Move the event handler on macos from poll to kqueue. Simpler and faster.

Move the event handler on macos from poll to kqueue. Simpler and faster.

R=sgjesse@google.com
BUG=
TEST=

Committed: https://code.google.com/p/dart/source/detail?r=4028

[Søren Gjesse, 2012-02-08 10:40:36]

lgtm

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
File runtime/bin/eventhandler_macos.cc (right):

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:47: static void RemoveFromKqueue(intptr_t
kqueue_fd_, SocketData* sd) {
Maybe add constant kMaxChanges = 2.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:60: if (changes > 0) {
And ASSERT(changes <= kMaxChanges)

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:73: intptr_t changes = 0;
Ditto constant.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:77: if (sd->HasReadEvent()) {
Is sd->read_tracked_by_kqueue() always false here? If so ASSERT otherwise skip
the EV_SET.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:87: if (sd->HasWriteEvent()) {
Ditto, is sd->write_tracked_by_kqueue() is always false?

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:242: intptr_t
EventHandlerImplementation::GetEvents(struct kevent* event,
Nice that we don't need special handling of pipes and need recv calls any more.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:274: // If the read end closed, close the
write end as well,
Maybe rephrase "If the read end closed, close the write end as well," to
something like "If the receiver closed for reading then close for writing".

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
File runtime/bin/eventhandler_macos.h (right):

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.h:9: #error Do not include eventhandler_linux.h
directly; use eventhandler.h instead.
Please revert this change.

[Mads Ager (google), 2012-02-08 12:15:15]

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
File runtime/bin/eventhandler_macos.cc (right):

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:47: static void RemoveFromKqueue(intptr_t
kqueue_fd_, SocketData* sd) {
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> Maybe add constant kMaxChanges = 2.

Done.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:60: if (changes > 0) {
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> And ASSERT(changes <= kMaxChanges)

Done.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:73: intptr_t changes = 0;
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> Ditto constant.

Done.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:77: if (sd->HasReadEvent()) {
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> Is sd->read_tracked_by_kqueue() always false here? If so ASSERT otherwise skip
> the EV_SET.

It is not always false. What is there now is correct because an ADD operation
means 'modify if exists'. However, by checking explicitly we can sometimes avoid
a system call. I have added the explicit checking.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:87: if (sd->HasWriteEvent()) {
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> Ditto, is sd->write_tracked_by_kqueue() is always false?

Same as above. Added explicit checking.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:242: intptr_t
EventHandlerImplementation::GetEvents(struct kevent* event,
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> Nice that we don't need special handling of pipes and need recv calls any
more.

Yes, this is so much cleaner. :)

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.cc:274: // If the read end closed, close the
write end as well,
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> Maybe rephrase "If the read end closed, close the write end as well," to
> something like "If the receiver closed for reading then close for writing".

Done.

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
File runtime/bin/eventhandler_macos.h (right):

https://chromiumcodereview.appspot.com/9365010/diff/2001/runtime/bin/eventhan...
runtime/bin/eventhandler_macos.h:9: #error Do not include eventhandler_linux.h
directly; use eventhandler.h instead.
On 2012/02/08 10:40:36, Søren Gjesse wrote:
> Please revert this change.

Whoops. Thanks! I started this with a copy from the linux file because the
structure would be similar to epoll. :)