On May 25, 2015 10:51:17 AM EDT, Niall Douglas wrote:

On 25 May 2015 at 8:27, Rob Stewart wrote:

...

...

In case you might want to know why a monadic return transport might be so useful as to be a whole new design idiom for C++ 11, try reading

https://svn.boost.org/trac/boost/wiki/BestPracticeHandbook#a8.DESIGN:Strongl....

Make the examples in that correct WRT shared_ptr. As written, the shared_ptrs will delete file handles rather than close them.

I would imagine destroying an open file handle would close it. Anyway, that really isn't important to the code examples given the topic, it was never mentioned what a handle_type is, I had an afio::async_io_handle in mind but it doesn't matter.

Your example calls ::open() and fd is an int. ___ Rob (Sent from my portable computation engine)

On May 25, 2015 6:26:48 PM EDT, Niall Douglas wrote:

On 25 May 2015 at 15:29, Rob Stewart wrote:

...

...

I would imagine destroying an open file handle would close it. Anyway, that really isn't important to the code examples given the topic, it was never mentioned what a handle_type is, I had an afio::async_io_handle in mind but it doesn't matter.

Your example calls ::open() and fd is an int.

Sorry, I must be missing something really obvious here.

I was assuming that handle_type consumes a valid fd and takes ownership of it. Are not the examples correct then, or if a particular example (of the four) is wrong, can you say which one?

The problem was in my inferences when reading the code. I wasn't thinking of the make_shared() expression as creating a handle_type from an implicit constructor taking an int when I read the examples. I saw, in effect, make_shared<int>(fd), with handle_type as a typedef for int. Since the reader isn't necessarily familiar with a class like your async_io_handle, it might be worth a sentence to indicate that handle_type takes ownership of the file descriptor to ensure people don't repeat my mistake. ___ Rob (Sent from my portable computation engine)

On 26 May 2015 at 10:02, Lee Clagett wrote:

But if the major complaint about Expected is the compile-time, what is causing that compile-time, and how will this new type avoid that?

This monad is ridiculously simple because it's fixed function variant, and instantiates exactly two types per instance, and currently two free function overloads. No partial specialisations at all. There will be R& and void specialisations later, but they are all hash lookups, not shortlisting lookups in the compiler. Expected instantiates dozens of types and functions per instance. Lots of partial specialisations too. All of that is very slow and involves lots of generating shortlists and iterating them for closest overload match. Slow. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 26 May 2015 at 23:26, Vicente J. Botet Escriba wrote:

So the question is, do we have the proposed interface for expected, or something else? Would be the 'to be named type" a literal type?

I would far prefer if Expected were layered with increasing amounts of complexity, so you only pay for what you use. I also think you need to await the WG21 variant design to become close to completion. It makes no sense to have an Expected not using that variant implementation, you're just duplicating work. Other than that, I found the interface generally good. I only ever needed about 10% of it personally, but it would be nice if the remainder were available by switching it on maybe with an explicit conversion to a more intricate subclass. The more intricate subclass would of course live in a separate header. Pay only for what you use. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 27 May 2015 at 7:02, Vicente J. Botet Escriba wrote:

...

...

So the question is, do we have the proposed interface for expected, or something else? Would be the 'to be named type" a literal type? I would far prefer if Expected were layered with increasing amounts of complexity, so you only pay for what you use.

I also think you need to await the WG21 variant design to become close to completion. It makes no sense to have an Expected not using that variant implementation, you're just duplicating work.

My apologies. No seriously, this are implementation details. Could you describe the interface changes that expected need.

That's a huge question Vicente. And it's very hard to answer in detail, because I don't really know. I don't think variant is an implementation detail. I suspect people will want to explicitly convert from an ordered variant into an expected and vice versa for example. They will also want to "repack" a variant from one ordering of type options into another without actually copying around any data. I can also see a Hana heterogeneous sequence or std::tuple being reduced into variant and/or into an expected. And so on. Until all this ecosystem stuff becomes more final, it is hard to imagine a new expected interface.

...

Other than that, I found the interface generally good. I only ever needed about 10% of it personally, but it would be nice if the remainder were available by switching it on maybe with an explicit conversion to a more intricate subclass. The more intricate subclass would of course live in a separate header. Pay only for what you use.

Could we define this interface?

If I remember rightly, in your WG21 paper you had a table somewhere where you compared futures to optional to expected with a tick list of features shared and features not shared. I think you need an orthogonal table of: 1. Lightweight monad. 2. Intermediate monad. 3. Featureful monad. ... and another table showing the progression from minimal monad to maximum monad. If you examine https://github.com/ned14/boost.spinlock/blob/master/include/boost/spin lock/future.hpp you should see that monad<>, from which future<> derives, implements simple then(), bind() and map(). future<> only implements then(), but I am planning for it to allow a then(F(monad)) which would allow future to also provide bind() and map(). My idea is you can eventually switch freely between asynchronous monadic programming and synchronous monadic programming using a simple cast, so I am implementing a "two layer" monad where the first is a synchronous monad and the second is an asynchronous monad. But I'm a long way away from any of that right now.

In other words, would the "to be named" type appear on the AFIO interface? Is for this reason that you need to name it? or it is really an implementation detail. This merits clarification.

Right now AFIO's synchronisation object is a struct async_io_op which contains a shared_future>. I'm going to replace async_io_op with a custom afio::future<T> which *always* carries a shared_ptr plus some T (or void). Internally that converts into a future> but that isn't important to know. That custom afio::future<T> will subclass the lightweight future<T> I am currently building. Strictly speaking, there is nothing stopping me building the same custom afio::future<T> right now using std::shared_future. However, I would need to implement continuations via .then(), and if I am to bother with that, I might as well do the whole lightweight future because as you know from Boost.Thread, most of the tricky hard to not deadlock work is in the continuations implementation. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 28 May 2015 at 8:00, Vicente J. Botet Escriba wrote:

...

...

My apologies. No seriously, this are implementation details. Could you describe the interface changes that expected need. That's a huge question Vicente. And it's very hard to answer in detail, because I don't really know. In other words, beside the slow compile time of expected, what you don't need from expected?

I would prefer to not have to think when writing code with expected. I think a monadic return type should be so primitive as to require no thought in usage even for the completely unfamiliar. This is why I borrow the API and semantics from future for the monad.

I see that you want to be able to store error_code or exception_ptr. Wouldn't the interface of expected> be enough?

One of the big attractions for me to the fixed variant design is implicit conversion from T, error_code and exception_ptr. No extra typing. No extra thought. Just return what you mean from a function, and let the compiler sort it out.

...

... and another table showing the progression from minimal monad to maximum monad.

I believe that we should stop using the word monad in this way. Just wondering what would be the operations of these monads. I'm aware of the operations of a Monad and the operations of an Error Monad. I'm aware of the operations of a Functor, an Applicative, .... These operations have nothing to be with the operations the concrete type provides.

I do agree - a monad is a monad or else it is not. Ok, I am being fuzzy now, but my feelings - rather than opinion - about Expected was it is too monolithic somehow. It's like this big thing when I don't want a big thing, at least not usually. I want Expected to be a programming primitive, and it never felt to me it was a primitive. I am sorry I am not much help here. Perhaps Lee can help? He had a different opinion of using Expected in big code bases to me.

My question was about the "to be named" type. Would this type be visible on the AFIO interface?

Yes. As a custom afio::future<T>. Users will notice they can't feed those futures into normal when_all/when_any.

I don't know nothing at all about AFIO, but why the following is not good for you, then?

template <class T> using your_name = future>;

I unfortunately need to carry more state than this. I suspect I am not the first to have this problem with futures, hence my desire for a lightweight future library easily extended with custom future types.

...

That custom afio::future<T> will subclass the lightweight future<T> I am currently building.

Strictly speaking, there is nothing stopping me building the same custom afio::future<T> right now using std::shared_future. However, I would need to implement continuations via .then(), and if I am to bother with that, I might as well do the whole lightweight future because as you know from Boost.Thread, most of the tricky hard to not deadlock work is in the continuations implementation.

Yes, I'm aware of the difficulty. Any help to make it more robust is welcome. As I said in another thread (Do we can a non backward compatible version with non-blocking futures?), I was working on a branch for non-blocking futures and a branch for lightweight executors. I must take the time to merge both together.

Up until now I have had the enormous advantage over you of being able to assume that there is an ASIO available to push work onto. That makes my implementation far easier, because when I need to tie break I just defer work onto ASIO. That said, I'm going to attempt an executor-less implementation this time round. We'll see how it goes. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 28 May 2015 at 14:38, Peter Dimov wrote:

...

I see that you want to be able to store error_code or exception_ptr. Wouldn't the interface of expected> be enough?

future<T> already adds the ability to store exception_ptr. There is no need to store another exception_ptr in the value. Niall's type is exactly equivalent to future>, as I already said (twice).

exception_ptr unfortunately must always do an atomic write to memory. That always forces code to be generated, and very substantially reduces optimisation opportunities because state must be reloaded around the exception_ptr. This, and the mandatory memory allocation, is a big reason current futures are not suitable for high performance ASIO. I also want a big semantic change that error returns are not exceptional. We hugely underuse std::error_code in STL C++ unfortunately. I am not one of those people who believes exceptions are evil, and ban them in the language as all the new system languages seem to. I also have no love for forcing everything through the return code as with C and Rust, but I do think there is a middle ground between good outcomes, bad outcomes, and exceptional outcomes which is easy to program, easy to conceptualise, and easy on the compiler. I therefore think some mongrel impure future/monad/optional type might just be the ticket. But I think I need to deliver real code used in a real world library to demonstrate its effectiveness first. If I don't persuade people of that simple effectiveness, then somebody else's solution is better.

...

I proposed to the C++ standard a function future<T>::next (bind) that takes a continuation having T as parameter (future<R>(T)), and and future<T>::catch_error that takes a as parameter a continuation that takes an errror as parameter. The proposal was not accepted.

future<T>::next( [](T){...} ) is actually pretty useful, catch_error() much less so. You should have proposed just the former. (Sean Parent's future implementation's ::then has ::next's semantics IIRC.)

I was going to have my continuations specialise on their parameter type, so a continuation might take T, error_code, exception_ptr, monad<T>, future<T> as consuming continuations. To make those non-consuming continuations simply make those a const lvalue ref. My only worry really on that feature is on compile times, so I may yet strip some of the automatedness. Also if I'm not going to regularly use a feature, then I'm removing that feature, I am aiming for a least possible implementation. We'll see how it goes. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 28 May 2015 at 16:25, Peter Dimov wrote:

...

