On top of "tiny allocator" (cl/38750047), reduces number of allocs by 1% on json.
No code must rely on zero termination. So will also make debugging simpler,
by uncovering issues earlier.
json-1
allocated 7949686 7915766 -0.43%
allocs 93778 92790 -1.05%
time 100957795 97250949 -3.67%
rest of the metrics are too noisy.
LGTM=r
R=golang-codereviews, r, bradfitz, iant
CC=golang-codereviews
https://golang.org/cl/40370061
There is more zeroing than I would like right now -
temporaries used for the new map and channel runtime
calls need to be eliminated - but it will do for now.
This CL only has an effect if you are building with
GOEXPERIMENT=precisestack ./all.bash
(or make.bash). It costs about 5% in the overall time
spent in all.bash. That number will come down before
we make it on by default, but this should be enough for
Keith to try using the precise maps for copying stacks.
amd64 only (and it's not really great generated code).
TBR=khr, iant
CC=golang-codereviews
https://golang.org/cl/56430043
The addition of TLS to ARM rewrote the MRC instruction
differently depending on whether we were using internal
or external linking mode. That's clearly not okay, since we
don't know that during compilation, which is when we now
generate the code. Also, because the change did not introduce
a real MRC instruction but instead just macro-expanded it
in the assembler, liblink is rewriting a WORD instruction that
may actually be looking for that specific constant, which would
lead to very unexpected results. It was also using one value
that happened to be 8 where a different value that also
happened to be 8 belonged. So the code was correct for those
values but not correct in general, and very confusing.
Throw it all away.
Replace with the following. There is a linker-provided symbol
runtime.tlsgm with a value (address) set to the offset from the
hardware-provided TLS base register to the g and m storage.
Any reference to that name emits an appropriate TLS relocation
to be resolved by either the internal linker or the external linker,
depending on the link mode. The relocation has exactly the
semantics of the R_ARM_TLS_LE32 relocation, which is what
the external linker provides.
This symbol is only used in two routines, runtime.load_gm and
runtime.save_gm. In both cases it is now used like this:
MRC 15, 0, R0, C13, C0, 3 // fetch TLS base pointer
MOVW $runtime·tlsgm(SB), R2
ADD R2, R0 // now R0 points at thread-local g+m storage
It is likely that this change breaks the generation of shared libraries
on ARM, because the MOVW needs to be rewritten to use the global
offset table and a different relocation type. But let's get the supported
functionality working again before we worry about unsupported
functionality.
LGTM=dave, iant
R=iant, dave
CC=golang-codereviews
https://golang.org/cl/56120043
The typo was introduced by one of Dmitriy's CLs this morning.
The fix makes the ARM build compile again; it still won't pass
its tests, but one thing at a time.
TBR=dvyukov
CC=golang-codereviews
https://golang.org/cl/55770044
Introduces two-phase goroutine parking mechanism -- prepare to park, commit park.
This mechanism does not require backing mutex to protect wait predicate.
Use it in netpoll. See comment in netpoll.goc for details.
This slightly reduces contention between reader, writer and read/write io notifications;
and just eliminates a bunch of mutex operations from hotpaths, thus making then faster.
benchmark old ns/op new ns/op delta
BenchmarkTCP4ConcurrentReadWrite 2109 1945 -7.78%
BenchmarkTCP4ConcurrentReadWrite-2 1162 1113 -4.22%
BenchmarkTCP4ConcurrentReadWrite-4 798 755 -5.39%
BenchmarkTCP4ConcurrentReadWrite-8 803 748 -6.85%
BenchmarkTCP4Persistent 9411 9240 -1.82%
BenchmarkTCP4Persistent-2 5888 5813 -1.27%
BenchmarkTCP4Persistent-4 4016 3968 -1.20%
BenchmarkTCP4Persistent-8 3943 3857 -2.18%
R=golang-codereviews, mikioh.mikioh, gobot, iant, rsc
CC=golang-codereviews, khr
https://golang.org/cl/45700043
- do not lose profiling signals when we have no mcache (possible for syscalls/cgo)
- do not lose any profiling signals on windows
- fix profiling of cgo programs on windows (they had no m->thread setup)
- properly setup tls in cgo programs on windows
- check _beginthread return value
Fixes#6417.
