06 Go 1.6 Release Notes

Go 1.6 Release Notes

Introduction to Go 1.6 Changes to the language Ports Tools Cgo Compiler Toolchain Gccgo Go command Go doc command Go vet command

Performance Core library HTTP/2 Runtime Reflect Sorting Templates Minor changes to the library

Introduction to Go 1.6

The latest Go release, version 1.6, arrives six months after 1.5. Most of its changes are in the implementation of the language, runtime, and libraries. There are no changes to the language specification. As always, the release maintains the Go 1 promise of compatibility. We expect almost all Go programs to continue to compile and run as before.

The release adds new ports to Linux on 64-bit MIPS and Android on 32-bit x86; defined and enforced rules for sharing Go pointers with C; transparent, automatic support for HTTP/2; and a new mechanism for template reuse.

Changes to the language

There are no language changes in this release.

Ports

Go 1.6 adds experimental ports to Linux on 64-bit MIPS ( linux/mips64 and linux/mips64le ). These ports support cgo but only with internal linking.

Go 1.6 also adds an experimental port to Android on 32-bit x86 ( android/386 ).

On FreeBSD, Go 1.6 defaults to using clang , not gcc , as the external C compiler.

On Linux on little-endian 64-bit PowerPC ( linux/ppc64le ), Go 1.6 now supports cgo with external linking and is roughly feature complete.

On NaCl, Go 1.5 required SDK version pepper-41. Go 1.6 adds support for later SDK versions.

On 32-bit x86 systems using the -dynlink or -shared compilation modes, the register CX is now overwritten by certain memory references and should be avoided in hand-written assembly. See the assembly documentationfor details.

Tools

Cgo

There is one major change to cgo , along with one minor change.

The major change is the definition of rules for sharing Go pointers with C code, to ensure that such C code can coexist with Go's garbage collector. Briefly, Go and C may share memory allocated by Go when a pointer to that memory is passed to C as part of a cgo call, provided that the memory itself contains no pointers to Go-allocated memory, and provided that C does not retain the pointer after the call returns. These rules are checked by the runtime during program execution: if the runtime detects a violation, it prints a diagnosis and crashes the program. The checks can be disabled by setting the environment variable GODEBUG=cgocheck=0 , but note that the vast majority of code identified by the checks is subtly incompatible with garbage collection in one way or another. Disabling the checks will typically only lead to more mysterious failure modes. Fixing the code in question should be strongly preferred over turning off the checks. See the cgo documentation for more details.

The minor change is the addition of explicit C.complexfloat and C.complexdouble types, separate from Go's complex64 and complex128 . Matching the other numeric types, C's complex types and Go's complex type are no longer interchangeable.

Compiler Toolchain

The compiler toolchain is mostly unchanged. Internally, the most significant change is that the parser is now hand-written instead of generated from yacc.

The compiler, linker, and go command have a new flag -msan , analogous to -race and only available on linux/amd64, that enables interoperation with the Clang MemorySanitizer. Such interoperation is useful mainly for testing a program containing suspect C or C++ code.

The linker has a new option -libgcc to set the expected location of the C compiler support library when linking cgo code. The option is only consulted when using -linkmode=internal , and it may be set to none to disable the use of a support library.

The implementation of build modes started in Go 1.5 has been expanded to more systems. This release adds support for the c-shared mode on android/386 , android/amd64 , android/arm64 , linux/386 , and linux/arm64 ; for the shared mode on linux/386 , linux/arm , linux/amd64 , and linux/ppc64le ; and for the new pie mode (generating position-independent executables) on android/386 , android/amd64 , android/arm , android/arm64 , linux/386 , linux/amd64 , linux/arm , linux/arm64 , and linux/ppc64le . See the design document for details.

As a reminder, the linker's -X flag changed in Go 1.5. In Go 1.4 and earlier, it took two arguments, as in

-X importpath.name value

Go 1.5 added an alternative syntax using a single argument that is itself a name=value pair:

-X importpath.name=value

In Go 1.5 the old syntax was still accepted, after printing a warning suggesting use of the new syntax instead. Go 1.6 continues to accept the old syntax and print the warning. Go 1.7 will remove support for the old syntax.

