UMTX_OP_WAIT expects that the address points to a uintptr, but
the code in lock_futex.c uses a uint32. UMTX_OP_WAIT_UINT is
just like UMTX_OP_WAIT, but the address points to a uint32.
This almost certainly makes no difference on a little-endian
system, but since the kernel supports it we should do the
right thing. And, who knows, maybe it matters.
R=golang-dev, bradfitz, r, ality
CC=golang-dev
https://golang.org/cl/8699043
The race detector uses a global lock to analyze atomic
operations. A panic in the middle of the code leaves the
lock acquired.
Similarly, the sync package may leave the race detectro
inconsistent when methods are called on nil pointers.
R=golang-dev, r, minux.ma, dvyukov, rsc, adg
CC=golang-dev
https://golang.org/cl/7981043
It's not trivial to make a comprehensive check
due to inferior pointers, reflect, gob, etc.
But this is essentially what I've used to debug
the GC issues.
Update #5193.
R=golang-dev, iant, 0xe2.0x9a.0x9b, r
CC=golang-dev
https://golang.org/cl/8455043
Use atomic operations on flags field to make sure we aren't
losing a flag update during parallel map operations.
R=golang-dev, dave, r
CC=golang-dev
https://golang.org/cl/8377046
The invariant is that there must be at least one running P or a thread polling network.
It was broken.
Fixes#5216.
R=golang-dev, bradfitz, r
CC=golang-dev
https://golang.org/cl/8459043
This makes it an unsafe.Pointer in Go so the garbage collector
will treat it as a pointer to untyped data, not a pointer to
bytes.
R=golang-dev, dvyukov
CC=golang-dev
https://golang.org/cl/8286045
If for whatever reason seh points into Go heap region,
the dangling pointer will cause memory corruption during GC.
Update #5193.
R=golang-dev, alex.brainman, iant
CC=golang-dev
https://golang.org/cl/8402045
Fixes#5175.
Race detector runtime expects values passed to MapShadow() to be page-aligned,
because they are used in mmap() call. If they are not aligned mmap() trims
either beginning or end of the mapping.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/8325043
This changes the map lookup behavior for string maps with 2-8 keys.
There was already previously a fastpath for 0 items and 1 item.
Now, if a string-keyed map has <= 8 items, first check all the
keys for length first. If only one has the right length, then
just check it for equality and avoid hashing altogether. Once
the map has more than 8 items, always hash like normal.
I don't know why some of the other non-string map benchmarks
got faster. This was with benchtime=2s, multiple times. I haven't
anything else getting slower, though.
benchmark old ns/op new ns/op delta
BenchmarkHashStringSpeed 37 34 -8.20%
BenchmarkHashInt32Speed 32 29 -10.67%
BenchmarkHashInt64Speed 31 27 -12.82%
BenchmarkHashStringArraySpeed 105 99 -5.43%
BenchmarkMegMap 274206 255153 -6.95%
BenchmarkMegOneMap 27 23 -14.80%
BenchmarkMegEqMap 148332 116089 -21.74%
BenchmarkMegEmptyMap 4 3 -12.72%
BenchmarkSmallStrMap 22 22 -0.89%
BenchmarkMapStringKeysEight_32 42 23 -43.71%
BenchmarkMapStringKeysEight_64 55 23 -56.96%
BenchmarkMapStringKeysEight_1M 279688 24 -99.99%
BenchmarkIntMap 16 15 -10.18%
BenchmarkRepeatedLookupStrMapKey32 40 37 -8.15%
BenchmarkRepeatedLookupStrMapKey1M 287918 272980 -5.19%
BenchmarkNewEmptyMap 156 130 -16.67%
R=golang-dev, khr
CC=golang-dev
https://golang.org/cl/7641057
The expected precision setting for the x87 on Win32 is 53-bit
but MinGW resets the floating point unit to 64-bit. Win32
object code generally expects values to be rounded to double,
not double extended, precision.
R=golang-dev, khr
CC=golang-dev
https://golang.org/cl/8175044
Doing grow work on reads is not multithreaded safe.
Changed code to do grow work only on inserts & deletes.
This is a short-term fix, eventually we'll want to do
grow work in parallel to recover the space of the old
table.
Fixes#5120.
R=bradfitz, khr
CC=golang-dev
https://golang.org/cl/8242043
Motivated by garbage profiling in HTTP benchmarks. This
changes means new empty maps are just one small allocation
(the HMap) instead the HMap + the relatively larger h->buckets
allocation. This helps maps which remain empty throughout
their life.
benchmark old ns/op new ns/op delta
BenchmarkNewEmptyMap 196 107 -45.41%
benchmark old allocs new allocs delta
BenchmarkNewEmptyMap 2 1 -50.00%
benchmark old bytes new bytes delta
BenchmarkNewEmptyMap 195 50 -74.36%
R=khr, golang-dev, r
CC=golang-dev
https://golang.org/cl/7722046
A HMUL node appears in some constant divisions, but
to observe a false negative in race detector the divisor must be
suitably chosen to make sure the only memory access is
done for HMUL.
R=dvyukov
CC=golang-dev
https://golang.org/cl/7935045
For Go 1.1, stop checking the rlimit, because it broke now
that mheap is allocated using SysAlloc. See issue 5049.
R=r
CC=golang-dev
https://golang.org/cl/7741050
The arm gentraceback mishandled frame linkage values pointing
to the assembly return function. This function is special as
its frame size is zero and it contains only one instruction.
These conditions would preserve the frame pointer and result
in an off by one error when unwinding the caller.
Fixes#5124
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/8023043
Prevents storm of error messages if something goes wrong.
In the case of issue 5073 the epoll fd was closed by the test.
Update #5073.
R=golang-dev, r, rsc
CC=golang-dev
https://golang.org/cl/7966043
Handle interface comparison correctly,
add a few more tests, mark more nodes as impossible.
R=dvyukov, golang-dev
CC=golang-dev
https://golang.org/cl/7942045
This keeps the logic about how to set the thread-local variables
m and g in code compiled and linked by the gc toolchain,
an important property for upcoming cgo changes.
It's also just a nice cleanup: one less place to update when
these details change.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7560048
The right operand of a && and || is only executed conditionnally,
so the instrumentation must be more careful. In particular
it should not turn nodes assumed to be cheap after walk into
expensive ones.
Update #4228
R=dvyukov, golang-dev
CC=golang-dev
https://golang.org/cl/7986043
The ARM implementation of runtime.cgocallback_gofunc diverged
from the calling convention by leaving a word of garbage at
the top of the stack and storing the return PC above the
locals. This change stores the return PC at the top of the
stack and removes the save area above the locals.
Update #5124
This CL fixes first part of the ARM issues and added the unwind test.
R=golang-dev, bradfitz, minux.ma, cshapiro, rsc
CC=golang-dev
https://golang.org/cl/7728045
Adds the new debugging constant 'checkgc'. If its value is non-zero
all calls to mallocgc() from hashmap.c will start a garbage collection.
Fixes#5074.
R=golang-dev, khr
CC=golang-dev, rsc
https://golang.org/cl/7663051
Fixes performance of the current windows network poller
with the new scheduler.
Gives runtime a hint when GetQueuedCompletionStatus() will block.
Fixes#5068.
benchmark old ns/op new ns/op delta
BenchmarkTCP4Persistent 4004000 33906 -99.15%
BenchmarkTCP4Persistent-2 21790 17513 -19.63%
BenchmarkTCP4Persistent-4 44760 34270 -23.44%
BenchmarkTCP4Persistent-6 45280 43000 -5.04%
R=golang-dev, alex.brainman, coocood, rsc
CC=golang-dev
https://golang.org/cl/7612045
I'm not sure how to write a test for this. The change in
behaviour is that if you somehow get a SIGBUS signal for an
address >= 0x1000, the program will now crash rather than
calling panic. As far as I know, on x86 GNU/Linux, the only
way to get a SIGBUS (rather than a SIGSEGV) is to set the
stack pointer to an invalid value.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7906045
On Darwin and FreeBSD, the mmap syscall return value is returned
unmodified. This means that the return value will either be a
valid address or a positive error number.
Also check return value from mmap in SysReserve - the callers of
SysReserve expect nil to be returned if the allocation failed.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/7871043
Rather than just checking for ENOMEM, check for a return value of less
than 4096, so that we catch other errors such as EACCES and EINVAL.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/7942043
Fixes#5061.
Current code relies on the fact that fd's are automatically removed from epoll set when closed. However, it is not true. Underlying file description is removed from epoll set only when *all* fd's referring to it are closed.
There are 2 bad consequences:
1. Kernel delivers notifications on already closed fd's.
2. The following sequence of events leads to error:
- add fd1 to epoll
- dup fd1 = fd2
- close fd1 (not removed from epoll since we've dup'ed the fd)
- dup fd2 = fd1 (get the same fd as fd1)
- add fd1 to epoll = EEXIST
So, if fd can be potentially dup'ed of fork'ed, it's necessary to explicitly remove the fd from epoll set.
R=golang-dev, bradfitz, dave
CC=golang-dev
https://golang.org/cl/7870043
Add missing CLOSUREVAR in switch.
