cargo/ops/cargo_compile/

mod.rs

//! The entry point for starting the compilation process for commands like
//! `build`, `test`, `doc`, `rustc`, etc.
//!
//! The [`compile`] function will do all the work to compile a workspace. A
//! rough outline is:
//!
//! 1. Resolve the dependency graph (see [`ops::resolve`]).
//! 2. Download any packages needed (see [`PackageSet`]).
//! 3. Generate a list of top-level "units" of work for the targets the user
//!   requested on the command-line. Each [`Unit`] corresponds to a compiler
//!   invocation. This is done in this module ([`UnitGenerator::generate_root_units`]).
//! 4. Starting from the root [`Unit`]s, generate the [`UnitGraph`] by walking the dependency graph
//!   from the resolver.  See also [`unit_dependencies`].
//! 5. Construct the [`BuildContext`] with all of the information collected so
//!   far. This is the end of the "front end" of compilation.
//! 6. Create a [`BuildRunner`] which coordinates the compilation process
//!   and will perform the following steps:
//!     1. Prepare the `target` directory (see [`Layout`]).
//!     2. Create a [`JobQueue`]. The queue checks the
//!       fingerprint of each `Unit` to determine if it should run or be
//!       skipped.
//!     3. Execute the queue via [`drain_the_queue`]. Each leaf in the queue's dependency graph is
//!        executed, and then removed from the graph when finished. This repeats until the queue is
//!        empty.  Note that this is the only point in cargo that currently uses threads.
//! 7. The result of the compilation is stored in the [`Compilation`] struct. This can be used for
//!    various things, such as running tests after the compilation  has finished.
//!
//! **Note**: "target" inside this module generally refers to ["Cargo Target"],
//! which corresponds to artifact that will be built in a package. Not to be
//! confused with target-tuple or target architecture.
//!
//! [`unit_dependencies`]: crate::compiler::unit_dependencies
//! [`Layout`]: crate::compiler::Layout
//! [`JobQueue`]: crate::compiler::job_queue
//! [`drain_the_queue`]: crate::compiler::job_queue
//! ["Cargo Target"]: https://doc.rust-lang.org/nightly/cargo/reference/cargo-targets.html

use cargo_util::collections::{HashMap, HashSet};
use std::hash::{Hash, Hasher};
use std::sync::Arc;

use crate::compiler::UserIntent;
use crate::compiler::unit_dependencies::build_unit_dependencies;
use crate::compiler::unit_graph::{self, UnitDep, UnitGraph};
use crate::compiler::{BuildConfig, BuildContext, BuildRunner, Compilation};
use crate::compiler::{CompileKind, CompileTarget, RustcTargetData, Unit};
use crate::compiler::{DefaultExecutor, Executor, UnitInterner};
use crate::compiler::{DepKindSet, UnitIndex};
use crate::compiler::{TargetInfo, apply_env_config, standard_lib};
use crate::context::{GlobalContext, WarningHandling};
use crate::drop_println;
use crate::ops;
use crate::ops::resolve::{SpecsAndResolvedFeatures, WorkspaceResolve};
use crate::resolver::features::{self, CliFeatures, FeaturesFor};
use crate::resolver::{ForceAllTargets, HasDevUnits, Resolve};
use crate::util::BuildLogger;
use crate::util::interning::InternedString;
use crate::util::log_message::LogMessage;
use crate::util::machine_message;
use crate::util::machine_message::Message as _;
use crate::util::{CargoResult, StableHasher};
use crate::workspace::profiles::Profiles;
use crate::workspace::{PackageId, PackageSet, SourceId, TargetKind, Workspace};

mod compile_filter;
use cargo_util_schemas::core::CrateType;
use cargo_util_terminal::report::{Group, Level, Origin};
pub use compile_filter::{CompileFilter, FilterRule, LibRule};

pub(super) mod unit_generator;
use itertools::Itertools as _;
use unit_generator::UnitGenerator;

mod packages;

pub use packages::Packages;

/// Contains information about how a package should be compiled.
///
/// Note on distinction between `CompileOptions` and [`BuildConfig`]:
/// `BuildConfig` contains values that need to be retained after
/// [`BuildContext`] is created. The other fields are no longer necessary. Think
/// of it as `CompileOptions` are high-level settings requested on the
/// command-line, and `BuildConfig` are low-level settings for actually
/// driving `rustc`.
#[derive(Debug, Clone)]
pub struct CompileOptions {
    /// Configuration information for a rustc build
    pub build_config: BuildConfig,
    /// Feature flags requested by the user.
    pub cli_features: CliFeatures,
    /// A set of packages to build.
    pub spec: Packages,
    /// Filter to apply to the root package to select which targets will be
    /// built.
    pub filter: CompileFilter,
    /// Extra arguments to be passed to rustdoc (single target only)
    pub target_rustdoc_args: Option<Vec<String>>,
    /// The specified target will be compiled with all the available arguments,
    /// note that this only accounts for the *final* invocation of rustc
    pub target_rustc_args: Option<Vec<String>>,
    /// Crate types to be passed to rustc (single target only)
    pub target_rustc_crate_types: Option<Vec<String>>,
    /// Whether the `--document-private-items` flags was specified and should
    /// be forwarded to `rustdoc`.
    pub rustdoc_document_private_items: bool,
    /// Whether the build process should check the minimum Rust version
    /// defined in the cargo metadata for a crate.
    pub honor_rust_version: Option<bool>,
}

impl CompileOptions {
    pub fn new(gctx: &GlobalContext, intent: UserIntent) -> CargoResult<CompileOptions> {
        let jobs = None;
        let keep_going = false;
        Ok(CompileOptions {
            build_config: BuildConfig::new(gctx, jobs, keep_going, &[], intent)?,
            cli_features: CliFeatures::new_all(false),
            spec: ops::Packages::Packages(Vec::new()),
            filter: CompileFilter::Default {
                required_features_filterable: false,
            },
            target_rustdoc_args: None,
            target_rustc_args: None,
            target_rustc_crate_types: None,
            rustdoc_document_private_items: false,
            honor_rust_version: None,
        })
    }
}

