Surprisingly, this sailed up as the major source of B/op in the Regal
linting itself benchmark. I guess we never paid attention to it in the
past as there were many worse culprits.
With help from @tsandall we tracked this down to Regal's frequent use
of walk, and how that'd unify the path-value array. Not sure if there
are any other noteworthy scenarios that lead on to this path, but at
least us `walk` enthusiasts get to start the weekend in the best possible
way.
No surprises wrt the fix here. Just cram everything into a single struct that
can be reused across requests via a sync pool. Not particularly elegant, but
fairly limited, so I hope to get this in even if the impact should be minimal
outside of `walk` heavy integrations.
RegalLintingItself benchmark, main vs change. Ns/op and # allocs largely
unchanged, while the difference in B/op is rather massive for a simple
change!
```
348015056 ns/op 1260364874 B/op 31868705 allocs/op
343963417 ns/op 1122163613 B/op 31874931 allocs/op
```
Signed-off-by: Anders Eknert <anders.eknert@apple.com>
Anders Eknert
397f9b0245