ColdString: A 1-word (8-byte) SSO string that saves up to 23 bytes over String
I’ve been working on a specialized string type called ColdString. The goal was to create the most memory-efficient string representation possible.
- Size: Exactly 1
usize(8 bytes on 64-bit). - Alignment: 1 byte (Uses
repr(transparent)around a[u8; 8]). - Inline Capacity: Up to 7 bytes (Small String Optimization).
- Heap Overhead: Only 1–9 bytes (VarInt length header) instead of the standard 16-byte
(pointer, length)pair.
Usage
use cold_string::ColdString;
let s = ColdString::new("qwerty");
assert_eq!(s.as_str(), "qwerty");
assert_eq!(std::mem::size_of::<ColdString>(), 8);
assert_eq!(std::mem::align_of::<ColdString>(), 1);
assert_eq!(std::mem::size_of::<(ColdString, u8)>(), 9);
assert_eq!(std::mem::align_of::<(ColdString, u8)>(), 1);
Memory Comparisons
(Average RSS size per string, in bytes, of 10 million ASCII strings).
| Crate | 0–4 chars | 0–8 chars | 0–16 chars | 0–32 chars | 0–64 chars |
|---|---|---|---|---|---|
std |
36.9 B | 38.4 B | 46.8 B | 55.3 B | 71.4 B |
smol_str |
24.0 B | 24.0 B | 24.0 B | 41.1 B | 72.2 B |
compact_str |
24.0 B | 24.0 B | 24.0 B | 35.4 B | 61.0 B |
compact_string |
24.1 B | 25.8 B | 32.6 B | 40.5 B | 56.5 B |
cold-string |
8.0 B | 11.2 B | 24.9 B | 36.5 B | 53.5 B |
How it works
ColdString uses a Tagged Pointer approach. Because we enforce an alignment of 2 for heap allocations, the least-significant bit (LSB) of any heap address is guaranteed to be 0.
- Inline Mode: If the LSB of the first byte is
1, the remaining bits in that byte represent the length (len<<1∣1), and the rest of the 8-byte array holds the UTF-8 data. - Heap Mode: If the LSB is
0, the 8 bytes are treated as ausizepointer. We useexpose_provenanceandwith_exposed_provenance(Stable as of 1.84+) to safely round-trip the pointer through the array. - Length Storage: To keep the struct at 8 bytes, we don't store the length in the struct. Instead, we use a VarInt (LEB128) encoded length header at the start of the heap allocation, immediately followed by the string data.
As always, any feedback welcome!