Why is the "better" digit-listing function slower? - performance

Question

I was playing around with Project Euler #34, and I wrote these functions:
import Data.Time.Clock.POSIX
import Data.Char
digits :: (Integral a) => a -> [Int]
digits x
| x < 10 = [fromIntegral x]
| otherwise = let (q, r) = x `quotRem` 10 in (fromIntegral r) : (digits q)
digitsByShow :: (Integral a, Show a) => a -> [Int]
digitsByShow = map (\x -> ord x - ord '0') . show
I thought that for sure digits has to be the faster one, as we don't convert to a String. I could not have been more wrong. I ran the two versions via pe034:
pe034 digitFunc = sum $ filter sumFactDigit [3..2540160]
where
sumFactDigit :: Int -> Bool
sumFactDigit n = n == (sum $ map sFact $ digitFunc n)
sFact :: Int -> Int
sFact n
| n == 0 = 1
| n == 1 = 1
| n == 2 = 2
| n == 3 = 6
| n == 4 = 24
| n == 5 = 120
| n == 6 = 720
| n == 7 = 5040
| n == 8 = 40320
| n == 9 = 362880
main = do
begin <- getPOSIXTime
print $ pe034 digitsByShow -- or digits
end <- getPOSIXTime
print $ end - begin
After compiling with ghc -O, digits consistently takes .5 seconds, while digitsByShow consistently takes .3 seconds. Why is this so? Why is the function which stays within Integer arithmetic slower, whereas the function which goes into string comparison is faster?
I ask this because I come from programming in Java and similar languages, where the % 10 trick of generating digits is way faster than the "convert to String" method. I haven't been able to wrap my head around the fact that converting to a string could be faster.

Answer 1

This is the best I can come up with.
digitsV2 :: (Integral a) => a -> [Int]
digitsV2 n = go n []
where
go x xs
| x < 10 = fromIntegral x : xs
| otherwise = case quotRem x 10 of
(q,r) -> go q (fromIntegral r : xs)
when compiled with -O2 and tested with Criterion
digits runs in 470.4 ms
digitsByShow runs in 421.8 ms
digitsV2 runs in 258.0 ms
results may vary
edit:
I am not sure why building the list like this helps so much.
But you can improve your codes speed by strictly evaluating quotRem x 10
You can do this with BangPatterns
| otherwise = let !(q, r) = x `quotRem` 10 in (fromIntegral r) : (digits q)
or with case
| otherwise = case quotRem x 10 of
(q,r) -> fromIntegral r : digits q
Doing this drops digits down to 323.5 ms
edit: time without using Criterion
digits = 464.3 ms
digitsStrict = 328.2 ms
digitsByShow = 259.2 ms
digitV2 = 252.5 ms
note: The criterion package measures software performance.