Mark MAKE, string conversion nodes as impossible.
Control statements do not need instrumentation.
Instrument COM and LROT nodes.
Instrument map length.
Update #4228
R=dvyukov, golang-dev
CC=golang-dev
https://golang.org/cl/7504047
Hashtable is arranged as an array of
8-entry buckets with chained overflow.
Each bucket has 8 extra hash bits
per key to provide quick lookup within
a bucket. Table is grown incrementally.
Update #3885
Go time drops from 0.51s to 0.34s.
R=r, rsc, m3b, dave, bradfitz, khr, ugorji, remyoudompheng
CC=golang-dev
https://golang.org/cl/7504044
Inserting a key-value pair into a hashmap storing keys or values
indirectly can cause the garbage collector to find the hashmap in
an inconsistent state.
Fixes#5074.
R=golang-dev, minux.ma, rsc
CC=golang-dev
https://golang.org/cl/7913043
On NetBSD tv_sec is already an int64 so no need for a test.
On OpenBSD, semasleep expects a Unix time as argument,
and 1<<30 is in 2004.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7810044
The current SysAlloc implementation suffers from a signed vs unsigned
comparision bug. Since the error code from mmap is negated, the
unsigned comparision of v < 4096 is always false on error. Fix this
by switching to the darwin/freebsd/linux mmap model and leave the mmap
return value unmodified.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7870044
This provides a way to generate core dumps when people need them.
The settings are:
GOTRACEBACK=0 no traceback on panic, just exit
GOTRACEBACK=1 default - traceback on panic, then exit
GOTRACEBACK=2 traceback including runtime frames on panic, then exit
GOTRACEBACK=crash traceback including runtime frames on panic, then crash
Fixes#3257.
R=golang-dev, devon.odell, r, daniel.morsing, ality
CC=golang-dev
https://golang.org/cl/7666044
NEGL does a negation of the bottom 32 bits and then zero-extends to 64 bits,
resulting in a negative 32-bit number but a positive 64-bit number.
NEGQ does a full 64-bit negation, so that the result is negative both as
a 32-bit and as a 64-bit number.
This doesn't matter for the functions that are declared to return int32.
It only matters for the ones that return int64 or void* [sic].
This will fix the current incorrect error in the OpenBSD/amd64 build.
The build will still be broken, but it won't report a bogus error.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/7536046
Bring net/fd_linux.go back (it was deleted this morning)
because it is still needed for ARM.
Fix a few typos in the runtime reorg.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7759046
thread_GOOS.c becomes os_GOOS.c.
signal_GOOS_GOARCH.c becomes os_GOOS_GOARCH.c,
but with non-GOARCH-specific code moved into os_GOOS.c.
The actual arch-specific signal handler moves into signal_GOARCH.c
to avoid per-GOOS duplication.
New files signal_GOOS_GOARCH.h provide macros for
accessing fields of the very system-specific signal info structs.
Lots moving, but nothing changing.
This is a preliminarly cleanup so I can work on the signal
handling code to fix some open issues without having to
make each change 13 times.
Tested on Linux and OS X, 386 and amd64.
Will fix Plan 9, Windows, and ARM after the fact if necessary.
(Plan 9 and Windows should be fine; ARM will probably have some typos.)
Net effect: -1081 lines of code.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7565048
With the global redefinition of runtime·open by CL 7543043,
we need to provide a third argument and remove the cast
to the string.
Fixes build on 386 version of Plan 9.
R=khr, rsc, rminnich, ality
CC=golang-dev
https://golang.org/cl/7644047
Uses AES hardware instructions on 386/amd64 to implement
a fast hash function. Incorporates a random key to
thwart hash collision DOS attacks.
Depends on CL#7548043 for new assembly instructions.
Update #3885
Helps some by making hashing faster. Go time drops from
0.65s to 0.51s.
R=rsc, r, bradfitz, remyoudompheng, khr, dsymonds, minux.ma, elias.naur
CC=golang-dev
https://golang.org/cl/7543043
The issue was that scvg is assigned *after* the scavenger goroutine is started,
so when the scavenger calls entersyscall() the g==scvg check can fail.
Fixes#5025.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7629045
The problem is that there are lots of dead G's from previous tests,
each dead G consumes 1 stack segment.
Fixes#5034.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/7749043
The call to the C function runtime.findnull() requires
that we provide the argument at 0(SP).
R=rsc, rminnich, ality
CC=golang-dev
https://golang.org/cl/7559047
Otherwise the next goroutine run on the m
can get inadvertently locked if it executes a cgo call
that turns on the internal lock.
While we're here, fix the cgo panic unwind to
decrement m->ncgo like the non-panic unwind does.
Fixes#4971.
R=golang-dev, iant, dvyukov
CC=golang-dev
https://golang.org/cl/7627043
Now the default startup is that the program begins at _rt0_386_$GOOS,
which behaves as if calling main(argc, argv). Main jumps to _rt0_386.
This makes the _rt0_386 entry match the expected semantics for
the standard C "main" function, which we can now provide for use when
linking against a standard C library.
386 analogue of https://golang.org/cl/7525043
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7551045
Change 231af8ac63aa (CL 7314062) made runtime.enteryscall()
set m->mcache = nil, which means that we can no longer use
syscall.errstr in syscall.Syscall and syscall.Syscall6, since it
requires a new buffer to be allocated for holding the error string.
Instead, we use pre-allocated per-M storage to hold error strings
from syscalls made while in entersyscall mode, and call
runtime.findnull to calculate the lengths.
Fixes#4994.
R=rsc, rminnich, ality, dvyukov, rminnich, r
CC=golang-dev
https://golang.org/cl/7567043
The deadlock episodically occurs on misc/cgo/test/TestCthread.
The problem is that starttheworld() leaves some P's with local work
without M's. Then all active M's enter into syscalls, but reject to
wake another M's due to the following check (both in entersyscallblock() and in retake()):
if(p->runqhead == p->runqtail &&
runtime·atomicload(&runtime·sched.nmspinning) +
runtime·atomicload(&runtime·sched.npidle) > 0)
continue;
R=rsc
CC=golang-dev
https://golang.org/cl/7424054
Still to do: non-linux and non-amd64.
It may work on other ELF-based amd64 systems too, but untested.
"go test -ldflags -hostobj $GOROOT/misc/cgo/test" passes.
Much may yet change, but this seems a reasonable checkpoint.
R=iant
CC=golang-dev
https://golang.org/cl/7369057
Now the default startup is that the program begins at _rt0_amd64_$GOOS,
which sets DI = argc, SI = argv and jumps to _rt0_amd64.
This makes the _rt0_amd64 entry match the expected semantics for
the standard C "main" function, which we can now provide for use when
linking against a standard C library.
R=golang-dev, devon.odell, minux.ma
CC=golang-dev
https://golang.org/cl/7525043
broke arm garbage collector
traceback_arm fails with a missing pc. It needs CL 7494043.
But that only makes the build break later, this time with
"invalid freelist". Roll back until it can be fixed correctly.
««« original CL description
runtime: restrict stack root scan to locals and arguments
R=rsc
CC=golang-dev
https://golang.org/cl/7301062
»»»
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/7493044
If the constant CollectStats is non-zero and GOGCTRACE=1
the garbage collector will print basic statistics about executed
GC instructions.
R=golang-dev, dvyukov
CC=golang-dev, rsc
https://golang.org/cl/7413049
Fixes#4904.
The problem was that when the test runs the heap had grown to ~100MB,
so GC allows it to grow to 200MB, and so the test fails.
Moving the test to a separate process makes it much more isolated and stable.
R=golang-dev, minux.ma
CC=golang-dev
https://golang.org/cl/7441046
Also make the crossover point an architecture-dependent constant,
although it's the same everywhere for now.
BenchmarkAppendStr1Byte 416 145 -65.14%
BenchmarkAppendStr4Bytes 743 217 -70.79%
BenchmarkAppendStr8Bytes 421 270 -35.87%
BenchmarkAppendStr16Bytes 415 403 -2.89%
BenchmarkAppendStr32Bytes 415 391 -5.78%
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7459044
Putting the M initialization in multiple places will not scale.
Various code assumes mstart is the start already. Make it so.
R=golang-dev, devon.odell
CC=golang-dev
https://golang.org/cl/7420048
There is a #pragma dynexport crosscall2, to help SWIG,
and 6l cannot export the symbol if it doesn't get to see it.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7448044
There are some function pointers declared by 6c in
package runtime without initialization and then also
declared in package runtime/cgo with initialization,
so that if runtime/cgo is linked in, the function pointers
are non-nil, and otherwise they are nil. We depend on
this property for implementing non-essential cgo hooks
in package runtime.
The declarations in package runtime are 6c-compiled
and end up in .6 files. The declarations in package runtime/cgo
are gcc-compiled and end up in .o files. Since 6l links the .6
and .o files together, this all works.
However, when we switch to "external linking" mode,
6l will not see the .o files, and it would be up to the host linker
to resolve the two into a single initialized symbol.
Not all host linkers will do this (in particular OS X gcc will not).
To fix this, move the cgo declarations into 6c-compiled code,
so that they end up in .6 files, so that 6l gets them no matter what.