/// Compiles!
///
/// This uses the [`DefaultExecutor`]. To use a custom [`Executor`], see [`compile_with_exec`].
pub fn compile<'a>(ws: &Workspace<'a>, options: &CompileOptions) -> CargoResult<Compilation<'a>> {
    let exec: Arc<dyn Executor> = Arc::new(DefaultExecutor);
    compile_with_exec(ws, options, &exec)
}

/// Like [`compile`] but allows specifying a custom [`Executor`]
/// that will be able to intercept build calls and add custom logic.
///
/// [`compile`] uses [`DefaultExecutor`] which just passes calls through.
pub fn compile_with_exec<'a>(
    ws: &Workspace<'a>,
    options: &CompileOptions,
    exec: &Arc<dyn Executor>,
) -> CargoResult<Compilation<'a>> {
    let parse_pass_output = crate::diagnostics::passes::emit_parse_diagnostics(
        ws,
        crate::diagnostics::rules::PARSE_PASS_RULES,
    )?;
    let compilation = compile_ws(ws, options, exec)?;
    if ws.gctx().warning_handling()? == WarningHandling::Deny
        && (compilation.lint_warning_count + parse_pass_output.lint_warning_count) > 0
    {
        anyhow::bail!("warnings are denied by `build.warnings` configuration")
    }
    Ok(compilation)
}

/// Like [`compile_with_exec`] but without warnings from manifest parsing.
#[tracing::instrument(skip_all)]
fn compile_ws<'a>(
    ws: &Workspace<'a>,
    options: &CompileOptions,
    exec: &Arc<dyn Executor>,
) -> CargoResult<Compilation<'a>> {
    let interner = UnitInterner::new();
    let logger = BuildLogger::maybe_new(ws, &options.build_config)?;

    if let Some(ref logger) = logger {
        let rustc = ws.gctx().load_global_rustc(Some(ws))?;
        let num_cpus = std::thread::available_parallelism()
            .ok()
            .map(|x| x.get() as u64);
        logger.log(LogMessage::BuildStarted {
            command: std::env::args_os()
                .map(|arg| arg.to_string_lossy().into_owned())
                .collect(),
            cwd: ws.gctx().cwd().to_path_buf(),
            host: rustc.host.to_string(),
            jobs: options.build_config.jobs,
            num_cpus,
            profile: options.build_config.requested_profile.to_string(),
            rustc_version: rustc.version.to_string(),
            rustc_version_verbose: rustc.verbose_version.clone(),
            target_dir: ws.target_dir().as_path_unlocked().to_path_buf(),
            workspace_root: ws.root().to_path_buf(),
        });

        if options.build_config.emit_json() {
            let run_id = logger.run_id().to_string();
            let msg = machine_message::BuildStarted { run_id: &run_id }.to_json_string();
            writeln!(ws.gctx().shell().out(), "{msg}")?;
        }
    }

    let bcx = create_bcx(ws, options, &interner, logger.as_ref())?;

    if options.build_config.unit_graph {
        unit_graph::emit_serialized_unit_graph(&bcx.roots, &bcx.unit_graph, ws.gctx())?;
        return Compilation::new(&bcx);
    }
    crate::workspace::gc::auto_gc(bcx.gctx);
    let build_runner = BuildRunner::new(&bcx)?;
    if options.build_config.dry_run {
        build_runner.dry_run()
    } else {
        build_runner.compile(exec)
    }
}

/// Executes `rustc --print <VALUE>`.
///
/// * `print_opt_value` is the VALUE passed through.
pub fn print<'a>(
    ws: &Workspace<'a>,
    options: &CompileOptions,
    print_opt_value: &str,
) -> CargoResult<()> {
    let CompileOptions {
        ref build_config,
        ref target_rustc_args,
        ..
    } = *options;
    let gctx = ws.gctx();
    let rustc = gctx.load_global_rustc(Some(ws))?;
    for (index, kind) in build_config.requested_kinds.iter().enumerate() {
        if index != 0 {
            drop_println!(gctx);
        }
        let target_info = TargetInfo::new(gctx, &build_config.requested_kinds, &rustc, *kind)?;
        let mut process = rustc.process();
        apply_env_config(gctx, &mut process)?;
        process.args(&target_info.rustflags);
        if let Some(args) = target_rustc_args {
            process.args(args);
        }
        kind.add_target_arg(&mut process);
        process.arg("--print").arg(print_opt_value);
        process.exec()?;
    }
    Ok(())
}

/// Prepares all required information for the actual compilation.
///
/// For how it works and what data it collects,
/// please see the [module-level documentation](self).
#[tracing::instrument(skip_all)]
pub fn create_bcx<'a, 'gctx>(
    ws: &'a Workspace<'gctx>,
    options: &'a CompileOptions,
    interner: &'a UnitInterner,
    logger: Option<&'a BuildLogger>,
) -> CargoResult<BuildContext<'a, 'gctx>> {
    let CompileOptions {
        ref build_config,
        ref spec,
        ref cli_features,
        ref filter,
        ref target_rustdoc_args,
        ref target_rustc_args,
        ref target_rustc_crate_types,
        rustdoc_document_private_items,
        honor_rust_version,
    } = *options;
    let gctx = ws.gctx();

