- Boost - lists.boost.org

Re: [boost] [type_erasure] Review started (July 18-27, 2012)
by Steven Watanabe 20 Aug '12

20 Aug '12

AMDG On 08/02/2012 12:29 PM, Larry Evans wrote: > On 07/18/12 00:13, Lorenzo Caminiti wrote: >> Hello all, >> >> *** The review of Steven Watanabe's proposed Boost.TypeErasure library >> begins on July 18, 2012 and ends on July 27, 2012. *** > [snip] > Should type_erasure be accepted into Boost? > > Answer: > No, not yet. > Why(in order of importance): > > 1) any<C,T>::any(U&, static_binding<Map>&) bug > > The bug reported here: > http://article.gmane.org/gmane.comp.lib.boost.devel/233101 > should be resolved. > > The bug is just too dangerous for the resolution to be delayed. > I added a static assert. > <snip> > > My review: > > 1. What is your evaluation of the design? > > Good, except for: > > [need variadic static_binding ctor]: > > The any CTOR: > > template<class U, class Map> > any(U & arg, const static_binding< Map > & binding); > > documented here: > > > http://steven_watanabe.users.sourceforge.net/type_erasure/libs/type_erasure… > > allows only a single argument to the constructor to the > underlying type. I'd suggest adding another CTOR with > interface: > > template<class Map, class... U> > explicit any(const static_binding< Map > & binding, U &&... arg); > > which is similar to: > > explicit any(const binding< Concept > & binding, U &&... arg); > > Then any constructor of the underlying could be used. > I think something like this is reasonable. I have to consider whether this is best approach, though. I don't consider this high priority, however, since the copy constructor should work most of the time. The STL got by just fine for years without emplace. > [too many any ctors]: > > According to `grep -e ' any('`, there are more than 45 any > constructors. That many makes it hard for the novice to know > which one to use. The Design notes: > The solution is simple. Don't use the source. Use the documentation page for any, which only shows 9 distinct constructors. > > http://steven_watanabe.users.sourceforge.net/type_erasure/libs/type_erasure… > > are some help, but those notes also mention ambiguities and > user rules, which sound like indications of usability problems > to me. > > Why wouldn't the following ctors be sufficent to do all the > current any constructors do (except for maybe requiring more > typing and maybe more constructor calls)? > > template > < class... U > > > any > ( static_binding<Map>const& > , U&&... > );//kind=binding constructor > > template > < class... U > > > any > ( binding<Concept>const& > , U&&... > );//kind=dispatching constructor > > template > < class Concept2 > , class Tag2 > > > any > ( any<Concept2,Tag2>U&& > , static_binding<Map>const& > );//kind=converting constructor > > template > < class Concept2 > , class Tag2 > > > any > ( any<Concept2,Tag2>U&& u > , binding<Concept2>const&b=binding_of(u) > );//kind=converting constructor > > Providing just these 4 constructors would improve the usabilty. Why? > Of course, I'm probably missing something, and if so, > explaining why the above would not work in the Design notes > would be helpful. > I think it should be obvious what's missing: a) any(const any&); The copy constructor is not a template. b) template<class U> any(const U&); This is the most commonly used constructor besides the copy constructor. It would be very inconvenient if it didn't exist. c) any(); The default constructor creates an empty any. d) A default argument can't refer to a prior argument. Your fourth constructor has to be two separate constructors. > 2. What is your evaluation of the implementation? > > [in-source comments]: > Very few in-source comments, as for example, the in the code: > > > http://steven_watanabe.users.sourceforge.net/type_erasure/boost/type_erasur… > > there are no in-source comments. This lack of in-source comments > made it very hard to understand the implementation. There is also > no tests in the directory: > > > > http://steven_watanabe.users.sourceforge.net/type_erasure/libs/type_erasure… > > which contain any string with 'normalize' in it; hence, I assume > there are no unit tests for any class or function with normalize > as part of its name. Such unit tests, hopefully, would have > provided more information of the purpose of the code in > */detail/normalize.hpp; hence, I'd recommend providing such tests > in addition to more in-source comments. > The tests only cover the public interface. I'm not going to add specific tests of the implementation details. > > 3. What is your evaluation of the documentation? > > Mostly good, but some documentation needs improvement. > For example: > > [vague:rebinding_any.html]: > > This doc: > > > http://steven_watanabe.users.sourceforge.net/type_erasure/libs/type_erasure… > > Claims rebind_any<Any,T>::type: > > returns any type corresponding to a placeholder. > > However, it doesn't say what the Any in rebind_any<Any,T> may be, > and if one assumes Any is a type_erasure::any<C,S> for some > concept, C, and placeholder, S, then that description makes no > sense because C does not contain any information about which > actual type is matched to which placeholder. Huh? As the documentation you quoted says, rebind_any returns a specialization of /any/. rebind_any<any<C, S>, T>::type is any<C, T>. > <snip> > > from Steven, in the reply: > > http://article.gmane.org/gmane.comp.lib.boost.devel/233067 > > did I finally conclude: > > The placeholders in concepts do *not* stand for the bound types > but instead for the any's containing those bound types. > That depends on the context. boost::type_erasure::call "unwraps" these any's. > Which is similar to the conclusion which I posted here: > > http://article.gmane.org/gmane.comp.lib.boost.devel/233078 > > Since the 233078 post got no replies, I assume that conclusion is > right. If so, then that should be made clear in the > documentation (including tutorials). In addition, the following > example code (and especially the comments), should reduce future > confusion on this topic: > > /* {***any_ctor_kinds.cpp*** */ > <snip> > I find this code almost completely unreadable. > If that conclusion is wrong, then the docs need to be modified > (I've no idea how) to clearly define "matching" as used in the > any.html#id2445089-bb reference. > > [incomplete:tuple.html]: > > The doc: > > > http://steven_watanabe.users.sourceforge.net/type_erasure/libs/type_erasure… > > makes no mention of how the binding produced by this call are > generated. > Already fixed. In Christ, Steven Watanabe