R=golang-dev
CC=golang-dev
https://golang.org/cl/7440045
The naming in this package is a disaster.
Make it all consistent.
Remove some 'static' from functions that will
be referred to from other files soon.
This CL is purely renames using global search and replace.
Submitting separately so that real changes will not
be drowned out by these renames in future CLs.
TBR=iant
CC=golang-dev
https://golang.org/cl/7416046
runtime: double-check that symbol table is sorted
If the symbol table is unsorted, the binary search in findfunc
will not find its func, which will make stack traces stop early.
When the garbage collector starts using the stack tracer,
that would be a serious problem.
The unsorted symbol addresses came from from two things:
1. The symbols in an ELF object are not necessarily sorted,
so sort them before adding them to the symbol list.
2. The __i686.get_pc_thunk.bx symbol is present in multiple
object files and was having its address adjusted multiple
times, producing an incorrect address in the symbol table.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7440044
Revision 6a88e1893941 corrupts the argument to
racefuncenter by pushing the data block pointer
to the stack.
Fixes#4885.
R=dvyukov, rsc
CC=golang-dev
https://golang.org/cl/7381053
add per-P cache of dead G's
add global runnable queue (not used for now)
add list of idle P's (not used for now)
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/7397061
The current code uses 64-bit pc-relative on 64-bit systems,
but in ELF linkers there is no such thing, so we cannot
express this in a .o file. Change to 32-bit.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7383055
This CL changes the encoding used for the Go symbol table,
stored in the binary and used at run time. It does not change
any of the semantics or structure: the bits are just packed
a little differently.
The comment at the top of runtime/symtab.c describes the new format.
Compared to the Go 1.0 format, the main changes are:
* Store symbol addresses as full-pointer-sized host-endian values.
(For 6g, this means addresses are 64-bit little-endian.)
* Store other values (frame sizes and so on) varint-encoded.
The second change more than compensates for the first:
for the godoc binary on OS X/amd64, the new symbol table
is 8% smaller than the old symbol table (1,425,668 down from 1,546,276).
This is a required step for allowing the host linker (gcc) to write
the final Go binary, since it will have to fill in the symbol address slots
(so the slots must be host-endian) and on 64-bit systems it may
choose addresses above 4 GB.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7403054
Separates the implementation of nanotime on 64-bit
version of Plan 9 from that on the 32-bit version.
The former uses a syscall.
R=rsc, rminnich, ality
CC=golang-dev
https://golang.org/cl/7379051
sigprocmask() is process-wide on darwin, so two concurrent
libcgo_sys_thread_start() can result in all signals permanently
blocked, which in particular blocks handling of nil derefs.
Fixes#4833.
R=golang-dev, dave, rsc
CC=golang-dev
https://golang.org/cl/7324058
For now, all the callbacks from C use top-level Go functions,
so they use the equivalent C function pointer, and will continue
to do so. But perhaps some day this will be useful for calling
a Go func value (at least if the type is already known).
More importantly, the Windows callback code needs to be able
to use cgocallback_gofunc to call a Go func value.
Should fix the Windows build.
R=ken2
CC=golang-dev
https://golang.org/cl/7388049
Change ARM context register to R7, to get out of the way
of the register allocator during the compilation of the
prologue statements (it wants to use R0 as a temporary).
Step 2 of http://golang.org/s/go11func.
R=ken2
CC=golang-dev
https://golang.org/cl/7369048
runtime: add context argument to gogocall
Too many other things use AX, and at least one
(stack zeroing) cannot be moved onto a different
register. Use the less special DX instead.
Preparation for step 2 of http://golang.org/s/go11func.
Nothing interesting here, just split out so that we can
see it's correct before moving on.
R=ken2
CC=golang-dev
https://golang.org/cl/7395050
Previously, the func structure contained an inaccurate value for
the args member and a 0 value for the locals member.
This change populates the func structure with args and locals
values computed by the compiler. The number of args was
already available in the ATEXT instruction. The number of
locals is now passed through in the new ALOCALS instruction.
This change also switches the unit of args and locals to be
bytes, just like the frame member, instead of 32-bit words.
R=golang-dev, bradfitz, cshapiro, dave, rsc
CC=golang-dev
https://golang.org/cl/7399045
mpreinit() is called on the parent thread and with mcache (can allocate memory),
minit() is called on the child thread and can not allocate memory.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/7389043
The removed code leads to the situation when M executes the same locked G again
and again.
This is https://golang.org/cl/7310096 but with return instead of break
in the nested switch.
Fixes#4820.
R=golang-dev, alex.brainman, rsc
CC=golang-dev
https://golang.org/cl/7304102
If a test can be placed in the same package ("internal"), it is placed
there. This facilitates testing of package-private details. Because of
dependency cycles some packages cannot be tested by internal tests.
R=golang-dev, rsc, mikioh.mikioh
CC=golang-dev, r
https://golang.org/cl/7323044
* Handle p==nil in signalstack by setting SS_DISABLE flag.
* Make minit only allocate a signal g if there's not one already.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7323072
Fix the sa_mask member of the sigaction struct - on FreeBSD this is
declared as a sigset_t, which is an array of four unsigned ints.
Replace the current int64 with Sigset from defs_freebsd_GOARCH, which
has the correct definition.
Unbreaks the FreeBSD builds.
R=golang-dev, dave, minux.ma
CC=golang-dev
https://golang.org/cl/7333047
broke windows build
««« original CL description
runtime: ensure forward progress of runtime.Gosched() for locked goroutines
The removed code leads to the situation when M executes the same locked G again and again.
Fixes#4820.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/7310096
»»»
TBR=dvyukov
CC=golang-dev
https://golang.org/cl/7343050
Arguably if this happens the program is buggy anyway,
but letting the panic continue looks better than interrupting it.
Otherwise things like this are possible, and confusing:
$ go run x.go
panic: $ echo $?
0
$
Fixes#3934.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/7322083
This is the same logic used in the standard tracebacks.
The caller pc is the pc after the call, so except in the
fake "call" caused by a panic, back up the pc enough
that the lookup will use the previous instruction.
Fixes#4150.
Fixes#4151.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7317047
Before, the mheap structure was in the bss,
but it's quite large (today, 256 MB, much of
which is never actually paged in), and it makes
Go binaries run afoul of exec-time bss size
limits on some BSD systems.
Fixes#4447.
R=golang-dev, dave, minux.ma, remyoudompheng, iant
CC=golang-dev
https://golang.org/cl/7307122
The removed code leads to the situation when M executes the same locked G again and again.
Fixes#4820.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/7310096
In addition to the compile failure fixed in signal*.c,
preserving the signal mask led to very strange crashes.
Testing shows that looking for SIG_IGN is all that
matters to get along with nohup, so reintroduce
sigset_zero instead of trying to preserve the signal mask.
TBR=iant
CC=golang-dev
https://golang.org/cl/7323067
There are two ways nohup(1) might be implemented:
it might mask away the signal, or it might set the handler
to SIG_IGN, both of which are inherited across fork+exec.
So two fixes:
* Make sure to preserve the inherited signal mask at
minit instead of clearing it.
* If the SIGHUP handler is SIG_IGN, leave it that way.
Fixes#4491.
R=golang-dev, mikioh.mikioh, iant
CC=golang-dev
https://golang.org/cl/7308102
No code changes.
This is mainly in preparation to scheduler changes,
oldstack/newstack are not related to scheduling.
R=golang-dev, minux.ma, rsc
CC=golang-dev
https://golang.org/cl/7311085
With the new scheduler races in the tests are reported during execution of other tests.
The change joins goroutines started during the tests.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/7310066
The problem happens when end=0, then end-1 is very big number.
Observed with the new scheduler.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7307073
Removes limit on maximum number of goroutines ever existed.
code.google.com/p/goexecutor tests now pass successfully.
Also slightly improves performance.
Before: $ time ./flate.test -test.short
real 0m9.314s
After: $ time ./flate.test -test.short
real 0m8.958s
Fixes#4286.
The runtime is built from llvm rev 174312.
R=rsc
CC=golang-dev
https://golang.org/cl/7218044
Change the stack unwinding code to compensate for the dynamic
relocation of symbols.
Change the gc instruction GC_CALL to use a relative offset instead of
an absolute address.
R=golang-dev
CC=golang-dev
https://golang.org/cl/7248048
* Separate internal and external LockOSThread, for cgo safety.
* Show goroutine that made faulting cgo call.
* Never start a panic due to a signal caused by a cgo call.
Fixes#3774.
Fixes#3775.
Fixes#3797.
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/7228081
Binary data in mprof.goc may prevent the garbage collector from freeing
memory blocks. This patch replaces all calls to runtime·mallocgc() with
calls to an allocator private to mprof.goc, thus making the private
memory invisible to the garbage collector. The addrhash variable is
moved outside of the .bss section.
R=golang-dev, dvyukov, rsc, minux.ma
CC=dave, golang-dev, remyoudompheng
https://golang.org/cl/7135063
This change also resolves some issues with note handling: we now make
sure that there is enough room at the bottom of every goroutine to
execute the note handler, and the `exitstatus' is no longer a global
entity, which resolves some race conditions.
