]> Cypherpunks repositories - gostls13.git/log
gostls13.git
9 years ago[dev.ssa] cmd/compile: use INC/DEC instead of add when we can
Ilya Tocar [Fri, 5 Feb 2016 16:24:53 +0000 (19:24 +0300)]
[dev.ssa] cmd/compile: use INC/DEC instead of add when we can

INC/DEC produces slightly faster and smaller code.

Change-Id: I329d9bdb01b90041be45e053d9df640818bf0c2d
Reviewed-on: https://go-review.googlesource.com/19238
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile/internal/ssa/gen: constant fold Neg*.
Alexandru Moșoi [Mon, 8 Feb 2016 17:55:56 +0000 (18:55 +0100)]
[dev.ssa] cmd/compile/internal/ssa/gen: constant fold Neg*.

Change-Id: Id51e5c97e9653b764b809bf3424f1a6d31b6ffea
Reviewed-on: https://go-review.googlesource.com/19338
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: fix for bug in cse speed improvements
David Chase [Mon, 8 Feb 2016 17:07:39 +0000 (12:07 -0500)]
[dev.ssa] cmd/compile: fix for bug in cse speed improvements

Problem was caused by use of Args[].Aux differences
in early partitioning.  This artificially separated
two equivalent expressions because sort ignores the
Aux field, hence things can end with equal things
separated by unequal things and thus the equal things
are split into more than one partition.  For example:
SliceLen(a), SliceLen(b), SliceLen(a).

Fix: don't use Args[].Aux in initial partitioning.

Left in a debugging flag and some debugging Fprintf's;
not sure if that is house style or not.  We'll probably
want to be more systematic in our naming conventions,
e.g. ssa.cse, ssa.scc, etc.

Change-Id: Ib1412539cc30d91ea542c0ac7b2f9b504108ca7f
Reviewed-on: https://go-review.googlesource.com/19316
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: speed up cse
Todd Neal [Sun, 7 Feb 2016 02:56:50 +0000 (20:56 -0600)]
[dev.ssa] cmd/compile: speed up cse

Examine both Aux and AuxInt to form more precise initial partitions.
Restructure loop to avoid repeated type.Equal() call.  Speeds up
compilation of testdata/gen/arithConst_ssa by 25%.

Change-Id: I3cfb1d254adf0601ee69239e1885b0cf2a23575b
Reviewed-on: https://go-review.googlesource.com/19313
Run-TryBot: Todd Neal <todd@tneal.org>
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: panic doesn't return
Keith Randall [Sun, 7 Feb 2016 06:35:34 +0000 (22:35 -0800)]
[dev.ssa] cmd/compile: panic doesn't return

Panic doesn't return, so record that we immediately exit after a panic
call.  This will help code analysis.

Change-Id: I4d1f67494f97b6aee130c43ff4e44307b2b0f149
Reviewed-on: https://go-review.googlesource.com/19303
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: strength-reduce 64-bit constant divides
Keith Randall [Sat, 6 Feb 2016 04:26:18 +0000 (20:26 -0800)]
[dev.ssa] cmd/compile: strength-reduce 64-bit constant divides

The frontend does this for 32 bits and below, but SSA needs
to do it for 64 bits.  The algorithms are all copied from
cgen.go:cgen_div.

Speeds up TimeFormat substantially: ~40% slower to ~10% slower.

Change-Id: I023ea2eb6040df98ccd9105e15ca6ea695610a7a
Reviewed-on: https://go-review.googlesource.com/19302
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
9 years ago[dev.ssa] cmd/compile: reorg write barriers a bit
Keith Randall [Sat, 30 Jan 2016 05:57:57 +0000 (21:57 -0800)]
[dev.ssa] cmd/compile: reorg write barriers a bit

Use just a single write barrier flag test, even if there
are multiple pointer fields in a struct.

This helps move more of the wb-specific code (like the LEA
needed to materialize the write address) into the unlikely path.

Change-Id: Ic7a67145904369c4ff031e464d51267d71281c8f
Reviewed-on: https://go-review.googlesource.com/19085
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: add store constant indexed operations
Keith Randall [Thu, 4 Feb 2016 23:53:33 +0000 (15:53 -0800)]
[dev.ssa] cmd/compile: add store constant indexed operations

Change-Id: Ifb8eba1929c79ee7a8cae2191613c55a3b8f74e5
Reviewed-on: https://go-review.googlesource.com/19236
Reviewed-by: Todd Neal <todd@tneal.org>
9 years ago[dev.ssa] cmd/compile: Use ADD instead of LEA when we can
Keith Randall [Thu, 4 Feb 2016 23:08:47 +0000 (15:08 -0800)]
[dev.ssa] cmd/compile: Use ADD instead of LEA when we can

If the output register is one of the input registers,
we can use a real add instead of LEA.

Change-Id: Ide58f1536afb077c0b939d3a8c7555807fd1c5e3
Reviewed-on: https://go-review.googlesource.com/19234
Reviewed-by: Alexandru Moșoi <alexandru@mosoi.ro>
9 years ago[dev.ssa] cmd/compile/internal/ssa/gen: enclose rules' code in a for loop
Alexandru Moșoi [Thu, 4 Feb 2016 18:52:10 +0000 (19:52 +0100)]
[dev.ssa] cmd/compile/internal/ssa/gen: enclose rules' code in a for loop

* Enclose each rule's code in a for with no condition
* The loop is ran at most once because it's always terminated by a return.
* Use break when matching condition fails
* Drop rule hashes
* Shaves about 3 lines of code per rule

The binary size is not afected.

Change-Id: I27c3e40dc8cae98dcd50739342dc38db2ef9c247
Reviewed-on: https://go-review.googlesource.com/19220
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: more combining of ops into instructions
Keith Randall [Tue, 2 Feb 2016 19:13:50 +0000 (11:13 -0800)]
[dev.ssa] cmd/compile: more combining of ops into instructions

Mostly indexed loads.  A few more LEA cases.

Change-Id: Idc1d447ed0dd6e906cd48e70307a95e77f61cf5f
Reviewed-on: https://go-review.googlesource.com/19172
Reviewed-by: Todd Neal <todd@tneal.org>
Run-TryBot: Keith Randall <khr@golang.org>

9 years ago[dev.ssa] cmd/internal/obj/x86: don't clobber flags with dynlink rewrite
Keith Randall [Thu, 4 Feb 2016 19:21:31 +0000 (11:21 -0800)]
[dev.ssa] cmd/internal/obj/x86: don't clobber flags with dynlink rewrite

LEAQ symbol+100(SB), AX

Under dynamic link, rewrites to

MOVQ symbol@GOT(SB), AX
ADDQ $100, AX

but ADDQ clobbers flags, whereas the original LEAQ (when not dynamic
linking) doesn't.

Use LEAQ instead of ADDQ to add that constant in so we preserve flags.

Change-Id: Ibb055403d94a4c5163e1c7d2f45da633ffd0b6a3
Reviewed-on: https://go-review.googlesource.com/19230
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years ago[dev.ssa] cmd/compile/internal/ssa/gen: move variable reset code into a function
Alexandru Moșoi [Thu, 4 Feb 2016 16:21:57 +0000 (17:21 +0100)]
[dev.ssa] cmd/compile/internal/ssa/gen: move variable reset code into a function

Shaves about 3 lines per generated rule.

Change-Id: I94adc94ab79f90ac5fd033f896ece3b1eddf0f3d
Reviewed-on: https://go-review.googlesource.com/19197
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: rewrites for constant shifts
Todd Neal [Wed, 3 Feb 2016 11:21:24 +0000 (06:21 -0500)]
[dev.ssa] cmd/compile: rewrites for constant shifts

Add rewrite rules to optimize constant shifts.

