internal/ld.

Good news: adding an unnecessary dependency on runtime/cgo does no
real harm.  We aren't going to force the linker to pull it in, we're
just going to build it so that its available if the linker wants it.

Change-Id: Ide676339d4e8b1c3d9792884a2cea921abb281b7
Reviewed-on: https://go-review.googlesource.com/9115
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>

10 years agoreflect: use arrayAt consistently
Sebastien Binet [Tue, 27 Jan 2015 09:04:11 +0000 (10:04 +0100)]
reflect: use arrayAt consistently

This change refactors reflect.Value to consistently use arrayAt when an element
of an array of bytes is indexed.

This effectively replaces:
 arr := unsafe.Pointer(...)
 arri := unsafe.Pointer(uintptr(arr) + uintptr(i)*elementSize)

with:
 arr := unsafe.Pointer(...)
 arri := arrayAt(arr, i, elementSize)

Change-Id: I53ffd0d6de693b43d5c10c0aa4cd6d4f5e95a1e3
Reviewed-on: https://go-review.googlesource.com/9183
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

10 years agocmd/internal/obj: reuse the varint encoding buffer
Josh Bleecher Snyder [Mon, 20 Apr 2015 17:20:36 +0000 (10:20 -0700)]
cmd/internal/obj: reuse the varint encoding buffer

This reduces the number of allocations in the compiler
while building the stdlib by 15.66%.

No functional changes. Passes toolstash -cmp.

Change-Id: Ia21b37134a8906a4e23d53fdc15235b4aa7bbb34
Reviewed-on: https://go-review.googlesource.com/9085
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
10 years agodoc/go1.5.txt: add reflect.ArrayOf
Sebastien Binet [Tue, 21 Apr 2015 15:30:49 +0000 (17:30 +0200)]
doc/go1.5.txt: add reflect.ArrayOf

Change-Id: I89704249218d4fdba11463c239c69143f8ad0051
Reviewed-on: https://go-review.googlesource.com/9185
Reviewed-by: Ian Lance Taylor <iant@golang.org>