Implement SIMD types for dart2js

Implement SIMD types for dart2js

R=vsm@google.com

Committed: https://code.google.com/p/dart/source/detail?r=25468

[vsm, 2013-07-24 20:53:36]

Are the typed_data tests actually running under dart2js?  I don't see a status
file change here.

[Cutch, 2013-07-24 20:55:31]

On 2013/07/24 20:53:36, vsm wrote:
> Are the typed_data tests actually running under dart2js?  I don't see a status
> file change here.

They are. I forgot to push up the status change.

[Cutch, 2013-07-24 20:56:52]

To be clear, I'm testing with the following command line:

$ ./tools/test.py -c dart2js -r d8 lib/typed_data
Test configuration: dart2js_d8_debug_ia32
[00:43 | 100% | +    8 | -    0]

[Cutch, 2013-07-24 20:59:08]

PTAL at Patch Set 2.

[vsm, 2013-07-24 21:24:58]

lgtm

My primary concern is the significant amount of heap allocation going on here
wrt to the native version.  Can you track the perf on this?

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
File sdk/lib/typed_data/dart2js/typed_data_dart2js.dart (right):

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:696: _storage[(i*4)+0] =
list[i].x;
It might be worth specializing this for when list is a Float32x4.  E.g., bulk
copy from list._storage to this._storage instead of allocating so many temporary
Float32x4 objects.

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:733: r._storage[(i*4)+0] =
_storage[((i+start)*4)+0];
How about something like r._storage = this._storage.sublist(...)?

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:910: _x = _x > _ux ? _ux :
_x;
Do you need to check that _lx <= _ux etc?

[Cutch, 2013-07-24 22:00:05]

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
File sdk/lib/typed_data/dart2js/typed_data_dart2js.dart (right):

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:696: _storage[(i*4)+0] =
list[i].x;
On 2013/07/24 21:24:58, vsm wrote:
> It might be worth specializing this for when list is a Float32x4.  E.g., bulk
> copy from list._storage to this._storage instead of allocating so many
temporary
> Float32x4 objects.

Done.

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:733: r._storage[(i*4)+0] =
_storage[((i+start)*4)+0];
On 2013/07/24 21:24:58, vsm wrote:
> How about something like r._storage = this._storage.sublist(...)?

Done.

https://codereview.chromium.org/19801005/diff/4001/sdk/lib/typed_data/dart2js...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:910: _x = _x > _ux ? _ux :
_x;
On 2013/07/24 21:24:58, vsm wrote:
> Do you need to check that _lx <= _ux etc?

We don't in the native C implementation nor in the inlined instructions on
IA32/X64/ARM.

[vsm, 2013-07-24 22:09:27]

Still LGTM

https://codereview.chromium.org/19801005/diff/12001/sdk/lib/typed_data/dart2j...
File sdk/lib/typed_data/dart2js/typed_data_dart2js.dart (right):

https://codereview.chromium.org/19801005/diff/12001/sdk/lib/typed_data/dart2j...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:699: // that this list is a
Float32x4List.
Would this work instead of iterating?:
_storage = new Float32List.fromList(storage);

https://codereview.chromium.org/19801005/diff/12001/sdk/lib/typed_data/dart2j...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:1198: Float32x4
withZWInXY(Float32x4 other) {
These are looking awfully specialized.  Are they really that useful?

[Cutch, 2013-07-24 22:19:31]

https://codereview.chromium.org/19801005/diff/12001/sdk/lib/typed_data/dart2j...
File sdk/lib/typed_data/dart2js/typed_data_dart2js.dart (right):

https://codereview.chromium.org/19801005/diff/12001/sdk/lib/typed_data/dart2j...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:699: // that this list is a
Float32x4List.
On 2013/07/24 22:09:27, vsm wrote:
> Would this work instead of iterating?:
> _storage = new Float32List.fromList(storage);

Done.

https://codereview.chromium.org/19801005/diff/12001/sdk/lib/typed_data/dart2j...
sdk/lib/typed_data/dart2js/typed_data_dart2js.dart:1198: Float32x4
withZWInXY(Float32x4 other) {
On 2013/07/24 22:09:27, vsm wrote:
> These are looking awfully specialized.  Are they really that useful?

Yes, this method and the new interleave methods are necessary to transpose a 4x4
matrix without leaving registers.

[vsm, 2013-07-25 00:03:17]

lgtm

[Cutch, 2013-07-25 14:46:46]

Committed patchset #6 manually as r25468 (presubmit successful).