bootstrap/

lib.rs

//! Implementation of bootstrap, the Rust build system.
//!
//! This module, and its descendants, are the implementation of the Rust build
//! system. Most of this build system is backed by Cargo but the outer layer
//! here serves as the ability to orchestrate calling Cargo, sequencing Cargo
//! builds, building artifacts like LLVM, etc. The goals of bootstrap are:
//!
//! * To be an easily understandable, easily extensible, and maintainable build
//!   system.
//! * Leverage standard tools in the Rust ecosystem to build the compiler, aka
//!   crates.io and Cargo.
//! * A standard interface to build across all platforms, including MSVC
//!
//! ## Further information
//!
//! More documentation can be found in each respective module below, and you can
//! also check out the `src/bootstrap/README.md` file for more information.
#![cfg_attr(test, allow(unused))]

use std::cell::Cell;
use std::collections::{BTreeSet, HashMap, HashSet};
use std::fmt::Display;
use std::path::{Path, PathBuf};
use std::sync::OnceLock;
use std::time::{Instant, SystemTime};
use std::{env, fs, io, str};

use build_helper::ci::gha;
use build_helper::exit;
use cc::Tool;
use termcolor::{ColorChoice, StandardStream, WriteColor};
use utils::build_stamp::BuildStamp;
use utils::channel::GitInfo;
use utils::exec::ExecutionContext;

use crate::core::builder;
use crate::core::builder::Kind;
use crate::core::config::{BootstrapOverrideLld, DryRun, LlvmLibunwind, TargetSelection, flags};
use crate::utils::exec::{BootstrapCommand, command};
use crate::utils::helpers::{self, dir_is_empty, exe, libdir, set_file_times, split_debuginfo};

mod core;
mod utils;

pub use core::builder::PathSet;
#[cfg(feature = "tracing")]
pub use core::builder::STEP_SPAN_TARGET;
pub use core::config::flags::{Flags, Subcommand};
pub use core::config::{ChangeId, Config};

#[cfg(feature = "tracing")]
use tracing::{instrument, span};
pub use utils::change_tracker::{
    CONFIG_CHANGE_HISTORY, find_recent_config_change_ids, human_readable_changes,
};
pub use utils::helpers::{PanicTracker, symlink_dir};
#[cfg(feature = "tracing")]
pub use utils::tracing::setup_tracing;

use crate::core::build_steps::vendor::VENDOR_DIR;

const LLVM_TOOLS: &[&str] = &[
    "llvm-cov",      // used to generate coverage report
    "llvm-nm",       // used to inspect binaries; it shows symbol names, their sizes and visibility
    "llvm-objcopy",  // used to transform ELFs into binary format which flashing tools consume
    "llvm-objdump",  // used to disassemble programs
    "llvm-profdata", // used to inspect and merge files generated by profiles
    "llvm-readobj",  // used to get information from ELFs/objects that the other tools don't provide
    "llvm-size",     // used to prints the size of the linker sections of a program
    "llvm-strip",    // used to discard symbols from binary files to reduce their size
    "llvm-ar",       // used for creating and modifying archive files
    "llvm-as",       // used to convert LLVM assembly to LLVM bitcode
    "llvm-dis",      // used to disassemble LLVM bitcode
    "llvm-link",     // Used to link LLVM bitcode
    "llc",           // used to compile LLVM bytecode
    "opt",           // used to optimize LLVM bytecode
];

/// LLD file names for all flavors.
const LLD_FILE_NAMES: &[&str] = &["ld.lld", "ld64.lld", "lld-link", "wasm-ld"];

/// Extra `--check-cfg` to add when building the compiler or tools
/// (Mode restriction, config name, config values (if any))
#[expect(clippy::type_complexity)] // It's fine for hard-coded list and type is explained above.
const EXTRA_CHECK_CFGS: &[(Option<Mode>, &str, Option<&[&'static str]>)] = &[
    (Some(Mode::Rustc), "bootstrap", None),
    (Some(Mode::Codegen), "bootstrap", None),
    (Some(Mode::ToolRustcPrivate), "bootstrap", None),
    (Some(Mode::ToolStd), "bootstrap", None),
    (Some(Mode::ToolRustcPrivate), "rust_analyzer", None),
    (Some(Mode::ToolStd), "rust_analyzer", None),
    // Any library specific cfgs like `target_os`, `target_arch` should be put in
    // priority the `[lints.rust.unexpected_cfgs.check-cfg]` table
    // in the appropriate `library/{std,alloc,core}/Cargo.toml`
];

/// A structure representing a Rust compiler.
///
/// Each compiler has a `stage` that it is associated with and a `host` that
/// corresponds to the platform the compiler runs on. This structure is used as
/// a parameter to many methods below.
#[derive(Eq, PartialOrd, Ord, Clone, Copy, Debug)]
pub struct Compiler {
    stage: u32,
    host: TargetSelection,
    /// Indicates whether the compiler was forced to use a specific stage.
    /// This field is ignored in `Hash` and `PartialEq` implementations as only the `stage`
    /// and `host` fields are relevant for those.
    forced_compiler: bool,
}

impl std::hash::Hash for Compiler {
    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
        self.stage.hash(state);
        self.host.hash(state);
    }
}

impl PartialEq for Compiler {
    fn eq(&self, other: &Self) -> bool {
        self.stage == other.stage && self.host == other.host
    }
}

/// Represents a codegen backend.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Default)]
pub enum CodegenBackendKind {
    #[default]
    Llvm,
    Cranelift,
    Gcc,
    Custom(String),
}

impl CodegenBackendKind {
    /// Name of the codegen backend, as identified in the `compiler` directory
    /// (`rustc_codegen_<name>`).
    pub fn name(&self) -> &str {
        match self {
            CodegenBackendKind::Llvm => "llvm",
            CodegenBackendKind::Cranelift => "cranelift",
            CodegenBackendKind::Gcc => "gcc",
            CodegenBackendKind::Custom(name) => name,
        }
    }

    /// Name of the codegen backend's crate, e.g. `rustc_codegen_cranelift`.
    pub fn crate_name(&self) -> String {
        format!("rustc_codegen_{}", self.name())
    }

    pub fn is_llvm(&self) -> bool {
        matches!(self, Self::Llvm)
    }

    pub fn is_cranelift(&self) -> bool {
        matches!(self, Self::Cranelift)
    }

    pub fn is_gcc(&self) -> bool {
        matches!(self, Self::Gcc)
    }
}

impl std::str::FromStr for CodegenBackendKind {
    type Err = &'static str;

    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s.to_lowercase().as_str() {
            "" => Err("Invalid empty backend name"),
            "gcc" => Ok(Self::Gcc),
            "llvm" => Ok(Self::Llvm),
            "cranelift" => Ok(Self::Cranelift),
            _ => Ok(Self::Custom(s.to_string())),
        }
    }
}

#[derive(PartialEq, Eq, Copy, Clone, Debug)]
pub enum TestTarget {
    /// Run unit, integration and doc tests (default).
    Default,
    /// Run unit, integration, doc tests, examples, bins, benchmarks (no doc tests).
    AllTargets,
    /// Only run doc tests.
    DocOnly,
    /// Only run unit and integration tests.
    Tests,
}

impl TestTarget {
    fn runs_doctests(&self) -> bool {
        matches!(self, TestTarget::DocOnly | TestTarget::Default)
    }
}

pub enum GitRepo {
    Rustc,
    Llvm,
}

/// Global configuration for the build system.
///
/// This structure transitively contains all configuration for the build system.
/// All filesystem-encoded configuration is in `config`, all flags are in
/// `flags`, and then parsed or probed information is listed in the keys below.
///
/// This structure is a parameter of almost all methods in the build system,
/// although most functions are implemented as free functions rather than
/// methods specifically on this structure itself (to make it easier to
/// organize).
pub struct Build {
    /// User-specified configuration from `bootstrap.toml`.
    config: Config,

    // Version information
    version: String,

    // Properties derived from the above configuration
    src: PathBuf,
    out: PathBuf,
    bootstrap_out: PathBuf,
    cargo_info: GitInfo,
    rust_analyzer_info: GitInfo,
    clippy_info: GitInfo,
    miri_info: GitInfo,
    rustfmt_info: GitInfo,
    enzyme_info: GitInfo,
    in_tree_llvm_info: GitInfo,
    in_tree_gcc_info: GitInfo,
    local_rebuild: bool,
    fail_fast: bool,
    test_target: TestTarget,
    verbosity: usize,

    /// Build triple for the pre-compiled snapshot compiler.
    host_target: TargetSelection,
    /// Which triples to produce a compiler toolchain for.
    hosts: Vec<TargetSelection>,
    /// Which triples to build libraries (core/alloc/std/test/proc_macro) for.
    targets: Vec<TargetSelection>,

    initial_rustc: PathBuf,
    initial_rustdoc: PathBuf,
    initial_cargo: PathBuf,
    initial_lld: PathBuf,
    initial_relative_libdir: PathBuf,
    initial_sysroot: PathBuf,

    // Runtime state filled in later on
    // C/C++ compilers and archiver for all targets
    cc: HashMap<TargetSelection, cc::Tool>,
    cxx: HashMap<TargetSelection, cc::Tool>,
    ar: HashMap<TargetSelection, PathBuf>,
    ranlib: HashMap<TargetSelection, PathBuf>,
    wasi_sdk_path: Option<PathBuf>,