Fixes#6986.
R=alex.brainman, rsc
CC=golang-codereviews
https://golang.org/cl/44820047
Now that liblink is compiled into the compilers and assemblers,
it must not refer to the "linkmode", since that is not known until
link time. This CL makes the ARM support no longer use linkmode,
which fixes a bug with cgo binaries that contain their own TLS
variables.
The x86 code must also remove linkmode; that is issue 7164.
Fixes#6992.
R=golang-codereviews, iant
CC=golang-codereviews
https://golang.org/cl/55160043
Currently we collect (add) all roots into a global array in a single-threaded GC phase.
This hinders parallelism.
With this change we just kick off parallel for for number_of_goroutines+5 iterations.
Then parallel for callback decides whether it needs to scan stack of a goroutine
scan data segment, scan finalizers, etc. This eliminates the single-threaded phase entirely.
This requires to store all goroutines in an array instead of a linked list
(to allow direct indexing).
This CL also removes DebugScan functionality. It is broken because it uses
unbounded stack, so it can not run on g0. When it was working, I've found
it helpless for debugging issues because the two algorithms are too different now.
This change would require updating the DebugScan, so it's simpler to just delete it.
With 8 threads this change reduces GC pause by ~6%, while keeping cputime roughly the same.
garbage-8
allocated 2987886 2989221 +0.04%
allocs 62885 62887 +0.00%
cputime 21286000 21272000 -0.07%
gc-pause-one 26633247 24885421 -6.56%
gc-pause-total 873570 811264 -7.13%
rss 242089984 242515968 +0.18%
sys-gc 13934336 13869056 -0.47%
sys-heap 205062144 205062144 +0.00%
sys-other 12628288 12628288 +0.00%
sys-stack 11534336 11927552 +3.41%
sys-total 243159104 243487040 +0.13%
time 2809477 2740795 -2.44%
R=golang-codereviews, rsc
CC=cshapiro, golang-codereviews, khr
https://golang.org/cl/46860043
Instead of a per-goroutine stack of defers for all sizes,
introduce per-P defer pool for argument sizes 8, 24, 40, 56, 72 bytes.
For a program that starts 1e6 goroutines and then joins then:
old: rss=6.6g virtmem=10.2g time=4.85s
new: rss=4.5g virtmem= 8.2g time=3.48s
R=golang-codereviews, rsc
CC=golang-codereviews
https://golang.org/cl/42750044
Currently for 2-word blocks we set the flag to clear the flag. Makes no sense.
In particular on 32-bits we call memclr always.
R=golang-codereviews, dave, iant
CC=golang-codereviews, khr, rsc
https://golang.org/cl/41170044
The test prints an excessive \n when /dev/null is not present.
R=golang-codereviews, bradfitz, dave
CC=golang-codereviews
https://golang.org/cl/54890043
What was happenning is as follows:
Each writer goroutine always triggers GC during its scheduling quntum.
After GC goroutines are shuffled so that the timer goroutine is always second in the queue.
This repeats infinitely, causing timer goroutine starvation.
Fixes#7126.
R=golang-codereviews, shanemhansen, khr, khr
CC=golang-codereviews
https://golang.org/cl/53080043
Vararg C calls present a problem for the GC because the
argument types are not derivable from the signature. Remove
them by passing pointers to channel elements instead of the
channel elements directly.
R=golang-codereviews, gobot, rsc, dvyukov
CC=golang-codereviews
https://golang.org/cl/53430043
Reflect used to communicate to the runtime using interface words,
which is bad for precise GC because sometimes iwords hold a pointer
and sometimes they don't. This change rewrites channel and select
operations to always pass pointers to the runtime.
reflect.Select gets somewhat more expensive, as we now do an allocation
per receive case instead of one allocation whose size is the max of
all the received types. This seems unavoidable to get preciseness
(unless we move the allocation into selectgo, which is a much bigger
change).
