Flux is a refinement type checker for Rust that lets you specify a range of correctness properties and have them be verified at compile time.

See the examples -- listed in the summary on the left -- to learn about Refinement types and Rust.

You can try it online here.

Installing Flux

Requirements

Be sure that the liquid-fixpoint and z3 executables are in your $PATH.

Installing

The only way to use flux is to build it from source.

First you need to clone the repository

git clone https://github.com/flux-rs/flux
cd flux

To build the source you need a nightly version of rustc. We pin the version using a toolchain file (more info here). If you are using rustup, no special action is needed as it should install the correct rustc version and components based on the information on that file.

Next, run the following to build and install flux binaries

cargo xtask install

This will install two binaries rustc-flux and cargo-flux in your cargo home. These two binaries should be used respectively to run flux on either a single file or on a project using cargo. The installation process will also copy some files to $HOME/.flux.

Running Flux

You can run flux on a single file or entire crate.

Running on a File: rustc-flux

You can use rustc-flux as you would use rustc. For example, the following command checks the file test.rs.

rustc-flux path/to/test.rs

The flux binary accepts the same flags as rustc. You could for example check a file as a library instead of a binary like so

rustc-flux --crate-type=lib path/to/test.rs

Running on a package: cargo-flux

Flux is integrated with cargo and can be invoked in a package as follows:

cargo flux

By default, Flux won't verify a package unless it's explicitly enabled in the manifest. To do so add the following to Cargo.toml:

[package.metadata.flux]
enabled = true

A tiny example

The following example declares a function inc that returns an integer greater than the input. We use the nightly feature register_tool to register the flux tool in order to add refinement annotations to functions.

#![allow(unused)]
fn main() {
#[flux::sig(fn(x: i32) -> i32{v: x < v})]
pub fn inc(x: i32) -> i32 {
    x - 1
}
}

You can save the above snippet in say test0.rs and then run

rustc-flux --crate-type=lib path/to/test0.rs

you should see in your output

error[FLUX]: postcondition might not hold
 --> test0.rs:3:5
  |
3 |     x - 1
  |     ^^^^^

as indeed x - 1 is not greater than x as required by the output refinement i32{v: x < v}.

If you fix the error by replacing x - 1 with x + 1, you should get no errors in the output (the output may be empty, but in this case no output is a good thing).

Read these chapters to learn more about what you specify and verify with flux.

A note about the flux-driver binary

The flux-driver binary is a rustc driver (similar to how clippy works) meaning it uses rustc as a library to "drive" compilation performing additional analysis along the way. Running the binary requires dynamically linking a correct version of librustc. Thus, to avoid the hassle you should never execute it directly. Instead, use rustc-flux or cargo-flux.

Editor Support

This section assumes you have installed flux, cargo-flux, and rustc-flux.

Rust-Analyzer in VSCode

Add this to the workspace settings i.e. .vscode/settings.json

{
  "rust-analyzer.check.overrideCommand": [
    "cargo",
    "flux",
    "--workspace",
    "--message-format=json-diagnostic-rendered-ansi"
  ]
}

Note: Make sure to edit the paths in the above snippet to point to the correct locations on your machine.

Configuration

Environment Variables

You can set various env variables to customize the behavior of flux.

  • FLUX_CONFIG tells flux where to find a config file for these settings.
    • By default, flux searches its directory for a flux.toml or .flux.toml.
  • FLUX_SYSROOT tells cargo-flux and rustc-flux where to find the flux-driver binary.
    • Defaults to the default installation location in ~/.flux.
  • FLUX_LOG_DIR=path/to/log/ sets the directory where constraints, timing and cache are saved. Defaults to ./log/.
  • FLUX_DUMP_CONSTRAINT=1 tell flux to dump constraints generated for each function.
  • FLUX_DUMP_CHECKER_TRACE=1 saves the checker's trace (useful for debugging!)
  • FLUX_DUMP_TIMINGS=1 saves the profile information
  • FLUX_DUMP_MIR=1 saves the low-level MIR for each analyzed function
  • FLUX_POINTER_WIDTH=N the size of (either 32 or 64), used to determine if an integer cast is lossy (default 64).
  • FLUX_CHECK_DEF=name only checks definitions containing name as a substring
  • FLUX_CACHE=1" switches on query caching and saves the cache in FLUX_CACHE_FILE
  • FLUX_CACHE_FILE=file.json customizes the cache file, default FLUX_LOG_DIR/cache.json
  • FLUX_CHECK_OVERFLOW=1 checks for over and underflow on arithmetic integer operations, default 0. When set to 0, it still checks for underflow on unsigned integer subtraction.

Config file

The config file is a .toml file that contains on each line the lowercase name of a flux command line flag without the FLUX_ prefix. Set environment variables take priority over the config file.

The config file should be in the project root.

For example, suppose your project root contains the following flux.toml.

log_dir = "./test"
dump_timings = true
dump_mir = true
cache = true

and you run in the project root

FLUX_DUMP_MIR=0 cargo-flux check

then flux will create the directory ./test/ and write ./test/timings, a file containing profiling information. It will not dump the MIR because that setting was overridden by setting the environment variable FLUX_DUMP_MIR=0.

Crate Config

Some flags can be configured on a per-crate basis using the custom inner attribute #![flux::cfg]. This annotation relies on the unstable custom inner attributes feature. To be able to use with a non-nightly compiler you have to put it under a cfg_attr. For example, to enable overflow checking:

#![allow(unused)]
#![cfg_attr(flux, flux::cfg(check_overflow = true))]
fn main() {
}

The only flag supported now is overflow checking.

