Re: RFC: Generic Fixed-Point IIR Filter

[Jeffrey Carter <spam.jrcarter.not@spam.not.acm.org>]

On 03/26/2015 02:12 PM, Randy Brukardt wrote:
> 
> Right. The rule specifically does not allow something like:
>     A * B * C
> because that could take an infinite number of bits to evaluate (just keeping 
> adding the * operators) - the number of bits needed is the sum of the 
> maximum bits for each value. Universal-fixed is not supposed to overflow, 
> some something has to make it implementable. In addition, most machines have 
> a double precision multiply operation (like the 32*32 bits gives 64 bits 
> operation the OP mentioned and a matching divide). So there's an obvious 
> implementation in many cases (not on the margins, as I noted), so long as we 
> are talking about a single multiply. Ada runtime math isn't supposed to 
> require some complex infinite precision operation (which admittedly is 
> interesting, since the compiler is required to do exactly that).

That makes sense. The main reason for fixed point in Ada 83 was to be faster
than SW floating-point on processors without FPUs. But since, as you note, the
compiler has to be able to do unbounded-precision math, then the compiler could
link in the library it uses to do that and do its intermediate calculations
using it. There could be a compiler switch to determine which is done. I think
GNAT has this.

What would be even better would be for the language to make the spec of that
unbounded-precision package part of the standard library.

-- 
Jeff Carter
"Since I strongly believe that overpopulation is by
far the greatest problem in the world, this [Soylent
Green] would be my only message movie."
Charleton Heston
123