There Is Life Before Main in Rust

A Rust developer explores what happens before the main function executes, detailing the runtime initialization phase and introducing novel techniques for mutable data. The post highlights how the Rust runtime builds atop C's runtime to configure panics, unwinding, and program arguments, and demonstrates the use of the linktime project for pre-main bootstrapping.

There Is Life Before Main in Rust permalink /blog/2026/06/11/life-before-main/ Disclosures 🧠 This post is 100% human-written /assets/2026/06/remarkable life after main.pdf . Claude was used for feedback and to assist with the linker symbol diagram. Cursor was used for feedback and to ensure examples were compilable. The author of this post is deeply interested in the topic of life-before-main: he is the author of the ctor https://crates.io/crates/ctor crate, and the creator of the project that we’ll be using in the examples below. https://github.com/mmastrac/linktime linktime Every Rust binary has one thing in common: fn main . If you come from the C world, that might be more familiar as int main argc, argv . Some platforms might obfuscate it a bit more, but under the hood, every binary has an entrypoint. We’re going to discuss what happens before main and what interesting things we can do there. In addition, we’ll be showing some novel techniques for mutable data that aren’t in common use in the Rust ecosystem today. This post is a deep dive into some technical details of how Rust source becomes a Rust binary. Some background knowledge may be helpful to the reader, including: Before main What might not be familiar to most developers is how you get into the main function. You see, under the hood for every language is the runtime . C has one: the C runtime that you might recognize as libc https://en.wikipedia.org/wiki/C standard library . Rust also has its own runtime: the Rust standard library. And because C is the lingua franca of runtimes for most executable code , Rust builds its own runtime atop of C’s, effectively building its own higher-level abstraction encapsulating C’s. 1 fn:0 A runtime is a bit fuzzy to define. It’s both the executable code that lives on disk and compilable headers and libraries used at compile time. But the purpose of a runtime is always the same: integrating developer code with the platform’s operating system. There’s an entire ecosystem of processing that happens before the function you declared as main starts up. C uses this to configure allocation, file access, thread-local storage and other C runtime services. Rust uses this time to configure parts of its own language and runtime. Specifically, Rust has infrastructure to handle panics and unwinding. Rust also needs to translate the C-style program arguments 2 into its own interface. The machinery for all this is https://doc.rust-lang.org/beta/std/env/fn.args.html std::env::args visible in the Rust compiler project https://github.com/rust-lang/rust/blob/main/compiler/rustc codegen ssa/src/base.rs L501 . Runtimes make use of this pre-main phase because it guarantees 1 running before user code, and 2 a single-threaded, highly-consistent and predictably-ordered environment, which allow for reliable and deterministic initialization. By not taking advantage of this environment, you are missing out on a very useful bootstrapping phase. We’ll see later on in this post how we can build some useful primitives making use of life before main. Entry Points A binary starts when the operating system’s loader 3 - the part of the OS that loads the binary into memory and sets up the environment - hands off control. The runtime is responsible for accepting the hand-off from the loader. There’s a platform-specific hook on every OS that accepts the hand-off - to some extent this is the real main. On Linux, the entry point is stored in the https://en.wikipedia.org/wiki/Executable and Linkable Format :~:text=e entry of the ELF header, and by default, the linker places the address of a symbol named e entry field start there. A similar hook exists on Windows https://learn.microsoft.com/en-us/windows/win32/debug/pe-format :~:text=AddressOfEntryPoint , and boots the executable in a function named https://stackoverflow.com/questions/1583193/what-functions-does-winmaincrtstartup-perform . At this point the C runtime has a chance to configure itself, and the way that all runtimes do this is via initialization functions. WinMainCRTStartup In early iterations of runtimes, bootstrapping was a static tree of function calls: initialize file I/O, initialize the allocator, etc. As runtimes became more complex, this tree of function calls became more complex, and binary sizes increased to absorb more C runtime functionality that they may or may not need. Over time, linkers developed the ability to discard unused code before even writing the binary to disk including unused parts of the C runtime , and with that came a need for a replacement for the static init call