Fixes#6490
R=golang-codereviews, dvyukov, rsc
CC=golang-codereviews
https://golang.org/cl/52900043
Example of output:
goroutine 4 [sleep for 3 min]:
time.Sleep(0x34630b8a000)
src/pkg/runtime/time.goc:31 +0x31
main.func·002()
block.go:16 +0x2c
created by main.main
block.go:17 +0x33
Full program and output are here:
http://play.golang.org/p/NEZdADI3TdFixes#6809.
R=golang-codereviews, khr, kamil.kisiel, bradfitz, rsc
CC=golang-codereviews
https://golang.org/cl/50420043
Use lock-free fixed-size ring for work queues
instead of an unbounded mutex-protected array.
The ring has single producer and multiple consumers.
If the ring overflows, work is put onto global queue.
benchmark old ns/op new ns/op delta
BenchmarkMatmult 7 5 -18.12%
BenchmarkMatmult-4 2 2 -18.98%
BenchmarkMatmult-16 1 0 -12.84%
BenchmarkCreateGoroutines 105 88 -16.10%
BenchmarkCreateGoroutines-4 376 219 -41.76%
BenchmarkCreateGoroutines-16 241 174 -27.80%
BenchmarkCreateGoroutinesParallel 103 87 -14.66%
BenchmarkCreateGoroutinesParallel-4 169 143 -15.38%
BenchmarkCreateGoroutinesParallel-16 158 151 -4.43%
R=golang-codereviews, rsc
CC=ddetlefs, devon.odell, golang-codereviews
https://golang.org/cl/46170044
Map iteration previously started from a random bucket, but walked each
bucket from the beginning. Now, iteration always starts from the first
bucket and walks each bucket starting at a random offset. For
performance, the random offset is selected at the start of iteration
and reused for each bucket.
Iteration over a map with 8 or fewer elements--a single bucket--will
now be non-deterministic. There will now be only 8 different possible
map iterations.
Significant benchmark changes, on my OS X laptop (rough but consistent):
benchmark old ns/op new ns/op delta
BenchmarkMapIter 128 121 -5.47%
BenchmarkMapIterEmpty 4.26 4.45 +4.46%
BenchmarkNewEmptyMap 114 111 -2.63%
Fixes#6719.
R=khr, bradfitz
CC=golang-codereviews
https://golang.org/cl/47370043
Profiling of multithreaded applications works correctly on OpenBSD
5.4-current, so enable the profiling test.
R=golang-codereviews, minux.ma
CC=golang-codereviews
https://golang.org/cl/50940043
Update Go so that it continues to work past the OpenBSD system ABI
break, with 64-bit time_t:
http://www.openbsd.org/faq/current.html#20130813
Note: this makes OpenBSD 5.5 (currently 5.4-current) the minimum
supported release for Go.
Fixes#7049.
R=golang-codereviews, mikioh.mikioh
CC=golang-codereviews
https://golang.org/cl/13368046
The spans array is allocated in runtime·mallocinit. On a
32-bit system the number of entries in the spans array is
MaxArena32 / PageSize, which (2U << 30) / (1 << 12) == (1 << 19).
So we are allocating an array that can hold 19 bits for an
index that can hold 20 bits. According to the comment in the
function, this is intentional: we only allocate enough spans
(and bitmaps) for a 2G arena, because allocating more would
probably be wasteful.
But since the span index is simply the upper 20 bits of the
memory address, this scheme only works if memory addresses are
limited to the low 2G of memory. That would be OK if we were
careful to enforce it, but we're not. What we are careful to
enforce, in functions like runtime·MHeap_SysAlloc, is that we
always return addresses between the heap's arena_start and
arena_start + MaxArena32.
We generally get away with it because we start allocating just
after the program end, so we only run into trouble with
programs that allocate a lot of memory, enough to get past
address 0x80000000.