    // Miscellaneous
    // allow bidirectional lookups: both name -> path and path -> name
    crates: HashMap<String, Crate>,
    crate_paths: HashMap<PathBuf, String>,
    is_sudo: bool,
    prerelease_version: Cell<Option<u32>>,

    #[cfg(feature = "build-metrics")]
    metrics: crate::utils::metrics::BuildMetrics,

    #[cfg(feature = "tracing")]
    step_graph: std::cell::RefCell<crate::utils::step_graph::StepGraph>,
}

#[derive(Debug, Clone)]
struct Crate {
    name: String,
    deps: HashSet<String>,
    path: PathBuf,
    features: Vec<String>,
}

impl Crate {
    fn local_path(&self, build: &Build) -> PathBuf {
        self.path.strip_prefix(&build.config.src).unwrap().into()
    }
}

/// When building Rust various objects are handled differently.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
pub enum DependencyType {
    /// Libraries originating from proc-macros.
    Host,
    /// Typical Rust libraries.
    Target,
    /// Non Rust libraries and objects shipped to ease usage of certain targets.
    TargetSelfContained,
}

/// The various "modes" of invoking Cargo.
///
/// These entries currently correspond to the various output directories of the
/// build system, with each mod generating output in a different directory.
#[derive(Debug, Hash, Clone, Copy, PartialEq, Eq)]
pub enum Mode {
    /// Build the standard library, placing output in the "stageN-std" directory.
    Std,

    /// Build librustc, and compiler libraries, placing output in the "stageN-rustc" directory.
    Rustc,

    /// Build a codegen backend for rustc, placing the output in the "stageN-codegen" directory.
    Codegen,

    /// Build a tool, placing output in the "bootstrap-tools"
    /// directory. This is for miscellaneous sets of tools that extend
    /// bootstrap.
    ///
    /// These tools are intended to be only executed on the host system that
    /// invokes bootstrap, and they thus cannot be cross-compiled.
    ///
    /// They are always built using the stage0 compiler, and they
    /// can be compiled with stable Rust.
    ///
    /// These tools also essentially do not participate in staging.
    ToolBootstrap,

    /// Build a cross-compilable helper tool. These tools do not depend on unstable features or
    /// compiler internals, but they might be cross-compilable (so we cannot build them using the
    /// stage0 compiler, unlike `ToolBootstrap`).
    ///
    /// Some of these tools are also shipped in our `dist` archives.
    /// While we could compile them using the stage0 compiler when not cross-compiling, we instead
    /// use the in-tree compiler (and std) to build them, so that we can ship e.g. std security
    /// fixes and avoid depending fully on stage0 for the artifacts that we ship.
    ///
    /// This mode is used e.g. for linkers and linker tools invoked by rustc on its host target.
    ToolTarget,

    /// Build a tool which uses the locally built std, placing output in the
    /// "stageN-tools" directory. Its usage is quite rare; historically it was
    /// needed by compiletest, but now it is mainly used by `test-float-parse`.
    ToolStd,

    /// Build a tool which uses the `rustc_private` mechanism, and thus
    /// the locally built rustc rlib artifacts,
    /// placing the output in the "stageN-tools" directory. This is used for
    /// everything that links to rustc as a library, such as rustdoc, clippy,
    /// rustfmt, miri, etc.
    ToolRustcPrivate,
}

impl Mode {
    pub fn must_support_dlopen(&self) -> bool {
        match self {
            Mode::Std | Mode::Codegen => true,
            Mode::ToolBootstrap
            | Mode::ToolRustcPrivate
            | Mode::ToolStd
            | Mode::ToolTarget
            | Mode::Rustc => false,
        }
    }
}

/// When `rust.rust_remap_debuginfo` is requested, the compiler needs to know how to
/// opportunistically unremap compiler vs non-compiler sources. We use two schemes,
/// [`RemapScheme::Compiler`] and [`RemapScheme::NonCompiler`].
pub enum RemapScheme {
    /// The [`RemapScheme::Compiler`] scheme will remap to `/rustc-dev/{hash}`.
    Compiler,
    /// The [`RemapScheme::NonCompiler`] scheme will remap to `/rustc/{hash}`.
    NonCompiler,
}

#[derive(Debug, Hash, Clone, Copy, PartialEq, Eq)]
pub enum CLang {
    C,
    Cxx,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum FileType {
    /// An executable binary file (like a `.exe`).
    Executable,
    /// A native, binary library file (like a `.so`, `.dll`, `.a`, `.lib` or `.o`).
    NativeLibrary,
    /// An executable (non-binary) script file (like a `.py` or `.sh`).
    Script,
    /// Any other regular file that is non-executable.
    Regular,
}

impl FileType {
    /// Get Unix permissions appropriate for this file type.
    pub fn perms(self) -> u32 {
        match self {
            FileType::Executable | FileType::Script => 0o755,
            FileType::Regular | FileType::NativeLibrary => 0o644,
        }
    }

    pub fn could_have_split_debuginfo(self) -> bool {
        match self {
            FileType::Executable | FileType::NativeLibrary => true,
            FileType::Script | FileType::Regular => false,
        }
    }
}

macro_rules! forward {
    ( $( $fn:ident( $($param:ident: $ty:ty),* ) $( -> $ret:ty)? ),+ $(,)? ) => {
        impl Build {
            $( fn $fn(&self, $($param: $ty),* ) $( -> $ret)? {
                self.config.$fn( $($param),* )
            } )+
        }
    }
}

forward! {
    do_if_verbose(f: impl Fn()),
    is_verbose() -> bool,
    create(path: &Path, s: &str),
    remove(f: &Path),
    tempdir() -> PathBuf,
    llvm_link_shared() -> bool,
    download_rustc() -> bool,
}

/// An alternative way of specifying what target and stage is involved in some bootstrap activity.
/// Ideally using a `Compiler` directly should be preferred.
struct TargetAndStage {
    target: TargetSelection,
    stage: u32,
}

impl From<(TargetSelection, u32)> for TargetAndStage {
    fn from((target, stage): (TargetSelection, u32)) -> Self {
        Self { target, stage }
    }
}

impl From<Compiler> for TargetAndStage {
    fn from(compiler: Compiler) -> Self {
        Self { target: compiler.host, stage: compiler.stage }
    }
}

impl Build {
    /// Creates a new set of build configuration from the `flags` on the command
    /// line and the filesystem `config`.
    ///
    /// By default all build output will be placed in the current directory.
    pub fn new(mut config: Config) -> Build {
        let src = config.src.clone();
        let out = config.out.clone();

        #[cfg(unix)]
        // keep this consistent with the equivalent check in x.py:
        // https://github.com/rust-lang/rust/blob/a8a33cf27166d3eabaffc58ed3799e054af3b0c6/src/bootstrap/bootstrap.py#L796-L797
        let is_sudo = match env::var_os("SUDO_USER") {
            Some(_sudo_user) => {
                // SAFETY: getuid() system call is always successful and no return value is reserved
                // to indicate an error.
                //
                // For more context, see https://man7.org/linux/man-pages/man2/geteuid.2.html
                let uid = unsafe { libc::getuid() };
                uid == 0
            }
            None => false,
        };
        #[cfg(not(unix))]
        let is_sudo = false;

        let rust_info = config.rust_info.clone();
        let cargo_info = config.cargo_info.clone();
        let rust_analyzer_info = config.rust_analyzer_info.clone();
        let clippy_info = config.clippy_info.clone();
        let miri_info = config.miri_info.clone();
        let rustfmt_info = config.rustfmt_info.clone();
        let enzyme_info = config.enzyme_info.clone();
        let in_tree_llvm_info = config.in_tree_llvm_info.clone();
        let in_tree_gcc_info = config.in_tree_gcc_info.clone();

        let initial_target_libdir = command(&config.initial_rustc)
            .run_in_dry_run()
            .args(["--print", "target-libdir"])
            .run_capture_stdout(&config)
            .stdout()
            .trim()
            .to_owned();

        let initial_target_dir = Path::new(&initial_target_libdir)
            .parent()
            .unwrap_or_else(|| panic!("{initial_target_libdir} has no parent"));

        let initial_lld = initial_target_dir.join("bin").join("rust-lld");

        let initial_relative_libdir = if cfg!(test) {
            // On tests, bootstrap uses the shim rustc, not the one from the stage0 toolchain.
            PathBuf::default()
        } else {
            let ancestor = initial_target_dir.ancestors().nth(2).unwrap_or_else(|| {
                panic!("Not enough ancestors for {}", initial_target_dir.display())
            });

            ancestor
                .strip_prefix(&config.initial_sysroot)
                .unwrap_or_else(|_| {
                    panic!(
                        "Couldn’t resolve the initial relative libdir from {}",
                        initial_target_dir.display()
                    )
                })
                .to_path_buf()
        };

        let version = std::fs::read_to_string(src.join("src").join("version"))
            .expect("failed to read src/version");
        let version = version.trim();