trees. The most popular method 4 of declaring init code came from GCC: attribute constructor . The way this worked was to place a list of init functions into a contiguous chunk of the binary on disk. When the C runtime started, it could walk through each of these functions and call them, allowing various bits of the C runtime to request initialization without strongly coupling subsystems, and allowing the linker to jettison unused subsystems, init code and all.Eventually the need for constructor ordering became important enough that constructors could be given a priority and run in a specific order, allowing the runtime to initialize subsystems before and after each other. E.g., the memory allocation malloc subsystem might be needed for buffered file I/O. On most platforms 5, the linker was called in to do the priority work: each platform ended up with a way to prioritize the order in which data gets written to sections, which allowed for the C runtime to end up with a well-ordered list of function pointers . 6 fn:4 We can even build an example of this by hand in Rust using the unsafe link section = "..." attribute try it in the Rust Playground https://play.rust-lang.org/?version=stable&mode=debug&edition=2024&gist=61a54d3cde75e732db52558f4ef9381c : /// Linux example: the modern glibc runtime uses .init array to hold function /// pointers, and a numeric suffix allows them to be ordered. Note that priorities /// less than or equal to 100 are reserved for the runtime itself, so any code that /// wants to use the C runtime must use a priority of 101 or higher. // On Linux, .init array holds function pointers , not functions. // We can convert a function to a function pointer with one of the below // blocks which is equivalent to this: // // used // <-- without this, Rust might decide the init function is unused and remove it // unsafe link section = ".init array.NNNNN" // <-- the section where we place the function pointer // static INIT ARRAY FN PTR: extern "C" fn // = function; // <-- the function pointer data: we assign the function to it // // extern "C" fn function { ... } // <-- the function itself used unsafe link section = ".init array.101" static INIT FN FIRST: extern "C" fn = const { extern "C" fn init { println "Initializing first " ; } init }; used unsafe link section = ".init array.201" static INIT FN SECOND: extern "C" fn = const { extern "C" fn init { println "Initializing second " ; } init }; fn main { println "Main " } linktime: ctor, link-section and more The examples in this post will work on Linux and various BSDs, but are not designed to be cross-platform examples. For example, macOS has start and stop symbols, but they are named differently 7. Windows does not support start and stop symbols, but has a set of rules for sorting sections that is effectively equivalent.Because platforms are so widely variable, we’ll be introducing the ctor https://crates.io/crates/ctor and crates from the https://crates.io/crates/link-section link-section as a way to abstract away platform-specific differences and hide the general complexity of linker work. https://github.com/mmastrac/linktime linktime projectThe excellent inventory https://crates.io/crates/inventory and are two other very popular crates built on the same principles, but have limitations https://crates.io/crates/linkme linkme that make them less suitable for the examples in this post. 8 fn:c1 If you’d like to learn more, the link-section crate contains a detailed report on platform-specific behaviour https://crates.io/crates/link-section :~:text=Platform%20Support . The ctor https://crates.io/crates/ctor crate is designed to handle all of the boilerplate of registering constructors in a cross-platform way. This allows us to simplify our examples above to: use ctor::ctor; ctor unsafe, priority = 101 fn init1 { println "Initializing first " ; } ctor unsafe, priority = 201 fn init2 { println "Initializing second " ; } fn main { println "Main " } Note that neither example explicitly calls the init functions. The linker organized them in a way that the C runtime called them for us Sections and Linker Scripts The process in which constructors are linked isn’t mysterious, though. In fact, compilers allow you to name the location in the binary on most platforms called a “section” you want to place any of your data and/or code. And by extension, and as we saw above, Rust allows this as well. The challenge, as we will see, is making use of this organizational feature. Linkers have been the key to C’s ability to target any form of binary for some time. Most linkers allow for developers to provide linker scripts - text files that live alongside your source code which is compiled to object files and instruct the linker on how those object files are assembled. Using a linker script, a single C file might become a Linux executable, or a block of raw assembly that lives in the boot sector of a hard drive. Linker scripts also allow for defining virtual