Introduction to Go 1.3

The latest Go release, version 1.3, arrives six months after 1.2, and contains no language changes. It focuses primarily on implementation work, providing precise garbage collection, a major refactoring of the compiler tool chain that results in faster builds, especially for large projects, significant performance improvements across the board, and support for DragonFly BSD, Solaris, Plan 9 and Google's Native Client architecture (NaCl). It also has an important refinement to the memory model regarding synchronization. As always, Go 1.3 keeps the promise of compatibility, and almost everything will continue to compile and run without change when moved to 1.3.

Changes to the supported operating systems and architectures

Removal of support for Windows 2000

Microsoft stopped supporting Windows 2000 in 2010. Since it has implementation difficulties regarding exception handling (signals in Unix terminology), as of Go 1.3 it is not supported by Go either.

Support for DragonFly BSD

Go 1.3 now includes experimental support for DragonFly BSD on the amd64 (64-bit x86) and 386 (32-bit x86) architectures. It uses DragonFly BSD 3.6 or above.

Support for FreeBSD

It was not announced at the time, but since the release of Go 1.2, support for Go on FreeBSD requires FreeBSD 8 or above.

As of Go 1.3, support for Go on FreeBSD requires that the kernel be compiled with the COMPAT_FREEBSD32 flag configured.

In concert with the switch to EABI syscalls for ARM platforms, Go 1.3 will run only on FreeBSD 10. The x86 platforms, 386 and amd64, are unaffected.

Support for Native Client

Support for the Native Client virtual machine architecture has returned to Go with the 1.3 release. It runs on the 32-bit Intel architectures (GOARCH=386) and also on 64-bit Intel, but using 32-bit pointers (GOARCH=amd64p32). There is not yet support for Native Client on ARM. Note that this is Native Client (NaCl), not Portable Native Client (PNaCl). Details about Native Client are here; how to set up the Go version is described here.

Support for NetBSD

As of Go 1.3, support for Go on NetBSD requires NetBSD 6.0 or above.

Support for OpenBSD

As of Go 1.3, support for Go on OpenBSD requires OpenBSD 5.5 or above.

Support for Plan 9

Go 1.3 now includes experimental support for Plan 9 on the 386 (32-bit x86) architecture. It requires the Tsemacquire syscall, which has been in Plan 9 since June, 2012.

Support for Solaris

Go 1.3 now includes experimental support for Solaris on the amd64 (64-bit x86) architecture. It requires illumos, Solaris 11 or above.

Changes to the memory model

The Go 1.3 memory model adds a new rule concerning sending and receiving on buffered channels, to make explicit that a buffered channel can be used as a simple semaphore, using a send into the channel to acquire and a receive from the channel to release. This is not a language change, just a clarification about an expected property of communication.

Changes to the implementations and tools

Stack

Go 1.3 has changed the implementation of goroutine stacks away from the old, "segmented" model to a contiguous model. When a goroutine needs more stack than is available, its stack is transferred to a larger single block of memory. The overhead of this transfer operation amortizes well and eliminates the old "hot spot" problem when a calculation repeatedly steps across a segment boundary. Details including performance numbers are in this design document.

Changes to the garbage collector

For a while now, the garbage collector has been precise when examining values in the heap; the Go 1.3 release adds equivalent precision to values on the stack. This means that a non-pointer Go value such as an integer will never be mistaken for a pointer and prevent unused memory from being reclaimed.

Starting with Go 1.3, the runtime assumes that values with pointer type contain pointers and other values do not. This assumption is fundamental to the precise behavior of both stack expansion and garbage collection. Programs that use package unsafe to store uintptrs in pointer values are illegal and will crash if the runtime detects the behavior. Programs that use package unsafe to store pointers in uintptr values are also illegal but more difficult to diagnose during execution. Because the pointers are hidden from the runtime, a stack expansion or garbage collection may reclaim the memory they point at, creating dangling pointers.

Updating: Code that converts a uintptr value stored in memory to unsafe.Pointer is illegal and must be rewritten. Such code can be identified by go vet.

Map iteration

Iterations over small maps no longer happen in a consistent order. Go 1 defines that “The iteration order over maps is not specified and is not guaranteed to be the same from one iteration to the next.” To keep code from depending on map iteration order, Go 1.0 started each map iteration at a random index in the map. A new map implementation introduced in Go 1.1 neglected to randomize iteration for maps with eight or fewer entries, although the iteration order can still vary from system to system. This has allowed people to write Go 1.1 and Go 1.2 programs that depend on small map iteration order and therefore only work reliably on certain systems. Go 1.3 reintroduces random iteration for small maps in order to flush out these bugs.

