On 19-Jun-06, at 1:15 AM, Per Bothner wrote:

> Marc Feeley wrote:
>> I repeat: in which implementations of Scheme will the multiple
>> values API be faster than the equivalent pair of operations which
>> return single values?
>
> Ones in which the operation can be inlined by a compiler.

Yes, that's what I meant, and specifically the fixnum operations.

> Of course inlining may be contrary to a strict reading of at least
> R5RS, unless you're doing whole-program compilation, which further
> reduces the value of standardizing these functions.

The module system in R6RS will (probably) forbid mutation of global
variables bound to predefined procedures, so this should not be a
problem.  In R5RS a compiler could use speculative inlining of the
predefined procedures (that's what Gambit-C does, I can email you a
paper on the subject if you are interested).

>
> > I find SRFI 77's complexity to be a real turn-off when I
>> consider the changes required to implement it properly.  Surely
>> I'm not the only implementer to feel this way.
>
> You're not.
>
> Though a tuned reference implementation that makes uses of a small
> number of implementation-optimized primitives would help.  But
> please note that in some implementations "efficient fixnums" may
> mean unboxed C/C++/Java values rather than general Scheme values.
> Some of us need to be able to classify at compile time whether
> a name or expression is a boxed object or an unboxed native number.
> Similar for floating point - that is even more common, I suspect.

I agree.

Marc