Answer 2

Let's investigate why #No_signal's solution is faster.
I made three runs of ghc:
ghc -O2 -ddump-simpl digits.hs >digits.txt
ghc -O2 -ddump-simpl digitsV2.hs >digitsV2.txt
ghc -O2 -ddump-simpl show.hs >show.txt
digits.hs
digits :: (Integral a) => a -> [Int]
digits x
| x < 10 = [fromIntegral x]
| otherwise = let (q, r) = x `quotRem` 10 in (fromIntegral r) : (digits q)
main = return $ digits 1
digitsV2.hs
digitsV2 :: (Integral a) => a -> [Int]
digitsV2 n = go n []
where
go x xs
| x < 10 = fromIntegral x : xs
| otherwise = let (q, r) = x `quotRem` 10 in go q (fromIntegral r : xs)
main = return $ digits 1
show.hs
import Data.Char
digitsByShow :: (Integral a, Show a) => a -> [Int]
digitsByShow = map (\x -> ord x - ord '0') . show
main = return $ digitsByShow 1
If you'd like to view the complete txt files, I placed them on ideone (rather than paste a 10000 char dump here):
digits.txt
digitsV2.txt
show.txt
If we carefully look through digits.txt, it appears that this is the relevant section:
lvl_r1qU = __integer 10
Rec {
Main.$w$sdigits [InlPrag=[0], Occ=LoopBreaker]
:: Integer -> (# Int, [Int] #)
[GblId, Arity=1, Str=DmdType <S,U>]
Main.$w$sdigits =
\ (w_s1pI :: Integer) ->
case integer-gmp-1.0.0.0:GHC.Integer.Type.ltInteger#
w_s1pI lvl_r1qU
of wild_a17q { __DEFAULT ->
case GHC.Prim.tagToEnum# # Bool wild_a17q of _ [Occ=Dead] {
False ->
let {
ds_s16Q [Dmd=<L,U(U,U)>] :: (Integer, Integer)
[LclId, Str=DmdType]
ds_s16Q =
case integer-gmp-1.0.0.0:GHC.Integer.Type.quotRemInteger
w_s1pI lvl_r1qU
of _ [Occ=Dead] { (# ipv_a17D, ipv1_a17E #) ->
(ipv_a17D, ipv1_a17E)
} } in
(# case ds_s16Q of _ [Occ=Dead] { (q_a11V, r_X12h) ->
case integer-gmp-1.0.0.0:GHC.Integer.Type.integerToInt r_X12h
of wild3_a17c { __DEFAULT ->
GHC.Types.I# wild3_a17c
}
},
case ds_s16Q of _ [Occ=Dead] { (q_X12h, r_X129) ->
case Main.$w$sdigits q_X12h
of _ [Occ=Dead] { (# ww1_s1pO, ww2_s1pP #) ->
GHC.Types.: # Int ww1_s1pO ww2_s1pP
}
} #);
True ->
(# GHC.Num.$fNumInt_$cfromInteger w_s1pI, GHC.Types.[] # Int #)
}
}
end Rec }
digitsV2.txt:
lvl_r1xl = __integer 10
Rec {
Main.$wgo [InlPrag=[0], Occ=LoopBreaker]
:: Integer -> [Int] -> (# Int, [Int] #)
[GblId, Arity=2, Str=DmdType <S,U><L,U>]
Main.$wgo =
\ (w_s1wh :: Integer) (w1_s1wi :: [Int]) ->
case integer-gmp-1.0.0.0:GHC.Integer.Type.ltInteger#
w_s1wh lvl_r1xl
of wild_a1dp { __DEFAULT ->
case GHC.Prim.tagToEnum# # Bool wild_a1dp of _ [Occ=Dead] {
False ->
case integer-gmp-1.0.0.0:GHC.Integer.Type.quotRemInteger
w_s1wh lvl_r1xl
of _ [Occ=Dead] { (# ipv_a1dB, ipv1_a1dC #) ->
Main.$wgo
ipv_a1dB
(GHC.Types.:
# Int
(case integer-gmp-1.0.0.0:GHC.Integer.Type.integerToInt ipv1_a1dC
of wild2_a1ea { __DEFAULT ->
GHC.Types.I# wild2_a1ea
})
w1_s1wi)
};
True -> (# GHC.Num.$fNumInt_$cfromInteger w_s1wh, w1_s1wi #)
}
}
end Rec }
I actually couldn't find the relevant section for show.txt. I'll work on that later.
Right off the bat, digitsV2.hs produces shorter code. That's probably a good sign for it.
digits.hs seems to be following this psuedocode:
def digits(w_s1pI):
if w_s1pI < 10: return [fromInteger(w_s1pI)]
else:
ds_s16Q = quotRem(w_s1pI, 10)
q_X12h = ds_s16Q[0]
r_X12h = ds_s16Q[1]
wild3_a17c = integerToInt(r_X12h)
ww1_s1pO = r_X12h
ww2_s1pP = digits(q_X12h)
ww2_s1pP.pushFront(ww1_s1pO)
return ww2_s1pP
digitsV2.hs seems to be following this psuedocode:
def digitsV2(w_s1wh, w1_s1wi=[]): # actually disguised as go(), as #No_signal wrote
if w_s1wh < 10:
w1_s1wi.pushFront(fromInteger(w_s1wh))
return w1_s1wi
else:
ipv_a1dB, ipv1_a1dC = quotRem(w_s1wh, 10)
w1_s1wi.pushFront(integerToIn(ipv1a1dC))
return digitsV2(ipv1_a1dC, w1_s1wi)
It might not be that these functions mutate lists like my psuedocode suggests, but this immediately suggests something: it looks as if digitsV2 is fully tail-recursive, whereas digits is actually not (may have to use some Haskell trampoline or something). It appears as if Haskell needs to store all the remainders in digits before pushing them all to the front of the list, whereas it can just push them and forget about them in digitsV2. This is purely speculation, but it is well-founded speculation.