We should check on the inlining properties of hashNubOn and ordNubOn.

[phadej]

See comment #77 (comment)

On the flip side, if witherM can fuse with a "good producer" of some sort, will that be a problem? Let's investigate.

Benchmark                 bench-lazy  bench-eta          bench-noinline          
fewstring/list/hashNub       1.65e-4    1.69e-4  +2.52%         2.46e-4   +49.12%
fewstring/list/hashOpt       1.59e-4    1.63e-4  +1.94%         1.59e-4    -0.58%
fewstring/list/ordNub        3.99e-4    3.90e-4  -2.10%         5.39e-4   +35.22%
fewstring/vector/hashNub     1.72e-4    1.75e-4  +1.56%         2.35e-4   +36.29%
fewstring/vector/ordNub      4.01e-4    3.96e-4  -1.37%         5.33e-4   +32.81%
string/list/hashNub          9.50e-2    9.60e-2  +1.10%         9.76e-2    +2.75%
string/list/hashOpt          9.81e-2       0.10  +2.12%         9.81e-2    -0.01%
string/list/ordNub              0.10       0.10  +1.16%            0.13   +30.47%
string/vector/hashNub        7.96e-2    8.09e-2  +1.59%         8.51e-2    +6.82%
string/vector/ordNub            0.11       0.11  +2.23%            0.14   +31.09%
text/list/hashNub            9.18e-2    9.04e-2  -1.57%         8.83e-2    -3.88%
text/list/hashOpt            9.04e-2    9.18e-2  +1.55%         9.08e-2    +0.40%
text/list/ordNub             7.82e-2    7.40e-2  -5.48%            0.13   +66.99%
text/vector/hashNub          9.60e-2    9.36e-2  -2.48%         8.41e-2   -12.40%
text/vector/ordNub           7.98e-2    7.78e-2  -2.55%            0.13   +63.05%
word64s/list/hashNub         6.06e-2    6.26e-2  +3.35%         6.67e-2   +10.21%
word64s/list/hashOpt         5.97e-2    6.09e-2  +2.06%         6.22e-2    +4.18%
word64s/list/ordNub          5.63e-2    5.69e-2  +1.15%            0.11  +101.46%
word64s/vector/hashNub       6.35e-2    6.40e-2  +0.71%         6.82e-2    +7.33%
word64s/vector/ordNub        5.79e-2    5.74e-2  -0.79%            0.11   +90.75%
zzz-geom-mean                1.85e-2    1.85e-2  +0.31%         2.29e-2   +24.03%

Where

bench-lazy is #77
bench-eta is with

-ordNubOn p t = evalState (witherM f t) Set.empty where
+ordNubOn p = \t -> evalState (witherM f t) Set.empty where

-hashNubOn p t = evalState (witherM f t) HSet.empty
+hashNubOn p = \t -> evalState (witherM f t) HSet.empty

and bench-noinline

is additionally dropping INLINE pragmas.

hashOpt is hashNub written manually for lists (without using witherM).

---

I can add this benchmark suite to the repository if it's fine (it will make CI slower as installing criterion takes some time when cache is cold).

---

(The table is made with criterion-compare-txt from aeson repository).