        let mut bootstrap_out = std::env::current_exe()
            .expect("could not determine path to running process")
            .parent()
            .unwrap()
            .to_path_buf();
        // Since bootstrap is hardlink to deps/bootstrap-*, Solaris can sometimes give
        // path with deps/ which is bad and needs to be avoided.
        if bootstrap_out.ends_with("deps") {
            bootstrap_out.pop();
        }
        if !bootstrap_out.join(exe("rustc", config.host_target)).exists() && !cfg!(test) {
            // this restriction can be lifted whenever https://github.com/rust-lang/rfcs/pull/3028 is implemented
            panic!(
                "`rustc` not found in {}, run `cargo build --bins` before `cargo run`",
                bootstrap_out.display()
            )
        }

        if rust_info.is_from_tarball() && config.description.is_none() {
            config.description = Some("built from a source tarball".to_owned());
        }

        let mut build = Build {
            initial_lld,
            initial_relative_libdir,
            initial_rustc: config.initial_rustc.clone(),
            initial_rustdoc: config.initial_rustdoc.clone(),
            initial_cargo: config.initial_cargo.clone(),
            initial_sysroot: config.initial_sysroot.clone(),
            local_rebuild: config.local_rebuild,
            fail_fast: config.cmd.fail_fast(),
            test_target: config.cmd.test_target(),
            verbosity: config.exec_ctx.verbosity as usize,

            host_target: config.host_target,
            hosts: config.hosts.clone(),
            targets: config.targets.clone(),

            config,
            version: version.to_string(),
            src,
            out,
            bootstrap_out,

            cargo_info,
            rust_analyzer_info,
            clippy_info,
            miri_info,
            rustfmt_info,
            enzyme_info,
            in_tree_llvm_info,
            in_tree_gcc_info,
            cc: HashMap::new(),
            cxx: HashMap::new(),
            ar: HashMap::new(),
            ranlib: HashMap::new(),
            wasi_sdk_path: env::var_os("WASI_SDK_PATH").map(PathBuf::from),
            crates: HashMap::new(),
            crate_paths: HashMap::new(),
            is_sudo,
            prerelease_version: Cell::new(None),

            #[cfg(feature = "build-metrics")]
            metrics: crate::utils::metrics::BuildMetrics::init(),

            #[cfg(feature = "tracing")]
            step_graph: std::cell::RefCell::new(crate::utils::step_graph::StepGraph::default()),
        };

        // If local-rust is the same major.minor as the current version, then force a
        // local-rebuild
        let local_version_verbose = command(&build.initial_rustc)
            .run_in_dry_run()
            .args(["--version", "--verbose"])
            .run_capture_stdout(&build)
            .stdout();
        let local_release = local_version_verbose
            .lines()
            .filter_map(|x| x.strip_prefix("release:"))
            .next()
            .unwrap()
            .trim();
        if local_release.split('.').take(2).eq(version.split('.').take(2)) {
            build.do_if_verbose(|| println!("auto-detected local-rebuild {local_release}"));
            build.local_rebuild = true;
        }

        build.do_if_verbose(|| println!("finding compilers"));
        utils::cc_detect::fill_compilers(&mut build);
        // When running `setup`, the profile is about to change, so any requirements we have now may
        // be different on the next invocation. Don't check for them until the next time x.py is
        // run. This is ok because `setup` never runs any build commands, so it won't fail if commands are missing.
        //
        // Similarly, for `setup` we don't actually need submodules or cargo metadata.
        if !matches!(build.config.cmd, Subcommand::Setup { .. }) {
            build.do_if_verbose(|| println!("running sanity check"));
            crate::core::sanity::check(&mut build);

            // Make sure we update these before gathering metadata so we don't get an error about missing
            // Cargo.toml files.
            let rust_submodules = ["library/backtrace"];
            for s in rust_submodules {
                build.require_submodule(
                    s,
                    Some(
                        "The submodule is required for the standard library \
                         and the main Cargo workspace.",
                    ),
                );
            }
            // Now, update all existing submodules.
            build.update_existing_submodules();

            build.do_if_verbose(|| println!("learning about cargo"));
            crate::core::metadata::build(&mut build);
        }

        // Create symbolic link to use host sysroot from a consistent path (e.g., in the rust-analyzer config file).
        let build_triple = build.out.join(build.host_target);
        t!(fs::create_dir_all(&build_triple));
        let host = build.out.join("host");
        if host.is_symlink() {
            // Left over from a previous build; overwrite it.
            // This matters if `build.build` has changed between invocations.
            #[cfg(windows)]
            t!(fs::remove_dir(&host));
            #[cfg(not(windows))]
            t!(fs::remove_file(&host));
        }
        t!(
            symlink_dir(&build.config, &build_triple, &host),
            format!("symlink_dir({} => {}) failed", host.display(), build_triple.display())
        );

        build
    }

    /// Updates a submodule, and exits with a failure if submodule management
    /// is disabled and the submodule does not exist.
    ///
    /// The given submodule name should be its path relative to the root of
    /// the main repository.
    ///
    /// The given `err_hint` will be shown to the user if the submodule is not
    /// checked out and submodule management is disabled.
    #[cfg_attr(
        feature = "tracing",
        instrument(
            level = "trace",
            name = "Build::require_submodule",
            skip_all,
            fields(submodule = submodule),
        ),
    )]
    pub fn require_submodule(&self, submodule: &str, err_hint: Option<&str>) {
        if self.rust_info().is_from_tarball() {
            return;
        }

        if self.config.dry_run() {
            return;
        }

        // When testing bootstrap itself, it is much faster to ignore
        // submodules. Almost all Steps work fine without their submodules.
        if cfg!(test) && !self.config.submodules() {
            return;
        }
        self.config.update_submodule(submodule);
        let absolute_path = self.config.src.join(submodule);
        if !absolute_path.exists() || dir_is_empty(&absolute_path) {
            let maybe_enable = if !self.config.submodules()
                && self.config.rust_info.is_managed_git_subrepository()
            {
                "\nConsider setting `build.submodules = true` or manually initializing the submodules."
            } else {
                ""
            };
            let err_hint = err_hint.map_or_else(String::new, |e| format!("\n{e}"));
            eprintln!(
                "submodule {submodule} does not appear to be checked out, \
                 but it is required for this step{maybe_enable}{err_hint}"
            );
            exit!(1);
        }
    }

    /// If any submodule has been initialized already, sync it unconditionally.
    /// This avoids contributors checking in a submodule change by accident.
    fn update_existing_submodules(&self) {
        // Avoid running git when there isn't a git checkout, or the user has
        // explicitly disabled submodules in `bootstrap.toml`.
        if !self.config.submodules() {
            return;
        }
        let output = helpers::git(Some(&self.src))
            .args(["config", "--file"])
            .arg(".gitmodules")
            .args(["--get-regexp", "path"])
            .run_capture(self)
            .stdout();
        std::thread::scope(|s| {
            // Look for `submodule.$name.path = $path`
            // Sample output: `submodule.src/rust-installer.path src/tools/rust-installer`
            for line in output.lines() {
                let submodule = line.split_once(' ').unwrap().1;
                let config = self.config.clone();
                s.spawn(move || {
                    Self::update_existing_submodule(&config, submodule);
                });
            }
        });
    }

    /// Updates the given submodule only if it's initialized already; nothing happens otherwise.
    pub fn update_existing_submodule(config: &Config, submodule: &str) {
        // Avoid running git when there isn't a git checkout.
        if !config.submodules() {
            return;
        }

        if config.git_info(false, Path::new(submodule)).is_managed_git_subrepository() {
            config.update_submodule(submodule);
        }
    }

    /// Executes the entire build, as configured by the flags and configuration.
    #[cfg_attr(feature = "tracing", instrument(level = "debug", name = "Build::build", skip_all))]
    pub fn build(&mut self) {
        trace!("setting up job management");
        unsafe {
            crate::utils::job::setup(self);
        }

        // Handle hard-coded subcommands.
        {
            #[cfg(feature = "tracing")]
            let _hardcoded_span =
                span!(tracing::Level::DEBUG, "handling hardcoded subcommands (Format, Perf)")
                    .entered();

            match &self.config.cmd {
                Subcommand::Format { check, all } => {
                    return core::build_steps::format::format(
                        &builder::Builder::new(self),
                        *check,
                        *all,
                        &self.config.paths,
                    );
                }
                Subcommand::Perf(args) => {
                    return core::build_steps::perf::perf(&builder::Builder::new(self), args);
                }
                _cmd => {
                    debug!(cmd = ?_cmd, "not a hardcoded subcommand; returning to normal handling");
                }
            }

            debug!("handling subcommand normally");
        }

        if !self.config.dry_run() {
            #[cfg(feature = "tracing")]
            let _real_run_span = span!(tracing::Level::DEBUG, "executing real run").entered();

            // We first do a dry-run. This is a sanity-check to ensure that
            // steps don't do anything expensive in the dry-run.
            {
                #[cfg(feature = "tracing")]
                let _sanity_check_span =
                    span!(tracing::Level::DEBUG, "(1) executing dry-run sanity-check").entered();
                self.config.set_dry_run(DryRun::SelfCheck);
                let builder = builder::Builder::new(self);
                builder.execute_cli();
            }