Query Caching

FLUX_CACHE=1 persistently caches the safe fixpoint queries for each DefId in FLUX_LOG_DIR/FLUX_CACHE_FILE, and on subsequent runs, skips queries that are already in the cache, which considerably speeds up cargo-flux check on an entire crate.

Flux Specification Guide

This is a WIP guide to writing specifications in flux.

Refinement Types

  • Indexed Type: An indexed type B[r] is composed of a base Rust type B and a refinement index r. The meaning of the index depends on the type. Some examples are

    • i32[n]: denotes the (singleton) set of i32 values equal to n.
    • List<T>[n]: values of type List<T> of length n.

  • Refinement parameter: Function signatures can be parametric on refinement variables. Refinement parameters are declared using the @n syntax. For example, the following signature:

    fn(i32[@n]) -> i32[n + 1]

    binds n over the entire scope of the function to specify that it takes an i32 equal to n and returns an i32 equal to n + 1. This is analogous to languages like Haskell where a lower case letter can be used to quantify over a type, e.g., the type a -> a in Haskell is polymorphic on the type a which is bound for the scope of the entire function type.

  • Existential Type: An existential type B{v: r(v)} is composed of a base type B, a refinement variable v and a refinement predicate r on v. Intuitively, a Rust value x has type B{v: r(v)} if there exists a refinement value a such that r(a) holds and x has type B[x].

    • i32{v: v > 0}: set of positive i32 values.
    • List<T>{v: v > 0}: set of non-empty lists.

  • Constraint Type: A constraint type has the form {T | r} where T is any type (not just a base type). Intuitively, a value has type {T | r} if it has type T and also r holds. They can be used to constraint a refinement parameter. For example, the following signature constraint the refinement parameter n to be less than n.

    fn({i32[@n] | n < 10}) -> i32[n + 1]

    Constraint types serve a similar role as existentials as they can also be used to constraint some refinement value with a predicate, but an existential type can only be used to constraint refinement variable that it bound locally, in contrast constraint types can be used to constraint a "non-local" parameter. This can be seen in the example above where the parameter n cannot be bound locally because it has to be used in the return type.

Argument Syntax

The @n syntax used to declare refinements parameters can be hard to read sometimes. Flux also supports a syntax that let you bind refinement parameters using colons similar to the syntax used to declare arguments in a function. We call this argument syntax. This syntax desugars to one of the refinements forms discussed above. For example, the following signature

fn(x: i32, y: i32) -> i32[x + y]

desugars to

fn(i32[@x], i32[@y]) -> i32[x + y]

It is also possible to attach some constraint to the parameters when using argument syntax. For example, to specify that y must be greater than x using argument syntax we can write:

fn(x: i32, y: i32{x > y}) -> i32[x + y]

This will desugar to:

fn(i32[@x], {i32[@y] | x > y}) -> i32[x + y]

Extern specs

Sometimes you may want to refine a struct or function that outside your code. We refer to such a specification as an "extern spec," which is short for "external specification."

Flux right now has rudimentary support for extern specs: they are supported for functions, impls, and structs. Impls are only supported for structs and if you have multiple impls for a struct (such as &[T] and [T]), those may conflict. Structs only support opaque refinements.

Import the procedural macros

In order to use an extern spec you need to add a dependency on flux_rs. Right now this needs to be done as a local dependency since it is not published. Below is an example of how you can include it, although the version may be different.

[dependencies]
flux-rs = { path = "path-to-flux/flux/flux-rs", version = "0.1.0" }

Then in your code you will need to include the extern_spec attribute macro.

use flux_rs::extern_spec;

Extern functions

An example of refining an extern function can be found here.

To define an extern spec on a function, you need to do three things, which happen to correspond to each of the below lines.

#[extern_spec(std::mem)]
#[flux::sig(fn(&mut i32[@a], &mut i32{v : a < v }) -> ())]
fn swap(a: &mut i32, b: &mut i32);
  1. Add the #[extern_spec] attribute. This attribute optionally takes a path; in the above example, this is std::mem. You can use this path to qualify the function. So in the above example, the function we are targeting has the full path of std::mem::swap.
  2. Add a #[flux::sig(...)] attribute. This is required for any extern spec on a function. This signature behaves as if the #[flux::trusted] attribute was added, because we can't actually check the implementation. We just verify some simple things, like that the function arguments have compatible types.
  3. Write a function stub that matches the external function.

If you do the above, you can use std::mem::swap as if it were refined by the above type.

You shouldn't need to know the details, but here's how the macro works. It parses the std::mem into a module path and then transforms the function into

#[flux::extern_spec]
#[flux::sig(fn(&mut i32[@a], &mut i32{v : a < v }) -> ())]
#[allow(unused, dead_code)]
fn __flux_extern_spec_swap(a: &mut i32, b: &mut i32) {
    std::mem::swap(a, b)
}

It does this to get information about the function std::mem::swap and its arguments (this turns out to be difficult to do without giving the compiler something to inspect and type check).

Extern structs and impls

An example of refining an extern struct and impl can be found here. A simpler example just involving structs can be found here.

The syntax for an extern spec on a struct is very similar to that for a function. Once again, each line in the example happens to correspond to a step.

#[extern_spec(std::string)]
#[flux::refined_by(len: int)]
struct String;
  1. Add the #[extern_spec] attribute. This attribute optionally takes a path; in the above example, this is std::string. You can use this path to qualify the function. So in the above example, the struct we are targeting has the full path of std::string::String.
  2. Add a #[flux::refined_by(...)] attribute. This is required for any extern spec on a struct. Right now these attributes behave as if they were opaque (#[flux::opaque]), although we may support non-opaque extern structs.
  3. Write a stub for the extern struct.

