Add optimized String.fromCharCodes path for Uint8List and Int8List.

Add optimized String.fromCharCodes path for Uint8List and Int8List.

BUG=
R=srdjan@google.com

Committed: https://code.google.com/p/dart/source/detail?r=30496

[Anders Johnsen, 2013-11-18 20:26:31]

Currently, there are no check in the C++ code - is it okay to make assumptions
like this?

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart
File runtime/lib/string_patch.dart (right):

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart...
runtime/lib/string_patch.dart:37: if (charCodes is Uint8List || charCodes is
Int8List) {
Should we use _classId here as well?

[srdjan, 2013-11-18 20:45:15]

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string.cc
File runtime/lib/string.cc (right):

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string.cc#newcode156
runtime/lib/string.cc:156: const Object& list =
Object::Handle(arguments->NativeArgAt(0));
Instance::CheckedHandle(arguments->NativeArgAt(0));

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart
File runtime/lib/string_patch.dart (right):

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart...
runtime/lib/string_patch.dart:37: if (charCodes is Uint8List || charCodes is
Int8List) {
On 2013/11/18 20:26:31, Anders Johnsen wrote:
> Should we use _classId here as well?

Yes, using cid-s is much quicker.

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart...
runtime/lib/string_patch.dart:59: return
_OneByteString._allocateFromOneByteList(charCodes);
Calling to native is quite slow, i.e., for small lists this will be quite more
expensive. Could you benchmark? Any golem numbers on existing benchmarks?

[Anders Johnsen, 2013-11-18 21:10:50]

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart
File runtime/lib/string_patch.dart (right):

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart...
runtime/lib/string_patch.dart:59: return
_OneByteString._allocateFromOneByteList(charCodes);
On 2013/11/18 20:45:15, srdjan wrote:
> Calling to native is quite slow, i.e., for small lists this will be quite more
> expensive. Could you benchmark? Any golem numbers on existing benchmarks?

Before, we already did the runtime-call to allocate the list. So, doing this is
actually faster (~25% for a small local test). Will upload golem numbers
tomorrow at the office.

[Anders Johnsen, 2013-11-19 07:38:42]

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string.cc
File runtime/lib/string.cc (right):

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string.cc#newcode156
runtime/lib/string.cc:156: const Object& list =
Object::Handle(arguments->NativeArgAt(0));
On 2013/11/18 20:45:15, srdjan wrote:
> Instance::CheckedHandle(arguments->NativeArgAt(0));

Done.

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart
File runtime/lib/string_patch.dart (right):

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart...
runtime/lib/string_patch.dart:37: if (charCodes is Uint8List || charCodes is
Int8List) {
On 2013/11/18 20:45:15, srdjan wrote:
> On 2013/11/18 20:26:31, Anders Johnsen wrote:
> > Should we use _classId here as well?
> 
> Yes, using cid-s is much quicker.

I encountered an issue here. These classes are in typed_data-patch, so:

- _Uint8List is not accessible from dart:core-patch.
- _cid is not accessible within dart:typed_data-patch.
- Uint8List._privateHelper is not accessible

What's your recommendation here?

[srdjan, 2013-11-19 16:47:38]

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart
File runtime/lib/string_patch.dart (right):

https://codereview.chromium.org/74423005/diff/1/runtime/lib/string_patch.dart...
runtime/lib/string_patch.dart:37: if (charCodes is Uint8List || charCodes is
Int8List) {
On 2013/11/19 07:38:43, Anders Johnsen wrote:
> On 2013/11/18 20:45:15, srdjan wrote:
> > On 2013/11/18 20:26:31, Anders Johnsen wrote:
> > > Should we use _classId here as well?
> > 
> > Yes, using cid-s is much quicker.
> 
> I encountered an issue here. These classes are in typed_data-patch, so:
> 
> - _Uint8List is not accessible from dart:core-patch.
> - _cid is not accessible within dart:typed_data-patch.
> - Uint8List._privateHelper is not accessible
> 
> What's your recommendation here?

This would not slow down existing code:

if ((ccid != _List._classId) && (ccid != _GrowableList._classId) && (ccid !=
_ImmutableList._classId)) {
  if (charCodes is Uint8List || charCodes is Int8List) {
    return _OneByteString._allocateFromOneByteList(charCodes);
  }
  charCodes = new List<int>.from(charCodes, growable: false);
}

You may want to benefit from skipping the call to native for small 8-bit lists
and skip the isOneByteString test.

You could also check first the charCodes length:

if (charCodes.length >128) {
  return _OneByteString._allocateFromOneByteList(charCodes);
}

We plan to make all cid's available through an internal VM class. Until then use
the instanceOf test is OK.

[Anders Johnsen, 2013-11-20 13:05:59]

Thanks, that's much better! PTAL

[srdjan, 2013-11-20 16:06:57]

LGTM with comments

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
File runtime/lib/string_patch.dart (right):

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
runtime/lib/string_patch.dart:43: if (charCodes is Uint8List || charCodes is
Int8List) {
Add parentheses around is tests.

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
runtime/lib/string_patch.dart:51: for (int i = 0; i < charCodes.length; i++) {
You may want to prefetch chatCodes.length and keep it in a final variable. It
may not help, but it would not hurt.

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
runtime/lib/string_patch.dart:61: // Allocate a one byte string.
Add a comment that above the threshold, the native code is faster. Is 128 the
proper threshold?

[Anders Johnsen, 2013-11-21 06:06:27]

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
File runtime/lib/string_patch.dart (right):

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
runtime/lib/string_patch.dart:43: if (charCodes is Uint8List || charCodes is
Int8List) {
On 2013/11/20 16:06:57, srdjan wrote:
> Add parentheses around is tests.

Done.

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
runtime/lib/string_patch.dart:51: for (int i = 0; i < charCodes.length; i++) {
On 2013/11/20 16:06:57, srdjan wrote:
> You may want to prefetch chatCodes.length and keep it in a final variable. It
> may not help, but it would not hurt.

Done.

https://codereview.chromium.org/74423005/diff/120001/runtime/lib/string_patch...
runtime/lib/string_patch.dart:61: // Allocate a one byte string.
On 2013/11/20 16:06:57, srdjan wrote:
> Add a comment that above the threshold, the native code is faster. Is 128 the
> proper threshold?

Just did a test with normal lists. 128 is indeed the magic number.

[Anders Johnsen, 2013-11-21 06:12:09]

Committed patchset #4 manually as r30496 (presubmit successful).