This changes the code that computes a span index to subtract
arena_start on 32-bit systems just as we currently do on
64-bit systems.
R=golang-codereviews, rsc
CC=golang-codereviews
https://golang.org/cl/49460043
NPTL uses SIGRTMIN (signal 32) to effect thread cancellation.
Go's runtime replaces NPTL's signal handler with its own, and
ends up aborting if a C library that ends up calling
pthread_cancel is used.
This patch prevents runtime from replacing NPTL's handler.
Fixes#6997.
R=golang-codereviews, iant, dvyukov
CC=golang-codereviews
https://golang.org/cl/47540043
This prevents callers from using reflect to create a new
instance of errorCString with an arbitrary value and calling
the Error method to examine arbitrary memory.
Fixes#7084.
R=golang-codereviews, minux.ma, bradfitz
CC=golang-codereviews
https://golang.org/cl/49600043
This lets stack splits work correctly when running under gdb
when gdb has inserted a breakpoint somewhere on the call
stack.
Fixes#6834.
R=golang-codereviews, minux.ma
CC=golang-codereviews
https://golang.org/cl/48650043
record finalizers and heap profile info. Enables
removing the special bit from the heap bitmap. Also
provides a generic mechanism for annotating occasional
heap objects.
finalizers
overhead per obj
old 680 B 80 B avg
new 16 B/span 48 B
profile
overhead per obj
old 32KB 24 B + hash tables
new 16 B/span 24 B
R=cshapiro, khr, dvyukov, gobot
CC=golang-codereviews
https://golang.org/cl/13314053
Some builders broke on this test; I'm guessing that was because
this test didn't try hard enough to find a different iteration order.
Update #6719
R=dave
CC=golang-codereviews
https://golang.org/cl/47300043
Technically the spec does not guarantee that the iteration order is random,
but it is a property that we have consciously pursued, and so it seems
right to verify that our implementation does indeed randomise.
Update #6719.
R=khr, bradfitz
CC=golang-codereviews
https://golang.org/cl/47010043
Fixes#6952.
runtime.asminit was incorrectly loading runtime.goarm as a word, not a uint8 which made it subject to alignment issues on arm5 platforms.
Alignment aside, this also meant that the top 3 bytes in R11 would have been garbage and could not be assumed to be setting up the FPU reliably.
R=iant, minux.ma
CC=golang-codereviews
https://golang.org/cl/46240043
This was done correctly for most targets but was missing from
FreeBSD/ARM and Linux/ARM.
R=golang-codereviews, dave
CC=golang-codereviews
https://golang.org/cl/45180043
sigprocmask use in a multithreaded environment is undefined so replace it with pthread_sigmask.
Fixes#6811.
R=jsing, iant
CC=golang-codereviews, golang-dev
https://golang.org/cl/30460043
Needed for precise gc and copying stacks.
reflect.Value now takes 4 words instead of 3.
Still to do:
- un-iword-ify channel ops.
- un-iword-ify method receivers.
R=golang-dev, iant, rsc, khr
CC=golang-dev
https://golang.org/cl/43040043
The runtime tests are executed 4 times in all.bash
and there is currently a 5-second delay each time.
R=golang-dev, minux.ma, khr, bradfitz
CC=golang-dev
https://golang.org/cl/42450043
On the plus side, we don't need to change the bits when mallocing
pointerless objects. On the other hand, we need to mark objects in the
free lists during GC. But the free lists are small at GC time, so it
should be a net win.
benchmark old ns/op new ns/op delta
BenchmarkMalloc8 40 33 -17.65%
BenchmarkMalloc16 45 38 -15.72%
BenchmarkMallocTypeInfo8 58 59 +0.85%
BenchmarkMallocTypeInfo16 63 64 +1.10%
R=golang-dev, rsc, dvyukov
CC=cshapiro, golang-dev
https://golang.org/cl/41040043
Adds the Pool type and docs, and use it in fmt.
This is a temporary implementation, until Dmitry
makes it fast.
Uses the API proposal from Russ in http://goo.gl/cCKeb2 but
adds an optional New field, as used in fmt and elsewhere.
