[squeak-dev] The defaullt implementation of isEmpty might do too much work

Chris Muller asqueaker at gmail.com
Tue Oct 25 03:18:15 UTC 2016


Since Smalltalk-80, sends to #do: have been considered expensive, and
sends to #size, inexpensive.  And so it has been okay for applications
which defined their own Collection subclasses, to add a few
milliseconds of cost to something that's already considered expensive
(#do:), and not to something which is expected to be fast (#size).

** For 30 years applications were built on these assumptions.  Even I
did it in Magma.  #isEmpty was built on that assumption.  Subclass
implementors have seemed to agree to these assumptions too, and made
their implementations true to them..

May we please reconsider this change?


On Mon, Oct 24, 2016 at 9:07 PM, Chris Muller <asqueaker at gmail.com> wrote:
> Normally its the Subclasses where implementation-specific
> optimizations are done, not in the top abstractions.  At the level of
> Collection, the code should be elegant, expressive, and defining of
> the responsibilities by the messages it sends.  #size is something of
> Collection, period, there is nothing wrong with sending it, and
> checking size = 0 is more elegant than a block activation with
> non-local return.
>
> As a community that respects its user base, we need to consider the
> *type* of damage that "Change it!" could cause to existing production
> applications.  These applications were built on the expressive
> implementation, but now they'll be sending #do: to that SQLCollection
> instead of #size, which will cause the database to retrieve a full
> ResultSet, send it back to the client, just so it can answer #isEmpty.
>
> Its this kind of "silent" damage that is the worst.  Because its
> invisible, not a an error, not a blow up.  Just an degradation in
> performance (ironic, given our motive here) which, believably, would
> not be caught in testing.
>
> I agree with Monty.  Collection is abstract, there will never be an
> instance of it.  No self-respecting subclass needs this optimized, and
> trying to optimize it way up there forces too many assumptions about
> implementation.  It makes the code less expressive while silently
> inflicting potentially performance-killing pain onto legacy
> applications until they're forced to sprinkle duplicate #isEmpty
> implementations all about..
>
> On Mon, Oct 24, 2016 at 3:31 AM, Bert Freudenberg <bert at freudenbergs.de> wrote:
>> On Monday, 24 October 2016, monty <monty2 at programmer.net> wrote:
>>>
>>> Why not just make #size a subclassResponsibility or add abstract
>>> superclasses for lazy or infinite collections that implement #isEmpty using
>>> #do: and change #size to shouldNotImplement?
>>
>>
>> Because #do: is the only method required to be overridden by Collection
>> subclasses. All other methods are implemented in terms of #do:.
>>
>> So yes, if your Collection subclass has an optimized implementation for
>> #isEmpty, then provide it. That is a small price to pay for making a class
>> work optimally across different code bases.
>>
>> - Bert -
>>
>>
>>


More information about the Squeak-dev mailing list