Re: [boost] [proto] transform for flattening an expression - how?

From: Eric Niebler (eric_at_[hidden])
Date: 2008-03-18 11:25:49

Markus Werle wrote:
> Eric Niebler <eric <at> boost-consulting.com> writes:
>> The reason I ask is because if you use expr<>, 
>> you are limited to BOOST_PROTO_MAX_ARITY elements, which is probably too 
>> low. And bumping BOOST_PROTO_MAX_ARITY gets expensive.
> 
> Alarm level red: You have measurements: what is the implication
> of setting BOOST_PROTO_MAX_ARITY to values greater than 10
> considering compile time?

Here are the compile times for libs/xpressive/test/test2.cpp, a fairly 
rigorous workout for Proto, for different max arities.

Arity	Compile time (sec)
-------------------------
5	5.132
10	5.725
15	6.895
20	8.517
25	11.091
30	14.492
35	18.922
40	25.100

As you can see, beyond 20 things begin to get bad, but not dramatically 
so. One of the big sources of slow-down is the operator() overloads on 
expr<> ... there are N specializations of expr<>, and each one has N 
overloads of operator(). That's O(N^2) growth. With a variadic 
operator(), I should be able to knock that down to O(N).

In picking a max arity for a DSEL, I think we reach the limits of human 
readability (a function that takes 20 arguments?! Really?) before we 
reach Proto's breaking point.

-- 
Eric Niebler
Boost Consulting
www.boost-consulting.com