Use csv's zero-copy API

[Dr-Emann]

While our use is not zero-copy, it allows us to avoid some re-allocations by reusing existing vectors.

This achieves a larger speedup in Rust 1.15, because of rust-lang/rust#38182, which specializes Vec<T>::extend when T : Copy

[emk]

Great idea! Do you have any idea how much speedup this produces? It would be fine to just capture the performance numbers printed by scrubcsv before and after your change.

I actually have some ideas on how we can eliminate those remaining allocations, and improve our handling of interior quotes, but it will require a bit more code. Basically we could have a single Vec<u8> as a working buffer, and a second Vec<&[u8]> which kept track of slices into it. Whenever we found a column, we'd append the relevant information to both vectors. This would require a fast Copy but no memory allocation in the inner loop. We'd also need to track what's going on with quotes in each CSV cell.

This should probably double or maybe even triple the speed of scrubcsv, with any luck. But your patch is a great first step!

[Dr-Emann]

On my box (Arch Linux, in a Virtualbox VM, rust 1.15), the speed before averaged at 75.51 MiB/s. After the change, the speed averaged at 90.75 MiB/s.
Both runs were averages of 3 runs.

[rossmeissl]

20% increase! Nice!

[Dr-Emann]

I don't think switching to a single Vec<u8> would considerably affect performance.

This implementation only allocates:

• For the first row, one Vec is allocated per column (at least 1KiB in size)
• If a column who's index is less than the number of columns in the first row is longer than the capacity for that column

Whereas a single Vec Implementation allocates:

• In piecemeal up to the sum of the lengths of the first row
• For every row in which the sum of the length of all columns (up to the length of the first row) exceeds the sum of lengths of previous rows.

I would guess that a single Vec implementation would be within 5-10% of this implementation.

[emk]

This was a brilliant idea, but it would need to be rethought for use with the stable csv API.