Add RULES for Data.IntMap.alterF

[m-renaud]

Adds afterF rewrite RULES for the Identity functor.

Effect on Performance

Identity

Since Identity a is isomorphic to a we can rewrite alterF to to use the more efficient alter which has complexity O(min(n,w)) compared to O(log n) of alterF. alter also has a smaller constant factor because it only needs to traverse the map once whereas alterF must do so twice (once to find the element and then again to delete).

Testing

• Adds unit tests for alterF for [], Identity, and Const
• Property test comparing the rewritten versions to versions operating on isomorphic structures

Benchmark

Benchmarks performed using GHC 8.2.2

benchmarking alterF_IdentityNoRewrite
time                 1.101 ms   (1.089 ms .. 1.113 ms)
                     0.999 R²   (0.998 R² .. 1.000 R²)
mean                 1.099 ms   (1.091 ms .. 1.111 ms)
std dev              31.16 μs   (22.66 μs .. 48.93 μs)
variance introduced by outliers: 18% (moderately inflated)
               
benchmarking alterF_IdentityRewrite
time                 903.7 μs   (865.3 μs .. 946.8 μs)
                     0.989 R²   (0.981 R² .. 0.996 R²)
mean                 864.9 μs   (848.6 μs .. 886.1 μs)
std dev              63.00 μs   (50.28 μs .. 95.92 μs)
variance introduced by outliers: 59% (severely inflated)
               
benchmarking alterF_Const
time                 19.25 μs   (18.28 μs .. 20.68 μs)
                     0.971 R²   (0.962 R² .. 0.985 R²)
mean                 20.66 μs   (19.74 μs .. 21.61 μs)
std dev              3.234 μs   (2.766 μs .. 3.837 μs)
variance introduced by outliers: 94% (severely inflated)

No rewrite RULES were added for Const because the benchmark results showed no meaningful difference.

This addresses #325.

[m-renaud]

Latest failure is due to CI issue, failed to install ghc for one of the versions.

[treeowl]

Very nice work so far! Just a few little quibbles.

[treeowl]

@m-renaud, you can click the restart button (circle with arrow) next to the test that failed in the Travis interface. Or at least I think you can. If not, let me know and I'll see about getting permissions fixed.

[m-renaud]

Found it, thanks! I originally wasn't logged in and was on my phone so it wasn't very apparent. Addressed your comments, PTAL.

[treeowl]

No need for coerce here! If you really need Const . getConst here, then use it, but maybe give it a comment or a type signature. We want to use coerce only when needed to prevent silly eta expansion from adding closures and indirection at runtime.

[treeowl]

Or, of course, to re-type a whole structure from the outside.

[m-renaud]

Yeah, either Const . getConst or coerce is needed because f :: Maybe a -> Const c (Maybe a) but the function returns Const c (IntMap a). Since the second type parameter for Const is phantom I would have thought that would be an appropriate use of coerce, but I also just learned about Coercible this morning so I may very well be wrong 😛.

I've changed it back to Const . getConst but I was wondering, would Const . getConst have runtime overhead or should the optimizer figure out that this is just changing a phantom parameter? It seems to me that Const . getConst is equivalent to coerce id for Const.

[treeowl]

Const, getConst, Const . getConst, and runIdentity should all be very reliably equivalent to coerce, at least when optimizations are enabled. The issue that comes with the Identity rule is that

runIdentity . f = coerce . f =?= coerce f

Why is that last equality questionable?

coerce . undefined = \x -> coerce (undefined x) = \ _ -> undefined  /=   undefined

Yuck! Because of this, GHC will (generally at least) only turn g (runIdentity . f) into coerce g f if it either

1. Knows that f has an arity of at least 1, or
2. Knows that g only uses f by applying it to an argument, and doesn't do something shady like using seq on it. I believe this currently requires that g be inlined, but I could be wrong.

In practice, GHC is pretty good at this stuff, but sometimes it goofs. In those circumstances where there's a risk, it's probably best to help it out. In first-order circumstances, there's generally no reason to bother (and make the code less clear).

