Re: MRP Summary & revised MRP 2ed (was: Monad of no `return` Proposal (MRP): Moving `return` out of `Monad`)

(>>) becomes slower by default in some cases and drastically faster in others, but since the definition is a member in a class it can actually be fixed by people. In situations where (<*>) is asymptotically more efficient than (>>=) then the default definition in terms of (<*>) wins. Right now, if you run through hackage there are lots of places where (>>) has been manually improved but the (*>) has not -- or vice versa. We have two places where people should apply an optimization and many have only realized that they should optimize one or the other. The key here is to encourage folks to actually define (*>) when it matters. -Edward On Wed, Nov 25, 2015 at 2:37 AM, Akio Takano wrote:

Hi Herbert,

On Thu, Nov 5, 2015 at 10:46 AM, Herbert Valerio Riedel wrote:

...

Hello everybody,

Based on the feedback gathered from the discussion, the proposal has been revised to address the raised concerns. The highlights are:

- A new stretched out transition scheme complying with the recently enacted 3-release policy (and beyond) has been devised.

- Unifying `>>`/`*>` has been incorporated into the proposal in the interest of bundling changes of obviously related changes.

- Moreover, the feasibility of automatic refactoring tooling was investigated and resulted in the working `Hs2010To201x` proof-of-concept.

Please re-read the revised proposal at

https://ghc.haskell.org/trac/ghc/wiki/Proposal/MonadOfNoReturn

for more details (or use the Wiki's diffing feature to see what changed relative to the original revision) if you want to comment, so we can focus on discussing the actual revised version.

I'm sorry to join the discussion so late, but I'd like to mention one thing that doesn't seem to have been brought up. If I understand correctly, this proposal can silently slow down existing programs, sometimes even asymptotically. This applies when a monad is defined with no explicit definition for (>>) nor for (*>).

The problem is this: currently, when a Monad instance lacks an explicit definition for (>>), a default implementation based on (>>=) is used. After this proposal, (>>) would be an alias for (*>), for which there is a default implementation based on (<*>). However, at least for some monads, the former is a much better default.

[1] is one example where it makes an asymptotic difference in runtime. This type of difference arises when (>>=) has to linearly traverse its LHS, but not its RHS.

[1] https://ghc.haskell.org/trac/ghc/ticket/10711#comment:1

I also constructed another example [2]. This is a standard implementation of the strict State monad, except that (>>=) is implemented as a NOINLINE function. You can see that using (*>) in place of (>>) changes the memory usage of the program from constant to linear. The difference here arises from the fact that the default implementation for (>>) can "tail-call" its RHS as long as (>>=) also has this tail-call property, but the default implementation of (*>) cannot.

[2] https://gist.github.com/takano-akio/7066c511b60d6ab090c5

In my opinion introducing this kind of performance regression is quite bad. Although I agree that it's frustrating that we are stuck with mapM, mapM_ etc., I believe this should be fixed in a way that doesn't make (>>) slower by default.

Regards, Takano Akio

...

Also, as per proposal guidelines, and more importantly, for the benefit of those that lost track of this rather complex discussion, I've tried to summarize the core of discussion over at

https://ghc.haskell.org/wiki/Proposal/MonadOfNoReturn/Discussion

More importantly, based on feedback gathered as well as concerns raised throughout the discussion, I've revised and extended the proposal into a "2nd edition MRP". I feel confident the new revised proposal addresses the major concerns as well as adhering to the recently requested 3-yr compatibility policy.

PS: One common oversight I noticed when reviewing the discussion is that the last-minute proposal addition -- of unifying `>>`/`*>` in the same vein as `pure`/`return` -- wasn't noticed by many who joined the debate late.

However, Unifying `>>`/`*>` and `pure`/`return` are in my opinion strongly related concerns as they share the same rationale about correctness issues and it doesn't make sense to do one without the other. However, the transition time-scales could be set slightly different to account for the more breaking nature of effectively changing `>>`'s current default method operational semantics.

Thanks, Herbert

_______________________________________________ Libraries mailing list Libraries@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries

_______________________________________________ Libraries mailing list Libraries@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries