]> Cypherpunks repositories - gostls13.git/log
gostls13.git
9 years agocmd/compile: remove most of the Lookupf users and garbage
Brad Fitzpatrick [Sun, 20 Mar 2016 01:17:58 +0000 (18:17 -0700)]
cmd/compile: remove most of the Lookupf users and garbage

Introduce garbage-free LookupN to replace most users of Lookupf.

Also, remove the string interning from LookupBytes which was hurting
more than helping.

name       old alloc/op    new alloc/op    delta
Template      63.0MB ± 0%     62.7MB ± 0%  -0.48%         (p=0.000 n=10+9)
Unicode       43.0MB ± 0%     43.0MB ± 0%  -0.17%         (p=0.000 n=10+7)
GoTypes        219MB ± 0%      218MB ± 0%  -0.14%        (p=0.000 n=10+10)
Compiler       992MB ± 0%      991MB ± 0%  -0.12%        (p=0.000 n=10+10)

name       old allocs/op   new allocs/op   delta
Template        683k ± 0%       681k ± 0%  -0.38%         (p=0.000 n=10+8)
Unicode         541k ± 0%       541k ± 0%  -0.11%        (p=0.000 n=10+10)
GoTypes        2.09M ± 0%      2.08M ± 0%  -0.40%        (p=0.000 n=10+10)
Compiler       9.28M ± 0%      9.24M ± 0%  -0.36%        (p=0.000 n=10+10)

Size of $GOROOT/pkg/darwin_amd64 drops from 40124 KB to 40100 KB too,
removing the zero padding as suggested by josharian.

Updates #6853

Change-Id: I3c557266e9325fe29c459cef8e5b8954913e7abb
Reviewed-on: https://go-review.googlesource.com/20931
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/dist: add "s390x" to okgoarch and cgoEnabled
Michael Munday [Fri, 18 Mar 2016 19:39:25 +0000 (15:39 -0400)]
cmd/dist: add "s390x" to okgoarch and cgoEnabled

Allows the compiler to recognise s390x specific files and s390x
build tags.

Change-Id: I7c62ab7361cf708181b1d9cfbe9b1fcb01be31e0
Reviewed-on: https://go-review.googlesource.com/20872
Reviewed-by: Minux Ma <minux@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoall: delete dead test code
Dominik Honnef [Sun, 20 Mar 2016 23:13:36 +0000 (00:13 +0100)]
all: delete dead test code

This deletes unused code and helpers from tests.

Change-Id: Ie31d46115f558ceb8da6efbf90c3c204e03b0d7e
Reviewed-on: https://go-review.googlesource.com/20927
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agotest/fixedbugs: add s390x case to issue11656
Michael Munday [Sun, 20 Mar 2016 23:25:58 +0000 (19:25 -0400)]
test/fixedbugs: add s390x case to issue11656

An instruction consisting of all 0s causes an illegal instruction
signal on s390x. Since 0s are the default in this test this CL just
makes it explicit.

Change-Id: Id6e060eed1a588f4b10a4e4861709fcd19b434ac
Reviewed-on: https://go-review.googlesource.com/20962
Reviewed-by: Minux Ma <minux@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/internal/objfile: add s390x support
Michael Munday [Fri, 18 Mar 2016 20:49:19 +0000 (16:49 -0400)]
cmd/internal/objfile: add s390x support

Change-Id: I39aa6569c9a6f327f7aaa01f8b4ace814fd5b766
Reviewed-on: https://go-review.googlesource.com/20943
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/link: delete more unreachable code
Michael Hudson-Doyle [Mon, 21 Mar 2016 06:32:57 +0000 (19:32 +1300)]
cmd/link: delete more unreachable code

Debugasm can never be set in cmd/link, so delete it and the code it enables.

Change-Id: If828db0b09f1a9e512dc660ac2750657a769094c
Reviewed-on: https://go-review.googlesource.com/20971
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/link: delete unreachable hash collision check
Michael Hudson-Doyle [Mon, 21 Mar 2016 06:27:35 +0000 (19:27 +1300)]
cmd/link: delete unreachable hash collision check

This expression in readsym:

    dup != nil && len(dup.P) > 0 && strings.HasPrefix(s.Name, "gclocals·")

can never be true: if dup != nil, then s.Name is ".dup" (and this is not new:
the same broken logic is present in 1.4, at least). Delete the whole block.

Change-Id: I33b14d9a82b292116d6fd79d22b38e3842501317
Reviewed-on: https://go-review.googlesource.com/20970
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/internal/obj: move Nocache helper to arm back end
Dave Cheney [Mon, 21 Mar 2016 03:19:57 +0000 (14:19 +1100)]
cmd/internal/obj: move Nocache helper to arm back end

The obj.Nocache helper was only used by the arm back end, move it there.

Change-Id: I5c9faf995499991ead1f3d8c8ffc3b6af7346876
Reviewed-on: https://go-review.googlesource.com/20868
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: unexport convlit
Josh Bleecher Snyder [Mon, 21 Mar 2016 02:41:43 +0000 (19:41 -0700)]
cmd/compile: unexport convlit

Add a special helper for its one external use.

This is in preparation for an upcoming CL.

Passes toolstash -cmp / buildall.

Change-Id: I9d3463792afe220cc4bc89269bdecf0279abd281
Reviewed-on: https://go-review.googlesource.com/20933
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/internal/obj: remove Link.Windows field
Dave Cheney [Mon, 21 Mar 2016 03:37:57 +0000 (14:37 +1100)]
cmd/internal/obj: remove Link.Windows field

This CL addresses a long standing CL by rsc by pushing the use of
Link.Windows down to its two users.

Link.Window was always initalised with the value of runtime.GOOS so
this does not affect cross compilation.

Change-Id: Ibbae068f8b5aad06336909691f094384caf12352
Reviewed-on: https://go-review.googlesource.com/20869
Run-TryBot: Dave Cheney <dave@cheney.net>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agomath/big: cleanup documentation for Format methods
Matthew Dempsky [Sun, 20 Mar 2016 21:54:13 +0000 (14:54 -0700)]
math/big: cleanup documentation for Format methods

'b' is a standard verb for floating point values. The runes like '+'
and '#' are called "flags" by package fmt's documentation. The flag
'-' controls left/right justification, not anything related to signs.

Change-Id: Ia9cf81b002df373f274ce635fe09b5bd0066aa1c
Reviewed-on: https://go-review.googlesource.com/20930
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/internal/obj, cmd/link: put all symbol data in one contiguous section
Michael Hudson-Doyle [Sun, 20 Mar 2016 21:55:20 +0000 (10:55 +1300)]
cmd/internal/obj, cmd/link: put all symbol data in one contiguous section

Another object file change, gives a reasonable improvement:

name       old s/op   new s/op   delta
LinkCmdGo  0.46 ± 3%  0.44 ± 9%  -3.34%  (p=0.000 n=98+82)
LinkJuju   4.09 ± 4%  3.92 ± 5%  -4.30%  (p=0.000 n=98+99)

I guess the data section could be mmap-ed instead of read, I haven't tried
that.

Change-Id: I959eee470a05526ab1579e3f5d3ede41c16c954f
Reviewed-on: https://go-review.googlesource.com/20928
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agonet/http: adaptive wait time in PersistConnLeak tests
Richard Miller [Sun, 20 Mar 2016 19:17:36 +0000 (19:17 +0000)]
net/http: adaptive wait time in PersistConnLeak tests

In tests TransportPersistConnLeak and TransportPersistConnLeakShortBody,
there's a fixed wait time (100ms and 400ms respectively) to allow
goroutines to exit after CloseIdleConnections is called. This
is sometimes too short on a slow host running many simultaneous
tests.

This CL replaces the fixed sleep in each test with a sequence of
shorter sleeps, testing the number of remaining goroutines until
it reaches the threshold or an overall time limit of 500ms expires.
This prevents some failures in the plan9_arm builder, while reducing
the test time on faster machines.

Fixes #14887

Change-Id: Ia5c871062df139e2667cdfb2ce8283e135435318
Reviewed-on: https://go-review.googlesource.com/20922
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/compile: add rules to simplify AddPtr
Alexandru Moșoi [Fri, 18 Mar 2016 09:50:00 +0000 (10:50 +0100)]
cmd/compile: add rules to simplify AddPtr

Fixes #14849

Change-Id: I86e2dc27ca73bb6b24261a68cbf0094a63167414
Reviewed-on: https://go-review.googlesource.com/20833
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Alexandru Moșoi <alexandru@mosoi.ro>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: fix varexpr handling of ODOT
Ian Lance Taylor [Sun, 20 Mar 2016 17:25:46 +0000 (10:25 -0700)]
cmd/compile: fix varexpr handling of ODOT

For a long time varexpr has handled ODOT incorrectly: it has always
returned false.  Before https://golang.org/cl/20890 this has been
because an ODOT had a Right field with an ONAME with no Class, for which
varexpr returns false.  CL 20890 preserved the behavior of varexpr for
ODOT, so that the change would pass toolstash -cmp.