symbols - that is, symbols that don’t exist in any source file but can be used by C code to access pointers to the underlying data in the loaded binary. Linker scripts are a complex topic and beyond the scope of this post, but we can easily find examples https://wiki.osdev.org/Linker Scripts of them in the wild: // Adapted from https://wiki.osdev.org/Linker Scripts SECTIONS { .text.start KERNEL BASE : { startup.o .text } .text : ALIGN CONSTANT MAXPAGESIZE { TEXT START = .; .text TEXT END = .; } .data : ALIGN CONSTANT MAXPAGESIZE { DATA START = .; .data DATA END = .; } } In the above example, the virtual symbols TEXT START and TEXT END are explicitly defined to point to the beginning and end of the .text section, respectively. The period in TEXT START = .; is a special syntax that refers to a location counter https://sourceware.org/binutils/docs/ld/Location-Counter.html that resolves roughly to the current output address in the binary. Linker Symbols This trips up most developers that encounter it for the first time, but the linker is setting the address of the start and end symbols , and therefore where the static with the same name is placed, and not setting the value of symbols that are pointers. That is to say: the start and stop symbols aren’t a const Type . The start and stop symbols carry no data themselves and are used for their addresses only The section consists of the range of data between the start inclusive and stop exclusive symbols. | Section | Static | Value | Linker symbol s | | |---|---|---|---|---| my numbers | DATA 1 | 11 | ⎫ ⎬ ⎭ | DATA 1, start my numbers | DATA 2 | 22 | DATA 2 | || DATA 3 | 33 | DATA 3 | || DATA 4 | 44 | DATA 4 | || past the end | ↤ | stop my numbers | Specifying start and end symbols for every section can be complex and tedious in linker scripts, so many linkers 9 eventually gained a feature where they could automatically define symbols bounding all sections in the executable. E.g., for GNU toolchains, a section named MY SECTION will automatically have symbols start MY SECTION and stop MY SECTION defined. macOS has a similar pattern https://discourse.llvm.org/t/lld-support-for-ld64-mach-o-linker-synthesised-symbols/45145 where it synthesizes a section$start and section$end symbol for each section.In the GNU linker, those sections not explicitly defined in the linker script are called “orphan sections” 10. One important thing to note: if and only if a section’s name is compatible with a C symbol name, the linker will automatically define a start - and stop -prefixed symbol for the section. In the example you’ll see below, the section name our strings that we used works, but if we had chosen our.strings or .our strings it would not have You’ll see in the example below that the start and stop symbols are MaybeUninit< . The boundary symbols contain no data, and only their address is significant. The ideal Rust type for these would be an “opaque external type” this would be implemented by the extern types feature https://doc.rust-lang.org/beta/unstable-book/language-features/extern-types.html . As these are not currently implemented in Stable Rust, MaybeUninit is a stand-in. It signifies to the compiler that the data is uninitialized, and generally not safe to read via reference. Since taking a to a https://blog.rust-lang.org/2024/10/17/Rust-1.82.0/ native-syntax-for-creating-a-raw-pointer &raw const pointer static item is always valid, however, we can still safely capture its address without ever reading its value. Try it in the Rust Playground https://play.rust-lang.org/?version=stable&mode=debug&edition=2024&gist=1696bdc67f02992cfde9de752e117e0e : use std::mem::MaybeUninit; used unsafe link section = "our strings" static FIRST STRING: &'static str = "Hello, "; used unsafe link section = "our strings" static SECOND STRING: &'static str = "world "; // Note: these are not pointers. Instead, the linker has placed // the boundary symbols STATIC STRING START and STATIC STRING END at // the start and end of the section unsafe extern "C" { link name = " start our strings" static STATIC STRING START: MaybeUninit< ; link name = " stop our strings" static STATIC STRING END: MaybeUninit< ; } fn main { let strings: &'static &'static str = unsafe { // SAFETY: get the addresses of the start and end symbols without // reading them. let start = &raw const STATIC STRING START as const &'static str; let end = &raw const STATIC STRING END as const &'static str; std::slice::from raw parts start, end.offset from start as usize }; // "Hello, world " println "String: {}", strings.join "" ; } The link-section https://crates.io/crates/link-section crate is designed to abstract away the details of these linker sections and convert them into traditional Rust slices with all standard slice operations available. We can use it to simplify the example above to: use link section::{in section, section}; section typed static OUR STRINGS: link section::TypedSection<&'static str ; in section OUR STRINGS