If you do the above, you can use std::string::String as if it were refined by an integer index.

The syntax for an extern impl is a little different than that for functions or structs.

#[extern_spec(std::string)]
impl String {
    #[flux::sig(fn() -> String[0])]
    fn new() -> String;

    #[flux::sig(fn(&String[@n]) -> usize[n])]
    fn len(s: &String) -> usize;
}
  1. You still need to add the #[extern_spec] attribute, with the same optional argument of the path as above.
  2. You need to write out the impl block for the struct you want to refine. This struct does not need an extern spec, since by refining the impl you're only refining its methods.
  3. Write an extern spec for each function you wish to refine (this may be a subset). This is written just like a function extern spec with the caveat that the self parameter is not presently supported. So for example, instead of writing fn len(&self) -> usize;, you need to write fn len(s: &String) -> usize;.

If you do the above, you can use the above methods ofstd::string::String as if they were refined.

You shouldn't need to know the details, but here's how the above two macros expand.

For structs:

#[flux::extern_spec]
#[allow(unused, dead_code)]
#[flux::refined_by(len: int)]
struct __FluxExternSpecString(std::string::String);

For impls (this was translated manually so there might be some bugs):

#[allow(unused, dead_code)]
struct __FluxExternImplStructString;

#[allow(unused, dead_code)]
impl __FluxExternImplStructString {
    #[flux::extern_spec]
    #[flux::sig(fn() -> String[0])]
    #[allow(unused, dead_code)]
    fn __flux_extern_spec_new() -> String {
       std::string::String::new::<>()
    }
    #[flux::extern_spec]
    #[flux::sig(fn(&String[@n]) -> usize[n])]
    #[allow(unused, dead_code)]
    fn __flux_extern_spec_len(s: &String) -> usize {
       std::string::String::len::<>(s)
    }
}

Grammar of Refinements

r ::= n                     // numbers 1,2,3...
    | x                     // identifiers x,y,z...
    | x.f                   // index-field access
    | r + r                 // addition
    | r - r                 // subtraction
    | n * e                 // multiplication by constant
    | if r { r } else { r } // if-then-else
    | f(r...)               // function application
    | true | false          // booleans
    | r == r                // equality
    | r != r                // not equal
    | r < r                 // less than
    | r <= r                // less than or equal
    | r > r                 // greater than
    | r >= r                // greater than or equal
    | r || r                // disjunction
    | r && r                // conjunction
    | r => r                // implication
    | !r                    // negation

Developer's Guide

Regression Tests

You can run the various regression tests in the tests/pos and tests/neg directories using cargo xtask test

This will build the flux binary and then run it against the entire test suite. You can optionally pass a filter to only run tests containing some substring. For example:

$ cargo xtask test impl_trait
   Compiling xtask v0.1.0 (/path/to/flux/xtask)
    Finished dev [unoptimized + debuginfo] target(s) in 0.29s
     Running `target/debug/xtask test impl_trait`
$ cargo build
    Finished dev [unoptimized + debuginfo] target(s) in 0.05s
$ cargo test -p tests -- --test-args impl_trait
   Compiling fluxtests v0.1.0 (/path/to/flux/tests)
    Finished test [unoptimized + debuginfo] target(s) in 0.62s
     Running tests/compiletest.rs (target/debug/deps/compiletest-1241128f1f51caa4)

running 5 tests
test [ui] pos/surface/impl_trait04.rs ... ok
test [ui] pos/surface/impl_trait03.rs ... ok
test [ui] pos/surface/impl_trait01.rs ... ok
test [ui] pos/surface/impl_trait00.rs ... ok
test [ui] pos/surface/impl_trait02.rs ... ok

test result: ok. 5 passed; 0 failed; 0 ignored; 0 measured; 191 filtered out; finished in 0.10s


running 2 tests
test [compile-fail] neg/surface/impl_trait00.rs ... ok
test [compile-fail] neg/surface/impl_trait02.rs ... ok

test result: ok. 2 passed; 0 failed; 0 ignored; 0 measured; 207 filtered out; finished in 0.09s

Testing Flux on a File

When working on Flux, you may want to test your changes by running it against a test file. You can use cargo xtask run <input> to run Flux on a single input file. The command will set appropriate flags to be able to use custom Flux attributes and macros, plus some extra flags useful for debugging. For example:

$ cat test.rs
#[flux::sig(fn(x: i32) -> i32[x + 1])]
fn add1(x: i32) -> i32 {
    x + 1
}
$ cargo xtask run test.rs

The command will use a super set of the flags passed when running regression tests. Thus, a common workflow is to identify a failing test and run it directly with cargo xtask run, or alternatively copy it to a different file.

You may also find useful to create a directory in the root of the project and add it to .git/info/exclude. You can keep files there, outside of version control, and test Flux against them. I have a directory called attic/ where I keep a file named playground.rs. To run Flux on it, I do cargo xtask run attic/playground.rs.

Reporting locations where errors are emitted

When you use cargo xtask run you'll see that we report the location an error was emitted, e.g.,

error[FLUX]: refinement type error
 --> attic/playground.rs:4:5
  |
4 |     0
  |     ^ a postcondition cannot be proved
-Ztrack-diagnostics: created at crates/flux-refineck/src/lib.rs:114:15   <------- this

You can also pass -Ztrack-diagnostics=y to enable it if you are not using cargo xtask run

Running outside the project

To run Flux in a package outside the flux repo you need to install the binaries globally. You can do that using cargo xtask install. If you are continuously testing new changes it could be annoying to do it each time. To deal with this, you can set the FLUX_SYSROOT environment variable to change the location where cargo-flux and rustc-flux load the flux-driver. You can set it globally to point to the target/debug directory inside your local copy of the repo. This way you won't have to run cargo xtask install after every change, and you can be sure you'll be using the latest local debug build. Just be aware that the rustc-flux and cargo-flux binaries are built for a specific toolchain, and you will get a dynamic linking error if the flux-driver was compiled with a different one. This is to say, you should at least run cargo xtask install every time after the toolchain is updated.