This CL fixes varexpr so that ODOT can return true in some cases.  This
breaks toolstash -cmp.  While the changed compiler allocates temporary
variables in a different order, I have not been able to find any
examples where the generated code is different, other than using
different stack offsets and, in some cases, registers.  It seems that
other parts of the compiler will force the ODOT into a temporary anyhow.

Still, this change is clearly correct, and is a minor compiler cleanup.

Change-Id: I71506877aa3c13966bb03c281aa16271ee7fe80a
Reviewed-on: https://go-review.googlesource.com/20907
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/link: patch up symbols only once per object file
Shahar Kohanim [Sun, 20 Mar 2016 15:22:57 +0000 (17:22 +0200)]
cmd/link: patch up symbols only once per object file

name       old s/op   new s/op   delta
LinkCmdGo  0.57 ± 5%  0.55 ± 6%  -2.37%  (p=0.000 n=97+98)

GOGC=off:

name       old s/op   new s/op   delta
LinkCmdGo  0.48 ± 3%  0.47 ± 3%  -2.90%  (p=0.000 n=100+100)

Change-Id: I1a36dbf84914cacb79842bc0ddb1e26b4c5a5828
Reviewed-on: https://go-review.googlesource.com/20917
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: don't penalize ODOT and friends when inlining
Ian Lance Taylor [Sat, 19 Mar 2016 01:12:14 +0000 (18:12 -0700)]
cmd/compile: don't penalize ODOT and friends when inlining

Historically ODOT and friends have been considered to cost an extra
budget point when deciding whether they should be inlined, because they
had an ONAME node that represented the name to the right of the dot.
This doesn't really make sense, as in general that symbol does not add
any extra instructions; it just affects the offset of the load or store
instruction.  And the ONAME node is gone now.  So, remove the extra
cost.

This does not pass toolstash -cmp, as it changes inlining decisions.
For example, mspan.init in runtime/mheap.go is now considered to be an
inlining candidate.

Change-Id: I5ad27f08c66fd5daa4c8472dd0795df989183f5e
Reviewed-on: https://go-review.googlesource.com/20891
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/link: use encbuf when writing integers
Shahar Kohanim [Sun, 20 Mar 2016 09:52:41 +0000 (11:52 +0200)]
cmd/link: use encbuf when writing integers

name       old s/op   new s/op   delta
LinkCmdGo  0.59 ± 6%  0.58 ± 5%  -1.61%  (p=0.000 n=99+99)

GOGC=off:
name       old s/op   new s/op   delta
LinkCmdGo  0.50 ± 3%  0.49 ± 3%  -1.28%  (p=0.000 n=98+99)

Change-Id: I737ae056214999441a210c69ec0cf4febc39a715
Reviewed-on: https://go-review.googlesource.com/20914
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile, cmd/link: remove unused fields from relocations
Shahar Kohanim [Sun, 20 Mar 2016 07:44:18 +0000 (09:44 +0200)]
cmd/compile, cmd/link: remove unused fields from relocations

Reduces size of archives in pkg/linux_amd64 by 3% from 41.5MB to 40.2MB

Change-Id: Id64ca7995de8dd84c9e7ce1985730927cf4bfd66
Reviewed-on: https://go-review.googlesource.com/20912
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agocmd/link: optimize int parsing
Shahar Kohanim [Sat, 19 Mar 2016 21:27:41 +0000 (23:27 +0200)]
cmd/link: optimize int parsing

Speeds up linking cmd/go by ~1.5%:

name       old s/op   new s/op   delta
LinkCmdGo  0.58 ± 6%  0.57 ± 5%  -1.21%  (p=0.000 n=98+99)

Less noisy benchmark, with garbage collection off:

name       old s/op   new s/op   delta
LinkCmdGo  0.49 ± 2%  0.49 ± 2%  -1.79%  (p=0.000 n=98+99)

Change-Id: I0123bcb66a87cbc4d703356e4c5a4035032012ec
Reviewed-on: https://go-review.googlesource.com/20916
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Run-TryBot: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agofmt: unify integer formatting
Martin Möhrmann [Fri, 18 Mar 2016 07:24:40 +0000 (08:24 +0100)]
fmt: unify integer formatting

Deduplicate the verb switch for signed and unsigned integer formatting.

Make names of integer related functions consistent
with names of other fmt functions.

Consolidate basic integer tests.

Change-Id: I0c19c24f1c2c06a3b1a4d7d377dcdac3b36bb0f5
Reviewed-on: https://go-review.googlesource.com/20831
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>
9 years agocmd/compile: ignore receiver name when checking duplicate methods
Matthew Dempsky [Sun, 20 Mar 2016 00:51:17 +0000 (17:51 -0700)]
cmd/compile: ignore receiver name when checking duplicate methods

In golang.org/cl/20602, I changed the semantics of Eqtype to stop
checking the receiver parameters for type equality, and pushed this
responsibility to addmethod (the only Eqtype caller that cared).
However, I accidentally made the check stricter by making it start
requiring that receiver names were identical.

In general, this is a non-problem because the receiver names in export
data will always match the original source. But running
GO_GCFLAGS=-newexport ./all.bash at one point tries to load both old
and new format export data for package sync, which reveals the
problem. (See golang.org/issue/14877 for details.)

Easy fix: just check the receiver type for type equality in addmethod,
instead of the entire receiver parameter list.

Fixes #14877.

Change-Id: If10b79f66ba58a1b7774622b4fbad1916aba32f1
Reviewed-on: https://go-review.googlesource.com/20906
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/compile: remove typechecklist
Josh Bleecher Snyder [Sun, 20 Mar 2016 00:02:01 +0000 (17:02 -0700)]
cmd/compile: remove typechecklist

Convert remaining uses to typecheckslice.

Passes toolstash -cmp.

Change-Id: I6ed0877386fb6c0b036e8ee5a228433343855abd
Reviewed-on: https://go-review.googlesource.com/20905
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Dave Cheney <dave@cheney.net>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: convert fmt.Sprintf to concatenation
Josh Bleecher Snyder [Sat, 19 Mar 2016 23:34:48 +0000 (16:34 -0700)]
cmd/compile: convert fmt.Sprintf to concatenation

There are plenty more, but these cover most of
the trivial cases, and all the cases that
showed up in profiling.

name       old time/op     new time/op     delta
Template       331ms ± 3%      327ms ± 6%    ~           (p=0.143 n=10+10)
Unicode        183ms ± 4%      180ms ± 2%    ~             (p=0.114 n=9+8)
GoTypes        1.12s ± 4%      1.07s ± 1%  -4.34%         (p=0.000 n=10+9)
Compiler       5.16s ± 2%      5.04s ± 2%  -2.24%         (p=0.001 n=10+9)
MakeBash       41.7s ± 2%      42.3s ± 1%  +1.51%        (p=0.000 n=10+10)

name       old alloc/op    new alloc/op    delta
Template      63.4MB ± 0%     63.1MB ± 0%  -0.42%        (p=0.000 n=10+10)
Unicode       43.2MB ± 0%     43.1MB ± 0%  -0.22%         (p=0.000 n=9+10)
GoTypes        220MB ± 0%      219MB ± 0%  -0.57%         (p=0.000 n=8+10)
Compiler       978MB ± 0%      975MB ± 0%  -0.30%        (p=0.000 n=10+10)

name       old allocs/op   new allocs/op   delta
Template        702k ± 0%       686k ± 0%  -2.35%        (p=0.000 n=10+10)
Unicode         548k ± 0%       542k ± 0%  -1.09%        (p=0.000 n=10+10)
GoTypes        2.17M ± 0%      2.09M ± 0%  -3.61%        (p=0.000 n=10+10)
Compiler       9.33M ± 0%      9.15M ± 0%  -1.93%        (p=0.000 n=10+10)

Change-Id: I3a3d7f2d56876427b04cfedc7302d7f496d5bb65
Reviewed-on: https://go-review.googlesource.com/20904
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Dave Cheney <dave@cheney.net>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: enforce 32-bit restrictions on ops
Keith Randall [Fri, 18 Mar 2016 17:45:30 +0000 (10:45 -0700)]
cmd/compile: enforce 32-bit restrictions on ops

Most 64-bit x86 ops can only take a signed 32-bit constant.
Clean up our rewrite rules to enforce this restriction.