Gccgo

The release schedules for the GCC and Go projects do not coincide. GCC release 5 contains the Go 1.4 version of gccgo. The next release, GCC 6, will have the Go 1.6.1 version of gccgo.

Go command

The go command's basic operation is unchanged, but there are a number of changes worth noting.

Go 1.5 introduced experimental support for vendoring, enabled by setting the GO15VENDOREXPERIMENT environment variable to 1 . Go 1.6 keeps the vendoring support, no longer considered experimental, and enables it by default. It can be disabled explicitly by setting the GO15VENDOREXPERIMENT environment variable to 0 . Go 1.7 will remove support for the environment variable.

The most likely problem caused by enabling vendoring by default happens in source trees containing an existing directory named vendor that does not expect to be interpreted according to new vendoring semantics. In this case, the simplest fix is to rename the directory to anything other than vendor and update any affected import paths.

For details about vendoring, see the documentation for the go command and the design document.

There is a new build flag, -msan , that compiles Go with support for the LLVM memory sanitizer. This is intended mainly for use when linking against C or C++ code that is being checked with the memory sanitizer.

Go doc command

Go 1.5 introduced the go doc command, which allows references to packages using only the package name, as in go doc http . In the event of ambiguity, the Go 1.5 behavior was to use the package with the lexicographically earliest import path. In Go 1.6, ambiguity is resolved by preferring import paths with fewer elements, breaking ties using lexicographic comparison. An important effect of this change is that original copies of packages are now preferred over vendored copies. Successful searches also tend to run faster.

Go vet command

The go vet command now diagnoses passing function or method values as arguments to Printf , such as when passing f where f() was intended.

Performance

As always, the changes are so general and varied that precise statements about performance are difficult to make. Some programs may run faster, some slower. On average the programs in the Go 1 benchmark suite run a few percent faster in Go 1.6 than they did in Go 1.5. The garbage collector's pauses are even lower than in Go 1.5, especially for programs using a large amount of memory.

There have been significant optimizations bringing more than 10% improvements to implementations of the compress/bzip2 , compress/gzip , crypto/aes , crypto/elliptic , crypto/ecdsa , and sort packages.

Core library

HTTP/2

Go 1.6 adds transparent support in the net/http package for the new HTTP/2 protocol. Go clients and servers will automatically use HTTP/2 as appropriate when using HTTPS. There is no exported API specific to details of the HTTP/2 protocol handling, just as there is no exported API specific to HTTP/1.1.

Programs that must disable HTTP/2 can do so by setting Transport.TLSNextProto (for clients) or Server.TLSNextProto (for servers) to a non-nil, empty map.

Programs that must adjust HTTP/2 protocol-specific details can import and use golang.org/x/net/http2 , in particular its ConfigureServer and ConfigureTransport functions.

Runtime

The runtime has added lightweight, best-effort detection of concurrent misuse of maps. As always, if one goroutine is writing to a map, no other goroutine should be reading or writing the map concurrently. If the runtime detects this condition, it prints a diagnosis and crashes the program. The best way to find out more about the problem is to run the program under the race detector, which will more reliably identify the race and give more detail.

For program-ending panics, the runtime now by default prints only the stack of the running goroutine, not all existing goroutines. Usually only the current goroutine is relevant to a panic, so omitting the others significantly reduces irrelevant output in a crash message. To see the stacks from all goroutines in crash messages, set the environment variable GOTRACEBACK to all or call debug.SetTraceback before the crash, and rerun the program. See the runtime documentation for details.

Updating : Uncaught panics intended to dump the state of the entire program, such as when a timeout is detected or when explicitly handling a received signal, should now call debug.SetTraceback("all") before panicking. Searching for uses of signal.Notify may help identify such code.

On Windows, Go programs in Go 1.5 and earlier forced the global Windows timer resolution to 1ms at startup by calling timeBeginPeriod(1) . Go no longer needs this for good scheduler performance, and changing the global timer resolution caused problems on some systems, so the call has been removed.

