libcollections: move Vec::push slow path out

[emberian]

Makes Vec::push considerably smaller: 25 instructions, rather than 42, on
x86_64.

[rust-highfive]

r? @aturon

(rust_highfive has picked a reviewer for you, use r? to override)

[emberian]

r? @pcwalton

[Gankra]

Vec already has a resize fn

[Gankra]

See also the private grow_capacity method for some pre-existing refactoring of this functionality.

[emberian]

@gankro using grow_capacity is a non-starter since the capacity calculation (and associated panic) don't want to be in the fast path. I'm not sure how this resize function doesn't conflict with the other one, but I've moved it to an inner helper function now (waiting for build to test if that still works as desired)

[emberian]

Yeah, new version works the same.

[huonw]

Do you happen to have some before/after benchmarks?

[Gankra]

@cmr I was moreso suggesting there might be a better refactoring to reduce duplication, but this certainly doesn't make it worse.

[emberian]

@huonw No, but I could fabricate one that is predicated on icache residency.

[emberian]

(@erickt also mentioned wanting this in serde)

[alexcrichton]

I would personally also prefer to see at least one benchmark showing one way or another here. This is unfortunately a loss in readability in the code and is an idiom we only rarely apply today, so I would prefer to have some numbers backing it up showing why the loss in readability is worth it.

[emberian]

Smaller function size is the main goal (this makes push 40% smaller), which can help inlining in complex functions and reduce icache pressure. Here's one silly benchmark:

externcrate test;#[bench]fnx(b:&mut test::Bencher){letmut v = Vec::with_capacity(100);
b.iter(|| { v.extend(0..100); v.truncate(0);});}

Before: test x ... bench: 144 ns/iter (+/- 7)
After: test x ... bench: 81 ns/iter (+/- 4)

[emberian]

For reference, against the C++ equivalent:

#include<vector>
#include<cstdint>extern"C" {
voidcpp_version() {
static std::vector<int32_t> v(100);
for (int i = 0; i < 100; i++) {
v.push_back(i);
}
v.resize(0);
}
}

externcrate test;#[bench]fnx(b:&mut test::Bencher){letmut v = Vec::with_capacity(100);
b.iter(|| { v.extend(0..100); v.truncate(0);});}#[link(name = "foo")]#[link(name = "stdc++")]extern"C"{fncpp_version();}#[bench]fncpp(b:&mut test::Bencher){
b.iter(|| unsafe{cpp_version()});}

$ clang++ -O2 -fPIC -std=c++11 -c foo.cpp
$ ar rc libfoo.a foo.o
$ LD_LIBRARY_PATH=x86_64-unknown-linux-gnu/stage1/lib ./x86_64-unknown-linux-gnu/stage1/bin/rustc --test -O -L native=. foo.rs
$ ./foo --bench
running 2 tests
test cpp ... bench: 130 ns/iter (+/- 9)
test x ... bench: 83 ns/iter (+/- 5)

Maybe framing it as "Make Vec::push faster than C++ vector::push_back" makes it go down better ;)

[emberian]

@bors: r=pcwalton rollup

[bors]

📌 Commit 0e838f7 has been approved by pcwalton

[bors]

⌛ Testing commit 0e838f7 with merge 928e2e2...

[bors]

☀️ Test successful - auto-linux-32-nopt-t, auto-linux-32-opt, auto-linux-64-nopt-t, auto-linux-64-opt, auto-linux-64-x-android-t, auto-mac-32-opt, auto-mac-64-nopt-t, auto-mac-64-opt, auto-win-32-nopt-t, auto-win-32-opt, auto-win-64-nopt-t, auto-win-64-opt