flexstr

A flexible, simple to use, immutable, clone-efficient String replacement for Rust. It unifies literals, inlined, and heap allocated strings into a single type.

Overview

Rust is great, but it's String type is optimized as a mutable string buffer, not for typical string use cases. Most string use cases don't modify their contents, often need to copy strings around as if they were cheap like integers, typically concatenate instead of modify, and often end up being cloned with identical contents. Additionally, String isn't able to wrap a string literal without additional allocation and copying.

Rust needs a new string type to unify usage of both literals and allocated strings in these typical use cases. This crate creates a new string type that is optimized for those use cases, while retaining the usage simplicity of String.

This type is not inherently "better" than String, but different. It works best in 'typical' string use cases (immutability, concatenation, cheap multi ownership) whereas String works better in "string buffer" use cases (mutability, string building, single ownership).

Installation

NOTE: The serde feature is optional and only included when specified.

[dependencies]
flexstr = { version = "0.5", features = ["serde"] }

Examples

use flexstr::{flex_fmt, FlexStr, IntoFlexStr, ToCase, ToFlexStr};

fn main() {
  // Use an `into` function to wrap a literal, no allocation or copying
  let static_str = "This will not allocate or copy".into_flex_str();
  assert!(static_str.is_static());

  // Strings up to 22 bytes (on 64-bit) will be inlined automatically 
  // (demo only, use `into` for literals as above)
  let inline_str = "inlined".to_flex_str();
  assert!(inline_str.is_inlined());

  // When a string is too long to be wrapped/inlined, it will heap allocate
  // (demo only, use `into` for literals as above)
  let rc_str = "This is too long to be inlined".to_flex_str();
  assert!(rc_str.is_heap());

  // You can efficiently create a new `FlexStr` (without creating a `String`)
  // This is equivalent to the stdlib `format!` macro
  let inline_str2 = flex_fmt!("in{}", "lined");
  assert!(inline_str2.is_inlined());
  assert_eq!(inline_str, inline_str2);

  // We can upper/lowercase strings without converting to a `String`
  // This doesn't heap allocate
  let inline_str3: FlexStr = "INLINED".to_ascii_lower();
  assert!(inline_str3.is_inlined());
  assert_eq!(inline_str, inline_str3);

  // Concatenation doesn't even copy if we can fit it in the inline string
  let inline_str4 = inline_str3 + "!!!";
  assert!(inline_str4.is_inlined());
  assert_eq!(inline_str4, "inlined!!!");
  
  // Clone is almost free, and never allocates
  // (at most it is a ref count increment for heap allocated strings)
  let static_str2 = static_str.clone();
  assert!(static_str2.is_static());

  // Regardless of storage type, these all operate seamlessly together 
  // and choose storage as required
  let heap_str2 = static_str2 + &inline_str;
  assert!(heap_str2.is_heap());
  assert_eq!(heap_str2, "This will not allocate or copyinlined");
}

How Does It Work?

Internally, FlexStr uses an enum with these variants:

• Static - A simple wrapper around a static string literal (&'static str)
• Inlined - An inlined string (no heap allocation for small strings)
• Heap - A heap allocated (reference counted) string

The type automatically chooses the best storage and allows you to use them interchangeably as a single string type.

Features

• Optimized for immutability and cheap cloning
• Allows for multiple ownership of the same string memory contents
• Serves as a universal string type (unifying literals and allocated strings)
• Doesn't allocate for literals and short strings (64-bit: up to 22 bytes)
• The same size as a String (64-bit: 24 bytes)
• Optional serde serialization support (feature = "serde")
• Compatible with embedded systems (doesn't use std)
• Efficient conditional ownership (borrows can take ownership without allocation/copying)
• Both single threaded compatible (FlexStr) and multi-thread safe (AFlexStr) options
• It is simple to use!

