pkgs/top-level/unit-files.nix: Benchmark concatMap-based evaluation

pkgs/top-level/unit-files.nix

@@ -78,8 +78,24 @@ let
   # m  m m  m m  m m  m
   #
   # Giving a total complexity of only O(log2(n)nm)
+
+  # Another potential approach is for `namesForShard` to return [ { name, value } ] instead, and then use builtins to merge it together:
+  #
+  #   listToAttrs (concatMap namesForShard shards)
+  #
+  # This way the shard merging can be done by Nix's C++ code.
+
+  # We benchmark the three approaches with the following command with 300 runs:
+  #
+  #   NIX_SHOW_STATS=1 nix-instantiate --eval --strict 'builtins.seq (import pkgs/top-level/unit-files.nix) null'
+  # Approach     | cpuTime [ms] | gc.totalBytes [MB]
+  # linear merge | 33.4 ± 4.8   | 126.075
+  # concatMap    | 21.9 ± 2.6   |   5.824
+  # binary merge | 16.6 ± 2.5   |   7.102
   #
-  # Benchmarking with the top-level attribute distribution in Nixpkgs shows that doing this reduces evaluation time by ~47% (from 36ms to 19ms) and allocated bytes by ~94% (from ~120MB to ~7MB)
+  # The lowest evaluation time is achieved with the binary merge algorithm, even despite the slightly higher memory usage.
+  # However if https://github.com/NixOS/nix/issues/7676 were implemented in the future, the concatMap-based approach could be much more performant.
+  # Alternatively, other Nix builtins to speed up this computation would help too.
 
 in
 if shards == [ ] then