Modify the assembler to fail if the offset does not fit
in the instruction.

That last check triggers a few times on weird testing code.
Suppress those errors if the compiler itself generated errors.

Fixes #14862

Change-Id: I76559af035b38483b1e59621a8029fc66b3a5d1e
Reviewed-on: https://go-review.googlesource.com/20815
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
9 years agofmt: integer formatting should not permanently change padding
Martin Möhrmann [Thu, 10 Mar 2016 11:33:33 +0000 (12:33 +0100)]
fmt: integer formatting should not permanently change padding

Changes the integer function to restore the original f.zero value
and therefore padding type before returning.

Change-Id: I456449259a3d39bd6d62e110553120c31ec63f23
Reviewed-on: https://go-review.googlesource.com/20512
Reviewed-by: Rob Pike <r@golang.org>
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agofmt: remove depth argument from handleMethods and printArg
Martin Möhrmann [Sat, 19 Mar 2016 12:18:43 +0000 (13:18 +0100)]
fmt: remove depth argument from handleMethods and printArg

handleMethods can format Error() and String() directly as its known
these return strings that can be directly printed using fmtString.
Remove the obsolete depth argument from handleMethods.

Remove the depth argument from printArg since it is only ever
called with depth set to 0. Recursion for formatting complex
arguments is handled only by printValue which keeps track of depth.

Change-Id: I4c4be588751de12ed999e7561a51bc168eb9eb2d
Reviewed-on: https://go-review.googlesource.com/20911
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>
9 years agocmd/compile: change ODOT and friends to use Sym, not Right
Ian Lance Taylor [Fri, 18 Mar 2016 23:52:30 +0000 (16:52 -0700)]
cmd/compile: change ODOT and friends to use Sym, not Right

The Node type ODOT and its variants all represent a selector, with a
simple name to the right of the dot.  Before this change this was
represented by using an ONAME Node in the Right field.  This ONAME node
served no useful purpose.  This CL changes these Node types to store the
symbol in the Sym field instead, thus not requiring allocating a Node
for each selector.

When compiling x/tools/go/types this CL eliminates nearly 5000 calls to
newname and reduces the total number of Nodes allocated by about 6.6%.
It seems to cut compilation time by 1 to 2 percent.

Getting this right was somewhat subtle, and I added two dubious changes
to produce the exact same output as before.  One is to ishairy in
inl.go: the ONAME node increased the cost of ODOT and friends by 1, and
I retained that, although really ODOT is not more expensive than any
other node.  The other is to varexpr in walk.go: because the ONAME in
the Right field of an ODOT has no class, varexpr would always return
false for an ODOT, although in fact for some ODOT's it seemingly ought
to return true; I added an && false for now.  I will send separate CLs,
that will break toolstash -cmp, to clean these up.

This CL passes toolstash -cmp.

Change-Id: I4af8a10cc59078c436130ce472f25abc3a9b2f80
Reviewed-on: https://go-review.googlesource.com/20890
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/compile: allow inlining of functions that declare a const
Todd Neal [Wed, 16 Mar 2016 23:44:17 +0000 (18:44 -0500)]
cmd/compile: allow inlining of functions that declare a const

Consider functions with an ODCLCONST for inlining and modify exprfmt to
ignore those nodes when exporting. Don't add symbols to the export list
if there is no definition.  This occurs when OLITERAL symbols are looked
up via Pkglookup for non-exported symbols.

Fixes #7655

Change-Id: I1de827850f4c69e58107447314fe7433e378e069
Reviewed-on: https://go-review.googlesource.com/20773
Run-TryBot: Todd Neal <todd@tneal.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
9 years agocmd/compile: merge shifts into LEAs
Keith Randall [Fri, 18 Mar 2016 21:54:18 +0000 (14:54 -0700)]
cmd/compile: merge shifts into LEAs

Change-Id: I5a43c354f36184ae64a52268023c3222da3026d8
Reviewed-on: https://go-review.googlesource.com/20880
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
9 years agoruntime: add space
Keith Randall [Fri, 18 Mar 2016 21:30:00 +0000 (14:30 -0700)]
runtime: add space

Missed this in review of 20812

Change-Id: I01e220499dcd58e1a7205e2a577dd9630a8b7174
Reviewed-on: https://go-review.googlesource.com/20819
Reviewed-by: Keith Randall <khr@golang.org>
9 years agofmt: simplify handling of reporting flags to formatters
Martin Möhrmann [Fri, 18 Mar 2016 15:56:47 +0000 (16:56 +0100)]
fmt: simplify handling of reporting flags to formatters

Remove rewriting of flags before calling formatters.
Change Flag method to directly take plusV and sharpV flags
into account when reporting if plus or sharp flag is set.

Change-Id: Ic3423881ad89e5a5f9fff5ab59e842062394ef6d
Reviewed-on: https://go-review.googlesource.com/20859
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>
9 years agoruntime: use unaligned loads on ppc64
Keith Randall [Thu, 17 Mar 2016 20:28:04 +0000 (13:28 -0700)]
runtime: use unaligned loads on ppc64

benchmark                      old ns/op     new ns/op     delta
BenchmarkAlignedLoad-160       8.67          7.42          -14.42%
BenchmarkUnalignedLoad-160     8.63          7.37          -14.60%

Change-Id: Id4609d7b4038c4d2ec332efc4fe6f1adfb61b82b
Reviewed-on: https://go-review.googlesource.com/20812
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/dist: redo flag-passing for bootstrap
David Chase [Fri, 18 Mar 2016 17:35:34 +0000 (13:35 -0400)]
cmd/dist: redo flag-passing for bootstrap

This ought to revert the bad effects of
https://go-review.googlesource.com/#/c/20775/
If you don't pass BOOT_GO_GCFLAGS, you get the
old behavior.

Tweaked to allow multiple space-separated flags in
BOOT_GO_GCFLAGS.

Change-Id: I2a22a04211b4535d1c5a8ec7a8a78cb051161c31
Reviewed-on: https://go-review.googlesource.com/20871
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agofmt: remove unused field from printer struct
Martin Möhrmann [Fri, 18 Mar 2016 11:10:47 +0000 (12:10 +0100)]
fmt: remove unused field from printer struct

Change-Id: I0ec775c51f461c6f0cbff88e796a7af55b736fcb
Reviewed-on: https://go-review.googlesource.com/20838
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agosync: don't assume b.N > 0
Marcel van Lohuizen [Fri, 18 Mar 2016 16:16:56 +0000 (17:16 +0100)]
sync: don't assume b.N > 0

Change-Id: I6eb91ea73ef887b025e5a8de1dd55f30618e1aa6
Reviewed-on: https://go-review.googlesource.com/20857
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/compile/internal/gc: support invalid types/constants in binary export data
Alan Donovan [Fri, 18 Mar 2016 15:13:24 +0000 (11:13 -0400)]
cmd/compile/internal/gc: support invalid types/constants in binary export data

(Corresponding x/tools/go/gcimporter change is https://go-review.googlesource.com/#/c/20827/)

Change-Id: I64e7fee2e273d387f1c51b87986294489978d250
Reviewed-on: https://go-review.googlesource.com/20828
Reviewed-by: Robert Griesemer <gri@golang.org>
9 years agotesting: disable tests that cause a hang on some platforms
Marcel van Lohuizen [Fri, 18 Mar 2016 13:05:54 +0000 (14:05 +0100)]
testing: disable tests that cause a hang on some platforms

plan9, nacl, and netbsd to be precise.

Only the first test causes a hang, but just to be sure.