Fixes #10637

Change-Id: I74b724d3e81aeb7098c696d02c050f7fdfd5b523
Reviewed-on: https://go-review.googlesource.com/19106
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Todd Neal <todd@tneal.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years ago[dev.ssa] src/cmd/compile/internal/ssa/gen: detect type earlier when generating rules.
Alexandru Moșoi [Wed, 3 Feb 2016 19:50:12 +0000 (20:50 +0100)]
[dev.ssa] src/cmd/compile/internal/ssa/gen: detect type earlier when generating rules.

Removes approx. one assignment per rule.

Change-Id: Ie9f0a7082ae12c4447ff6b4d40678cd92bdbb6f2
Reviewed-on: https://go-review.googlesource.com/19194
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile/internal/ssa: simplify comparisons with constants
Alexandru Moșoi [Wed, 3 Feb 2016 18:43:46 +0000 (19:43 +0100)]
[dev.ssa] cmd/compile/internal/ssa: simplify comparisons with constants

* Simplify comparisons of form a + const1 == const2 or a + const1 != const2.
* Canonicalize Eq, Neq, Add, Sub to have a constant as first argument.
Needed for the above new rules and helps constant folding.

Change-Id: I8078702a5daa706da57106073a3e9f640a67f486
Reviewed-on: https://go-review.googlesource.com/19192
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: use sparsetree in checkFunc
Todd Neal [Thu, 4 Feb 2016 02:06:21 +0000 (21:06 -0500)]
[dev.ssa] cmd/compile: use sparsetree in checkFunc

Modify the simple domCheck to use the sparse tree code.  This
speeds up compilation of one of the generated test cases from
1m48s to 17s.

Change-Id: If577410ee77b54918147a66917a8e3721297ee0a
Reviewed-on: https://go-review.googlesource.com/19187
Run-TryBot: Todd Neal <todd@tneal.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: remove dead code
Todd Neal [Thu, 4 Feb 2016 00:27:43 +0000 (19:27 -0500)]
[dev.ssa] cmd/compile: remove dead code

Change-Id: I1738e3af7de0972c54d74325d80781059d0796d8
Reviewed-on: https://go-review.googlesource.com/19186
Run-TryBot: Todd Neal <todd@tneal.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: cache sparse sets in Config
Todd Neal [Tue, 2 Feb 2016 11:35:34 +0000 (06:35 -0500)]
[dev.ssa] cmd/compile: cache sparse sets in Config

Move the cached sparse sets to the Config.  I tested make.bash with
pre-allocating sets of size 150 and not caching very small sets, but the
difference between this implementation (no min size, no preallocation)
and a min size with preallocation was fairly negligible:

Number of sparse sets allocated:
Cached in Config w/none preallocated no min size    3684 *this CL*
Cached in Config w/three preallocated no min size   3370
Cached in Config w/three preallocated min size=150  3370
Cached in Config w/none preallocated min size=150  15947
Cached in Func,  w/no min                          96996 *previous code*

Change-Id: I7f9de8a7cae192648a7413bfb18a6690fad34375
Reviewed-on: https://go-review.googlesource.com/19152
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: add aux typing, flags to ops
Keith Randall [Sun, 31 Jan 2016 19:39:39 +0000 (11:39 -0800)]
[dev.ssa] cmd/compile: add aux typing, flags to ops

Add the aux type to opcodes.
Add rematerializeable as a flag.

Change-Id: I906e19281498f3ee51bb136299bf26e13a54b2ec
Reviewed-on: https://go-review.googlesource.com/19088
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
9 years ago[dev.ssa] cmd/compile: reducing alloc footprint of dominator calc
David Chase [Sat, 30 Jan 2016 22:37:38 +0000 (17:37 -0500)]
[dev.ssa] cmd/compile: reducing alloc footprint of dominator calc

Converted working slices of pointer into slices of pointer
index.  Half the size (on 64-bit machine) and no pointers
to trace if GC occurs while they're live.

TODO - could expose slice mapping ID->*Block; some dom
clients also construct these.

Minor optimization in regalloc that cuts allocation count.

Minor optimization in compile.go that cuts calls to Sprintf.

Change-Id: I28f0bfed422b7344af333dc52ea272441e28e463
Reviewed-on: https://go-review.googlesource.com/19104
Run-TryBot: Todd Neal <todd@tneal.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
9 years ago[dev.ssa] cmd/compile: regenerate code
Todd Neal [Mon, 1 Feb 2016 03:18:55 +0000 (21:18 -0600)]
[dev.ssa] cmd/compile: regenerate code

These changes were left out of the 1cc5789df9 commit.

Change-Id: Id7f49973da281a567b69228dbaea31846b82b4af
Reviewed-on: https://go-review.googlesource.com/19105
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: lots of small rewrite optimizations
Keith Randall [Sat, 30 Jan 2016 19:25:38 +0000 (11:25 -0800)]
[dev.ssa] cmd/compile: lots of small rewrite optimizations

Small optimizations I noticed while looking at Giovanni's test cases.

More shifts by constants.
Indexed stores for smaller types.
Fold LEA into loads/stores.
More extending loads.
CMP $0 of AND -> TEST

Fix order of TEST ops.

Giovanni's test cases at https://gist.github.com/rasky/62fba94e3a20d1b05b2a

Change-Id: I7077bc0b5319bf05767eeb39f401f4bb4b39f635
Reviewed-on: https://go-review.googlesource.com/19086
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Todd Neal <todd@tneal.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: reuse sparse sets across compiler passes
Todd Neal [Fri, 29 Jan 2016 04:19:46 +0000 (22:19 -0600)]
[dev.ssa] cmd/compile: reuse sparse sets across compiler passes

Cache sparse sets in the function so they can be reused by subsequent
compiler passes.

benchmark                        old ns/op     new ns/op     delta
BenchmarkDSEPass-8               206945        180022        -13.01%
BenchmarkDSEPassBlock-8          5286103       2614054       -50.55%
BenchmarkCSEPass-8               1790277       1790655       +0.02%
BenchmarkCSEPassBlock-8          18083588      18112771      +0.16%
BenchmarkDeadcodePass-8          59837         41375         -30.85%
BenchmarkDeadcodePassBlock-8     1651575       511169        -69.05%
BenchmarkMultiPass-8             531529        427506        -19.57%
BenchmarkMultiPassBlock-8        7033496       4487814       -36.19%

benchmark                        old allocs     new allocs     delta
BenchmarkDSEPass-8               11             4              -63.64%
BenchmarkDSEPassBlock-8          599            120            -79.97%
BenchmarkCSEPass-8               18             18             +0.00%
BenchmarkCSEPassBlock-8          2700           2700           +0.00%
BenchmarkDeadcodePass-8          4              3              -25.00%
BenchmarkDeadcodePassBlock-8     30             9              -70.00%
BenchmarkMultiPass-8             24             20             -16.67%
BenchmarkMultiPassBlock-8        1800           1000           -44.44%

benchmark                        old bytes     new bytes     delta
BenchmarkDSEPass-8               221367        142           -99.94%
BenchmarkDSEPassBlock-8          3695207       3846          -99.90%
BenchmarkCSEPass-8               303328        303328        +0.00%
BenchmarkCSEPassBlock-8          5006400       5006400       +0.00%
BenchmarkDeadcodePass-8          84232         10506         -87.53%
BenchmarkDeadcodePassBlock-8     1274940       163680        -87.16%
BenchmarkMultiPass-8             608674        313834        -48.44%
BenchmarkMultiPassBlock-8        9906001       5003450       -49.49%

Change-Id: Ib1fa58c7f494b374d1a4bb9cffbc2c48377b59d3
Reviewed-on: https://go-review.googlesource.com/19100
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: exposed do-log boolean to reduce allocations
David Chase [Fri, 29 Jan 2016 19:44:15 +0000 (14:44 -0500)]
[dev.ssa] cmd/compile: exposed do-log boolean to reduce allocations

From memory profiling, about 3% reduction in allocation count.

