Improve Vec::truncate performance in no-opt builds

[veddan]

This significantly improves the performance of Vec::truncate for types
without destructors when building with --opt-level=0 or --opt-level=1.
Fixes #17633.

[veddan]

Benchmark code

extern crate test;
use test::Bencher;

#[deriving(Clone)]
struct Foo { x: u8  }

impl Drop for Foo {
    fn drop(&mut self) { }
}

#[bench]
fn bench_no_drop(b: &mut Bencher) {
    let v = Vec::from_elem(10000, 0u8);
    b.iter(|| { let mut x = v.clone(); x.truncate(0); });
}

#[bench]
fn bench_drop(b: &mut Bencher) {
    let v = Vec::from_elem(10000, Foo { x: 0 });
    b.iter(|| { let mut x = v.clone(); x.truncate(0); });
}

Results

Before, -O0

test bench_drop    ... bench:    799459 ns/iter (+/- 33636)
test bench_no_drop ... bench:    667411 ns/iter (+/- 27678)

After, -O0

test bench_drop    ... bench:    791541 ns/iter (+/- 28567)
test bench_no_drop ... bench:    476229 ns/iter (+/- 19815)

Before, -O1

test bench_drop    ... bench:    394395 ns/iter (+/- 17525)
test bench_no_drop ... bench:    388002 ns/iter (+/- 21695)

After, -O1

test bench_drop    ... bench:    406982 ns/iter (+/- 30450)
test bench_no_drop ... bench:    288509 ns/iter (+/- 13999)

[thestinger]

I don't like the idea of adding complexity to improve runtime performance without optimizations. I don't think a small constant factor slowdown for --opt-level=1 is important. It could be fixed at --opt-level=1 with careful application of #[inline(always)]. In the worst case, the lack of this specialization will double the constant factor for each paired initialization / destruction.

[pcwalton]

Agreed with @thestinger, closing. Many thanks for the patch, however :)