Russ Cox [Wed, 29 Jun 2011 13:52:34 +0000 (09:52 -0400)]
reflect: support for struct tag use by multiple packages
Each package using struct field tags assumes that
it is the only package storing data in the tag.
This CL adds support in package reflect for sharing
tags between multiple packages. In this scheme, the
tags must be of the form
key:"value" key2:"value2"
(raw strings help when writing that tag in Go source).
reflect.StructField's Tag field now has type StructTag
(a string type), which has method Get(key string) string
that returns the associated value.
Clients of json and xml will need to be updated.
Code that says
type T struct {
X int "name"
}
should become
type T struct {
X int `json:"name"` // or `xml:"name"`
}
Use govet to identify struct tags that need to be changed
to use the new syntax.
Russ Cox [Wed, 29 Jun 2011 03:58:35 +0000 (23:58 -0400)]
gc: fix package quoting logic
The decision for when to say "hash/crc32".New instead of
crc32.New in an error was double-counting imports
from different packages or indirect imports, so it was
quoting even when there was no ambiguity.
Gustavo Niemeyer [Tue, 28 Jun 2011 21:28:30 +0000 (22:28 +0100)]
ld: fix ELF strip by removing overlap of sections
The gosymtab and gopclntab sections were pointing to the proper
data, but that data was already owned by the rodata section.
Some ELF references explicitly prohibit multiple sections from
owning the same data, and strip behaves accordingly.
The data for these sections was moved to after rodata, and the
gosymtab and gopclntab sections now own their respective ranges.
This change makes strip happy both with and without -s being
provided at link time. Note that it won't remove these sections
because they are still allocated, and that's by design since
they are necessary at runtime for generating proper backtraces
and similar introspection operations.
Unlike the previous behavior, -s will now maintain zero-sized
gosymtab and gopclntab sections. This makes the implementation
slightly cleaner.
Dmitriy Vyukov [Tue, 28 Jun 2011 19:09:53 +0000 (15:09 -0400)]
runtime: replace Semacquire/Semrelease implementation
1. The implementation uses distributed hash table of waitlists instead of a centralized one.
It significantly improves scalability for uncontended semaphores.
2. The implementation provides wait-free fast-path for signalers.
3. The implementation uses less locks (1 lock/unlock instead of 5 for Semacquire).
4. runtime·ready() call is moved out of critical section.
5. Semacquire() does not call semwake().
Benchmark results on HP Z600 (2 x Xeon E5620, 8 HT cores, 2.40GHz)
are as follows:
benchmark old ns/op new ns/op delta
runtime_test.BenchmarkSemaUncontended 58.20 36.30 -37.63%
runtime_test.BenchmarkSemaUncontended-2 199.00 18.30 -90.80%
runtime_test.BenchmarkSemaUncontended-4 327.00 9.20 -97.19%
runtime_test.BenchmarkSemaUncontended-8 491.00 5.32 -98.92%
runtime_test.BenchmarkSemaUncontended-16 946.00 4.18 -99.56%
Dmitriy Vyukov [Tue, 28 Jun 2011 13:43:01 +0000 (09:43 -0400)]
sync: add fast path to Once
The implementation does not grab the lock,
if Once is already initalized.
Benchmark results on HP Z600 (2 x Xeon E5620, 8 HT cores, 2.40GHz)
are as follows:
benchmark old ns/op new ns/op delta
sync_test.BenchmarkOnce 187.00 14.00 -92.51%
sync_test.BenchmarkOnce-2 909.00 21.40 -97.65%
sync_test.BenchmarkOnce-4 3684.00 20.90 -99.43%
sync_test.BenchmarkOnce-8 5987.00 23.00 -99.62%
sync_test.BenchmarkOnce-16 5051.00 21.60 -99.57%
Rob Pike [Tue, 28 Jun 2011 13:04:08 +0000 (23:04 +1000)]
Add the beginnings of the template execution code. Lots still to do,
including evaluation up the data tree (in this code all fields must be
in dot itself), plus more control structure, but the basics are in place.
R=rsc, r
CC=golang-dev
https://golang.org/cl/4665041
Dmitriy Vyukov [Tue, 28 Jun 2011 12:14:54 +0000 (08:14 -0400)]
sync: replace Mutex benchmarks
For both contended and uncontended case:
- support arbitrary number of cpus (not just 2)
- dynamic load balancing (improves stability)
- periodic execution of Gosched() to work around non-preemptiviness
For uncontended case eliminates possible false-sharing.
For contended case includes additional variation with some
amount of local work between mutex operations.
Russ Cox [Tue, 28 Jun 2011 02:42:34 +0000 (22:42 -0400)]
cc: add two new #pragma varargck
#pragma varargck countpos f 1
says that the first argument to f is
the count of variadic arguments that follow.
#pragma varargck type f t
says that t is one of the allowed types for
a variadic argument to f.