10 24

[multiprecision] Some rvalue reference experiments (performance inhancements #1)
by John Maddock 20 Aug '12

20 Aug '12

Folks, I've started to work through some of the points that came up in the multiprecision review, starting with further rvalue reference support. The idea (due to Marc Glisse) was to add operator overloads that look something like: Number&& operator+(Number&&a, const Number& b) { return static_cast<Number&&>(a += b); } Using my Polynomial Horner evaluation scheme as an example the number of temporaries dropped as follows: Type # temporaries No expression templates, no rvalue refs 24 No expression templates, rvalue refs 1 Expression template 0 Which is pretty impressive, but not necessarily typical usage, so by way of "real world" tests, I compared the times taken to evaluate Boost.Math's Bessel function test cases to 50 decimal places: Library Time/s # Allocations MPFR/mpreal (no ET's or rvalue refs) 9.4 13226029 MPFR/mpfr_class (ET's) 6.3 3948316 MPFR/mp_number(ET's) 6.0 2683646 GMP/mp_number(no ET's) 5.2 4186977 GMP/mp_number(ET's) 5.0 2594441 As you can see the number of allocations required when using mp_number with ET's turned off is almost double that without, I guess this is because many statements are similar to: a = b op c; Where a temporary is still required when relying on rvalue refs, but not for expression templates. Having said that the times are getting much closer now.... John.

8 25

Is there any interest in getting pseudo random number generation using preprocessor at compile time into the Preprocessor library?
by Alexander Stoyan 20 Aug '12

20 Aug '12

The complete code, approach and efficiency proofs I described in my blog: http://alexander-stoyan.blogspot.com/2012/07/getting-pseudo-random-numbers-a t.html Need your opinion regarding the relevance and the approach correctness of the feature. Thank you in advance.

4 7

[1.51.0] release candidates available
by Marshall Clow 19 Aug '12

19 Aug '12

Release candidate files for 1.51.0 are available at http://boost.cowic.de/rc/ At this point, the release branch is closed for changes without explicit permission from the release managers. As always, the release managers would appreciate it if you download the candidate of your choice and give building it a try. Please report both success and failure, and anything else that is noteworthy. This helps ensure the candidates build OK before we push them out to SourceForge. The files (and associated md5s) are: 6a1f32d902203ac70fbec78af95b3cf8 *boost_1_51_0_rc1.tar.gz 4b6bd483b692fd138aef84ed2c8eb679 *boost_1_51_0_rc1.tar.bz2 ee8112e48088b05c248d68329cd5d908 *boost_1_51_0_rc1.zip e87d3747966ab7e0c9487884326d7d17 *boost_1_51_0_rc1.7z Thanks, --The release managers

