Do not grow the container if an insertion is on a tombstone.

[alkis]

No description provided.

[Amanieu]

Can you provide benchmark results for this change?

[alkis]

Changes are in the noise. Unsurprisingly since the benchmarks don't test insert/erase workloads. The insert/erase inserts a bunch and removes a bunch so the difference in the code doesn't matter. It does verify it does not make them slower.

Before:

test insert_erase_i32 ... bench:      24,370 ns/iter (+/- 4,587)
test insert_erase_i64 ... bench:      24,403 ns/iter (+/- 1,900)
test insert_i32       ... bench:      19,143 ns/iter (+/- 2,207)
test insert_i64       ... bench:      20,110 ns/iter (+/- 1,531)
test iter_i32         ... bench:       1,643 ns/iter (+/- 304)
test iter_i64         ... bench:       1,640 ns/iter (+/- 54)
test lookup_fail_i32  ... bench:       1,759 ns/iter (+/- 562)
test lookup_fail_i64  ... bench:       1,912 ns/iter (+/- 57)
test lookup_i32       ... bench:       2,447 ns/iter (+/- 174)
test lookup_i64       ... bench:       2,358 ns/iter (+/- 74)

After

test insert_erase_i32 ... bench:      24,310 ns/iter (+/- 1,222)
test insert_erase_i64 ... bench:      25,604 ns/iter (+/- 15,702)
test insert_i32       ... bench:      19,896 ns/iter (+/- 13,620)
test insert_i64       ... bench:      19,917 ns/iter (+/- 816)
test iter_i32         ... bench:       1,655 ns/iter (+/- 131)
test iter_i64         ... bench:       1,649 ns/iter (+/- 153)
test lookup_fail_i32  ... bench:       1,749 ns/iter (+/- 104)
test lookup_fail_i64  ... bench:       1,995 ns/iter (+/- 796)
test lookup_i32       ... bench:       2,439 ns/iter (+/- 126)
test lookup_i64       ... bench:       2,404 ns/iter (+/- 650)

Any idea why insert_no_grow is now detected as dead even if it is used by rustc_entry.rs? Ideas to fix this?

[alkis]

Perhaps we can run some kind of real-life benchmark? Rustc maybe? Any instructions on how to do that?

[Amanieu]

You can try using https://github.com/rust-lang-nursery/rustc-perf to see the impact of the change on a wide range of crates (including rustc). You will need to change src/libstd/Cargo.toml to point to your modified hashbrown instead of the one from crates.io, and then build a custom rustc toolchain using that libstd.

[Amanieu]

CI is failing because insert_no_grow is an unused function. Can you mark it with #[cfg(feature = "rustc-dep-of-std")] since it is only used by rustc_entry?

[Amanieu]

Closing in favor of #106.