            // Actual run.
            {
                #[cfg(feature = "tracing")]
                let _actual_run_span =
                    span!(tracing::Level::DEBUG, "(2) executing actual run").entered();
                self.config.set_dry_run(DryRun::Disabled);
                let builder = builder::Builder::new(self);
                builder.execute_cli();
            }
        } else {
            #[cfg(feature = "tracing")]
            let _dry_run_span = span!(tracing::Level::DEBUG, "executing dry run").entered();

            let builder = builder::Builder::new(self);
            builder.execute_cli();
        }

        #[cfg(feature = "tracing")]
        debug!("checking for postponed test failures from `test  --no-fail-fast`");

        // Check for postponed failures from `test --no-fail-fast`.
        self.config.exec_ctx().report_failures_and_exit();

        #[cfg(feature = "build-metrics")]
        self.metrics.persist(self);
    }

    fn rust_info(&self) -> &GitInfo {
        &self.config.rust_info
    }

    /// Gets the space-separated set of activated features for the standard library.
    /// This can be configured with the `std-features` key in bootstrap.toml.
    fn std_features(&self, target: TargetSelection) -> String {
        let mut features: BTreeSet<&str> =
            self.config.rust_std_features.iter().map(|s| s.as_str()).collect();

        match self.config.llvm_libunwind(target) {
            LlvmLibunwind::InTree => features.insert("llvm-libunwind"),
            LlvmLibunwind::System => features.insert("system-llvm-libunwind"),
            LlvmLibunwind::No => false,
        };

        if self.config.backtrace {
            features.insert("backtrace");
        }

        if self.config.profiler_enabled(target) {
            features.insert("profiler");
        }

        // If zkvm target, generate memcpy, etc.
        if target.contains("zkvm") {
            features.insert("compiler-builtins-mem");
        }

        if self.config.llvm_enzyme {
            features.insert("llvm_enzyme");
        }

        features.into_iter().collect::<Vec<_>>().join(" ")
    }

    /// Gets the space-separated set of activated features for the compiler.
    fn rustc_features(&self, kind: Kind, target: TargetSelection, crates: &[String]) -> String {
        let possible_features_by_crates: HashSet<_> = crates
            .iter()
            .flat_map(|krate| &self.crates[krate].features)
            .map(std::ops::Deref::deref)
            .collect();
        let check = |feature: &str| -> bool {
            crates.is_empty() || possible_features_by_crates.contains(feature)
        };
        let mut features = vec![];
        if self.config.jemalloc(target) && check("jemalloc") {
            features.push("jemalloc");
        }
        if (self.config.llvm_enabled(target) || kind == Kind::Check) && check("llvm") {
            features.push("llvm");
        }
        if self.config.llvm_enzyme {
            features.push("llvm_enzyme");
        }
        if self.config.llvm_offload {
            features.push("llvm_offload");
        }
        // keep in sync with `bootstrap/compile.rs:rustc_cargo_env`
        if self.config.rust_randomize_layout && check("rustc_randomized_layouts") {
            features.push("rustc_randomized_layouts");
        }
        if self.config.compile_time_deps && kind == Kind::Check {
            features.push("check_only");
        }

        if crates.iter().any(|c| c == "rustc_transmute") {
            // for `x test rustc_transmute`, this feature isn't enabled automatically by a
            // dependent crate.
            features.push("rustc");
        }

        // If debug logging is on, then we want the default for tracing:
        // https://github.com/tokio-rs/tracing/blob/3dd5c03d907afdf2c39444a29931833335171554/tracing/src/level_filters.rs#L26
        // which is everything (including debug/trace/etc.)
        // if its unset, if debug_assertions is on, then debug_logging will also be on
        // as well as tracing *ignoring* this feature when debug_assertions is on
        if !self.config.rust_debug_logging && check("max_level_info") {
            features.push("max_level_info");
        }

        features.join(" ")
    }

    /// Component directory that Cargo will produce output into (e.g.
    /// release/debug)
    fn cargo_dir(&self, mode: Mode) -> &'static str {
        match (mode, self.config.rust_optimize.is_release()) {
            (Mode::Std, _) => "dist",
            (_, true) => "release",
            (_, false) => "debug",
        }
    }

    fn tools_dir(&self, build_compiler: Compiler) -> PathBuf {
        let out = self
            .out
            .join(build_compiler.host)
            .join(format!("stage{}-tools-bin", build_compiler.stage + 1));
        t!(fs::create_dir_all(&out));
        out
    }

    /// Returns the root directory for all output generated in a particular
    /// stage when being built with a particular build compiler.
    ///
    /// The mode indicates what the root directory is for.
    fn stage_out(&self, build_compiler: Compiler, mode: Mode) -> PathBuf {
        use std::fmt::Write;

        fn bootstrap_tool() -> (Option<u32>, &'static str) {
            (None, "bootstrap-tools")
        }
        fn staged_tool(build_compiler: Compiler) -> (Option<u32>, &'static str) {
            (Some(build_compiler.stage + 1), "tools")
        }

        let (stage, suffix) = match mode {
            // Std is special, stage N std is built with stage N rustc
            Mode::Std => (Some(build_compiler.stage), "std"),
            // The rest of things are built with stage N-1 rustc
            Mode::Rustc => (Some(build_compiler.stage + 1), "rustc"),
            Mode::Codegen => (Some(build_compiler.stage + 1), "codegen"),
            Mode::ToolBootstrap => bootstrap_tool(),
            Mode::ToolStd | Mode::ToolRustcPrivate => (Some(build_compiler.stage + 1), "tools"),
            Mode::ToolTarget => {
                // If we're not cross-compiling (the common case), share the target directory with
                // bootstrap tools to reuse the build cache.
                if build_compiler.stage == 0 {
                    bootstrap_tool()
                } else {
                    staged_tool(build_compiler)
                }
            }
        };
        let path = self.out.join(build_compiler.host);
        let mut dir_name = String::new();
        if let Some(stage) = stage {
            write!(dir_name, "stage{stage}-").unwrap();
        }
        dir_name.push_str(suffix);
        path.join(dir_name)
    }

    /// Returns the root output directory for all Cargo output in a given stage,
    /// running a particular compiler, whether or not we're building the
    /// standard library, and targeting the specified architecture.
    fn cargo_out(&self, build_compiler: Compiler, mode: Mode, target: TargetSelection) -> PathBuf {
        self.stage_out(build_compiler, mode).join(target).join(self.cargo_dir(mode))
    }

    /// Root output directory of LLVM for `target`
    ///
    /// Note that if LLVM is configured externally then the directory returned
    /// will likely be empty.
    fn llvm_out(&self, target: TargetSelection) -> PathBuf {
        if self.config.llvm_from_ci && self.config.is_host_target(target) {
            self.config.ci_llvm_root()
        } else {
            self.out.join(target).join("llvm")
        }
    }

    fn enzyme_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(&*target.triple).join("enzyme")
    }

    fn offload_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(&*target.triple).join("offload")
    }

    fn lld_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(target).join("lld")
    }

    /// Output directory for all documentation for a target
    fn doc_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(target).join("doc")
    }

    /// Output directory for all JSON-formatted documentation for a target
    fn json_doc_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(target).join("json-doc")
    }

    fn test_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(target).join("test")
    }

    /// Output directory for all documentation for a target
    fn compiler_doc_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(target).join("compiler-doc")
    }

    /// Output directory for some generated md crate documentation for a target (temporary)
    fn md_doc_out(&self, target: TargetSelection) -> PathBuf {
        self.out.join(target).join("md-doc")
    }

    /// Path to the vendored Rust crates.
    fn vendored_crates_path(&self) -> Option<PathBuf> {
        if self.config.vendor { Some(self.src.join(VENDOR_DIR)) } else { None }
    }

    /// Returns the path to `FileCheck` binary for the specified target
    fn llvm_filecheck(&self, target: TargetSelection) -> PathBuf {
        let target_config = self.config.target_config.get(&target);
        if let Some(s) = target_config.and_then(|c| c.llvm_filecheck.as_ref()) {
            s.to_path_buf()
        } else if let Some(s) = target_config.and_then(|c| c.llvm_config.as_ref()) {
            let llvm_bindir = command(s).arg("--bindir").run_capture_stdout(self).stdout();
            let filecheck = Path::new(llvm_bindir.trim()).join(exe("FileCheck", target));
            if filecheck.exists() {
                filecheck
            } else {
                // On Fedora the system LLVM installs FileCheck in the
                // llvm subdirectory of the libdir.
                let llvm_libdir = command(s).arg("--libdir").run_capture_stdout(self).stdout();
                let lib_filecheck =
                    Path::new(llvm_libdir.trim()).join("llvm").join(exe("FileCheck", target));
                if lib_filecheck.exists() {
                    lib_filecheck
                } else {
                    // Return the most normal file name, even though
                    // it doesn't exist, so that any error message
                    // refers to that.
                    filecheck
                }
            }
        } else {
            let base = self.llvm_out(target).join("build");
            let base = if !self.ninja() && target.is_msvc() {
                if self.config.llvm_optimize {
                    if self.config.llvm_release_debuginfo {
                        base.join("RelWithDebInfo")
                    } else {
                        base.join("Release")
                    }
                } else {
                    base.join("Debug")
                }
            } else {
                base
            };
            base.join("bin").join(exe("FileCheck", target))
        }
    }

    /// Directory for libraries built from C/C++ code and shared between stages.
    fn native_dir(&self, target: TargetSelection) -> PathBuf {
        self.out.join(target).join("native")
    }