When using -buildmode=c-archive or -buildmode=c-shared to build an archive or a shared library, the handling of signals has changed. In Go 1.5 the archive or shared library would install a signal handler for most signals. In Go 1.6 it will only install a signal handler for the synchronous signals needed to handle run-time panics in Go code: SIGBUS, SIGFPE, SIGSEGV. See the os/signal package for more details.

Reflect

The reflect package has resolved a long-standing incompatibility between the gc and gccgo toolchains regarding embedded unexported struct types containing exported fields. Code that walks data structures using reflection, especially to implement serialization in the spirit of the encoding/json and encoding/xml packages, may need to be updated.

The problem arises when using reflection to walk through an embedded unexported struct-typed field into an exported field of that struct. In this case, reflect had incorrectly reported the embedded field as exported, by returning an empty Field.PkgPath . Now it correctly reports the field as unexported but ignores that fact when evaluating access to exported fields contained within the struct.

Updating : Typically, code that previously walked over structs and used

f.PkgPath != ""

to exclude inaccessible fields should now use

f.PkgPath != "" && !f.Anonymous

For example, see the changes to the implementations of encoding/json and encoding/xml .

Sorting

In the sort package, the implementation of Sort has been rewritten to make about 10% fewer calls to the Interface 's Less and Swap methods, with a corresponding overall time savings. The new algorithm does choose a different ordering than before for values that compare equal (those pairs for which Less(i, j) and Less(j, i) are false).

Updating : The definition of Sort makes no guarantee about the final order of equal values, but the new behavior may still break programs that expect a specific order. Such programs should either refine their Less implementations to report the desired order or should switch to Stable , which preserves the original input order of equal values.

Templates

In the text/template package, there are two significant new features to make writing templates easier.

First, it is now possible to trim spaces around template actions, which can make template definitions more readable. A minus sign at the beginning of an action says to trim space before the action, and a minus sign at the end of an action says to trim space after the action. For example, the template

{{23 -}}

   <

{{- 45}}

formats as 23<45 .

Second, the new {{block}} action, combined with allowing redefinition of named templates, provides a simple way to define pieces of a template that can be replaced in different instantiations. There is an example in the text/template package that demonstrates this new feature.