static FIRST STRING: &'static str = "Hello, "; in section OUR STRINGS static SECOND STRING: &'static str = "world "; fn main { println "String: {}", OUR STRINGS.join "" ; } In these examples we’re submitting items to the link section in a single module within a single crate, but that’s not a requirement. In fact, the power of link sections is that you can submit items to a link section from any crate that contributes code to a binary - the linker will gather them all together just before writing the final binary. Dependency Injection The registration pattern we’re about to build is Dependency Injection https://en.wikipedia.org/wiki/Dependency injection by another name. This is a well-known pattern: frameworks like Dagger https://dagger.dev/ and Spring https://spring.io/ are built on the same principle that consumers of registration data should not be coupled to the providers of that data. A provider registers data at its definition site, a consumer simply reads the registry. What’s somewhat different with linker sections versus traditional DI is that in DI the framework often needs to walk the module graph or scan loaded classes at startup to discover both providers and consumer sites. With linker sections, this magic is handled when the binary is written. The linker is the one that gathers all of the provider data and makes it trivially available to the consumer. The example below uses a link section::section to register CLI subcommands and is an instance of this pattern. More complex projects like Turbopack https://github.com/vercel/next.js/blob/canary/turbopack/ use this pattern to register string-pool constants, and the registration machinery used for serialization/deserialization and turbotask incremental compilation functions https://web.archive.org/web/20250222021941/https://turbo.build/pack/docs/incremental-computation . A hypothetical webserver could make use of this pattern to register routes and middleware that are automatically collected at build time. The core mechanism is the same: the contributors place data into a shared registration system from any crate in the dependency tree, and the consumer reads the collected data without having to know where it was provided from. Using Sections for Registration One advantage we have in doing work before main is that it is well-behaved. No threads are running unless we start them. This means we are able to avoid the complexity of locks and other synchronization primitives in many cases, and that we can explicitly split our writable and immutable phase of our data’s lifecycle clearly: before and after main. And because of that, accessing data in the running program can become both simpler and more efficient by avoiding the need to acquire and release locks. First, we’ll define our subcommand, a const constructor function, and a section to collect them: use std::collections::VecDeque; use std::path::Path; use link section::{in section, section}; struct CliSubcommand { is default: bool, name: &'static str, description: &'static str, f: fn &Path, & String , } impl CliSubcommand { const fn new name: &'static str, description: &'static str, f: fn &Path, & String - Self { Self { is default: false, name, description, f } } const fn new default name: &'static str, description: &'static str, f: fn &Path, & String - Self { Self { is default: true, name, description, f } } } section typed static CLI SUBCOMMANDS: link section::TypedSection<CliSubcommand ; Then we’ll register subcommands - these can live anywhere in your code: mod list { in section CLI SUBCOMMANDS static CLI SUBCOMMAND LIST: CliSubcommand = CliSubcommand::new "list", "List all items", | exe, args| { println "Listing all items" ; } ; } mod add { in section CLI SUBCOMMANDS static CLI SUBCOMMAND ADD: CliSubcommand = CliSubcommand::new "add", "Add a new item", | exe, args| { println "Adding a new item" ; } ; } mod help { in section CLI SUBCOMMANDS static CLI SUBCOMMAND HELP: CliSubcommand = CliSubcommand::new default "help", "Show help", |exe, args| { println "Usage: {} <subcommand options ", exe.display ; println ; println "Subcommands:" ; for subcommand in CLI SUBCOMMANDS { println " {}: {}", subcommand.name, subcommand.description ; } } ; } And then in our main function we can dynamically dispatch to any registered subcommand without ever having to know what they are or where they live. It only needs to be able to see the CLI SUBCOMMANDS section definition: js fn main { let mut args: VecDeque<String = std::env::args .collect ; let exe = args.pop front .expect "No executable name provided" ; let exe = Path::new &exe ; let subcommand name = args.pop front .unwrap or default ; let rest: Vec<String = args.into ; // Try to find the subcommand by name for cmd in CLI SUBCOMMANDS { if cmd.name == subcommand name { cmd.f exe, &rest ; return; } } // If no subcommand was found, fall back to the default subcommand for cmd in CLI SUBCOMMANDS { if cmd.is default { cmd.f exe, &rest ; return; } } } Running the code