    /// Root output directory for rust_test_helpers library compiled for
    /// `target`
    fn test_helpers_out(&self, target: TargetSelection) -> PathBuf {
        self.native_dir(target).join("rust-test-helpers")
    }

    /// Adds the `RUST_TEST_THREADS` env var if necessary
    fn add_rust_test_threads(&self, cmd: &mut BootstrapCommand) {
        if env::var_os("RUST_TEST_THREADS").is_none() {
            cmd.env("RUST_TEST_THREADS", self.jobs().to_string());
        }
    }

    /// Returns the libdir of the snapshot compiler.
    fn rustc_snapshot_libdir(&self) -> PathBuf {
        self.rustc_snapshot_sysroot().join(libdir(self.config.host_target))
    }

    /// Returns the sysroot of the snapshot compiler.
    fn rustc_snapshot_sysroot(&self) -> &Path {
        static SYSROOT_CACHE: OnceLock<PathBuf> = OnceLock::new();
        SYSROOT_CACHE.get_or_init(|| {
            command(&self.initial_rustc)
                .run_in_dry_run()
                .args(["--print", "sysroot"])
                .run_capture_stdout(self)
                .stdout()
                .trim()
                .to_owned()
                .into()
        })
    }

    fn info(&self, msg: &str) {
        match self.config.get_dry_run() {
            DryRun::SelfCheck => (),
            DryRun::Disabled | DryRun::UserSelected => {
                println!("{msg}");
            }
        }
    }

    /// Return a `Group` guard for a [`Step`] that:
    /// - Performs `action`
    ///   - If the action is `Kind::Test`, use [`Build::msg_test`] instead.
    /// - On `what`
    ///   - Where `what` possibly corresponds to a `mode`
    /// - `action` is performed with/on the given compiler (`target_and_stage`).
    ///   - Since for some steps it is not possible to pass a single compiler here, it is also
    ///     possible to pass the host and stage explicitly.
    /// - With a given `target`.
    ///
    /// [`Step`]: crate::core::builder::Step
    #[must_use = "Groups should not be dropped until the Step finishes running"]
    #[track_caller]
    fn msg(
        &self,
        action: impl Into<Kind>,
        what: impl Display,
        mode: impl Into<Option<Mode>>,
        target_and_stage: impl Into<TargetAndStage>,
        target: impl Into<Option<TargetSelection>>,
    ) -> Option<gha::Group> {
        let target_and_stage = target_and_stage.into();
        let action = action.into();
        assert!(
            action != Kind::Test,
            "Please use `Build::msg_test` instead of `Build::msg(Kind::Test)`"
        );

        let actual_stage = match mode.into() {
            // Std has the same stage as the compiler that builds it
            Some(Mode::Std) => target_and_stage.stage,
            // Other things have stage corresponding to their build compiler + 1
            Some(
                Mode::Rustc
                | Mode::Codegen
                | Mode::ToolBootstrap
                | Mode::ToolTarget
                | Mode::ToolStd
                | Mode::ToolRustcPrivate,
            )
            | None => target_and_stage.stage + 1,
        };

        let action = action.description();
        let what = what.to_string();
        let msg = |fmt| {
            let space = if !what.is_empty() { " " } else { "" };
            format!("{action} stage{actual_stage} {what}{space}{fmt}")
        };
        let msg = if let Some(target) = target.into() {
            let build_stage = target_and_stage.stage;
            let host = target_and_stage.target;
            if host == target {
                msg(format_args!("(stage{build_stage} -> stage{actual_stage}, {target})"))
            } else {
                msg(format_args!("(stage{build_stage}:{host} -> stage{actual_stage}:{target})"))
            }
        } else {
            msg(format_args!(""))
        };
        self.group(&msg)
    }

    /// Return a `Group` guard for a [`Step`] that tests `what` with the given `stage` and `target`.
    /// Use this instead of [`Build::msg`] for test steps, because for them it is not always clear
    /// what exactly is a build compiler.
    ///
    /// [`Step`]: crate::core::builder::Step
    #[must_use = "Groups should not be dropped until the Step finishes running"]
    #[track_caller]
    fn msg_test(
        &self,
        what: impl Display,
        target: TargetSelection,
        stage: u32,
    ) -> Option<gha::Group> {
        let action = Kind::Test.description();
        let msg = format!("{action} stage{stage} {what} ({target})");
        self.group(&msg)
    }

    /// Return a `Group` guard for a [`Step`] that is only built once and isn't affected by `--stage`.
    ///
    /// [`Step`]: crate::core::builder::Step
    #[must_use = "Groups should not be dropped until the Step finishes running"]
    #[track_caller]
    fn msg_unstaged(
        &self,
        action: impl Into<Kind>,
        what: impl Display,
        target: TargetSelection,
    ) -> Option<gha::Group> {
        let action = action.into().description();
        let msg = format!("{action} {what} for {target}");
        self.group(&msg)
    }

    #[track_caller]
    fn group(&self, msg: &str) -> Option<gha::Group> {
        match self.config.get_dry_run() {
            DryRun::SelfCheck => None,
            DryRun::Disabled | DryRun::UserSelected => Some(gha::group(msg)),
        }
    }

    /// Returns the number of parallel jobs that have been configured for this
    /// build.
    fn jobs(&self) -> u32 {
        self.config.jobs.unwrap_or_else(|| {
            std::thread::available_parallelism().map_or(1, std::num::NonZeroUsize::get) as u32
        })
    }

    fn debuginfo_map_to(&self, which: GitRepo, remap_scheme: RemapScheme) -> Option<String> {
        if !self.config.rust_remap_debuginfo {
            return None;
        }

        match which {
            GitRepo::Rustc => {
                let sha = self.rust_sha().unwrap_or(&self.version);

                match remap_scheme {
                    RemapScheme::Compiler => {
                        // For compiler sources, remap via `/rustc-dev/{sha}` to allow
                        // distinguishing between compiler sources vs library sources, since
                        // `rustc-dev` dist component places them under
                        // `$sysroot/lib/rustlib/rustc-src/rust` as opposed to `rust-src`'s
                        // `$sysroot/lib/rustlib/src/rust`.
                        //
                        // Keep this scheme in sync with `rustc_metadata::rmeta::decoder`'s
                        // `try_to_translate_virtual_to_real`.
                        Some(format!("/rustc-dev/{sha}"))
                    }
                    RemapScheme::NonCompiler => {
                        // For non-compiler sources, use `/rustc/{sha}` remapping scheme.
                        Some(format!("/rustc/{sha}"))
                    }
                }
            }
            GitRepo::Llvm => Some(String::from("/rustc/llvm")),
        }
    }

    /// Returns the path to the C compiler for the target specified.
    fn cc(&self, target: TargetSelection) -> PathBuf {
        if self.config.dry_run() {
            return PathBuf::new();
        }
        self.cc[&target].path().into()
    }

    /// Returns the internal `cc::Tool` for the C compiler.
    fn cc_tool(&self, target: TargetSelection) -> Tool {
        self.cc[&target].clone()
    }

    /// Returns the internal `cc::Tool` for the C++ compiler.
    fn cxx_tool(&self, target: TargetSelection) -> Tool {
        self.cxx[&target].clone()
    }

    /// Returns C flags that `cc-rs` thinks should be enabled for the
    /// specified target by default.
    fn cc_handled_clags(&self, target: TargetSelection, c: CLang) -> Vec<String> {
        if self.config.dry_run() {
            return Vec::new();
        }
        let base = match c {
            CLang::C => self.cc[&target].clone(),
            CLang::Cxx => self.cxx[&target].clone(),
        };

        // Filter out -O and /O (the optimization flags) that we picked up
        // from cc-rs, that's up to the caller to figure out.
        base.args()
            .iter()
            .map(|s| s.to_string_lossy().into_owned())
            .filter(|s| !s.starts_with("-O") && !s.starts_with("/O"))
            .collect::<Vec<String>>()
    }

    /// Returns extra C flags that `cc-rs` doesn't handle.
    fn cc_unhandled_cflags(
        &self,
        target: TargetSelection,
        which: GitRepo,
        c: CLang,
    ) -> Vec<String> {
        let mut base = Vec::new();

        // If we're compiling C++ on macOS then we add a flag indicating that
        // we want libc++ (more filled out than libstdc++), ensuring that
        // LLVM/etc are all properly compiled.
        if matches!(c, CLang::Cxx) && target.contains("apple-darwin") {
            base.push("-stdlib=libc++".into());
        }

        // Work around an apparently bad MinGW / GCC optimization,
        // See: https://lists.llvm.org/pipermail/cfe-dev/2016-December/051980.html
        // See: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78936
        if &*target.triple == "i686-pc-windows-gnu" {
            base.push("-fno-omit-frame-pointer".into());
        }

        if let Some(map_to) = self.debuginfo_map_to(which, RemapScheme::NonCompiler) {
            let map = format!("{}={}", self.src.display(), map_to);
            let cc = self.cc(target);
            if cc.ends_with("clang") || cc.ends_with("gcc") {
                base.push(format!("-fdebug-prefix-map={map}"));
            } else if cc.ends_with("clang-cl.exe") {
                base.push("-Xclang".into());
                base.push(format!("-fdebug-prefix-map={map}"));
            }
        }
        base
    }