I also want a big semantic change that error returns are not exceptional. We hugely underuse std::error_code in STL C++ unfortunately. I am not one of those people who believes exceptions are evil, and ban them in the language as all the new system languages seem to. I also have no love for forcing everything through the return code as with C and Rust, but I do think there is a middle ground between good outcomes, bad outcomes, and exceptional outcomes which is easy to program, easy to conceptualise, and easy on the compiler.

There could be a misunderstanding. When I say that your type is future>, I don't mean any specific future or expected implementations such as std::future or boost::expected. What I mean is a future-like type, having the interface of std/boost::future, and an expected-like type, having (a subset of) the interface of std/boost::expected.

For reference, I originally began all of this nine months ago with a future, error_code>>. I found it obtusive and irritating to work with because of the nested layers of API e.g. if(f.has_value()) f.get().value_or(...) etc. I found myself wanting to wrap future, error_code>> into a class which implements a flat convenience API.

You could still implement your own future<> and expected<> if the existing ones are unfit. My point is purely that these are independent concepts and there is no real need to couple them into one type, with a hardcoded error_code to boot.

error_code is pretty great in fact. I've tried to come up with an error transmitting type which is better for the same light weight spec, and I failed after several months of effort. I actually now believe error_code has an optimal design which cannot be improved upon without losing that its storage is just an int and a pointer and is therefore a completely trivial type. That is a very rare opinion for me to hold on any part of the STL, but I became quite impressed with the design once I failed to improve on it. My sole thing I would change if I could is that I really wish it stored a ssize_t not an int. An additional void * in its storage could also be useful.

So far, you've stated that you like that your type is constructible from T, error_code or exception_ptr. I suppose a future> may not be as convenient.

It wasn't fun and convenient to use when I started out down this path. I also didn't like how many opcodes were generated by the compiler for very simple operations, especially on MSVC.

Are there other ways in which the interface of future> is deficient?

I very much like the expected API and design. Why I choose not to use it comes down to the lack of a flat API as mentioned above, and compile times. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

Vicente J. Botet Escriba wrote:

Without catch_error(), you need yet .then() as you want to be able to recover from errors in the same way.

I wasn't suggesting removing .then, just adding .next. It's useful when I need to perform the same action regardless of whether I got a value or an exception.

Niall Douglas wrote:

On 28 May 2015 at 21:14, Peter Dimov wrote:

...

I wasn't suggesting removing .then, just adding .next. It's useful when I need to perform the same action regardless of whether I got a value or an exception.

I'm a bit confused. If you want some action to occur regardless, you surely ignore the future you are passed in your continuation?

Or by next(), do you mean that the continuation must fire even if no value nor exception is ever set, and is therefore fired on future destruction?

What I mean is: .then( []( future<T> ft ) { /*...*/ } ) Used when I want to, for instance, send ft to someone via message passing .next( []( T t ) { /* ... */ } ) Fires when ready() && has_value(). Used when I only care about results. So, for example, when I do auto r = async(f1).next(f2).next(f3).next(f4).get(); I don't care which of the four steps has failed with an exception. That's as if I write auto r = f4(f3(f2(f1()))); If one of these functions throws, I just get an exception, I don't care from where. It's the same with .next.

On 28 May 2015 at 21:34, Peter Dimov wrote:

...

...

I wasn't suggesting removing .then, just adding .next. It's useful when I need to perform the same action regardless of whether I got a value or an exception.

I'm a bit confused. If you want some action to occur regardless, you surely ignore the future you are passed in your continuation?

Or by next(), do you mean that the continuation must fire even if no value nor exception is ever set, and is therefore fired on future destruction?

What I mean is:

.then( []( future<T> ft ) { /*...*/ } )

Used when I want to, for instance, send ft to someone via message passing

.next( []( T t ) { /* ... */ } )

Fires when ready() && has_value(). Used when I only care about results.

So, for example, when I do

auto r = async(f1).next(f2).next(f3).next(f4).get();

I don't care which of the four steps has failed with an exception.

Ok, let me rewrite that so we know we are on the same page: future<T>.next(R(T)) is the same effect as: future<T>.then([](future<T> f){ return R(f.get()); // f.get() rethrows any exception, doesn't execute R(T) }); Your .next() is the same as Vicente's expected.bind() right? If so, I was going to have .then() cope with a R(T) callable, but that requires metaprogramming to examine the input and determine the overload. A .next(R(T)) might be easier on build times. I think I saw you didn't like Vicente's catch_error(E)? Was there a reason why? Is there anything wrong with a .next(R(E))? Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 29 May 2015 at 11:05, Peter Dimov wrote:

...

If so, I was going to have .then() cope with a R(T) callable, but that requires metaprogramming to examine the input and determine the overload.

You're not going to be able to tell the difference between .then and .next when given [](auto x) { ... }.

Surely with Expression SFINAE you can? But sure, it's sounding slower than a simple overload. And I'd like to retain VS2015 compatibility.

...

I think I saw you didn't like Vicente's catch_error(E)? Was there a reason why? Is there anything wrong with a .next(R(E))?

It's not that I don't like it, I don't see sufficient motivation for it.

Oh okay. In my situation, T and error_code are considered equals, only exception_ptr is exceptional.

I'm also not quite sure how it's to be used. Vicente's wording in N4048 is a bit unclear.

future<X> f1 = async( ... ); future<Y> f2 = f1.catch_error( []( exception_ptr e ) { return Y(); } );

What happens when f1 has a value?

Surely for catch_error() one must always return the same type of future as the input? Otherwise it couldn't work when f1 has a value. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 30 May 2015 at 1:26, Peter Dimov wrote:

...

...

...

It's not that I don't like it, I don't see sufficient motivation for it.

Actually I've changed my mind and now I don't like it. It's a catch(...) without a rethrow, and while those are occasionally necessary, they are not likable.

To be clear, I'm talking about the N4048's .catch_error that takes an exception_ptr, not an error_code.

I am sympathetic to this line of thought. Once you allow an error_code, it becomes hard to imagine where treating exception_ptr outcomes as just another value returned can be wise. Ok, so if .next() filters for T, would .next_error() filter for error_code? I can't say I like the name .next_error(). It suggests "the next error", not the current one. .on_error() maybe? But then .next() should be .on_value() instead right?

...

...

I'm also not quite sure how it's to be used. Vicente's wording in N4048 is a bit unclear.

future<X> f1 = async( ... ); future<Y> f2 = f1.catch_error( []( exception_ptr e ) { return Y(); } );

What happens when f1 has a value?

Surely for catch_error() one must always return the same type of future as the input? Otherwise it couldn't work when f1 has a value.

Vicente in N4048 says that it's not required to return the same type.

Oh. I don't know how that can work then. Maybe Vicente can clarify? Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

Vicente J. Botet Escriba wrote:

...

...

Surely for catch_error() one must always return the same type of future as the input? Otherwise it couldn't work when f1 has a value.

Vicente in N4048 says that it's not required to return the same type.

If I said that it was an error. Could you point me where?

Input Parameters: • Lambda function: The lambda function on map()/bind() takes a future<t>::value_type. The lambda function on catch_error() takes an exception_ptr. Both could return whatever type. This makes propagating exceptions straightforward. This approach also simplifies the chaining of continuations.

On Sun, May 31, 2015 at 6:25 PM, Vicente J. Botet Escriba wrote:

Le 30/05/15 22:49, Peter Dimov a écrit :

...

Vicente J. Botet Escriba wrote:

...

...

...

Surely for catch_error() one must always return the same type of future >> as the input? Otherwise it couldn't work when f1 has a value.

Vicente in N4048 says that it's not required to return the same type.

If I said that it was an error. Could you point me where?

Input Parameters: • Lambda function: The lambda function on map()/bind() takes a future<t>::value_type. The lambda function on catch_error() takes an exception_ptr. Both could return whatever type. This makes propagating exceptions straightforward. This approach also simplifies the chaining of continuations.

Thanks. I suspect I was referring to map and bind. future<T>::cath_error must return T or future<T>. I recognize that it is not clear in the N4048.

Initially catch_error() was called recover(). It can be used to recover from errors and then the function returns a value or a future ready with a value and also to change the reported error, in this case the function should return future<T> ready with an exception.

I can imagine catch_error() returning something _convertible_

future<X> f1 = async( ... ); future<Y> f2 = f1.catch_error( []( exception_ptr e ) { return Y(); } );

Y needs to be convertible from X. Not sure there is much value in that, however. P.S. can I specialize std::future<MyType> ? And does that just cause headaches for the STL? Tony

...

I can imagine catch_error() returning something _convertible_

...

future<X> f1 = async( ... ); future<Y> f2 = f1.catch_error( []( exception_ptr e ) { return Y(); } );

Y needs to be convertible from X.

The opposite is also possible. The continuation can return a type Y convertible to X

future<X> f1 = async( ... ); future<X> f2 = f1.catch_error( []( exception_ptr e ) { return Y(); } );

...

Not sure there is much value in that, however.

I have not needed it.

...

P.S. can I specialize std::future<MyType> ? And does that just cause headaches for the STL?

No and I don't know. Do you have a use case that would need specialization?

No. I fear specialization here. I think we should ban it now before anyone starts doing it, unless someone can come up with a really compelling reason.

On 05/25/2015 05:57 PM, Niall Douglas wrote:

On 25 May 2015 at 15:44, Avi Kivity wrote:

...

I believe error_code is unneeded. Exceptions are expected to be slow. If you want another type of variant return, let the user encapsulate it in T (could be optional<T>, or expected, or whatever). Please read the rationale at https://svn.boost.org/trac/boost/wiki/BestPracticeHandbook#a8.DESIGN:S tronglyconsiderusingconstexprsemanticwrappertransporttypestoreturnstat esfromfunctions as was requested.

In particular, error_code is fast, and unexpected returns are not exceptional and must be as fast as expected returns.

As I mentioned, in this case the user can use expected<> or similar themselves. Otherwise, what's the type of error_code? There could be an infinite amount of error_code types to choose from (starting with simple 'enum class'es and continuing with error codes that include more information about the error (nonscalar objects).

Also, any monadic transport would default construct to an unexpected state of a null error_code in fact, which is constexpr. This lets one work around a number of exception safety irritations where move constructor of T is not noexcept more easily.