    // Perform some pre-flight validation.
    match build_config.intent {
        UserIntent::Test | UserIntent::Build | UserIntent::Check { .. } | UserIntent::Bench => {
            if ws.gctx().get_env("RUST_FLAGS").is_ok() {
                gctx.shell().print_report(
                    &[Level::WARNING
                        .secondary_title("ignoring environment variable `RUST_FLAGS`")
                        .element(Level::HELP.message("rust flags are passed via `RUSTFLAGS`"))],
                    false,
                )?;
            }
        }
        UserIntent::Doc { .. } | UserIntent::Doctest => {
            if ws.gctx().get_env("RUSTDOC_FLAGS").is_ok() {
                gctx.shell().print_report(
                    &[Level::WARNING
                        .secondary_title("ignoring environment variable `RUSTDOC_FLAGS`")
                        .element(
                            Level::HELP.message("rustdoc flags are passed via `RUSTDOCFLAGS`"),
                        )],
                    false,
                )?;
            }
        }
    }
    gctx.validate_term_config()?;

    let mut target_data = RustcTargetData::new(ws, &build_config.requested_kinds)?;

    let specs = spec.to_package_id_specs(ws)?;
    let has_dev_units = {
        // Rustdoc itself doesn't need dev-dependencies. But to scrape examples from packages in the
        // workspace, if any of those packages need dev-dependencies, then we need include dev-dependencies
        // to scrape those packages.
        let any_pkg_has_scrape_enabled = ws
            .members_with_features(&specs, cli_features)?
            .iter()
            .any(|(pkg, _)| {
                pkg.targets()
                    .iter()
                    .any(|target| target.is_example() && target.doc_scrape_examples().is_enabled())
            });

        if filter.need_dev_deps(build_config.intent)
            || (build_config.intent.is_doc() && any_pkg_has_scrape_enabled)
        {
            HasDevUnits::Yes
        } else {
            HasDevUnits::No
        }
    };
    let dry_run = false;

    if let Some(logger) = logger {
        let elapsed = ws.gctx().invocation_instant().elapsed().as_secs_f64();
        logger.log(LogMessage::ResolutionStarted { elapsed });
    }

    let resolve = ops::resolve_ws_with_opts(
        ws,
        &mut target_data,
        &build_config.requested_kinds,
        cli_features,
        &specs,
        has_dev_units,
        ForceAllTargets::No,
        dry_run,
    )?;
    let WorkspaceResolve {
        mut pkg_set,
        workspace_resolve,
        targeted_resolve: resolve,
        specs_and_features,
    } = resolve;

    if let Some(logger) = logger {
        let elapsed = ws.gctx().invocation_instant().elapsed().as_secs_f64();
        logger.log(LogMessage::ResolutionFinished { elapsed });
    }

    let std_resolve_features = if let Some(crates) = &gctx.cli_unstable().build_std {
        let (std_package_set, std_resolve, std_features) = standard_lib::resolve_std(
            ws,
            &mut target_data,
            &build_config,
            crates,
            &build_config.requested_kinds,
        )?;
        pkg_set.add_set(std_package_set);
        Some((std_resolve, std_features))
    } else {
        None
    };

    // Find the packages in the resolver that the user wants to build (those
    // passed in with `-p` or the defaults from the workspace), and convert
    // Vec<PackageIdSpec> to a Vec<PackageId>.
    let to_build_ids = resolve.specs_to_ids(&specs)?;
    // Now get the `Package` for each `PackageId`. This may trigger a download
    // if the user specified `-p` for a dependency that is not downloaded.
    // Dependencies will be downloaded during build_unit_dependencies.
    let mut to_builds = pkg_set.get_many(to_build_ids)?;

    // The ordering here affects some error messages coming out of cargo, so
    // let's be test and CLI friendly by always printing in the same order if
    // there's an error.
    to_builds.sort_by_key(|p| p.package_id());

    for pkg in to_builds.iter() {
        pkg.manifest().print_teapot(gctx);

        if build_config.intent.is_any_test()
            && !ws.is_member(pkg)
            && pkg.dependencies().iter().any(|dep| !dep.is_transitive())
        {
            anyhow::bail!(
                "package `{}` cannot be tested because it requires dev-dependencies \
                 and is not a member of the workspace",
                pkg.name()
            );
        }
    }

    let (extra_args, extra_args_name) = match (target_rustc_args, target_rustdoc_args) {
        (Some(args), _) => (Some(args.clone()), "rustc"),
        (_, Some(args)) => (Some(args.clone()), "rustdoc"),
        _ => (None, ""),
    };

    if extra_args.is_some() && to_builds.len() != 1 {
        panic!(
            "`{}` should not accept multiple `-p` flags",
            extra_args_name
        );
    }

    let profiles = Profiles::new(ws, build_config.requested_profile)?;
    profiles.validate_packages(
        ws.profiles(),
        &mut gctx.shell(),
        workspace_resolve.as_ref().unwrap_or(&resolve),
    )?;