Change-Id: I400bb356ee2a0cf12c8666c95af79c924d1629aa
Reviewed-on: https://go-review.googlesource.com/20839
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agotesting: always ignore RunParallel in probe phase
Marcel van Lohuizen [Fri, 18 Mar 2016 15:43:43 +0000 (16:43 +0100)]
testing: always ignore RunParallel in probe phase

Change-Id: If45410a2d7e48d1c9e6800cd98f81dd89024832c
Reviewed-on: https://go-review.googlesource.com/20852
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agoruntime: don't assume b.N > 0
Marcel van Lohuizen [Fri, 18 Mar 2016 15:41:51 +0000 (16:41 +0100)]
runtime: don't assume b.N > 0

Change-Id: I2e26717f2563d7633ffd15f4adf63c3d0ee3403f
Reviewed-on: https://go-review.googlesource.com/20856
Run-TryBot: Marcel van Lohuizen <mpvl@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet/url: don't assume b.N > 0
Marcel van Lohuizen [Fri, 18 Mar 2016 15:41:38 +0000 (16:41 +0100)]
net/url: don't assume b.N > 0

Change-Id: Ie79c16d6e61b3baa274069528cf883b22fd255fe
Reviewed-on: https://go-review.googlesource.com/20855
Run-TryBot: Marcel van Lohuizen <mpvl@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet/rpc: don't assume b.N > 0
Marcel van Lohuizen [Fri, 18 Mar 2016 15:41:21 +0000 (16:41 +0100)]
net/rpc: don't assume b.N > 0

Change-Id: I58c4a75168fd1f797a25735c4151f501f0475332
Reviewed-on: https://go-review.googlesource.com/20854
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agoencoding/binary: don't assume b.N > 0
Marcel van Lohuizen [Fri, 18 Mar 2016 15:40:46 +0000 (16:40 +0100)]
encoding/binary: don't assume b.N > 0

Change-Id: I9e887a0b32baf0adc85fa9e4b85b319e8ef333e9
Reviewed-on: https://go-review.googlesource.com/20853
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agoencoding/binary: improve error messages for benchmarks
Marcel van Lohuizen [Fri, 18 Mar 2016 13:27:29 +0000 (14:27 +0100)]
encoding/binary: improve error messages for benchmarks

Change-Id: I0f4b6752ecc8b4945ecfde627cdec13fc4bb6a69
Reviewed-on: https://go-review.googlesource.com/20850
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/compile: perform minimal phi elimination during critical
Todd Neal [Fri, 18 Mar 2016 12:05:58 +0000 (07:05 -0500)]
cmd/compile: perform minimal phi elimination during critical

Phi splitting sometimes leads to a phi with only a single predecessor.
This must be replaced with a copy to maintain a valid SSA form.

Fixes #14857

Change-Id: I5ab2423fb6c85a061928e3206b02185ea8c79cd7
Reviewed-on: https://go-review.googlesource.com/20826
Reviewed-by: Keith Randall <khr@golang.org>
9 years agotesting: finish implementation of subtests
Marcel van Lohuizen [Mon, 25 Jan 2016 15:27:23 +0000 (16:27 +0100)]
testing: finish implementation of subtests

API not exposed yet.

Change-Id: Iaba0adc0fa1ae8075e6b56796f99ee8db9177a78
Reviewed-on: https://go-review.googlesource.com/18896
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agotesting: implementation of subbenchmarks
Marcel van Lohuizen [Wed, 20 Jan 2016 14:47:54 +0000 (15:47 +0100)]
testing: implementation of subbenchmarks

API is not exposed yet.

Change-Id: I729360ef2be1d8ea683ca93cdb1763897cc8657c
Reviewed-on: https://go-review.googlesource.com/18895
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agotesting: hoisted chunks of code to prepare for Run method
Marcel van Lohuizen [Fri, 29 Jan 2016 15:55:35 +0000 (16:55 +0100)]
testing: hoisted chunks of code to prepare for Run method

testing.go:
- run method will evolve into the Run method.
- added level field in common

benchmark.go:
- benchContext will be central to distinguish handling of benchmarks
  between normal Run methods and ones called from within Benchmark
  function.
- expandCPU will evolve into the processing hook for Run methods
  called within normal processing.
- runBench will evolve into the Run method.

Change-Id: I1816f9985d5ba94deb0ad062302ea9aee0bb5338
Reviewed-on: https://go-review.googlesource.com/18894
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agotesting: prepare for the introduction of Run methods
Marcel van Lohuizen [Tue, 19 Jan 2016 21:43:52 +0000 (22:43 +0100)]
testing: prepare for the introduction of Run methods

The biggest change is that each test is now responsible for managing
the starting and stopping of its parallel subtests.

The "Main" test could be run as a tRunner as well. This shows that
the introduction of subtests is merely a generalization of and
consistent with the current semantics.

Change-Id: Ibf8388c08f85d4b2c0df69c069326762ed36a72e
Reviewed-on: https://go-review.googlesource.com/18893
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agoregexp: avoid copying mutex in (*Regexp).Copy.
David Symonds [Thu, 17 Mar 2016 22:55:18 +0000 (09:55 +1100)]
regexp: avoid copying mutex in (*Regexp).Copy.

There's nothing guaranteeing that the *Regexp isn't in active use,
and so copying the sync.Mutex value is invalid.

Updates #14839.

Change-Id: Iddf52bf69df1b563377922399f64a571f76b95dd
Reviewed-on: https://go-review.googlesource.com/20841
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agocmd/compile: use loop information in regalloc
David Chase [Thu, 10 Mar 2016 19:42:52 +0000 (14:42 -0500)]
cmd/compile: use loop information in regalloc

This seems to help the problem reported in #14606; this
change seems to produce about a 4% improvement (mostly
for the 128-8192 shards).

Fixes #14789.

Change-Id: I1bd52c82d4ca81d9d5e9ab371fdfc860d7e8af50
Reviewed-on: https://go-review.googlesource.com/20660
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
9 years agocmd/go: fix TestShadowingLogic fails when GOROOT path has spaces
Christopher Nelson [Tue, 15 Mar 2016 13:14:22 +0000 (09:14 -0400)]
cmd/go: fix TestShadowingLogic fails when GOROOT path has spaces

Improve the test by also translating " " to "_".

Fixes #14671.

Change-Id: Ie5997934b93c7663d7b8432244fad47bb5d3ffbe
Reviewed-on: https://go-review.googlesource.com/20714
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/compile: get gcflags to bootstrap; ssa debug opts for "all"
David Chase [Thu, 17 Mar 2016 18:12:12 +0000 (14:12 -0400)]
cmd/compile: get gcflags to bootstrap; ssa debug opts for "all"

This is intended to help debug compiler problems that pop
up in the bootstrap phase of make.bash.  GO_GCFLAGS does not
normally apply there.  Options-for-all phases is intended
to allow crude tracing (and full timing) by turning on timing
for all phases, not just one.

Phase names can also be specified using a regular expression,
for example
BOOT_GO_GCFLAGS=-d='ssa/~^.*scc$/off' \
GO_GCFLAGS='-d=ssa/~^.*scc$/off' ./make.bash

I just added this because it was the fastest way to get
me to a place where I could easily debug the compiler.

Change-Id: I0781f3e7c19651ae7452fa25c2d54c9a245ef62d
Reviewed-on: https://go-review.googlesource.com/20775
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile/test: replace switch{} with go:noinline
Todd Neal [Thu, 17 Mar 2016 02:22:41 +0000 (21:22 -0500)]
cmd/compile/test: replace switch{} with go:noinline

Change-Id: Ic40449b2e4b4f18cbe5b5d4c3d51ea7b05ac674d
Reviewed-on: https://go-review.googlesource.com/20823
Run-TryBot: Todd Neal <todd@tneal.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agonet: make unexposed methods start with lowercase letters
Mikio Hara [Wed, 16 Mar 2016 20:33:13 +0000 (05:33 +0900)]
net: make unexposed methods start with lowercase letters

This change makes unexposed methods start with lowercase letters for
avoiding unnecessary confusion because the net package uses many
embedding structures and intrefaces for controlling exposure of APIs.

Note that this change leaves DNS-related methods as they are.

Change-Id: I253758d1659175c5d0af6b2efcd30ce83f46543d
Reviewed-on: https://go-review.googlesource.com/20784
Run-TryBot: Mikio Hara <mikioh.mikioh@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/pack,vet: use `go doc` instead of `godoc` in doc
Hyang-Ah Hana Kim [Thu, 17 Mar 2016 19:32:20 +0000 (15:32 -0400)]
cmd/pack,vet: use `go doc` instead of `godoc` in doc

Change-Id: Ic5f62a7d0a5c090da69213d1d0187af0ea48e358
Reviewed-on: https://go-review.googlesource.com/20820
Reviewed-by: Rob Pike <r@golang.org>
9 years agocmd/compile: correct maintain use count when phi args merge
David Chase [Thu, 17 Mar 2016 20:14:41 +0000 (16:14 -0400)]
cmd/compile: correct maintain use count when phi args merge

The critical phase did not correctly maintain the use count
when two predecessors of a new critical block transmit the
same value.

Change-Id: Iba802c98ebb84e36a410721ec32c867140efb6d4
Reviewed-on: https://go-review.googlesource.com/20822
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
9 years agoencoding/binary: remove bound checks from conversions.
Alexandru Moșoi [Sun, 13 Mar 2016 21:12:03 +0000 (22:12 +0100)]
encoding/binary: remove bound checks from conversions.

* This the simplest solution I could came up with
that doesn't required changing the compiler.
* The bound checks become constants now
so they are removed during opt phase.