Minor changes to the library

• The archive/tar package's implementation corrects many bugs in rare corner cases of the file format. One visible change is that the Reader type's Read method now presents the content of special file types as being empty, returning io.EOF immediately.
• In the archive/zip package, the Reader type now has a RegisterDecompressor method, and the Writer type now has a RegisterCompressor method, enabling control over compression options for individual zip files. These take precedence over the pre-existing global RegisterDecompressor and RegisterCompressor functions.
• The bufio package's Scanner type now has a Buffer method, to specify an initial buffer and maximum buffer size to use during scanning. This makes it possible, when needed, to scan tokens larger than MaxScanTokenSize . Also for the Scanner , the package now defines the ErrFinalToken error value, for use bysplit functions to abort processing or to return a final empty token.
• The compress/flate package has deprecated its ReadError and WriteError error implementations. In Go 1.5 they were only rarely returned when an error was encountered; now they are never returned, although they remain defined for compatibility.
• The compress/flate , compress/gzip , and compress/zlib packages now report io.ErrUnexpectedEOF for truncated input streams, instead of io.EOF .
• The crypto/cipher package now overwrites the destination buffer in the event of a GCM decryption failure. This is to allow the AESNI code to avoid using a temporary buffer.
• The crypto/tls package has a variety of minor changes. It now allows Listen to succeed when the Config has a nil Certificates , as long as the GetCertificate callback is set, it adds support for RSA with AES-GCM cipher suites, and it adds a RecordHeaderError to allow clients (in particular, the net/http package) to report a better error when attempting a TLS connection to a non-TLS server.
• The crypto/x509 package now permits certificates to contain negative serial numbers (technically an error, but unfortunately common in practice), and it defines a new InsecureAlgorithmError to give a better error message when rejecting a certificate signed with an insecure algorithm like MD5.
• The debug/dwarf and debug/elf packages together add support for compressed DWARF sections. User code needs no updating: the sections are decompressed automatically when read.
• The debug/elf package adds support for general compressed ELF sections. User code needs no updating: the sections are decompressed automatically when read. However, compressed Sections do not support random access: they have a nil ReaderAt field.
• The encoding/asn1 package now exports tag and class constants useful for advanced parsing of ASN.1 structures.
• Also in the encoding/asn1 package, Unmarshal now rejects various non-standard integer and length encodings.
• The encoding/base64 package's Decoder has been fixed to process the final bytes of its input. Previously it processed as many four-byte tokens as possible but ignored the remainder, up to three bytes. The Decoder therefore now handles inputs in unpadded encodings (like RawURLEncoding) correctly, but it also rejects inputs in padded encodings that are truncated or end with invalid bytes, such as trailing spaces.
• The encoding/json package now checks the syntax of a Number before marshaling it, requiring that it conforms to the JSON specification for numeric values. As in previous releases, the zero Number (an empty string) is marshaled as a literal 0 (zero).
• The encoding/xml package's Marshal function now supports a cdata attribute, such as chardata but encoding its argument in one or more <![CDATA[ ... ]]> tags.
• Also in the encoding/xml package, Decoder 's Token method now reports an error when encountering EOF before seeing all open tags closed, consistent with its general requirement that tags in the input be properly matched. To avoid that requirement, use RawToken .
• The fmt package now allows any integer type as an argument to Printf 's * width and precision specification. In previous releases, the argument to * was required to have type int .
• Also in the fmt package, Scanf can now scan hexadecimal strings using %X, as an alias for %x. Both formats accept any mix of upper- and lower-case hexadecimal.
• The image and image/color packages add NYCbCrA and NYCbCrA types, to support Y'CbCr images with non-premultiplied alpha.
• The io package's MultiWriter implementation now implements a WriteString method, for use by WriteString .
• In the math/big package, Int adds Append and Text methods to give more control over printing.
• Also in the math/big package, Float now implements encoding.TextMarshaler and encoding.TextUnmarshaler , allowing it to be serialized in a natural form by the encoding/json and encoding/xml packages.
• Also in the math/big package, Float 's Append method now supports the special precision argument -1. As in strconv.ParseFloat , precision -1 means to use the smallest number of digits necessary such that Parse reading the result into a Float of the same precision will yield the original value.
• The math/rand package adds a Read function, and likewise Rand adds a Read method. These make it easier to generate pseudorandom test data. Note that, like the rest of the package, these should not be used in cryptographic settings; for such purposes, use the crypto/rand package instead.
• The net package's ParseMAC function now accepts 20-byte IP-over-InfiniBand (IPoIB) link-layer addresses.
• Also in the net package, there have been a few changes to DNS lookups. First, the DNSError error implementation now implements Error , and in particular its new IsTemporary method returns true for DNS server errors. Second, DNS lookup functions such as LookupAddr now return rooted domain names (with a trailing dot) on Plan 9 and Windows, to match the behavior of Go on Unix systems.