5 6

Modularized Boost: How the files in libs/libXX must be included?
by Vicente J. Botet Escriba 19 Aug '12

19 Aug '12

Hi, the following ticket https://svn.boost.org/trac/boost/ticket/7200 report an issue while building Boost.Thread using the modularized cmake. What is the correct way to include the files in lib/libXXX? Best, Vicente

2 1

[type_traits][parameter] Inconsistent boost::is_convertible between gcc and clang
by John Bytheway 19 Aug '12

19 Aug '12

I have been tracking down an issue with some of my code which compiles fine with gcc but not clang. It manifested as a failure in Boost.Parameter but I eventually pinned it down to an inconsistency in the behaviour of boost::is_convertible. The following program demonstrates the issue: ============== #include <boost/type_traits/is_convertible.hpp> #include <boost/static_assert.hpp> class A {}; int main() { BOOST_STATIC_ASSERT((boost::is_convertible<A, A&>::value)); } ============== This compiles fine under g++ (4.7.1), but the assertion fires under clang++ (r160940). This is not too surprising because is_convertible is implemented in completely different ways on the two compilers: - on clang it is in terms of the __is_convertible intrinsic, which behaves like std::is_convertible and fails because rvalue references don't convert to lvalue references. - on gcc it has a complex custom implementation which at some point performs add_reference<> on the first argument and thus determines that the conversion is OK. Reading the docs for boost::is_convertible I think the clang interpretation is closer to the intention, in which case the gcc implementation is wrong, and so is Boost.Parameter (which relies on this behaviour). On the other hand, I suspect there's probably other code out there that depends on this in a similar way to Boost.Parameter, so perhaps it would be safer to tweak the clang implementation instead, and deviate from std::is_convertible. Any thoughts? John Bytheway

5 13

[odeint] Iterator semantics
by Karsten Ahnert 19 Aug '12

19 Aug '12

Hi, we have posted a review request for odeint (a library for solving ordinary differential equations) on this list some days ago. An ODE is dx/dt = f(x(t)) and one is looking for the solution x(t) which is a function of t. Solving ODEs is usually an iterative task, so one starts with x(0) and iteratively applies a solver to obtain a discretized (and approximated) representation of the solution: x(0) -> x(dt) -> x(2*dt) -> x(3*dt) -> ... I played around with an iterator which is doing exactly this iterative procedure. For example, it can be used via odeint::runge_kutta4< state_type > stepper; // ode solver std::array< double , 3 > x = {{ 10.0 , 10.0 , 10.0 }}; // initial state double res = std::accumulate( // parameters : solver , ode , state , time , time_step make_const_step_iterator_begin(stepper, lorenz(), x, 0.0, 0.01) , make_const_step_iterator_end( stepper, lorenz(), x, 1.0, 0.01) , 0.0 , []( double sum, const state_type &x) { return sum + x[0]; } ); The first iterator increments the time until the time of the second iterator is rearched t=0.0 -> t=0.01 -> t=0.02 -> .. t=1.0 and the first component of the solution is accumulated. In principle, this approach works quite well, but there are some problems at least semantically. Maybe someone here has some comments or ideas on this: The defintion of this iterator is: template< class Stepper , class System , class StepperTag > class const_step_iterator : public boost::iterator_facade < const_step_iterator< Stepper , System , StepperTag > , typename Stepper::state_type const , boost::single_pass_traversal_tag > { ... }; It is a single pass iterator, so it can only check for equality of two iterators. This is problematic since equality here means that the time (t) of the begin iterator is smaller then the time of the end iterator. To ensure commutativity of the != or == operation I needed to flag the begin and end iterator explicitly. This is the reason for the two factory functions make_const_step_iterator_begin and make_const_step_iterator_end, which are doing this flagging. The second semantically problem is that the end iterator in principle does not need to know the stepper as well as the system (lorenz() in the above example). But all algorithms from the standard library and Boost.Range assume that the begin and end iterator are of same type. Therefore you have to put the stepper and the system into the end iterator too. Of course, there are techniques like type erasure which might solve this problem, but they introduce some run-time overhead which we want to avoid here. Any ideas or comments about this? P.S. The code can be found in the iterator branch in odeint: boost/numeric/odeint/iterator/const_step_iterator.hpp and some examples libs/numeric/odeint/examples/const_step_iterator.cpp git clone -b iterator git@github.com:headmyshoulder/odeint-v2.git