    // If `--target` has not been specified, then the unit graph is built
    // assuming `--target $HOST` was specified. See
    // `rebuild_unit_graph_shared` for more on why this is done.
    let explicit_host_kind = CompileKind::Target(CompileTarget::new(
        &target_data.rustc.host,
        gctx.cli_unstable().json_target_spec,
    )?);
    let explicit_host_kinds: Vec<_> = build_config
        .requested_kinds
        .iter()
        .map(|kind| match kind {
            CompileKind::Host => explicit_host_kind,
            CompileKind::Target(t) => CompileKind::Target(*t),
        })
        .collect();

    let mut root_units = Vec::new();
    let mut unit_graph = HashMap::default();
    let mut scrape_units = Vec::new();

    if let Some(logger) = logger {
        let elapsed = ws.gctx().invocation_instant().elapsed().as_secs_f64();
        logger.log(LogMessage::UnitGraphStarted { elapsed });
    }

    let mut selected_dep_kinds = DepKindSet::default();
    for SpecsAndResolvedFeatures {
        specs,
        resolved_features,
    } in &specs_and_features
    {
        // Passing `build_config.requested_kinds` instead of
        // `explicit_host_kinds` here so that `generate_root_units` can do
        // its own special handling of `CompileKind::Host`. It will
        // internally replace the host kind by the `explicit_host_kind`
        // before setting as a unit.
        let spec_names = specs.iter().map(|spec| spec.name()).collect::<Vec<_>>();
        let packages = to_builds
            .iter()
            .filter(|package| spec_names.contains(&package.name().as_str()))
            .cloned()
            .collect::<Vec<_>>();
        let generator = UnitGenerator {
            ws,
            packages: &packages,
            spec,
            target_data: &target_data,
            filter,
            requested_kinds: &build_config.requested_kinds,
            explicit_host_kind,
            intent: build_config.intent,
            resolve: &resolve,
            workspace_resolve: &workspace_resolve,
            resolved_features: &resolved_features,
            package_set: &pkg_set,
            profiles: &profiles,
            interner,
            has_dev_units,
        };
        let (mut targeted_root_units, curr_selected_dep_kinds) = generator.generate_root_units()?;
        // Should be fine as the loop iterate is independent of target selection
        selected_dep_kinds = curr_selected_dep_kinds;

        if let Some(args) = target_rustc_crate_types {
            override_rustc_crate_types(&mut targeted_root_units, args, interner)?;
        }

        let should_scrape =
            build_config.intent.is_doc() && gctx.cli_unstable().rustdoc_scrape_examples;
        let targeted_scrape_units = if should_scrape {
            generator.generate_scrape_units(&targeted_root_units)?
        } else {
            Vec::new()
        };

        let std_roots = if let Some(crates) = gctx.cli_unstable().build_std.as_ref() {
            let (std_resolve, std_features) = std_resolve_features.as_ref().unwrap();
            standard_lib::generate_std_roots(
                &crates,
                &targeted_root_units,
                std_resolve,
                std_features,
                &explicit_host_kinds,
                &pkg_set,
                interner,
                &profiles,
                &target_data,
            )?
        } else {
            Default::default()
        };

        unit_graph.extend(build_unit_dependencies(
            ws,
            &pkg_set,
            &resolve,
            &resolved_features,
            std_resolve_features.as_ref(),
            &targeted_root_units,
            &targeted_scrape_units,
            &std_roots,
            build_config.intent,
            &target_data,
            &profiles,
            interner,
        )?);
        root_units.extend(targeted_root_units);
        scrape_units.extend(targeted_scrape_units);
    }

    // TODO: In theory, Cargo should also dedupe the roots, but I'm uncertain
    // what heuristics to use in that case.
    if build_config.intent.wants_deps_docs() {
        remove_duplicate_doc(build_config, &root_units, &mut unit_graph);
    }

    let host_kind_requested = build_config
        .requested_kinds
        .iter()
        .any(CompileKind::is_host);
    // Rebuild the unit graph, replacing the explicit host targets with
    // CompileKind::Host, removing `artifact_target_for_features` and merging any dependencies
    // shared with build and artifact dependencies.
    //
    // NOTE: after this point, all units and the unit graph must be immutable.
    let (root_units, scrape_units, unit_graph) = rebuild_unit_graph_shared(
        interner,
        unit_graph,
        &root_units,
        &scrape_units,
        host_kind_requested.then_some(explicit_host_kind),
        build_config.compile_time_deps_only,
    );

    let units: Vec<_> = unit_graph.keys().sorted().collect();
    let unit_to_index: HashMap<_, _> = units
        .iter()
        .enumerate()
        .map(|(i, &unit)| (unit.clone(), UnitIndex(i as u64)))
        .collect();

    if let Some(logger) = logger {
        let root_unit_indexes: HashSet<_> =
            root_units.iter().map(|unit| unit_to_index[&unit]).collect();

        for (index, unit) in units.into_iter().enumerate() {
            let index = UnitIndex(index as u64);
            let dependencies = unit_graph
                .get(unit)
                .map(|deps| {
                    deps.iter()
                        .filter_map(|dep| unit_to_index.get(&dep.unit).copied())
                        .collect()
                })
                .unwrap_or_default();
            logger.log(LogMessage::UnitRegistered {
                package_id: unit.pkg.package_id().to_spec(),
                target: (&unit.target).into(),
                mode: unit.mode,
                platform: target_data.short_name(&unit.kind).to_owned(),
                index,
                features: unit
                    .features
                    .iter()
                    .map(|s| s.as_str().to_owned())
                    .collect(),
                requested: root_unit_indexes.contains(&index),
                dependencies,
            });
        }
        let elapsed = ws.gctx().invocation_instant().elapsed().as_secs_f64();
        logger.log(LogMessage::UnitGraphFinished { elapsed });
    }

    let mut extra_compiler_args = HashMap::default();
    if let Some(args) = extra_args {
        if root_units.len() != 1 {
            anyhow::bail!(
                "extra arguments to `{}` can only be passed to one \
                 target, consider filtering\nthe package by passing, \
                 e.g., `--lib` or `--bin NAME` to specify a single target",
                extra_args_name
            );
        }
        extra_compiler_args.insert(root_units[0].clone(), args);
    }