R=rminnich, npe, rsc, ality
CC=golang-dev
https://golang.org/cl/6569068
Range access functions are already available in TSan library
but were not yet used.
Time for go test -race -short:
Before:
compress/flate 24.244s
exp/norm >200s
go/printer 78.268s
After:
compress/flate 17.760s
exp/norm 5.537s
go/printer 5.738s
Fixes#4250.
R=dvyukov, golang-dev, fullung
CC=golang-dev
https://golang.org/cl/7229044
Useful for debugging of runtime bugs.
+ Do not print "stack segment boundary" unless GOTRACEBACK>1.
+ Do not traceback system goroutines unless GOTRACEBACK>1.
R=rsc, minux.ma
CC=golang-dev
https://golang.org/cl/7098050
Mark candidate spans one GC pass earlier.
Move scavenger's code out from mgc0 and constrain it into mheap (where it belongs).
R=rsc, dvyukov, minux.ma
CC=golang-dev
https://golang.org/cl/7002049
This is for SPARC64, a 64-bit processor that uses all 64-bits
of virtual addresses. The idea is to use the low order 3 bits
to at least get a small ABA counter. That should work since
pointers are aligned. The idea is for SPARC64 to set CNT_MASK
== 7, PTR_BITS == 0, PTR_MASK == 0xffffffffffffff8.
Also add uintptr casts to avoid GCC warnings. The gccgo
runtime code is compiled with GCC, and GCC warns when casting
between a pointer and a type of a different size.
R=dvyukov
CC=golang-dev
https://golang.org/cl/7225043
Currently it's summed to mark phase.
The change makes it easier to diagnose long stop-the-world phases.
R=golang-dev, bradfitz, dave
CC=golang-dev
https://golang.org/cl/7182043
If the scanned block has no typeinfo the garbage collector will attempt
to get the actual type of the block.
R=golang-dev, bradfitz, rsc
CC=golang-dev
https://golang.org/cl/7093045
Also undo revision a5b96b602690 used to workaround the bug.
Fixes#4643.
R=rsc, golang-dev, dave, minux.ma, lucio.dere, bradfitz
CC=golang-dev
https://golang.org/cl/7090043
The Plan 9 symbol table format defines big-endian symbol values
for portability, but we want to be able to generate an ELF object file
and let the host linker link it, as part of the solution to issue 4069.
The symbol table itself, since it is loaded into memory at run time,
must be filled in by the final host linker, using relocation directives
to set the symbol values. On a little-endian machine, the linker will
only fill in little-endian values during relocation, so we are forced
to use little-endian symbol values.
To preserve most of the original portability of the symbol table
format, we make the table itself say whether it uses big- or
little-endian values. If the table begins with the magic sequence
fe ff ff ff 00 00
then the actual table begins after those six bytes and contains
little-endian symbol values. Otherwise, the table is in the original
format and contains big-endian symbol values. The magic sequence
looks like an "end of table" entry (the fifth byte is zero), so legacy
readers will see a little-endian table as an empty table.
All the gc architectures are little-endian today, so the practical
effect of this CL is to make all the generated tables little-endian,
but if a big-endian system comes along, ld will not generate
the magic sequence, and the various readers will fall back to the
original big-endian interpretation.
R=ken2
CC=golang-dev
https://golang.org/cl/7066043
sysarch requires arguments to be passed on the stack, not in registers.
Credit to Shenghou Ma (minux) for the fix.
R=minux.ma, devon.odell
CC=golang-dev
https://golang.org/cl/7037043
Used to then die on a nil pointer situation. Most Linux standard setups are rather
restrictive regarding the default amount of lockable memory.
R=minux.ma, rsc
CC=golang-dev
https://golang.org/cl/6997049
Currently it silently "succeeds" saying that it run 0 tests
if there are compilations errors.
With this change it fails and outputs the compilation error.
R=golang-dev, remyoudompheng
CC=golang-dev
https://golang.org/cl/7002058
When we release memory to the OS, if the OS doesn't want us
to release it (for example, because the program executed
mlockall(MCL_FUTURE)), madvise will fail. Ignore the failure
instead of crashing.
Fixes#3435.
R=ken2
CC=golang-dev
https://golang.org/cl/6998052
This disables checks for limited address space
and unlimited stack. They are not required for Go.
Fixes#4577.
R=golang-dev, iant
CC=golang-dev, kamil.kisiel, minux.ma
https://golang.org/cl/7003045
Enable cgo on OpenBSD.
The OpenBSD ld.so(1) does not currently support PT_TLS sections. Work
around this by fixing up the TCB that has been provided by librthread
and reallocating a TCB with additional space for TLS. Also provide a
wrapper for pthread_create, allowing zeroed TLS to be allocated for
threads created externally to Go.
Joint work with Shenghou Ma (minux).
Requires change 6846064.
Fixes#3205.
R=golang-dev, minux.ma, iant, rsc, iant
CC=golang-dev
https://golang.org/cl/6853059
With this change the runtime can now read GOMAXPROCS, GOGC, etc.
I'm not quite sure how we missed this.
R=seed, lucio.dere, rsc
CC=golang-dev
https://golang.org/cl/6935062
The code:
func main() {
v := make([]int64, 10)
i := 1
_ = v[(i*4)/3]
}
crashes compiler with:
Program received signal SIGSEGV, Segmentation fault.
0x000000000043c274 in walkexpr (np=0x7fffffffc9b8, init=0x0) at src/cmd/gc/walk.c:587
587 *init = concat(*init, n->ninit);
(gdb) bt
#0 0x000000000043c274 in walkexpr (np=0x7fffffffc9b8, init=0x0) at src/cmd/gc/walk.c:587
#1 0x0000000000432d15 in copyexpr (n=0x7ffff7f69a48, t=<optimized out>, init=0x0) at src/cmd/gc/subr.c:2020
#2 0x000000000043f281 in walkdiv (init=0x0, np=0x7fffffffca70) at src/cmd/gc/walk.c:2901
#3 walkexpr (np=0x7ffff7f69760, init=0x0) at src/cmd/gc/walk.c:956
#4 0x000000000043d801 in walkexpr (np=0x7ffff7f69bc0, init=0x0) at src/cmd/gc/walk.c:988
#5 0x000000000043cc9b in walkexpr (np=0x7ffff7f69d38, init=0x0) at src/cmd/gc/walk.c:1068
#6 0x000000000043c50b in walkexpr (np=0x7ffff7f69f50, init=0x0) at src/cmd/gc/walk.c:879
#7 0x000000000043c50b in walkexpr (np=0x7ffff7f6a0c8, init=0x0) at src/cmd/gc/walk.c:879
#8 0x0000000000440a53 in walkexprlist (l=0x7ffff7f6a0c8, init=0x0) at src/cmd/gc/walk.c:357
#9 0x000000000043d0bf in walkexpr (np=0x7fffffffd318, init=0x0) at src/cmd/gc/walk.c:566
#10 0x00000000004402bf in vmkcall (fn=<optimized out>, t=0x0, init=0x0, va=0x7fffffffd368) at src/cmd/gc/walk.c:2275
#11 0x000000000044059a in mkcall (name=<optimized out>, t=0x0, init=0x0) at src/cmd/gc/walk.c:2287
#12 0x000000000042862b in callinstr (np=0x7fffffffd4c8, init=0x7fffffffd568, wr=0, skip=<optimized out>) at src/cmd/gc/racewalk.c:478
#13 0x00000000004288b7 in racewalknode (np=0x7ffff7f68108, init=0x7fffffffd568, wr=0, skip=0) at src/cmd/gc/racewalk.c:287
#14 0x0000000000428781 in racewalknode (np=0x7ffff7f65840, init=0x7fffffffd568, wr=0, skip=0) at src/cmd/gc/racewalk.c:302
#15 0x0000000000428abd in racewalklist (l=0x7ffff7f65840, init=0x0) at src/cmd/gc/racewalk.c:97
#16 0x0000000000428d0b in racewalk (fn=0x7ffff7f5f010) at src/cmd/gc/racewalk.c:63
#17 0x0000000000402b9c in compile (fn=0x7ffff7f5f010) at src/cmd/6g/../gc/pgen.c:67
#18 0x0000000000419f86 in funccompile (n=0x7ffff7f5f010, isclosure=0) at src/cmd/gc/dcl.c:1414
#19 0x0000000000424161 in p9main (argc=<optimized out>, argv=<optimized out>) at src/cmd/gc/lex.c:431
#20 0x0000000000401739 in main (argc=<optimized out>, argv=<optimized out>) at src/lib9/main.c:35
The problem is nil init passed to mkcall().
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6940045
Details:
- This CL is the conceptual skeleton of code found in CL 6114046
- The garbage collector uses struct Obj to specify memory blocks
- scanblock() is putting found memory blocks into an intermediate buffer
(xbuf) before adding/flushing them to the main work buffer (wbuf)
- The main loop in scanblock() is replaced with a skeleton code that
in the future will be able to recognize the type of objects and
thus will improve the garbage collector's precision.
For now, all objects are simply sequences of pointers so
the precision of the garbage collector remains unchanged.