I'm not sure how the default constructor of future<> and the move constructor of T are related. I'm not even sure why future<> would require a default constructor. Seastar's doesn't have one.

...

...

... turns into a "mov $5, %eax", so future<T> is now also a lightweight monadic return transport capable of being directly constructed. Can you post the code? I'd be very interested in comparing it with seastar's non-allocating futures. I may do so once I've got the functioning constexpr reduction being unit tested per commit.

I'm looking forward to it! I've been bitten by the same compile time explosion problems and I'm curious to see how you solved them.

Niall Douglas wrote:

I've got everything working except the sequence:

promise<int> p; p.set_value(5); return p.get_future().get();

This should reduce to a mov $5, %eax, but currently does not for an unknown reason. I'm just about to go experiment and see why.

I'm really struggling to see how all that could work. Where is the result stored? In the promise? Wouldn't this require the promise to outlive the future<>? This doesn't hold in general. How is it to be guaranteed?

Avi Kivity wrote:

On 05/25/2015 09:04 PM, Peter Dimov wrote:

...

Niall Douglas wrote:

...

I've got everything working except the sequence:

promise<int> p; p.set_value(5); return p.get_future().get();

This should reduce to a mov $5, %eax, but currently does not for an unknown reason. I'm just about to go experiment and see why.

I'm really struggling to see how all that could work. Where is the result stored? In the promise? Wouldn't this require the promise to outlive the future<>? This doesn't hold in general. How is it to be guaranteed?

In seastar, we achieved this by reserving space for the result in both future and promise, and by having future and promise track each other (so they can survive moves).

Doesn't this introduce a race between ~promise and ~future? ~future checks if( promise_ ), ~promise checks if( future_ ), and things get ugly. Fixing that requires synchronization, which makes mov $5, %eax impossible.

On 25 May 2015 at 21:04, Peter Dimov wrote:

...

I've got everything working except the sequence:

promise<int> p; p.set_value(5); return p.get_future().get();

This should reduce to a mov $5, %eax, but currently does not for an unknown reason. I'm just about to go experiment and see why.

I'm really struggling to see how all that could work. Where is the result stored? In the promise? Wouldn't this require the promise to outlive the future<>? This doesn't hold in general. How is it to be guaranteed?

The promise storage is an unrestricted union holding any of the types returnable by the future, or a future<T> * or a shared_ptr (for shared_future). Ideally you'd fold the internal spinlock into the unrestricted union too as you only need it when the storage is selected to the future<T> *, but it complicates the source code very significantly, so for now I've left that out. Peter Dimov wrote:

...

In seastar, we achieved this by reserving space for the result in both future and promise, and by having future and promise track each other (so they can survive moves).

Doesn't this introduce a race between ~promise and ~future? ~future checks if( promise_ ), ~promise checks if( future_ ), and things get ugly.

Fixing that requires synchronization, which makes mov $5, %eax impossible.

In my implementation, if you never call promise.get_future() you never get synchronisation. The constexpr folding has the compiler elide all that. Even though clang spills a lot of ops, it still spills no synchronisation. MSVC unfortunately spills everything, but no synchronisation would ever be executed by the CPU. Also if you call promise.set_value() before promise.get_future(), you never get synchronisation as futures are single shot. I took a very simple synchronisation solution - a spinlock in both promise and future. Both are locked before any changes happen in either, if synchronisation has been switched on. Another optimisation I have yet to do is to detach the future-promise after the value has been set to avoid unnecessary synchronisation. Giovanni Piero Deretta wrote:

So, the future/promise pair can be optimized out if the work can be completed synchronously (i.e. immediately or at get time). But then, why use a future at all? What is the use case you are trying to optimize for? do you have an example?

For me my main purpose is making AFIO allocate four malloc/frees per op instead of eight. I also make heavy use of make_ready_future(), and by definition that is now optimally fast. Also, as I mentioned, the lion's share of the future implementation is actually reusable as a monadic transport. That's currently a monad base class in my code, but I am asking for bike shedding here on what to name a user facing specialisation.

I believe that trying to design a future that can fulfill everybody's requirements is a lost cause. The c++ way is to define concepts and algorithms that work on concepts. The types we want to generalize are std::future, expected, possibly optional and all the other futures that have been cropping up in the meantime. The algorithms are of course those required for composition: then, when_all, when_any plus probably get and wait.

I think it might not be a lost cause in a world with concepts and especially modules. Until then it's going to be unacceptably slow. And that's years away, and I need this now. Regarding heterogeneous future type wait composition, yes this is hoped to be the foundation for such an eventual outcome. That's more Vicente's boat than mine though.

We could just take a page from Haskell and call the concept Monad, but maybe we want something more specific, like Result.

I'll take that as a vote for result<T>. Bjorn suggested holder<T> and value<T> as well. I think the former too generic, and the latter suggests the thing is a value and can decay into one without get(). Or maybe, that some might think that.

Then, in addition to the algorithms themselves, there is certainly space in boost for a library for helping build custom futures, a-la boost.iterator.

I originally intended that, my second round of experiments showed it could be very exciting. But I choose to bow out in favour of something I can deliver in weeks rather than months. I need all this ready mid-June so I can start retrofitting AFIO. Besides, simpler I think may well win the race. More complex means more compiler quirks. I am dealing enough with those already! Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

Seems to me that future and promise both first need to take their spinlock and then the other's, which creates the potential for deadlock.

You could fix that by preallocating a sufficiently big static array of spinlocks, then generating an index when creating the promise and using spinlock_[i] in both. Collisions are possible but harmless. This non-allocating implementation is an interesting argument in favor of the current "unique future", which I've long disliked. I prefer futures to be shared_futures. That may too be possible to implement without allocation, but it's kind of complicated.

On 26 May 2015 at 3:50, Gottlob Frege wrote:

Alternatively, with a lock in each and pointers pointing to each other, you avoid deadlock by first setting (via CAS) your own state to be "I'm moving" then (if successful) setting your partner's flag to "see ya later", then (if successful) moving.

No one moves without telling the other first. You can get a live lock, but not a dead lock. The live Lock can be dealt with (particularly easily since the relationship (promise vs future) is asymmetrical - just say future always goes first, for example).

I found via empirical testing that the number of occasions when two threads both access promise and future concurrently is extremely rare. The stupidly simple method of locking both for every access I didn't find was a problem. And it's easy to test, and verify with the thread sanitiser. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 27 May 2015 at 12:21, Gottlob Frege wrote:

...

...

Alternatively, with a lock in each and pointers pointing to each other, you avoid deadlock by first setting (via CAS) your own state to be "I'm moving" then (if successful) setting your partner's flag to "see ya later", then (if successful) moving.

No one moves without telling the other first. You can get a live lock, but not a dead lock. The live Lock can be dealt with (particularly easily since the relationship (promise vs future) is asymmetrical - just say future always goes first, for example).

I found via empirical testing that the number of occasions when two threads both access promise and future concurrently is extremely rare. The stupidly simple method of locking both for every access I didn't find was a problem. And it's easy to test, and verify with the thread sanitiser.

Same algorithm though, right?

Lock yourself, then try to lock your partner, unlock yourself if you can't lock your partner, repeat?

I think yours has bias towards the future? I found the logic for handling that given the rarity of live lock wasn't worth it. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On Tue, May 26, 2015 at 6:36 AM, Peter Dimov wrote:

Gottlob Frege wrote:

...

Can't use a spinlock pool. Because you might hold the spin lock when moving/copying T, and T could be anything, then, in theory at least, T's constructor might also use a future/promise, etc, etc, such that a collision could cause a deadlock.

Yes, good point. Need recursive spinlocks.

I don't think a pool of recursive spinlocks works, because T could decide to wait on another thread, which is waiting on a future, which could - behind the scenes - be using the same spinlock as the original thread. You could probably still use a pool somehow, with more management, but it gets complicated. You probably either need at least as many locks as threads, or a fall back of allocating a lock if none are available - but the original point was to not allocate... (Or a way to know that Thread A is waiting on Thread B...) Tony

Giovanni Piero Deretta wrote:

On 26 May 2015 12:59 am, "Peter Dimov" wrote:

...

This non-allocating implementation is an interesting argument in favor of the current "unique future", which I've long disliked. I prefer futures to be shared_futures.

Interesting, why do you dislike the unique future design?

I dislike the unique/shared future split, which requires all algorithms to be duplicated/manyplicated. auto f = when_any( f1, f2, f3, f4 ); // 2^4 options I prefer a single future<T> type, and I want this single type to be shared because I want the above when_any to not std::move my futures, leaving me with empty shells.

A shared future pretty much requires holding a shared pointer and needs heavy weight synchronisation ( a muted+condvar or equivalent). On the other hand a unique future need no internal mutual exclusion and the only synchronisation is needed for the handoff between producer and consumer; the reference count is implicit (just have the consumer always deallocate the object). The implementation can be significantly more light weight.

I already acknowledged that these implementations are an argument in favor of unique future. But I don't think I agree with what you're saying. I don't see how mutex+condvar is required by shared future but not by unique future (how do you implement wait() is not related to sharedness), neither do I see how unique futures require no synchronization (they obviously do in Niall's implementation.)

Giovanni Piero Deretta wrote:

...

auto f = when_any( f1, f2, f3, f4 ); // 2^4 options

The algorithm can be generic of course.

Can it be generic? Do you have a generic implementation of when_any?

You'll have the same problem if you want to mix different futures from separate libraries.

Yes, I know. Hence my preference for a single std::future that is _the_ future type.

You are right of course. The readiness notification is orthogonal, but in a shared future you need a mutex to synchronise the various consumers access to the shared state, so a condition variable becomes the obvious choice, while you can be more creative with plain futures.

No, in fact I do not need a mutex to synchronize consumers, as far I can see. Where would I need it? Consumers don't interfere with one another.

Giovanni Piero Deretta wrote:

You need to protect concurrent calls to then...

Hm, good point, I hadn't thought of that. Note however that if remove promise::get_future and create a promise/future pair in one go, I can then in principle check that the future is unique and omit the mutex lock.

You might as well use it to protect the refcount.

Eh, no. Refcounts have been protecting themselves for decades. :-)

Giovanni Piero Deretta wrote:

...

Can it be generic? Do you have a generic implementation of when_any?

You do need to standardise a generic asynchronous wait protocol of course.