[m-renaud]

Thanks for the explanation! Do you have any more comments, or is this ready to be merged?

[treeowl]

I don't understand the benchmark results. What does each of them mean? Please clarify in the commit message and PR text. How does the rewrite rule change performance for Identity and for Const? Be careful; if you define your own versions of Identity and Const to test with, be sure to copy the Functor and Applicative instances from base; there's some coercion magic about, as I recall.

[treeowl]

For Identity:

instance Functor Identity where
  fmap = coerce
instance Applicative Identity where
  pure = Identity
  (<*>) = coerce
  liftA2 = coerce

For Const,

instance Monoid m => Applicative (Const m) where
    pure _ = Const mempty
    liftA2 _ (Const x) (Const y) = Const (x `mappend` y)
    (<*>) = coerce (mappend :: m -> m -> m)

[m-renaud]

Added a section to the PR text, I can add it to the squashed commit when the branch is merged as well.

be sure to copy the Functor and Applicative instances from base

Good catch, thanks for pointing that out! The TestConst functor instance was the same, but I've updated TestIdentityto use coerce. There's no need to define an applicative instance because alterF only requires Functor.

[treeowl]

No, I mean what does each line in the benchmark results actually represent? Most especially, what does the alterF line mean?

[m-renaud]

Ohhh, sorry, I misunderstood. I'll rename (and rework the benchmarks) as soon as I get a chance. I was thinking:

• alterF_IdentityNoRewrite (using TestIdentity)
• alterF_IdentityRewrite (using Identity)
• alterF_ConstNoRewrite (using TestConst)
• alterF_ConstRewrite (using Const)

Does that sound better?

[treeowl]

Yes, much. On Jan 4, 2018 11:16 AM, "Matt Renaud" <[email protected]> wrote: Ohhh, sorry, I misunderstood. I'll rename (and rework the benchmarks) as soon as I get a chance. I was thinking: - alterF_IdentityNoRewrite (using TestIdentity) - alterF_IdentityRewrite (using Identity) - alterF_ConstNoRewrite (using TestConst) - alterF_ConstRewrite (using Const) Does that sound better? — You are receiving this because your review was requested. Reply to this email directly, view it on GitHub <#467 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABzi_cXl_HsvhkxRTZSHW-Plp-DMFVT8ks5tHPljgaJpZM4RPdHz> .

[treeowl]

You have bona fide CI failures. It looks like you forgot to import Data.Coerce (coerce) some places.

[treeowl]

By the way, there are buttons to restart or cancel CI runs on the Travis page. Do you see those? Please restart your own when there are silly problems, and please try to cancel your own when they're out of date or doomed to fail. Canceling useless builds saves time for the whole Haskell organization. Speaking of which, as I think I said, you should use [ci skip] where appropriate. In particular, your whole tutorial project should be [ci skip] unless and until we make Travis build those docs. Since we don't even make Travis build the Haddocks, I don't think that's going to happen anytime soon.

[treeowl]

Can you fix up the tiny problem so this can pass CI? I want it merged!

[m-renaud]

Sorry for the delay, I only had a few minutes here and there over the weekend to look at this stuff. I've re-organized the benchmarks as discussed and fixed the import.

It also appears that the import for Data.Coerce in IntMap/Internal.hs was redundant since it's been included in __GLASGOW_HASKELL__ >= 709, do I need to add another conditional to import if base >= 4.8 and GLASGOW_HASKELL > 709?

[treeowl]

Um.... I wasn't telling you to just change that. I was suggesting that you check what that does to the benchmarks. I want to understand why there's a difference for Const, which I don't think there should be.

[m-renaud]

I originally had a todo to see if alterF should be marked INLINABLE [1] or NOINLINE [1] but it somehow got clobbered (I went with NOINLINE to be safe). The difference in Const mostly goes away (see benchmark results in commit description), I think the remaining difference may just be noise (but I wanted to see what you thought). With such a small difference I'm not sure the custom rules for Const are necessary now, wdyt?