    for unit in root_units
        .iter()
        .filter(|unit| unit.mode.is_doc() || unit.mode.is_doc_test())
        .filter(|unit| rustdoc_document_private_items || unit.target.is_bin())
    {
        // Add `--document-private-items` rustdoc flag if requested or if
        // the target is a binary. Binary crates get their private items
        // documented by default.
        let mut args = vec!["--document-private-items".into()];
        if unit.target.is_bin() {
            // This warning only makes sense if it's possible to document private items
            // sometimes and ignore them at other times. But cargo consistently passes
            // `--document-private-items`, so the warning isn't useful.
            args.push("-Arustdoc::private-intra-doc-links".into());
        }
        extra_compiler_args
            .entry(unit.clone())
            .or_default()
            .extend(args);
    }

    // Validate target src path for each root unit
    let mut error_count: usize = 0;
    for unit in &root_units {
        if let Some(target_src_path) = unit.target.src_path().path() {
            validate_target_path_as_source_file(
                gctx,
                target_src_path,
                unit.target.name(),
                unit.target.kind(),
                unit.pkg.manifest_path(),
                &mut error_count,
            )?
        }
    }
    if error_count > 0 {
        let plural: &str = if error_count > 1 { "s" } else { "" };
        anyhow::bail!(
            "could not compile due to {error_count} previous target resolution error{plural}"
        );
    }

    if honor_rust_version.unwrap_or(true) {
        let rustc_version = target_data.rustc.version.clone().into();

        let mut incompatible = Vec::new();
        let mut local_incompatible = false;
        for unit in unit_graph.keys() {
            let Some(pkg_msrv) = unit.pkg.rust_version() else {
                continue;
            };

            if pkg_msrv.is_compatible_with(&rustc_version) {
                continue;
            }

            local_incompatible |= unit.is_local();
            incompatible.push((unit, pkg_msrv));
        }
        if !incompatible.is_empty() {
            use std::fmt::Write as _;

            let plural = if incompatible.len() == 1 { "" } else { "s" };
            let mut message = format!(
                "rustc {rustc_version} is not supported by the following package{plural}:\n"
            );
            incompatible.sort_by_key(|(unit, _)| (unit.pkg.name(), unit.pkg.version()));
            for (unit, msrv) in incompatible {
                let name = &unit.pkg.name();
                let version = &unit.pkg.version();
                writeln!(&mut message, "  {name}@{version} requires rustc {msrv}").unwrap();
            }
            if !ws.is_ephemeral() && !local_incompatible {
                writeln!(
                    &mut message,
                    "Either upgrade rustc or select compatible dependency versions with
`cargo update <name>@<current-ver> --precise <compatible-ver>`
where `<compatible-ver>` is the latest version supporting rustc {rustc_version}",
                )
                .unwrap();
            }
            return Err(anyhow::Error::msg(message));
        }
    }

    let bcx = BuildContext::new(
        ws,
        logger,
        pkg_set,
        build_config,
        selected_dep_kinds,
        profiles,
        extra_compiler_args,
        target_data,
        root_units,
        unit_graph,
        unit_to_index,
        scrape_units,
    )?;

    Ok(bcx)
}

// Checks if a target path exists and is a source file, not a directory
fn validate_target_path_as_source_file(
    gctx: &GlobalContext,
    target_path: &std::path::Path,
    target_name: &str,
    target_kind: &TargetKind,
    unit_manifest_path: &std::path::Path,
    error_count: &mut usize,
) -> CargoResult<()> {
    if !target_path.exists() {
        *error_count += 1;

        let err_msg = format!(
            "can't find {} `{}` at path `{}`",
            target_kind.description(),
            target_name,
            target_path.display()
        );

        let group = Group::with_title(Level::ERROR.primary_title(err_msg)).element(Origin::path(
            unit_manifest_path.to_str().unwrap_or_default(),
        ));

        gctx.shell().print_report(&[group], true)?;
    } else if target_path.is_dir() {
        *error_count += 1;

        // suggest setting the path to a likely entrypoint
        let main_rs = target_path.join("main.rs");
        let lib_rs = target_path.join("lib.rs");

        let suggested_files_opt = match target_kind {
            TargetKind::Lib(_) => {
                if lib_rs.exists() {
                    Some(format!("`{}`", lib_rs.display()))
                } else {
                    None
                }
            }
            TargetKind::Bin => {
                if main_rs.exists() {
                    Some(format!("`{}`", main_rs.display()))
                } else {
                    None
                }
            }
            TargetKind::Test => {
                if main_rs.exists() {
                    Some(format!("`{}`", main_rs.display()))
                } else {
                    None
                }
            }
            TargetKind::ExampleBin => {
                if main_rs.exists() {
                    Some(format!("`{}`", main_rs.display()))
                } else {
                    None
                }
            }
            TargetKind::Bench => {
                if main_rs.exists() {
                    Some(format!("`{}`", main_rs.display()))
                } else {
                    None
                }
            }
            TargetKind::ExampleLib(_) => {
                if lib_rs.exists() {
                    Some(format!("`{}`", lib_rs.display()))
                } else {
                    None
                }
            }
            TargetKind::CustomBuild => None,
        };

        let err_msg = format!(
            "path `{}` for {} `{}` is a directory, but a source file was expected.",
            target_path.display(),
            target_kind.description(),
            target_name,
        );
        let mut group = Group::with_title(Level::ERROR.primary_title(err_msg)).element(
            Origin::path(unit_manifest_path.to_str().unwrap_or_default()),
        );

