Boost logo

Proto :

Subject: Re: [proto] van Wijngaarden grammars
From: joel falcou (joel.falcou_at_[hidden])
Date: 2010-07-18 14:08:18

Eric Niebler wrote:
> Is it that a vW evaluator is free to try *any* rule in the grammar in
> place of the T? If so, that seems FAR too powerful to be useful. Indeed
> the wikipedia entry says as much. Instead, it points to CDL
> ( and affix
> grammars as more practical for defining programming languages.
It exactly does that: trues any rules in place of T
> What is the "parameter" of a grammar? Do you mean the expression against
> which you're evaluating the grammar?
The grammar parameter is the "type" rule in the vW grammar
>> so i have a grammar:
>> template<class T> struct simd_grammar : proto::or_< ... > {};
> OK. What is T?
Any type that fit in a SIMD terminal ...
>> and a meta-function:
>> template<class T> struct is_vectorizable;
> What is T, and how is this implemented?
and actually matches this meta-function.

is_vectorizable itself is defined so it check that on current platform,
T has a SIMD equivalent.

> Maybe once I understand what you're doing we can find a better way.
> I already know how difficult it is to, for instance, ensure that all
> terminals in an expression have the same type. That's the problem the
> poster who suggested vW grammars had. It doesn't sound like your
> technique can help in instances like this.
Here is the simd::grammar<T> grammar:

namespace nt2{ namespace simd
    namespace bp=boost::proto;

    // grammar is a template class to avoid a complex, recursive type
    // to ensure that SIMD expression of different native source are not
    template<class T, class Cardinal> struct grammar
      : bp::or_< bp::and_< bp::terminal< data<T,Cardinal> >
                                      , bp::if_<
meta::is_vectorizable<T,Cardinal>() >
                     , meta::lambda_term< boost::is_arithmetic >
                     , bp::nary_expr< bp::_, bp::vararg<
grammar<T,CArdinal> > >
> {};
} }

simd::vector<T,Cardinal> then inherit from an expression based on
grammar<T,Cardinal> in a proper subdomain.
Then vector<float,4> can be mixed together but vector<T,4> +
vector<char,16> can't

We are sure all terminal have the same type cause we can't build
expression in a way they will not.

Proto list run by eric at