I'm very interested in patching MPL to be able to use variadic templates.

There are many C++ libraries and code that use MPL. Those libraries can not switch from MPL to Metal or Hana because: * they have to be usable in C++98/C++03 and C++11 * they have a huge code base that can not be simply migrated to new library * they depend on third party libraries that use MPL and need the ability to interact with those libraries

Moreover, making MPL variadic will not just speed up compilation, but also will: * reduce executable size and improve startup times if MPL structure is part of an exported entity * improve runtime speed if MPL structure is part of an entity that is used in operations involving typeid()

So please, continue the work and make a pull request.

-- Best regards, Antony Polukhin

For what it's worth, I very much agree with all of the above. -- chris

On 03/01/2017 05:07 PM, Larry Evans via Boost wrote:

On 03/01/2017 12:56 PM, Antony Polukhin via Boost wrote:

...

2017-02-21 13:52 GMT+03:00 Михаил Максимов via Boost :

...

Dear community,

I've recently started contributing to boost::variant. To speed up variant's compilation I'm implementing mpl::vector on variadic templates. For now status of new vector implementation is: <...> made me wonder, is there interest in variadic templates implementation? Are there limitations for it's usefulness, which I did not see?

I'm very interested in patching MPL to be able to use variadic templates.

There was a variadic template mpl try from 7 years ago:

https://github.com/cppljevans/variadic_templates/tree/master/boost/mpl

Maybe it could give you some ideas. OTOH, maybe it's *too* old :(

One problem with the above cppljevans mpl is there's no at.hpp. Instead, the non-variadic boost/mpl/at.hpp was used. The reason no variadic at was created was because, AFAICT, there was no non-recursive method for picking the I-th element from T... , and, IIUC, recursive templates cause compile-time slow downs. Since then, I've stumbled across other methods as shown in the attached (where, type_at is used instead of at). In addition there have been language proposals to do essentially the same: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0535r0.html#paramet... http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4235.htm http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0327r1.pdf I don't know of any compiler's that implement those "tuple slicing" operations; however, if someone does know of some, I'd like to see their use as an alternative to the existing mpl::at. Mike, if you find such an tuple-slicing compiler, maybe in your variadic mpl implementation you could provide a flag which uses that or one of the methods in the type_at.hpp attachment. Then, you could profile them to demonstrate which is best :) -regards, Larry

On 03/02/2017 11:27 AM, Peter Dimov via Boost wrote:

Larry Evans wrote:

...

One problem with the above cppljevans mpl is there's no at.hpp. Instead, the non-variadic boost/mpl/at.hpp was used. The reason no variadic at was created was because, AFAICT, there was no non-recursive method for picking the I-th element from T... , and, IIUC, recursive templates cause compile-time slow downs.

Have you read

http://pdimov.com/cpp2/simple_cxx11_metaprogramming_2.html

?

Nope. Thanks *very much* for the link. I'm impressed (especially with the way you actually cited the parts of the standard to guide your search for the best method!).

Search for mp_at.

Thanks for that tip. I'm a bit surprised that the mp_map_from_list was fastest. I would have thought that large template classes mean slow compile times, but I guess not. Thanks again, Peter. -regards, Larry

On Thu, Mar 2, 2017 at 7:42 PM, Larry Evans via Boost

wrote:

On 03/02/2017 11:54 AM, Larry Evans via Boost wrote:

...

On 03/02/2017 11:27 AM, Peter Dimov via Boost wrote:

...

Larry Evans wrote:

One problem with the above cppljevans mpl is there's no at.hpp.

...

Instead, the non-variadic boost/mpl/at.hpp was used. The reason no variadic at was created was because, AFAICT, there was no non-recursive method for picking the I-th element from T... , and, IIUC, recursive templates cause compile-time slow downs.

Have you read

http://pdimov.com/cpp2/simple_cxx11_metaprogramming_2.html

?

Nope. Thanks *very much* for the link. I'm impressed (especially with the way you actually cited the parts of the standard to guide your search for the best method!).

...

Search for mp_at.

Thanks for that tip.

I'm a bit surprised that the mp_map_from_list was fastest. I would have thought that large template classes mean slow compile times, but I guess not.