Change-Id: I4b662d55b8a94fe724759a2b22f05a08d0bf40f8
Reviewed-on: https://go-review.googlesource.com/19103
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: tweak init function prologue
Keith Randall [Tue, 26 Jan 2016 23:55:05 +0000 (15:55 -0800)]
[dev.ssa] cmd/compile: tweak init function prologue

We used to compare the init state with == to 0 and 2, which
requires 2 comparisons.  Instead, compare with 1 and use
<, ==.  That requires only one comparison.

This isn't a big deal performance-wise, as it is just init
code.  But there is a fair amount of init code, so this
should help a bit with code size.

Change-Id: I4a2765f1005776f0edce28ac143f4b7596d95a68
Reviewed-on: https://go-review.googlesource.com/18948
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: prepare for some load+op combining
Keith Randall [Tue, 26 Jan 2016 23:47:08 +0000 (15:47 -0800)]
[dev.ssa] cmd/compile: prepare for some load+op combining

Rename StoreConst to ValAndOff so we can use it for other ops.
Make ValAndOff print nicely.

Add some notes & checks related to my aborted attempt to
implement combined CMP+load ops.

Change-Id: I2f901d12d42bc5a82879af0334806aa184a97e27
Reviewed-on: https://go-review.googlesource.com/18947
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: fix write barriers for SSA
Keith Randall [Tue, 26 Jan 2016 01:06:54 +0000 (17:06 -0800)]
[dev.ssa] cmd/compile: fix write barriers for SSA

The old write barriers used _nostore versions, which
don't work for Ian's cgo checker.  Instead, we adopt the
same write barrier pattern as the default compiler.

It's a bit trickier to code up but should be more efficient.

Change-Id: I6696c3656cf179e28f800b0e096b7259bd5f3bb7
Reviewed-on: https://go-review.googlesource.com/18941
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: optimization for && and || expressions
Keith Randall [Mon, 25 Jan 2016 17:21:17 +0000 (09:21 -0800)]
[dev.ssa] cmd/compile: optimization for && and || expressions

Compiling && and || expressions often leads to control
flow of the following form:

p:
  If a goto b else c
b: <- p ...
  x = phi(a, ...)
  If x goto t else u

Note that if we take the edge p->b, then we are guaranteed
to take the edge b->t also.  So in this situation, we might
as well go directly from p to t.

Change-Id: I6974f1e6367119a2ddf2014f9741fdb490edcc12
Reviewed-on: https://go-review.googlesource.com/18910
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: simple forward-looking register allocation tweak
Keith Randall [Tue, 19 Jan 2016 04:00:15 +0000 (20:00 -0800)]
[dev.ssa] cmd/compile: simple forward-looking register allocation tweak

For each value that needs to be in a fixed register at the end of the
block, and try to pick that fixed register when the instruction
generating that value is scheduled (or restored from a spill).

Just used for end-of-block register requirements for now.
Fixed-register instruction requirements (e.g. shift in ecx) can be
added later.  Also two-instruction constraints (input reg == output
reg) might be recorded in a similar manner.

Change-Id: I59916e2e7f73657bb4fc3e3b65389749d7a23fa8
Reviewed-on: https://go-review.googlesource.com/18774
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: fix -N build
Keith Randall [Fri, 29 Jan 2016 01:43:45 +0000 (17:43 -0800)]
[dev.ssa] cmd/compile: fix -N build

The OpSB hack didn't quite work.  We need to really
CSE these ops to make regalloc happy.

Change-Id: I9f4d7bfb0929407c84ee60c9e25ff0c0fbea84af
Reviewed-on: https://go-review.googlesource.com/19083
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: add backing store buffers for block.{Preds,Succs,Values}
Keith Randall [Thu, 28 Jan 2016 23:54:45 +0000 (15:54 -0800)]
[dev.ssa] cmd/compile: add backing store buffers for block.{Preds,Succs,Values}

Speeds up compilation by 6%.

Change-Id: Ibaad95710323ddbe13c1b0351843fe43a48d776e
Reviewed-on: https://go-review.googlesource.com/19080
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years ago[dev.ssa] cmd/compile: fix build
Keith Randall [Fri, 29 Jan 2016 00:11:56 +0000 (16:11 -0800)]
[dev.ssa] cmd/compile: fix build

Some tests make multiple Funcs per Config at once.
With value & block caching, we can't do that any more.

Change-Id: Ibdb60aa2fcf478f1726b3be0fcaa06b04433eb67
Reviewed-on: https://go-review.googlesource.com/19081
Reviewed-by: Keith Randall <khr@golang.org>
9 years ago[dev.ssa] cmd/compile: preallocate small-numbered values and blocks
Keith Randall [Thu, 28 Jan 2016 21:46:30 +0000 (13:46 -0800)]
[dev.ssa] cmd/compile: preallocate small-numbered values and blocks

Speeds up the compiler ~5%.

Change-Id: Ia5cf0bcd58701fd14018ec77d01f03d5c7d6385b
Reviewed-on: https://go-review.googlesource.com/19060
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years ago[dev.ssa] cmd/compile: make cse faster
Keith Randall [Thu, 28 Jan 2016 00:47:23 +0000 (16:47 -0800)]
[dev.ssa] cmd/compile: make cse faster

It is one of the slowest compiler phases right now, and we
run two of them.

Instead of using a map to make the initial partition, use a sort.
It is much less memory intensive.

Do a few optimizations to avoid work for size-1 equivalence classes.

Implement -N.

Change-Id: I1d2d85d3771abc918db4dd7cc30b0b2d854b15e1
Reviewed-on: https://go-review.googlesource.com/19024
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: disable xor clearing when flags must be preserved
Keith Randall [Fri, 22 Jan 2016 21:44:58 +0000 (13:44 -0800)]
[dev.ssa] cmd/compile: disable xor clearing when flags must be preserved

The x86 backend automatically rewrites MOV $0, AX to
XOR AX, AX.  That rewrite isn't ok when the flags register
is live across the MOV.  Keep track of which moves care
about preserving flags, then disable this rewrite for them.

On x86, Prog.Mark was being used to hold the length of the
instruction.  We already store that in Prog.Isize, so no
need to store it in Prog.Mark also.  This frees up Prog.Mark
to hold a bitmask on x86 just like all the other architectures.

Update #12405

Change-Id: Ibad8a8f41fc6222bec1e4904221887d3cc3ca029
Reviewed-on: https://go-review.googlesource.com/18861
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years ago[dev.ssa] cmd/compile: get rid of +0.0 hack
Keith Randall [Tue, 26 Jan 2016 04:26:06 +0000 (20:26 -0800)]
[dev.ssa] cmd/compile: get rid of +0.0 hack

The conversion from -0.0 to +0.0 happens inside mpgetflt now.
The SSA code doesn't need this fix any more.

Change-Id: I6cd4f4a4e75b13cf284ebbb95b08af050ed9891c
Reviewed-on: https://go-review.googlesource.com/18942
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>

9 years ago[dev.ssa] cmd/compile: short-circuit empty blocks
Keith Randall [Thu, 21 Jan 2016 21:27:01 +0000 (13:27 -0800)]
[dev.ssa] cmd/compile: short-circuit empty blocks

Empty blocks are introduced to remove critical edges.
After regalloc, we can remove any of the added blocks
that are still empty.

Change-Id: I0b40e95ac3a6cc1e632a479443479532b6c5ccd9
Reviewed-on: https://go-review.googlesource.com/18833
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: update SSA TODOs
Keith Randall [Wed, 20 Jan 2016 22:06:17 +0000 (14:06 -0800)]
[dev.ssa] cmd/compile: update SSA TODOs

Change-Id: I78743987dcb45d821212caf95a00ae15b7a6cfd8
Reviewed-on: https://go-review.googlesource.com/18773
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years ago[dev.ssa] cmd/compile: Allow structs to be SSAd
Keith Randall [Tue, 12 Jan 2016 05:05:33 +0000 (21:05 -0800)]
[dev.ssa] cmd/compile: Allow structs to be SSAd

Break small structs up into their components so they
can be registerized.

Change StructSelect to use field indexes instead of
field offsets, as field offsets aren't unique in the
presence of zero-sized fields.