• The net/http package has a number of minor additions beyond the HTTP/2 support already discussed. First, the FileServer now sorts its generated directory listings by file name. Second, the ServeFile function now refuses to serve a result if the request's URL path contains “..” (dot-dot) as a path element. Programs should typically use FileServer and Dir instead of calling ServeFile directly. Programs that need to serve file content in response to requests for URLs containing dot-dot can still call ServeContent . Third, the Client now allows user code to set the Expect: 100-continue header (see Transport.ExpectContinueTimeout ). Fourth, there arefive new error codes: StatusPreconditionRequired (428), StatusTooManyRequests (429), StatusRequestHeaderFieldsTooLarge (431), and StatusNetworkAuthenticationRequired (511) from RFC 6585, as well as the recently-approved StatusUnavailableForLegalReasons (451). Fifth, the implementation and documentation of CloseNotifier has been substantially changed. The Hijacker interface now works correctly on connections that have previously been used with CloseNotifier . The documentation now describes when CloseNotifier is expected to work.
• Also in the net/http package, there are a few changes related to the handling of a Request data structure with its Method field set to the empty string. An empty Method field has always been documented as an alias for "GET" and it remains so. However, Go 1.6 fixes a few routines that did not treat an empty Method the same as an explicit "GET" . Most notably, in previous releases Client followed redirects only with Method set explicitly to "GET" ; in Go 1.6 Client also follows redirects for the empty Method . Finally, NewRequest accepts a method argument that has not been documented as allowed to be empty. In past releases, passing an empty method argument resulted in a Request with an empty Method field. In Go 1.6, the resulting Request always has an initialized Method field: if its argument is an empty string, NewRequest sets the Method field in the returned Request to "GET" .
• The net/http/httptest package's ResponseRecorder now initializes a default Content-Type header using the same content-sniffing algorithm as in http.Server .
• The net/url package's Parse is now stricter and more spec-compliant regarding the parsing of host names. For example, spaces in the host name are no longer accepted.
• Also in the net/url package, the Error type now implements net.Error .
• The os package's IsExist , IsNotExist , and IsPermission now return correct results when inquiring about an SyscallError .
• On Unix-like systems, when a write to os.Stdout or os.Stderr (more precisely, an os.File opened for file descriptor 1 or 2) fails due to a broken pipe error, the program will raise a SIGPIPE signal. By default this will cause the program to exit; this may be changed by calling the os/signal Notify function for syscall.SIGPIPE . A write to a broken pipe on a file descriptor other 1 or 2 will simply return syscall.EPIPE (possibly wrapped in os.PathError and/or os.SyscallError ) to the caller. The old behavior of raising an uncatchable SIGPIPE signal after 10 consecutive writes to a broken pipe no longer occurs.
• In the os/exec package, Cmd 's Output method continues to return an ExitError when a command exits with an unsuccessful status. If standard error would otherwise have been discarded, the returned ExitError now holds a prefix and suffix (currently 32 kB) of the failed command's standard error output, for debugging or for inclusion in error messages. The ExitError 's String method does not show the captured standard error; programs must retrieve it from the data structure separately.
• On Windows, the path/filepath package's Join function now correctly handles the case when the base is a relative drive path. For example, Join(`c:`, `a`) now returns `c:a` instead of `c:\a` as in past releases. This may affect code that expects the incorrect result.
• In the regexp package, the Regexp type has always been safe for use by concurrent goroutines. It uses a sync.Mutex to protect a cache of scratch spaces used during regular expression searches. Some high-concurrency servers using the same Regexp from many goroutines have seen degraded performance due to contention on that mutex. To help such servers, Regexp now has a Copy method, which makes a copy of a Regexp that shares most of the structure of the original but has its own scratch space cache. Two goroutines can use different copies of a Regexp without mutex contention. A copy does have additional space overhead, so Copy should only be used when contention has been observed.
• The strconv package adds IsGraphic , similar to IsPrint . It also adds QuoteToGraphic , QuoteRuneToGraphic , AppendQuoteToGraphic , and AppendQuoteRuneToGraphic , analogous to QuoteToASCII , QuoteRuneToASCII , and so on. The ASCII family escapes all space characters except ASCII space (U+0020). In contrast, the Graphic family does not escape any Unicode space characters (category Zs).
• In the testing package, when a test calls t.Parallel, that test is paused until all non-parallel tests complete, and then that test continues execution with all other parallel tests. Go 1.6 changes the time reported for such a test: previously the time counted only the parallel execution, but now it also counts the time from the start of testing until the call to t.Parallel .
• The text/template package contains two minor changes, in addition to the major changes described above. First, it adds a new ExecError type returned for any error during Execute that does not originate in a Write to the underlying writer. Callers can distinguish template usage errors from I/O errors by checking for ExecError . Second, the Funcs method now checks that the names used as keys in the FuncMap are identifiers that can appear in a template function invocation. If not, Funcs panics.
• The time package's Parse function has always rejected any day of month larger than 31, such as January 32. In Go 1.6, Parse now also rejects February 29 in non-leap years, February 30, February 31, April 31, June 31, September 31, and November 31.