- The code plugs .gcdata and .gcbss sections into the garbage collector.
scanblock() in this CL is unable to make any use of this.
R=rsc, dvyukov, remyoudompheng
CC=dave, golang-dev, minux.ma
https://golang.org/cl/6856121
This includes GORACE history_size and log_path flags.
R=golang-dev, bradfitz, rsc, remyoudompheng, minux.ma
CC=golang-dev
https://golang.org/cl/6947046
When a race happens inside of runtime (chan, slice, etc),
currently reports contain only user file:line.
If the line contains a complex expression,
it's difficult to figure out where the race exactly.
This change adds one more top frame with exact
runtime function (e.g. runtime.chansend, runtime.mapaccess).
R=golang-dev
CC=golang-dev
https://golang.org/cl/6851125
Garbage collection code (to be merged later) is calling functions
which have many local variables. This increases the probability that
the stack capacity won't be big enough to hold the local variables.
So, start gc() on a bigger stack to eliminate a potentially large number
of calls to runtime·morestack().
R=rsc, remyoudompheng, dsymonds, minux.ma, iant, iant
CC=golang-dev
https://golang.org/cl/6846044
madvise was missing so implement it in assembler. This change
needs to be extended to the other BSD variantes (Net and Open)
Without this change the scavenger will attempt to pass memory back
to the operating system when it has become idle, but the memory is
not returned and for long running Go processes the total memory used
can grow until OOM occurs.
I have only been able to test the code on FreeBSD AMD64. The ARM
platforms needs testing.
R=golang-dev, mikioh.mikioh, dave, jgc, minux.ma
CC=golang-dev
https://golang.org/cl/6850081
Update OpenBSD runtime to use the new version of the sys___tfork
syscall and switch TLS initialisation from sys_arch to sys___set_tcb
(note that both of these syscalls are available in OpenBSD 5.2).
R=golang-dev, minux.ma
CC=golang-dev
https://golang.org/cl/6843058
This enables to loop over some goroutines, e.g. to print the
backtrace of goroutines 1 to 9:
set $i = 1
while $i < 10
printf "backtrace of goroutine %d:\n", $i
goroutine $i++ bt
end
R=lvd, lvd
CC=golang-dev
https://golang.org/cl/6843071
This significantly decreases amount of shadow memory
mapped by race detector.
I haven't tested on Windows, but on Linux it reduces
virtual memory size from 1351m to 330m for fmt.test.
Fixes#4379.
R=golang-dev, alex.brainman, iant
CC=golang-dev
https://golang.org/cl/6849057
Currently race detector runtime just disables race detection in the finalizer goroutine.
It has false positives when a finalizer writes to shared memory -- the race with finalizer is reported in a normal goroutine that accesses the same memory.
After this change I am going to synchronize the finalizer goroutine with the rest of the world in racefingo(). This is closer to what happens in reality and so
does not have false positives.
And also add README file with instructions how to build the runtime.
R=golang-dev, minux.ma, rsc
CC=golang-dev
https://golang.org/cl/6810095
It allows to catch e.g. a data race between atomic write and non-atomic write,
or Mutex.Lock() and mutex overwrite (e.g. mu = Mutex{}).
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6817103
In order to add these, we need to be able to find references
to such types that already exist in the binary. To do that, introduce
a new linker section holding a list of the types corresponding to
arrays, chans, maps, and slices.
To offset the storage cost of this list, and to simplify the code,
remove the interface{} header from the representation of a
runtime type. It was used in early versions of the code but was
made obsolete by the kind field: a switch on kind is more efficient
than a type switch.
In the godoc binary, removing the interface{} header cuts two
words from each of about 10,000 types. Adding back the list of pointers
to array, chan, map, and slice types reintroduces one word for
each of about 500 types. On a 64-bit machine, then, this CL *removes*
a net 156 kB of read-only data from the binary.
This CL does not include the needed support for precise garbage
collection. I have created issue 4375 to track that.
This CL also does not set the 'algorithm' - specifically the equality
and copy functions - for a new array correctly, so I have unexported
ArrayOf for now. That is also part of issue 4375.
Fixes#2339.
R=r, remyoudompheng, mirtchovski, iant
CC=golang-dev
https://golang.org/cl/6572043
Otherwise a poorly timed GC can collect the memory before it
is returned to the Go program.
R=golang-dev, dave, dvyukov, minux.ma
CC=golang-dev
https://golang.org/cl/6819119
Re-enable the crash tests on NetBSD now that the issue has been
identified and fixed.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/6813100
Currently race detector runtime maps shadow memory eagerly at process startup.
It works poorly on Windows, because Windows requires reservation in swap file
(especially problematic if several Go program runs at the same, each consuming GBs
of memory).
With this change race detector maps shadow memory lazily, so Go runtime must notify
about all new heap memory.
It will help with Windows port, but also eliminates scary 16TB virtual mememory
consumption in top output (which sometimes confuses some monitoring scripts).
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6811085
When the first result of a type assertion is blank, the compiler would still copy out a potentially large non-interface type.
Fixes#1021.
R=golang-dev, bradfitz, rsc
CC=golang-dev
https://golang.org/cl/6812079
It speedups the race detector somewhat, but also prevents
getcallerpc() from obtaining lessstack().
R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/6812091
The deadlock occurs when another goroutine requests GC
during the test. When wait=true the test expects physical parallelism,
that is, that P goroutines are all active at the same time.
If GC is requested, then part of the goroutines are not scheduled,
so other goroutines deadlock.
With wait=false, goroutines finish parallel for w/o waiting for all
other goroutines.
Fixes#3954.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/6820098
The race detector does not understand ParFor synchronization, because it's implemented in C.
If run with -cpu=2 currently race detector says:
WARNING: DATA RACE
Read by goroutine 5:
runtime_test.TestParForParallel()
src/pkg/runtime/parfor_test.go:118 +0x2e0
testing.tRunner()
src/pkg/testing/testing.go:301 +0x8f
Previous write by goroutine 6:
runtime_test.func·024()
src/pkg/runtime/parfor_test.go:111 +0x52
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/6811082
PauseNs is a circular buffer of recent pause times, and the
most recent one is at [((NumGC-1)+256)%256].
Also fix comments cross-linking the Go and C definition of
various structs.
R=golang-dev, rsc, bradfitz
CC=golang-dev
https://golang.org/cl/6657047
If source are not available, then the stack looks like:
stack_test.go:40: /tmp/gobuilder/linux-amd64-race-72b15c5d6f65/go/src/pkg/runtime/debug/bla-bla-bla/src/pkg/runtime/debug/stack_test.go:15 (0x43fb11)
stack_test.go:40: /tmp/gobuilder/linux-amd64-race-72b15c5d6f65/go/src/pkg/runtime/debug/bla-bla-bla/src/pkg/runtime/debug/stack_test.go:18 (0x43fb7a)
stack_test.go:40: /tmp/gobuilder/linux-amd64-race-72b15c5d6f65/go/src/pkg/runtime/debug/bla-bla-bla/src/pkg/runtime/debug/stack_test.go:37 (0x43fbf4)
stack_test.go:40: /tmp/gobuilder/linux-amd64-race-72b15c5d6f65/go/src/pkg/testing/bla-bla-bla/src/pkg/testing/testing.go:301 (0x43b5ba)
stack_test.go:40: /tmp/gobuilder/linux-amd64-race-72b15c5d6f65/go/src/pkg/runtime/bla-bla-bla/src/pkg/runtime/proc.c:276 (0x410670)
stack_test.go:40:
which is 6 lines.
R=golang-dev, minux.ma
CC=golang-dev
https://golang.org/cl/6637060
Check for specific, important misalignment in garbage collector.
Not a complete fix for issue 599 but an important workaround.
Update #599.
R=golang-dev, iant, dvyukov
CC=golang-dev
https://golang.org/cl/6641049
Also add call to GC() to make it easier to re-enable the test.
Update #4155.
When we have precise GC merged, re-enable this test.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/6622058
The profiler collects goroutine blocking information similar to Google Perf Tools.
You may see an example of the profile (converted to svg) attached to
http://code.google.com/p/go/issues/detail?id=3946
The public API changes are:
+pkg runtime, func BlockProfile([]BlockProfileRecord) (int, bool)
+pkg runtime, func SetBlockProfileRate(int)
+pkg runtime, method (*BlockProfileRecord) Stack() []uintptr
+pkg runtime, type BlockProfileRecord struct
+pkg runtime, type BlockProfileRecord struct, Count int64
+pkg runtime, type BlockProfileRecord struct, Cycles int64
+pkg runtime, type BlockProfileRecord struct, embedded StackRecord
R=rsc, dave, minux.ma, r
CC=gobot, golang-dev, r, remyoudompheng
https://golang.org/cl/6443115
The Go run-time assumes that all SSE floating-point exceptions
are masked so that Go programs are not broken by such invalid
operations. By default, the 64-bit version of the Plan 9 kernel
masks only some SSE floating-point exceptions. Here, we mask
them all on a per-thread basis.
R=rsc, rminnich, minux.ma
CC=golang-dev
https://golang.org/cl/6592056
The assembly offsets were converted mechanically using
code.google.com/p/rsc/cmd/asmlint. The instruction
changes were done by hand.
