Re: Vector Extensions in GCC

[Mark Mitchell <mark at codesourcery dot com>]

> Do you have something more concrete?  Over a third of the C front end is
> now common code used also by C++, a lot more than in 2.95, and a lot more
> is sufficiently similar that it ought to be shared, which is some sort of
> incremental merge but hardly likely to yield a single front end in a
> reasonable amount of time.  What about Objective-C?

We are in discussions with people about doing this work, including 
Objective-C.  If that works out, it will happen in time measurable in 
quarters.  If not, then it will just remain my pipe-dream.

> Do you have an example of how templates would be used to implement one of
> the <tgmath.h> functions?

I forget the exact <tgmat.h> syntax/semantics, but here's what I had in 
mind.

  template <typename T> T acos (T);
  template <> double acos(double d) { return cos (d); }
  template <> float acos(float f) { return cosf (f); }
  template <> long double acos(long double l) { return cosl (l); }

If you built up this representation internally, things would just
work.  Using templates is a little cheesy here -- overloading would
work well.  But, if you had an algorithm that you wanted to encode
(for example, an algorithm for computing `acos', independent of the
precision), then you could omit the specializations, and just
let the compiler instantiate the generic method for the appropriate
types.

For example:

  template <typename T> T fmax (T t1, T t2)
  { return t1 ? t2 : t1 : t2; }

I am not sure what the right semantics are if `t1' and `t2' do
not have the same type.  You might still need specializations to
handle all the arithmetic promotions, etc.

So, you could be right that overloading is more what you want.

-- 
Mark Mitchell                mark@codesourcery.com
CodeSourcery, LLC            http://www.codesourcery.com