Change-Id: I2f1dc89f7fa58e1cf58aa1a32b238959d53f62e4
Reviewed-on: https://go-review.googlesource.com/18570
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: optimize phi ops
Keith Randall [Fri, 15 Jan 2016 00:02:23 +0000 (16:02 -0800)]
[dev.ssa] cmd/compile: optimize phi ops

Redo how we keep track of forward references when building SSA.
When the forward reference is resolved, update the Value node
in place.

Improve the phi elimination pass so it can simplify phis of phis.

Give SSA package access to decoded line numbers.  Fix line numbers
for constant booleans.

Change-Id: I3dc9896148d260be2f3dd14cbe5db639ec9fa6b7
Reviewed-on: https://go-review.googlesource.com/18674
Reviewed-by: David Chase <drchase@google.com>
Run-TryBot: Keith Randall <khr@golang.org>

9 years ago[dev.ssa] Merge remote-tracking branch 'origin/master' into mergebranch
Keith Randall [Tue, 19 Jan 2016 17:59:21 +0000 (09:59 -0800)]
[dev.ssa] Merge remote-tracking branch 'origin/master' into mergebranch

Semi-regular merge from tip to dev.ssa.

Conflicts:
src/runtime/sys_windows_amd64.s

Change-Id: I5f733130049c810e6ceacd46dad85faebca52b29

9 years ago[dev.ssa] cmd/compile: use wider move instruction for floats
Keith Randall [Fri, 15 Jan 2016 16:45:47 +0000 (08:45 -0800)]
[dev.ssa] cmd/compile: use wider move instruction for floats

Distinguish move/load/store ops.  Unify some of this code a bit.

Reduces Mandelbrot slowdown with SSA from 58% to 12%.

Change-Id: I3276eaebcbcdd9de3f8299c79b5f25c0429194c4
Reviewed-on: https://go-review.googlesource.com/18677
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: report better line numbers for Unimplemented/Fatal
Keith Randall [Wed, 13 Jan 2016 19:14:57 +0000 (11:14 -0800)]
[dev.ssa] cmd/compile: report better line numbers for Unimplemented/Fatal

If a failure occurs in SSA processing, we always report the
last line of the function we're compiling.  Modify the callbacks
from SSA to the GC compiler so we can pass a line number back
and use it in Fatalf.

Change-Id: Ifbfad50d5e167e997e0a96f0775bcc369f5c397e
Reviewed-on: https://go-review.googlesource.com/18599
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years agonet: add timeout to DNS requests sent by TestSpecialDomainName
Michael Munday [Thu, 14 Jan 2016 21:36:43 +0000 (16:36 -0500)]
net: add timeout to DNS requests sent by TestSpecialDomainName

The timeout means that TestSpecialDomainName will not hang if
the DNS server does not respond to the request.

Fixes #13939

Change-Id: I46e30bbd3c11b6c560656134e704331cf6f8af3f
Reviewed-on: https://go-review.googlesource.com/18661
Reviewed-by: Mikio Hara <mikioh.mikioh@gmail.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agonet/http: make hidden http2 Transport respect remaining Transport fields
Brad Fitzpatrick [Tue, 19 Jan 2016 05:10:58 +0000 (05:10 +0000)]
net/http: make hidden http2 Transport respect remaining Transport fields

Updates x/net/http2 to git rev 72aa00c6 for https://golang.org/cl/18721
(but actually at https://golang.org/cl/18722 now)

Fixes #14008

Change-Id: If05d5ad51ec0ba5ba7e4fe16605c0a83f0484bc8
Reviewed-on: https://go-review.googlesource.com/18723
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Andrew Gerrand <adg@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agobuild: test linux/mips64 and linux/mips64le with buildall.bash
Shenghou Ma [Tue, 19 Jan 2016 03:55:11 +0000 (22:55 -0500)]
build: test linux/mips64 and linux/mips64le with buildall.bash

Both mips64 architectures share the same runtime/rt0 file, so
we have to hardcode them in buildall.bash.

Ideally we should have cmd/dist report all supported platforms,
see #12270.

Change-Id: I08ce35cfe0a831af5e1e8255b305efd38386fa52
Reviewed-on: https://go-review.googlesource.com/18687
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: print address as hex in messages
Shenghou Ma [Mon, 18 Jan 2016 22:55:06 +0000 (17:55 -0500)]
runtime: print address as hex in messages

Change-Id: I7ccf1b5001d77c4390479f53c0137ab02f98595b
Reviewed-on: https://go-review.googlesource.com/18685
Run-TryBot: Minux Ma <minux@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agonet/http: clarify docs on Request.Proto field
Brad Fitzpatrick [Mon, 18 Jan 2016 06:38:56 +0000 (06:38 +0000)]
net/http: clarify docs on Request.Proto field

No need to say "by default" because there is no alternative and no way
to override. Always HTTP/2.0 is officially spelled HTTP/2 these days.

Fixes #13985 harder

Change-Id: Ib1ec03cec171ca865342b8e7452cd4c707d7b770
Reviewed-on: https://go-review.googlesource.com/18720
Reviewed-by: Rob Pike <r@golang.org>
9 years agonet/http: make http2 Transport send Content Length
Brad Fitzpatrick [Mon, 18 Jan 2016 22:50:52 +0000 (14:50 -0800)]
net/http: make http2 Transport send Content Length

Updates x/net/http2 to git rev 5c0dae8 for https://golang.org/cl/18709

Fixes #14003

Change-Id: I8bc205d6d089107b017e3458bbc7e05f6d0cae60
Reviewed-on: https://go-review.googlesource.com/18730
Reviewed-by: Andrew Gerrand <adg@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agonet/http: panic on bogus use of CloseNotifier or Hijacker
Brad Fitzpatrick [Mon, 18 Jan 2016 19:43:32 +0000 (11:43 -0800)]
net/http: panic on bogus use of CloseNotifier or Hijacker

Fixes #14001

Change-Id: I6f9bc3028345081758d8f537c3aaddb2e254e69e
Reviewed-on: https://go-review.googlesource.com/18708
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 agonet: disable TestInterfaceAddrsWithNetsh on windows
Mikio Hara [Sun, 17 Jan 2016 00:04:58 +0000 (09:04 +0900)]
net: disable TestInterfaceAddrsWithNetsh on windows

Updates #13981.

Change-Id: Id8f3cd56a81a7a993cea5c757e619407da491fed
Reviewed-on: https://go-review.googlesource.com/18710
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agoruntime: don't run TestEnsureDropM on windows or plan9
Ian Lance Taylor [Mon, 18 Jan 2016 17:01:48 +0000 (09:01 -0800)]
runtime: don't run TestEnsureDropM on windows or plan9

This is testing code in asm_GOARCH.s, so it's not necessary to run the
test on systems where it doesn't build.

Fixes #13991.

Change-Id: Ia7a2d3a34b32e6987dc67428c1e09e63baf0518a
Reviewed-on: https://go-review.googlesource.com/18707
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agonet/http: update bundled http2, pass through Transport.CloseIdleConnections
Brad Fitzpatrick [Fri, 15 Jan 2016 18:53:51 +0000 (10:53 -0800)]
net/http: update bundled http2, pass through Transport.CloseIdleConnections

Wire up Transport.CloseIdleConnections to http2.Transport.CloseIdleConnections.

Updates x/net/http2 to git rev c92cdcb0 for https://golang.org/cl/18678

Fixes #13975

Change-Id: I1183a31256104ff95ae7621e5788cfeee741b1aa
Reviewed-on: https://go-review.googlesource.com/18679
Reviewed-by: Andrew Gerrand <adg@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agoruntime: fix sleep/wakeup race for GC assists
Austin Clements [Fri, 15 Jan 2016 18:28:41 +0000 (13:28 -0500)]
runtime: fix sleep/wakeup race for GC assists

GC assists check gcBlackenEnabled under the assist queue lock to avoid
going to sleep after gcWakeAllAssists has already woken all assists.
However, currently we clear gcBlackenEnabled shortly *after* waking
all assists, which opens a window where this exact race can happen.