Fixes#2188.
R=iant, r, bradfitz, remyoudompheng
CC=golang-dev
https://golang.org/cl/6550058
This CL makes the runtime understand that the type of
the len or cap of a map, slice, or string is 'int', not 'int32',
and it is also careful to distinguish between function arguments
and results of type 'int' vs type 'int32'.
In the runtime, the new typedefs 'intgo' and 'uintgo' refer
to Go int and uint. The C types int and uint continue to be
unavailable (cause intentional compile errors).
This CL does not change the meaning of int, but it should make
the eventual change of the meaning of int on amd64 a bit
smoother.
Update #2188.
R=iant, r, dave, remyoudompheng
CC=golang-dev
https://golang.org/cl/6551067
Using offsets from Tos is cumbersome and we've had problems
in the past. Since it's only being used to grab the PID, we'll just
get that from the default TLS instead.
R=rsc, rminnich, npe
CC=golang-dev
https://golang.org/cl/6543049
The change is a preparation for the new scheduler.
It introduces runtime.park() function,
that will atomically unlock the mutex and park the goroutine.
It will allow to remove the racy readyonstop flag
that is difficult to implement w/o the global scheduler mutex.
R=rsc, remyoudompheng, dave
CC=golang-dev
https://golang.org/cl/6501077
Fixes#3456.
This proposal is a reformulation of CL 5987063. This CL resets
the default GOARM value to 6 and allows the use of the VFPv3
optimisation if GOARM=7. Binaries built with this CL in place
will abort if GOARM=7 was used and the target host does not
support VFPv3.
R=minux.ma, rsc, ajstarks
CC=golang-dev
https://golang.org/cl/6501099
Fixes#3911.
Requires CL 6449127.
dfc@qnap:~$ ./runtime.test
runtime: this CPU has no floating point hardware, so it cannot run
this GOARM=7 binary. Recompile using GOARM=5.
R=rsc, minux.ma
CC=golang-dev
https://golang.org/cl/6442109
Signal handlers are global resources but many language
environments (Go, C++ at Google, etc) assume they have sole
ownership of a particular handler. Signal handlers in
mixed-language applications must therefore be robust against
unexpected delivery of certain signals, such as SIGPROF.
The default Go signal handler runtime·sigtramp assumes that it
will never be called on a non-Go thread, but this assumption
is violated by when linking in C++ code that spawns threads.
Specifically, the handler asserts the thread has an associated
"m" (Go scheduler).
This CL is a very simple workaround: discard SIGPROF delivered to non-Go threads. runtime.badsignal(int32) now receives the signal number; if it returns without panicking (e.g. sig==SIGPROF) the signal is discarded.
I don't think there is any really satisfactory solution to the
problem of signal-based profiling in a mixed-language
application. It's not only the issue of handler clobbering,
but also that a C++ SIGPROF handler called in a Go thread
can't unwind the Go stack (and vice versa). The best we can
hope for is not crashing.
Note:
- I've ported this to all POSIX platforms, except ARM-linux which already ignores unexpected signals on m-less threads.
- I've avoided tail-calling runtime.badsignal because AFAICT the 6a/6l don't support it.
- I've avoided hoisting 'push sig' (common to both function calls) because it makes the code harder to read.
- Fixed an (apparently incorrect?) docstring.
R=iant, rsc, minux.ma
CC=golang-dev
https://golang.org/cl/6498057
Reverts part of CL 6460082.
If a doc comment describes a type by explaining the
meaning of one instance of the type, a leading article
is fine and makes the text less awkward.
Compare:
// A dog is a kind of animal.
// Dog is a kind of animal.
R=golang-dev, dsymonds, dvyukov, r
CC=golang-dev
https://golang.org/cl/6494066
This set of changes extends the Plan 9 support
to include the AMD64 architecture and should
work on all versions of Plan 9.
R=golang-dev, rminnich, noah.evans, rsc, minux.ma, npe
CC=akskuma, golang-dev, jfflore, noah.evans
https://golang.org/cl/6479052
Use version 2 of the NetBSD signal ABI - both version 2 and version 3
are supported by the kernel, with near identical behaviour. However,
the netbsd32 compat code does not allow version 3 to be used, which
prevents Go netbsd/386 binaries from running in compat mode on a
NetBSD amd64 kernel. Switch to version 2 of the ABI, which is the
same version currently used by NetBSD's libc.
R=minux.ma
CC=golang-dev
https://golang.org/cl/6476068
When manipulating the stack pointer use the UESP register instead
of the ESP register, since the UESP register is the one that gets
restored from the machine context. Fixes broken tests on netbsd/386.
R=golang-dev, minux.ma, r, bsiegert
CC=golang-dev
https://golang.org/cl/6465054
Disable the crash handler test on NetBSD until I can figure out why
it triggers failures in later tests.
R=golang-dev, minux.ma
CC=golang-dev
https://golang.org/cl/6460090
Surrogate halves are part of UTF-16 and should never appear in UTF-8.
(The rune that two combined halves represent in UTF-16 should
be encoded directly.)
Encoding: encode as RuneError.
Decoding: convert to RuneError, consume one byte.
This requires changing:
package unicode/utf8
runtime for range over string
Also added utf8.ValidRune and fixed bug in utf.RuneLen.
Fixes#3927.
R=golang-dev, rsc, bsiegert
CC=golang-dev
https://golang.org/cl/6458099
Depends on CL 6197045.
Result obtained on Core i7 620M, Darwin/amd64:
benchmark old ns/op new ns/op delta
BenchmarkComplex128DivNormal 57 28 -50.78%
BenchmarkComplex128DivNisNaN 49 15 -68.90%
BenchmarkComplex128DivDisNaN 49 15 -67.88%
BenchmarkComplex128DivNisInf 40 12 -68.50%
BenchmarkComplex128DivDisInf 33 13 -61.06%
Result obtained on Core i7 620M, Darwin/386:
benchmark old ns/op new ns/op delta
BenchmarkComplex128DivNormal 89 50 -44.05%
BenchmarkComplex128DivNisNaN 307 802 +161.24%
BenchmarkComplex128DivDisNaN 309 788 +155.02%
BenchmarkComplex128DivNisInf 278 237 -14.75%
BenchmarkComplex128DivDisInf 46 22 -52.46%
Result obtained on 700MHz OMAP4460, Linux/ARM:
benchmark old ns/op new ns/op delta
BenchmarkComplex128DivNormal 1557 465 -70.13%
BenchmarkComplex128DivNisNaN 1443 220 -84.75%
BenchmarkComplex128DivDisNaN 1481 218 -85.28%
BenchmarkComplex128DivNisInf 952 216 -77.31%
BenchmarkComplex128DivDisInf 861 231 -73.17%
The 386 version has a performance regression, but as we have
decided to use SSE2 instead of x87 FPU for 386 too (issue 3912),
I won't address this issue.
R=dsymonds, mchaten, iant, dave, mtj, rsc, r
CC=golang-dev
https://golang.org/cl/6024045
our old choice is not working properly at least on VFPv2 in
ARM1136JF-S (it's not preserved across float64->float32 conversions).
Fixes#3745.
R=dave, rsc
CC=golang-dev
https://golang.org/cl/6344078
Since NUL usually terminates strings in underlying syscalls, allowing
it when converting string arguments is a security risk, especially
when dealing with filenames. For example, a program might reason that
filename like "/root/..\x00/" is a subdirectory or "/root/" and allow
access to it, while underlying syscall will treat "\x00" as an end of
that string and the actual filename will be "/root/..", which might
be unexpected. Returning EINVAL when string arguments have NUL in
them makes sure this attack vector is unusable.
R=golang-dev, r, bradfitz, fullung, rsc, minux.ma
CC=golang-dev
https://golang.org/cl/6458050
When a cgo program calls setuid, setgid, etc., the GNU/Linux
pthread library sends signal SIGSETXID to each thread to tell
it to update its UID info. If Go is permitted to intercept
the default SIGSETXID signal handler, the program will hang.
This patch tells the runtime package to not try to intercept
SIGSETXID on GNU/Linux. This will be odd if a Go program
wants to try to use that signal, but it means that cgo
programs that call setuid, etc., won't hang.
Fixes#3871.
R=rsc, r, minux.ma, bradfitz
CC=golang-dev
https://golang.org/cl/6455050
Move panic/defer/recover-related stuff from proc.c/runtime.c to a new file panic.c.
No semantic changes.
proc.c is 1800+ LOC and is a bit difficult to work with.
R=golang-dev, dave, r
CC=golang-dev
https://golang.org/cl/6343071
1. Rename 'g' and 'm' local vars to 'gp' and 'mp' (convention already used in some functions)
'g' and 'm' are global vars that mean current goroutine and current machine,
when they are shadowed by local vars, it's confusing, no ability to debug log both, etc.
2. White-space shuffling.
No semantic changes.
In preparation to bigger changes.