    /// Returns the path to the `ar` archive utility for the target specified.
    fn ar(&self, target: TargetSelection) -> Option<PathBuf> {
        if self.config.dry_run() {
            return None;
        }
        self.ar.get(&target).cloned()
    }

    /// Returns the path to the `ranlib` utility for the target specified.
    fn ranlib(&self, target: TargetSelection) -> Option<PathBuf> {
        if self.config.dry_run() {
            return None;
        }
        self.ranlib.get(&target).cloned()
    }

    /// Returns the path to the C++ compiler for the target specified.
    fn cxx(&self, target: TargetSelection) -> Result<PathBuf, String> {
        if self.config.dry_run() {
            return Ok(PathBuf::new());
        }
        match self.cxx.get(&target) {
            Some(p) => Ok(p.path().into()),
            None => Err(format!("target `{target}` is not configured as a host, only as a target")),
        }
    }

    /// Returns the path to the linker for the given target if it needs to be overridden.
    fn linker(&self, target: TargetSelection) -> Option<PathBuf> {
        if self.config.dry_run() {
            return Some(PathBuf::new());
        }
        if let Some(linker) = self.config.target_config.get(&target).and_then(|c| c.linker.clone())
        {
            Some(linker)
        } else if target.contains("vxworks") {
            // need to use CXX compiler as linker to resolve the exception functions
            // that are only existed in CXX libraries
            Some(self.cxx[&target].path().into())
        } else if !self.config.is_host_target(target)
            && helpers::use_host_linker(target)
            && !target.is_msvc()
        {
            Some(self.cc(target))
        } else if self.config.bootstrap_override_lld.is_used()
            && self.is_lld_direct_linker(target)
            && self.host_target == target
        {
            match self.config.bootstrap_override_lld {
                BootstrapOverrideLld::SelfContained => Some(self.initial_lld.clone()),
                BootstrapOverrideLld::External => Some("lld".into()),
                BootstrapOverrideLld::None => None,
            }
        } else {
            None
        }
    }

    // Is LLD configured directly through `-Clinker`?
    // Only MSVC targets use LLD directly at the moment.
    fn is_lld_direct_linker(&self, target: TargetSelection) -> bool {
        target.is_msvc()
    }

    /// Returns if this target should statically link the C runtime, if specified
    fn crt_static(&self, target: TargetSelection) -> Option<bool> {
        if target.contains("pc-windows-msvc") {
            Some(true)
        } else {
            self.config.target_config.get(&target).and_then(|t| t.crt_static)
        }
    }

    /// Returns the "musl root" for this `target`, if defined.
    ///
    /// If this is a native target (host is also musl) and no musl-root is given,
    /// it falls back to the system toolchain in /usr.
    fn musl_root(&self, target: TargetSelection) -> Option<&Path> {
        let configured_root = self
            .config
            .target_config
            .get(&target)
            .and_then(|t| t.musl_root.as_ref())
            .or(self.config.musl_root.as_ref())
            .map(|p| &**p);

        if self.config.is_host_target(target) && configured_root.is_none() {
            Some(Path::new("/usr"))
        } else {
            configured_root
        }
    }

    /// Returns the "musl libdir" for this `target`.
    fn musl_libdir(&self, target: TargetSelection) -> Option<PathBuf> {
        self.config
            .target_config
            .get(&target)
            .and_then(|t| t.musl_libdir.clone())
            .or_else(|| self.musl_root(target).map(|root| root.join("lib")))
    }

    /// Returns the `lib` directory for the WASI target specified, if
    /// configured.
    ///
    /// This first consults `wasi-root` as configured in per-target
    /// configuration, and failing that it assumes that `$WASI_SDK_PATH` is
    /// set in the environment, and failing that `None` is returned.
    fn wasi_libdir(&self, target: TargetSelection) -> Option<PathBuf> {
        let configured =
            self.config.target_config.get(&target).and_then(|t| t.wasi_root.as_ref()).map(|p| &**p);
        if let Some(path) = configured {
            return Some(path.join("lib").join(target.to_string()));
        }
        let mut env_root = self.wasi_sdk_path.clone()?;
        env_root.push("share");
        env_root.push("wasi-sysroot");
        env_root.push("lib");
        env_root.push(target.to_string());
        Some(env_root)
    }

    /// Returns `true` if this is a no-std `target`, if defined
    fn no_std(&self, target: TargetSelection) -> Option<bool> {
        self.config.target_config.get(&target).map(|t| t.no_std)
    }

    /// Returns `true` if the target will be tested using the `remote-test-client`
    /// and `remote-test-server` binaries.
    fn remote_tested(&self, target: TargetSelection) -> bool {
        self.qemu_rootfs(target).is_some()
            || target.contains("android")
            || env::var_os("TEST_DEVICE_ADDR").is_some()
    }

    /// Returns an optional "runner" to pass to `compiletest` when executing
    /// test binaries.
    ///
    /// An example of this would be a WebAssembly runtime when testing the wasm
    /// targets.
    fn runner(&self, target: TargetSelection) -> Option<String> {
        let configured_runner =
            self.config.target_config.get(&target).and_then(|t| t.runner.as_ref()).map(|p| &**p);
        if let Some(runner) = configured_runner {
            return Some(runner.to_owned());
        }

        if target.starts_with("wasm") && target.contains("wasi") {
            self.default_wasi_runner(target)
        } else {
            None
        }
    }

    /// When a `runner` configuration is not provided and a WASI-looking target
    /// is being tested this is consulted to prove the environment to see if
    /// there's a runtime already lying around that seems reasonable to use.
    fn default_wasi_runner(&self, target: TargetSelection) -> Option<String> {
        let mut finder = crate::core::sanity::Finder::new();

        // Look for Wasmtime, and for its default options be sure to disable
        // its caching system since we're executing quite a lot of tests and
        // ideally shouldn't pollute the cache too much.
        if let Some(path) = finder.maybe_have("wasmtime")
            && let Ok(mut path) = path.into_os_string().into_string()
        {
            path.push_str(" run -C cache=n --dir .");
            // Make sure that tests have access to RUSTC_BOOTSTRAP. This (for example) is
            // required for libtest to work on beta/stable channels.
            //
            // NB: with Wasmtime 20 this can change to `-S inherit-env` to
            // inherit the entire environment rather than just this single
            // environment variable.
            path.push_str(" --env RUSTC_BOOTSTRAP");

            if target.contains("wasip2") {
                path.push_str(" --wasi inherit-network --wasi allow-ip-name-lookup");
            }

            return Some(path);
        }

        None
    }

    /// Returns whether the specified tool is configured as part of this build.
    ///
    /// This requires that both the `extended` key is set and the `tools` key is
    /// either unset or specifically contains the specified tool.
    fn tool_enabled(&self, tool: &str) -> bool {
        if !self.config.extended {
            return false;
        }
        match &self.config.tools {
            Some(set) => set.contains(tool),
            None => true,
        }
    }

    /// Returns the root of the "rootfs" image that this target will be using,
    /// if one was configured.
    ///
    /// If `Some` is returned then that means that tests for this target are
    /// emulated with QEMU and binaries will need to be shipped to the emulator.
    fn qemu_rootfs(&self, target: TargetSelection) -> Option<&Path> {
        self.config.target_config.get(&target).and_then(|t| t.qemu_rootfs.as_ref()).map(|p| &**p)
    }

    /// Temporary directory that extended error information is emitted to.
    fn extended_error_dir(&self) -> PathBuf {
        self.out.join("tmp/extended-error-metadata")
    }

    /// Tests whether the `compiler` compiling for `target` should be forced to
    /// use a stage1 compiler instead.
    ///
    /// Currently, by default, the build system does not perform a "full
    /// bootstrap" by default where we compile the compiler three times.
    /// Instead, we compile the compiler two times. The final stage (stage2)
    /// just copies the libraries from the previous stage, which is what this
    /// method detects.
    ///
    /// Here we return `true` if:
    ///
    /// * The build isn't performing a full bootstrap
    /// * The `compiler` is in the final stage, 2
    /// * We're not cross-compiling, so the artifacts are already available in
    ///   stage1
    ///
    /// When all of these conditions are met the build will lift artifacts from
    /// the previous stage forward.
    fn force_use_stage1(&self, stage: u32, target: TargetSelection) -> bool {
        !self.config.full_bootstrap
            && !self.config.download_rustc()
            && stage >= 2
            && (self.hosts.contains(&target) || target == self.host_target)
    }

    /// Checks whether the `compiler` compiling for `target` should be forced to
    /// use a stage2 compiler instead.
    ///
    /// When we download the pre-compiled version of rustc and compiler stage is >= 2,
    /// it should be forced to use a stage2 compiler.
    fn force_use_stage2(&self, stage: u32) -> bool {
        self.config.download_rustc() && stage >= 2
    }