[treeowl]

Oh, I missed the new benchmark results in your commit. I think you should run more tests. See what happens if you reverse the order the benchmarks run in, for example. If the difference remains, it would be nice to understand why, but we should probably keep the rule. If the difference is really just noise, scrap the rule.

[m-renaud]

I ran it several more times and changed the order of benchmarks but the results remain about the same.

[treeowl]

Interesting. Let's keep the rule for now but maybe try to dig into Core and such some other time. We should get this merged.

…

On Mon, Jan 8, 2018 at 4:06 PM, Matt Renaud ***@***.***> wrote: I ran it several more times and changed the order of benchmarks but the results remain about the same. — You are receiving this because your review was requested. Reply to this email directly, view it on GitHub <#467 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABzi_cuF1OY557-Wgz68NGXgefyI3o_lks5tIoNugaJpZM4RPdHz> .

[m-renaud]

Woah woah woah, I ran it a few more times for good measure and got the below results which indicate that the rewrite is now slower than the non-rewritten version :/

benchmarking alterF_ConstRewrite
time                 23.26 μs   (22.48 μs .. 23.70 μs)
                     0.993 R²   (0.986 R² .. 0.997 R²)
mean                 22.07 μs   (21.30 μs .. 22.63 μs)
std dev              2.207 μs   (1.692 μs .. 2.816 μs)
variance introduced by outliers: 85% (severely inflated)
             
benchmarking alterF_ConstNoRewrite
time                 18.75 μs   (17.94 μs .. 19.86 μs)
                     0.978 R²   (0.968 R² .. 0.992 R²)
mean                 20.02 μs   (19.11 μs .. 21.07 μs)
std dev              3.218 μs   (2.715 μs .. 3.618 μs)
variance introduced by outliers: 94% (severely inflated)

[treeowl]

This sounds like some sort of silliness. Maybe some code alignment nonsense or something. It sounds like they're probably pretty much the same in reality, so get rid of the Const rule and we can merge.

…

On Mon, Jan 8, 2018 at 5:22 PM, Matt Renaud ***@***.***> wrote: Woah woah woah, I ran it a few more times for good measure and got the below results which indicate that the rewrite is now slower than the non-rewritten version :/ benchmarking alterF_ConstRewrite time 23.26 μs (22.48 μs .. 23.70 μs) 0.993 R² (0.986 R² .. 0.997 R²) mean 22.07 μs (21.30 μs .. 22.63 μs) std dev 2.207 μs (1.692 μs .. 2.816 μs) variance introduced by outliers: 85% (severely inflated) benchmarking alterF_ConstNoRewrite time 18.75 μs (17.94 μs .. 19.86 μs) 0.978 R² (0.968 R² .. 0.992 R²) mean 20.02 μs (19.11 μs .. 21.07 μs) std dev 3.218 μs (2.715 μs .. 3.618 μs) variance introduced by outliers: 94% (severely inflated) — You are receiving this because your review was requested. Reply to this email directly, view it on GitHub <#467 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABzi_bdxGFiD5IRtsrpj0cilZMyprF8gks5tIpUZgaJpZM4RPdHz> .

[m-renaud]

Done. I've left a few of the tests and benchmarks in case we want to revisit this in the future.

[treeowl]

Waait a minute. You're only testing one way of using alterF! Please copy over the relevant Data.Map benchmarks and test properly! Also, pure . id is the same as pure.....

[treeowl]

Also, please indicate in the commit what GHC version you used for benchmarking, and use GHC >= 8.2. I don't know if the new join point system affects this stuff, but I would not be at all surprised. NB: when you squash this thing, make sure the final commit message is the one with the latest/correctest benchmarks.

[m-renaud]

Updated PR description with new benchmark results and note on what GHC version used. Is rewording an already pushed commit allowed? I thought it may mess up history or something.

[treeowl]

Speaking of GHC changes, I should probably rerun the benchmarks that led to my choice of alterF Identity specialization for Data.Map.