Profiling Flux

Set FLUX_DUMP_TIMINGS=true to have flux write timing diagnostics to ./log/timings.

Right now this is extremely simple, it just provides some details for the spans under flux_typeck and flux_driver.

Sample output

Below is a sample output for an invocation of cargo-flux check that took 19 seconds. The missing 2 seconds approximately accounts for the time it takes for cargo check to run.

Note that check_crate contains everything running under check_top, which is why the sum of the spans is greater than 19 seconds.

check_top
  Checker::infer
    num events:   205
    min non-zero: 0.52ms
    1st quartile: 0.52ms
    2nd quartile: 1.05ms
    3rd quartile: 2.62ms
    max:          24.12ms
    total time:   229.64ms
  Checker::check
    num events:   205
    min non-zero: 0.52ms
    1st quartile: 0.52ms
    2nd quartile: 1.05ms
    3rd quartile: 5.24ms
    max:          159.91ms
    total time:   2028.47ms
  FixpointCtx::check
    num events:   205
    min non-zero: 22.02ms
    1st quartile: 26.21ms
    2nd quartile: 28.31ms
    3rd quartile: 40.37ms
    max:          1867.51ms
    total time:   9106.36ms
total time: 11364.47ms

check_crate
  Callbacks::check_wf
    num events:   1
    min non-zero: 18.35ms
    1st quartile: 18.87ms
    2nd quartile: 18.87ms
    3rd quartile: 18.87ms
    max:          18.87ms
    total time:   18.87ms
  Callbacks::check_crate
    num events:   1
    min non-zero: 16986.93ms
    1st quartile: 16995.32ms
    2nd quartile: 16995.32ms
    3rd quartile: 16995.32ms
    max:          16995.32ms
    total time:   16995.32ms
total time: 17014.19ms

Macro expansion

For example if you have code like in path/to/file.rs

#![allow(unused)]
fn main() {
#[extern_spec]
#[flux::refined_by(elems: Set<T>)]
struct HashSet<T, S = RandomState>;
}

and you want to see what the extern_spec macro expands it out to, then run

cargo x run -- -Zunpretty=expanded path/to/file.rs

Or you can run the xtask command directly

cargo x expand path/to/file.rs

High-level Architecture

Flux is implemented as a compiler driver. We hook into the compiler by implementing the Callbacks trait. The implementation is located is in the flux-driver crate, and it is the main entry point to Flux.

Crates

  • crates/flux-bin: Contains the cargo-flux and rustc-flux binaries used to launch the flux-driver.
  • crates/flux-common: Common utility definitions used across all crates.
  • crates/flux-config: Crate containing logic associated with global configuration flags that change the behavior of Flux, e.g, to enable or disable overflow checking.
  • crates/flux-desugar: Implementation of name resolution and desugaring from Flux surface syntax into Flux high-level intermediate representation (fhir). This includes name resolution.
  • crates/flux-driver: Main entry point to Flux. It contains the flux-driver binary and the implementation of the Callbacks trait.
  • crates/flux-errors: Utility definitions for user facing error reporting.
  • crates/flux-fhir-analysis: Implements the "analyses" performed in the fhir, most notably well-formedness checking and conversion from fhir into rty.
  • crates/flux-fixpoint: Code to interact with the Liquid Fixpoint binary.
  • crates/flux-macros: Procedural macros used internally to implement Flux.
  • crates/flux-metadata: Logic for saving Flux crate metadata that can be used to import refined signatures from external crates.
  • crates/flux-middle: This crate contains common type definitions that are used by the rest of Flux like the rty and fhir intermediate representations. Akin to rustc_middle.
  • crates/flux-refineck: Implementation of refinement type checking.
  • crates/flux-syntax: Definition of the surface syntax AST and parser.
  • tests: Flux regression tests.
  • lib/flux-attrs: Implementation of user facing procedural macros for annotating programs with Flux specs.
  • lib/flux-rs: This is just a re-export of the macros implemented in flux-attrs. The intention is to eventually put Flux "standard library" here, i.e., a set of definitions that are useful when working with Flux.

Intermediate Representations

Flux has several intermediate representations (IR) for types. They represent a refined version of an equivalent type in some rustc IR. We have picked a distinct verb to refer to the process of going between these different representations to make it easier to refer to them. The following image summarizes all the IRs and the process for going between them.

IRs diagram

Surface

The surface IR represents source level Flux annotations. It corresponds to the rustc_ast data structures in rustc. The definition as well as the parser is located in the flux-syntax crate.

Fhir

The Flux High-Level Intermediate Representation (fhir) is a refined version of rustc's hir. The definition is located in the flux_middle crate inside the fhir module. The process of going from surface to fhir is called desugaring, and it is implemented in the flux-desugar crate.

Rty

The definition in the flux_middle::rty module correspond to a refined version of the main rustc representation for types defined in rustc_middle::ty. The process of going from fhir to rty is called conversion, and it is implemented in the flux_fhir_analysis::conv module.

Simplified Rustc

The definition in the flux_middle::rustc module correspond to simplified version of data structures in rustc. They can be understood as the currently supported subset of Rust. The process of going from a definition in rustc_middle into flux_middle::rustc is called lowering and it is implemented in flux_middle::rustc::lowering.

