Re: Imitation is the sincerest form of flattery

[Robert A Duff <bobduff@shell01.TheWorld.com>]

"Frank J. Lhota" <NOSPAM.lhota.adarose@verizon.net> writes:

> If there is a way to get the effect of tagged scalar types without a big hit
> on efficiency, however, it would certainly be helpful.

Well, I think efficiency per se is irrelevant to language design.
What matters is distributed overhead -- you don't want to pay
an efficiency cost for features you don't use.

If a feature is useful, then presumably implementers will implement it
in the most efficient way that's reasonable.  It doesn't make sense to
leave out a feature purely because it is inherently slow -- if it's
useful, then programmers will have to implement it themselves, and it
will *still* be slow.  It makes sense to leave out a feature if it makes
*other* features slower than they would otherwise need to be.

So the question is whether a language can provide extensible scalars
with dispatching calls without slowing code that doesn't need that
power.  I think the answer is yes: don't store the tag with every
object.  Instead, gin it up when calling a class-wide operation.
So "+" on normal (non-class-wide) integers can be a single machine
instruction, and integers can be stored as a single word.  The compiler
can determine when to gin up a tag -- no need for explicit boxing
operations.

- Bob