R=golang-dev, dave
CC=golang-dev
https://golang.org/cl/6355061
There may be further savings if convT2I can avoid the function call
if the cache is good and T is uintptr-shaped, a la convT2E, but that
will be a follow-up CL.
src/pkg/runtime:
benchmark old ns/op new ns/op delta
BenchmarkConvT2ISmall 43 15 -64.01%
BenchmarkConvT2IUintptr 45 14 -67.48%
BenchmarkConvT2ILarge 130 101 -22.31%
test/bench/go1:
benchmark old ns/op new ns/op delta
BenchmarkBinaryTree17 8588997000 8499058000 -1.05%
BenchmarkFannkuch11 5300392000 5358093000 +1.09%
BenchmarkGobDecode 30295580 31040190 +2.46%
BenchmarkGobEncode 18102070 17675650 -2.36%
BenchmarkGzip 774191400 771591400 -0.34%
BenchmarkGunzip 245915100 247464100 +0.63%
BenchmarkJSONEncode 123577000 121423050 -1.74%
BenchmarkJSONDecode 451969800 596256200 +31.92%
BenchmarkMandelbrot200 10060050 10072880 +0.13%
BenchmarkParse 10989840 11037710 +0.44%
BenchmarkRevcomp 1782666000 1716864000 -3.69%
BenchmarkTemplate 798286600 723234400 -9.40%
R=rsc, bradfitz, go.peter.90, daniel.morsing, dave, uriel
CC=golang-dev
https://golang.org/cl/6337058
The issue seems to not be triggered right now,
but I've seen the deadlock after some other legal
modifications to runtime.
So I think we are safer this way.
R=rsc, r
CC=golang-dev
https://golang.org/cl/6339051
This can only happen if the hash function we're using is getting
far more than it's fair share of collisions, but that has happened
to us repeatedly as we've expanded the allowed use cases for
hash tables (issue 1544, issue 2609, issue 2630, issue 2883, issue 3695).
Maybe this will help the next time we try something new.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/6306083
if we were to use sizeof(sa.sa_mask) instead of 8 as the last argument
to rt_sigaction, we would have already fixed this bug, so also updated
Linux/386 and Linux/amd64 files to use that; also test the return value
of rt_sigaction.
R=dave, rsc
CC=golang-dev
https://golang.org/cl/6297087
On netbsd/386, tv_sec is a 64-bit integer for both timeval and timespec.
Fix the time handling code so that it works correctly.
R=golang-dev, rsc, m4dh4tt3r
CC=golang-dev
https://golang.org/cl/6256056
Thanks to Dave Cheney for the magic words "comm page".
benchmark old ns/op new ns/op delta
BenchmarkNow 197 33 -83.05%
This should make profiling a little better on OS X.
The raw time saved is unlikely to matter: what likely matters
more is that it seems like OS X sends profiling signals on the
way out of system calls more often than it should; avoiding
the system call should increase the accuracy of cpu profiles.
The 386 version would be similar but needs to do different
math for CPU speeds less than 1 GHz. (Apparently Apple has
never shipped a 64-bit CPU with such a slow clock.)
R=golang-dev, bradfitz, dave, minux.ma, r
CC=golang-dev
https://golang.org/cl/6275056
amd64 was done in CL 6275056.
We don't attempt to handle machines with clock speeds
less than 1 GHz. Those will fall back to the system call.
benchmark old ns/op new ns/op delta
BenchmarkNow 364 38 -89.53%
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/6307045
Using an int64 for a block size doesn't make
sense on 32bit platforms but extracts a performance
penalty dealing with double word quantities on Arm.
linux/arm
benchmark old ns/op new ns/op delta
BenchmarkGobDecode 155401600 144589300 -6.96%
BenchmarkGobEncode 72772220 62460940 -14.17%
BenchmarkGzip 5822632 2604797 -55.26%
BenchmarkGunzip 326321 151721 -53.51%
benchmark old MB/s new MB/s speedup
BenchmarkGobDecode 4.94 5.31 1.07x
BenchmarkGobEncode 10.55 12.29 1.16x
R=golang-dev, rsc, bradfitz
CC=golang-dev
https://golang.org/cl/6272047
The previous code was preparing arrays of entries that would be
filled if there was one entry every 128 bytes. Moving to a 4096
byte interval reduces the overhead per megabyte of address space
to 2kB from 64kB (on 64-bit systems).
The performance impact will be negative for very small MemProfileRate.
test/bench/garbage/tree2 -heapsize 800000000 (default memprofilerate)
Before: mprof 65993056 bytes (1664 bucketmem + 65991392 addrmem)
After: mprof 1989984 bytes (1680 bucketmem + 1988304 addrmem)
R=golang-dev, rsc
CC=golang-dev, remy
https://golang.org/cl/6257069
The previous heap profile format did not include buckets with
zero used bytes. Also add several missing MemStats fields in
debug mode.
R=golang-dev, rsc
CC=golang-dev, remy
https://golang.org/cl/6249068
The correct procid is needed for unparking LWPs on NetBSD - always
initialise procid in minit() so that cgo works correctly. The non-cgo
case already works correctly since procid is initialised via
lwp_create().
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6257071
A block with finalizer might also be profiled. The special bit
is needed to unregister the block from the profile. It will be
unset only when the block is freed.
Fixes#3668.
R=golang-dev, rsc
CC=golang-dev, remy
https://golang.org/cl/6249066
It's sad to introduce a new macro, but rnd shows up consistently
in profiles, and the function call overwhelms the two arithmetic
instructions it performs.
R=r
CC=golang-dev
https://golang.org/cl/6260051
This is from CL 5451105 but was dropped from that CL.
See also CL 6137051.
The only change compared to 5451105 is to check for
h != nil in reflect·mapiterinit; allowing use of nil maps
must have happened after that original CL.
Fixes#3573.
R=golang-dev, dave, r
CC=golang-dev
https://golang.org/cl/6215078
The previous attempt to explain this got it backwards (all the more reason to be
sad we couldn't make the two functions behave the same).
Fixes#3669.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6249051
This fixes occasional 64-bit failures.
Maybe it will fix the 32-bit failures too,
so re-enable on 32-bit for now.
R=golang-dev, bradfitz, r, dvyukov
CC=golang-dev
https://golang.org/cl/6218050
With the timed semacquire patch
(kernel-tsemacquire) for Plan 9,
we can now properly do a timed
wait for the semaphore, in
semasleep.
R=golang-dev, rsc, rminnich, ality, r
CC=0intro, golang-dev, john, mirtchovski
https://golang.org/cl/6197046
Implement getcontext and sigprocmask for NetBSD - these will soon be
used by the thread handling code.
Also fix netbsd/386 signal handling - there is no sigreturn, just
return so that we hit the trampoline.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6215049
Update/correct NetBSD signal handling - most of this is needed due to
the correctly generated runtime definitions.
R=golang-dev, m4dh4tt3r, rsc
CC=golang-dev
https://golang.org/cl/6195079
Fix and regenerate runtime defs for NetBSD.
Whilst the mcontext struct can be handled across architectures,
the registers are provided as defines that index an array, rather
than as members of the struct. Since these are architecture
dependent, include them via a defs_netbsd_<arch>.go file.
R=golang-dev, m4dh4tt3r, rsc
CC=golang-dev
https://golang.org/cl/6190070
Parallel GC needs to know in advance how many helper threads will be there.
Hopefully it's the last patch before I can tackle parallel sweep phase.
The benchmarks are unaffected.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6200064
Set the TLS base using the _lwp_setprivate() syscall, instead of via
sysarch(). NetBSD tracks the pointer passed to _lwp_setprivate() and
restores this value when restoring mcontext. If sysarch() is used
directly, restoring an mcontext trashes the FS/GS value, resulting
in a segfault when we next try to access the TLS.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/6206062
we can't add the division result to n during iteration, because it might
turn n into NaN or Inf.
R=golang-dev, rsc, iant, iant
CC=golang-dev
https://golang.org/cl/6197045
Use correct syscall numbers and arguments for NetBSD.
Provide a trampoline for signal returns (using signal API 3).
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6209048
Decode AT_RANDOM, AT_HWCAP, and AT_PLATFORM.
This CL only make use of AT_RANDOM, but future CLs will make use of the others.
R=dave, rsc
CC=golang-dev
https://golang.org/cl/5978051
1. In CL 5989057, I made a mistake in the last minute change.
"MOVW.W R4, -4(SP)" should really be "MOVW.W R4, -4(R13)",
as 5l will rewrite offset for SP.
2. misc/cgo/test/issue1560.go tests for parallel sleep of 1s,
but on ARM, the deadline is frequently missed, so change sleep
time to 2s on ARM.
R=golang-dev, dave, rsc
CC=golang-dev
https://golang.org/cl/6202043
This adds proper note handling for Plan 9,
and fixes the issue of properly killing go procs.
Without this change, the first go proc that dies
(using runtime·exit()) would kill all the running
go procs. Proper signal handling is needed.
R=golang-dev, ality, rminnich, rsc
CC=golang-dev, john, mirtchovski
https://golang.org/cl/5617048
This makes it possible to inline the prefetch of upcoming
memory addresses during garbage collection, instead of
needing to flush registers, make a function call, and
reload registers. On garbage collection-heavy workloads,
this results in a 5% speedup.
Fixes#3493.
