Update fannkuchredux benchmark

[brson]

From the discussion on reddit:
http://www.reddit.com/r/rust/comments/2fenlg/benchmark_improvement_fannkuchredux/

This adds two variants: the primary, that uses an unsafe block, and a secondary
that is completely safe.

The one with the unsafe block matches clang's performance and beats gcc's.

[huonw]

Hm, is adjusting slice::reverse to use the unsafe code for performance acceptable? (This should leave the -safe one as fast as the other.)

[brson]

@huonw I'm testing that change to slice::reverse now.

[brson]

@huonw Optimized reverse so that the safe fannkuchredux matches C++ perf!

[alexcrichton]

It'd be awesome if we would show off idiomatic rust code beating/matching C++ rather than straight ports of the C++ benchmarks, and I'm curious, with the reverse improvements, is the current rust version that much slower?

[huonw]

http://www.reddit.com/r/rust/comments/2fenlg/benchmark_improvement_fannkuchredux/ck8swcx?context=1

Iirc the 'original' listed there uses @dotdash's reverse (the fast one), but I can't remember exactly which cfg I used. If so, it is still slower. (I imagine that removing the heap allocation from the original Rust one would give a speed up, but that would basically be the same as the C++ anyway.)

[dotdash]

result -> reverse in the second commit message. That had me quite confused for a moment.

[huonw]

Ah, no, I think that 'original' was the truly original code (makes sense...), but just substituting in the faster reverse takes 1.55s (vs. 1.16 for the fastest Rust).

[Gankra]

Could this just be mem::swap instead? That way there's no raw ptrs.

[arthurprs]

Shouldn't we prefer the neater alternative, also by @brson

Edit: I don't think we need to leave a note saying it improves a benchmark. It's a valid optimization as observed in other functions as well.

    let p = self.as_mut_ptr();
    unsafe {
        for off in range(0, k / 2) {
            std::ptr::swap(p.offset(off), p.offset(k - 1 - off));
        }
    }

[brson]

I'm not clear on what people are suggesting happens here.

[brson]

Updated the description of the second commit.

[brson]

I removed the comment about fankuch. mem::swap can't be used here because of borrowing conflicts.

[huonw]

I would r+ but I wrote the code, basically, so that feels bad. :)