If `shallow_clone` is called with `&mut self`, and `Bytes` contains
`Vec`, then expensive CAS can be avoided, because no other thread
have references to this `Bytes` object.
Bench `split_off_and_drop` difference:
Before the diff:
```
test split_off_and_drop ... bench: 91,858 ns/iter (+/- 17,401)
```
With the diff:
```
test split_off_and_drop ... bench: 81,162 ns/iter (+/- 17,603)
```
* Compact Bytes original capacity representation
In order to avoid unnecessary allocations, a `Bytes` structure remembers
the capacity with which it was first created. When a reserve operation
is issued, this original capacity value is used to as a baseline for
reallocating new storage.
Previously, this original capacity value was stored in its raw form. In
other words, the original capacity `usize` was stored as is. In order to
reclaim some `Bytes` internal storage space for additional features,
this original capacity value is compressed from requiring 16 bits to 3.
To do this, instead of storing the exact original capacity. The original
capacity is rounded down to the nearest power of two. If the original
capacity is less than 1024, then it is rounded down to zero. This
roughly means that the original capacity is now stored as a table:
0 => 0
1 => 1k
2 => 2k
3 => 4k
4 => 8k
5 => 16k
6 => 32k
7 => 64k
For the purposes that the original capacity feature was introduced, this
is sufficient granularity.
* Provide `advance` on Bytes and BytesMut
This is the `advance` function that would be part of a `Buf`
implementation. However, `Bytes` and `BytesMut` cannot impl `Buf` until
the next breaking release.
The implementation uses the additional storage made available by the
previous commit to store the number of bytes that the view was advanced.
The `ptr` pointer will point to the start of the window, avoiding any
pointer arithmetic when dereferencing the `Bytes` handle.
* Inner: make uninitialized construction explicit
* Remove Inner2
* Remove unnecessary transmutes
* Use AtomicPtr::get_mut where possible
* Some minor tweaks
Slice operation should return inline when possible
It is cheaper than atomic increment/decrement.
Before this patch:
```
test slice_avg_le_inline_from_arc ... bench: 28,582 ns/iter (+/- 3,880)
test slice_empty ... bench: 8,797 ns/iter (+/- 1,325)
test slice_large_le_inline_from_arc ... bench: 27,684 ns/iter (+/- 5,920)
test slice_short_from_arc ... bench: 27,439 ns/iter (+/- 5,783)
```
After this patch:
```
test slice_avg_le_inline_from_arc ... bench: 18,872 ns/iter (+/- 2,937)
test slice_empty ... bench: 9,136 ns/iter (+/- 1,908)
test slice_large_le_inline_from_arc ... bench: 18,052 ns/iter (+/- 2,981)
test slice_short_from_arc ... bench: 18,200 ns/iter (+/- 2,534)
```
Saves the cognitive load of having to wrap them in slices to compare
them when that seems like what one would expect.
Signed-off-by: Clint Byrum <clint@fewbar.com>
Return empty `Bytes` object
Bench for `slice_empty` difference is
```
55 ns/iter (+/- 1) # before this patch
17 ns/iter (+/- 5) # with this patch
```
Bench for `slice_not_empty` is
```
25,058 ns/iter (+/- 1,099) # before this patch
25,072 ns/iter (+/- 1,593) # with this patch
```
Round up to power of 2 is not necessary, because `reserve` already
doubles previous capacity in
```
new_cap = cmp::max(
cmp::max(v.capacity() << 1, new_cap),
original_capacity);
```
which makes `reserve` calls constant in average. Avoiding rounding
up prevents `reserve` from wasting space when caller knows exactly
what space they need.
Patch adds three tests which would fail before this test. The most
important is this:
```
#[test]
fn reserve_in_arc_unique_does_not_overallocate() {
let mut bytes = BytesMut::with_capacity(1000);
bytes.take();
// now bytes is Arc and refcount == 1
assert_eq!(1000, bytes.capacity());
bytes.reserve(2001);
assert_eq!(2001, bytes.capacity());
}
```
It asserts that when user requests more than double of current
capacity, exactly the requested amount of memory is allocated and
is not wasted to next power of two.
`extend_with_slice` is super-convenient operation on `Bytes`.
While `put_u8` would be expensive on `Bytes`, `extend_from_slice`
is OK, because it is batch, and it checks for kind only once.
Patch also adds `impl Extend for Bytes`.
cc #116
The shared debug_assert is to ensure that the internal Bytes
representation is such that offset views are supported. The only
representation that does not support offset views is vec.
Fixes#97
Before this commit `Bytes::split_{off,to}` always created a shallow copy if `self` is arc or vec.
However, in certain cases `split_off` or `split_to` is called with `len` or `0` parameter. E. g. if you are reading a frame from buffered stream, it is likely that buffer contains exactly the frame size bytes, so `split_to` will be called with `len` param.
Although, `split_off` and `split_to` functions are `O(1)`, shallow copy have downsides:
* shallow copy on vector does malloc and atomic cmpxchg
* after shallow copy, following operations (e. g. `drop`) on both `bytes` objects require atomics
* memory will be probably released to the system later
* `try_mut` will fail
* [into_vec](https://github.com/carllerche/bytes/issues/86) will copy
Standard `Debug` implementation for `[u8]` is comma separated list
of numbers. Since large amount of byte strings are in fact ASCII
strings or contain a lot of ASCII strings (e. g. HTTP), it is
convenient to print strings as ASCII when possible.
Limit the number of threads when using qemu to 1. Also, don't bother
running the stress test as this will trigger qemu bugs. Finally, also
make the stress test actually stress test.
I found this significantly improved a
[benchmark](https://gist.github.com/danburkert/34a7d6680d97bc86dca7f396eb8d0abf)
which calls `bytes_mut`, writes 1 byte, and advances the pointer with
`advance_mut` in a pretty tight loop. In particular, it seems to be the
inline annotation on `bytes_mut` which had the most effect. I also took
the opportunity to simplify the bounds checking in advance_mut.
before:
```
test encode_varint_small ... bench: 540 ns/iter (+/- 85) = 1481 MB/s
```
after:
```
test encode_varint_small ... bench: 422 ns/iter (+/- 24) = 1895 MB/s
```
As you can see, the variance is also significantly improved.
Interestingly, I tried to change the last statement in `bytes_mut` from
```
&mut slice::from_raw_parts_mut(ptr, cap)[len..]
```
to
```
slice::from_raw_parts_mut(ptr.offset(len as isize), cap - len)
```
but, this caused a very measurable perf regression (almost completely
negating the gains from marking bytes_mut inline).
Stepan Koltsov