above works as you’d expect: $ ./cli Usage: ./cli <subcommand options Subcommands: list: List all items add: Add a new item help: Show help $ ./cli list Listing all items Beyond Immutable Data This section deals with some more advanced topics. Familiarity with Rust Atomics and Locks https://marabos.nl/atomics/ , or at least reading the first chapter on the basics of Rust concurrency, is recommended The example above assumes that the linked data is immutable. But that’s only half the power of using linker organization for data. Mutability in global static data is a common problem with well-known solutions https://doc.rust-lang.org/reference/interior-mutability.html in standard Rust. We could potentially use Rust’s built-in tools for interior mutability like mutexes, or atomic types, for example. Each of those comes with some runtime cost. If they are “uncontended” they aren’t expensive, but they are not necessarily free. 11 fn:7 But what if we want to minimize the overhead of runtime data access? Immutable data is trivial: Rust allows safe concurrent access to immutable data by default 12. Rust has strict requirements for mutable data, however. There are two requirements to safely mutate data: 1 the modifications must be done in a thread-safe manner, and 2 , there must never be more than one reference to the data if a mutable reference exists. At the beginning of this post, we mentioned that life-before-main is a useful place to bootstrap because no threads are running unless we start them. And the solution to 1 is trivial if the data is currently accessible to a single thread only We don’t need to do anything atomically; we only need to ensure that all of the changes we make to that data “happen before” https://marabos.nl/atomics/ any reads to the data. In a single-threaded environment, “happens before” is automatic 13. This means that we can mutate data in a link-section before main and it will be safe to access, lock free, from any thread after main. The resolution for 2 is similar: as long as we only ever take a mutable reference and only a mutable reference before main , there will never be more than one reference to the data when a mutable reference exists. The pre-main environment satisfies both 1 and 2 , without needing to reach for locks or other synchronization primitives. There’s also one additional gotcha with linker sections we need to be very careful with: the slice that contains all of the items in a section is an alias to the static item that lives in the section. The rules about aliasing apply to both the slice and the static item, and you must ensure that static items are placed in UnsafeCell to safely mutate them from the slice 14. Rust does not allow a static item to be modified through other means. With static items that aren’t wrapped in an UnsafeCell , LLVM may consider itself free to cache, reorder or otherwise make assumptions about the data itself. UnsafeCell itself is not Sync , so you’ll need to add your own wrapper types on top of this Note that in the example below, we’re now using MaybeUninit<SyncUnsafeCell<... for the boundary symbols and SyncUnsafeCell<... for the items. Because we’re planning on sorting the slice, we need to tell Rust that the slice items are not immutable so the data doesn’t end up in read-only memory. By using a type that includes UnsafeCell - a semantic signal that Rust uses to indicate interior mutability https://doc.rust-lang.org/reference/interior-mutability.html - the Rust compiler will then know to place that data in a part of the binary that can be written. On some platforms Windows in particular , omitting this from the data items will cause segmentation faults when trying to sort the slice. On other platforms AIX for example , section mutability is part of a section’s identifier, so the boundary symbols’ mutability needs to match the section’s mutability Let’s walk through an example of how we might otherwise do something like this. We’re going to build a string interning pool defined entirely at link-time, and add a wrinkle: we want to be able to sort the slice of interned strings at runtime so we can quickly intern a string by value if needed via binary search try it in the Rust Playground https://play.rust-lang.org/?version=stable&mode=debug&edition=2024&gist=cc4df8dc5aba46357a5f6c64ce11e587 : use std::cell::UnsafeCell; use std::mem::MaybeUninit; cfg debug assertions use std::sync::atomic::{AtomicBool, Ordering}; /// Nightly Rust offers a built-in SyncUnsafeCell . This is a minimal /// reimplementation of that: /// <https://doc.rust-lang.org/std/cell/struct.SyncUnsafeCell.html repr transparent struct SyncUnsafeCell<T: ?Sized UnsafeCell<T ; // SAFETY: safety burden of UnsafeCell is placed entirely on the user unsafe impl<T: ?Sized + Sync Sync for SyncUnsafeCell<T {} macro rules intern string { $name:ident, $string:literal = { allow unused const $name: &'static str = const { // This is not a common pattern, but it's entirely valid // to nest