Re: [boost] math/tools/rational.hpp gcc 3.2 problem

Subject: Re: [boost] math/tools/rational.hpp gcc 3.2 problem
From: John Maddock (john_at_[hidden])
Date: 2009-10-14 10:18:12

>> gcc 3.2 doesn't compile this code (works with gcc 3.3 and up):
>>
>> #include <boost/math/distributions.hpp>
>> int
>> main()
>> {
>>  boost::math::students_t_distribution<double> dist(1.);
>>  dist.find_degrees_of_freedom(1.,2.,3.,4.,5.);
>>  return 0;
>> }
>>
>>
>>
>> Reduced:
>>
>> template <unsigned N, class T, class V>
>> inline V evaluate_polynomial(const T(&a)[N], const V& val)
>>
>
> I will be wrong linguistically, but I hate the definition of 'a' here. 
> What
> is it to be? I assume it is a reference to an array, I guess. But then why
> not just take a reference to the first element and a count, or use a range
> (pointer + size)?
>
> Sorry, I don't know anything about boost/math/distributions. But this kind
> of non-problem being a problem is in fact, a problem.
>
> If 'a' is just an array, say it like it is. Why make a fuss?

Because it makes a difference, both for maintenance and performance:

* For maintenance, we're much less likely to introduce subtle bugs if the 
function "knows" how big the array is, rather than us having to specify the 
size at the call site, as well as at the definition of the array.
* For performance: for small arrays of known size we can omit the for-loop 
and just evaluate the polynomial using near-optimal code.  Since 
polynomial/rational function evaluation is often the pinch-point for 
functions implemented using rational-approximations it makes sense to 
optimize this as far as possible even if it only gains us 20-30%.

We could of course code the polynomial evaluation separately each time it's 
needed - and without a separate function - that's what most other similar 
libraries do, but as well as being untidy, it precludes library wide 
improvements to this core function, unless of course you change every usage 
:-(

As for the original problem: I'm afraid I don't have an obvious workaround 
that springs to mind, I will try and install an early gcc version to test 
with, but last time I tried you can't even *build* gcc-3.x on recent Linux 
builds :-(

John.