Updating: If code assumes a fixed iteration order for small maps, it will break and must be rewritten not to make that assumption. Because only small maps are affected, the problem arises most often in tests.

As part of the general overhaul to the Go linker, the compilers and linkers have been refactored. The linker is still a C program, but now the instruction selection phase that was part of the linker has been moved to the compiler through the creation of a new library called liblink. By doing instruction selection only once, when the package is first compiled, this can speed up compilation of large projects significantly.

Updating: Although this is a major internal change, it should have no effect on programs.

Status of gccgo

GCC release 4.9 will contain the Go 1.2 (not 1.3) version of gccgo. The release schedules for the GCC and Go projects do not coincide, which means that 1.3 will be available in the development branch but that the next GCC release, 4.10, will likely have the Go 1.4 version of gccgo.

Changes to the go command

The cmd/go command has several new features. The go run and go test subcommands support a new -exec option to specify an alternate way to run the resulting binary. Its immediate purpose is to support NaCl.

The test coverage support of the go test subcommand now automatically sets the coverage mode to -atomic when the race detector is enabled, to eliminate false reports about unsafe access to coverage counters.

The go test subcommand now always builds the package, even if it has no test files. Previously, it would do nothing if no test files were present.

The go build subcommand supports a new -i option to install dependencies of the specified target, but not the target itself.

Cross compiling with cgo enabled is now supported. The CC_FOR_TARGET and CXX_FOR_TARGET environment variables are used when running all.bash to specify the cross compilers for C and C++ code, respectively.

Finally, the go command now supports packages that import Objective-C files (suffixed .m) through cgo.

Changes to cgo

The cmd/cgo command, which processes import "C" declarations in Go packages, has corrected a serious bug that may cause some packages to stop compiling. Previously, all pointers to incomplete struct types translated to the Go type *[0]byte, with the effect that the Go compiler could not diagnose passing one kind of struct pointer to a function expecting another. Go 1.3 corrects this mistake by translating each different incomplete struct to a different named type. However, some Go code took advantage of this bug to pass (for example) a *C.FILE from one package to another. This is not legal and no longer works: in general Go packages should avoid exposing C types and names in their APIs.

Updating: Code confusing pointers to incomplete types or passing them across package boundaries will no longer compile and must be rewritten. If the conversion is correct and must be preserved, use an explicit conversion via unsafe.Pointer.

SWIG 3.0 required for programs that use SWIG

For Go programs that use SWIG, SWIG version 3.0 is now required. The cmd/go command will now link the SWIG generated object files directly into the binary, rather than building and linking with a shared library.

Command-line flag parsing

In the gc tool chain, the assemblers now use the same command-line flag parsing rules as the Go flag package, a departure from the traditional Unix flag parsing. This may affect scripts that invoke the tool directly. For example, go tool 6a -SDfoo must now be written go tool 6a -S -D foo. (The same change was made to the compilers and linkers in Go 1.1.)

Miscellany

The program misc/benchcmp that compares performance across benchmarking runs has been rewritten. Once a shell and awk script in the main repository, it is now a Go program in the go.tools repo. Documentation is here.

For the few of us that build Go distributions, the tool misc/dist has been moved and renamed; it now lives in misc/makerelease, still in the main repository.

Performance

The performance of Go binaries for this release has improved in many cases due to changes in the runtime and garbage collection, plus some changes to libraries. Significant instances include:

Also, the runtime now includes in stack dumps how long a goroutine has been blocked, which can be useful information when debugging deadlocks or performance issues.

Changes to the standard library

New packages

A new package debug/plan9obj was added to the standard library. It implements access to Plan 9 a.out object files.

Major changes to the library

A previous bug in crypto/tls made it possible to skip verification in TLS inadvertently. In Go 1.3, the bug is fixed: one must specify either ServerName or InsecureSkipVerify, and if ServerName is specified it is enforced. This may break existing code that incorrectly depended on insecure behavior.

There is an important new type added to the standard library: sync.Pool. It provides an efficient mechanism for implementing certain types of caches whose memory can be reclaimed automatically by the system.

The testing package's benchmarking helper, B, now has a RunParallel method to make it easier to run benchmarks that exercise multiple CPUs.

Updating: The crypto/tls fix may break existing code, but such code was erroneous and should be updated.

Minor changes to the library

The following list summarizes a number of minor changes to the library, mostly additions. See the relevant package documentation for more information about each change.