precompute offsets ahead-of-time rather than on each dereference

[Ekleog-NEAR]

Criterion results on a test related to near/nearcore-private-1#2:

compile                 time:   [9.7873 ms 9.8477 ms 9.9119 ms]
                        change: [-5.2388% -4.2981% -3.2369%] (p = 0.00 < 0.05)
                        Performance has improved.

Manual testing on the exact test of near/nearcore-private-1#2 confirms the ~5% speedup:

Non-rayon goes down to 24.5-25ms from 25.5-26ms
Rayon goes down to 12ms from 12.5ms

This also feels more like a code cleanup than added complexity for optimization’s sake to me, but I would understand if other people thought the added invariants are added complexity

[matklad]

Rayon goes down to 12ms from 12.5ms

I think we agrdeed to just kill rayon (near/nearcore#8948), could you send a quick PR to do that as a follow up?

[nagisa]

This is so much more comprehensible than the previous code already, but I do wonder if there’s an opportunity to further simplify this code with some helpers.

For instance: instead of manually having checked_add and checked_mul, having that in a utility function along the lines of fma (fused multiply-add).

That way we’d be looking at something like

self.vmctx_imported_functions_begin = offset_fma(
    self.num_signature_ids, 
    u32::from(self.size_of_vmshared_signature_index())
    self.vmctx_signature_ids_begin
);

which seems like less noise overall. (names/argument ordering at your choice)

[Ekleog-NEAR]

Great idea thanks! I’ve just pushed a commit doing this change :)