Types

• FlexStr - regular usage
  • Heap storage based on Rc
• AFlexStr- provides Send / Sync for multi-threaded use
  • Heap storage based on Arc

Usage

Hello World

use flexstr::IntoFlexStr;

fn main() {
  // From literal - no copying or allocation
  let world = "world!".into_flex_str();

  println!("Hello {world}");
}

Conversions

use flexstr::{IntoAFlexStr, IntoFlexStr, ToFlexStr};

fn main() {
  // From literal - no copying or allocation
  // NOTE: `to_flex_str` will copy, so use `into_flex_str` for literals
  let literal = "literal".into_flex_str();

  // From borrowed string - Copied into inline string
  let owned = "inlined".to_string();
  let str_to_inlined = (&owned).to_flex_str();

  // From borrowed String - copied into `str` wrapped in `Rc`
  let owned = "A bit too long to be inlined!!!".to_string();
  let str_to_wrapped = (&owned).to_flex_str();

  // From String - copied into inline string (`String` storage released)
  let inlined = "inlined".to_string().into_flex_str();

  // From String - `str` wrapped in `Rc` (`String` storage released)
  let counted = "A bit too long to be inlined!!!".to_string().into_flex_str();

  // *** If you want a Send/Sync type you need `AFlexStr` instead ***

  // From FlexStr wrapped literal - no copying or allocation
  let literal = literal.into_a_flex_str();

  // From FlexStr inlined string - no allocation
  let inlined = inlined.into_a_flex_str();

  // From FlexStr `Rc` wrapped `str` - copies into `str` wrapped in `Arc`
  let counted = counted.into_a_flex_str();
}

Passing FlexStr to Conditional Ownership Functions

This has always been a confusing situation in Rust, but it is easy with FlexStr since multi ownership is cheap.

use flexstr::{IntoFlexStr, FlexStr};

struct MyStruct {
  s: FlexStr
}

impl MyStruct {
  fn to_own_or_not_to_own(s: &FlexStr) -> Self {
    let s = if s == "own_me" {
      // Since a wrapped literal, no copy or allocation
      s.clone()
    } else {
      // Wrapped literal - no copy or allocation
      "own_me".into()
    };

    Self { s }
  }
}

fn main() {
  // Wrapped literals - no copy or allocation
  let s = "borrow me".into_flex_str();
  let s2 = "own me".into_flex_str();

  let struct1 = MyStruct::to_own_or_not_to_own(&s);
  let struct2 = MyStruct::to_own_or_not_to_own(&s2);

  assert_eq!(s2, struct1.s);
  assert_eq!(s2, struct2.s);
}

Performance Characteristics

• Clones are cheap and never allocate
  • At minimum, they are just a copy of the enum and at max an additional reference count increment
• Literals are just wrapped when used with into() and never copied
• Calling into() on a String will result in an inline string (if short) otherwise copied into a str wrapped in Rc/Arc (which will allocate, copy, and then release original String storage)
• into_flex_str() and into_a_flex_str() are equivalent to calling into() on both literals and String (they are present primarily for let bindings so there is no need to declare a type)
• to_flex_str() and to_a_flex_str() are meant for taking ownership of borrowed strings and always copy into either an inline string (for short strings) or an Rc/Arc wrapped str (which will allocate)
• to_string always copies into a new String
• Conversions back and forth between AFlexStr and FlexStr using into() are cheap when using wrapped literals or inlined strings
  • Inlined strings and wrapped literals just create a new enum wrapper
  • Reference counted wrapped strings will always require an allocation and copy for the new Rc or Arc

Benchmarks

Create

Heap creates are fairly expensive still compared to String (apparently due to the overhead of creating the enum?), Rc<str> andArc<str>, but inline/static creation is very fast as expected.

