perf: outline and make reserve as cold

[fereidani]

This outlines and mark reserve as cold as it rarely happens during execution of a program.
It helps both with performance and binary size.

Benchmarks results are mixed, some improvements and some regressions, but overall as rust standard library does this too, I think these changes are good for the project:

     Running benches/bench.rs (target/release/build/smallvec/30c2fd0ca48816cd/out/bench-30c2fd0ca48816cd)
Gnuplot not found, using plotters backend
bench_push              time:   [382.76 ns 387.09 ns 392.10 ns]
                        change: [+10.972% +13.856% +16.887%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 14 outliers among 100 measurements (14.00%)
  14 (14.00%) high mild

bench_push_small        time:   [47.911 ns 48.409 ns 48.958 ns]
                        change: [−9.7960% −7.7322% −5.6598%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 3 outliers among 100 measurements (3.00%)
  3 (3.00%) high mild

bench_insert_push       time:   [460.18 ns 469.63 ns 479.96 ns]
                        change: [+1.5551% +4.3700% +7.4103%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 12 outliers among 100 measurements (12.00%)
  2 (2.00%) high mild
  10 (10.00%) high severe

bench_insert_push_small time:   [58.995 ns 59.779 ns 60.670 ns]
                        change: [+5.9883% +7.1482% +8.4367%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 3 outliers among 100 measurements (3.00%)
  2 (2.00%) high mild
  1 (1.00%) high severe

bench_pushpop           time:   [316.26 ns 320.23 ns 324.42 ns]
                        change: [−11.452% −9.5607% −7.6188%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 5 outliers among 100 measurements (5.00%)
  4 (4.00%) high mild
  1 (1.00%) high severe

bench_push_vec          time:   [330.65 ns 334.85 ns 339.32 ns]
                        change: [−4.1151% −3.1676% −2.2502%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
  2 (2.00%) high mild
  2 (2.00%) high severe

bench_push_vec_small    time:   [40.818 ns 41.165 ns 41.570 ns]
                        change: [−0.9266% +0.5175% +1.8925%] (p = 0.47 > 0.05)
                        No change in performance detected.
Found 15 outliers among 100 measurements (15.00%)
  15 (15.00%) high mild

bench_insert_push_vec   time:   [415.17 ns 418.68 ns 422.85 ns]
                        change: [−2.4608% −1.2247% +0.0281%] (p = 0.06 > 0.05)
                        No change in performance detected.
Found 7 outliers among 100 measurements (7.00%)
  4 (4.00%) high mild
  3 (3.00%) high severe

bench_insert_push_vec_small
                        time:   [48.515 ns 48.592 ns 48.681 ns]
                        change: [−12.541% −10.992% −9.6939%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 5 outliers among 100 measurements (5.00%)
  3 (3.00%) high mild
  2 (2.00%) high severe

bench_pushpop_vec       time:   [262.56 ns 263.35 ns 264.20 ns]
                        change: [−9.0990% −8.2617% −7.4276%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 4 outliers among 100 measurements (4.00%)
  2 (2.00%) high mild
  2 (2.00%) high severe

[alejandro-vaz]

I have the intuition that benchmarks are no more reliable than coin flips

it makes sense that marking it as cold makes it faster, and std also does it

I side on making it cold

but why add new methods?? just keep the existing ones and add #[cold]

[fereidani]

I think there are instances that we want to inline the reserve when we are sure that we need that extra space, or maybe call the reserve in another inline(never) function which it will be not wise to outline the action once again.

[alejandro-vaz]

it makes sense, but we are adding too many functions for something so simple

is there any way we can do that without having to make a whole new function but just with an attribute??

also, the cold function does not do the check additional > self.capacity() - self.len() when called directly

[fereidani]

Give me some time to think about it. the additional > self.capacity() - self.len() is responsibility of the caller.
I also suggest we add needs_to_grow function like rust standard library instead of additional > self.capacity() - self.len() everytime.

[fereidani]

Another note if we move additional > self.capacity() - self.len() to the cold reserve that cold call happens every-time which defeats benefit of outlining.

[alejandro-vaz]

alright I made up my mind, I think this is my stance for now https://github.com/servo/rust-smallvec/wiki/on-v2%E2%80%90alpha-performance-contributions

I'll recover this PR if applicable by then in beta, though push concerns me