    /// Given `num` in the form "a.b.c" return a "release string" which
    /// describes the release version number.
    ///
    /// For example on nightly this returns "a.b.c-nightly", on beta it returns
    /// "a.b.c-beta.1" and on stable it just returns "a.b.c".
    fn release(&self, num: &str) -> String {
        match &self.config.channel[..] {
            "stable" => num.to_string(),
            "beta" => {
                if !self.config.omit_git_hash {
                    format!("{}-beta.{}", num, self.beta_prerelease_version())
                } else {
                    format!("{num}-beta")
                }
            }
            "nightly" => format!("{num}-nightly"),
            _ => format!("{num}-dev"),
        }
    }

    fn beta_prerelease_version(&self) -> u32 {
        fn extract_beta_rev_from_file<P: AsRef<Path>>(version_file: P) -> Option<String> {
            let version = fs::read_to_string(version_file).ok()?;

            helpers::extract_beta_rev(&version)
        }

        if let Some(s) = self.prerelease_version.get() {
            return s;
        }

        // First check if there is a version file available.
        // If available, we read the beta revision from that file.
        // This only happens when building from a source tarball when Git should not be used.
        let count = extract_beta_rev_from_file(self.src.join("version")).unwrap_or_else(|| {
            // Figure out how many merge commits happened since we branched off main.
            // That's our beta number!
            // (Note that we use a `..` range, not the `...` symmetric difference.)
            helpers::git(Some(&self.src))
                .arg("rev-list")
                .arg("--count")
                .arg("--merges")
                .arg(format!(
                    "refs/remotes/origin/{}..HEAD",
                    self.config.stage0_metadata.config.nightly_branch
                ))
                .run_in_dry_run()
                .run_capture(self)
                .stdout()
        });
        let n = count.trim().parse().unwrap();
        self.prerelease_version.set(Some(n));
        n
    }

    /// Returns the value of `release` above for Rust itself.
    fn rust_release(&self) -> String {
        self.release(&self.version)
    }

    /// Returns the "package version" for a component.
    ///
    /// The package version is typically what shows up in the names of tarballs.
    /// For channels like beta/nightly it's just the channel name, otherwise it's the release
    /// version.
    fn rust_package_vers(&self) -> String {
        match &self.config.channel[..] {
            "stable" => self.version.to_string(),
            "beta" => "beta".to_string(),
            "nightly" => "nightly".to_string(),
            _ => format!("{}-dev", self.version),
        }
    }

    /// Returns the `version` string associated with this compiler for Rust
    /// itself.
    ///
    /// Note that this is a descriptive string which includes the commit date,
    /// sha, version, etc.
    fn rust_version(&self) -> String {
        let mut version = self.rust_info().version(self, &self.version);
        if let Some(ref s) = self.config.description
            && !s.is_empty()
        {
            version.push_str(" (");
            version.push_str(s);
            version.push(')');
        }
        version
    }

    /// Returns the full commit hash.
    fn rust_sha(&self) -> Option<&str> {
        self.rust_info().sha()
    }

    /// Returns the `a.b.c` version that the given package is at.
    fn release_num(&self, package: &str) -> String {
        let toml_file_name = self.src.join(format!("src/tools/{package}/Cargo.toml"));
        let toml = t!(fs::read_to_string(toml_file_name));
        for line in toml.lines() {
            if let Some(stripped) =
                line.strip_prefix("version = \"").and_then(|s| s.strip_suffix('"'))
            {
                return stripped.to_owned();
            }
        }

        panic!("failed to find version in {package}'s Cargo.toml")
    }

    /// Returns `true` if unstable features should be enabled for the compiler
    /// we're building.
    fn unstable_features(&self) -> bool {
        !matches!(&self.config.channel[..], "stable" | "beta")
    }

    /// Returns a Vec of all the dependencies of the given root crate,
    /// including transitive dependencies and the root itself. Only includes
    /// "local" crates (those in the local source tree, not from a registry).
    fn in_tree_crates(&self, root: &str, target: Option<TargetSelection>) -> Vec<&Crate> {
        let mut ret = Vec::new();
        let mut list = vec![root.to_owned()];
        let mut visited = HashSet::new();
        while let Some(krate) = list.pop() {
            let krate = self
                .crates
                .get(&krate)
                .unwrap_or_else(|| panic!("metadata missing for {krate}: {:?}", self.crates));
            ret.push(krate);
            for dep in &krate.deps {
                if !self.crates.contains_key(dep) {
                    // Ignore non-workspace members.
                    continue;
                }
                // Don't include optional deps if their features are not
                // enabled. Ideally this would be computed from `cargo
                // metadata --features …`, but that is somewhat slow. In
                // the future, we may want to consider just filtering all
                // build and dev dependencies in metadata::build.
                if visited.insert(dep)
                    && (dep != "profiler_builtins"
                        || target
                            .map(|t| self.config.profiler_enabled(t))
                            .unwrap_or_else(|| self.config.any_profiler_enabled()))
                    && (dep != "rustc_codegen_llvm"
                        || self.config.hosts.iter().any(|host| self.config.llvm_enabled(*host)))
                {
                    list.push(dep.clone());
                }
            }
        }
        ret.sort_unstable_by_key(|krate| krate.name.clone()); // reproducible order needed for tests
        ret
    }

    fn read_stamp_file(&self, stamp: &BuildStamp) -> Vec<(PathBuf, DependencyType)> {
        if self.config.dry_run() {
            return Vec::new();
        }

        if !stamp.path().exists() {
            eprintln!(
                "ERROR: Unable to find the stamp file {}, did you try to keep a nonexistent build stage?",
                stamp.path().display()
            );
            crate::exit!(1);
        }

        let mut paths = Vec::new();
        let contents = t!(fs::read(stamp.path()), stamp.path());
        // This is the method we use for extracting paths from the stamp file passed to us. See
        // run_cargo for more information (in compile.rs).
        for part in contents.split(|b| *b == 0) {
            if part.is_empty() {
                continue;
            }
            let dependency_type = match part[0] as char {
                'h' => DependencyType::Host,
                's' => DependencyType::TargetSelfContained,
                't' => DependencyType::Target,
                _ => unreachable!(),
            };
            let path = PathBuf::from(t!(str::from_utf8(&part[1..])));
            paths.push((path, dependency_type));
        }
        paths
    }

    /// Copies a file from `src` to `dst`.
    ///
    /// If `src` is a symlink, `src` will be resolved to the actual path
    /// and copied to `dst` instead of the symlink itself.
    #[track_caller]
    pub fn resolve_symlink_and_copy(&self, src: &Path, dst: &Path) {
        self.copy_link_internal(src, dst, true);
    }

    /// Links a file from `src` to `dst`.
    /// Attempts to use hard links if possible, falling back to copying.
    /// You can neither rely on this being a copy nor it being a link,
    /// so do not write to dst.
    #[track_caller]
    pub fn copy_link(&self, src: &Path, dst: &Path, file_type: FileType) {
        self.copy_link_internal(src, dst, false);

        if file_type.could_have_split_debuginfo()
            && let Some(dbg_file) = split_debuginfo(src)
        {
            self.copy_link_internal(
                &dbg_file,
                &dst.with_extension(dbg_file.extension().unwrap()),
                false,
            );
        }
    }

    #[track_caller]
    fn copy_link_internal(&self, src: &Path, dst: &Path, dereference_symlinks: bool) {
        if self.config.dry_run() {
            return;
        }
        if src == dst {
            return;
        }

        #[cfg(feature = "tracing")]
        let _span = trace_io!("file-copy-link", ?src, ?dst);

        if let Err(e) = fs::remove_file(dst)
            && cfg!(windows)
            && e.kind() != io::ErrorKind::NotFound
        {
            // workaround for https://github.com/rust-lang/rust/issues/127126
            // if removing the file fails, attempt to rename it instead.
            let now = t!(SystemTime::now().duration_since(SystemTime::UNIX_EPOCH));
            let _ = fs::rename(dst, format!("{}-{}", dst.display(), now.as_nanos()));
        }
        let mut metadata = t!(src.symlink_metadata(), format!("src = {}", src.display()));
        let mut src = src.to_path_buf();
        if metadata.file_type().is_symlink() {
            if dereference_symlinks {
                src = t!(fs::canonicalize(src));
                metadata = t!(fs::metadata(&src), format!("target = {}", src.display()));
            } else {
                let link = t!(fs::read_link(src));
                t!(self.symlink_file(link, dst));
                return;
            }
        }
        if let Ok(()) = fs::hard_link(&src, dst) {
            // Attempt to "easy copy" by creating a hard link (symlinks are privileged on windows),
            // but if that fails just fall back to a slow `copy` operation.
        } else {
            if let Err(e) = fs::copy(&src, dst) {
                panic!("failed to copy `{}` to `{}`: {}", src.display(), dst.display(), e)
            }
            t!(fs::set_permissions(dst, metadata.permissions()));

            // Restore file times because changing permissions on e.g. Linux using `chmod` can cause
            // file access time to change.
            let file_times = fs::FileTimes::new()
                .set_accessed(t!(metadata.accessed()))
                .set_modified(t!(metadata.modified()));
            t!(set_file_times(dst, file_times));
        }
    }

    /// Links the `src` directory recursively to `dst`. Both are assumed to exist
    /// when this function is called.
    /// Will attempt to use hard links if possible and fall back to copying.
    #[track_caller]
    pub fn cp_link_r(&self, src: &Path, dst: &Path) {
        if self.config.dry_run() {
            return;
        }
        for f in self.read_dir(src) {
            let path = f.path();
            let name = path.file_name().unwrap();
            let dst = dst.join(name);
            if t!(f.file_type()).is_dir() {
                t!(fs::create_dir_all(&dst));
                self.cp_link_r(&path, &dst);
            } else {
                self.copy_link(&path, &dst, FileType::Regular);
            }
        }
    }

