Speed up the Haskell implementation

[maoe]

This PR speeds up the Haskell implementation significantly using the LLVM backend.

Yesterday I saw someone wondering why the Haskell implementation was significantly slower than the other implementations in this test, so I took a stab at optimizing it.

The key points are:

• Use unboxed vectors instead of (singly linked) lists
• Replace the div/mod functions, which are known to be slow, with quot/rem
• Switch from the native code generator to the LLVM backend, which is known to generate faster code at expense of slower compilation

Here's the result on my MacBook Pro 16'':

% hyperfine {c,cpp,rust,haskell}/main
Benchmark #1: c/main
  Time (mean ± σ):      2.327 s ±  0.051 s    [User: 2.302 s, System: 0.018 s]
  Range (min … max):    2.272 s …  2.418 s    10 runs

Benchmark #2: cpp/main
  Time (mean ± σ):      3.556 s ±  0.069 s    [User: 3.519 s, System: 0.028 s]
  Range (min … max):    3.470 s …  3.688 s    10 runs

Benchmark #3: rust/main
  Time (mean ± σ):      2.633 s ±  0.079 s    [User: 2.584 s, System: 0.019 s]
  Range (min … max):    2.582 s …  2.848 s    10 runs

Benchmark #4: haskell/main
  Time (mean ± σ):      3.105 s ±  0.058 s    [User: 3.049 s, System: 0.028 s]
  Range (min … max):    3.064 s …  3.264 s    10 runs

Summary
  'c/main' ran
    1.13 ± 0.04 times faster than 'rust/main'
    1.33 ± 0.04 times faster than 'haskell/main'
    1.53 ± 0.04 times faster than 'cpp/main'

System prerequisites are:

• LLVM >= 9 && < 13
• GHC 8.10.7 and cabal-install are installed using ghcup
  • Install ghcup then ghcup install ghc 8.10.7 && ghcup install cabal latest

As noted in the commit message, the Makefile expects that llc and opt are available in $PATH. If not please set OPT and LLC accordingly when running make.

[jabbalaci]

Thanks. Nice optimization! Please check the dir. haskell/, I made some changes. I changed the any.base ==4.14.3.0 to any.base ==4.14.2.0 for instance, otherwise the compilation failed on my system.

[maoe]

I guess that's because you're probably not using GHC 8.10.7. If that change works for you, I think it's fine.

[maoe]

Ah apparently you're using 8.10.5 according to the updated README. In that case we should update

SpeedTests/haskell/Makefile

Line 13 in d4090d8

ghcup set ghc 8.10.7

accordingly.

[jabbalaci]

Found it. In the PATH, I had to put ~/.ghcup/bin before /bin, and now it finds GHC 8.10.7 . I restored any.base ==4.14.3.0 and re-run the tests. There's no change in the runtime but the EXE is much bigger than the EXE produced by 8.10.5 . Thanks for the contribution!