5 19

Request for a typedef for the larger integer type that supports atomic loads and stores.
by Vicente J. Botet Escriba 18 Aug '12

18 Aug '12

Hi, I'm locking for a portable integer type that is the larger integer type that supports atomic loads and stores. The sig_atomic_t type defined in <csignal> supports atomic loads and stores but I'm not sure is the larger one. Does boost define such a typedef? If not, could such typedef be included in boost and where? Best, Vicente

2 1

What to do with platform defining 'null' as preprocessor symbol?
by Jan Hudec 17 Aug '12

17 Aug '12

Hello Folks, At work we develop, among others, for Bada. The system library provided is unfortunately somewhat flaky. A tell-tale sign is, that it's native API headers define lowercase 'null' as preprocessor symbol expanding to 0. Unfortunately boost uses identifier 'null'. So far we only hit the two cases in boost/test/utils/basic_cstring/basic_cstring.hpp, where I fixed it by renaming the symbols in our local copy of that header, but I noticed there are many more instances in other modules. Would it make sense to rename those identifiers and avoid that name like boost avoids 'min' and 'max'? Or perhaps I could try simply undefining it in config/platform somewhere and see whether Bada headers would break? Unfortunately I don't think gcc has ability to redefine macros temporarily to reinstate the original definition at the end of boost. Generally speaking, Bada provides it's own native API, which is different from anything else. It's object-oriented and looks nice at first sight, but sometimes is inflexible in surprising places. It looks like they wrote whatever they needed to implement applications that ship with the phones and stopped there. The platform also provides ANSI C library and some parts of POSIX, but many functions don't really work. Fortunately the platform is going to be merged with Titzen, so it should at least get working POSIX kernel (Linux) and standard libraries. That makes supporting this obscure platform even less worth big trouble. -- Jan 'Bulb' Hudec <bulb(a)ucw.cz>

1 0

[general] Endianity detection for ARM
by Jan Hudec 17 Aug '12

17 Aug '12

Hello All, Something in Boost.Test started using detail/endian.hpp internally between 1.43 and 1.50. Since at work I build Boost.Test-based tests for several mobile platforms, all of which use ARM, I needed to fix it. It is easy for all platforms built with either gcc or clang (Android, iOS and Bada). Gcc defines __ARMEB__ for big-endian ARM and __ARMEL__ for little-endian ARM. Windows CE is however another issue. Microsoft usually defines rather characteristic _M_<arch> macros, but _M_ARM is *not* defined in Windows CE, though it appears to be defined for Windows 8 ARM (I don't work with those yet). The WinCE projects define ARM and _ARM_, but those may appear on big-endian ARM platforms too. However since there are no big-endian Win32 platforms and they are unlikely to be created, I suggest just hardcoding little endian for anything that defines _WIN32 (the _M_* macros could then possibly be removed, but they don't cause any problems there). I therefore propose this patch: --- a/boost/detail/endian.hpp +++ b/boost/detail/endian.hpp @@ -57,7 +57,7 @@ || defined(_POWER) || defined(__powerpc__) \ || defined(__ppc__) || defined(__hpux) || defined(__hppa) \ || defined(_MIPSEB) || defined(_POWER) \ - || defined(__s390__) + || defined(__s390__) || defined(__ARMEB__) # define BOOST_BIG_ENDIAN # define BOOST_BYTE_ORDER 4321 #elif defined(__i386__) || defined(__alpha__) \ @@ -66,7 +66,9 @@ || defined(_M_ALPHA) || defined(__amd64) \ || defined(__amd64__) || defined(_M_AMD64) \ || defined(__x86_64) || defined(__x86_64__) \ - || defined(_M_X64) || defined(__bfin__) + || defined(_M_X64) || defined(__bfin__) \ + || defined(__ARMEL__) \ + || defined(_WIN32) // ARM Windows CE don't define __ARMEL__ nor _M_ARM, but there are no big-endian Windows versions # define BOOST_LITTLE_ENDIAN # define BOOST_BYTE_ORDER 1234 -- Jan 'Bulb' Hudec <bulb(a)ucw.cz>

1 0