Updates #14808

Change-Id: If32c33d7ec08bb400321b465015d152f0a5d3001
Reviewed-on: https://go-review.googlesource.com/20654
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Alexandru Moșoi <alexandru@mosoi.ro>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: eliminate NumFields wrapper functions
Matthew Dempsky [Thu, 17 Mar 2016 20:26:08 +0000 (13:26 -0700)]
cmd/compile: eliminate NumFields wrapper functions

Change-Id: I3c6035559288cfdc33857216f50241b81932c8a4
Reviewed-on: https://go-review.googlesource.com/20811
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: enforce nowritebarrier in SSA compiler
Keith Randall [Thu, 17 Mar 2016 04:51:17 +0000 (21:51 -0700)]
cmd/compile: enforce nowritebarrier in SSA compiler

Make sure we don't generate write barriers in runtime
code that is marked to forbid write barriers.

Implement the optimization that if we're writing a sliced
slice back to the location it came from, we don't need a
write barrier.

Fixes #14784

Change-Id: I04b6a3b2ac303c19817e932a36a3b006de103aaa
Reviewed-on: https://go-review.googlesource.com/20791
Reviewed-by: Austin Clements <austin@google.com>
9 years agocmd/compile: deduplicate symbol references
Shahar Kohanim [Thu, 17 Mar 2016 11:18:34 +0000 (13:18 +0200)]
cmd/compile: deduplicate symbol references

Reduces size of archives in pkg/linux_amd64 by 1.4MB (3.2%),
slightly improving link time.

name       old s/op   new s/op   delta
LinkCmdGo  0.52 ± 3%  0.51 ± 2%  -0.65%  (p=0.000 n=98+99)

Change-Id: I7e265f4d4dd08967c5c5d55c1045e533466bbbec
Reviewed-on: https://go-review.googlesource.com/20802
Run-TryBot: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/compile: fix comment
Ingo Oeser [Thu, 17 Mar 2016 19:40:02 +0000 (20:40 +0100)]
cmd/compile: fix comment

Change-Id: I32fd5c36f055fdb1dfe56524085676aa4111089a
Reviewed-on: https://go-review.googlesource.com/20830
Reviewed-by: David Chase <drchase@google.com>
9 years agocmd/compile: get rid of Type's {This,In,Out}tuple fields
Matthew Dempsky [Thu, 17 Mar 2016 08:47:16 +0000 (01:47 -0700)]
cmd/compile: get rid of Type's {This,In,Out}tuple fields

Boolean expressions involving t.Thistuple were converted to use
t.Recv(), because it's a bit clearer and will hopefully reveal cases
where we could remove redundant calls to t.Recv() (in followup CLs).

The other cases were all converted to use t.Recvs().NumFields(),
t.Params().NumFields(), or t.Results().NumFields().

Change-Id: I4df91762e7dc4b2ddae35995f8dd604a52c09b09
Reviewed-on: https://go-review.googlesource.com/20796
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>

9 years agocmd/compile: simplify typehash
Matthew Dempsky [Thu, 17 Mar 2016 08:39:59 +0000 (01:39 -0700)]
cmd/compile: simplify typehash

We never need a type hash for a method type, so skip trying to
overwrite Thistuple.

Change-Id: I8de6480ba5fd321dfa134facf7661461d298840e
Reviewed-on: https://go-review.googlesource.com/20795
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/compile: eliminate a bunch of IterFields/IterMethods calls
Matthew Dempsky [Thu, 17 Mar 2016 08:32:18 +0000 (01:32 -0700)]
cmd/compile: eliminate a bunch of IterFields/IterMethods calls

This is an automated rewrite of all the calls of the form:

    for f, it := IterFields(t); f != nil; f = it.Next() { ... }

Followup CLs will work on cleaning up the remaining cases.

Change-Id: Ic1005ad45ae0b50c63e815e34e507e2d2644ba1a
Reviewed-on: https://go-review.googlesource.com/20794
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agocmd/compile: add and use new Fields type
Matthew Dempsky [Thu, 17 Mar 2016 07:44:07 +0000 (00:44 -0700)]
cmd/compile: add and use new Fields type

Analogous to the Nodes type used as a more space efficient []*Node
representation.

Passes toolstash -cmp.

Change-Id: I8341e45304777d6e4200bd36dadc935b07ccf3ff
Reviewed-on: https://go-review.googlesource.com/20793
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: document sudog
Austin Clements [Thu, 17 Mar 2016 18:08:32 +0000 (14:08 -0400)]
runtime: document sudog

Change-Id: I85c0bcf02842cc32dbc9bfdcea27efe871173574
Reviewed-on: https://go-review.googlesource.com/20774
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/compile: stop constructing sudog type
Matthew Dempsky [Thu, 17 Mar 2016 11:58:34 +0000 (04:58 -0700)]
cmd/compile: stop constructing sudog type

The compiler doesn't care about the runtime's sudog type. Stop
constructing it.

Change-Id: If1885fe30b2e215a08d17662eab5ea6d81fe58ab
Reviewed-on: https://go-review.googlesource.com/20797
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
9 years agocmd/compile: don't create 2 Sym's and 2 Node's for every string
Ian Lance Taylor [Thu, 17 Mar 2016 05:22:58 +0000 (22:22 -0700)]
cmd/compile: don't create 2 Sym's and 2 Node's for every string

For every string constant the compiler was creating 2 Sym's and 2
Node's.  It would never refer to them again, but would keep them alive
in gostringpkg.  This changes the code to just use obj.LSym's instead.

When compiling x/tools/go/types, this yields about a 15% reduction in
the number of calls to newname and a 3% reduction in the total number of
Node objects.  Unfortunately I couldn't see any change in compile time,
but reducing memory usage is desirable anyhow.

Passes toolstash -cmp.

Change-Id: I24f1cb1e6cff0a3afba4ca66f7166874917a036b
Reviewed-on: https://go-review.googlesource.com/20792
Reviewed-by: David Crawshaw <crawshaw@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: escape analysis explanations added to -m -m output
David Chase [Mon, 29 Feb 2016 15:43:18 +0000 (10:43 -0500)]
cmd/compile: escape analysis explanations added to -m -m output

This should probably be considered "experimental" at this stage, but
what it needs is feedback from adventurous adopters.  I think the data
structure used for describing escape reasons might be extendable to
allow a cleanup of the underlying algorithms, which suffers from
insufficiently separated concerns (the graph does not deal well with
escape level adjustments, so it is augmented by a second custom-walk
portion of the "flood" phase. It would be better to put it all,
including level adjustments, in a single graph structure, and then
simply flood the graph.

Tweaked to avoid allocations in the no-logging case.

Modified run.go to ignore lines with leading "#" in the output (since
it can never match a line), and in -update_errors to ignore leading
tabs in output lines and to normalize embedded filenames.

Currently requires -m -m because otherwise the noise/update
burden for the other escape tests is considerable.

There is a partial test.  Existing escape analysis tests seem to
cover all except the panic case and what looks like it might be
unreachable code in escape analysis.

Fixes #10526.

Change-Id: I2524fdec54facae48b00b2548e25d9e46fcaf832
Reviewed-on: https://go-review.googlesource.com/18041
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/compile: reuse blocks in critical pass
Todd Neal [Wed, 16 Mar 2016 02:14:04 +0000 (21:14 -0500)]
cmd/compile: reuse blocks in critical pass

If a phi has duplicate arguments, then the new block that is constructed
to remove the critical edge can be used for all of the duplicate
arguments.

read-only data = -904 bytes (-0.058308%)
global text (code) = -2240 bytes (-0.060056%)
Total difference -3144 bytes (-0.056218%)

Change-Id: Iee3762744d6a8c9d26cdfa880bb23feb62b03c9c
Reviewed-on: https://go-review.googlesource.com/20746
Run-TryBot: Todd Neal <todd@tneal.org>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agosyscall: avoid failure in plan9 StartProcess from fd close race
Richard Miller [Thu, 17 Mar 2016 10:20:54 +0000 (10:20 +0000)]
syscall: avoid failure in plan9 StartProcess from fd close race

Between the enumeration of fdsToClose in the parent and the
closing of fds in the child, it's possible for a file to be
closed in another thread. If that file descriptor is reused
when opening the child-parent status pipe, it will be closed
prematurely in the child and the forkExec gets out of sync.
This has been observed to cause failures in builder tests
when the link step of a build is started before the compile
step has run, with "file does not exist" messages as the
visible symptom.

The simple workaround is to check against closing the pipe.
A more comprehensive solution would be to rewrite the fd
closing code to avoid races, along the lines of the long
ago proposed https://golang.org/cl/57890043 - but meanwhile
this correction will prevent some builder failures.