(can be repeated)
combined, these can be used to check the
runtime.stdcall functions in the windows port
or in any other port that needs a vararg list of
uintptrs even on a 64-bit platform (where it is
very easy to pass a less-than-uintptr in the ...).
demo:
typedef unsigned int uintptr;
#pragma varargck countpos f 1
#pragma varargck type f uintptr
#pragma varargck type f void*
int f(int count, ...);
void *v;
char *p;
void
main(void)
{
f(1, v); // ok
f(1, main); // ok
f(1, p); // ok
f(2, v, v); // ok
f(2, v); // found 1 argument after count 2
f(1, 'a'); // invalid type INT in call to f
f(1, 0); // invalid type INT in call to f
}
Rob Pike [Mon, 27 Jun 2011 23:43:14 +0000 (09:43 +1000)]
strings.Split: make the default to split all.
Change the signature of Split to have no count,
assuming a full split, and rename the existing
Split with a count to SplitN.
Do the same to package bytes.
Add a gofix module.
Kyle Lemons [Mon, 27 Jun 2011 23:07:28 +0000 (19:07 -0400)]
xml: add Marshal and MarshalIndent
I have written up a Marshal and MarshalIndent pair that should
closely reflect the way that Unmarshal works. I would love feedback
on making this code more accessible and efficient... I haven't used
reflecton on this scale before, so there is probably a lot of work
that can be done on that.
Some potentially controversial things:
- All tag names are lower-cased by default.
- Zero-valued struct values are skipped.
- No namespace prefix (o:tag, etc) mechanism is supplied.
- You are allowed to marshal non-struct values (even though unmarshal
cannot handle them).
- A tag for a non-XMLName struct field that isn't "attr", "chardata",
or "innerxml" is used as the name of the tag. This could wreak
havoc if you try to marshal a protobuf struct.
- The "innerxml" and "chardata" are inserted verbatim. If you try to
marshal something straight from unmarshal, the results could be
unexpected (remove "innerxml" support from Marshal would be one
possible solution).
Graham Miller [Mon, 27 Jun 2011 20:12:04 +0000 (16:12 -0400)]
bufio: do not cache Read errors
Reader previously had cached an error from the underlying reader
and would return it on every subsequent call to Read. The Reader
will now return the error only once, and subsequent calls will result
in a new Read call to the underlying Reader.
Lucio De Re [Mon, 27 Jun 2011 18:42:18 +0000 (14:42 -0400)]
8a: fixes for Plan 9 build
8a/a.h:
. Removed <u.h> and <libc.h> includes as they work better in "a.y".
. Made definition of EOF conditional as it's defined in the Plan 9
header files, but not elsewhere.
8a/a.y:
. Added <u.h> and <libc.h> because <stdio.h> in Plan 9 needs them.
Sequence <u.h>, <stdio.h>, <libc.h> recommended by RSC.
8a/lex.c:
. Added <u.h> and <libc.h> as now needed by "a.h".
. Dropped <ctype.h>.
Brad Fitzpatrick [Mon, 27 Jun 2011 17:37:33 +0000 (10:37 -0700)]
http: do TLS handshake explicitly before copying TLS state
Previously we were snapshotting the TLS state into *Request
before we did the HTTP ReadRequest, the first Read of which
triggered the TLS handshake implicitly.
Dmitriy Vyukov [Mon, 27 Jun 2011 17:31:40 +0000 (13:31 -0400)]
gotest: add -test.benchtime and -test.cpu flags.
-test.benchtime allows to specify benchmark execution time.
-test.cpu allows to execute tests/benchmarks for several
values of GOMAXPROCS.
Michael T. Jones [Sat, 25 Jun 2011 00:26:45 +0000 (17:26 -0700)]
fmt: Added SkipSpace() function to fmt's ScanState interface.
Users of the Scan() infrastructure that employ ReadRune() rather than
Token() need a way to skip leading spaces and newlines as set by the
the parent, Fscan(), Fscanln, or Fscanf(). As the internal methods and
boolean flags are not exported, this new function was added here and
in the Int and Nat Scan() functions of the big package. (fmt.Rat did
not need change since it uses Token()) Also added Printf style format
code support to int types and tests for same to int_test.go
Brad Fitzpatrick [Fri, 24 Jun 2011 23:46:14 +0000 (16:46 -0700)]
http: better handling of 0-length Request.Body
As rsc suggested after change 58a6bdac3d12 was committed, we
now read the first byte of Request.Body when the
Request.ContentLength is 0 to disambiguate between a truly
zero-length body and a body of unknown length where the user
didn't set the ContentLength field.
This was also causing the reverse proxy problem where incoming
requests (which always have a body, of private type http.body,
even for 0-lengthed requests) were being relayed to the http
Transport for fetching, which was serializing the request as a
chunked request (since ContentLength was 0 and Body was
non-nil)
Gustavo Niemeyer [Fri, 24 Jun 2011 03:29:59 +0000 (00:29 -0300)]
runtime: don't use twice the memory with grsec-like kernels
grsec needs the FIXED flag to be provided to mmap, which
works now. That said, when the allocation fails to be made
in the specific address, we're still given back a writable
page. This change will unmap that page to avoid using
twice the amount of memory needed.