Or, I can make promises shared as well, and then use .then. when_any( f1, f2, f3 ) { future f; promise p(f); f1.then( [p](auto){ p.set_value(0); } ); f2.then( [p](auto){ p.set_value(1); } ); f3.then( [p](auto){ p.set_value(2); } ); return f.then( []( future index ) { return make_pair(index.get(), make_tuple(f1, f2, f3)); }); } Something like that. (I still don't need a mutex for the multiple set_value calls, by the way - if(rn_++ == 0) { set the value; publish/notify; }.)

Not quite sure how this would work. Seems to me that future and promise both first need to take their spinlock and then the other's, which creates the potential for deadlock. But I may be missing something.

I saw the code, you try_lock and back off.

On 05/26/2015 12:20 PM, Niall Douglas wrote:

...

...

The constexpr folding has the compiler elide all that. I suspect that what you call "constexpr folding" has nothing to do with constexpr, it's just inlining and ordinary non-constexpr optimization. KAI C++ was famous for doing such folding miracles decades ago. You're right it's not in the standard. Well, actually it is, but indirectly.

Let me explain. If you read what is allowed for constexpr at http://en.cppreference.com/w/cpp/language/constexpr, and then write logic which provides outcome paths which do nothing not constexpr, then the compiler will completely elide code entirely at compile time when those paths are followed. If you examine my code closely, you'll see I always predicate the construction of anything with a non-trivial destructor (e.g. exception_ptr) because non-trivial destructors appear to force code output in current compiler technology, though it may also be the atomic writes that exception_ptr does.

This probably is not by accident. The machinery inside the compiler to implement constexpr is probably reused for optimisation, or rather, vice versa. Maybe a compiler vendor might chime in here to tell us?

I believe it is an accident. While the compiler is required to fold a constexpr expression in a constexpr context (like an array dimension or non-type template argument), it isn't required to do so in non-constexpr contexts, and won't do so if optimization is not enabled or if it makes a bad inlining decision (say, because the function was too large). The optimization machinery for folding constants is actually a superset of constexpr, and as a result you will sometimes get a constexpr expression not optimized, and sometimes a non-constexpr expression will be folded to nothing.

...

...

Also if you call promise.set_value() before promise.get_future(), you never get synchronisation as futures are single shot. Another interesting case for which I've trouble imagining a practical use. :-) Its main benefit for me is as a unit test smoke test. However resumable functions, as currently proposed, should hugely benefit from this pattern. I've emailed Gor about it.

I am also waiting for resumables (and concepts, and modules, and ranges...) with bated breath.

On 26 May 2015 at 13:50, Peter Dimov wrote:

The use case I was wondering above is that you create a promise, but never get a future from it.

The present resumable functions proposal requires a promise to be created on entry to every resumable function. The first time the function is suspended promise.get_future() is called, and the future returned immediately. If the resumable function never suspends, promise.get_future() ought to be called straight after the set_value() OR make_ready_future will be used instead. At least, that will be the case of Gor agrees to tweak the operation ordering in the TS as I have requested (I can't see it being a problem). If that ordering is implemented, then with these future-promises a resumable function which was never suspended has identical runtime overheads to a non-resumable function, and that is a HUGE win over the present situation which is much less efficient. All this assumes I understand the current TS before the committee correctly and its exact boilerplate expansion from the await/yield etc keywords. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 26 May 2015 12:09 am, "Niall Douglas" wrote:

Giovanni Piero Deretta wrote:

...

So, the future/promise pair can be optimized out if the work can be completed synchronously (i.e. immediately or at get time). But then, why use a future at all? What is the use case you are trying to optimize for? do you have an example?

For me my main purpose is making AFIO allocate four malloc/frees per op instead of eight. I also make heavy use of make_ready_future(), and by definition that is now optimally fast.

The ready future path is taken unconditionally or with a condition often known at compile time? If yes that seems strange to me and would love to see an example. If no then you are probably optimising for the wrong case.

Also, as I mentioned, the lion's share of the future implementation is actually reusable as a monadic transport. That's currently a monad base class in my code, but I am asking for bike shedding here on what to name a user facing specialisation.

I think you are selling your code wrongly. What you have is a potentially zero overhead expected-like result wrapper; this is something people would want. What you are advertising is a zero overhead future in the uninteresting case and people can't see the point of it. The fact that your future is internally implemented using your result wrapper is nice but not critical. Also why does the future inherit from the result object? I would expect it to contain something like variant.

...

I believe that trying to design a future that can fulfill everybody's requirements is a lost cause. The c++ way is to define concepts and algorithms that work on concepts. The types we want to generalize are std::future, expected, possibly optional and all the other futures that have been cropping up in the meantime. The algorithms are of course those required for composition: then, when_all, when_any plus probably get and wait.

I think it might not be a lost cause in a world with concepts and especially modules. Until then it's going to be unacceptably slow.

I'm not following. How do concepts and modules allows to have a catch all future?

And that's years away, and I need this now.

Regarding heterogeneous future type wait composition, yes this is hoped to be the foundation for such an eventual outcome. That's more Vicente's boat than mine though.

...

We could just take a page from Haskell and call the concept Monad, but maybe we want something more specific, like Result.

I'll take that as a vote for result<T>.

Eh, I meant it as a concept name, not actual class, but sure.

Bjorn suggested holder<T> and value<T> as well. I think the former too generic, and the latter suggests the thing is a value and can decay into one without get(). Or maybe, that some might think that.

You could leave the return type unspecified and just list the type requirements.

...

Then, in addition to the algorithms themselves, there is certainly space in boost for a library for helping build custom futures, a-la boost.iterator.

I originally intended that, my second round of experiments showed it could be very exciting. But I choose to bow out in favour of something I can deliver in weeks rather than months. I need all this ready mid-June so I can start retrofitting AFIO.

If you need this week just use boost.thread future and get afio reviewed with that. Have the magic future as an optional unstable preview only interface. Boost reviews usually focus on interfaces more than implementation. As long as the reviewers believe that your interface can be implemented efficiently you'll be fine. In fact having yet another non interoperable future in afio might be a negative point during review. After afio is accepted, you can lobby to get boost.thread future do what you want, or better add generalised future composition to boost.

Besides, simpler I think may well win the race. More complex means more compiler quirks. I am dealing enough with those already!

Niall

-- ned Productions Limited Consulting http:// http://www.nedproductions.biz/www.nedproductions.biz/

http://www.nedproductions.biz/

http:// http://ie.linkedin.com/in/nialldouglas/ie.linkedin.com http://ie.linkedin.com/in/nialldouglas//in/ http://ie.linkedin.com/in/nialldouglas/nialldouglas http://ie.linkedin.com/in/nialldouglas// http://ie.linkedin.com/in/nialldouglas/

_______________________________________________ Unsubscribe & other changes: http:// http://lists.boost.org/mailman/listinfo.cgi/boostlists.boost.org http://lists.boost.org/mailman/listinfo.cgi/boost/mailman/ http://lists.boost.org/mailman/listinfo.cgi/boostlistinfo.cgi http://lists.boost.org/mailman/listinfo.cgi/boost/boost http://lists.boost.org/mailman/listinfo.cgi/boost

On 05/25/2015 07:37 PM, Niall Douglas wrote:

On 25 May 2015 at 18:33, Avi Kivity wrote:

...

...

In particular, error_code is fast, and unexpected returns are not exceptional and must be as fast as expected returns. As I mentioned, in this case the user can use expected<> or similar themselves. As I mentioned, expected<> is too hard on compile times for large code bases. It's also way overkill for what 98% of use cases need.

...

Otherwise, what's the type of error_code? There could be an infinite amount of error_code types to choose from (starting with simple 'enum class'es and continuing with error codes that include more information about the error (nonscalar objects). It's std::error_code. Same as ASIO uses. If being compiled as part of Boost, I expect boost::error_code and boost::exception_ptr will work too as additional variant options.

...

...

Also, any monadic transport would default construct to an unexpected state of a null error_code in fact, which is constexpr. This lets one work around a number of exception safety irritations where move constructor of T is not noexcept more easily. I'm not sure how the default constructor of future<> and the move constructor of T are related. Well, let's assume we're really talking about a maybe<T>, and future<T> subclasses maybe<T> with additional thread safety stuff.

In this situation a maybe<T> doesn't need a default constructor, but because it's a fixed variant we always know that error_code is available, and error_code (a) doesn't allocate memory and (b) is STL container friendly, so it seems sensible to make maybe<T> also STL container friendly by letting it default to error_code.

The problem, as with the WG21 proposed variant, is getting move assignment to not undefine behaviour the existing contents if the throwing move constructor throws. Boost Variant's copy constructor dynamically allocates a temporary copy of itself internally to give that strong guarantee - this is unacceptable overhead for mine. So I need some well defined state to default to if during move assignment my move constructor throws after I have destructed my current state. Defaulting to an error_code saying the move constructor threw is a reasonable well defined outcome.

...

I'm not even sure why future<> would require a default constructor. Seastar's doesn't have one. My future promise is as close to a strict superset of the Concurrency TS as is possible. It should be drop in replaceable in 99% of use cases, with the only point of failure being if you are trying to use allocators with your futures.

My future promise is also intended to enter the Boost.Thread rewrite as the next gen future promise, if it proves popular.

...

...

...

...

... turns into a "mov $5, %eax", so future<T> is now also a lightweight monadic return transport capable of being directly constructed. I'm looking forward to it! I've been bitten by the same compile time explosion problems and I'm curious to see how you solved them. With a great deal of concentrated study of compiler diagnostics and trial and error!

Once they are working and they are being unit tested per commit, I'll get a CI failure every time I break it. That should make things enormously easier going. Lots of machinery and scripting to come before that though.

I've got everything working except the sequence:

promise<int> p; p.set_value(5); return p.get_future().get();

This should reduce to a mov $5, %eax, but currently does not for an unknown reason. I'm just about to go experiment and see why.

I managed to get very close to this by sprinking always_inline attributes, mostly at destructors. As soon as the compiler makes a bad inlining decision, it loses track of the values it propagated and basically has to undo all optimization. It's still not perfect (one extra instruction) though. (using gcc 5).

On 26/05/2015 13:48, Hartmut Kaiser wrote:

...

I've got everything working except the sequence:

promise<int> p; p.set_value(5); return p.get_future().get();

This should reduce to a mov $5, %eax, but currently does not for an unknown reason. I'm just about to go experiment and see why.

I asked this question many times before: what is this good for in practice except for demonstrating some impressive compiler optimization capabilities? If I need to return the number '5' I'd usually write

return 5;

in the first place...