Fix this by clearing gcBlackenEnabled before waking blocked assists.
However, it's unlikely this actually matters because the world is
stopped between waking assists and clearing gcBlackenEnabled and there
aren't any obvious allocations during this window, so I don't think an
assist could actually slip in to this race window.

Updates #13645.

Change-Id: I7571f059530481dc781d8fd96a1a40aadebecb0d
Reviewed-on: https://go-review.googlesource.com/18682
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
9 years agoruntime: readjust signal code for dragonfly-2.6 and above
Mikio Hara [Fri, 15 Jan 2016 07:37:47 +0000 (16:37 +0900)]
runtime: readjust signal code for dragonfly-2.6 and above

Also adds missing nosplit to unminit.

Fixes #13964.

Change-Id: I07d93a8c872a255a89f91f808b66c889f0a6a69c
Reviewed-on: https://go-review.googlesource.com/18658
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agonet/http: update bundled http2
Brad Fitzpatrick [Fri, 15 Jan 2016 00:29:24 +0000 (16:29 -0800)]
net/http: update bundled http2

If a user starts two HTTP requests when no http2 connection is
available, both end up creating new TCP connections, since the
server's protocol (h1 or h2) isn't yet known. Once it turns out that
the server supports h2, one of the connections is useless. Previously
we kept upgrading both TLS connections to h2 (SETTINGS frame exchange,
etc).  Now the unnecessary connections are closed instead, before the
h2 preface/SETTINGS.

Updates x/net/http2 to git rev a8e212f3d for https://golang.org/cl/18675

This CL contains the tests for https://golang.org/cl/18675

Semi-related change noticed while writing the tests: now that we have
TLSNextProto in Go 1.6, which consults the TLS
ConnectionState.NegotiatedProtocol, we have to gurantee that the TLS
handshake has been done before we look at the ConnectionState. So add
that check after the DialTLS hook. (we never documented that users
have to call Handshake, so do it for them, now that it matters)

Updates #13957

Change-Id: I9a70e9d1282fe937ea654d9b1269c984c4e366c0
Reviewed-on: https://go-review.googlesource.com/18676
Reviewed-by: Andrew Gerrand <adg@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years ago[dev.ssa] runtime: fix windows build
Keith Randall [Thu, 14 Jan 2016 23:28:26 +0000 (15:28 -0800)]
[dev.ssa] runtime: fix windows build

Pull the relevant part of https://go-review.googlesource.com/#/c/18304/
into the SSA branch.  This fixes the windows SSA build.

Change-Id: Iee3834d8e7019cd90307e32029f3d1032ceb46c1
Reviewed-on: https://go-review.googlesource.com/18673
Reviewed-by: Austin Clements <austin@google.com>
9 years agoruntime: use at least "system" traceback level for runtime tests
Austin Clements [Thu, 14 Jan 2016 21:43:40 +0000 (16:43 -0500)]
runtime: use at least "system" traceback level for runtime tests

While the default behavior of eliding runtime frames from tracebacks
usually makes sense, this is not the case when you're trying to test
the runtime itself. Fix this by forcing the traceback level to at
least "system" in the runtime tests.

This will specifically help with debugging issue #13645, which has
proven remarkably resistant to reproduction outside of the build
dashboard itself.

Change-Id: I2a8356ba6c3c5badba8bb3330fc527357ec0d296
Reviewed-on: https://go-review.googlesource.com/18648
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Alex Brainman <alex.brainman@gmail.com>
9 years agoruntime: fix darwin/arm64 build
Matthew Dempsky [Thu, 14 Jan 2016 20:14:05 +0000 (12:14 -0800)]
runtime: fix darwin/arm64 build

Fixes #13916.

Change-Id: If2cad5473a749460909519ac20aca19dea8a2e7a
Reviewed-on: https://go-review.googlesource.com/18671
Reviewed-by: David Crawshaw <crawshaw@golang.org>
9 years agocmd/go: fix vendor directory check on Windows
Russ Cox [Thu, 14 Jan 2016 19:40:54 +0000 (14:40 -0500)]
cmd/go: fix vendor directory check on Windows

Fixes build.

Change-Id: Idd7d87f0bf3fe553b3ab00dd32eee4925a4fe3ff
Reviewed-on: https://go-review.googlesource.com/18647
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
9 years agoruntime: remove erroneous go:noescape declaration
Ian Lance Taylor [Thu, 14 Jan 2016 17:10:57 +0000 (09:10 -0800)]
runtime: remove erroneous go:noescape declaration

Change-Id: I6b1dc789e54a385c958961e7ba16bfd9d0f3b313
Reviewed-on: https://go-review.googlesource.com/18629
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agonet/http: fix flaky TestServerValidatesHostHeader
Brad Fitzpatrick [Thu, 14 Jan 2016 05:36:03 +0000 (05:36 +0000)]
net/http: fix flaky TestServerValidatesHostHeader

Passes with go test -race -count=1000 -name=TestServerValidatesHostHeader now
without hanging.

Fixes #13950

Change-Id: I41c3a555c642595c95c8c52f19a05a4c68e67630
Reviewed-on: https://go-review.googlesource.com/18660
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agoruntime: minimize time between lockextra/unlockextra
Ian Lance Taylor [Tue, 12 Jan 2016 23:34:03 +0000 (15:34 -0800)]
runtime: minimize time between lockextra/unlockextra

This doesn't fix a bug, but may improve performance in programs that
have many concurrent calls from C to Go.  The old code made several
system calls between lockextra and unlockextra.  That could be happening
while another thread is spinning acquiring lockextra.  This changes the
code to not make any system calls while holding the lock.

Change-Id: I50576478e478670c3d6429ad4e1b7d80f98a19d8
Reviewed-on: https://go-review.googlesource.com/18548
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/go: respect internal directories during 'go run'
Russ Cox [Thu, 14 Jan 2016 03:35:17 +0000 (22:35 -0500)]
cmd/go: respect internal directories during 'go run'

Fixes #12217.

Change-Id: I5ee6cb18eaa66bdec1affe689aa531c05e719fc9
Reviewed-on: https://go-review.googlesource.com/18645
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agoruntime: fix several issues in TestFutexsleep
Austin Clements [Wed, 13 Jan 2016 20:14:26 +0000 (15:14 -0500)]
runtime: fix several issues in TestFutexsleep

TestFutexsleep is supposed to clean up before returning by waking up
the goroutines it started and left blocked in futex sleeps. However,
it currently fails at this in several ways:

1. Both the sleep and wakeup are done on the address of tt.mtx, but in
   both cases tt is a *local copy* of the futexsleepTest created by a
   loop, so the sleep and wakeup happen on completely different
   addresses. Fix this by making them both use the address of the
   global tt.mtx.

2. If the sleep happens after the wakeup (not likely, but not
   impossible), it won't wake up. Fix this by using the futex protocol
   properly: sleep if the mutex's value is 0, and set the mutex's
   value to non-zero before doing the wakeup.

3. If TestFutexsleep runs more than once, channels and mutex values
   left over from the first run will interfere with later runs. Fix
   this by clearing the mutex value and creating a new channel for
   each test and waiting for goroutines to finish before returning
   (lest they send their completion to the channel for the next run).

As an added bonus, this test now actually tests that futex
sleep/wakeup work. Previously this test would have been satisfied if
futexsleep was an infinite loop and futexwakeup was a no-op.

Change-Id: I1cbc6871cc9dcb8f4601b3621913bec2b79b0fc3
Reviewed-on: https://go-review.googlesource.com/18617
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Mikio Hara <mikioh.mikioh@gmail.com>
9 years agocmd/go, go/build: ignore vendor directories with no Go source files
Russ Cox [Thu, 14 Jan 2016 02:34:13 +0000 (21:34 -0500)]
cmd/go, go/build: ignore vendor directories with no Go source files

Otherwise it is impossible to vendor a/b/c without hiding the real a/b.
I also updated golang.org/s/go15vendor.

