Unlifted standard classes

[yaitskov]

Is your feature request related to a problem? Please describe.

GHC is shipped with UnliftedDatatypes extension, but base library has similar issue like with LinearTypes extension i.e. doesn't support it well. Even basic standard classes (Show/Eq/Ord/Num) are not supported. Unlifted and Linear types are both extensions are driven by performance ghost.

There is no much action around making better support of Unlifted types. I spot 1 semi stale discussion https://github.com/ghc-proposals/ghc-proposals/pull/30 and conclusion to reimplement a library of standard classes with levity polymorphic support in 3rd party library.

https://github.com/ekmett/unboxed

I looked at linear-base it just reexports Show from Prelude.

I curious why Unlifted extension is not used in linear-base - the library already has different class hierarchy - so (e.g. how Num is defined)

Performance gain from Unlifted extension is noticable, but combining code with different levity without common classes is hard:

benchmarking Fold Lifted/Left
time                 218.2 μs   (216.1 μs .. 220.4 μs)
                     0.999 R²   (0.999 R² .. 1.000 R²)
mean                 222.1 μs   (220.4 μs .. 224.3 μs)
std dev              6.753 μs   (4.599 μs .. 11.13 μs)
variance introduced by outliers: 26% (moderately inflated)

benchmarking Fold Unlifted/Left
time                 11.63 μs   (11.50 μs .. 11.80 μs)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 11.82 μs   (11.67 μs .. 12.06 μs)
std dev              689.3 ns   (459.1 ns .. 1.024 μs)
variance introduced by outliers: 67% (severely inflated)

https://github.com/yaitskov/lifted-vs-unlifted-benchmark/tree/6354d13bf4e36fd1aa93ec70bf19c2b8cb583608

[aspiwack]

The reason for this is simple: linear-base has existed for longer than -XUnliftedDatatypes.

I haven't thought of what it would take to have better support for unlifted datatypes. What would you like to see?

PS: does the difference in performance persists if you use foldl' instead of foldl in the lifted case?

[yaitskov]

foldl' does better than just foldl, but the unlifted version keeps the lead.

benchmarking Fold Lifted Prime/Left
time                 48.69 μs   (48.18 μs .. 49.43 μs)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 48.32 μs   (47.86 μs .. 49.13 μs)
std dev              1.773 μs   (1.192 μs .. 2.662 μs)
variance introduced by outliers: 39% (moderately inflated)

benchmarking Fold Unlifted/Left
time                 11.47 μs   (11.25 μs .. 11.70 μs)
                     0.997 R²   (0.995 R² .. 0.999 R²)
mean                 11.48 μs   (11.34 μs .. 11.78 μs)
std dev              672.5 ns   (467.2 ns .. 1.016 μs)
variance introduced by outliers: 68% (severely inflated)

Both Unlifted and Linear features are elephants. It is logical they started in separated sandboxes, but eventually they should merge gradually. I would start with redefining basic classes such Show, Ord etc one by one in linear-base or reimporting from Unboxed.