performance regression introduced in nightly-2018-11-04

[tomcraven]

debugging a performance regression in my code, found it was introduced between nightly-2018-11-03 and nightly-2018-11-04 - 8b09631...04fdb44

the best minimum example i could come up with:

    #[bench]
    fn bench_something(b: &mut Bencher) {
        b.iter(|| {
            (0..1000)
                .for_each(|a| {
                    let mut vec = Vec::with_capacity(10);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                    vec.push(0);
                })
        });
    }

test between the two versions:

$ rustup override set nightly-2018-11-03 && cargo bench something && rustup override set nightly-2018-11-04 && cargo bench something
info: using existing install for 'nightly-2018-11-03-x86_64-apple-darwin'
info: override toolchain for '<redacted>' set to 'nightly-2018-11-03-x86_64-apple-darwin'
  nightly-2018-11-03-x86_64-apple-darwin unchanged - rustc 1.32.0-nightly (8b096314a 2018-11-02)
    Finished release [optimized] target(s) in 1.27s
     Running target/release/deps/<redacted>-08b08ca17f1f5108
running 1 test
test benchmarks::bench_something                                        ... bench:      27,958 ns/iter (+/- 3,807)
test result: ok. 0 passed; 0 failed; 0 ignored; 1 measured; 80 filtered out
info: using existing install for 'nightly-2018-11-04-x86_64-apple-darwin'
info: override toolchain for '<redacted>' set to 'nightly-2018-11-04-x86_64-apple-darwin'
  nightly-2018-11-04-x86_64-apple-darwin unchanged - rustc 1.32.0-nightly (04fdb44f5 2018-11-03)
    Finished release [optimized] target(s) in 1.25s
     Running target/release/deps/<redacted>-307045d1760e56ff
running 1 test
test benchmarks::bench_something                                        ... bench:      75,034 ns/iter (+/- 16,036)
test result: ok. 0 passed; 0 failed; 0 ignored; 1 measured; 80 filtered out

and same results for current nightly:

$ rustup override set nightly
info: using existing install for 'nightly-x86_64-apple-darwin'
info: override toolchain for '<redacted>' set to 'nightly-x86_64-apple-darwin'
  nightly-x86_64-apple-darwin unchanged - rustc 1.32.0-nightly (14997d56a 2018-12-05)

$ cargo bench something
    Finished release [optimized] target(s) in 0.25s
     Running target/release/deps/<redacted>-bda65b0f58ecd966
running 1 test
test benchmarks::bench_something                                        ... bench:      78,035 ns/iter (+/- 14,081)
test result: ok. 0 passed; 0 failed; 0 ignored; 1 measured; 80 filtered out

[nikic]

In this nightly the default allocator was switched from jemalloc to the system allocator. It is expected that this may lead to performance regressions in some cases. See also #56283 for a similar report and in particular the comment #56283 (comment) showing how you can explicitly enable jemalloc again.