Lifting and Refining

Besides the different translation between Flux intermediate representations, there are two ways to get a refined version from a rust type. The process of going from a type in hir into a type in fhir is called lifting, and it is implemented in flux_middle::fhir::lift. The process for going from a type in flux_middle::rustc::ty into a flux_middle::rty is called refining, and it is implemented flux_middle::rty::refining.

Introducing Flux

Online demo

Types bring order to code. For example, if a variable i:usize then we know i is a number that can be used to index a vector. Similarly, if v:vec<&str> then we can be sure that v is a collection of strings which may be indexed but of course, not used as an index. However, by itself usize doesn't tell us how big or small the number and hence the programmer must still rely on their own wits, a lot of tests, and a dash of optimism, to ensure that all the different bits fit properly at run-time.

Refinements are a promising new way to extend type checkers with logical constraints that specify additional correctness requirements that can be verified by the compiler, thereby entirely eliminating various classes of run-time problems.

We're excited to introduce Flux, a refinement type checker plugin that brings this technology to Rust.

Indexed Types

The most basic form of refinement type in flux is a type that is indexed by a logical value. For example

TypeMeaning
i32[10]The (singleton) set of i32 values equal to 10
bool[true]The (singleton) set of bool values equal to true

Post-Conditions

We can already start using these indexed types to start writing (and checking) code. For example we can write the following specification which says that the value returned by mk_ten must in fact be 10

#![allow(unused)]
fn main() {
#[flux::sig(fn() -> i32[10])]
pub fn mk_ten() -> i32 {
    5 + 4
}
}

but when you compile it, flux will say

error[FLUX]: postcondition might not hold
 --> src/basics.rs:7:5
  |
7 |     5 + 4
  |     ^^^^^

The error says that that the postcondition might not hold which means that the output produced by mk_ten may not in fact be an i32[10] as indeed, in this case, the result is 9! You can eliminate the error by replacing the body with 5 + 5 or just 10.

Pre-Conditions

Here's a second example that shows how you can use an index to restrict the space of inputs that a function expects.

#![allow(unused)]
fn main() {
#[flux::sig(fn (b:bool[true]))]
pub fn assert(b:bool) {
  if !b { panic!("assertion failed") }
}
}

Here, the refined specification for assert says that you can only call it with true as the input. So if you write

#![allow(unused)]
fn main() {
fn test(){
  assert(2 + 2 == 4);
  assert(2 + 2 == 5); // fails to type check
}
}

then flux will complain that

#![allow(unused)]
fn main() {
error[FLUX]: precondition might not hold
  --> src/basics.rs:12:5
   |
12 |     assert(2 + 2 == 5); // fails to type check
   |     ^^^^^^^^^^^^^^^^^^
}

meaning that the call to assert fails to establish that the input is indeed true (as of course, in this case, it is not!)

Index Parameters and Expressions

It's not terribly exciting to only talk about fixed values like 10 or true. To be more useful, flux lets you index types by refinement parameters. For example, you can write

#![allow(unused)]
fn main() {
#[flux::sig(fn(n:i32) -> bool[0 < n])]
pub fn is_pos(n: i32) -> bool {
    if 0 < n {
        true
    } else {
        false
    }
}
}

Here, the type says that is_pos

  • takes as input some i32 indexed by n
  • returns as output the bool indexed by 0 < n

in other words, the output is true exactly when 0 < n.

We might use this function to check that:

#![allow(unused)]
fn main() {
pub fn test_pos(n: i32) {
  let m = if is_pos(n) { n - 1 } else { 0 };
  assert(0 <= m);
}
}

Existential Types

Often we don't care about the exact value of a thing -- but just care about some properties that it may have. For example, we don't care that an i32 is equal to 5 or 10 or n but that it is non-negative.

TypeMeaning
i32{v: 0 < v}The set of i32 values that positive
i32{v: n <= v}The set of i32 values greater than or equal to n

Flux allows such specifications by pairing plain Rust types with assertions 1 that constrain the value. For example, we can rewrite mk_10 with the output type i32{v:0<v} that specifies a weaker property: the value returned by mk_ten is positive.

#![allow(unused)]
fn main() {
#[flux::sig(fn() -> i32{v: 0 < v})]
pub fn mk_ten() -> i32 {
    5 + 5
}
}

Similarly, you might specify that a function that computes the absolute value of an i32 with a type which says the result is non-negative and exceeds the input n.

#![allow(unused)]
fn main() {
#[flux::sig(fn (n:i32) -> i32{v:0<=v && n<=v})]
pub fn abs(n: i32) -> i32 {
    if 0 <= n {
        n
    } else {
        0 - n
    }
}
}

As a last example, you might write a function to compute the factorial of n

#![allow(unused)]
fn main() {
#[flux::sig(fn (n:i32) -> i32{v:1<=v && n<=v})]
pub fn factorial(n: i32) -> i32 {
    let mut i = 0;
    let mut res = 1;
    while i < n {
        i += 1;
        res = res * i;
    }
    res
}
}

Here the specification says the input must be non-negative, and the output is at least as large as the input. Note, that unlike the previous examples, here we're actually changing the values of i and res.

Can you guess why the copilot suggestions failed to pass flux, and what refinements were inferred for i and res in the fixed code at the end?

Summary

In this post, we saw how Flux lets you

  1. decorate basic Rust types like i32 and bool with indices and constraints that let you respectively refine the sets of values that inhabit that type, and

  2. specify contracts on functions that state pre-conditions on the sets of legal inputs that they accept, and post-conditions that describe the outputs that they produce.

The whole point of Rust, of course, is to allow for efficient imperative sharing and updates, without sacrificing thread- or memory-safety. Next time, we'll see how Flux melds refinements and Rust's ownership to make refinements happily coexist with imperative code.

