efficiently checking that all the elements of a (big) list are the same

Question

Problem

Let us suppose that we have a list xs (possibly a very big one), and we want to check that all its elements are the same.

I came up wi

Answer 1

First of all, I don't think you want to be working with lists. A lot of your algorithms rely upon calculating the length, which is bad. You may want to consider the vector package, which will give you O(1) length compared to O(n) for a list. Vectors are also much more memory efficient, particularly if you can use Unboxed or Storable variants.

That being said, you really need to consider traversals and usage patterns in your code. Haskell's lists are very efficient if they can be generated on demand and consumed once. This means that you shouldn't hold on to references to a list. Something like this:

average xs = sum xs / length xs

requires that the entire list be retained in memory (by either sum or length) until both traversals are completed. If you can do your list traversal in one step, it'll be much more efficient.

Of course, you may need to retain the list anyway, such as to check if all the elements are equal, and if they aren't, do something else with the data. In this case, with lists of any size you're probably better off with a more compact data structure (e.g. vector).

Now that this is out of they way, here's a look at each of these functions. Where I show core, it was generated with ghc-7.0.3 -O -ddump-simpl. Also, don't bother judging Haskell code performance when compiled with -O0. Compile it with the flags you would actually use for production code, typically at least -O and maybe other options too.

Solution 0

allTheSame :: (Eq a) => [a] -> Bool
allTheSame xs = and $ map (== head xs) (tail xs)

GHC produces this Core:

Test.allTheSame
  :: forall a_abG. GHC.Classes.Eq a_abG => [a_abG] -> GHC.Bool.Bool
[GblId,
 Arity=2,
 Str=DmdType LS,
 Unf=Unf{Src=, TopLvl=True, Arity=2, Value=True,
         ConLike=True, Cheap=True, Expandable=True,
         Guidance=IF_ARGS [3 3] 16 0}]
Test.allTheSame =
  \ (@ a_awM)
    ($dEq_awN :: GHC.Classes.Eq a_awM)
    (xs_abH :: [a_awM]) ->
    case xs_abH of _ {
      [] ->
        GHC.List.tail1
        `cast` (CoUnsafe (forall a1_axH. [a1_axH]) GHC.Bool.Bool
                :: (forall a1_axH. [a1_axH]) ~ GHC.Bool.Bool);
      : ds1_axJ xs1_axK ->
        letrec {
          go_sDv [Occ=LoopBreaker] :: [a_awM] -> GHC.Bool.Bool
          [LclId, Arity=1, Str=DmdType S]
          go_sDv =
            \ (ds_azk :: [a_awM]) ->
              case ds_azk of _ {
                [] -> GHC.Bool.True;
                : y_azp ys_azq ->
                  case GHC.Classes.== @ a_awM $dEq_awN y_azp ds1_axJ of _ {
                    GHC.Bool.False -> GHC.Bool.False; GHC.Bool.True -> go_sDv ys_azq
                  }
              }; } in
        go_sDv xs1_axK
    }

This looks pretty good, actually. It will produce an error with an empty list, but that's easily fixed. This is the case xs_abH of _ { [] ->. After this GHC performed a worker/wrapper transformation, the recursive worker function is the letrec { go_sDv binding. The worker examines its argument. If [], it's reached the end of the list and returns True. Otherwise it compares the head of the remaining to the first element and either returns False or checks the rest of the list.

Three other features.

1. The map was entirely fused away and doesn't allocate a temporary list.
2. Near the top of the definition notice the Cheap=True statement. This means GHC considers the function "cheap", and thus a candidate for inlining. At a call site, if a concrete argument type can be determined, GHC will probably inline allTheSame and produce a very tight inner loop, completely bypassing the Eq dictionary lookup.
3. The worker function is tail-recursive.

Verdict: Very strong contender.

Solution 1

allTheSame' :: (Eq a) => [a] -> Bool
allTheSame' xs = (length xs) == (length $ takeWhile (== head xs) xs)

Even without looking at core I know this won't be as good. The list is traversed more than once, first by length xs then by length $ takeWhile. Not only do you have the extra overhead of multiple traversals, it means that the list must be retained in memory after the first traversal and can't be GC'd. For a big list, this is a serious problem.