Fixes #13832.

Change-Id: Iee3d53c11ea870721803f6e8e67845b405686e79
Reviewed-on: https://go-review.googlesource.com/18644
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agobuild: use consistent shebang line
Shenghou Ma [Thu, 14 Jan 2016 02:33:33 +0000 (21:33 -0500)]
build: use consistent shebang line

Fixes #13948.

Change-Id: I37d734c1241f3d72d8fe33dfdf79b573e5476b1f
Reviewed-on: https://go-review.googlesource.com/18643
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agodoc: revise help page with categories and more links
Andrew Gerrand [Thu, 14 Jan 2016 02:44:45 +0000 (13:44 +1100)]
doc: revise help page with categories and more links

Fixes #12489

Change-Id: I25dd3f76e4cfe9a71b987c3b31445724568391e9
Reviewed-on: https://go-review.googlesource.com/18625
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agodoc: document archive/tar changes
Russ Cox [Thu, 14 Jan 2016 02:13:02 +0000 (21:13 -0500)]
doc: document archive/tar changes

Fixes #13647.

Change-Id: I28df7ade9b5abd79ce6b9c3d14ceaa988e86fc01
Reviewed-on: https://go-review.googlesource.com/18642
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agodebug/dwarf: fix nil pointer dereference in cyclic type structures
Austin Clements [Fri, 8 Jan 2016 21:25:29 +0000 (16:25 -0500)]
debug/dwarf: fix nil pointer dereference in cyclic type structures

Currently readType simultaneously constructs a type graph and resolves
the sizes of the types. However, these two operations are
fundamentally at odds: the order we parse a cyclic structure in may be
different than the order we need to resolve type sizes in. As a
result, it's possible that when readType attempts to resolve the size
of a typedef, it may dereference a nil Type field of another typedef
retrieved from the type cache that's only partially constructed.

To fix this, we delay resolving typedef sizes until the end of the
readType recursion, when the full type graph is constructed.

Fixes #13039.

Change-Id: I9889af37fb3be5437995030fdd61e45871319d07
Reviewed-on: https://go-review.googlesource.com/18459
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: add AVARLIVE to peep for arm, arm64, mips64, ppc64
Russ Cox [Thu, 14 Jan 2016 02:02:12 +0000 (21:02 -0500)]
cmd/compile: add AVARLIVE to peep for arm, arm64, mips64, ppc64

Fixes build on those systems.

Also fix printing of AVARLIVE.

Change-Id: I1b38cca0125689bc08e4e1bdd0d0c140b1ea079a
Reviewed-on: https://go-review.googlesource.com/18641
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/internal/obj: separate code layout from object writing
Russ Cox [Tue, 5 Jan 2016 14:27:40 +0000 (09:27 -0500)]
cmd/internal/obj: separate code layout from object writing

This will allow the compiler to crunch Prog lists down to code as each
function is compiled, instead of waiting until the end, which should
reduce the working set of the compiler. But not until Go 1.7.

This also makes it easier to write some machine code output tests
for the assembler, which is why it's being done now.

For #13822.

Change-Id: I0811123bc6e5717cebb8948f9cea18e1b9baf6f7
Reviewed-on: https://go-review.googlesource.com/18311
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agocmd/compile: recognize Syscall-like functions for liveness analysis
Russ Cox [Wed, 13 Jan 2016 05:46:28 +0000 (00:46 -0500)]
cmd/compile: recognize Syscall-like functions for liveness analysis

Consider this code:

func f(*int)

func g() {
p := new(int)
f(p)
}

where f is an assembly function.
In general liveness analysis assumes that during the call to f, p is dead
in this frame. If f has retained p, p will be found alive in f's frame and keep
the new(int) from being garbage collected. This is all correct and works.
We use the Go func declaration for f to give the assembly function
liveness information (the arguments are assumed live for the entire call).

Now consider this code:

func h1() {
p := new(int)
syscall.Syscall(1, 2, 3, uintptr(unsafe.Pointer(p)))
}

Here syscall.Syscall is taking the place of f, but because its arguments
are uintptr, the liveness analysis and the garbage collector ignore them.
Since p is no longer live in h once the call starts, if the garbage collector
scans the stack while the system call is blocked, it will find no reference
to the new(int) and reclaim it. If the kernel is going to write to *p once
the call finishes, reclaiming the memory is a mistake.

We can't change the arguments or the liveness information for
syscall.Syscall itself, both for compatibility and because sometimes the
arguments really are integers, and the garbage collector will get quite upset
if it finds an integer where it expects a pointer. The problem is that
these arguments are fundamentally untyped.

The solution we have taken in the syscall package's wrappers in past
releases is to insert a call to a dummy function named "use", to make
it look like the argument is live during the call to syscall.Syscall:

func h2() {
p := new(int)
syscall.Syscall(1, 2, 3, uintptr(unsafe.Pointer(p)))
use(unsafe.Pointer(p))
}

Keeping p alive during the call means that if the garbage collector
scans the stack during the system call now, it will find the reference to p.

Unfortunately, this approach is not available to users outside syscall,
because 'use' is unexported, and people also have to realize they need
to use it and do so. There is much existing code using syscall.Syscall
without a 'use'-like function. That code will fail very occasionally in
mysterious ways (see #13372).

This CL fixes all that existing code by making the compiler do the right
thing automatically, without any code modifications. That is, it takes h1
above, which is incorrect code today, and makes it correct code.

Specifically, if the compiler sees a foreign func definition (one
without a body) that has uintptr arguments, it marks those arguments
as "unsafe uintptrs". If it later sees the function being called
with uintptr(unsafe.Pointer(x)) as an argument, it arranges to mark x
as having escaped, and it makes sure to hold x in a live temporary
variable until the call returns, so that the garbage collector cannot
reclaim whatever heap memory x points to.

For now I am leaving the explicit calls to use in package syscall,
but they can be removed early in a future cycle (likely Go 1.7).

The rule has no effect on escape analysis, only on liveness analysis.

Fixes #13372.

Change-Id: I2addb83f70d08db08c64d394f9d06ff0a063c500
Reviewed-on: https://go-review.googlesource.com/18584
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agogo/types: rename Importer2 to ImporterFrom go1.6beta2
Brad Fitzpatrick [Wed, 13 Jan 2016 23:20:20 +0000 (23:20 +0000)]
go/types: rename Importer2 to ImporterFrom

Per https://groups.google.com/forum/#!topic/golang-dev/javNmryAh0I

Change-Id: I08d7cbc94da4fc61c848f3dbee4637bf8fcfeb01
Reviewed-on: https://go-review.googlesource.com/18630
Reviewed-by: Alan Donovan <adonovan@google.com>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Chris Broadfoot <cbro@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>
9 years agoruntime: don't use CMOV for 386
Keith Randall [Wed, 13 Jan 2016 21:09:46 +0000 (13:09 -0800)]
runtime: don't use CMOV for 386

CMOVs were not introduced until P6.  We need 386 to run on
Pentium MMX.

Fixes #13923

Change-Id: Iee9572cd83e64c3a1336bc1e6b300b048fbcc996
Reviewed-on: https://go-review.googlesource.com/18621
Reviewed-by: Minux Ma <minux@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

9 years agonet/http: update bundled http2
Brad Fitzpatrick [Wed, 13 Jan 2016 22:14:40 +0000 (22:14 +0000)]
net/http: update bundled http2

Updates x/net/http2 to git rev 341cd08 for https://golang.org/cl/18576

Change-Id: If5dcb60ac449b798c34fe332ede5ec74e66eb9db
Reviewed-on: https://go-review.googlesource.com/18579
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

9 years agonet/http: clarify ConnState StateActive docs for HTTP/2
Brad Fitzpatrick [Wed, 13 Jan 2016 19:36:51 +0000 (19:36 +0000)]
net/http: clarify ConnState StateActive docs for HTTP/2

Update #13925

Change-Id: I7cd0625fad841eb0e3f364629f9bc225aa2fdce9
Reviewed-on: https://go-review.googlesource.com/18575
Reviewed-by: Andrew Gerrand <adg@golang.org>
9 years agogo/importer: fix field/method package for binary importer
Robert Griesemer [Wed, 13 Jan 2016 18:52:56 +0000 (10:52 -0800)]
go/importer: fix field/method package for binary importer

This is the equivalent of https://golang.org/cl/18549 for
the binary importer (which is usually not used because by
default the gc compiler produces the traditional textual
export format).