Since you seem to be interested on benchmarking compilation times, check out http://metaben.ch.

This:

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4235.htm

claims:

template instantiations use memory that is never freed during the compilation process

so, maybe mp_map_from_list should be used with caution. If many mp_map_from_list instantiations are used, I'm guessing the compiler might become pressed for memory. Maybe a benchmark showing compile time vs number of instantiations would show this. For example, instead of N=800, see what happens when 2 instantiations with N=400 happen, and compare with a similar test for mp_repeat_c.

This is actually not the case, in fact the exact opposite is true. It turns out mp_map_from_list is actually very cheap to the compiler in terms of memory allocation, because it takes advantage of template memoization, so if you retrieve every single element of a list using this trick, the compiler needs to instantiate it only once. On the other hand, the trick employing void pointers can't be expressed in a way that take advantage of memoization, which means that, for every element retrieved from the list, the entire pack of void pointers must be instantiated again. I'm not just guessing here, in fact, benchmarks for the algorithm `at` on metaben.ch prove it by plotting the time it takes to compile a metaprogram that retrieves every element in a list as a function of its size, from 0 up to 500 elements. Brigand and Meta (and likely also Hana, but I'm not sure) rely on the void pointers trick, as you can verify in their source codes, and as you can see on metaben.ch their benchmarks appear to vary with the square of the list size. This is precisely because the entire variadic pack of void pointers must be generated for each of the N elements retrieved. On the other hand, the fastest implementation on all compilers(*) is that of Metal, which seems to vary approximately linearly with the list size. That is because it relies on a variant of the trick to retrieve an element from a map and was carefully designed to take advantage of memoization, as you can verify here https://github.com/brunocodutra/metal/blob/master/include/metal/detail/looku... (*) Specifically on Clang, Metal actually relies on a compiler intrinsic called __type_pack_element to get the N-th type in a variadic pack. This is arguably as fast as `at` can possibly get and still it is only slightly faster than the implementation on GCC, which does not rely on any compiler intrinsic. Bruno

On Fri, Mar 3, 2017 at 2:46 PM, Larry Evans via Boost

wrote:

On 03/02/2017 01:54 PM, Bruno Dutra via Boost wrote:

...

On Thu, Mar 2, 2017 at 7:42 PM, Larry Evans via Boost < boost@lists.boost.org

...

wrote:

On 03/02/2017 11:54 AM, Larry Evans via Boost wrote:

...

On 03/02/2017 11:27 AM, Peter Dimov via Boost wrote:

...

Larry Evans wrote:

...

One problem with the above cppljevans mpl is there's no at.hpp.

...

Instead, the non-variadic boost/mpl/at.hpp was used. The reason no variadic at was created was because, AFAICT, there was no non-recursive method for picking the I-th element from T... , and, IIUC, recursive templates cause compile-time slow downs.

Have you read

http://pdimov.com/cpp2/simple_cxx11_metaprogramming_2.html

?

Nope. Thanks *very much* for the link. I'm impressed (especially with the way you actually cited the parts of the standard to guide your search for the best method!).

Search for mp_at.

...

Thanks for that tip.

I'm a bit surprised that the mp_map_from_list was fastest. I would have thought that large template classes mean slow compile times, but I guess not.

Since you seem to be interested on benchmarking compilation times, check

out http://metaben.ch.

The at benchmark there indicates mpl.vector is the fastest in the, I guess you'd call it, the "performance at given number_of_elements pop-up list", when the cursor is moved over to the rhs at number_of_elements=500. Hmmm. Now I see there's actually no mpl.vector graph at that number_of_elements. OOPS. I see that that's because mpl.vector was not run for number_of_elements > 50, but that's only obvious when the subtract_baseline radio button is turned-off. To avoid this confusion, the graph should give some indication, in the "performance at given number_of_elements pop-up list", which methods were actually run at the given number_of_elements.

This is actually a known feature of NVD3, which is the tool we use to plot data, see: https://github.com/ldionne/metabench/issues/158 I agree it is not very intuitive the way it is, but it is not obvious how to fix it without hacking NVD3. I appreciate thoughts on that.

This:

...

...

