Use Set.alterF in nubOrd

[meooow25]

Compared to the previous implementation of member followed by insert, alterF uses memberIndex and insertAt. This avoids all the element comparisons in insert, at the cost of memberIndex being a little more expensive than member. I expect this change to be an improvement in typical use cases.

Related: #1205

---

Benchmarks with GHC 9.14:

Name                                Time - - - - - - - -    Allocated - - - - -
                                         A       B     %         A       B     %
nubOrd.issue1202_distinct            48 μs   46 μs   -4%    672 KB  672 KB   +0%
nubOrd.issue1202_repeat             3.8 μs  4.0 μs   +3%     16 KB   16 KB   +0%
nubOrd.noFusion_distinct             44 μs   43 μs   -1%    633 KB  633 KB   +0%
nubOrd.noFusion_repeat              1.5 μs  1.9 μs  +19%    152 B   152 B    +0%
nubOrd.strings_noFusion_distinct    166 μs  112 μs  -32%    611 KB  626 KB   +2%
nubOrd.strings_noFusion_repeat      3.6 μs  4.2 μs  +16%    136 B   152 B   +11%

[sjakobi]

I think it's very good to reduce the number of Ord comparisons, but it would be better to if the new implementation didn't allocate all these Maybes.

[meooow25]

Ah but it doesn't, the Maybes disappear after optimizations. You'll notice that the allocations numbers haven't increased (except for nubOrd.strings_noFusion_repeat, I haven't checked what's up with that but the Core looks good so I am satisfied).

[sjakobi]

Ah but it doesn't, the Maybes disappear after optimizations. You'll notice that the allocations numbers haven't increased (except for nubOrd.strings_noFusion_repeat, I haven't checked what's up with that but the Core looks good so I am satisfied).

Ah, I missed that! And it appears that you managed to reduce the allocations in #1229! 👍