Akshat Kumar [Fri, 18 Jan 2013 21:43:25 +0000 (16:43 -0500)]
syscall, os: fix a fork-exec/wait race in Plan 9.
On Plan 9, only the parent of a given process can enter its wait
queue. When a Go program tries to fork-exec a child process
and subsequently waits for it to finish, the goroutines doing
these two tasks do not necessarily tie themselves to the same
(or any single) OS thread. In the case that the fork and the wait
system calls happen on different OS threads (say, due to a
goroutine being rescheduled somewhere along the way), the
wait() will either return an error or end up waiting for a
completely different child than was intended.
This change forces the fork and wait syscalls to happen in the
same goroutine and ties that goroutine to its OS thread until
the child exits. The PID of the child is recorded upon fork and
exit, and de-queued once the child's wait message has been read.
The Wait API, then, is translated into a synthetic implementation
that simply waits for the requested PID to show up in the queue
and then reads the associated stats.
Rémy Oudompheng [Fri, 18 Jan 2013 21:40:32 +0000 (22:40 +0100)]
cmd/gc: fix handling of struct padding in hash/eq.
The test case of issue 4585 was not passing due to
miscalculation of memequal args, and the previous fix
does not handle padding at the end of a struct.
Handling of padding at end of structs also fixes the case
of [n]T where T is such a padded struct.
Russ Cox [Fri, 18 Jan 2013 20:10:36 +0000 (15:10 -0500)]
build: change GO386=sse to GO386=sse2
sse2 is a more precise description of the requirement,
and it matches what people will see in, for example
grep sse2 /proc/cpuinfo # linux
sysctl hw.optional.sse2 # os x
Raph Levien [Fri, 18 Jan 2013 20:09:42 +0000 (15:09 -0500)]
compress/flate: Performance improvement for inflate
Decode as much as possible of a Huffman symbol in a single table
lookup (much like the zlib implementation), filling more bits
(conservatively, so we don't consume past the end of the stream)
when the code prefix indicates more bits are needed. This
results in about a 50% performance gain in speed benchmarks.
The following set is benchcmp done on a retina MacBook Pro:
Andrew Balholm [Thu, 17 Jan 2013 01:06:04 +0000 (12:06 +1100)]
exp/html: remove "INCOMPLETE" comment
I think that the parser is complete enough to take that warning out.
It passes the test suite.
There may be incompatible API changes, but being in the exp directory
is warning enough for that.
Robert Griesemer [Mon, 14 Jan 2013 23:19:32 +0000 (15:19 -0800)]
go/types: various minor fixes
- always set the Pkg field in QualifiedIdents
- call Context.Ident for all identifiers in the AST that denote
a types.Object (bug fix)
- added test that Context.Ident is called for all such identifiers
Andrew Gerrand [Sun, 13 Jan 2013 22:35:04 +0000 (09:35 +1100)]
cmd/godoc: support m=text parameter for text files
It's possible to view the package docs in plain text, eg:
http://golang.org/pkg/time/?m=text
and this CL introduces the ability to do the same for files:
http://golang.org/src/pkg/time/time.go?m=text
Robert Griesemer [Sun, 13 Jan 2013 18:33:08 +0000 (10:33 -0800)]
go/types: Moving from *ast.Objects to types.Objects (step 2).
Completely removed *ast.Objects from being exposed by the
types API. *ast.Objects are still required internally for
resolution, but now the door is open for an internal-only
rewrite of identifier resolution entirely at type-check
time. Once that is done, ASTs can be type-checked whether
they have been created via the go/parser or otherwise,
and type-checking does not require *ast.Object or scope
invariants to be maintained externally.
Dave Cheney [Sat, 12 Jan 2013 06:52:52 +0000 (17:52 +1100)]
test: limit runoutput tests on arm platforms
runoutput styles tests generally consume a lot of memory. On arm platforms rotate?.go consume around 200mb each to compile, and as tests are sorted alphabetically, they all tend to run at once.
This change limits the number of runoutput jobs to 2 on arm platforms.
Robert Griesemer [Fri, 11 Jan 2013 21:53:38 +0000 (13:53 -0800)]
go/types: Moving from *ast.Objects to types.Objects (step 1).
The existing type checker was relying on augmenting ast.Object
fields (empty interfaces) for its purposes. While this worked
for some time now, it has become increasingly brittle. Also,
the need for package information for Fields and Methods would
have required a new field in each ast.Object. Rather than making
them bigger and the code even more subtle, in this CL we are moving
away from ast.Objects.
The types packge now defines its own objects for different
language entities (Const, Var, TypeName, Func), and they
implement the types.Object interface. Imported packages
create a Package object which holds the exported entities
in a types.Scope of types.Objects.