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4235.htm

claims:

template instantiations use memory that is never freed during the compilation process

so, maybe mp_map_from_list should be used with caution. If many mp_map_from_list instantiations are used, I'm guessing the compiler might become pressed for memory. Maybe a benchmark showing compile time vs number of instantiations would show this. For example, instead of N=800, see what happens when 2 instantiations with N=400 happen, and compare with a similar test for mp_repeat_c.

This is actually not the case, in fact the exact opposite is true.

So n4235.htm was wrong or maybe I've misinterpreted it?

It turns

...

out mp_map_from_list is actually very cheap to the compiler in terms of memory allocation, because it takes advantage of template memoization, so if you retrieve every single element of a list using this trick, the compiler needs to instantiate it only once. On the other hand, the trick employing void pointers can't be expressed in a way that take advantage of memoization, which means that, for every element retrieved from the list, the entire pack of void pointers must be instantiated again.

So how do you resolve what n4235 is saying and what you say above?

I'm guessing that 4235 was actually referring to this memoization when it said:

template instantiations use memory that is never freed

The memoizations do use memory which, IIUC, is never freed until the compiler is finished. OTOH, when you say:

...

for every element retrieved from the list, the entire pack of void pointers must be instantiated again.

That means more compile time is needed for the repeated instantiations; however, no memory is *permanently* used as is the case for memoization. IOW, this is a classic case of trading memory for speed:

https://en.wikipedia.org/wiki/Space%E2%80%93time_tradeoff

OTOH, my point was that as these memoization's increase, the compiler has less memory to use; hence, might slow down; whereas, since, the memory used by the vec_of_void* method is reused, memory pressure is less, and, although it's slower for 1 instance of a pack, it might become faster as many packs are used. http://lists.boost.org/mailman/listinfo.cgi/boost

I believe you've misinterpreted n4235 and/or what I was trying to say. Memoization is not something the programmer can tell the compiler to do or not, compilers just do it, because that is what they are made for: instantiate templates and record information about them so that they can be used elsewhere in the program where they happen to be visible, so it is true that memory is never freed once a template is instantiated, because in general it must not. Now let us take a look at the problem at hand from this point of view. Using the type-map idea, for a given list of N elements, the compiler instantiates the *same* type-map of N key-value pairs whenever any element from this list is retrieved. Since information about this type is memoized, the compiler doesn't need to repeatedly instantiate it, it can simply look it up. This means a single extra template instantiation is needed to retrieve all N elements of a list. OTOH the void* trick requires that a pack of M void pointers be instantiated when the element at position M is retrieved from a list, so, for a list of N elements, the compiler must instantiate N *distinct* packs of void pointers, because each of these packs have a different size and is thus a distinct type. This means the compiler does not benefit from memoization at all. Also remember that a bunch of intermediary instantiations are required in order to generate a pack of void pointers, even if we use the logarithmic trick, so we end up forcing the compiler to store a bunch of useless information about all these dummy types. That is what compromises compilation times when using the void pointer trick. Bruno

On Wed, 2017-03-01 at 21:56 +0300, Antony Polukhin via Boost wrote:

2017-02-21 13:52 GMT+03:00 Михаил Максимов via Boost :

...

Dear community,

I've recently started contributing to boost::variant. To speed up variant's compilation I'm implementing mpl::vector on variadic templates. For now status of new vector implementation is:

<...>

...

made me wonder, is there interest in variadic templates implementation? Are there limitations for it's usefulness, which I did not see?

I'm very interested in patching MPL to be able to use variadic templates.

There are many C++ libraries and code that use MPL. Those libraries can not switch from MPL to Metal or Hana because: * they have to be usable in C++98/C++03 and C++11 * they have a huge code base that can not be simply migrated to new library * they depend on third party libraries that use MPL and need the ability to interact with those libraries

Moreover, making MPL variadic will not just speed up compilation, but also will: * reduce executable size and improve startup times if MPL structure is part of an exported entity * improve runtime speed if MPL structure is part of an entity that is used in operations involving typeid()

So please, continue the work and make a pull request.

I wonder if instead an MPL implementation could be written with a Metal backend for newer compilers and use the legacy MPL for C++03 compiler. I thought Bruno was developing something like that at one point. So building off of that could be useful. Although, Metal is not in boost yet, hopefully it will be soon. Paul