static items inside of const blocks. // You can think of this as a way to hide the symbols // in a completely anonymous namespace. const VALUE: &str = $string; // Safety note: this static must never be used. This is // purely a submission to the linker and any access to it // may be UB. used unsafe link section = "our strings" static ITEM: SyncUnsafeCell<&'static str = SyncUnsafeCell UnsafeCell::new VALUE ; VALUE }; }; } intern string WORLD, "world" ; intern string EXCLAMATION, " " ; intern string HELLO, "hello" ; intern string FROM, "from" ; intern string RUST, "Rust" ; unsafe extern "C" { link name = " start our strings" static STATIC STRING START: MaybeUninit<SyncUnsafeCell< ; link name = " stop our strings" static STATIC STRING END: MaybeUninit<SyncUnsafeCell< ; } /// Debug check to make sure the slice is sorted once and only once. This could /// be enabled in release mode without any major performance impact, but we have enough /// guarantees in place. Note that atomic access does establish some memory ordering /// guarantees, but the soundness guarantees are upheld with or without this atomic check. cfg debug assertions static SLICE IS SORTED: AtomicBool = AtomicBool::new false ; // Implementation note: this function must not be called before SORT STRINGS CTOR has // run. fn interned strings - &'static &'static str { // We use Acquire/Release pairing as a double-initialization check cfg debug assertions debug assert SLICE IS SORTED.load Ordering::Acquire , "Oh no Slice was not sorted " ; // SAFETY: we are calling this after main and we can guarantee that no // mutable reference is still alive. Since we know that no other code // is running before main, and that SORT STRINGS CTOR will run before main, // we can guarantee creating these slices is safe as 1 the sort "happens-before" // any access and 2 the mutable reference has been closed before any read-reference // access satisfying aliasing XOR mutability requirement . let strings: &'static &'static str = unsafe { let start = &raw const STATIC STRING START as const &'static str; let end = &raw const STATIC STRING END as const &'static str; std::slice::from raw parts start, end.offset from start as usize }; strings } // Implementation note: this function assumes the slice has been sorted. See // the guarantee above on interned strings for reasoning. fn maybe intern string s: impl AsRef<str - Option<&'static str { let s = s.as ref ; let strings = interned strings ; strings.binary search &s .ok .map |index| strings index } // SAFETY: We use the reserved .init array.0 priority because we do not // access any C runtime functions sort unstable does not allocate and we // want to run before all other code. .init array.101 would work in our // case, but this prevents other early-init code from accidentally running // in the wrong order. SLICE IS SORTED is a debug check to make sure that // doesn't happen. Note that all early-init code is tagged with unsafe so // it always needs to be aware of safety guarantees of all APIs it touches. used unsafe link section = ".init array.0" static SORT STRINGS CTOR: extern "C" fn = const { extern "C" fn sort strings { // We use Acquire/Release pairing as a double-initialization check cfg debug assertions debug assert SLICE IS SORTED.load Ordering::Acquire , "Oh no Sorted twice? ?" ; // SAFETY: we are calling this before main and we can guarantee that no // reference from interned strings exists yet because we know no other // threads will be running, and we're not calling interned strings yet. let strings: &mut &'static str = unsafe { // SAFETY: the bounds markers are not mutable, but we can safely // cast them to mutable pointers because we know the data behind // them is stored within UnsafeCell which is Rust's way of // giving us interior mutability. let start = &raw const STATIC STRING START as mut &'static str; let end = &raw const STATIC STRING END as mut &'static str; std::slice::from raw parts mut start, end.offset from start as usize }; strings.sort unstable ; cfg debug assertions SLICE IS SORTED.store true, Ordering::Release ; } sort strings }; fn main { for i, s in interned strings .iter .enumerate { println " {i} : {s}" ; } println "{}, {}{}", HELLO, WORLD, EXCLAMATION ; println "{}, {}{}", maybe intern string "hello" .unwrap , maybe intern string "world" .unwrap , maybe intern string " " .unwrap ; } The above example is pretty heavy and a bit thicker thanks to the generous commentary , but it’s a good example of how much boilerplate crates like ctor and link-section can save you. The equivalent using those crates can make use of the TypedMutableSection and a ctor to ensure the items are sorted before main. Note that the requirements for TypedMutableSection are that the items must be const - the reason is that the mutable section uses a similar style of code to the manually-implemented example above. // String interning pool using ctor and link-section . use ctor::ctor; use link section::{in section, section}; section