    /// Copies the `src` directory recursively to `dst`. Both are assumed to exist
    /// when this function is called.
    /// Will attempt to use hard links if possible and fall back to copying.
    /// Unwanted files or directories can be skipped
    /// by returning `false` from the filter function.
    #[track_caller]
    pub fn cp_link_filtered(&self, src: &Path, dst: &Path, filter: &dyn Fn(&Path) -> bool) {
        // Immediately recurse with an empty relative path
        self.cp_link_filtered_recurse(src, dst, Path::new(""), filter)
    }

    // Inner function does the actual work
    #[track_caller]
    fn cp_link_filtered_recurse(
        &self,
        src: &Path,
        dst: &Path,
        relative: &Path,
        filter: &dyn Fn(&Path) -> bool,
    ) {
        for f in self.read_dir(src) {
            let path = f.path();
            let name = path.file_name().unwrap();
            let dst = dst.join(name);
            let relative = relative.join(name);
            // Only copy file or directory if the filter function returns true
            if filter(&relative) {
                if t!(f.file_type()).is_dir() {
                    let _ = fs::remove_dir_all(&dst);
                    self.create_dir(&dst);
                    self.cp_link_filtered_recurse(&path, &dst, &relative, filter);
                } else {
                    self.copy_link(&path, &dst, FileType::Regular);
                }
            }
        }
    }

    fn copy_link_to_folder(&self, src: &Path, dest_folder: &Path) {
        let file_name = src.file_name().unwrap();
        let dest = dest_folder.join(file_name);
        self.copy_link(src, &dest, FileType::Regular);
    }

    fn install(&self, src: &Path, dstdir: &Path, file_type: FileType) {
        if self.config.dry_run() {
            return;
        }
        let dst = dstdir.join(src.file_name().unwrap());

        #[cfg(feature = "tracing")]
        let _span = trace_io!("install", ?src, ?dst);

        t!(fs::create_dir_all(dstdir));
        if !src.exists() {
            panic!("ERROR: File \"{}\" not found!", src.display());
        }

        self.copy_link_internal(src, &dst, true);
        chmod(&dst, file_type.perms());

        // If this file can have debuginfo, look for split debuginfo and install it too.
        if file_type.could_have_split_debuginfo()
            && let Some(dbg_file) = split_debuginfo(src)
        {
            self.install(&dbg_file, dstdir, FileType::Regular);
        }
    }

    fn read(&self, path: &Path) -> String {
        if self.config.dry_run() {
            return String::new();
        }
        t!(fs::read_to_string(path))
    }

    #[track_caller]
    fn create_dir(&self, dir: &Path) {
        if self.config.dry_run() {
            return;
        }

        #[cfg(feature = "tracing")]
        let _span = trace_io!("dir-create", ?dir);

        t!(fs::create_dir_all(dir))
    }

    fn remove_dir(&self, dir: &Path) {
        if self.config.dry_run() {
            return;
        }

        #[cfg(feature = "tracing")]
        let _span = trace_io!("dir-remove", ?dir);

        t!(fs::remove_dir_all(dir))
    }

    /// Make sure that `dir` will be an empty existing directory after this function ends.
    /// If it existed before, it will be first deleted.
    fn clear_dir(&self, dir: &Path) {
        if self.config.dry_run() {
            return;
        }

        #[cfg(feature = "tracing")]
        let _span = trace_io!("dir-clear", ?dir);

        let _ = std::fs::remove_dir_all(dir);
        self.create_dir(dir);
    }

    fn read_dir(&self, dir: &Path) -> impl Iterator<Item = fs::DirEntry> {
        let iter = match fs::read_dir(dir) {
            Ok(v) => v,
            Err(_) if self.config.dry_run() => return vec![].into_iter(),
            Err(err) => panic!("could not read dir {dir:?}: {err:?}"),
        };
        iter.map(|e| t!(e)).collect::<Vec<_>>().into_iter()
    }

    fn symlink_file<P: AsRef<Path>, Q: AsRef<Path>>(&self, src: P, link: Q) -> io::Result<()> {
        #[cfg(unix)]
        use std::os::unix::fs::symlink as symlink_file;
        #[cfg(windows)]
        use std::os::windows::fs::symlink_file;
        if !self.config.dry_run() { symlink_file(src.as_ref(), link.as_ref()) } else { Ok(()) }
    }

    /// Returns if config.ninja is enabled, and checks for ninja existence,
    /// exiting with a nicer error message if not.
    fn ninja(&self) -> bool {
        let mut cmd_finder = crate::core::sanity::Finder::new();

        if self.config.ninja_in_file {
            // Some Linux distros rename `ninja` to `ninja-build`.
            // CMake can work with either binary name.
            if cmd_finder.maybe_have("ninja-build").is_none()
                && cmd_finder.maybe_have("ninja").is_none()
            {
                eprintln!(
                    "
Couldn't find required command: ninja (or ninja-build)

You should install ninja as described at
<https://github.com/ninja-build/ninja/wiki/Pre-built-Ninja-packages>,
or set `ninja = false` in the `[llvm]` section of `bootstrap.toml`.
Alternatively, set `download-ci-llvm = true` in that `[llvm]` section
to download LLVM rather than building it.
"
                );
                exit!(1);
            }
        }

        // If ninja isn't enabled but we're building for MSVC then we try
        // doubly hard to enable it. It was realized in #43767 that the msbuild
        // CMake generator for MSVC doesn't respect configuration options like
        // disabling LLVM assertions, which can often be quite important!
        //
        // In these cases we automatically enable Ninja if we find it in the
        // environment.
        if !self.config.ninja_in_file
            && self.config.host_target.is_msvc()
            && cmd_finder.maybe_have("ninja").is_some()
        {
            return true;
        }

        self.config.ninja_in_file
    }

    pub fn colored_stdout<R, F: FnOnce(&mut dyn WriteColor) -> R>(&self, f: F) -> R {
        self.colored_stream_inner(StandardStream::stdout, self.config.stdout_is_tty, f)
    }

    pub fn colored_stderr<R, F: FnOnce(&mut dyn WriteColor) -> R>(&self, f: F) -> R {
        self.colored_stream_inner(StandardStream::stderr, self.config.stderr_is_tty, f)
    }

    fn colored_stream_inner<R, F, C>(&self, constructor: C, is_tty: bool, f: F) -> R
    where
        C: Fn(ColorChoice) -> StandardStream,
        F: FnOnce(&mut dyn WriteColor) -> R,
    {
        let choice = match self.config.color {
            flags::Color::Always => ColorChoice::Always,
            flags::Color::Never => ColorChoice::Never,
            flags::Color::Auto if !is_tty => ColorChoice::Never,
            flags::Color::Auto => ColorChoice::Auto,
        };
        let mut stream = constructor(choice);
        let result = f(&mut stream);
        stream.reset().unwrap();
        result
    }

    pub fn exec_ctx(&self) -> &ExecutionContext {
        &self.config.exec_ctx
    }

    pub fn report_summary(&self, path: &Path, start_time: Instant) {
        self.config.exec_ctx.profiler().report_summary(path, start_time);
    }

    #[cfg(feature = "tracing")]
    pub fn report_step_graph(self, directory: &Path) {
        self.step_graph.into_inner().store_to_dot_files(directory);
    }
}

impl AsRef<ExecutionContext> for Build {
    fn as_ref(&self) -> &ExecutionContext {
        &self.config.exec_ctx
    }
}

#[cfg(unix)]
fn chmod(path: &Path, perms: u32) {
    use std::os::unix::fs::*;
    t!(fs::set_permissions(path, fs::Permissions::from_mode(perms)));
}
#[cfg(windows)]
fn chmod(_path: &Path, _perms: u32) {}

impl Compiler {
    pub fn new(stage: u32, host: TargetSelection) -> Self {
        Self { stage, host, forced_compiler: false }
    }

    pub fn forced_compiler(&mut self, forced_compiler: bool) {
        self.forced_compiler = forced_compiler;
    }

    /// Returns `true` if this is a snapshot compiler for `build`'s configuration
    pub fn is_snapshot(&self, build: &Build) -> bool {
        self.stage == 0 && self.host == build.host_target
    }

    /// Indicates whether the compiler was forced to use a specific stage.
    pub fn is_forced_compiler(&self) -> bool {
        self.forced_compiler
    }
}

fn envify(s: &str) -> String {
    s.chars()
        .map(|c| match c {
            '-' => '_',
            c => c,
        })
        .flat_map(|c| c.to_uppercase())
        .collect()
}

/// Ensures that the behavior dump directory is properly initialized.
pub fn prepare_behaviour_dump_dir(build: &Build) {
    static INITIALIZED: OnceLock<bool> = OnceLock::new();

    let dump_path = build.out.join("bootstrap-shims-dump");

    let initialized = INITIALIZED.get().unwrap_or(&false);
    if !initialized {
        // clear old dumps
        if dump_path.exists() {
            t!(fs::remove_dir_all(&dump_path));
        }

        t!(fs::create_dir_all(&dump_path));

        t!(INITIALIZED.set(true));
    }
}