[treeowl]

You've pushed these commits to a branch. You're absolutely free to amend them, rebase them, etc., etc. Once they're actually merged to master, neither you nor I has the authority to touch them. BTW, my git-fu is horrible (I made a very serious hash of the git history of Data.Sequence last year by moving a file the wrong way :-( ). But I can tell you that `git rebase -i` is *very useful*. Just type `git rebase -i base`, where base is the last commit before your branch, and you'll get a pretty much totally self-explanatory way to fix everything. If you only want one commit, you can set the first one to "pick" and all the rest to "squash"; this will open an editor with all the texts of all your commits, and you can delete/condense/etc. to get one good message. If you're doing something where multiple commits make sense (here, that might be one commit for the benchmarks followed by another for the changes), you can pick a couple, squash others, etc. If you want to modify something and push that back in time, you can add a new commit for it, rebase -i, and move the new commit up higher and then (perhaps) squash it onto a previous one.

…

On Mon, Jan 8, 2018 at 7:31 PM, Matt Renaud ***@***.***> wrote: Updated PR description with new benchmark results and note on what GHC version used. Is rewording an already pushed commit allowed? I thought it may mess up history or something. — You are receiving this because your review was requested. Reply to this email directly, view it on GitHub <#467 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABzi_X7nBX85o5KkNglpCR3fLe37mMbRks5tIrNjgaJpZM4RPdHz> .

[treeowl]

You still need the additional benchmarks ported from Data.Map.

[m-renaud]

Sooo, these benchmarks tell a different story than the other ones. TL;DR; rewrite rules perform
better in casses when the element being altered is present, but worse when the element is absent. See the commit message for more details.

[treeowl]

Hrmmm. Yeah, this kind of weirdness complicated my choices for Map as well. I ended up rewriting the Identity case not to alter but to something that (for reasons I no longer remember) had performance characteristics somewhat more closely matched to the alterF version. Or something like that. Been a long time. Maybe you can come up with something? I think we should probably consider this longer and probably let it be till the next release. The benefit isn't nearly as great as for Map because IntMap lookups are really, really cheap.

[m-renaud]

Sounds good, I'll keep poking at it and see if I can come up with something better. And as you said, we shouldn't block the next release on this.

…

On Tue, Jan 9, 2018, 10:08 AM David Feuer ***@***.***> wrote: Hrmmm. Yeah, this kind of weirdness complicated my choices for Map as well. I ended up rewriting the Identity case not to alter but to something that (for reasons I no longer remember) had performance characteristics somewhat more closely matched to the alterF version. Or something like that. Been a long time. Maybe you can come up with something? I think we should probably consider this longer and probably let it be till the next release. The benefit isn't nearly as great as for Map because IntMap lookups are really, really cheap. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#467 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAqw5mcwn4tQPY0v_N6s8BcNJ7rC45e1ks5tI6sfgaJpZM4RPdHz> .

[treeowl]

The biggest reason you're running into trouble is that if the passed function takes Nothing to Nothing and the key is absent, then

• alterF does nothing
• alter rebuilds a path in the map

One option is to use pointer equality in alter the way we do for many functions in Data.Map: if the new subtree is the same as the old, we return our argument unchanged. Another thing to experiment with is GHC's new unboxed sum feature.

type Maybe# a = (# Void# | a #)

pattern Nothing# :: Maybe# a
pattern Nothing# <- (# _ | #) where
  Nothing# = (# void# | #)

pattern Just# :: a -> Maybe# a
pattern Just# a = (# | a #)

Now we can make recursive calls in alter return Just# a new map or Nothing#, without having to allocate a real Just constructor like we would for Maybe. We'll have to experiment to see how that performs.

[treeowl]

To avoid confusion: Void# is not an unboxed version of Void. For hysterical raisins, it's an unboxed version of ().

[m-renaud]

So just using ptrEq in alter's zero and otherwise case makes the benchmarks worse (somewhat expected, since it's doing a ptrEq at every level). I'm going to try using unboxed sum approach and see if that helps.

[treeowl]

One option is to implement alter the way it's done in containers--by hand. I'm not sure if it's worth it.

[m-renaud]

Well, if anything I was able to make alter twice as fast using unboxed sums :) I'll move that to its own PR.