Generally this sort of pattern comes up when implementing a generic method that is constrained (by base class or concept) to return a future<T>, but where the actual implementation can complete synchronously without waiting (eg. a filesystem API that normally reads data asynchronously, but in this particular case is being implemented by an in-memory filesystem that can reply instantly). I'm assuming that the code Niall posted isn't literally written like that but is instead produced after inlining such generic method calls for a particular test case.

On 26/05/2015 13:48, Hartmut Kaiser wrote:

...

...

I've got everything working except the sequence:

promise<int> p; p.set_value(5); return p.get_future().get();

This should reduce to a mov $5, %eax, but currently does not for an unknown reason. I'm just about to go experiment and see why.

I asked this question many times before: what is this good for in

practice

...

except for demonstrating some impressive compiler optimization capabilities? If I need to return the number '5' I'd usually write

return 5;

in the first place...

Generally this sort of pattern comes up when implementing a generic method that is constrained (by base class or concept) to return a future<T>, but where the actual implementation can complete synchronously without waiting (eg. a filesystem API that normally reads data asynchronously, but in this particular case is being implemented by an in-memory filesystem that can reply instantly).

Ok, so what? Just use make_ready_future - done. Do we even know how large the overheads of this are? Or do you just 'assume' for the overheads to be unacceptable? Can I see some numbers from a real world application?

I'm assuming that the code Niall posted isn't literally written like that but is instead produced after inlining such generic method calls for a particular test case.

I doubt any speedup based on all of this fancy-pants optimizations would be even measurable in the context of file system operations. I'm still highly doubtful of all of this. As said, give me a real world use case with real world measurements showing at least some speedup over 'conventional' futures. Otherwise all of this is an empty exercise. Regards Hartmut --------------- http://boost-spirit.com http://stellar.cct.lsu.edu

On 26 May 2015 at 20:35, Avi Kivity wrote:

It can be important for O_DIRECT AIO operations. I agree that for buffered I/O, the filesystem overhead will dominate (and, on Linux, you don't have a way to implement futures over buffered I/O without resorting to threads, which will slow things down further).

Actually, on recent Linuces with ext4 reads from page cached data are now kaio wait free. It makes a big difference for warm cache filesystem when you're doing lots of small reads. Writes unfortunately still lock, and moreover exclude readers. Linux has a very long way to go to reach BSD and especially Windows for async i/o. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On Tue, May 26, 2015 at 5:29 PM, Hartmut Kaiser wrote:

...

...

It can be important for O_DIRECT AIO operations. I agree that for buffered I/O, the filesystem overhead will dominate (and, on Linux, you don't have a way to implement futures over buffered I/O without resorting to threads, which will slow things down further).

Actually, on recent Linuces with ext4 reads from page cached data are now kaio wait free. It makes a big difference for warm cache filesystem when you're doing lots of small reads. Writes unfortunately still lock, and moreover exclude readers. Linux has a very long way to go to reach BSD and especially Windows for async i/o.

Optimizing away one allocation to create a promise/future pair (or just a future for make_ready_future) will have no measurable impact in the context of any I/O, be it wait free or asynchronous or both.

In general, all I'm hearing on this thread is 'it could be helpful', 'it should be faster', 'it can be important', or 'makes a big difference', etc. I was hoping that we as a Boost community can do better!

Nobody so far has shown the impact of this optimization technique on a real world applications (measurements). Or at least, measurement results from artificial benchmarks under heavy concurrency conditions (using decent multi-threaded allocators like jemalloc or tcmalloc). I'd venture to say that there will be no measurable speedup (unless proven otherwise).

+1 I'd add that not only there should be a measurable improvement, but it should be a measurable improvement compared to other reasonable optimizations, for example allocations can be optimized through custom allocators. As an analogy, it's not sufficient to show that shared_ptr is "too slow" or "allocates too much" compared to some other smart pointer type -- it also must be shown that the slowness can't be trivially dealt with by implementing a custom allocator for some shared_ptr instances, if needed. -- Emil Dotchevski Reverge Studios, Inc. http://www.revergestudios.com/reblog/index.php?n=ReCode

...

Optimizing away one allocation to create a promise/future pair (or just a future for make_ready_future) will have no measurable impact in the context of any I/O, be it wait free or asynchronous or both.

No one here is claiming better futures make any effect on i/o performance except you.

From the AFIO docs:

AFIO is 'A C++ library which lets you schedule an ordered dependency graph of file input/output operations...' (http://rrsd.com/blincubator.com/bi_library/afio/?gform_post_id=938). And you said more than once that a) the new 'futures' are 'much' faster and b) that you want to use those for AFIO. Go figure. You have mentioned wait free algorithms in the ext4 file system caches which would make file system IO so fast that using your 'futures' would be beneficial (see your mail: http://thread.gmane.org/gmane.comp.lib.boost.devel/260307/focus=260404). You even mention it further down in the mail Iäm answering to. Go figure. Others have drawn direct connections on this thread between futures and IO as well (see for instance Gavin Lambert: http://thread.gmane.org/gmane.comp.lib.boost.devel/260307/focus=260345, or Avi Kivity: http://thread.gmane.org/gmane.comp.lib.boost.devel/260307/focus=260391). Go figure. You clearly should decide what you want besides trolling the Boost-devel mailing list.

You are reading only the parts of the thread you want to in order to believe what you already believe (apparently that I have some master plan of "taking over" Boost).

I have not said or implied anything like that in this thread. Also, you know best what your 'master plan' is.

You drew the link here between i/o and futures, none of us claimed it. The thread earlier was clearly about two entirely separate topics. You conflated them to make your own personal point.

I have not conflated anything.

...

In general, all I'm hearing on this thread is 'it could be helpful', 'it should be faster', 'it can be important', or 'makes a big difference', etc. I was hoping that we as a Boost community can do better!

Constant cherry picking of thread topics just to nay say and put down any discussion of alternative idioms and designs isn't being positive nor helpful.

Constant? Really? Just because I believe that your ideas have flaws and your solutions are wrong because you start from incorrect assumptions?

...

Nobody so far has shown the impact of this optimization technique on a real world applications (measurements). Or at least, measurement results from artificial benchmarks under heavy concurrency conditions (using decent multi-threaded allocators like jemalloc or tcmalloc). I'd venture to say that there will be no measurable speedup (unless proven otherwise).

Again nobody claimed that. I was quite clear I primarily want single op code reduction as part of unit testing. That's its main purpose for me as a per-commit CI test that I am writing perfectly optimal code, not just mostly optimal code.

What is 'op code reduction' all about if not to achieve speedup? All I asked is to give us numbers showing whether those 'op code reductions' have any _significant_ impact on the overall performance of real world applications. All you gave us so far are conjectures.

A happy consequence is a potential runtime cost optimal monadic transport, and that's what I came here to bikeshed a name for, and see if there is interest in such a development. Feedback on both has been both positive, and useful, so I will proceed.

Sure. I'm off of this thread. Happy bikeshed-ing! Regards Hartmut --------------- http://boost-spirit.com http://stellar.cct.lsu.edu

I have also been very clear that this new design solves my major problems, not *the* major problems with existing futures. That's what I designed it to do. I believe it also solves the same problems as face futures in ASIO. Once finished and deployed, if others find it solves their problems too then it has a great chance on becoming a next gen Boost future. If it doesn't, then it won't.

I have been working on this replacement future design since October, with multiple presentations of my code experiments here to gain feedback. Others have presented their code experiments here too. We have all reviewed each other's design ideas and code, and evolved our own designs and code in response. If this exchange of code experiments between people all with similar problems with existing futures isn't what Boost is exactly all about, then I don't know what negative and cynical vision of Boost you have. Multiple people here have problems with futures, and multiple people are experimenting with improvements. This is something that should be welcomed, not constantly put down with negativity.

I would expect you'll see me present benchmarks here in due course once the implementation is drop in replaceable. I am expecting about a 5% performance improvement in AFIO as a drop in, and a 20% improvement once I replace AFIO's continuations infrastructure with .then() and remove the central spinlocked unordered_map. This should help further close the gap between AFIO and ASIO which is currently between 15% and 32%. That gain is what I am developing these futures for after all - to solve my problems, and maybe as a happy consequence solve other people's problems too.

Niall

-- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

Niall Douglas wrote:

However I am being successful in this persuasion - the recent changes in policy announced at C++ Now are all movements in the direction of what I have been advocating for years.

What are those recent changes in policy?

On 30 May 2015 at 0:26, Peter Dimov wrote:

...

However I am being successful in this persuasion - the recent changes in policy announced at C++ Now are all movements in the direction of what I have been advocating for years.

What are those recent changes in policy?

Topic changed as per Tony's request. https://groups.google.com/forum/#!forum/boost-steering Read all posts from May 14th onwards. There is even more exciting, possibly revolutionary, potential changes being pondered for the future, also mentioned in passing on the above list. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 30 May 2015 at 14:36, Peter Dimov wrote:

...

Read all posts from May 14th onwards. There is even more exciting, possibly revolutionary, potential changes being pondered for the future, also mentioned in passing on the above list.

I read the posts, but was actually unable to deduce what these policy changes will be. There must be something I'm missing. Perhaps you can provide a summary.

I am very hesitant to do so. Someone from the steering committee is the right person to do this, not me as I am not privy to committee thinking. However, the three key seismic changes from my perspective are (i) in how new maintainers for unmaintained libraries will now be found (ii) what will happen to unmaintained libraries for which no maintainer can be found after the new policy has been attempted (ii) the new infrastructure grants system whereby anyone can petition the steering committee with a proposal and a cash amount to be paid for its completion. Read the posts again with those three headings in mind. You'll see how profound a cultural change it is for Boost. In particular, I do not believe any of those three groups of changes could have happened whilst Dave was still here, me and him argued about those topics by private email over an extended period of approaching a decade and he was profoundly against my opinion on those issues. It was his opposition to my opinion which in part caused me to not become involved with Boost until 2012 as we would just have fought on it all the time (and on other matters we disgreed on).

On an unrelated note, it was very heart-warming to read this:

https://groups.google.com/forum/#!topic/boost-steering/WWM6nQ4szSY

and not see bpm (an existing and working implementation of what this initiative is intended to produce) get even a mention.