FlexStr

create_static_normal    time:   [3.7062 ns 3.7213 ns 3.7422 ns]
create_inline_small     time:   [3.8932 ns 3.9004 ns 3.9084 ns]
create_heap_normal      time:   [13.533 ns 13.557 ns 13.587 ns]
create_heap_large       time:   [18.605 ns 18.635 ns 18.664 ns]
create_heap_arc_normal  time:   [18.535 ns 18.551 ns 18.568 ns]
create_heap_arc_large   time:   [26.794 ns 26.861 ns 26.937 ns]

Comparables

create_string_small     time:   [7.4377 ns 7.4572 ns 7.4794 ns]
create_string_normal    time:   [8.0550 ns 8.0605 ns 8.0667 ns]
create_string_large     time:   [12.940 ns 12.955 ns 12.973 ns]
create_rc_small         time:   [8.0525 ns 8.0577 ns 8.0639 ns]
create_rc_normal        time:   [8.2438 ns 8.2512 ns 8.2604 ns]
create_rc_large         time:   [13.139 ns 13.153 ns 13.168 ns]
create_arc_small        time:   [8.7128 ns 8.7231 ns 8.7341 ns]
create_arc_normal       time:   [8.7454 ns 8.7851 ns 8.8446 ns]
create_arc_large        time:   [13.827 ns 13.855 ns 13.886 ns]

Clone

Clones are MUCH cheaper than String (except when using Arc). Interested to find out why the enum wrapper and single branch op causes such a large differential between the wrapped Rc<str>/Arc<str> and the raw version.

FlexStr

clone_static_normal     time:   [3.9540 ns 3.9572 ns 3.9610 ns]
clone_inline_small      time:   [4.4717 ns 4.4763 ns 4.4819 ns]
clone_heap_normal       time:   [4.4738 ns 4.4839 ns 4.4965 ns]
clone_heap_arc_normal   time:   [10.596 ns 10.607 ns 10.618 ns]

Comparables

clone_string_small      time:   [11.774 ns 11.789 ns 11.807 ns]
clone_string_normal     time:   [12.289 ns 12.422 ns 12.540 ns]
clone_string_large      time:   [14.931 ns 15.013 ns 15.116 ns]
clone_rc_normal         time:   [652.97 ps 653.58 ps 654.30 ps]
clone_arc_normal        time:   [3.2948 ns 3.2986 ns 3.3021 ns]

Conversions

Thanks (mostly) to itoa and ryu our conversions are much faster than String.

FlexStr

convert_bool            time:   [3.7080 ns 3.7094 ns 3.7109 ns]
convert_char            time:   [3.8104 ns 3.8159 ns 3.8222 ns]
convert_i8              time:   [3.2817 ns 3.2827 ns 3.2838 ns]
convert_i16             time:   [3.5285 ns 3.5379 ns 3.5511 ns]
convert_i32             time:   [10.568 ns 10.575 ns 10.582 ns]
convert_i64             time:   [7.6351 ns 7.6390 ns 7.6430 ns]
convert_i128            time:   [38.756 ns 38.787 ns 38.821 ns]
convert_f32             time:   [24.669 ns 24.692 ns 24.721 ns]
convert_f64             time:   [33.105 ns 33.145 ns 33.184 ns]

Comparables

convert_string_bool     time:   [18.466 ns 18.505 ns 18.538 ns]
convert_string_char     time:   [7.2933 ns 7.2966 ns 7.3003 ns]
convert_string_i8       time:   [7.3838 ns 7.4546 ns 7.5457 ns]
convert_string_i16      time:   [23.087 ns 23.477 ns 24.025 ns]
convert_string_i32      time:   [38.577 ns 38.624 ns 38.683 ns]
convert_string_i64      time:   [43.348 ns 43.396 ns 43.446 ns]
convert_string_i128     time:   [71.120 ns 71.174 ns 71.225 ns]
convert_string_f32      time:   [100.24 ns 100.50 ns 100.78 ns]
convert_string_f64      time:   [179.86 ns 180.00 ns 180.14 ns]

Negatives

There is no free lunch:

• Due to usage of Rc (or Arc), when on-boarding String it will need to reallocate and copy
• Due to the enum wrapper, every string operation has the overhead of an extra branching operation
• Since FlexStr is not Send or Sync, there is a need to consider single-threaded (FlexStr) and multi-threaded (AFlexStr) use cases and convert accordingly

Status

This is currently beta quality and still needs testing. The API may very possibly change but semantic versioning will be followed.

License

This project is licensed optionally under either:

• Apache License, Version 2.0, (LICENSE-APACHE or https://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or https://opensource.org/licenses/MIT)