mutable static INTERNED STRINGS: link section::TypedMutableSection<&'static str ; in section INTERNED STRINGS const WORLD: &'static str = "world"; in section INTERNED STRINGS const EXCLAMATION: &'static str = " "; in section INTERNED STRINGS const HELLO: &'static str = "hello"; in section INTERNED STRINGS const FROM: &'static str = "from"; in section INTERNED STRINGS const RUST: &'static str = "Rust"; ctor unsafe fn sort strings { let strings: &mut &'static str = unsafe { INTERNED STRINGS.as mut slice }; strings.sort unstable ; } fn maybe intern string s: impl AsRef<str - Option<&'static str { let s = s.as ref ; let strings = INTERNED STRINGS.as slice ; strings.binary search &s .ok .map |index| strings index } fn main { for i, s in INTERNED STRINGS.iter .enumerate { println " {i} : {s}" ; } println "{}, {}{}", HELLO, WORLD, EXCLAMATION ; println "{}, {}{}", maybe intern string "hello" .unwrap , maybe intern string "world" .unwrap , maybe intern string " " .unwrap ; } This particular example isn’t impossible without link sections, of course. What we get from the patterns we’ve discussed in this post are three things: 1 the guaranteed aggregation of tagged items, with all data pre-allocated and contiguous in memory; 2 the ability to distribute registrations anywhere in the code; and 3 a guaranteed count of the items in the section. One major benefit that falls out of the three advantages above is that link sections require no allocations. If we were to rewrite this without link sections we’d be allocating a HashMap , Vec or other data structure, and potentially resizing it a number of times as we gather items because we don’t actually know how many items we’ll have until runtime . The second major benefit that falls out is the Inversion of Control . The dependency graph for a traditional “collection” approach looks like below, with shared types deeply nested in the dependency graph, modules depending on that shared types module, and then a collector module that depends on all those modules to collect their types: The change doesn’t seem large, but there is a large impact: the collector can now live anywhere and no longer needs to care what modules are contributing data: And of course, we aren’t just limited to slices: you’ll find that there are analogues to many data-structures with link-time support in the scattered-collect crate https://docs.rs/scattered-collect/latest/scattered collect/index.html : Scattered Slice : Various Vec -like structures that provide slices and optionally sorting . ScatteredMap / ScatteredSet : An analogue to HashMap / HashSet that provides hashed key-to-value lookup with some minimal pre-main initialization. But Seriously: When Not to Use This Link-time computation is fun and powerful, and it’s not always the right tool for the job. There is often a non-link-time equivalent: manually collecting data in crates that have visibility into each crate that wishes to contribute data. This can be inconvenient at times - instead of the contributors seeing a single contribution point “upstream” in a core crate, a “collector” crate with lots of crate references is required to collect them all. Dead-code elimination becomes challenging: the link-section crate and the linkme equivalent both decorate all items using used , so the linker is disallowed from pruning unused data. Figuring out how to make link-time collection and dead-code elimination work well together is a complex problem beyond the scope of this post. For smaller bits of data like interned string atoms this might not be a problem, but if a program wants to intern larger chunks of data like chunks of raw JSON/JavaScript, or extensive data structures, this may add up to a lot of dead-code that may be difficult to identify. Pre-main constructor functions have limitations https://docs.rs/ctor/latest/ctor/life before main/index.html : they cannot panic 15, Rust does not guarantee that all stdlib functions are available, and the order that the initialization functions are called within a given priority level is not guaranteed and highly platform-dependent. With careful planning, these limitations may be worked around but life-before-main may not be correct for subtle and difficult-to-debug reasons. At this time, Miri https://github.com/rust-lang/miri/ is not fully compatible with all pre-main constructors and link-section constructions: it has a very basic view of pre-main execution, and does not model link sections at all. This may improve over time, but as of the time of writing, LLVM sanitizers ASan https://clang.llvm.org/docs/AddressSanitizer.html , TSan https://clang.llvm.org/docs/ThreadSanitizer.html , and others are the recommended way to test for undefined behaviour. The Inversion of Control pattern also has a cost: it makes it potentially harder to audit all the places that contribute data to a link section. In reality, many widely-deployed and heavily-used Rust programs already rely on pre-main functionality: the