        if let Some(suggested_files) = suggested_files_opt {
            group = group.element(
                Level::HELP.message(format!("an entry point exists at {}", suggested_files)),
            );
        }

        gctx.shell().print_report(&[group], true)?;
    }

    Ok(())
}

/// This is used to rebuild the unit graph, sharing host dependencies if possible,
/// and applying other unit adjustments based on the whole graph.
///
/// This will translate any unit's `CompileKind::Target(host)` to
/// `CompileKind::Host` if `to_host` is not `None` and the kind is equal to `to_host`.
/// This also handles generating the unit `dep_hash`, and merging shared units if possible.
///
/// This is necessary because if normal dependencies used `CompileKind::Host`,
/// there would be no way to distinguish those units from build-dependency
/// units or artifact dependency units.
/// This can cause a problem if a shared normal/build/artifact dependency needs
/// to link to another dependency whose features differ based on whether or
/// not it is a normal, build or artifact dependency. If all units used
/// `CompileKind::Host`, then they would end up being identical, causing a
/// collision in the `UnitGraph`, and Cargo would end up randomly choosing one
/// value or the other.
///
/// The solution is to keep normal, build and artifact dependencies separate when
/// building the unit graph, and then run this second pass which will try to
/// combine shared dependencies safely. By adding a hash of the dependencies
/// to the `Unit`, this allows the `CompileKind` to be changed back to `Host`
/// and `artifact_target_for_features` to be removed without fear of an unwanted
/// collision for build or artifact dependencies.
///
/// This is also responsible for adjusting the `strip` profile option to
/// opportunistically strip if debug is 0 for all dependencies. This helps
/// remove debuginfo added by the standard library.
///
/// This is also responsible for adjusting the `debug` setting for host
/// dependencies, turning off debug if the user has not explicitly enabled it,
/// and the unit is not shared with a target unit.
///
/// This is also responsible for adjusting whether each unit should be compiled
/// or not regarding `--compile-time-deps` flag.
fn rebuild_unit_graph_shared(
    interner: &UnitInterner,
    unit_graph: UnitGraph,
    roots: &[Unit],
    scrape_units: &[Unit],
    to_host: Option<CompileKind>,
    compile_time_deps_only: bool,
) -> (Vec<Unit>, Vec<Unit>, UnitGraph) {
    let mut result = UnitGraph::default();

    // Map of the old unit to the new unit, used to avoid recursing into units
    // that have already been computed to improve performance.
    let mut memo = HashMap::default();

    // Map of the unit to whether the unit requires debuginfo
    let mut debuginfo_memo = HashMap::default();

    let new_roots = roots
        .iter()
        .map(|root| {
            traverse_and_share(
                interner,
                &mut memo,
                &mut debuginfo_memo,
                &mut result,
                &unit_graph,
                root,
                true,
                false,
                to_host,
                compile_time_deps_only,
            )
        })
        .collect();
    // If no unit in the unit graph ended up having scrape units attached as dependencies,
    // then they won't have been discovered in traverse_and_share and hence won't be in
    // memo. So we filter out missing scrape units.
    let new_scrape_units = scrape_units
        .iter()
        .map(|unit| memo.get(unit).unwrap().clone())
        .collect();
    (new_roots, new_scrape_units, result)
}

/// Recursive function for rebuilding the graph.
///
/// This walks `unit_graph`, starting at the given `unit`. It inserts the new
/// units into `new_graph`, and returns a new updated version of the given
/// unit (`dep_hash` is filled in, and `kind` switched if necessary).
fn traverse_and_share(
    interner: &UnitInterner,
    memo: &mut HashMap<Unit, Unit>,
    debuginfo_memo: &mut HashMap<Unit, bool>,
    new_graph: &mut UnitGraph,
    unit_graph: &UnitGraph,
    unit: &Unit,
    unit_is_root: bool,
    unit_is_for_host: bool,
    to_host: Option<CompileKind>,
    compile_time_deps_only: bool,
) -> Unit {
    if let Some(new_unit) = memo.get(unit) {
        // Already computed, no need to recompute.
        return new_unit.clone();
    }
    let mut dep_hash = StableHasher::new();
    let skip_non_compile_time_deps = compile_time_deps_only
        && (!unit.target.is_compile_time_dependency() ||
        // Root unit is not a dependency unless other units are dependant
        // to it.
        unit_is_root);
    let new_deps: Vec<_> = unit_graph[unit]
        .iter()
        .map(|dep| {
            let new_dep_unit = traverse_and_share(
                interner,
                memo,
                debuginfo_memo,
                new_graph,
                unit_graph,
                &dep.unit,
                false,
                dep.unit_for.is_for_host(),
                to_host,
                // If we should compile the current unit, we should also compile
                // its dependencies. And if not, we should compile compile time
                // dependencies only.
                skip_non_compile_time_deps,
            );
            new_dep_unit.hash(&mut dep_hash);
            UnitDep {
                unit: new_dep_unit,
                ..dep.clone()
            }
        })
        .collect();
    // Here, we have recursively traversed this unit's dependencies, and hashed them: we can
    // finalize the dep hash.
    let new_dep_hash = Hasher::finish(&dep_hash);

    // This is the key part of the sharing process: if the unit is a runtime dependency, whose
    // target is the same as the host, we canonicalize the compile kind to `CompileKind::Host`.
    // A possible host dependency counterpart to this unit would have that kind, and if such a unit
    // exists in the current `unit_graph`, they will unify in the new unit graph map `new_graph`.
    // The resulting unit graph will be optimized with less units, thanks to sharing these host
    // dependencies.
    let canonical_kind = match to_host {
        Some(to_host) if to_host == unit.kind => CompileKind::Host,
        _ => unit.kind,
    };

    let mut profile = unit.profile.clone();

