[squeak-dev] The Trunk: Collections-ul.668.mcz

Levente Uzonyi leves at elte.hu
Tue Oct 13 18:05:37 UTC 2015

What we can do is to use class (or species) checks. That way we can be 
consistent with the current hash implementation, and ensure symmetry, but 
different dictionaries holding the same pairs will not be equivalent. 
PluggableDictionary needs its own implementation of course.
And Sets can use the same kind of check as well.


On Tue, 13 Oct 2015, Chris Muller wrote:

>>> Exactly. It can happen with other kind of dictionaries as well. And class checks would not enough:
>>> d1 := PluggableDictionary new.
>>> d2 := PluggableDictionary new equalBlock: #==.
>>> d1 at: 'test' copy put: 4.
>>> d2 at: 'test' put: 4.
>>> {d1 = d2.
>>> d2 = d1}
>>> So, performance goes down in some cases, but I guess we agree on #= has to be symmetric.
> For simple value types, yes, we should be able to make that
> assumption, but symmetry of equivalence starts to break down for
> objects as complex as a Dictionary..
>> We can speed up some cases if the dictionaries can tell if #= is guaranteed to be symmetric
>> when they have the same class.
>> But #hash still uses species, that means, we can have a Dictionary and an IdentityDictionary to be equal, but with different hash.
>> s := 'a'
>> ed := Dictionary with: s -> $a.
>> id := IdentityDictionary with: s -> $a.
>> { ed = id. id = ed. ed hash = id hash } "->  #(true true false)"
> Equivalence of complex objects is always a "loose" definition, of
> which there can be several.  Additional API like #hasEqualElements:
> (of SequenceableCollection) could be added as needed to accomodate
> other application needs, but we only get ONE particular implementation
> which can also be suitable for use in HashedCollections, because they
> send #=.  Changing it should be really scrutinized.
> I'm wondering if Levente could get by with an app-specific equivalence
> check without needing it for use in HashedCllections?  Seems like a
> lot of the #= in the system require class or species equivalence, a
> few allow isKindOf:, but only one or two allow #is message response.
> I'm wondering now if those last kind are ill-advised...
> BTW, with this change, I guess we're now somewhat inconsistent with Set,
> | ed id |
> ed := Set with: 'a' copy.
> id := IdentitySet with: 'a'.
> { ed = id. id = ed }   " #(false true)"
> My production app is going to be punished by Levente's change,
> unnecessarily.  I think we should consider an alternative like adding
> #hasEqualElements: for your app...

More information about the Squeak-dev mailing list