EncodingMap uses 0 as its empty value but 0 is a valid encoding ID (MacRoman), which is actually added to the map by buildDictionaries(). I think the empty value should be changed to something else (but not InvalidEncoding, which is used as the deleted value), and emptyValueIsZero should be defined as false.
Created attachment 8367 [details] A fix
Comment on attachment 8367 [details] A fix I'm really not sure about this one (except that it fixes the bug, of course).
Comment on attachment 8367 [details] A fix Actually, the specialization of the HashKeyStorageTraits can and should be removed entirely if we have to change the empty value. The reason that's there is to make the hash table implementation get shared. But that sharing can't take place if the traits are not the standard traits for the integer type. Otherwise, this patch looks fine.
This second invalid encoding should probably go in the TextEncoding.h header and have a symbolic name, too. I agree that the value 0xFFFFFFFEU is a good value to use. We can probably call it InvalidEncoding2 or some other lovely name like that.
Created attachment 8391 [details] Patch w/change log InvalidEncoding2 it is!
Comment on attachment 8391 [details] Patch w/change log Looks great, r=me. One comment for the future. When you have an assert that has an && in it, it's usually better to have two asserts. That way each one has a smaller expression in it, and also if one fails you know which half failed.
Committed revision 14465.