For #13898.

Change-Id: Idb6b515f2ee49e6d0362c71846994b0bd4dae8f7
Reviewed-on: https://go-review.googlesource.com/18598
Reviewed-by: Alan Donovan <adonovan@google.com>
Run-TryBot: Robert Griesemer <gri@golang.org>

9 years agocmd/go: fix gccSupportsNoPie for old GCC's that don't exit 0
Ian Lance Taylor [Wed, 13 Jan 2016 18:45:54 +0000 (10:45 -0800)]
cmd/go: fix gccSupportsNoPie for old GCC's that don't exit 0

GCC 4.8 exits 1 on an unrecognized option, but GCC 4.4 and 4.5 exit 0.
I didn't check other versions, or try to figure out just when this
changed.

Fixes #13937.

Change-Id: If193e9053fbb535999c9bde99f430f465a8c7c57
Reviewed-on: https://go-review.googlesource.com/18597
Run-TryBot: Ian Lance Taylor <iant@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agogo/importer: revert incorrect change that slipped in prior CL
Robert Griesemer [Wed, 13 Jan 2016 18:41:37 +0000 (10:41 -0800)]
go/importer: revert incorrect change that slipped in prior CL

The package of anonymous fields is the package in which they were
declared, not the package of the anonymous field's type. Was correct
before and incorrectly changed with https://golang.org/cl/18549.

Change-Id: I9fd5bfbe9d0498c8733b6ca7b134a85defe16113
Reviewed-on: https://go-review.googlesource.com/18596
Reviewed-by: Alan Donovan <adonovan@google.com>
9 years agocmd/link: add LC_VERSION_MIN_MACOSX to linkmode=internal OS X binaries
Russ Cox [Wed, 13 Jan 2016 14:59:16 +0000 (09:59 -0500)]
cmd/link: add LC_VERSION_MIN_MACOSX to linkmode=internal OS X binaries

This makes lldb willing to debug them.
The minimum version is hard-coded at OS X 10.7,
because that is the minimum that Go requires.
For more control over the version, users can
use linkmode=external and pass the relevant flags to the host linker.

Fixes #12941.

Change-Id: I20027be8aa034d07dd2a3326828f75170afe905f
Reviewed-on: https://go-review.googlesource.com/18588
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agoruntime: allow for C pointers between arena_start and arena_used in cgo check
Russ Cox [Wed, 13 Jan 2016 17:23:44 +0000 (12:23 -0500)]
runtime: allow for C pointers between arena_start and arena_used in cgo check

Fixes #13928.

Change-Id: Ia04c6bdef5ae6924d03982682ee195048f8f387f
Reviewed-on: https://go-review.googlesource.com/18611
Reviewed-by: Ian Lance Taylor <iant@golang.org>
9 years agoruntime: arrange to show a few local variables when cgoCheckUnknownPointer panics
Russ Cox [Wed, 13 Jan 2016 17:45:52 +0000 (12:45 -0500)]
runtime: arrange to show a few local variables when cgoCheckUnknownPointer panics

For #13934.

Change-Id: Id399e35598def96f8bb89b9fcd1bf14ee06e2e62
Reviewed-on: https://go-review.googlesource.com/18612
Reviewed-by: Ian Lance Taylor <iant@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years ago[dev.ssa] cmd/compile: clean up comparisons
Keith Randall [Tue, 5 Jan 2016 22:56:26 +0000 (14:56 -0800)]
[dev.ssa] cmd/compile: clean up comparisons

Add new constant-flags opcodes.  These can be generated from
comparisons that we know the result of, like x&31 < 32.

Constant-fold the constant-flags opcodes into all flag users.

Reorder some CMPxconst args so they read in the comparison direction.

Reorg deadcode removal a bit - it needs to remove the OpCopy ops it
generates when strength-reducing Phi ops.  So it needs to splice out all
the dead blocks and do a copy elimination before it computes live
values.

Change-Id: Ie922602033592ad8212efe4345394973d3b94d9f
Reviewed-on: https://go-review.googlesource.com/18267
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
9 years ago[dev.ssa] cmd/compile: fix spill sizes
Keith Randall [Mon, 4 Jan 2016 21:34:54 +0000 (13:34 -0800)]
[dev.ssa] cmd/compile: fix spill sizes

In code that does:

    var x, z int32
    var y int64
    z = phi(x, int32(y))

We silently drop the int32 cast because truncation is a no-op.
The phi operation needs to make sure it uses the size of the
phi, not the size of its arguments, when generating spills.

Change-Id: I1f7baf44f019256977a46fdd3dad1972be209042
Reviewed-on: https://go-review.googlesource.com/18390
Reviewed-by: David Chase <drchase@google.com>
9 years agogo/importer: associate exported field and interface methods with correct package
Robert Griesemer [Wed, 13 Jan 2016 01:02:32 +0000 (17:02 -0800)]
go/importer: associate exported field and interface methods with correct package

In gc export data, exported struct field and interface method names appear
in unqualified form (i.e., w/o package name). The (gc)importer assumed that
unqualified exported names automatically belong to the package being imported.
This is not the case if the field or method belongs to a struct or interface
that was declared in another package and re-exported.

The issue becomes visible if a type T (say an interface with a method M)
is declared in a package A, indirectly re-exported by a package B (which
imports A), and then imported in C. If C imports both A and B, if A is
imported before B, T.M gets associated with the correct package A. If B
is imported before A, T.M appears to be exported by B (even though T itself
is correctly marked as coming from A). If T.M is imported again via the
import of A if gets dropped (as it should) because it was imported already.

The fix is to pass down the parent package when we parse imported types
so that the importer can use the correct package when creating fields
and methods.

Fixes #13898.

Change-Id: I7ec2ee2dda15859c582b65db221c3841899776e1
Reviewed-on: https://go-review.googlesource.com/18549
Reviewed-by: Alan Donovan <adonovan@google.com>
9 years agocmd/go: -buildmode=c-shared should work on darwin/386
Joe Sylve [Mon, 11 Jan 2016 05:23:51 +0000 (23:23 -0600)]
cmd/go: -buildmode=c-shared should work on darwin/386

* Enable c-shared buildmode on darwin/386
* dyld does not support text relocation on i386. Add -read_only_relocs suppress flag to linker

Fixes #13904

Change-Id: I9adbd20d3f36ce9bbccf1bffb746b391780d088f
Reviewed-on: https://go-review.googlesource.com/18500
Reviewed-by: David Crawshaw <crawshaw@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet/http: update bundled http2
Brad Fitzpatrick [Wed, 13 Jan 2016 17:53:57 +0000 (17:53 +0000)]
net/http: update bundled http2

Updates x/net/http2 to git rev c93a9b4f2a for https://golang.org/cl/18474

Forgot to submit this four days ago.