It isn't personal Peter, not by a long stretch. The notes written there were Robert's best effort and remembering what was discussed to help jog our memories later on. They weren't mean as a technical plan, or even for communicating with the community as to what was discussed. We were definitely fully aware of your work, in fact your work was repeatedly brought up in detail at the meeting itself and there was some discussion about how complete it was if I remember, and how much it would take to finish the job, or whether instead just to draw a line between those libraries it works for and those it does not and make two Boost distros out of that. The steering committee has gone off to ponder all those issues raised, and I would assume will decide on more stuff around CppCon time. What technical approach is actually taken isn't as important as how Boost decides to implement the resourcing of any technical approaches taken, if that makes sense. Once they decide on the resourcing method, then would start any technical approach decision making, and I would assume that would mean some form of review and writing of a preliminary report. But as I mentioned, I am not privy to committee thinking, and everything I just said is speculation. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 30 May 2015 at 17:05, Peter Dimov wrote:

...

We were definitely fully aware of your work, in fact your work was repeatedly brought up in detail at the meeting itself and there was some discussion about how complete it was if I remember, and how much it would take to finish the job, or whether instead just to draw a line between those libraries it works for and those it does not and make two Boost distros out of that.

There is no such line; it works the same way for all libraries. And it's basically finished. What remains to be done is not any work on bpm itself (to a first approximation) but tackling the remaining vestiges of centralization in the Boost release process.

Sorry, I meant by discussion of completeness whether you could arbitrarily fetch any individual Boost library with ideal dependencies fetched, and that the result compiled and worked perfectly. Some Boost libraries work perfectly with bpm and others do not, or at least someone at the meeting said so. Your personal investment in disentangling Boost libraries was mentioned, and was recognised and appreciated by all in the room. As a personal statement, whilst I disagree with the whole effort as a misallocation of resources given the in my opinion better alternatives, I absolutely take off my hat and bow to you for the investiture of effort in an area underappreciated by the community, and with likely little future recognition once it is complete. Just because I disagree with some here doesn't mean I don't totally respect them and their service. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 30 May 2015 at 17:57, Peter Dimov wrote:

...

Sorry, I meant by discussion of completeness whether you could arbitrarily fetch any individual Boost library with ideal dependencies fetched, and that the result compiled and worked perfectly. Some Boost libraries work perfectly with bpm and others do not, or at least someone at the meeting said so.

The dependency graph in my example bpm distribution doesn't track the dependencies in the test/ directories. This is on purpose as it results in fewer dependencies being installed if you just want to use the library and not run its tests.

Hang on, are you saying here that bpm is fully functional for all Boost libraries right now? If so, why aren't we shipping bpm based fully modular Boost right now?

(Bringing in Boost.Test as a dependency brings in the world.) But if you're using the ability to run the library's tests as an indication of whether bpm has worked, then yes, you'll find that it hasn't worked for some.

If the sole roadblock remaining to a bpm based fully modular Boost release is Boost.Test, then I think none of us at that meeting were aware of that. And if that is indeed the case, then that puts a *very* different spin on things indeed, and you should tell the Boost steering committee via that mailing list as that has enormous relevance to their deliberations. I'll put this another way round: most libraries using Boost.Test use like 2% of the functionality. A quick and dirty emulation of Boost.Test like the one in APIBind could replace the Boost.Test dependency for almost all libraries in Boost with relatively little work. In other words, if I have understood you right, Boost could be a fully modular distro as early as end of this year if Boost were to award a small grant for the grunt work involved. Do let me know. I am now excited. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On Saturday 30 May 2015 16:29:07 Niall Douglas wrote:

On 30 May 2015 at 17:57, Peter Dimov wrote:

...

(Bringing in Boost.Test as a dependency brings in the world.) But if you're using the ability to run the library's tests as an indication of whether bpm has worked, then yes, you'll find that it hasn't worked for some.

If the sole roadblock remaining to a bpm based fully modular Boost release is Boost.Test, then I think none of us at that meeting were aware of that.

And if that is indeed the case, then that puts a *very* different spin on things indeed, and you should tell the Boost steering committee via that mailing list as that has enormous relevance to their deliberations.

I'll put this another way round: most libraries using Boost.Test use like 2% of the functionality. A quick and dirty emulation of Boost.Test like the one in APIBind could replace the Boost.Test dependency for almost all libraries in Boost with relatively little work.

In other words, if I have understood you right, Boost could be a fully modular distro as early as end of this year if Boost were to award a small grant for the grunt work involved.

Boost.Test is only manifestation of the problem. As has been said in multiple discussions on this list, there are generally three sets of dependencies: the dependencies needed to use the library, to build it and to test it (well, there is fourth also: to build its docs, but we'll ignore it for now). The three sets can be very different. If I'm not mistaken (let Peter correct me if I am), bpm takes care of the first two sets without distinguishing between them and ignores the third one because it typically adds many more dependencies that aren't needed if you don't intend to run the library tests. Trying to emulate Boost.Test is not the solution (and we already have lightweight_test.hpp for that). I think, bpm has to distinguish between the three sets of dependencies and let the user decide what he wants to install.

As has been said in multiple discussions on this list, there are generally three sets of dependencies: the dependencies needed to use the library, to build it and to test it (well,

Andrey Semashev wrote: there is fourth also: to build its docs, but we'll ignore it for now). The three sets can be very different. If I'm not mistaken (let Peter correct me if I am), bpm takes care of the first two sets without distinguishing between them and ignores the third one because it typically adds many more dependencies that aren't needed if you don't intend to run the library tests. bpm currently does whatever the dependency file tells it to. For a source release, this dependency file would presumably tell it to fetch all libraries need to build (and it does for the example distribution). For a binary release, the dependency file would tell it to fetch all libraries needed for use. I could in principle add an option to fetch the test dependencies and use them instead. The dependency file for the example distribution is http://pdimov.com/tmp/pkg-develop-1612497/dependencies.txt.lzma created by the line dist\bin\boostdep.exe --track-sources --list-dependencies | xz --format=lzma

%OUTDIR%\dependencies.txt.lzma

in the package.bat script.

On 05/30/2015 04:36 AM, Peter Dimov wrote:

On an unrelated note, it was very heart-warming to read this:

https://groups.google.com/forum/#!topic/boost-steering/WWM6nQ4szSY

and not see bpm (an existing and working implementation of what this initiative is intended to produce) get even a mention.

The Steering Committee discussed bpm and mentioned it multiple times to Robert. We were/are encouraging the exploration/use of existing tools to achieve modularization. Beeman specifically spent some time describing bpm and his experience using the tool. I wonder if bpm got confused with bcp. michael <with-his-steering-committee-member-hat-on> -- Michael Caisse ciere consulting ciere.com

On 30 May 2015 at 0:19, Thomas Heller wrote:

However, what is proposed in this thread is hardly usable in any context where concurrent operations may occur. What you missed is the message that it is not memory allocation or the existance of exception handling code that makes the futures "slow". In fact what makes futures slow is the mechanism to start asynchronous tasks (If you have the finest task granularity you can imagine).

And that has nothing to do with futures per se. Your claim (and that of those involved with HPX) is that optimising futures beyond what HPX has done is pointless because you found no cost benefit [relative to how the rest of HPX is implemented]. Which is fine, for HPX, but in your C++ Now presentation as soon as you said your futures were always using jemalloc I instantly knew you had just undermined all your claims about your design of futures. Only FreeBSD has jemalloc as its system allocator. Generic C++ code, especially library code, can make zero assumptions about dynamic memory allocation except that it can be catastrophically slow. I mentally budget about 1000 cycles for any malloc or free call, overkill but not wrong on a low end ARM CPU. The part in the square brackets is unstated by you and your colleagues, and you and your colleagues have a nasty habit of assuming that HPX is the end all and be all of perfect design, and therefore the part in brackets need not be stated because it is now some absolute truth. You then proceed to make sweeping assumptions that all other work on futures is futile as if the case were closed, and that we are not listening to you "the experts", and that therefore discussion or even thought that you are not perfectly correct in the general design is verging on blasphemy. But let me put this to you from a Boost community perspective: 1. Why haven't any of HPX's design choices been peer reviewed here? 2. Why isn't HPX being contributed to Boost, or the parts of HPX where it would be sensible to do so? Why do HPX authors repeatedly keep the new and often very useful libraries they write away from Boost? Even the non-HPX libraries? 3. Why do those working on HPX not contribute to the Boost community in general as they once used to in the past? Where are the reviews of libraries? Where are the pull requests fixing bugs? What mentoring of GSoC projects has been done in the last three years? What gives you, or any of your colleagues, the right to lecture people here about design imprimatur when none of your designs nor code have passed peer review here, and the last time I can see any of you contributed substantially to the community here was late 2012? If you or your colleagues were the active maintainer of the definitive Boost library implementing these then your lecturing would be taken very seriously. But all your work is occurring in an ivory tower very far removed from peer review here, and the fact you as a group are not engaging with the processes nor community here substantially weakens your authority. Nobody is undervaluing the depth of experience and the talent in the HPX team, nor you as a group's previous very substantial contributions to Boost and indeed the C++ standard. I for one have a huge respect for all of you, though as a group you make it hard sometimes. Your attitude as a group of people, well it stinks, and the bad attitude was noticed at C++ Now by a number of people. If as a group you did more positive contributing to the Boost community and less negative hectoring, especially in private conversations when the person you are beating down is not there, we'd all get along much better.

What's stupid is the assumption that your task decomposition has to go down to a single instruction on a single data elemtn.

Sigh. It's like talking to a brick wall sometimes. You appear to be actively refusing to change your preheld convictions that whatever I am doing I must be wrong. This is despite at least six people contributing positive ideas and feedback to this thread, at least three of whom have written and shown the community code implementing those new ideas. For the record, it's not about task decomposition, nor was it ever. It's about affording the maximum possible scope to the compiler's optimiser by not using design patterns which get in the way. That's why turning thousands of lines of code into a single mov $5, %eax instruction is important, because that's a unit testable proxy for the right design. I am the first to realise that in 95% of the use cases the compiler can't do such dramatic reductions, but to discount such optimiser friendly design out of hand as being automatically not worth the effort seems churlish. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On 30 May 2015 at 8:57, Glen Fernandes wrote:

Why is this degenerating into a level of pettiness? What makes you so defensive about the current topic, which results in you straying into such territory? First with Hartmut and now with Thomas.

Hartmut publicly claimed here I had a secret agenda and that I am trolling, which implies I am being deceptive. Additional things have been claimed about me in private which aren't important here, and yes I am feeling very defensive towards that group who are publicly and privately politicking against me for some reason. Thomas works with Hartmut, and has in the past argued similarly to Hartmut. Theirs are not arm's length individual opinions.

