Dart: Performance improvements to Dart's handle watcher.

Dart: Performance improvements to Dart's handle watcher.

- Keeps state for the call to MojoWaitMany in external TypedData arrays
  to reduce copying.
- Processes all ready signals to avoid repeated calls.
- Puts more active handles at the front of the handles array.
- Disables checked mode for the vm service and handle watcher isolates

On my machine:
C++ <=> C++: ~40-50us
Dart <=> Dart: ~150-160us => 120-130us
Dart <=> C++: ~90-100us => 60-70us

So, on desktop savings of ~30us for round trips involving Dart code.

BUG=
R=johnmccutchan@google.com

Committed: https://chromium.googlesource.com/external/mojo/+/e54a91983350239d275cc1f0b12a1bdb03946513

[zra, 2015-10-05 20:57:59]

zra@google.com changed reviewers:
+ johnmccutchan@google.com

[Cutch, 2015-10-05 21:12:26]

General question- what was the measured improvement from this change?

https://codereview.chromium.org/1378163004/diff/80001/mojo/dart/embedder/mojo...
File mojo/dart/embedder/mojo_natives.cc (right):

https://codereview.chromium.org/1378163004/diff/80001/mojo/dart/embedder/mojo...
mojo/dart/embedder/mojo_natives.cc:728: static std::vector<mojo::Handle>
mwm_handles;
Please create a class:

class MojoWaitState {
}


With these statics.

Also, drop the mojo wait many 'mwm' prefix.

https://codereview.chromium.org/1378163004/diff/80001/mojo/dart/embedder/mojo...
mojo/dart/embedder/mojo_natives.cc:796: uint32_t* handles =
reinterpret_cast<uint32_t*>(mwm_handles.data());
Seems like mwm_handles and mwm_signals should just be std::vector<uint32_t>

https://codereview.chromium.org/1378163004/diff/80001/mojo/public/dart/mojo/s...
File mojo/public/dart/mojo/sdk_ext/src/wait_many_state.dart (right):

https://codereview.chromium.org/1378163004/diff/80001/mojo/public/dart/mojo/s...
mojo/public/dart/mojo/sdk_ext/src/wait_many_state.dart:48: // more active
handles will tend to be at the front of the list.
What performance benefit have you measured by putting active handles at the
front of the list?

https://codereview.chromium.org/1378163004/diff/80001/mojo/public/dart/mojo/s...
mojo/public/dart/mojo/sdk_ext/src/wait_many_state.dart:67: _ports[i + 1] =
_ports[i];
We are copying the state possibly twice here, in the call to '_grow' we copy and
then we copy again here.

[zra, 2015-10-05 22:34:10]

https://codereview.chromium.org/1378163004/diff/80001/mojo/dart/embedder/mojo...
File mojo/dart/embedder/mojo_natives.cc (right):

https://codereview.chromium.org/1378163004/diff/80001/mojo/dart/embedder/mojo...
mojo/dart/embedder/mojo_natives.cc:728: static std::vector<mojo::Handle>
mwm_handles;
On 2015/10/05 21:12:26, Cutch wrote:
> Please create a class:
> 
> class MojoWaitState {
> }
> 
> 
> With these statics.
> 
> Also, drop the mojo wait many 'mwm' prefix.

Done.

https://codereview.chromium.org/1378163004/diff/80001/mojo/dart/embedder/mojo...
mojo/dart/embedder/mojo_natives.cc:796: uint32_t* handles =
reinterpret_cast<uint32_t*>(mwm_handles.data());
On 2015/10/05 21:12:26, Cutch wrote:
> Seems like mwm_handles and mwm_signals should just be std::vector<uint32_t>

Done.

https://codereview.chromium.org/1378163004/diff/80001/mojo/public/dart/mojo/s...
File mojo/public/dart/mojo/sdk_ext/src/wait_many_state.dart (right):

https://codereview.chromium.org/1378163004/diff/80001/mojo/public/dart/mojo/s...
mojo/public/dart/mojo/sdk_ext/src/wait_many_state.dart:48: // more active
handles will tend to be at the front of the list.
On 2015/10/05 21:12:26, Cutch wrote:
> What performance benefit have you measured by putting active handles at the
> front of the list?

10+ us per round trip on average.

https://codereview.chromium.org/1378163004/diff/80001/mojo/public/dart/mojo/s...
mojo/public/dart/mojo/sdk_ext/src/wait_many_state.dart:67: _ports[i + 1] =
_ports[i];
On 2015/10/05 21:12:26, Cutch wrote:
> We are copying the state possibly twice here, in the call to '_grow' we copy
and
> then we copy again here.

The extra copy will happen only once for every 100 additional handles. Instead
of doing backflips to avoid the copy in this CL, I plan to add a benchmark that
will generate many more handles in a forthcoming CL to see how the handle
watcher behaves. In that case addFront may no longer help, but I'd rather make
that decision with some numbers to back it up.

[zra, 2015-10-06 16:21:38]

Found a bad bug in this CL. I'll ping again when it's ready for another look.

[zra, 2015-10-06 20:31:18]

Bug fixed. PTAL.

[tonyg, 2015-10-06 20:55:13]

On 2015/10/05 21:12:26, Cutch wrote:
> General question- what was the measured improvement from this change?

Random drive-by: I'd love to see before & after benchmark results in the patch
description.

[zra, 2015-10-06 21:12:14]

On 2015/10/06 20:55:13, tonyg wrote:
> On 2015/10/05 21:12:26, Cutch wrote:
> > General question- what was the measured improvement from this change?
> 
> Random drive-by: I'd love to see before & after benchmark results in the patch
> description.

Updated description.

[zra, 2015-10-07 04:50:36]

Disabling checked mode for the vm service and handle watcher isolates improves
round trip times slightly, and improves startup time by 10-20ms on my machine.

[Cutch, 2015-10-07 13:52:37]

lgtm

[zra, 2015-10-07 16:27:46]

Committed patchset #10 (id:180001) manually as
e54a91983350239d275cc1f0b12a1bdb03946513 (presubmit successful).