Almost all callers want that.
Update #4720
R=golang-dev, rsc, cshapiro, iant, r, dvyukov, khr
CC=golang-dev
https://golang.org/cl/41860043
Hash tables currently store an evacuated bit in the low bit
of the overflow pointer. That's probably not sustainable in the
long term as GC wants correctly typed & aligned pointers. It is
also a pain to move any of this code to Go in the current state.
This change moves the evacuated bit into the tophash entries.
Performance change is negligable.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/14412043
And document it explicitly, even though it already said
it wasn't guaranteed.
Fixes#6857
R=golang-dev, khr
CC=golang-dev
https://golang.org/cl/43580043
Fixes a regression introduced in revision 4cb93e2900d0.
That revision changed runtime·memmove to use SSE MOVOU
instructions for sizes between 17 and 256 bytes. We were
using memmove to save a copy of the note string during
the note handler. The Plan 9 kernel does not allow the
use of floating point in note handlers (which includes
MOVOU since it touches the XMM registers).
Arguably, runtime·memmove should not be using MOVOU when
GO386=387 but that wouldn't help us on amd64. It's very
important that we guard against any future changes so we
use a simple copy loop instead.
This change is extracted from CL 9796043 (since that CL
is still being ironed out).
R=rsc
CC=golang-dev
https://golang.org/cl/34640045
The funcdata symbol incorrectly named the dead value map the
dead pointer map. The dead value map identifies all dead
values, including pointers and non-pointers, in a stack frame.
The purpose of this map is to allow the runtime to poison
locations of dead data to catch lost invariants.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/38670043
The new linker will disallow this on arm
(it is already disallowed on amd64 and 386)
in order to be able to lay out each function
separately.
The restriction is only for jumps into the middle
of a function; jumps to the beginning of a function
remain fine.
Prereq for linker cleanup (golang.org/s/go13linker).
R=iant, r, minux.ma
CC=golang-dev
https://golang.org/cl/35800043
When enabled this new debugging mode will allocate objects on
their own page and never recycle memory addresses. This is an
essential tool to root cause a broad class of heap corruption.
R=golang-dev, dave, daniel.morsing, dvyukov, rsc, iant, cshapiro
CC=golang-dev
https://golang.org/cl/22060046
This change allows the garbage collector to examine stack
slots that are determined as live and containing a pointer
value by the garbage collector. This results in a mean
reduction of 65% in the number of stack slots scanned during
an invocation of "GOGC=1 all.bash".
Unfortunately, this does not yet allow garbage collection to
be precise for the stack slots computed as live. Pointers
confound the determination of what definitions reach a given
instruction. In general, this problem is not solvable without
runtime cost but some advanced cooperation from the compiler
might mitigate common cases.
R=golang-dev, rsc, cshapiro
CC=golang-dev
https://golang.org/cl/14430048
Pass as a slice of strings instead. For 2-5 strings, implement
dedicated routines so no slices are needed.
static call counts in the go binary:
2 strings: 342 occurrences
3 strings: 98
4 strings: 30
5 strings: 13
6+ strings: 14
Why? C varags, bad for stack scanning and copying.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/36380043
Now that the map implementation is reading the keys and values from
arbitrary memory (instead of from stack slots), it needs to tell the
race detector when it does so.
Fixes#6875.
R=golang-dev, dave
CC=golang-dev
https://golang.org/cl/36360043
This change is part of the plan to get rid of all vararg C calls
which are a pain for getting exact stack scanning.
We allocate a chunk of zero memory to return a pointer to when a
map access doesn't find the key. This is simpler than returning nil
and fixing things up in the caller. Linker magic allocates a single
zero memory area that is shared by all (non-reflect-generated) map
types.