Change-Id: I4ef5eaea70c21d00f4df0e0847a1c5b2966de7da
Reviewed-on: https://go-review.googlesource.com/20800
Run-TryBot: David du Colombier <0intro@gmail.com>
Reviewed-by: David du Colombier <0intro@gmail.com>
9 years agofmt: separate unicode and integer formatting
Martin Möhrmann [Fri, 11 Mar 2016 12:29:23 +0000 (13:29 +0100)]
fmt: separate unicode and integer formatting

Separate unicode formatting into its own fmt_unicode function.
Remove the fmtUnicode wrapper and the f.unicode and f.uniQuote
flags that are not needed anymore. Remove mangling and restoring
of the precision and sharp flags.

Removes the buffer copy needed for %#U by moving
the character encoding before the number encoding.

Changes the behavior of plus and space flag to have
no effect instead of printing a plus or space before "U+".

Always print at least four digits after "U+"
even if precision is set to less than 4.

Change-Id: If9a0ee79e9eca2c76f06a4e0fdd75d98393899ac
Reviewed-on: https://go-review.googlesource.com/20574
Run-TryBot: Rob Pike <r@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rob Pike <r@golang.org>
9 years agocmd/compile: keep value use counts in SSA
Keith Randall [Wed, 16 Mar 2016 03:45:50 +0000 (20:45 -0700)]
cmd/compile: keep value use counts in SSA

Keep track of how many uses each Value has.  Each appearance in
Value.Args and in Block.Control counts once.

The number of uses of a value is generically useful to
constrain rewrite rules.  For instance, we might want to
prevent merging index operations into loads if the same
index expression is used lots of times.

But I have one use in particular for which the use count is required.
We must make sure we don't combine ops with loads if the load has
more than one use.  Otherwise, we may split a single load
into multiple loads and that breaks perceived behavior in
the presence of races.  In particular, the load of m.state
in sync/mutex.go:Lock can't be done twice.  (I have a separate
CL which triggers the mutex failure.  This CL has a test which
demonstrates a similar failure.)

Change-Id: Icaafa479239f48632a069d0c3f624e6ebc6b1f0e
Reviewed-on: https://go-review.googlesource.com/20790
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
9 years agocmd/compile/internal/gc: disable logProgs debug flag
Dave Cheney [Thu, 17 Mar 2016 02:46:43 +0000 (13:46 +1100)]
cmd/compile/internal/gc: disable logProgs debug flag

Spotted while splunking in the compiler with GOGC=off.

name       old time/op     new time/op     delta
Template       407ms ± 5%      402ms ± 6%     ~           (p=0.301 n=20+20)
GoTypes        1.33s ± 2%      1.29s ± 1%   -3.47%        (p=0.000 n=20+20)
Compiler       6.21s ± 1%      5.91s ± 2%   -4.83%        (p=0.000 n=20+20)

name       old alloc/op    new alloc/op    delta
Template      66.8MB ± 0%     63.9MB ± 0%   -4.46%        (p=0.000 n=19+20)
GoTypes        232MB ± 0%      220MB ± 0%   -5.16%        (p=0.000 n=19+17)
Compiler      1.02GB ± 0%     0.97GB ± 0%   -5.81%        (p=0.000 n=20+20)

name       old allocs/op   new allocs/op   delta
Template        789k ± 0%       708k ± 0%  -10.28%        (p=0.000 n=19+20)
GoTypes        2.49M ± 0%      2.20M ± 0%  -11.57%        (p=0.000 n=20+20)
Compiler       10.8M ± 0%       9.4M ± 0%  -12.82%        (p=0.000 n=20+20)

Change-Id: I76615cab912dde10595ca6ab9979ff6c5f1aec49
Reviewed-on: https://go-review.googlesource.com/20782
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
9 years agocmd/link: do not add duplicate symbols to Allsym
Michael Hudson-Doyle [Wed, 16 Mar 2016 23:38:51 +0000 (12:38 +1300)]
cmd/link: do not add duplicate symbols to Allsym

When building shared libraries, all symbols on Allsym are marked reachable.
What I didn't realize was that this includes the ".dup" symbols created when
"dupok" symbols are read from multiple package files. This breaks now because
deadcode makes some assumptions that fail for these ".dup" symbols, but in any
case was a bad idea -- I suspect this change makes libstd.so a bunch smaller,
but creating it was broken before this CL so I can't be sure.

This change simply stops adding these symbols to Allsym, which might make some
of the many iterations over Allsym the linker does a touch quicker, although
that's not the motivation here.

Add a test that no symbols called ".dup" makes it into the runtime shared
library.

Fixes #14841

Change-Id: I65dd6e88d150a770db2d01b75cfe5db5fd4f8d25
Reviewed-on: https://go-review.googlesource.com/20780
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: ignore receiver parameters in Eqtype
Matthew Dempsky [Fri, 11 Mar 2016 22:38:16 +0000 (14:38 -0800)]
cmd/compile: ignore receiver parameters in Eqtype

Receiver parameters generally aren't relevant to the function
signature type. In particular:

  1. When checking whether a type's method implements an interface's
     method, we specifically want to ignore the receiver parameters,
     because they'll be different.

  2. When checking interface type equality, interface methods always
     use the same "fakethis" *struct{} type as their receiver.

  3. Finally, method expressions and method values degenerate into
     receiver-less function types.

The only case where we care about receiver types matching is in
addmethod, which is easily handled by adding an extra Eqtype check of
the receiver parameters. Also, added a test for this, since
(surprisingly) there weren't any.

As precedence, go/types.Identical ignores receiver parameters when
comparing go/types.Signature values.

Notably, this allows us to slightly simplify the "implements"
function, which is used for checking whether type/interface t
implements interface iface. Currently, cmd/compile actually works
around Eqtype's receiver parameter checking by creating new throwaway
TFUNC Types without the receiver parameter.

(Worse, the compiler currently only provides APIs to build TFUNC Types
from Nod syntax trees, so building those throwaway types also involves
first building throwaway syntax trees.)

Passes toolstash -cmp.

Change-Id: Ib07289c66feacee284e016bc312e8c5ff674714f
Reviewed-on: https://go-review.googlesource.com/20602
Reviewed-by: Robert Griesemer <gri@golang.org>
9 years agocmd/compile: further sinit.go cleanup
Josh Bleecher Snyder [Mon, 14 Mar 2016 21:03:57 +0000 (14:03 -0700)]
cmd/compile: further sinit.go cleanup

Follow-up to CL 20674.

Passes toolstash -cmp.

Change-Id: I065fd4cd80d996c1e6566773189401ca4630c1ca
Reviewed-on: https://go-review.googlesource.com/20692
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Dave Cheney <dave@cheney.net>
9 years agocmd/internal/obj/x86: estimate text size
Josh Bleecher Snyder [Mon, 14 Mar 2016 19:49:58 +0000 (12:49 -0700)]
cmd/internal/obj/x86: estimate text size

We can’t perfectly predict how large the function
will be, but we can make a safe overestimate.
No significant CPU time changes.

name       old alloc/op    new alloc/op    delta
Template      67.7MB ± 0%     67.5MB ± 0%   -0.24%          (p=0.029 n=4+4)
Unicode       43.9MB ± 0%     43.8MB ± 0%   -0.13%          (p=0.029 n=4+4)
GoTypes        244MB ± 0%      244MB ± 0%   -0.28%          (p=0.029 n=4+4)
Compiler      1.05GB ± 0%     1.05GB ± 0%   -0.38%          (p=0.029 n=4+4)

name       old allocs/op   new allocs/op   delta
Template        795k ± 0%       794k ± 0%   -0.14%          (p=0.029 n=4+4)
Unicode         569k ± 0%       569k ± 0%     ~             (p=0.114 n=4+4)
GoTypes        2.59M ± 0%      2.58M ± 0%   -0.11%          (p=0.029 n=4+4)
Compiler       11.0M ± 0%      11.0M ± 0%   -0.09%          (p=0.029 n=4+4)

Passes toolstash -cmp.

Change-Id: I0a92ab04cba7520540ec58fe7189666d0e771454
Reviewed-on: https://go-review.googlesource.com/20771
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>

9 years agocmd/internal/obj: convert Symgrow to a method
Josh Bleecher Snyder [Wed, 16 Mar 2016 21:10:51 +0000 (14:10 -0700)]
cmd/internal/obj: convert Symgrow to a method

Passes toolstash -cmp.

