| bdd17cde8cf675f5d5703aad4dfb565932fb4c50 |
|
08-Feb-2012 |
Elliott Hughes <enh@google.com> |
Fix an ICS regression with hasArray on a JNI-allocated DirectByteBuffer.
Bug: http://code.google.com/p/android/issues/detail?id=24327
Change-Id: Ib31b60fe01e3d0b5a1ed7125ddf047e09fd64056
|
| 5a1e8cc38e1cc6c3f2fc66f039ca0fc5309a0dc3 |
|
16-Dec-2010 |
Elliott Hughes <enh@google.com> |
MappedByteBuffer get(byte[]) and get(byte[], int, int) weren't updating 'position'.
The majority of this change is work to extend our tests to cover
MappedByteBuffer too, plus new tests to check that 'position' is
always updated correctly by ByteBuffer's relative get/put methods.
Bug: 3291927
Change-Id: I58217de45ba65675730f30771fba592cb2af28ff
|
| 43a9f774d075e0e441d8b996e3f6c81ea483ec89 |
|
19-Oct-2010 |
Elliott Hughes <enh@google.com> |
BufferIterator support for byte[].
Bug: 3032515
Change-Id: I040a92c3ce241aa97719ff94e3cc31c62473bf46
|
| ae704b984c10a63883cc366e823d53902d6ac7a9 |
|
21-Sep-2010 |
Elliott Hughes <enh@google.com> |
Minor tidy-up.
The nio implementation is still a bit of a mess because Java doesn't have any
kind of "module" concept, nor "friend". But this makes it a little less untidy,
and makes the deliberate access control subversions a little more explicit, and
makes various things less accessible than they were.
Change-Id: I514b1e24000b91f86b768ed408fedae115dc3080
|
| f714d3767875c96c5122683ef8c75cc9d5f1259f |
|
13-Sep-2010 |
Elliott Hughes <enh@google.com> |
Make MappedByteBuffer.get(byte[], int, int) orders of magnitude faster.
Even with the misleading logarithmic scale graphs, this stood out like
a sore thumb. Other than this, our MappedByteBuffer looks like it's just a
layer or two of missing inlining away from DirectByteBuffer performance.
Bug: 2935622
Change-Id: Ie4ed153fad6d19a2c83748fee2e364c3e86d40f3
|
| dbb973b7f1dc6d3e86c6f9528a2aadff1e01c0cb |
|
10-Sep-2010 |
Elliott Hughes <enh@google.com> |
Fix MappedByteBufferAdapter.
I broke MappedByteBufferAdapter in https://android-git.corp.google.com/g/65226
because I didn't move the position-setting to _before_ the delegated gets (the
puts were correct). This was spotted in review, but I'd already submitted to
unblock a later change.
The upstream tests didn't spot this because they mainly test buffers full of
zeros (which aren't very interesting from an endian point of view). I've raised
http://b/2991877 to get the upstream tests fixed.
I've also added a motivating comment for this admittedly weird class.
Bug: 2935622
Change-Id: I6bbf00636a5e05e7396cf94ee6985bcf50fc4552
|
| 0e1b748ecabf720065a632f28330f5d4d037d5ae |
|
10-Sep-2010 |
Elliott Hughes <enh@google.com> |
Consistent naming for get/put/wrap arguments.
Bug: 2935622
Change-Id: Icb39392f9106f9fe5ba70ccdcb6b059dbf5617bc
|
| 0c53cf8b2c46deb41e91db50ddc17d598cc64a10 |
|
09-Sep-2010 |
Elliott Hughes <enh@google.com> |
Speed up MappedByteBuffer.
Make it possible to override limit(int) and position(int) internally, so I can
avoid updating MappedByteBufferAdapter's wrapped ByteBuffer's limit and
position on every get or put. I still need to do extra work with position if
you use the get/put overloads that don't take an index, but even that becomes
cheaper. The indexed overloads benefit the most, though. Before:
benchmark before after (both us)
ByteBuffer_getByte 830 757
ByteBuffer_getChar 940 875
ByteBuffer_getByte_indexed 800 509
ByteBuffer_getChar_indexed 936 643
Also factor out duplicated toString implementations.
