| smallvec.rs |
Benchmarks that compare TinyVec to SmallVec
All the following commentary is based on the latest nightly at the time:
rustc 1.55.0 (c8dfcfe04 2021-09-06).
Some of these benchmarks are just a few instructions, so we put our own for
loop inside the criterion::Bencher::iter call. This seems to improve the
stability of measurements, and it has the wonderful side effect of making
the emitted assembly easier to follow. Some of these benchmarks are totally
inlined so that there are no calls at all in the hot path, so finding
this for loop is an easy way to find your way around the emitted assembly.
The clear method is cheaper to call for arrays of elements without a Drop
impl, so wherever possible we reuse a single object in the benchmark loop,
with a clear + black_box on each iteration in an attempt to not make that
visible to the optimizer.
We always call black_box(&v), instead of v = black_box(v) because the latter
does a move of the inline array, which is linear in the size of the array
and thus varies based on the array type being benchmarked, and this move can
be more expensive than the function we're trying to benchmark.
We also black_box the input to each method call. This has a significant
effect on the assembly emitted, for example if we do not black_box the range
we iterate over in the ::push benchmarks, the loop is unrolled. It's not
entirely clear if it's better to black_box the iterator that yields the
items being pushed, or to black_box at a deeper level: v.push(black_box(i))
for example. Anecdotally, it seems like the latter approach produces
unreasonably bad assembly. |
12866 |