Passing things by reference gets rid of some copies, so it speeds
up code with big keys/values.
benchmark old ns/op new ns/op delta
BenchmarkBigKeyMap 34 31 -8.48%
BenchmarkBigValMap 37 30 -18.62%
BenchmarkSmallKeyMap 26 23 -11.28%
R=golang-dev, dvyukov, khr, rsc
CC=golang-dev
https://golang.org/cl/14794043
Two bugs:
1. The first iteration of the traceback always uses LR when provided,
which it is (only) during a profiling signal, but in fact LR is correct
only if the stack frame has not been allocated yet. Otherwise an
intervening call may have changed LR, and the saved copy in the stack
frame should be used. Fix in traceback_arm.c.
2. The division runtime call adds 8 bytes to the stack. In order to
keep the traceback routines happy, it must copy the saved LR into
the new 0(SP). Change
SUB $8, SP
into
MOVW 0(SP), R11 // r11 is temporary, for use by linker
MOVW.W R11, -8(SP)
to update SP and 0(SP) atomically, so that the traceback always
sees a saved LR at 0(SP).
Fixes#6681.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/19910044
The CL causes misc/cgo/test to fail randomly.
I suspect that the problem is the use of a division instruction
in usleep, which can be called while trying to acquire an m
and therefore cannot store the denominator in m.
The solution to that would be to rewrite the code to use a
magic multiply instead of a divide, but now we're getting
pretty far off the original code.
Go back to the original in preparation for a different,
less efficient but simpler fix.
««« original CL description
cmd/5l, runtime: make ARM integer division profiler-friendly
The implementation of division constructed non-standard
stack frames that could not be handled by the traceback
routines.
CL 13239052 left the frames non-standard but fixed them
for the specific case of a divide-by-zero panic.
A profiling signal can arrive at any time, so that fix
is not sufficient.
Change the division to store the extra argument in the M struct
instead of in a new stack slot. That keeps the frames bog standard
at all times.
Also fix a related bug in the traceback code: when starting
a traceback, the LR register should be ignored if the current
function has already allocated its stack frame and saved the
original LR on the stack. The stack copy should be used, as the
LR register may have been modified.
Combined, these make the torture test from issue 6681 pass.
Fixes#6681.
R=golang-dev, r, josharian
CC=golang-dev
https://golang.org/cl/19810043
»»»
TBR=r
CC=golang-dev
https://golang.org/cl/20350043
The implementation of division constructed non-standard
stack frames that could not be handled by the traceback
routines.
CL 13239052 left the frames non-standard but fixed them
for the specific case of a divide-by-zero panic.
A profiling signal can arrive at any time, so that fix
is not sufficient.
Change the division to store the extra argument in the M struct
instead of in a new stack slot. That keeps the frames bog standard
at all times.
Also fix a related bug in the traceback code: when starting
a traceback, the LR register should be ignored if the current
function has already allocated its stack frame and saved the
original LR on the stack. The stack copy should be used, as the
LR register may have been modified.
Combined, these make the torture test from issue 6681 pass.
Fixes#6681.
R=golang-dev, r, josharian
CC=golang-dev
https://golang.org/cl/19810043
The case can happen when starttheworld is calling acquirep
to get things moving again and acquirep gets preempted.
The stack trace is in golang.org/issue/6644.
It is difficult to build a short test case for this, but
the person who reported issue 6644 confirms that this
solves the problem.
Fixes#6644.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/18740044
Nomemprof seems to be unneeded now, there is no recursion.
If the recursion will be re-introduced, it will break loudly by deadlocking.
Fixes#6566.
R=golang-dev, minux.ma, rsc
CC=golang-dev
https://golang.org/cl/14695044
Instead of adding an -march=armv5t flag to the gcc command
line, the same effect is obtained with an ".arch armv5t"
pseudo op in the assembly file that uses armv5t instructions.
R=golang-dev, iant, dave
CC=golang-dev
https://golang.org/cl/14511044
If an iterator is started while a map is in the middle of a grow,
and the map has NaN keys, then those keys might get returned by
the iterator more than once. This fix makes the evacuation decision
deterministic and repeatable for NaN keys so each one gets returned
only once.
R=golang-dev, r, khr, iant
CC=golang-dev
https://golang.org/cl/14367043