1

These are not arbitrary Rust expressions but a subset of expressions from logics that can be efficiently decided by SMT Solvers

Ownership in Flux

Online demo

Previously we saw how to refine basic Rust types like i32 and bool with indices and constraints to constrain the set of values described by those types.

The whole point of Rust, of course, is to allow for efficient imperative sharing and updates, via the clever type system that keeps an eye on the ownership of resources to make sure that aliasing and mutation cannot happen at the same time.

Next, lets see how Flux melds refinements and Rust's ownership mechanisms to make refinements work in the imperative setting.

Exclusive Ownership

Rust's most basic form of ownership is exclusive ownership, in which exactly one variable in a function has the right to mutate a memory location. When a location is exclusively owned, we can be sure that there are no other references to it, which lets flux update the type whenever the location is changed. For example, consider the program

#![allow(unused)]
fn main() {
#[flux::sig(fn () -> i32[3])]
pub fn mk_three() -> i32 {
    let mut r = 0;  // r: i32[0]
    r += 1;
    assert(r == 1); // r: i32[1]
    r += 1;
    assert(r == 2); // r: i32[2]
    r += 1;
    assert(r == 3); // r: i32[3]
    r
}
}

The variable r has different types at each point inside mk_three. It starts off as i32[0]. The first increment changes it to i32[1], then i32[2] and finally, the returned type i32[3].

This exclusive ownership mechanism is at work in the factorial example we signed off with previously

#![allow(unused)]
fn main() {
#[flux::sig(fn (n:i32{0 <= n}) -> i32{v:n <= v})]
pub fn factorial(n: i32) -> i32 {
    let mut i = 0;  // i: i32[0]
    let mut r = 1;  // r: i32[1]
    while i < n {
                    // i: i32{v:0<=v<=n}
                    // r: i32{v:1<=v && i<=v}
        i += 1;
        r = r * i;
    }
    r
}
}

In the above code, i and r start off at 0 and 1 but then Rust infers (a story for another day) that inside the while-loop1

  • i has type i32{v:0<=v && v < n}
  • r has type i32{v:1<=v && i <= v}

and hence, upon exit since i == n we get that the result is at least n.

Borrowing: Shared References

Exclusive ownership suffices for simple local updates like in factorial. However, for more complex data, functions must temporarily relinquish ownership to allow other functions to mutate the data. Rust cleverly allows this via the notion of borrowing using two kinds of references that give callees temporary access to a memory location.

The simplest kind of references are of the form &T which denote read-only access to a value of type T. For example, we might write abs to take a shared reference to an i32

#![allow(unused)]
fn main() {
#[flux::sig(fn (p: &i32[@n]) -> i32{v:0<=v && n<=v})]
pub fn abs(p: &i32) -> i32 {
    let n = *p;
    if 0 <= n {
        n
    } else {
        0 - n
    }
}
}

Notice that the input type has changed: the function now

  • Accepts p a reference to an i32 whose value is n as denoted by @n
  • Returns an i32 that is non-negative and larger than n

The @ marks the n as a refinement parameter whose value is automatically computed by flux during type checking.

So, for example, Flux can check the below code by automatically determining that the refinement parameter at the call-site is 10.

#![allow(unused)]
fn main() {
pub fn test_abs() {
    let z = 10;
    assert(0 <= abs(&z))
    assert(10 <= abs(&z))
}
}

Refinement Parameters

As an aside, we have secretly been using refinement parameters like @n all along. For example, Flux automatically desugars the signature fn(n:i32{0 <= n} -> ... that we wrote for factorial into

#![allow(unused)]
fn main() {
fn ({i32[@n] : 0 <= n}) -> i32{v:n <= v}
}

where @n is a refinement parameter that is implicitly determined from the rust parameter n:i32. However, explicit parameters are essential to name the value of what a reference points to. In abs the rust parameter p names the reference but the @n names the (input) value and lets us use it to provide more information about the output of abs.

Flux is modular in that the only information it knows about the implementation of abs is the signature: for example if we remove the fact that the output exceeds n then Flux will reject the assertion 10 <= abs(&z).

Borrowing: Mutable References

References of type &mut T denote mutable references that can be used to (read and) write or update the contents of a T value. Crucially, Rust ensures that while there may be multiple read-only (shared) references to a location, there is at most one active writeable (mutable) reference at any point in time.

Flux exploits the semantics of &mut T to treat T as an invariant of the underlying data. As an example, consider the following function that decrements the value of a mutable reference while ensuring the data is non-negative:

#![allow(unused)]
fn main() {
#[flux::sig(fn(p: &mut i32{v:0 <= v}))]
pub fn decr(p: &mut i32) {
    *p = *p - 1;
}
}

Flux will complain with the following message

#![allow(unused)]
fn main() {
error[FLUX]: assignment might be unsafe
  --> src/basics.rs:13:9
   |
13 |         *p = *p - 1;
   |         ^^^^^^^^^^^
}

as in fact, we may be writing a negative value into *p if, for example, the old value was zero. We can fix this code by guarding the update with a test that ensures the original contents are in fact non-zero

#![allow(unused)]
fn main() {
#[flux::sig(fn(p: &mut i32{v:0 <= v}))]
pub fn decr(p: &mut i32) {
    let n = *p;
    if n != 0 {
        *p = n - 1;
    }
}
}

at which point Flux is happy to sign off on the code.

