Re: [Proposal] Strict `sum` and `product`

I don't think this example shows that a lazy fold would be better. Even if the combining argument of the fold is not strict in the second argument, evaluating something like "sum [Infinity, 1, 2, 3]" with the infinity-shorting double will still touch every element in the list. As I understand it, to actually benefit from laziness in these functions they'd need both a lazy combining function AND a lazy accumulator type (which Double is not). For a practical example of where a shortcut might be possible, consider a case like "if (product [1,2,3,0,4,5]) < 20 then f x else g y". Evaluation of the list could theoretically stop after it encounters the zero [0] since after that any further elements in the list will no longer change the value of the outcome. I don't think that this actually occurs however and doubt the various fold functions could be rewritten in such a way as to handle this in the general case as long as the accumulator type is strict. As was mentioned earlier, all of the Num types in base are strict and both `sum` and `product` have a Num constraint. If lazy shortcutting is important, a small custom function that recurses over the list and tests for problem-specific edge cases will still be possible. I disagree btw with the statement that we should not change this even though nobody can think of a compelling reason not to because we might think of such a reason later. That standpoint in its extreme can only lead to paralysis, since it is not always possible to prove no reasons will be thought of later. Davids mail came in as I was writing this, I do agree with many of his points regarding laziness but would like to point out that as Haskell was/is intended to explore laziness in programming, we should eventually be able to get the benefits from these explorations as well. Negative results can be as valuable as positive results after all, so if "we-the-library-writers" discover a place where laziness is not ideal it should be possible (encouraged even) to apply this learning and change our language for the better. In this specific case (sum and product only), the use of strict accumulator types results in strict behavior already and therefore we lose nothing by changing from lazy to strict foldl. The proposal is NOT about removing lazy folds in general, just these two particular cases where they do not bring any benefits but do bring space leaks. Wander [0]: This is an optimization we see in the implementation of Integer multiplication https://hackage.haskell.org/package/integer-gmp-1.0.3.0/docs/src/GHC.Integer.... It does not work for `product` though, only for `(*)`. Op do 22 okt. 2020 om 11:36 schreef Jon Fairbairn < jon.fairbairn@cl.cam.ac.uk>:

Sylvain Henry writes:

...

Do we have an actual example of the use of a lazy sum in the wild?

That’s not a very good argument, though I’ve seen it used quite a lot. The problem is that just because no-one can currently think of a “real” example it doesn’t mean that there are no cases where laziness is required.

For a forced example, consider a wrapping of Double with a test for infinity on the left argument

… a + b | isInfinite a = a …

which doesn’t evaluate it’s second argument if the first is infinite.

...

If the most common case is the strict one, making `sum` strict would be a better default.

Haskell should remain a lazy language, so the defaults should be lazy unless they can be proved strict anyway.

...

If need be we could also provide `lazySum` or something, but is there really a need?

Who can truly say? I’d support sum' as the strict version as that is consistent with other usages.

— Jón

...

Sylvain

On 18/10/2020 22:16, Oleg Grenrus wrote:

...

Sorry I wasn't clear myself. Option C is "add sum'"

For lists:

sum = foldr (+) 0 sum' = foldl' (+) 0

For Foldable

sum = getSum #. foldMap Sum sum' = getSum #. foldMap' Sum

- Oleg

On 18.10.2020 23.11, Hécate wrote:

...

Indeed, and I initially went to suggest a foldMap'`-based implementation to keep with the current implementation of many Foldable functions that are based on `foldMap` rather than a raw `fold`.

On 18/10/2020 22:04, Oleg Grenrus wrote:

...

For the sake of bigger audience I didn't bother mentioning #. which is a coercion helper. It's essentially better (.) when the first argument is newtype constructor (i.e. coerce).

So with Option A (strict):

sum = getSum #. foldMap' Sum

Or Option B (lazy)

sum = getSum #. foldMap Sum

---

There is also third option, Option C:

sum = foldr (+) 0 sum' = foldl' (+) 0

I don't think this is worthwhile, but it is an option. (to rehash, I don't consider maintaining status quo to be an option at all).

- Oleg On 18.10.2020 22.54, Vanessa McHale wrote:

...

It's

sum = getSum #. foldMap Sum

in base.

On 10/18/20 2:49 PM, Oleg Grenrus wrote: > > The problem is the current definition of sum for lists > which uses foldl, i.e non-strict left fold > > sum = foldl (+) 0 > > It's utterly broken. Either we should change it to > foldl' to work on some types where addition is strict, > Option A: > > sum = foldl' (+) 0 > > or alternatively (to make people using lazy > accumulator types), Option B: > > sum = foldr (+) 0 > > The current state is no good for anyone or anything. > > --- > > Related issue which Hecate didn't clearly mention, is > that Foldable class default implementation has > > class Foldable f where > ... > sum = getSum . foldMap Sum -- this is "good" lazy definition > > If we select option A, then I argue that for > consistency the default `Foldable.sum` should be > > sum = getSum . foldMap' Sum -- strict foldMap' > > If we select option B, Foldable definition doesn't need to be

changed.

...

...

...

...

...

> > --- > > I repeat, using non-strict left fold, foldl, for sum > and product is not good for anything. > Either foldr or foldl'. > > I have no strong preference. Current state is unacceptable. > > - Oleg > > On 18.10.2020 22.24, Henning Thielemann wrote: >> >> On Sun, 18 Oct 2020, Hécate wrote: >> >>> In conclusion, leaving things to the optimiser that >>> could be trivially made fast every time seems >>> needlessly risky. >> >> `seq` is still a hack. A strict 'sum' and 'product' >> would still fail on a lazy accumulator type, say a >> lazy pair type. If at all, sum and product should be >> deepseq-strict. So currently, letting the optimiser >> make a lazy sum strict is still the smaller hack. >> >> _______________________________________________ >> 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

_______________________________________________ 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 -- Hécate ✨ IRC: Uniaika WWW:https://glitchbra.in RUN: BSD

_______________________________________________ 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

---

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

-- Jón Fairbairn Jon.Fairbairn@cl.cam.ac.uk http://www.chaos.org.uk/~jf/Stuff-I-dont-want.html (updated 2014-04-05)

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