You mentioned Thomas' presentation, he voiced his opinion, and because it wasn't to your favor, this thread strays into the unrelated territory of HPX developer involvement in Boost?

This is fair. I did tar Thomas with the Hartmut brush without sufficient reason. I apologise. Incidentally, of that group of people the only one I've ever talked with personally is Thomas, and it was he who initiated that not me. We disagree on almost everything :), but I'll acknowledge right now that he alone out of that group made the effort to reach out for which I am appreciative. And I reiterate my respect for that group's coding abilities. I've learned a lot from their code, and will continue to do so.

Yes, there and are more than six people interested in the discussion here. At least three people who positively affirmed your work. Two people that criticized it, or didn't find any value in it. But they stopped at the criticism of your work, and didn't question anything else (your contributions to Boost libraries, or otherwise). I hope you can do the same.

Fair enough. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

:

On 30 May 2015 at 0:19, Thomas Heller wrote:

...

However, what is proposed in this thread is hardly usable in any context where concurrent operations may occur. What you missed is

Am 30.05.2015 3:22 nachm. schrieb "Niall Douglas"

...

message that it is not memory allocation or the existance of exception handling code that makes the futures "slow". In fact what makes futures slow is the mechanism to start asynchronous tasks (If you have the finest task granularity you can imagine).

And that has nothing to do with futures per se.

Your claim (and that of those involved with HPX) is that optimising futures beyond what HPX has done is pointless because you found no cost benefit [relative to how the rest of HPX is implemented]. Which is fine, for HPX, but in your C++ Now presentation as soon as you said your futures were always using jemalloc I instantly knew you had just undermined all your claims about your design of futures. Only FreeBSD has jemalloc as its system allocator. Generic C++ code, especially library code, can make zero assumptions about dynamic memory allocation except that it can be catastrophically slow. I mentally budget about 1000 cycles for any malloc or free call, overkill but not wrong on a low end ARM CPU.

Well, using things like jemalloc is just an optimization which is not tied to freebsd. You can easily and non intrusively apply that to any existing code. The 1000 cycles for memory allocation might be a good upper bound for you. Just one question: how many cycles do you account for accessing the hard disk? Last I checked that's an order of magnitude more. So yes, I'm claiming that memory allocation plays a insignificant role in any future implementation that does more than just returning a ready future from a function.

The part in the square brackets is unstated by you and your colleagues, and you and your colleagues have a nasty habit of assuming that HPX is the end all and be all of perfect design, and therefore the part in brackets need not be stated because it is now some absolute truth. You then proceed to make sweeping assumptions that all other work on futures is futile as if the case were closed, and that we are not listening to you "the experts", and that therefore discussion or even thought that you are not perfectly correct in the general design is verging on blasphemy.

We merely state our experience with the use of an api that heavily relies on futures.

But let me put this to you from a Boost community perspective:

1. Why haven't any of HPX's design choices been peer reviewed here?

2. Why isn't HPX being contributed to Boost, or the parts of HPX where it would be sensible to do so? Why do HPX authors repeatedly keep the new and often very useful libraries they write away from Boost? Even the non-HPX libraries?

3. Why do those working on HPX not contribute to the Boost community in general as they once used to in the past? Where are the reviews of libraries? Where are the pull requests fixing bugs? What mentoring of GSoC projects has been done in the last three years?

What gives you, or any of your colleagues, the right to lecture people here about design imprimatur when none of your designs nor code have passed peer review here, and the last time I can see any of you contributed substantially to the community here was late 2012? If you or your colleagues were the active maintainer of the definitive Boost library implementing these then your lecturing would be taken very seriously. But all your work is occurring in an ivory tower very far removed from peer review here, and the fact you as a group are not engaging with the processes nor community here substantially weakens your authority.

Nobody is undervaluing the depth of experience and the talent in the HPX team, nor you as a group's previous very substantial contributions to Boost and indeed the C++ standard. I for one have a huge respect for all of you, though as a group you make it hard sometimes. Your attitude as a group of people, well it stinks, and the bad attitude was noticed at C++ Now by a number of people. If as a group you did more positive contributing to the Boost community and less negative hectoring, especially in private conversations when the person you are beating down is not there, we'd all get along much better.

OK, we might want to take that of list. This has nothing to do with that discussion and I have no clue what you are talking about here.

...

What's stupid is the assumption that your task decomposition has to go

down

...

to a single instruction on a single data elemtn.

Sigh. It's like talking to a brick wall sometimes. You appear to be actively refusing to change your preheld convictions that whatever I am doing I must be wrong. This is despite at least six people contributing positive ideas and feedback to this thread, at least three of whom have written and shown the community code implementing those new ideas.

For the record, it's not about task decomposition, nor was it ever. It's about affording the maximum possible scope to the compiler's optimiser by not using design patterns which get in the way. That's why turning thousands of lines of code into a single mov $5, %eax instruction is important, because that's a unit testable proxy for the right design. I am the first to realise that in 95% of the use cases the compiler can't do such dramatic reductions, but to discount such optimiser friendly design out of hand as being automatically not worth the effort seems churlish.

Yet you haven't shown those thousands lines of code where your efforts really make a difference. Did I miss something?

Niall

-- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

_______________________________________________ Unsubscribe & other changes:

http://lists.boost.org/mailman/listinfo.cgi/boost

On 05/26/2015 03:14 AM, Niall Douglas wrote:

On 25 May 2015 at 18:33, Avi Kivity wrote:

...

...

...

...

... turns into a "mov $5, %eax", so future<T> is now also a lightweight monadic return transport capable of being directly constructed. Can you post the code? I'd be very interested in comparing it with seastar's non-allocating futures. I may do so once I've got the functioning constexpr reduction being unit tested per commit. I'm looking forward to it! I've been bitten by the same compile time explosion problems and I'm curious to see how you solved them. Now fully functional at:

https://github.com/ned14/boost.spinlock/tree/master/test/constexprs

I have the clang and GCC disassembler dumps there too so you can see they all reduce to single opcodes on GCC only. clang isn't so good, and I'm just about to send a large bug report to LLVM (https://llvm.org/bugs/show_bug.cgi?id=23652) followed by email to Chandler. The unit testing will later do a diff compare to ensure that single opcode reduction always remains so per commit.

VS2015 RC can't cope with exception_ptr being inside an unrestricted union, so it doesn't work there yet. I have a hacky macro workaround, but I may just wait for VS2015 RTM.

Thanks! Our compile time explosions are related to chaining, which isn't there yet, so I'll wait for those. Some notes: - get() waits by spinning, how do you plan to fix that? Seems like you'd need a mutex/cond_var pair, which can dwarf the future/promise pair in terms of size and costs. - _lock_buffer isn't aligned, should use a union or aligned_storage or something. - _is_consuming causes value_storage::reset() to be called, but even if !_is_consuming, the value will be moved away, effectively consuming it. Like Peter mentioned, the compiler optimizations aren't related to constexpr, but to aggressive inlining. I expect you'll get the same generated code even if you drop all constexpr annotations.

On Tue, May 26, 2015 at 3:08 AM, Niall Douglas wrote:

...

- _lock_buffer isn't aligned, should use a union or aligned_storage or something.

I believe this doesn't matter on any major architecture as it will always be four byte aligned thanks to compiler padding.

I'm surprised that this is the response. Why would you not write an "alignas(alignof(spinlock<bool>))" for _lock_buffer (or use aligned_storage, as Avi suggested, and have correct C++ code, instead of "probably correct C++ code depending on implementation"? Is this code not part of what you intend to submit for inclusion in Boost? Glen

On 25 May 2015 at 15:51, Peter Dimov wrote:

Niall Douglas wrote:

...

https://svn.boost.org/trac/boost/wiki/BestPracticeHandbook#a8.DESIGN:Strongl...

This section is called

"(Strongly) consider using constexpr semantic wrapper transport types to return states from functions"

but I see nothing constexpr in it - the result of calling ::open can never be constexpr.

Could you perhaps elaborate a bit on the "constexpr semantic" part?

I was referring to the return type of: std::expected< std::expected< std::shared_ptr, std::error_code>, std::exception_ptr> This should constexpr reduce into no assembler output where the compiler can see the implementation. In other words, returing an error_code is as if returning a naked error_code from the point of view of opcodes generated - std::expected "disappears". Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/

On May 25, 2015 7:09:57 PM EDT, Niall Douglas wrote:

On 25 May 2015 at 23:35, Vicente J. Botet Escriba wrote:

Names suggested so far are maybe, result, holder, value.

I'm still trying to understand use cases to help guide naming. However, among those choices, "result" seems best. That said, one often refers to the result of a function, so discussing a function that returns a "result" would get awkward. Perhaps "retval" would convey the idea well enough and be less awkward?

I am simply proposing subclassing monad with a thread unsafe subclass called <insert name here>. It has almost the same API as future as it shares a common code implementation.

...

...

Options are:

* result<T> sync or async result?

A result<T> has most of the future<T> API with the set APIs from promise<T>. So you might do:

result<int> v(5); assert(v.get()==5); v.set_value(6); assert(v.get()==6); v.set_exception(foo()); v.get(); // throws foo v.set_error(error_code); v.get(); // throws system_error(error_code)

Can one ask whether it contains an exception or an error_code? Can one retrieve the exception or error_code without calling get() with, say, get_exception() and get_error_code()?

...

...

* maybe<T> we have already optional, isn't it?

True. But I think a monadic transport supersets optional as it can have no value. So:

result<int> v; assert(!v.is_ready()); assert(!v.has_value()); v.get(); // throws no_state.

Here you show has_value(). Are there has_exception() and has_error_code()? ___ Rob (Sent from my portable computation engine)

-----Original Message----- From: Boost [mailto:boost-bounces@lists.boost.org] On Behalf Of Niall Douglas Sent: 26 May 2015 11:23 To: boost@lists.boost.org Subject: Re: [boost] [next gen future-promise] What to call the monadic return type?

On 26 May 2015 at 5:07, Rob Stewart wrote:

...

On May 25, 2015 7:09:57 PM EDT, Niall Douglas wrote:

...

On 25 May 2015 at 23:35, Vicente J. Botet Escriba wrote:

Names suggested so far are maybe, result, holder, value.

I'm still trying to understand use cases to help guide naming.