    // If this is a build dependency, and it's not shared with runtime dependencies, we can weaken
    // its debuginfo level to optimize build times. We do nothing if it's an artifact dependency,
    // as it and its debuginfo may end up embedded in the main program.
    if unit_is_for_host
        && to_host.is_some()
        && profile.debuginfo.is_deferred()
        && !unit.artifact.is_true()
    {
        // We create a "probe" test to see if a unit with the same explicit debuginfo level exists
        // in the graph. This is the level we'd expect if it was set manually or the default value
        // set by a profile for a runtime dependency: its canonical value.
        let canonical_debuginfo = profile.debuginfo.finalize();
        let mut canonical_profile = profile.clone();
        canonical_profile.debuginfo = canonical_debuginfo;
        let unit_probe = interner.intern(
            &unit.pkg,
            &unit.target,
            canonical_profile,
            to_host.unwrap(),
            unit.mode,
            unit.features.clone(),
            unit.rustflags.clone(),
            unit.rustdocflags.clone(),
            unit.links_overrides.clone(),
            unit.is_std,
            unit.dep_hash,
            unit.artifact,
            unit.artifact_target_for_features,
            unit.skip_non_compile_time_dep,
        );

        // We can now turn the deferred value into its actual final value.
        profile.debuginfo = if unit_graph.contains_key(&unit_probe) {
            // The unit is present in both build time and runtime subgraphs: we canonicalize its
            // level to the other unit's, thus ensuring reuse between the two to optimize build times.
            canonical_debuginfo
        } else {
            // The unit is only present in the build time subgraph, we can weaken its debuginfo
            // level to optimize build times.
            canonical_debuginfo.weaken()
        }
    }

    if profile.strip.is_deferred() {
        // If strip was not manually set, and all dependencies of this unit together
        // with this unit have debuginfo turned off, we enable debuginfo stripping.
        // This will remove pre-existing debug symbols coming from the standard library.
        if !profile.debuginfo.is_turned_on()
            && !deps_have_debuginfo(&new_deps, new_graph, debuginfo_memo)
        {
            profile.strip = profile.strip.strip_debuginfo();
        }
    }

    let new_unit = interner.intern(
        &unit.pkg,
        &unit.target,
        profile,
        canonical_kind,
        unit.mode,
        unit.features.clone(),
        unit.rustflags.clone(),
        unit.rustdocflags.clone(),
        unit.links_overrides.clone(),
        unit.is_std,
        new_dep_hash,
        unit.artifact,
        // Since `dep_hash` is now filled in, there's no need to specify the artifact target
        // for target-dependent feature resolution
        None,
        skip_non_compile_time_deps,
    );
    if !unit_is_root || !compile_time_deps_only {
        assert!(memo.insert(unit.clone(), new_unit.clone()).is_none());
    }
    new_graph.entry(new_unit.clone()).or_insert(new_deps);
    new_unit
}

fn deps_have_debuginfo(
    deps: &[UnitDep],
    graph: &UnitGraph,
    memo: &mut HashMap<Unit, bool>,
) -> bool {
    deps.iter().any(|dep| {
        let unit = &dep.unit;
        // Host deps are linked separately from their dependents.
        if unit.target.for_host() {
            return false;
        }
        if unit.profile.debuginfo.is_turned_on() {
            return true;
        }
        if let Some(&has_debuginfo) = memo.get(unit) {
            return has_debuginfo;
        }
        let has_debuginfo = deps_have_debuginfo(&graph[unit], graph, memo);
        memo.insert(unit.clone(), has_debuginfo);
        has_debuginfo
    })
}