It'd also be pretty easy to avoid the extra system calls
once we detected that the flag is needed, but I'm not sure
if that edge case is worth the effort.
Ken Rockot [Wed, 22 Jun 2011 22:07:20 +0000 (18:07 -0400)]
syscall: add tty support to StartProcess
These changes add a Ctty int field to the Unix syscall.ProcAttr which,
if set >= 0 in conjuction with Setsid=true, will be used by
forkAndExecInChild as the file descriptor for the new child's
controlling terminal.
Necessary changes have been made to mkerrors.sh to generate defs for
TIOC*, though changes to its output files are not included here.
The changes made should support Linux, FreeBSD and Darwin, at least.
Gustavo Niemeyer [Wed, 22 Jun 2011 19:12:22 +0000 (15:12 -0400)]
ld: don't attempt to build dynamic sections unnecessarily
This prevents ld from generating zeroed symtab entries for
sections that aren't going to be generated because dynamic
linkage has been disabled (-d was used or no dynamic libs
were seen). Even though they were not explicitly added by
doelf, the section creation process was making them
reachable again.
The windows head is being disconsidered for this because
apparently it's not taking into account debug['d'].
Robert Griesemer [Wed, 22 Jun 2011 17:52:47 +0000 (10:52 -0700)]
os.Error API: don't export os.ErrorString, use os.NewError consistently
This is a core API change.
1) gofix misc src
2) Manual adjustments to the following files under src/pkg:
gob/decode.go
rpc/client.go
os/error.go
io/io.go
bufio/bufio.go
http/request.go
websocket/client.go
as well as:
src/cmd/gofix/testdata/*.go.in (reverted)
test/fixedbugs/bug243.go
3) Implemented gofix patch (oserrorstring.go) and test case (oserrorstring_test.go)
Russ Cox [Wed, 22 Jun 2011 01:45:13 +0000 (21:45 -0400)]
godefs: remove test from build
The test is only defined on darwin/amd64, and it fails
with recent versions of Xcode, which do not support
-gstabs+ debugging output. At some point godefs will
have to be replaced, perhaps merged with cgo.
Godefs is not needed during builds anyway (its output files
are checked into the repository in src/pkg/runtime),
so its failure on the newer Xcode is a distraction from an
otherwise usable build. Disable the test.
Fixes #1985.
R=golang-dev, gri, robert.hencke, r
CC=golang-dev
https://golang.org/cl/4638053
Lucio De Re [Tue, 21 Jun 2011 16:14:32 +0000 (12:14 -0400)]
8l: more fixes for Plan 9
Once these changes are effected, it is possible to construct
"8l" native on a (386?) Plan 9 system, albeit with assistance
from modules such as mkfiles that are not (yet) included in any
public patches.
8l/asm.c:
. Corrected some format qualifiers.
8l/list.c:
. Cast a print() argument to (int) to match the given format.
It may be possible to change the format (%R), but I have not
looked into it.
8l/obj.c:
. Removed some unused code.
8l/span.c:
. Removed unnecessary incrementation on "bp".
. Corrected some format qualifiers.
ld/data.c:
. Corrected some format qualifiers.
. Cast print argument to (int): used as field size.
. Use braces to suppress warning about empty if() statements.
ld/dwarf.c:
. Trivial spelling mistake in comment.
ld/ldelf.c:
. Added USED() statements to silence warnings.
. Dropped redundant address (&) operators.
. corrected some format qualifiers.
. Cast to (int) for switch selection variable.
ld/macho.c:
. Added USED() statements to silence warnings.
ld/ldpe.c:
. Added USED() statements to silence warnings.
. More careful use of "sect" variable.
. Corrected some format qualifiers.
. Removed redundant assignments.
. Minor fix dropped as it was submitted separately.
ld/pe.c:
. Dropped <time.h> which is now in <u.h>.
. Dropped redundant address (&) operators.
. Added a missing variable initialisation.
ld/symtab.c:
. Added USED() statements to silence warnings.
. Removed redundant incrementation.
. Corrected some format qualifiers.
All the above have been tested against a (very) recent release
and do not seem to trigger any regressions.
Rob Pike [Mon, 20 Jun 2011 22:31:02 +0000 (08:31 +1000)]
fmt: catch panics from calls to String etc.
This change causes Print et al. to catch panics generated by
calls to String, GoString, and Format. The panic is formatted
into the output stream as an error, but the program continues.
As a special case, if the argument was a nil pointer, the
result is just "<nil>", because that's almost certainly enough
information and handles the very common case of String
methods that don't guard against nil.
Scan does not want this change. Input must work; output can
be for debugging and it's nice to get output even when you
make a mistake.