Change-Id: I77a415a4e5d8de7eb902fb0866aaf8783259485a
Reviewed-on: https://go-review.googlesource.com/20770
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/cgo: add C.CBytes
James Bardin [Wed, 16 Mar 2016 17:53:53 +0000 (13:53 -0400)]
cmd/cgo: add C.CBytes

Add a C.CBytes function to copy a Go byte slice into C memory. This
returns an unsafe.Pointer, since that is what needs to be passed to
C.free, and the data is often opaque bytes anyway.

Fixes #14838

Change-Id: Ic7bc29637eb6f1f5ee409b3898c702a59833a85a
Reviewed-on: https://go-review.googlesource.com/20762
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/compile: omit write barrier when assigning global function
Austin Clements [Wed, 16 Mar 2016 22:22:58 +0000 (18:22 -0400)]
cmd/compile: omit write barrier when assigning global function

Currently we generate write barriers when the right side of an
assignment is a global function. This doesn't fall into the existing
case of storing an address of a global because we haven't lowered the
function to a pointer yet.

This write barrier is unnecessary, so eliminate it.

Fixes #13901.

Change-Id: Ibc10e00a8803db0fd75224b66ab94c3737842a79
Reviewed-on: https://go-review.googlesource.com/20772
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
9 years agocmd/compile: make sinit consts Go-ish
Josh Bleecher Snyder [Mon, 14 Mar 2016 00:06:10 +0000 (17:06 -0700)]
cmd/compile: make sinit consts Go-ish

Passes toolstash -cmp.

Change-Id: Ie11912a16d2cd54500e2f6e84316519b80e7c304
Reviewed-on: https://go-review.googlesource.com/20672
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/internal/obj/x86: clean up asm buffer
Josh Bleecher Snyder [Wed, 16 Mar 2016 00:26:28 +0000 (17:26 -0700)]
cmd/internal/obj/x86: clean up asm buffer

c2go translated writing and advancing a pointer using slices.
Switch to something more idiomatic.
It is also more efficient, but not enough to matter.

Change-Id: I67709632ac53253615a35365824ae97bbe5458d5
Reviewed-on: https://go-review.googlesource.com/20767
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocmd/internal/obj/x86: clean up part of span6
Josh Bleecher Snyder [Mon, 14 Mar 2016 19:48:24 +0000 (12:48 -0700)]
cmd/internal/obj/x86: clean up part of span6

Passes toolstash -cmp.

Change-Id: I38eb507de2e9dc2cf01822e420bf31a91fb1b720
Reviewed-on: https://go-review.googlesource.com/20766
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: remove dead code handling '~' operator
Robert Griesemer [Wed, 16 Mar 2016 21:02:56 +0000 (14:02 -0700)]
cmd/compile: remove dead code handling '~' operator

The parser code was not reachable ever since some of the lexer cleanups.
We could recognize '~' in the lexer, complain, and return a '^' instead,
but it's been a few years since Go was new and this may have been a use-
ful error. The lexer complains with "illegal character U+007E '~'" which
is good enough.

For #13244.

Change-Id: Ie3283738486eb6f8462d594f2728ac98333c0520
Reviewed-on: https://go-review.googlesource.com/20768
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
9 years agonet/http: remove init func reference to ServeMux
Brad Fitzpatrick [Wed, 16 Mar 2016 20:14:13 +0000 (20:14 +0000)]
net/http: remove init func reference to ServeMux

Shrinks cmd/go by 30KB.

Change-Id: Ied31192e85af76ebac743f8cc12bd9ef6ec5048f
Reviewed-on: https://go-review.googlesource.com/20765
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

9 years agocmd/compile: move LSym.RefIdx for better packing
Josh Bleecher Snyder [Wed, 16 Mar 2016 19:41:55 +0000 (12:41 -0700)]
cmd/compile: move LSym.RefIdx for better packing

Change-Id: I0516d49ee8381c5e022d77c2fb41515c01c8a631
Reviewed-on: https://go-review.googlesource.com/20764
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agocmd/internal/obj: remove LSym.Etext
Josh Bleecher Snyder [Tue, 15 Mar 2016 05:28:04 +0000 (22:28 -0700)]
cmd/internal/obj: remove LSym.Etext

Use a local variable instead.

Passes toolstash -cmp.

Change-Id: I9623a40ff0d568f11afd1279b6aaa1c33eda644c
Reviewed-on: https://go-review.googlesource.com/20730
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/internal/obj: remove LSym.Next
Josh Bleecher Snyder [Tue, 15 Mar 2016 04:51:09 +0000 (21:51 -0700)]
cmd/internal/obj: remove LSym.Next

Instead, use a slice.

Passes toolstash -cmp.

Change-Id: I889fdb4ae997416f907522f549b96506be13bec7
Reviewed-on: https://go-review.googlesource.com/20699
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/internal/obj: remove LSym.Value
Josh Bleecher Snyder [Tue, 15 Mar 2016 04:14:15 +0000 (21:14 -0700)]
cmd/internal/obj: remove LSym.Value

It is unused.

Passes toolstash -cmp.

Change-Id: I22ae2bb432ce6be377dea43cf018ffccb6e95f37
Reviewed-on: https://go-review.googlesource.com/20698
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: shrink stacks during concurrent mark
Austin Clements [Mon, 15 Feb 2016 23:30:48 +0000 (18:30 -0500)]
runtime: shrink stacks during concurrent mark

Currently we shrink stacks during STW mark termination because it used
to be unsafe to shrink them concurrently. For some programs, this
significantly increases pause time: stack shrinking costs ~5ms/MB
copied plus 2µs/shrink.

Now that we've made it safe to shrink a stack without the world being
stopped, shrink them during the concurrent mark phase.

This reduces the STW time in the program from issue #12967 by an order
of magnitude and brings it from over the 10ms goal to well under:

name           old 95%ile-markTerm-time  new 95%ile-markTerm-time  delta
Stackshrink-4               23.8ms ±60%               1.80ms ±39%  -92.44%  (p=0.008 n=5+5)

Fixes #12967.

This slows down the go1 and garbage benchmarks overall by < 0.5%.

name              old time/op  new time/op  delta
XBenchGarbage-12  2.48ms ± 1%  2.49ms ± 1%  +0.45%  (p=0.005 n=25+21)

name                      old time/op    new time/op    delta
BinaryTree17-12              2.93s ± 2%     2.97s ± 2%  +1.34%  (p=0.002 n=19+20)
Fannkuch11-12                2.51s ± 1%     2.59s ± 0%  +3.09%  (p=0.000 n=18+18)
FmtFprintfEmpty-12          51.1ns ± 2%    51.5ns ± 1%    ~     (p=0.280 n=20+17)
FmtFprintfString-12          175ns ± 1%     169ns ± 1%  -3.01%  (p=0.000 n=20+20)
FmtFprintfInt-12             160ns ± 1%     160ns ± 0%  +0.53%  (p=0.000 n=20+20)
FmtFprintfIntInt-12          265ns ± 0%     266ns ± 1%  +0.59%  (p=0.000 n=20+20)
FmtFprintfPrefixedInt-12     237ns ± 1%     238ns ± 1%  +0.44%  (p=0.000 n=20+20)
FmtFprintfFloat-12           326ns ± 1%     341ns ± 1%  +4.55%  (p=0.000 n=20+19)
FmtManyArgs-12              1.01µs ± 0%    1.02µs ± 0%  +0.43%  (p=0.000 n=20+19)
GobDecode-12                8.41ms ± 1%    8.30ms ± 2%  -1.22%  (p=0.000 n=20+19)
GobEncode-12                6.66ms ± 1%    6.68ms ± 0%  +0.30%  (p=0.000 n=18+19)
Gzip-12                      322ms ± 1%     322ms ± 1%    ~     (p=1.000 n=20+20)
Gunzip-12                   42.8ms ± 0%    42.9ms ± 0%    ~     (p=0.174 n=20+20)
HTTPClientServer-12         69.7µs ± 1%    70.6µs ± 1%  +1.20%  (p=0.000 n=20+20)
JSONEncode-12               16.8ms ± 0%    16.8ms ± 1%    ~     (p=0.154 n=19+19)
JSONDecode-12               65.1ms ± 0%    65.3ms ± 1%  +0.34%  (p=0.003 n=20+20)
Mandelbrot200-12            3.93ms ± 0%    3.92ms ± 0%    ~     (p=0.396 n=19+20)
GoParse-12                  3.66ms ± 1%    3.65ms ± 1%    ~     (p=0.117 n=16+18)
RegexpMatchEasy0_32-12      85.0ns ± 2%    85.5ns ± 2%    ~     (p=0.143 n=20+20)
RegexpMatchEasy0_1K-12       267ns ± 1%     267ns ± 1%    ~     (p=0.867 n=20+17)
RegexpMatchEasy1_32-12      83.3ns ± 2%    83.8ns ± 1%    ~     (p=0.068 n=20+20)
RegexpMatchEasy1_1K-12       432ns ± 1%     432ns ± 1%    ~     (p=0.804 n=20+19)
RegexpMatchMedium_32-12      133ns ± 0%     133ns ± 0%    ~     (p=1.000 n=20+20)
RegexpMatchMedium_1K-12     40.3µs ± 1%    40.4µs ± 1%    ~     (p=0.319 n=20+19)
RegexpMatchHard_32-12       2.10µs ± 1%    2.10µs ± 1%    ~     (p=0.723 n=20+18)
RegexpMatchHard_1K-12       63.0µs ± 0%    63.0µs ± 0%    ~     (p=0.158 n=19+17)
Revcomp-12                   461ms ± 1%     476ms ± 8%  +3.29%  (p=0.002 n=20+20)
Template-12                 80.1ms ± 1%    79.3ms ± 1%  -1.00%  (p=0.000 n=20+20)
TimeParse-12                 360ns ± 0%     360ns ± 0%    ~     (p=0.802 n=18+19)
TimeFormat-12                374ns ± 1%     372ns ± 0%  -0.77%  (p=0.000 n=20+19)
[Geo mean]                  61.8µs         62.0µs       +0.40%