Aliased References

Flux uses Rust's borrowing rules to track invariants even when there may be aliasing. As an example, consider the function

#![allow(unused)]
fn main() {
#[flux::sig(fn (bool) -> i32{v:0 <= v})]
fn test_alias(z: bool) -> i32 {
    let mut x = 1;  // x: i32[1]
    let mut y = 2;  // y: i32[2]
    let r = if z { &mut x } else { &mut y };
                    // r: &mut i32{v:0 <= v}
    decr(r);
    *r
}
}

The reference r could point to either x or y depending on the (unknown) value of the boolean z. Nevertheless, Flux determines that both references &mut x and &mut y point to values of the more general type i32{v:0<=v} and hence, infers r : &mut i32{v:0<=v} which allows us it to then call decr with the reference and guarantee the result (after decr) is still non-negative.

Borrowing: Strong References

In many situations, we want to lend a value to another function that actually changes the value's (refinement) type upon exit. For example, consider the following function to increment a reference to a non-negative i32

#![allow(unused)]
fn main() {
#[flux::sig(fn (p: &mut i32{v:0 <= v}))]
fn incr(p: &mut i32) {
  *p += 1
}
}

Recall that Flux is modular in that the only information it has about incr is what is said in the signature. The signature for incr only says p remains non-negative: Flux does not know that incr actually increments the value of p.

Hence, Flux fusses that the following assert may fail even though its patently obvious that it will succeed!

To verify test_incr we need a signature for incr that says that its output is indeed one greater2 than its input. Flux extends Rust with the notion of strong references of the form &strg T which refine Rust's &mut T to grant exclusive access to the underlying T. Crucially, strong references also let us specify how the type is updated when the function exits3. Thus, we can use strong references to type incr as

#![allow(unused)]
fn main() {
#[flux::sig(fn(p: &strg i32[@n]) ensures p:i32[n+1])]
fn incr(p: &mut i32) {
  *p += 1
}
}

The Flux signature refines the plain Rust one to specify that

  1. p is a strong reference to an i32,
  2. the input type of *p is i32[n], and
  3. the output type of *p is i32[n+1].

With this specification, Flux merrily checks test_incr, by determining that the refinement parameter @n is 10 and hence, that upon return x: i32[11].

Summary

To sum up, Flux exploits Rust's ownership mechanisms to track properties of shared (&T) and mutable (&mut T) references, and additionally adds a strong (&strg T) reference -- a special case of &mut -- to support the cases where the type itself is changed by a call.

Next, we'll see how refinements and ownership yield a simple refined API for vectors that lets Flux check bounds safety at compile time...

1

For those familiar with the term, these types are loop invariants

2

Setting aside the issue of overflows for now

3

Thereby allowing so-called strong updates in the type specifications

Refined Vectors

Online demo

While rustc has a keen eye for spotting nasty bugs at compile time, it is not omniscient. We've all groaned in dismay at seeing deployed code crash with messages like

panicked at 'index out of bounds: the len is ... but the index is ...'

Next, lets see how flux's refinement and ownership mechanisms let us write a refined vector API whose types track vector sizes and ensure --- at compile time --- that vector accesses cannot fail at runtime.

Refining Vectors to Track their Size

To begin with, we will defined a refined vector type which is simply a wrapper around the standard Vec type

#![allow(unused)]
fn main() {
#[flux::refined_by(len: int)]
pub struct RVec<T> {
    inner: Vec<T>,
}
}

The #[flux::refined_by(len: int)] attribute tells flux that the type RVec<T> struct is indexed by a len refinement which tracks the size of the underlying vector, just like the indices for i32 and bool tracked the actual value of the underlying integer or boolean). The idea is that the type

  • RVec<i32>[10] represents a vector of i32 size 10, and
  • RVec<bool>{v:0 < v} represents a non-empty vector of bool, and
  • RVec<RVec<f32>[n]>[m] represents a vector of vectors of f32 of size m and each of whose elements is a vector of size n.

Creating Vectors

Now that we can talk about the size of a vector, lets build up an API for creating and manipulating vectors. I suppose one must start with nothing: an empty vector.

#![allow(unused)]
fn main() {
impl<T> RVec<T> {
    #[flux::trusted]
    #[flux::sig(fn() -> RVec<T>[0])]
    pub fn new() -> Self {
        Self { inner: Vec::new() }
    }
}
}

The above implements RVec::new as a wrapper around Vec::new. The #[flux::trusted] attribute tells Flux there is nothing to "check" here, as we are defining the API itself and trusting that the implementation (using vec is correct). However, the signature says that callers of the RVec::new get back a vector indexed with 0 i.e. an empty vector.

Pushing Values

An empty vector is a rather desolate thing.

To be of any use, we need to be able to push values into the container, like so

#![allow(unused)]
fn main() {
#[flux::trusted]
#[flux::sig(fn(self: &strg RVec<T>[@n], T)
            ensures self: RVec<T>[n+1])]
pub fn push(&mut self, item: T) {
    self.inner.push(item);
}
}

The refined type for push says that it takes a strong reference (self) --- where strg means the refined type may be changed by the function --- to an RVec<T> of size n and a value T and upon exit, the size of self is increased by 1.

Popping Values

Not much point stuffing things into a vector if we can't get them out again.

For that, we might implement a pop method that returns the last element of the vector. Aha, but what if the vector is empty? You could return an Option<T> or since we're tracking sizes, we could require that pop only be called with non-empty vectors.