/// Removes duplicate `CompileMode::Doc` units that would cause problems with
/// filename collisions.
///
/// Rustdoc only separates units by crate name in the file directory
/// structure. If any two units with the same crate name exist, this would
/// cause a filename collision, causing different rustdoc invocations to stomp
/// on one another's files.
///
/// Unfortunately this does not remove all duplicates, as some of them are
/// either user error, or difficult to remove. Cases that I can think of:
///
/// - Same target name in different packages. See the `collision_doc` test.
/// - Different sources. See `collision_doc_sources` test.
///
/// Ideally this would not be necessary.
fn remove_duplicate_doc(
    build_config: &BuildConfig,
    root_units: &[Unit],
    unit_graph: &mut UnitGraph,
) {
    // First, create a mapping of crate_name -> Unit so we can see where the
    // duplicates are.
    let mut all_docs: HashMap<String, Vec<Unit>> = HashMap::default();
    for unit in unit_graph.keys() {
        if unit.mode.is_doc() {
            all_docs
                .entry(unit.target.crate_name())
                .or_default()
                .push(unit.clone());
        }
    }
    // Keep track of units to remove so that they can be efficiently removed
    // from the unit_deps.
    let mut removed_units: HashSet<Unit> = HashSet::default();
    let mut remove = |units: Vec<Unit>, reason: &str, cb: &dyn Fn(&Unit) -> bool| -> Vec<Unit> {
        let (to_remove, remaining_units): (Vec<Unit>, Vec<Unit>) = units
            .into_iter()
            .partition(|unit| cb(unit) && !root_units.contains(unit));
        for unit in to_remove {
            tracing::debug!(
                "removing duplicate doc due to {} for package {} target `{}`",
                reason,
                unit.pkg,
                unit.target.name()
            );
            unit_graph.remove(&unit);
            removed_units.insert(unit);
        }
        remaining_units
    };
    // Iterate over the duplicates and try to remove them from unit_graph.
    for (_crate_name, mut units) in all_docs {
        if units.len() == 1 {
            continue;
        }
        // Prefer target over host if --target was not specified.
        if build_config
            .requested_kinds
            .iter()
            .all(CompileKind::is_host)
        {
            // Note these duplicates may not be real duplicates, since they
            // might get merged in rebuild_unit_graph_shared. Either way, it
            // shouldn't hurt to remove them early (although the report in the
            // log might be confusing).
            units = remove(units, "host/target merger", &|unit| unit.kind.is_host());
            if units.len() == 1 {
                continue;
            }
        }
        // Prefer newer versions over older.
        let mut source_map: HashMap<(InternedString, SourceId, CompileKind), Vec<Unit>> =
            HashMap::default();
        for unit in units {
            let pkg_id = unit.pkg.package_id();
            // Note, this does not detect duplicates from different sources.
            source_map
                .entry((pkg_id.name(), pkg_id.source_id(), unit.kind))
                .or_default()
                .push(unit);
        }
        let mut remaining_units = Vec::new();
        for (_key, mut units) in source_map {
            if units.len() > 1 {
                units.sort_by(|a, b| a.pkg.version().partial_cmp(b.pkg.version()).unwrap());
                // Remove any entries with version < newest.
                let newest_version = units.last().unwrap().pkg.version().clone();
                let keep_units = remove(units, "older version", &|unit| {
                    unit.pkg.version() < &newest_version
                });
                remaining_units.extend(keep_units);
            } else {
                remaining_units.extend(units);
            }
        }
        if remaining_units.len() == 1 {
            continue;
        }
        // Are there other heuristics to remove duplicates that would make
        // sense? Maybe prefer path sources over all others?
    }
    // Also remove units from the unit_deps so there aren't any dangling edges.
    for unit_deps in unit_graph.values_mut() {
        unit_deps.retain(|unit_dep| !removed_units.contains(&unit_dep.unit));
    }
    // Remove any orphan units that were detached from the graph.
    let mut visited = HashSet::default();
    fn visit(unit: &Unit, graph: &UnitGraph, visited: &mut HashSet<Unit>) {
        if !visited.insert(unit.clone()) {
            return;
        }
        for dep in &graph[unit] {
            visit(&dep.unit, graph, visited);
        }
    }
    for unit in root_units {
        visit(unit, unit_graph, &mut visited);
    }
    unit_graph.retain(|unit, _| visited.contains(unit));
}

/// Override crate types for given units.
///
/// This is primarily used by `cargo rustc --crate-type`.
fn override_rustc_crate_types(
    units: &mut [Unit],
    args: &[String],
    interner: &UnitInterner,
) -> CargoResult<()> {
    if units.len() != 1 {
        anyhow::bail!(
            "crate types to rustc can only be passed to one \
            target, consider filtering\nthe package by passing, \
            e.g., `--lib` or `--example` to specify a single target"
        );
    }

    let unit = &units[0];
    let override_unit = |f: fn(Vec<CrateType>) -> TargetKind| {
        let crate_types = args.iter().map(|s| s.into()).collect();
        let mut target = unit.target.clone();
        target.set_kind(f(crate_types));
        interner.intern(
            &unit.pkg,
            &target,
            unit.profile.clone(),
            unit.kind,
            unit.mode,
            unit.features.clone(),
            unit.rustflags.clone(),
            unit.rustdocflags.clone(),
            unit.links_overrides.clone(),
            unit.is_std,
            unit.dep_hash,
            unit.artifact,
            unit.artifact_target_for_features,
            unit.skip_non_compile_time_dep,
        )
    };
    units[0] = match unit.target.kind() {
        TargetKind::Lib(_) => override_unit(TargetKind::Lib),
        TargetKind::ExampleLib(_) => override_unit(TargetKind::ExampleLib),
        _ => {
            anyhow::bail!(
                "crate types can only be specified for libraries and example libraries.\n\
                Binaries, tests, and benchmarks are always the `bin` crate type"
            );
        }
    };

    Ok(())
}

/// Gets all of the features enabled for a package, plus its dependencies'
/// features.
///
/// Dependencies are added as `dep_name/feat_name` because `required-features`
/// wants to support that syntax.
pub fn resolve_all_features(
    resolve_with_overrides: &Resolve,
    resolved_features: &features::ResolvedFeatures,
    package_set: &PackageSet<'_>,
    package_id: PackageId,
    has_dev_units: HasDevUnits,
    requested_kinds: &[CompileKind],
    target_data: &RustcTargetData<'_>,
    force_all_targets: ForceAllTargets,
) -> HashSet<String> {
    let mut features: HashSet<String> = resolved_features
        .activated_features(package_id, FeaturesFor::NormalOrDev)
        .iter()
        .map(|s| s.to_string())
        .collect();

    // Include features enabled for use by dependencies so targets can also use them with the
    // required-features field when deciding whether to be built or skipped.
    let filtered_deps = PackageSet::filter_deps(
        package_id,
        resolve_with_overrides,
        has_dev_units,
        requested_kinds,
        target_data,
        force_all_targets,
    );
    for (dep_id, deps) in filtered_deps {
        let is_proc_macro = package_set
            .get_one(dep_id)
            .expect("packages downloaded")
            .proc_macro();
        for dep in deps {
            let features_for = FeaturesFor::from_for_host(is_proc_macro || dep.is_build());
            for feature in resolved_features
                .activated_features_unverified(dep_id, features_for)
                .unwrap_or_default()
            {
                features.insert(format!("{}/{}", dep.name_in_toml(), feature));
            }
        }
    }

    features
}