Change-Id: Ib60cd46b7a4987e07670eb271d22f6cee5802842
Reviewed-on: https://go-review.googlesource.com/20044
Reviewed-by: Keith Randall <khr@golang.org>
9 years agoruntime: generalize work.finalizersDone to work.markrootDone
Austin Clements [Mon, 15 Feb 2016 23:24:06 +0000 (18:24 -0500)]
runtime: generalize work.finalizersDone to work.markrootDone

We're about to add another root marking job that needs to happen only
during the first markroot pass (whether that's concurrent or STW),
just like finalizer scanning. Rather than introducing another flag
that has the same value as finalizersDone, just rename finalizersDone
to markrootDone.

Change-Id: I535356c6ea1f3734cb5b6add264cb7bf48de95e8
Reviewed-on: https://go-review.googlesource.com/20043
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 shrinkstack concurrent-safe
Austin Clements [Mon, 15 Feb 2016 22:38:06 +0000 (17:38 -0500)]
runtime: make shrinkstack concurrent-safe

Currently shinkstack is only safe during STW because it adjusts
channel-related stack pointers and moves send/receive stack slots
without synchronizing with the channel code. Make it safe to use when
the world isn't stopped by:

1) Locking all channels the G is blocked on while adjusting the sudogs
   and copying the area of the stack that may contain send/receive
   slots.

2) For any stack frames that may contain send/receive slot, using an
   atomic CAS to adjust pointers to prevent races between adjusting a
   pointer in a receive slot and a concurrent send writing to that
   receive slot.

In principle, the synchronization could be finer-grained. For example,
we considered synchronizing around the sudogs, which would allow
channel operations involving other Gs to continue if the G being
shrunk was far enough down the send/receive queue. However, using the
channel lock means no additional locks are necessary in the channel
code. Furthermore, the stack shrinking code holds the channel lock for
a very short time (much less than the time required to shrink the
stack).

This does not yet make stack shrinking concurrent; it merely makes
doing so safe.

This has negligible effect on the go1 and garbage benchmarks.

For #12967.

Change-Id: Ia49df3a8a7be4b36e365aac4155a2416b94b988c
Reviewed-on: https://go-review.googlesource.com/20042
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

9 years agoruntime: define lock order between G status and channel lock
Austin Clements [Sun, 21 Feb 2016 15:40:39 +0000 (10:40 -0500)]
runtime: define lock order between G status and channel lock

Currently, locking a G's stack by setting its status to _Gcopystack or
_Gscan is unordered with respect to channel locks. However, when we
make stack shrinking concurrent, stack shrinking will need to lock the
G and then acquire channel locks, which imposes an order on these.

Document this lock ordering and fix closechan to respect it.
Everything else already happens to respect it.

For #12967.

Change-Id: I4dd02675efffb3e7daa5285cf75bf24f987d90d4
Reviewed-on: https://go-review.googlesource.com/20041
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: protect sudog.elem with hchan.lock
Austin Clements [Thu, 18 Feb 2016 14:34:43 +0000 (09:34 -0500)]
runtime: protect sudog.elem with hchan.lock

Currently sudog.elem is never accessed concurrently, so in several
cases we drop the channel lock just before reading/writing the
sent/received value from/to sudog.elem. However, concurrent stack
shrinking is going to have to adjust sudog.elem to point to the new
stack, which means it needs a way to synchronize with accesses to
sudog.elem. Hence, add sudog.elem to the fields protected by
hchan.lock and scoot the unlocks down past the uses of sudog.elem.

While we're here, better document the channel synchronization rules.

For #12967.

Change-Id: I3ad0ca71f0a74b0716c261aef21b2f7f13f74917
Reviewed-on: https://go-review.googlesource.com/20040
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: fix transient _Gwaiting states in newstack
Austin Clements [Thu, 25 Feb 2016 20:37:40 +0000 (15:37 -0500)]
runtime: fix transient _Gwaiting states in newstack

With concurrent stack shrinking, the stack can move the instant after
a G enters _Gwaiting. There are only two places that put a G into
_Gwaiting: gopark and newstack. We fixed uses of gopark. This commit
fixes newstack by simplifying its G transitions and, in particular,
eliminating or narrowing the transient _Gwaiting states it passes
through so it's clear nothing in the G is accessed while in _Gwaiting.

For #12967.

Change-Id: I2440ead411d2bc61beb1e2ab020ebe3cb3481af9
Reviewed-on: https://go-review.googlesource.com/20039
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: never pass stack pointers to gopark
Austin Clements [Fri, 26 Feb 2016 15:50:54 +0000 (10:50 -0500)]
runtime: never pass stack pointers to gopark

gopark calls the unlock function after setting the G to _Gwaiting.
This means it's generally unsafe to access the G's stack from the
unlock function because the G may start running on another P. Once we
start shrinking stacks concurrently, a stack shrink could also move
the stack the moment after it enters _Gwaiting and before the unlock
function is called.

Document this restriction and fix the two places where we currently
violate it.

This is unlikely to be a problem in practice for these two places
right now, but they're already skating on thin ice. For example, the
following sequence could in principle cause corruption, deadlock, or a
panic in the select code:

On M1/P1:
1. G1 selects on channels A and B.
2. selectgoImpl calls gopark.
3. gopark puts G1 in _Gwaiting.
4. gopark calls selparkcommit.
5. selparkcommit releases the lock on channel A.

On M2/P2:
6. G2 sends to channel A.
7. The send puts G1 in _Grunnable and puts it on P2's run queue.
8. The scheduler runs, selects G1, puts it in _Grunning, and resumes G1.
9. On G1, the sellock immediately following the gopark gets called.
10. sellock grows and moves the stack.

On M1/P1:
11. selparkcommit continues to scan the lock order for the next
channel to unlock, but it's now reading from a freed (and possibly
reused) stack.

This shouldn't happen in practice because step 10 isn't the first call
to sellock, so the stack should already be big enough. However, once
we start shrinking stacks concurrently, this reasoning won't work any
more.

For #12967.

Change-Id: I3660c5be37e5be9f87433cb8141bdfdf37fadc4c
Reviewed-on: https://go-review.googlesource.com/20038
Reviewed-by: Rick Hudson <rlh@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: put g.waiting list in lock order
Austin Clements [Tue, 16 Feb 2016 16:06:00 +0000 (11:06 -0500)]
runtime: put g.waiting list in lock order

Currently the g.waiting list created by a select is in poll order.
However, nothing depends on this, and we're going to need access to
the channel lock order in other places shortly, so modify select to
put the waiting list in channel lock order.

For #12967.

Change-Id: If0d38816216ecbb37a36624d9b25dd96e0a775ec
Reviewed-on: https://go-review.googlesource.com/20037
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>

9 years agoruntime: use indexes for select lock order
Austin Clements [Tue, 16 Feb 2016 04:50:58 +0000 (23:50 -0500)]
runtime: use indexes for select lock order

Currently the select lock order is a []*hchan. We're going to need to
refer to things other than the channel itself in lock order shortly,
so switch this to a []uint16 of indexes into the select cases. This
parallels the existing representation for the poll order.

Change-Id: I89262223fe20b4ddf5321592655ba9eac489cda1
Reviewed-on: https://go-review.googlesource.com/20036
Reviewed-by: Rick Hudson <rlh@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>