4 releases
0.3.0 | Jul 21, 2022 |
---|---|
0.2.2 | Jul 2, 2022 |
0.2.1 | Jul 1, 2022 |
0.2.0 | Jul 1, 2022 |
0.1.0 |
|
#505 in Embedded development
293 downloads per month
Used in 2 crates
(via usbd-human-interface-devi…)
25KB
295 lines
A Block of Optionals!
The option-block
crate provides a simple primitive for fixed-size blocks of optional types. Formally speaking, it's a direct-address table with a fixed-size array as the storage medium.
Importantly, this is not to be confused with the popular slab
crate, which internally uses the dynamically-sized, heap-allocated Vec
. Although both crates provide indexed accesses and map-like features, option-block
operates at a lower level.
Specifically, option-block
does not keep track of the next empty slot in the allocation upon insertion (unlike slab
). Instead, option-block
is simply a wrapper around an array and a bit mask. The array contains the (maybe uninitialized) data while the bit mask keeps track of the valid (i.e. initialized) entries in the allocation. Again, it's basically a direct-address table.
This crate is compatible with
no_std
environments! Neitherstd
noralloc
is necessary.
Example
let mut block = option_block::Block8::<u8>::default();
assert!(block.is_empty());
assert!(block.insert(0, 10).is_none());
assert!(block.insert(1, 20).is_none());
assert_eq!(block.insert(0, 100), Some(10));
assert_eq!(block.insert(1, 200), Some(20));
assert_eq!(block.get(0), Some(&100));
assert_eq!(block.get(1), Some(&200));
assert_eq!(block.remove(0), Some(100));
assert_eq!(block.remove(1), Some(200));
assert!(block.is_empty());
assert_eq!(block.get(0), None);
assert_eq!(block.get(1), None);
assert_eq!(block.remove(0), None);
assert_eq!(block.remove(1), None);
Motivation
The Nullable Pointer Optimization
Sometimes, a direct-address table with a fixed-size allocation on the stack is sufficient for simple look-ups. That is, a heap-allocated HashMap
and Vec
may be overkill. Intuitively, one may be inclined to implement such a table using an array of Option<T>
(for some type T
). This is not ideal, however, because for most types, the size of an Option<T>
(in bytes) is unnecessarily large.
Certain types in Rust take advantage of the nullable pointer optimization. For some enum
types (like Option
), the compiler can do clever tricks to minimize its memory footprint. For instance, consider an Option<&T>
. Assuming a 64-bit target without the nullable pointer optimization enabled, the compiler may naively allocate 16 bytes for a single Option<&T>
: 8 bytes for the reference (i.e. the actual pointer) plus 8 bytes for the enum
discriminant. This is indeed rather wasteful.
To resolve these issues, recall that all references in Rust are never null. The compiler can take advantage of this fact by assigning the None::<&T>
variant to be the actual null pointer instead. Hence, we say that &T
is None
if the reference is null; otherwise, it is the Some
variant (which has a valid reference). The enum
discriminant is thus no longer necessary. An Option<&T>
is now just 8 bytes!
The Rustonomicon discusses more examples that enable the optimization. The point is: some types have properties and assumptions that allow the compiler to forego some size overhead. But what if this size optimization cannot happen?
Double the Memory Footprint
Consider an Option<u64>
. The core::mem::size_of
function tells us that a single Option<u64>
takes up 16 bytes of memory! The first 8 bytes belong to the u64
itself while the other 8 bytes belong to the enum
discriminant. Again, this is rather wasteful.
To resolve the enum
discriminant overhead, the standard library provides the core::num::NonZeroU64
type. The NonZeroU64
is a zero-cost wrapper for u64
that is assumed to be non-zero (as its name suggests).
This assumption makes NonZeroU64
eligible for the nullable pointer optimization. That is, an Option<NonZeroU64>
is None
if it contains 0
; otherwise, it is the Some
variant (which has a valid non-zero value). We may thus remove the overhead since the value already implicitly encodes the discriminant. An Option<NonZeroU64>
is now just 8 bytes!
use core::{mem::size_of, num::NonZeroU64};
assert_eq!(size_of::<Option<u64>>(), 16);
assert_eq!(size_of::<Option<NonZeroU64>>(), 8);
For this reason, a direct-address table which internally uses an array of Option<T>
values will inevitably consume more memory than necessary. Unless the inner type is conveniently eligible for the nullable pointer optimization, the enum
discriminant overhead will (at most) double the memory footprint.
A New Crate is Born!
However, not all hope is lost. Observe that the discriminant for the Option
type may actually be stored as a single bit. Therefore, it is possible to store multiple discriminants (for an array of optional values) in a single bit mask. This is exactly the abstraction that the option-block
crate provides.
This crate provides five primitives: Block8
, Block16
, Block32
, Block64
, and Block128
. As its name suggests, a Block8
is a block of at most 8 optional values, where the internal bit mask is a u8
(one for each cell). The rest of the primitives are basically the 16-, 32-, 64-, and 128-element analogs of the Block8
.
use core::mem::size_of;
use option_block::Block16;
assert_eq!(size_of::<[Option<u16>; 16]>(), 64);
assert_eq!(size_of::<Block16<u16>>(), 34);
Implementation Details
Further internal details are explained in narrative format in a supplementary article titled "Dipping Toes into Unsafe Code".
Stack Limitations
Since option-block
allocates on the stack, one must handle the Block64
and Block128
types with care. In the extreme case of the Block128
type, it allocates 128 instances of the inner data type plus 16 more bytes for the bit mask. Stack memory usage can easily skyrocket if too many are created. Thus, it is advised to use the larger block variants sparingly.