For type-checking, the current package is still using ast.Objects
to make this transition manageable. In a next step, the type-
checker will also use types.Objects instead, which opens the door
door to resolving ASTs entirely by the type checker. As a result,
the AST and type checker become less entangled, and ASTs can be
manipulated "by hand" or programmatically w/o having to worry
about scope and object invariants that are very hard to maintain.
(As a consequence, a future parser can do less work, and a
future AST will not need to define objects and scopes anymore.
Also, object resolution which is now split across the parser,
the ast, (ast.NewPackage), and even the type checker (for composite
literal keys) can be done in a single place which will be simpler
and more efficient.)
Change details:
- Check now takes a []*ast.File instead of a map[string]*ast.File.
It's easier to handle (I deleted code at all use sites) and does
not suffer from undefined order (which is a pain for testing).
- ast.Object.Data is now a *types.Package rather then an *ast.Scope
if the object is a package (obj.Kind == ast.Pkg). Eventually this
will go away altogether.
- Instead of an ast.Importer, Check now uses a types.Importer
(which returns a *types.Package).
- types.NamedType has two object fields (Obj Object and obj *ast.Object);
eventually there will be only Obj. The *ast.Object is needed during
this transition since a NamedType may refer to either an imported
(using types.Object) or locally defined (using *ast.Object) type.
- ast.NewPackage is not used anymore - there's a local copy for
package-level resolution of imports.
- struct fields now take the package origin into account.
- The GcImporter is now returning a *types.Package. It cannot be
used with ast.NewPackage anymore. If that functionality is still
used, a copy of the old GcImporter should be made locally (note
that GcImporter was part of exp/types and it's API was not frozen).
- dot-imports are not handled for the time being (this will come back).
Brad Fitzpatrick [Fri, 11 Jan 2013 18:03:43 +0000 (10:03 -0800)]
net/http: buffer before chunking
This introduces a buffer between writing from a handler and
writing chunks. Further, it delays writing the header until
the first full chunk is ready. In the case where the first
full chunk is also the final chunk (for small responses), that
means we can also compute a Content-Length, which is a nice
side effect for certain benchmarks.
Dmitriy Vyukov [Thu, 10 Jan 2013 05:57:06 +0000 (09:57 +0400)]
runtime: less aggressive per-thread stack segment caching
Introduce global stack segment cache and limit per-thread cache size.
This greatly reduces StackSys memory on workloads that create lots of threads.
Anthony Martin [Wed, 9 Jan 2013 23:05:22 +0000 (15:05 -0800)]
cmd/ld: fix incompatible type signatures on Plan 9
Changeset f483bfe81114 moved ELF generation to the architecture
independent code and in doing so added a Section* to the Sym
type and an Elf64_Shdr* to the Section type.
This caused the Plan 9 compilers to complain about incompatible
type signatures in the many files that reference the Sym type.
R=rsc, dave
CC=golang-dev
https://golang.org/cl/7057058
Dave Cheney [Wed, 9 Jan 2013 22:57:01 +0000 (09:57 +1100)]
cmd/go: remove $GOROOT as a go get target
Fixes #4186.
Back in the day, before the Go 1.0 release, $GOROOT was mandatory for building from source. Fast forward to now and $GOPATH is mandatory and $GOROOT is optional, and mainly used by those who use the binary distribution in uncommon places.
For example, most novices at least know about `sudo` as they would have used it to install the binary tarball into /usr/local. It is logical they would use the `sudo` hammer to `go get` other Go packages when faced with a permission error talking about the path they just had to use `sudo` on last time.
Even if they had read the documentation and set $GOPATH, go get will not work as expected as `sudo` masks most environment variables.
Rémy Oudompheng [Wed, 9 Jan 2013 21:02:53 +0000 (22:02 +0100)]
cmd/gc: add space to export data to match linker expectations
The linker split PKGDEF into (prefix, name, def) pairs,
and defines def to begin after a space following the identifier.
This is totally wrong for the following export data:
func "".FunctionName()
var SomethingCompletelyUnrelated int
The linker would parse
name=`"".FunctionName()\n\tvar`
def=`SomethingCompletelyUnrelated int`
since there is no space after FunctionName.
Dave Cheney [Wed, 9 Jan 2013 21:00:03 +0000 (08:00 +1100)]
cmd/dist: drop unneeded clang flags
Our source no longer needs these flags set to build cleanly using clang.
Tested with
* Ubuntu clang version 3.0-6ubuntu3 (tags/RELEASE_30/final) (based on LLVM 3.0) on i386
* clang version 3.2 (tags/RELEASE_32/final) on amd64 cross compiling all platforms