Test.allTheSame'
  :: forall a_abF. GHC.Classes.Eq a_abF => [a_abF] -> GHC.Bool.Bool
[GblId,
 Arity=2,
 Str=DmdType LS,
 Unf=Unf{Src=, TopLvl=True, Arity=2, Value=True,
         ConLike=True, Cheap=True, Expandable=True,
         Guidance=IF_ARGS [3 3] 20 0}]
Test.allTheSame' =
  \ (@ a_awF)
    ($dEq_awG :: GHC.Classes.Eq a_awF)
    (xs_abI :: [a_awF]) ->
    case GHC.List.$wlen @ a_awF xs_abI 0 of ww_aC6 { __DEFAULT ->
    case GHC.List.$wlen
           @ a_awF
           (GHC.List.takeWhile
              @ a_awF
              (let {
                 ds_sDq :: a_awF
                 [LclId, Str=DmdType]
                 ds_sDq =
                   case xs_abI of _ {
                     [] -> GHC.List.badHead @ a_awF; : x_axk ds1_axl -> x_axk
                   } } in
               \ (ds1_dxa :: a_awF) ->
                 GHC.Classes.== @ a_awF $dEq_awG ds1_dxa ds_sDq)
              xs_abI)
           0
    of ww1_XCn { __DEFAULT ->
    GHC.Prim.==# ww_aC6 ww1_XCn
    }
    }

Looking at the core doesn't tell much beyond that. However, note these lines:

case GHC.List.$wlen @ a_awF xs_abI 0 of ww_aC6 { __DEFAULT ->
        case GHC.List.$wlen

This is where the list traversals happen. The first gets the length of the outer list and binds it to ww_aC6. The second gets the length of the inner list, but the binding doesn't happen until near the bottom, at

of ww1_XCn { __DEFAULT ->
GHC.Prim.==# ww_aC6 ww1_XCn

The lengths (both Ints) can be unboxed and compared by a primop, but that's a small consolation after the overhead that's been introduced.

Verdict: Not good.

Solution 2

allTheSame'' :: (Eq a) => [a] -> Bool
allTheSame'' xs
  | n == 0 = False
  | n == 1 = True
  | n == 2 = xs !! 0 == xs !! 1
  | otherwise = (xs !! 0 == xs !! 1) && (allTheSame'' $ snd $ splitAt 2 xs)
    where  n = length xs

This has the same problem as solution 1. The list is traversed multiple times, and it can't be GC'd. It's worse here though, because now the length is calculated for each sub-list. I'd expect this to have the worst performance of all on lists of any significant size. Also, why are you special-casing lists of 1 and 2 elements when you're expecting the list to be big?

Verdict: Don't even think about it.

Solution 3

allTheSame''' :: (Eq a) => [a] -> Bool
allTheSame''' xs
  | n == 0 = False
  | n == 1 = True
  | n == 2 = xs !! 0 == xs !! 1
  | n == 3 = xs !! 0 == xs !! 1 && xs !! 1 == xs !! 2
  | otherwise = allTheSame''' (fst split) && allTheSame''' (snd split)
    where n = length xs
          split = splitAt (n `div` 2) xs

This has the same problem as Solution 2. Namely, the list is traversed multiple times by length. I'm not certain a divide-and-conquer approach is a good choice for this problem, it could end up taking longer than a simple scan. It would depend on the data though, and be worth testing.

Verdict: Maybe, if you used a different data structure.

Solution 4

allTheSame'''' :: (Eq a) => [a] -> Bool
allTheSame'''' xs = all (== head xs) (tail xs)

This was basically my first thought. Let's check the core again.

Test.allTheSame''''
  :: forall a_abC. GHC.Classes.Eq a_abC => [a_abC] -> GHC.Bool.Bool
[GblId,
 Arity=2,
 Str=DmdType LS,
 Unf=Unf{Src=, TopLvl=True, Arity=2, Value=True,
         ConLike=True, Cheap=True, Expandable=True,
         Guidance=IF_ARGS [3 3] 10 0}]
Test.allTheSame'''' =
  \ (@ a_am5)
    ($dEq_am6 :: GHC.Classes.Eq a_am5)
    (xs_alK :: [a_am5]) ->
    case xs_alK of _ {
      [] ->
        GHC.List.tail1
        `cast` (CoUnsafe (forall a1_axH. [a1_axH]) GHC.Bool.Bool
                :: (forall a1_axH. [a1_axH]) ~ GHC.Bool.Bool);
      : ds1_axJ xs1_axK ->
        GHC.List.all
          @ a_am5
          (\ (ds_dwU :: a_am5) ->
             GHC.Classes.== @ a_am5 $dEq_am6 ds_dwU ds1_axJ)
          xs1_axK
    }

Ok, not too bad. Like solution 1, this will error on empty lists. The list traversal is hidden in GHC.List.all, but it will probably be expanded to good code at a call site.

Verdict: Another strong contender.

So between all of these, with lists I'd expect that Solutions 0 and 4 are the only ones worth using, and they are pretty much the same. I might consider Option 3 in some cases.

Edit: in both cases, the errors on empty lists can be simply fixed as in @augustss's answer.

The next step would be to do some time profiling with criterion.