<AWT Dev> [9] Review Request: 8027148 SystemFlavorMap.getNativesForFlavor returns list of native formats in incorrect order

[Petr Pchelko]

Hello, AWT Team.

Please review the fix for the issue:
https://bugs.openjdk.java.net/browse/JDK-8027148
The fix is available at:
http://cr.openjdk.java.net/~pchelko/9/8027148/webrev/

Sorry for the long text, but it’s quite tangled)

The problem:
The flavor map contains some predefined mappings which are stored in the flavormap.properties file. But we can extend these mappings using 
addUnencodedNativeForFlavor method. Javadoc states, that these new mappings would have lower priority that standard mappings. But in the current implementation this was not the case, because getNativesForFlavor method relied on the fact, that standard text mappings were stored as FlavorBaseType<->Native and newly added mappings were stored as DataFlavor<->Native, but after some fix in Java 8 this is not the case any more. Everything is stored as a DataFlavor as a key. This is important only for text flavors, because we support different text charsets and can reencode the text on the fly. So each native text format could be represented in many different DataFlavors with different encodings and representation classes. When we generate the set of DataFlavor’s that a text native can be translated to we no longer know how to distinguish the standard mappings and additional mappings and the get shuffled when we generate missing mappings for text formats.

The solution:
I’ve added an additional map for standard text mappings. With this map we can now take natives for mime types directly and not "find all flavors for a mime-type and than all natives for each flavor". This is not only faster, but we can distinguish standard text mappings from custom and return the list in the correct order. The new hash map contains only a few elements.

Also I’ve replaced the ArrayList as a collection of natives for a particular Flavor with a LinkedHashSet, because this list could not contain duplicated which we were enforcing ourselves. Now it works out of the box and some code can be removed.

I’ve measured the performance of a couple of most hot methods and on average the new implementation is 1.7 times faster. 

The test is being open sources.

I’ve tested this with JCK and our regression tests, everything looks good. Also I’ve tested with a couple of hand-made toys.

Thank you.
With best regards. Petr.