Change-Id: Id96ab164ec765911c31874cca39b44aa55e80153
Reviewed-on: https://go-review.googlesource.com/18574
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>

9 years agodatabase/sql/driver: clarify DefaultParameterConverter docs
Brad Fitzpatrick [Mon, 11 Jan 2016 19:54:29 +0000 (11:54 -0800)]
database/sql/driver: clarify DefaultParameterConverter docs

Fixes #11489

Change-Id: I887ebac2dcb772e73ee393891c487f694028aaf2
Reviewed-on: https://go-review.googlesource.com/18520
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet/http: fix Transport crash when abandoning dial which upgrades protos
Brad Fitzpatrick [Wed, 13 Jan 2016 16:30:00 +0000 (16:30 +0000)]
net/http: fix Transport crash when abandoning dial which upgrades protos

When the Transport was creating an bound HTTP connection (protocol
unknown initially) and then ends up deciding it doesn't need it, a
goroutine sits around to clean up whatever the result was. That
goroutine made the false assumption that the result was always an
HTTP/1 connection or an error. It may also be an alternate protocol
in which case the *persistConn.conn net.Conn field is nil, and the
alt field is non-nil.

Fixes #13839

Change-Id: Ia4972e5eb1ad53fa00410b3466d4129c753e0871
Reviewed-on: https://go-review.googlesource.com/18573
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agodoc: document Go 1.5.3
Russ Cox [Wed, 13 Jan 2016 11:52:27 +0000 (06:52 -0500)]
doc: document Go 1.5.3

Change-Id: I9b4b76abfba66ff655aef55b43d9b4721aba604a
Reviewed-on: https://go-review.googlesource.com/18587
Reviewed-by: Chris Broadfoot <cbro@golang.org>
9 years agonet/http: reject non three digit status codes in ReadResponse
Emmanuel Odeke [Mon, 11 Jan 2016 06:20:06 +0000 (23:20 -0700)]
net/http: reject non three digit status codes in ReadResponse

Change-Id: If4a90c4017ef4b5c9f497cf117c8ad62b7e15c62
Reviewed-on: https://go-review.googlesource.com/18501
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: stop using fucomi* ops for 387 builds
Keith Randall [Tue, 12 Jan 2016 20:42:28 +0000 (12:42 -0800)]
cmd/compile: stop using fucomi* ops for 387 builds

The fucomi* opcodes were only introduced for the Pentium Pro.
They do not exist for an MMX Pentium.  Use the fucom* instructions
instead and move the condition codes from the fp flags register to
the integer flags register explicitly.

The use of fucomi* opcodes in ggen.go was introduced in 1.5 (CL 8738).
The bad ops were generated for 64-bit floating-point comparisons.

The use of fucomi* opcodes in gsubr.go dates back to at least 1.1.
The bad ops were generated for float{32,64} to uint64 conversions.

Fixes #13923

Change-Id: I5290599f5edea8abf8fb18036f44fa78bd1fc9e6
Reviewed-on: https://go-review.googlesource.com/18590
Reviewed-by: Minux Ma <minux@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/go: fixes for -compiler=gccgo -buildmode=c-shared
Ian Lance Taylor [Wed, 13 Jan 2016 05:24:49 +0000 (21:24 -0800)]
cmd/go: fixes for -compiler=gccgo -buildmode=c-shared

Install pkg.h rather than libpkg.h.

Link against -lc.

Fixes #13860.

Change-Id: I4e429426f8363712a5dbbd2655b9aab802ab2888
Reviewed-on: https://go-review.googlesource.com/18592
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Michael Hudson-Doyle <michael.hudson@canonical.com>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agocmd/internal/obj/x86: add new instructions, cleanup.
Ilya Tocar [Wed, 13 Jan 2016 13:43:22 +0000 (16:43 +0300)]
cmd/internal/obj/x86: add new instructions, cleanup.

Add several instructions that were used via BYTE and use them.
Instructions added: PEXTRB, PEXTRD, PEXTRQ, PINSRB, XGETBV, POPCNT.

Change-Id: I5a80cd390dc01f3555dbbe856a475f74b5e6df65
Reviewed-on: https://go-review.googlesource.com/18593
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet: LookupAddr("127.0.0.1") is "localhost" not "localhost." on Plan 9 and Windows
Mikio Hara [Wed, 13 Jan 2016 12:04:10 +0000 (21:04 +0900)]
net: LookupAddr("127.0.0.1") is "localhost" not "localhost." on Plan 9 and Windows

This change applies the fix for #13564 to Plan 9 and Windows.
Also enables Lookup API test cases on builders.

Updates #13564.

Change-Id: I863f03c7cb6fbe58b3a55223bfa0ac5f9bf9c3df
Reviewed-on: https://go-review.googlesource.com/18559
Run-TryBot: Mikio Hara <mikioh.mikioh@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
9 years agonet/http2: update bundled http2
Brad Fitzpatrick [Wed, 13 Jan 2016 04:51:08 +0000 (20:51 -0800)]
net/http2: update bundled http2

Update bundled http2 to git rev 76365a4 for https://golang.org/issue/18571

Fixes golang/go#13924

Change-Id: Ibb48cd6935b35d9965df70fb8761be5986d79ffc
Reviewed-on: https://go-review.googlesource.com/18591
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
9 years agocrypto/tls: don't block in Conn.Close if Writes are in-flight
Brad Fitzpatrick [Tue, 12 Jan 2016 21:15:51 +0000 (21:15 +0000)]
crypto/tls: don't block in Conn.Close if Writes are in-flight

Conn.Close sends an encrypted "close notify" to signal secure EOF.
But writing that involves acquiring mutexes (handshake mutex + the
c.out mutex) and writing to the network. But if the reason we're
calling Conn.Close is because the network is already being
problematic, then Close might block, waiting for one of those mutexes.

Instead of blocking, and instead of introducing new API (at least for
now), distinguish between a normal Close (one that sends a secure EOF)
and a resource-releasing destructor-style Close based on whether there
are existing Write calls in-flight.

Because io.Writer and io.Closer aren't defined with respect to
concurrent usage, a Close with active Writes is already undefined, and
should only be used during teardown after failures (e.g. deadlines or
cancelations by HTTP users). A normal user will do a Write then
serially do a Close, and things are unchanged for that case.

This should fix the leaked goroutines and hung net/http.Transport
requests when there are network errors while making TLS requests.

Change-Id: If3f8c69d6fdcebf8c70227f41ad042ccc3f20ac9
Reviewed-on: https://go-review.googlesource.com/18572
Reviewed-by: Adam Langley <agl@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

9 years agocmd/compile: better modeling of escape across loop levels
David Chase [Mon, 4 Jan 2016 21:44:20 +0000 (16:44 -0500)]
cmd/compile: better modeling of escape across loop levels

Brief background on "why heap allocate".  Things can be
forced to the heap for the following reasons:

1) address published, hence lifetime unknown.
2) size unknown/too large, cannot be stack allocated
3) multiplicity unknown/too large, cannot be stack allocated
4) reachable from heap (not necessarily published)

The bug here is a case of failing to enforce 4) when an
object Y was reachable from a heap allocation X forced
because of 3).  It was found in the case of a closure
allocated within a loop (X) and assigned to a variable
outside the loop (multiplicity unknown) where the closure
also captured a map (Y) declared outside the loop (reachable
from heap). Note the variable declared outside the loop (Y)
is not published, has known size, and known multiplicity
(one). The only reason for heap allocation is that it was
reached from a heap allocated item (X), but because that was
not forced by publication, it has to be tracked by loop
level, but escape-loop level was not tracked and thus a bug
results.

The fix is that when a heap allocation is newly discovered,
use its looplevel as the minimum loop level for downstream
escape flooding.

Every attempt to generalize this bug to X-in-loop-
references-Y-outside loop succeeded, so the fix was aimed
to be general.  Anywhere that loop level forces heap
allocation, the loop level is tracked.  This is not yet
tested for all possible X and Y, but it is correctness-
conservative and because it caused only one trivial
regression in the escape tests, it is probably also
performance-conservative.

The new test checks the following:
1) in the map case, that if fn escapes, so does the map.
2) in the map case, if fn does not escape, neither does the map.
3) in the &x case, that if fn escapes, so does &x.
4) in the &x case, if fn does not escape, neither does &x.

Fixes #13799.

Change-Id: Ie280bef2bb86ec869c7c206789d0b68f080c3fdb
Reviewed-on: https://go-review.googlesource.com/18234
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Russ Cox <rsc@golang.org>