R=dvyukov, ken, r, dave
CC=golang-dev
https://golang.org/cl/5990066
Switch from using the rfork() syscall on OpenBSD, to the __tfork()
syscall. The __tfork() syscall is the preferred way of creating
system threads and the rfork() syscall has recently been removed.
Note: this will break compatibility with OpenBSD releases prior to 5.1.
R=golang-dev, bradfitz, devon.odell, rsc
CC=golang-dev
https://golang.org/cl/6037048
This lets the test pass on PPC64 GNU/Linux, which uses a much
larger page size and thus uses more memory to hold blocks
allocated for memory profiling.
R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/6048054
Update runtime defs for openbsd. Add struct __tfork, which will be
needed by an upcoming change.
R=golang-dev, adg
CC=golang-dev
https://golang.org/cl/6007050
Update the threxit and thrsleep syscalls to match the ABI of the
OpenBSD 5.1 kernel. These changes are backwards compatible with
older kernels.
Fixes#3311.
R=golang-dev, rsc, devon.odell
CC=golang-dev
https://golang.org/cl/5777079
This leads to ~30kB improvement on code size for ARM machines with VFP/NEON.
Example: go test -c math
GOARM=5 GOARM=6
Old: 1884200 1839144
New: 1884165 1805245
-: 35 33899
R=rsc, bradfitz, dave, kai.backman
CC=golang-dev
https://golang.org/cl/5975060
Change 5660047 moved an FLDCW instruction
that disables invalid operand traps into
runtime·asminit, which is called from
runtime·mstart. Thus, runtime·check is being
called prior to setting the appropriate control bits,
which on any QNaN comparison will cause Plan 9
to take an invalid operand trap. This change loads
the control bits (for Plan 9) prior to runtime·check.
Ideally, this should be done before the QNaN checks
on any system, but possibly other kernels simply
don't ever trap on invalid operands.
R=golang-dev, rminnich
CC=golang-dev, john, rsc
https://golang.org/cl/5939045
Block signals during thread creation, otherwise the new thread can
receive a signal prior to initialisation completing.
Fixes#3102.
R=golang-dev, rsc, devon.odell, minux.ma
CC=golang-dev
https://golang.org/cl/5757064
It is a bug that Caller and Callers disagree about the offset of the skip
parameter. Document the bug.
R=rsc, dsymonds, r, iant
CC=golang-dev
https://golang.org/cl/5976064
Not a complete fix for issue 3342, but fixes the trivial case.
There may still be a race in the instants before and after
a scavenger-induced garbage collection.
Intended to be "obviously safe": a call to runtime·gosched
before main.main is no different than a call to runtime.Gosched
at the beginning of main.main, and it is (or had better be)
safe to call runtime.Gosched at any point during main.
Update #3342.
R=iant
CC=golang-dev
https://golang.org/cl/5919052
Breaks closure test when GOMAXPROCS=2 or more.
««« original CL description
runtime: restore deadlock detection in the simplest case.
Fixes#3342.
R=iant, r, dave, rsc
CC=golang-dev, remy
https://golang.org/cl/5844051
»»»
R=rsc
CC=golang-dev
https://golang.org/cl/5924045
There was a small window during program initialization
where a signal could come in before the handling mechanisms
were set up to handle it. Delay the signal-handler installation
until we're ready for the signals.
Fixes#3314.
R=golang-dev, dsymonds, mikioh.mikioh
CC=golang-dev
https://golang.org/cl/5833049
This function uses 48-byte of precious non-split stack for every callback
function, and without this CL, it can easily overflow the non-split stack.
I encountered this when trying to enable misc/cgo/test on windows/amd64.
R=rsc
CC=golang-dev
https://golang.org/cl/5784075
It's the best we can do before Go 1.
For issue 3250; not a fix but at least less mysterious.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/5797068
It may have to switch stacks, since we are calling
a DLL instead of a system call.
badcallback says where it is, because it is being called
on a Windows stack already.
R=golang-dev, alex.brainman
CC=golang-dev
https://golang.org/cl/5782060
Implement runtime·write, like on the other systems,
and also runtime·badcallback, in assembly to reduce
stack footprint.
TBR=golang-dev
CC=golang-dev
https://golang.org/cl/5785055
When a very low-level system call that should never fail
does fail, we call notok, which crashes the program.
Often, we are then left with only the program counter as
information about the crash, and it is in notok.
Instead, inline calls to notok (it is just one instruction
on most systems) so that the program counter will
tell us which system call is unhappy.
R=golang-dev, gri, minux.ma, bradfitz
CC=golang-dev
https://golang.org/cl/5792048
Before:
$ go run x.go
signal 11 (core dumped)
$
After:
$ go run x.go
runtime: cgo callback on thread not created by Go.
signal 11 (core dumped)
$
For issue 3068.
Not a fix, but as much of a fix as we can do before Go 1.
R=golang-dev, rogpeppe, gri
CC=golang-dev
https://golang.org/cl/5781047
If it didn't reach the limit, we can try extending the arena
before resorting to random memory mappings and praying for the
kernel to be kind.
Fixes#3173.
R=rsc, rsc
CC=golang-dev
https://golang.org/cl/5725045
Don't try to print obviously corrupt slices or interfaces.
Doesn't actually solve 3047 or 2818, but seems a good idea anyway.
R=rsc, bsiegert
CC=golang-dev
https://golang.org/cl/5708061
Work around profiling kernel bug with signal masks.
Still broken on 64-bit Snow Leopard kernel,
but I think we can ignore that one and let people
upgrade to Lion.
Add new trivial tools addr2line and objdump to take
the place of the GNU tools of the same name, since
those are not installed on OS X.
Adapt pprof to invoke 'go tool addr2line' and
'go tool objdump' if the system tools do not exist.
Clean up disassembly of base register on amd64.
Fixes#2008.
R=golang-dev, bradfitz, mikioh.mikioh, r, iant
CC=golang-dev
https://golang.org/cl/5697066
Makes it possible for client code to maintain its own profiles,
and also reduces the API surface by giving us a type that
models built-in profiles.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/5684056
Ignore signals while we are spawning a new thread. Previously, a
signal arriving just before runtime.minit setting up the signal
handler triggers a "double fault" in signal trampolining.
Fixes#3017.
R=rsc, mikioh.mikioh, minux.ma, adg
CC=golang-dev
https://golang.org/cl/5684060
cc: add #pragma textflag to set it
runtime: mark mheap to go into noptr-bss.
remove special case in garbage collector
Remove the ARM from.flag field created by CL 5687044.
The DUPOK flag was already in p->reg, so keep using that.
Otherwise test/nilptr.go creates a very large binary.
Should fix the arm build.
Diagnosed by minux.ma; replacement for CL 5690044.
R=golang-dev, minux.ma, r
CC=golang-dev
https://golang.org/cl/5686060
A fault during malloc might lead to the program's
first call to findfunc, which would in turn call malloc.
Don't do that.
Fixes#1777.
R=golang-dev, gri
CC=golang-dev
https://golang.org/cl/5689047
morebuf holds a pc/sp from the last stack split or
reflect.call or panic/recover. If the pc is a closure,
the reference will keep it from being collected.
moreargp holds a pointer to the arguments from the
last stack split or reflect.call or panic/recover.
Normally it is a stack pointer and thus not of interest,
but in the case of reflect.call it is an allocated argument
list and holds up the arguments to the call.
R=golang-dev, r
CC=golang-dev
https://golang.org/cl/5674109
The m->cret word holds the C return value when returning
across a stack split boundary. It was not being cleared after
use, which means that the return value (if a C function)
or else the value of AX/R0 at the time of the last stack unsplit
was being kept alive longer than necessary. Clear it.
I think the effect here should be very small, but worth fixing
anyway.
R=golang-dev, bradfitz
CC=golang-dev
https://golang.org/cl/5677092
Periodically browse MHeap's freelists for long unused spans and release them if any.
Current hardcoded settings:
- GC is forced if none occured over the last 2 minutes.
- spans are handed back after 5 minutes of uselessness.
SysUnused (for Unix) is a wrapper on madvise MADV_DONTNEED on Linux and MADV_FREE on BSDs.
R=rsc, dvyukov, remyoudompheng
CC=golang-dev
https://golang.org/cl/5451057
Changeset 36c9c7810f14 broke support for grsec-patched kernels.
Those do not give back the address requested without MAP_FIXED,
so when verifying an mmap without this flag for success, the
resulting address must not be compared against the requested
address since it may have succeeded at a different location.
R=golang-dev, rsc, gustavo, iant
CC=golang-dev
https://golang.org/cl/5650072
before after
go test -short time 4.144s 1.215s
go test -short runtime 1.315s 0.351s
go test -short -cpu=1,2,4 runtime 4.376s 1.266s
Partially solves issue 3015.
R=golang-dev, r
CC=golang-dev, remy
https://golang.org/cl/5673045
It is possible that Linux and Windows copy the FP control word
from the parent thread when creating a new thread. Empirically,
Darwin does not. Reset the FP control world in all cases.
Enable the floating-point strconv test.
Fixes#2917 (again).
R=golang-dev, r, iant
CC=golang-dev
https://golang.org/cl/5660047
