perf(createSelector): gather input selector results without extra allocations

[veksa]

dependenciesChecker calls collectInputSelectorResults(dependencies, arguments) on every dependency recomputation. Inlining that loop is worth about a fifth of the call, for two separate reasons:

• arguments stops escaping. Passing it to another function forces V8 to materialize it on the heap every call. Used only as the second operand of .apply, it can stay in the frame. The dev-mode block below still hands it around, but that block is compiled out of production builds, which is the only build this matters in.
• The array is allocated at its final size instead of being grown from [] by push. This is the larger half. I assumed it would be the smaller one.

I also tried arity-specialized array literals on top of this — [d0(...)], [d0(...), d1(...)], and so on. They measured no better than a single new Array(length) loop at 1 and 3 inputs, and at 6 inputs the ladder was slightly worse than the fallback, so there's one path for every dependency count instead of a cutoff I couldn't justify.

dependencies and its length are still read per call rather than hoisted, so a selector whose dependencies array is mutated behaves as before.

Measured with the harness from #770 (paired against master in one process, min of 15 interleaved rounds, ns/call):

case	before	after
1 input, (state, props)	177.9	143.4	1.24x
3 inputs, (state, props)	213.6	182.8	1.17x
6 inputs, (state, props)	313.1	277.1	1.13x
same as first, argsMemoize: lruMemoize	77.7	70.5	1.10x
1 input, (state)	11.9	11.6	noise
nested (2 output selectors)	12.1	12.0	noise

The last two rows are selectors called with the state alone, so they hit the argument cache and never reach this code — nothing to gain there. The win shrinks as the input count grows, which makes sense: the allocation saved is fixed, the input selector calls aren't.

Recomputation counts are identical everywhere, so it's the same work throughout.

A profile is what pointed here, and it also killed my first two guesses — arity dispatch and the array allocation looked like the obvious targets, and dependenciesChecker was 1.3% of self time. The array turned out to matter anyway, but through allocation and GC rather than through the code being slow.

[netlify]

✅ Deploy Preview for reselect-docs canceled.

Name	Link
🔨 Latest commit	394247f
🔍 Latest deploy log	https://app.netlify.com/projects/reselect-docs/deploys/6a7596f7005c9f00083d6763

[codesandbox-ci]

This pull request is automatically built and testable in CodeSandbox.

To see build info of the built libraries, click here or the icon next to each commit SHA.