Exactly why I asked for people to bikeshed here on the naming. I am also unsure.

outcome? Paul --- Paul A. Bristow Prizet Farmhouse Kendal UK LA8 8AB +44 (0) 1539 561830

On Tue, May 26, 2015 at 6:23 AM, Niall Douglas wrote:

On 26 May 2015 at 5:07, Rob Stewart wrote:

...

On May 25, 2015 7:09:57 PM EDT, Niall Douglas wrote:

...

On 25 May 2015 at 23:35, Vicente J. Botet Escriba wrote:

Names suggested so far are maybe, result, holder, value.

I'm still trying to understand use cases to help guide naming.

Exactly why I asked for people to bikeshed here on the naming. I am also unsure.

...

However, among those choices, "result" seems best. That said, one often refers to the result of a function, so discussing a function that returns a "result" would get awkward. Perhaps "retval" would convey the idea well enough and be less awkward?

One vote for result<T>. One vote for retval<T>. Okay.

I would suggest taking

...

That said, one often refers to the result of a function, so discussing a function that returns a "result" would get awkward.

as a vote *against* the name 'result'. Or take my vote as being against 'result'. Or both. Same thing came up in committee with "dumb_ptr" - any names like "raw_ptr", etc, (and in your case 'result<>') are bad because they cause confusion when spoken,

Le 26/05/15 01:09, Niall Douglas a écrit :

On 25 May 2015 at 23:35, Vicente J. Botet Escriba wrote:

...

...

However, future<T> doesn't seem named very "monadic", Why? Because we don't have mbind or the proposed next? No, merely the name "future<T>"!

future<T> is fine for a thread safe monad. But for a faster, thread unsafe one, I was asking here purely for names.

Names suggested so far are maybe, result, holder, value.

Neither result, holder nor value conveys the fact that the value can not be there. maybe<T> has already the connotation T or nothing as optional. There is no place for error_code and exception_ptr. If the type is not related to thread synchronization, here they are some alternative names: probably_value<T>, probable<T>?

...

...

so I am inclined to turn future<T> into a subclass of a type better named. Sub-classing should be an implementation detail and I don't see how a future could be a sub-class of a class that is not asynchronous itself. It's not a problem. My future<T> subclasses an internal implementation type monad which does most of the work of a monad already. monad<> has no knowledge of synchronisation.

I am simply proposing subclassing monad with a thread unsafe subclass called <insert name here>. It has almost the same API as future as it shares a common code implementation.

I had the impression that you wanted to me future a subclass of <insert name here>. I don't share this design.

...

...

Options are:

* result<T> sync or async result? A result<T> has most of the future<T> API with the set APIs from promise<T>. So you might do:

result<int> v(5); assert(v.get()==5); v.set_value(6); assert(v.get()==6); v.set_exception(foo()); v.get(); // throws foo v.set_error(error_code); v.get(); // throws system_error(error_code)

This is quite close to expected, isn't it?

...

...

* maybe<T> we have already optional, isn't it? True. But I think a monadic transport supersets optional as it can have no value. So:

result<int> v; assert(!v.is_ready()); assert(!v.has_value()); v.get(); // throws no_state.

The compiler treats a default initialised monad identically to a void return i.e. zero overhead.

Ah, then the to be named type has the additional ready state. This is closer to a valid future with the addition of the promise (setting) interface.

...

Some comments, not always directly related to your future/promise/expected design, but about the interaction between future and expected.

IMO, a future is not an expected (nor result or maybe). We can say that a ready future behaves like an expected, but a future has an additional state. Ready or not. The standard proposal and the future in Boost.Thread has yet an additional state, valid or not. So future has the following states invalid, not ready, valued or exceptional. We should be able to get an implementation that performs better if we have less states. Would the future you want have all these states? My aim is to track, as closely as possible, the Concurrency TS. Including all its bad decisions which aren't too awful. So yes, I'd keep the standard states. I agree absolutely that makes my monad not expected, nor even a proper monad. I'd call it a "bastard C++ monad type" of the kind purists dislike.

I have no problem. There are not proper and dirty monads. Purist will give you the good name. I see several nested monads. IIUC, the type doesn't has the invalid state, isn't it?

...

A future can store itself the shared state when the state is not shared, I suppose this is your idea and I think it is a good one.Let me know if I'm wrong. Clearly this future doesn't need allocators, nor memory allocation. Yes, either the promise or the future can keep the shared state. It always prefers to use the future where possible though.

...

We could have a conversion from an expected to a future. A future<T> could be constructed from an expected<T>. Absolutely agreed.

...

I believe that we could have an future operation that extracts an expected from a ready future or that it blocks until the future is ready. In the same way we have future<T>::shared() that returns a shared_future<T>, we could have a future<T>::expected() function that returns an expected<T> (waiting if needed).

If a continuation RetExpectedC returns an expected<C>, the decltype(f1) could be future<C>

auto f1 = f.then(RetExpectedC);

We could also have a when_all/match that could be applied to any probable valued type, including optional, expected, future, ...

optional<int> a; auto f4 = when_all(a, f).match( [](int i, int j ) { return 1; }, [](...) make_unexpected(MyException) ; } );

the type of f4 would be future<int>. The previous could be equivalent to

auto f4 = when_all(f).then([a](future<int> b) { return inspect(a, b.expected()).match( [](int a, int b ) { return a + b; }, [](nullopt_ i, auto const &j ) { return ???; } ); });

auto f4 = when_all(a, f).next( [](int i, int j ) { return 1; } );

but the result of when_all will be a future.

The inspect(a, b, c) could be seen as a when_all applied to probably valued instances that are all ready. expected integration is very far away for me. I'm even a fair distance from continuations, because getting a std::vector to constexpr collapse is tricky, and you need a std::vectorstd::function to hold the continuations.

Why do you need to store store a vector of continuations. You have just one. What am I missing?

My main goal is getting AFIO past peer review for now.

However, I have been speaking with Gor @ Microsoft, and if I understand how his resumable functions implementation expands into boilerplate then non-allocating future-promise means he can simplify his implementation quite considerably, and improve its efficiency. No magic tricks for future shared state allocation needed anymore. As others I'm waiting for a written proposal. But I think this is not on your plans.

Vicente

On 26 May 2015 at 19:56, Vicente J. Botet Escriba wrote:

...

...

Names suggested so far are maybe, result, holder, value. Neither result, holder nor value conveys the fact that the value can not be there. maybe<T> has already the connotation T or nothing as optional. There is no place for error_code and exception_ptr. If the type is not related to thread synchronization, here they are some alternative names: probably_value<T>, probable<T>? probable<T> vs result<T> vs maybe<T>. Hmm, I'll have to think about it.

...

...

result<int> v(5); assert(v.get()==5); v.set_value(6); assert(v.get()==6); v.set_exception(foo()); v.get(); // throws foo v.set_error(error_code); v.get(); // throws system_error(error_code) This is quite close to expected, isn't it? I lean on expected's design yes. I like the expected design. No point inventing a new API after all, we know yours works.

...

...

The compiler treats a default initialised monad identically to a void return i.e. zero overhead. Ah, then the to be named type has the additional ready state. This is closer to a valid future with the addition of the promise (setting) interface. Exactly right. No surprises.

...

I have no problem. There are not proper and dirty monads. Purist will give you the good name. I see several nested monads. IIUC, the type doesn't has the invalid state, isn't it? It has valid() on the future<T>. But not on the monad. So I suppose not. Then you to be named type is between expected and future. if it can be used asynchronously it could be seen as a valid_future<T>. Then a future coul dbe seen as a specialization of

Le 27/05/15 01:36, Niall Douglas a écrit : optional. Otherwise, if it is synchronous, it could be seen as an specialization of optional. Does your "to be named" type provide wait(), then() , ....?

...

...

expected integration is very far away for me. I'm even a fair distance from continuations, because getting a std::vector to constexpr collapse is tricky, and you need a std::vectorstd::function to hold the continuations. Why do you need to store store a vector of continuations. You have just one. What am I missing? It makes my implementation a lot easier if I can add an unknown number of continuations (non-consuming, therefore they are really C type callbacks) to some future. Remember I am optionally supporting Boost.Fiber later on, and I need an extensible and generic way of firing off coroutines on future signal with minimal #ifdef code branches all over the place.

Under the Concurrency TS you'd implement this by creating a callable struct which contains the vector of more continuations, so as an example:

struct many_continuations { std::vector> nonconsuming; std::function consuming; void operator()(std::future v) { for(auto &c : nonconsuming) c(v); } consuming(v); };

... and now do future.then(many_continuations());

I'm simply building that same functionality straight into my future instead of having external code bolt it on.

This is a functionality that shared_future<T> accepts already. Note that shared here is related to the underlying T value. As you said, don't pay for what you don't use. The many_continualions functor is only one way to achieve that. If you (the user) are looking for performances you can have a more intrusive solution. Does your "to be named" type provide this kind of consuming/not-consuming continuations? It is difficult to nane something we don't know which interface and semantics it has.

...

...

My main goal is getting AFIO past peer review for now.

As others I'm waiting for a written proposal. But I think this is not on your plans. Nope. My priority is AFIO. I have about two months to deliver it. We'll go from there after.

If we are discussing here of internal implementation details of AFIO, I would suggest you spent this time to improve the interface and the documentation and documenting the performances. You will have all the time to improve the implementation and its performances before going to a release. If we are discussing here of another possible boost classes (even if only part of AFIO), then it is different. In other words, would the "to be named" type appear on the AFIO interface? Is for this reason that you need to name it? or it is really an implementation detail. This merits clarification. Best, Vicente

On 26 May 2015 at 14:07, Peter Dimov wrote:

I also disagree with the implicit expectation that a programming style based on expected will take the world by storm. It won't. Exceptions are much more convenient and make for clearer code. C++ programmers are not Haskell programmers and don't want to be; they don't use monads and do-statements. There is no need.

Get some experience programming in Rust and come back to me. I think you'll realise that monadic programming is going to become huge in C++ 11/14 in the near future in those use case scenarios where it is far better than all other idioms. We just need a decent not-Haskell monad implementation that doesn't get in the way of C++ programming. But for sure, this is a personal opinion, and yours is just is valid as mine right now until we see where the code trends to. BTW for reference the Rust Result<> monad *does* get in the way. I think we can do a lot better in C++, something more natural to program with and use. Niall -- ned Productions Limited Consulting http://www.nedproductions.biz/ http://ie.linkedin.com/in/nialldouglas/