#![allow(unused)]
fn main() {
#[flux::trusted]
#[flux::sig(fn(self: &strg {RVec<T>[@n] | 0 < n}) -> T
            ensures self: RVec<T>[n-1])]
pub fn pop(&mut self) -> T {
  self.inner.pop().unwrap()
}
}

Using the API

Now already flux can start checking some code, for example if you push two elements, then you can pop twice, but flux will reject the third pop at compile-time

In fact, the error message from flux will point to exact condition that does not hold

#![allow(unused)]
fn main() {
error[FLUX]: precondition might not hold
  --> src/vectors.rs:24:5
   |
24 |     v.pop();
   |     ^^^^^^^ call site
   |
   = note: a precondition cannot be proved at this call site
note: this is the condition that cannot be proved
  --> src/rvec.rs:78:47
   |
78 |     #[flux::sig(fn(self: &strg {RVec<T>[@n] | 0 < n}) -> T
   |                                               ^^^^^
}

Querying the Size

Perhaps we should peek at the size of the vector to make sure its not empty before we pop it. We can do that with a len method whose type says that the returned usize is, in fact, the size of the input vector

#![allow(unused)]
fn main() {
#[flux::trusted]
#[flux::sig(fn(&RVec<T>[@n]) -> usize[n])]
pub fn len(&self) -> usize {
    self.inner.len()
}
}

Now, flux "knows" that after two pushes, the size of the vector is 2 and after the two pops, the size is 0 again

Random Access

Of course, vectors are not just stacks, they also allow random access to their elements which is where those pesky panics occur, and where the refined vector API gets rather useful. Since we're tracking sizes, we can require that the method to get an element only be called with a valid index that is between 0 and the vector's size

#![allow(unused)]
fn main() {
#[flux::sig(fn(&RVec<T>[@n], i: usize{i < n}) -> &T)]
pub fn get(&self, i: usize) -> &T {
    &self.inner[i]
}

#[flux::sig(fn(&mut RVec<T>[@n], i: usize{i < n}) -> &mut T)]
pub fn get_mut(&mut self, i: usize) -> &mut T {
    &mut self.inner[i]
}
}

With these refined get methods, flux can now spot the ``off-by-one'' error in the following code and accepts the fix 1

Its a bit gross to use get and get_mut directly, so instead we implement the Index and IndexMut traits for RVec which allows us to use the [] operator to access elements

#![allow(unused)]
fn main() {
impl<T> std::ops::Index<usize> for RVec<T> {
    type Output = T;
    #[flux::sig(fn(&RVec<T>[@n], i:usize{i < n}) -> &T)]
    fn index(&self, index: usize) -> &T {
        self.get(index)
    }
}

impl<T> std::ops::IndexMut<usize> for RVec<T> {
    #[flux::sig(fn(&mut RVec<T>[@n], i:usize{i < n}) -> &mut T)]
    fn index_mut(&mut self, index: usize) -> &mut T {
        self.get_mut(index)
    }
}
}

And now the above vec_sum example looks a little nicer

Memoization

Lets put the whole API to work in this "memoized" version of the fibonacci function which uses a vector to store the results of previous calls

#![allow(unused)]
fn main() {
pub fn fib(n: usize) -> i32 {
    let mut r = RVec::new();
    let mut i = 0;
    while i < n {
        if i == 0 {
            r.push(0);
        } else if i == 1 {
            r.push(1);
        } else {
            let a = r[i - 1];
            let b = r[i - 2];
            r.push(a + b);
        }
        i += 1;
    }
    r.pop()
}
}

Oops, flux is not happy with the call to pop at the end of the function which returns the last value as the result.

#![allow(unused)]
fn main() {
error[FLUX]: precondition might not hold
  --> src/vectors.rs:40:5
   |
40 |     r.pop()
   |     ^^^^^^^
}

Flux complains that the vector may be empty and so the pop call may fail ... but why? Can you spot the problem?

Indeed, we missed a "corner" case -- when n is 0 we skip the loop and so the vector is empty! Once we add a test for that, flux is happy.

As a last example, lets look at a simplified version of the binary_search method from std::vec, into which I've snuck a tiny little bug

#![allow(unused)]
fn main() {
pub fn binary_search(vec: &RVec<i32>, x: i32) -> Result<usize, usize> {
    let mut size = vec.len();
    let mut left = 0;
    let mut right = size;
    while left <= right {
        let mid = left + size / 2;
        let val = vec[mid];
        if val < x {
            left = mid + 1;
        } else if x < val {
            right = mid;
        } else {
            return Ok(mid);
        }
        size = right - left;
    }
    Err(left)
}
}

Flux complains in two places

#![allow(unused)]
fn main() {
error[FLUX]: precondition might not hold
   --> src/vectors.rs:152:19
    |
152 |         let val = vec[mid];
    |                   ^^^^^^^^ call site
    |
    = note: a precondition cannot be proved at this call site
note: this is the condition that cannot be proved
   --> src/rvec.rs:189:44
    |
189 |     #[flux::sig(fn(&RVec<T>[@n], usize{v : v < n}) -> &T)]
    |                                            ^^^^^

error[FLUX]: arithmetic operation may overflow
   --> src/vectors.rs:160:9
    |
160 |         size = right - left;
    |         ^^^^^^^^^^^^^^^^^^^
}

  • The vector access may be unsafe as mid could be out of bounds!

  • The size variable may underflow as left may exceed right!

Can you the spot off-by-one and figure out a fix?

Summary

So, we saw how Flux's index and constraint mechanisms combine with Rust's ownership to let us write a refined vector API that ensures the safety of all accesses at compile time.

Next time, we'll see how these mechanisms are compositional in that we can use standard type machinery to build up compound structures and APIs from simple ones.

1

Why not use an iterator? We'll get there in due course!

About Flux

Flux is a research project described in the paper

Team

Flux is being developed by

Code

Flux is open-source and available here

Thanks

This work was supported by the National Science Foundation, European Research Council, and by generous gifts from Microsoft Research.

Limitations

This is a prototype! Use at your own risk. Everything could break and it will break.