Bug: 2935622
Change-Id: I4cfe206fde638bddb5a7b4a62223e2588f44b8ac
|
| 0440837fd0013373ba3476283151299e7be9e5a3 |
|
08-Sep-2010 |
Elliott Hughes <enh@google.com> |
Move the MemoryBlock up into java.nio.Buffer.
This lets me kill getBaseAddress and the (bogus) DirectBuffer interface.
Bug: 2935622
Change-Id: I51c758e7687a6cfa4a3a64ef089fb600b847b70f
|
| eb29579498a8860f81fd38275f8657c21bb67abb |
|
08-Sep-2010 |
Elliott Hughes <enh@google.com> |
Rename PlatformAddress to MemoryBlock.
Bug: 2935622
Change-Id: I8a09cb9be0b0ea468278f7808f1a18c3ce820b49
|
| 0f335c4ed1a0cb419c8ce82183fabb241d5dd032 |
|
07-Sep-2010 |
Elliott Hughes <enh@google.com> |
Kill getEffectiveAddress.
Set the corresponding field in the constructors.
Requires a dalvikvm change: https://android-git.corp.google.com/g/64866
Bug: 2935622
Change-Id: I26663bfcdb80db75672be71395b861c1a79c24fc
|
| c73a5be50cdd804ff3c12e7b43da08c33cdd6f21 |
|
04-Sep-2010 |
Elliott Hughes <enh@google.com> |
Move some of the nio implementation into java.nio.
Another stepping-stone to simplifying the PlatformAddress/OSMemory situation.
Bug: 2935622
Change-Id: Ic5b941ee7c17d6b5aeda53c3c2cdf2c43cebff55
|
| c3c38fbbccaf428200acdd819b1f7799edac54d2 |
|
04-Sep-2010 |
Elliott Hughes <enh@google.com> |
Remove nio cruft, move the implementation cost of 'free' where it belongs...
...and inline some getBaseAddress calls. The JIT doesn't manage to do that, it
seems, but being the JIT is not really my purpose: I'm trying to kill
getBaseAddress.
Bug: 2935622
Change-Id: I3cd99cd131a2c09796f5ff5bb31332287db2130f
|
| 29dbfe19b113a13b712be2bc762ef1c81cd06c47 |
|
03-Sep-2010 |
Elliott Hughes <enh@google.com> |
Use finalizers to free direct buffers.
Also switch to a cheaper scheme for avoiding double-frees (which could only
happen if library code is broken; at the moment the only caller is the
iovec-style stuff). This halves the cost of direct buffer access, at the
unimportant cost of not throwing an exception if we try to access a buffer
we freed while we still had a reference to it.
Also use named constants for the sizes of primitive types.
Also kill BufferFactory.
Bug: 2935622
Change-Id: I6021ce7531f49d8cff7635d6d2cc54851bacaf69
|
| f46d2ce4d4c92d16e2574f57f70fc5477dc12697 |
|
02-Sep-2010 |
Elliott Hughes <enh@google.com> |
Make DirectBuffer.getEffectiveAddress cheaper.
DirectBuffer.getEffectiveAddress is used to implement the JNI function
GetDirectBufferAddress. Currently it uses reference types unnecessarily,
and we have an object pool to reduce the cost of this. Switch to using
int instead.
I'll remove the pool in a later change.
Note that this change requires a corresponding change to the
GetDirectBufferAddress implementation in dalvik:
https://android-git.corp.google.com/g/64328
Bug: 2935622
Change-Id: Ia0c20bd58c5355550f6af840feb08acff14fd4e9
|
| 8510524dab13e0acc1babf22cbc55002fb122777 |
|
17-Aug-2010 |
Elliott Hughes <enh@google.com> |
Some nio tidying.
Remove indirection on OSMemory, and use public ByteOrder and MapMode types
throughout the implementation, rather than adding our own duplication. Also
remove some unnecessary Java methods that just forwarded to identical native
methods, and an unnecessary factory class for MappedByteBuffer instances.
Also get rid of ICommonDataTypes, most of which was unused cruft.
Change-Id: I3240b02bcc19941b1b1cbba351ae7f7c1cdfc5b4
|
| 3ff29eb11a241d58c54ecc84230bb0672f4aa148 |
|
23-Apr-2010 |
Peter Hallam <peterhal@google.com> |
More merging of nio into luni
Change-Id: Ib4356b3cd57490df8ff246ea6a4bc1c269ba848b
|