Boost logo

Boost :

From: Eric Niebler (eric_at_[hidden])
Date: 2008-03-10 15:41:40

Larry Evans wrote:
> On 03/10/08 11:29, Eric Niebler wrote:
>> Steven Watanabe wrote:
>>> AMDG
>>> Markus Werle wrote:
> [snip]
>> The typelist used there is not Fusion-compatible, but expr<> as a whole
>> is. For a while, I was using mpl::vector<> as the second template
>> parameter to expr<>, but I changed to a very simple custom type because
>> it improved compile times.
> Could you give some estimate as to how much it improved
> compile times? Do you have handy a benchmark test to show
> this? I would like to try this benchmark with the variadic
> template compiler to see how close they are.

I switched away from mpl containers very early on. I don't have any of
the old code for you to test. I recall compile times improved by about
~8-10%. The issue is accessing the members. Going thru mpl::at_c<> to
get at each child forces an extra template instantiation. Ditto for
getting the length via mpl::size<>. These template instantiations add up
and cause a measurable perf hit.

> I know you've already tried the variadic compiler, but I'd
> like to try the benchmark myself and maybe show the results
> and code to Douglas Gregor so that maybe he can figure out
> why the variadic compiler is not improving compile times.
> As you noted in a private email to me, the basic
> reason is:
> Instantiating templates is expensive. By comparison,
> expanding a macro is cheap.
> However, maybe the benchmark would prompt Douglas to look
> for ways to improve the compiler's template instantiation
> speed.

I've already spoken with Doug about it. Variadics are great and do
improve compile times for variadic functions, esp. when combined with
rvalue refs. Variadic class templates are another matter. The only way
to metaprogram with them is with recursive template instantiations. That
means instantiating lots of templates. The preprocessor can improve the
situation by doing loop unrolling, but in the limit, the number of
template instantiations is going to be linear with the length of the
parameter pack.

If you want to play with the variadic Proto code, you can find it at
branches/proto/v3. I recently resurrected this branch because some
people on the std committee were interested in it. It's not production
code, though.

Eric Niebler
Boost Consulting

Boost list run by bdawes at, gregod at, cpdaniel at, john at