ctor , link-section , inventory and linkme crates are used by many downstream crates today. Briefly, on WASM The examples above omitted a fairly important platform, though for good reason. WASM does not currently support linker sections natively because of an inconvenient choice many years back 51088 https://github.com/rust-lang/rust/issues/51088 and 52353 https://github.com/rust-lang/rust/pull/52353 for more details . Instead of allowing the link section annotation to place items in a true code section, the items are placed in a WASM custom section which is inaccessible to the WASM code itself The linktime crates do support WASM and have an emulation workaround https://docs.rs/crate/link-section/latest :~:text=On%20WASM%20platforms that makes the approaches work for WASM binaries, but the author of this post hopes to make a suggestion in the near future on how proper WASM support could be added Conclusions You can do a lot before main, and the benefits of doing so are pretty significant for certain cases. It’s a highly-ordered, highly-controllable environment that lets you more confidently do a lot of work without locks, atomics and other synchronization primitives. Link sections give you arbitrary aggregation and co-location of related data across your whole binary without awkward crate dependency order. In a lot of cases you can even avoid allocations completely which helps keep you away from one of the worst allocator sins: churn of allocations leading to fragmentation. For further reading, check out the various crates discussed in this post: : Module initialization functions that run before ctor main : Not discussed in this post, but the shutdown analogue to dtor ctor : Linker-managed typed slices and untyped sections, with mutability support. link-section : Linker-managed higher-level collections: slices, sorted slices, maps scattered-collect Thanks Thanks to my lovely wife Mia https://github.com/ojeda-e , Benjamin Woodruff https://x.com/ bgwoodruff and Luke Sandberg https://x.com/lukeisandberg , as well as @ssokolow https://lobste.rs/s/rs1t8s/there is life before main rust c qoljqy for their feedback and review. This post would not have been possible without their help. - Go is a notable exception in that it avoids the C runtime on some platforms, only using libc as an ABI stability boundary on platforms that require it e.g., Apple’s libSystem.dylib and OpenBSD’s libc . ↩ fnref:0 - On Windows, these are DOS-style which were in turn derived from CP/M-style arguments. ↩ fnref:0b - Before the loader runs, the program is just some bytes on a disk and the loader which can be the kernel itself or a user-space system component like ld.so on Linux maps those bytes into memory and hands off control. ↩ fnref:1 - The most popular method… in the humble author’s opinion. ↩ fnref:2 - macOS does not support this. The C runtime does its own initialization and then just runs every user constructor function in the order the linker saw them. ↩ fnref:3 - AIX has a special symbol naming convention for constructor functions: the sinit prefix, followed by a hexadecimal priority value. ↩ fnref:4 - This will be discussed later in the post, but macOS synthesizes a section$start and section$end symbol for each section instead of a start and stop symbol. ↩ fnref:c0 - linkme creates distributed slices, but does not currently support WASM, and does not support mutable section data required to sort a section. inventory supports WASM, but requires a ctor -like function per item in the section. ↩ fnref:c1 - The Windows linker does not support this feature, but instead defines an overall sort order for https://devblogs.microsoft.com/oldnewthing/20181107-00/?p=100155 that is effectively equivalent. symbols ↩ fnref:5 - Orphan sections have a complicated algorithm for placement https://maskray.me/blog/2024-06-02-understanding-orphan-sections . ↩ fnref:6 - For example, an atomic value must always be re-read, and that may incur use of CPU cache which is pretty numerous these days, but definitely not infinite. ↩ fnref:7 - As long as it’s https://doc.rust-lang.org/book/ch16-04-extensible-concurrency-sync-and-send.html . Sync Sync means that it’s safe to share a reference to the data between threads. ↩ fnref:8 - This is a complex topic, and Rust Atomics and Locks https://marabos.nl/atomics/ is the best resource for learning more. Starting a new thread means that all the previous writes “happen before” anything on the new thread, but we’ll leave the proof of this for the reader or possibly a future post . ↩ fnref:9 - Even taking a reference to a mutable static is disallowed by default in Rust 2024 https://doc.rust-lang.org/edition-guide/rust-2024/static-mut-references.html ↩ fnref:10 - Well, more accurately they can panic but they shouldn’t panic. In fact you’ll get a double panic: thread '<unnamed ' panicked at ...: pre-main panic and then immediately thread '<unnamed ' panicked at ...: panic in a function that cannot unwind . ↩ fnref:11