Fast Thread Locals In Rust

Rust thread-locals are slower than they could be. This is because they violate zero-cost abstraction principle, specifically the you dont pay for what you dont use bit.

Rusts thread-local implementation( 1, 2 ) comes with built-in support for laziness thread locals are initialized on the first access. Sometimes this overhead is a big deal, as thread locals are a common tool for writing high-performance code. For example, allocator fast path often involves looking into thread-local heap.

Theres an unstable #[thread_local] attribute for a zero-cost implementation (see the tracking issue).

Lets see how much is thread local initialized? check costs by comparing these two programs:

thread_local! {
  static COUNTER: Cell<u32> = Cell::new(0);

const STEPS: u32 = 1_000_000_000;
fn sum_rust() -> u32 {
  for step in 0..STEPS {
    COUNTER.with(|it| {
      let inc = step.wrapping_mul(step) ^ step;
  COUNTER.with(|it| it.get())

fn main() {
  let t = Instant::now();
  let r = sum_rust();
  eprintln!("Rust:   {} {}ms", r, t.elapsed().as_millis());
#define _POSIX_C_SOURCE 200809L

#include "inttypes.h"
#include "stdint.h"
#include "stdio.h"
#include "threads.h"
#include "time.h"

thread_local uint32_t COUNTER = 0;

const uint32_t STEPS = 1000000000;

uint32_t sum_c() {
  for (uint32_t step = 0; step < STEPS; step++) {
    uint32_t inc = (step * step) ^ step;
    COUNTER += inc;
  return COUNTER;

uint64_t now_ms() {
  struct timespec spec;
  clock_gettime(CLOCK_MONOTONIC, &spec);
  return spec.tv_sec * 1000 + spec.tv_nsec / 1000000;

int main(void) {
  uint64_t t = now_ms();
  uint32_t r = sum_c();
  printf("C:      %" PRIu32 " %"PRIu64"ms\n", r, now_ms() - t);
  return 0;

In this test, we declare an integer thread-local variable, and use it as an accumulator for the summation.

We use non-trivial summation term: (step * step) ^ step this is to prevent LLVM from evaluating the sum at compile time. If a term of a summation is a polynomial (like 1, step or step * step), then the sum itself is a one degree higher polynomial, and LLVM can figure this out! We rely on wrapping overflow of unsigned integers in C, and use wrapping_mul and wrapping_add in Rust. To make sure that both programs are equivalent, we also print the result.

One optimization we specifically dont protect from is caching thread-local access. That is, instead of doing a billion of thread-local loads and stores, the compiler could generate code to compute the sum into the local variable, and do a single store at the end. This is because can the compiler optimize thread-local access? is exactly the property we want to measure.

Theres no standard way to get monotonic wall-clock time in C, so the C version is not cross-platform.

This code gives the following results on my machine:

$ cargo build --release -q        && ./target/release/ftl
Rust:   62565888 487ms
$ clang -std=c17 -O3 ./src/main.c && ./a.out
C:      62565888 239ms

This benchmark doesnt allow to measure the cost of thread-local access per se, but the overall time is about 2x longer for Rust.

Can we make Rust faster? I dont know how to do that, but I know how to cheat. We can apply a general Rust extension trick write some C code and link it with Rust!

Lets implement a simple C library which declares a thread-local and provides access to it:

#include "stdint.h"
#include "threads.h"

thread_local uint32_t COUNTER = 0;

uint32_t* get_thread_local() {
  return &COUNTER;

Link it with Rust:

use std::{env, path::Path, process::Command};

fn main() {
  let out_dir = env::var("OUT_DIR").unwrap();

    .args(&[ "src/thread_local.c", "-O3", "-c", "-o"])
    .arg(&format!("{}/thread_local.o", out_dir))
    .args(&["crus", "libthread_local.a", "thread_local.o"])

  println!("cargo:rustc-link-search=native={}", out_dir);

And use it:

fn with_counter<T>(f: impl FnOnce(&Cell<u32>) -> T) -> T {
  extern "C" { fn get_thread_local() -> *mut u32; }
  let counter =
    unsafe { &*(get_thread_local() as *mut Cell<u32>) };

fn sum_rust_c() -> u32 {
  for step in 0..STEPS {
    with_counter(|it| {
      let inc = step.wrapping_mul(step) ^ step;
  with_counter(|it| it.get())

The result are underwhelming:

C:               62565888 239ms
Rust:            62565888 485ms
Rust/C:          62565888 1198ms

This is expected we replaced access to a thread local with a function call. As we are crossing the language boundary, the compiler cant inline it, which destroys performance. However, theres a way around that: Rust allows cross-language Link Time Optimization (docs). That is, Rust and C compilers can cooperate, to allow the linker to do inlining across the languages.

This requires to manually align a bunch of stars:

Now, just recompiling the old code gives the same performance for C and Rust:

C:               62565888 240ms
Rust:            62565888 495ms
Rust/C:          62565888 241ms

Interestingly, this is the same performance we get without any thread-locals at all:

fn sum_local() -> u32 {
  let mut counter = 0u32;
  for step in 0..STEPS {
    let inc = step.wrapping_mul(step) ^ step;
    counter = counter.wrapping_add(inc)

So, either the compiler/linker was able to lift thread-local access out of the loop, or its cost is masked by arithmetics.

Full code for the benchmarks is available at Note that this research only scratches the surface of the topic: thread locals are implemented differently on different OSes. Even on a single OS, there are be differences depending on compilation flags (dynamic libraries differ from static libraries, for example). Looking at the generated assembly could also be illuminating (code on Compiler Explorer).

Discussion on /r/rust.