Separating the back folder between backend-agnostic and LLVM-specific code

14 files changed, 4023 insertions, 49 deletions

diff --git a/src/librustc_codegen_ssa/Cargo.toml b/src/librustc_codegen_ssa/Cargo.toml
index ae187c5e3e0..a158c34f9d1 100644
--- a/src/librustc_codegen_ssa/Cargo.toml
+++ b/src/librustc_codegen_ssa/Cargo.toml
@@ -9,3 +9,7 @@ path = "lib.rs"
 test = false
 
 [dependencies]
+cc = "1.0.1"
+num_cpus = "1.0"
+rustc-demangle = "0.1.4"
+memmap = "0.6"
diff --git a/src/librustc_codegen_ssa/back/archive.rs b/src/librustc_codegen_ssa/back/archive.rs
new file mode 100644
index 00000000000..b5e1deb0d5d
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/archive.rs
@@ -0,0 +1,36 @@
+// Copyright 2018 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use rustc::session::Session;
+
+use std::path::PathBuf;
+
+pub fn find_library(name: &str, search_paths: &[PathBuf], sess: &Session)
+                    -> PathBuf {
+    // On Windows, static libraries sometimes show up as libfoo.a and other
+    // times show up as foo.lib
+    let oslibname = format!("{}{}{}",
+                            sess.target.target.options.staticlib_prefix,
+                            name,
+                            sess.target.target.options.staticlib_suffix);
+    let unixlibname = format!("lib{}.a", name);
+
+    for path in search_paths {
+        debug!("looking for {} inside {:?}", name, path);
+        let test = path.join(&oslibname);
+        if test.exists() { return test }
+        if oslibname != unixlibname {
+            let test = path.join(&unixlibname);
+            if test.exists() { return test }
+        }
+    }
+    sess.fatal(&format!("could not find native static library `{}`, \
+                         perhaps an -L flag is missing?", name));
+}
diff --git a/src/librustc_codegen_ssa/back/command.rs b/src/librustc_codegen_ssa/back/command.rs
new file mode 100644
index 00000000000..9ebbdd7c3c9
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/command.rs
@@ -0,0 +1,175 @@
+// Copyright 2017 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+//! A thin wrapper around `Command` in the standard library which allows us to
+//! read the arguments that are built up.
+
+use std::ffi::{OsStr, OsString};
+use std::fmt;
+use std::io;
+use std::mem;
+use std::process::{self, Output};
+
+use rustc_target::spec::LldFlavor;
+
+#[derive(Clone)]
+pub struct Command {
+    program: Program,
+    args: Vec<OsString>,
+    env: Vec<(OsString, OsString)>,
+}
+
+#[derive(Clone)]
+enum Program {
+    Normal(OsString),
+    CmdBatScript(OsString),
+    Lld(OsString, LldFlavor)
+}
+
+impl Command {
+    pub fn new<P: AsRef<OsStr>>(program: P) -> Command {
+        Command::_new(Program::Normal(program.as_ref().to_owned()))
+    }
+
+    pub fn bat_script<P: AsRef<OsStr>>(program: P) -> Command {
+        Command::_new(Program::CmdBatScript(program.as_ref().to_owned()))
+    }
+
+    pub fn lld<P: AsRef<OsStr>>(program: P, flavor: LldFlavor) -> Command {
+        Command::_new(Program::Lld(program.as_ref().to_owned(), flavor))
+    }
+
+    fn _new(program: Program) -> Command {
+        Command {
+            program,
+            args: Vec::new(),
+            env: Vec::new(),
+        }
+    }
+
+    pub fn arg<P: AsRef<OsStr>>(&mut self, arg: P) -> &mut Command {
+        self._arg(arg.as_ref());
+        self
+    }
+
+    pub fn args<I>(&mut self, args: I) -> &mut Command
+        where I: IntoIterator,
+              I::Item: AsRef<OsStr>,
+    {
+        for arg in args {
+            self._arg(arg.as_ref());
+        }
+        self
+    }
+
+    fn _arg(&mut self, arg: &OsStr) {
+        self.args.push(arg.to_owned());
+    }
+
+    pub fn env<K, V>(&mut self, key: K, value: V) -> &mut Command
+        where K: AsRef<OsStr>,
+              V: AsRef<OsStr>
+    {
+        self._env(key.as_ref(), value.as_ref());
+        self
+    }
+
+    fn _env(&mut self, key: &OsStr, value: &OsStr) {
+        self.env.push((key.to_owned(), value.to_owned()));
+    }
+
+    pub fn output(&mut self) -> io::Result<Output> {
+        self.command().output()
+    }
+
+    pub fn command(&self) -> process::Command {
+        let mut ret = match self.program {
+            Program::Normal(ref p) => process::Command::new(p),
+            Program::CmdBatScript(ref p) => {
+                let mut c = process::Command::new("cmd");
+                c.arg("/c").arg(p);
+                c
+            }
+            Program::Lld(ref p, flavor) => {
+                let mut c = process::Command::new(p);
+                c.arg("-flavor").arg(match flavor {
+                    LldFlavor::Wasm => "wasm",
+                    LldFlavor::Ld => "gnu",
+                    LldFlavor::Link => "link",
+                    LldFlavor::Ld64 => "darwin",
+                });
+                c
+            }
+        };
+        ret.args(&self.args);
+        ret.envs(self.env.clone());
+        return ret
+    }
+
+    // extensions
+
+    pub fn get_args(&self) -> &[OsString] {
+        &self.args
+    }
+
+    pub fn take_args(&mut self) -> Vec<OsString> {
+        mem::replace(&mut self.args, Vec::new())
+    }
+
+    /// Returns a `true` if we're pretty sure that this'll blow OS spawn limits,
+    /// or `false` if we should attempt to spawn and see what the OS says.
+    pub fn very_likely_to_exceed_some_spawn_limit(&self) -> bool {
+        // We mostly only care about Windows in this method, on Unix the limits
+        // can be gargantuan anyway so we're pretty unlikely to hit them
+        if cfg!(unix) {
+            return false
+        }
+
+        // Right now LLD doesn't support the `@` syntax of passing an argument
+        // through files, so regardless of the platform we try to go to the OS
+        // on this one.
+        if let Program::Lld(..) = self.program {
+            return false
+        }
+
+        // Ok so on Windows to spawn a process is 32,768 characters in its
+        // command line [1]. Unfortunately we don't actually have access to that
+        // as it's calculated just before spawning. Instead we perform a
+        // poor-man's guess as to how long our command line will be. We're
+        // assuming here that we don't have to escape every character...
+        //
+        // Turns out though that `cmd.exe` has even smaller limits, 8192
+        // characters [2]. Linkers can often be batch scripts (for example
+        // Emscripten, Gecko's current build system) which means that we're
+        // running through batch scripts. These linkers often just forward
+        // arguments elsewhere (and maybe tack on more), so if we blow 8192
+        // bytes we'll typically cause them to blow as well.
+        //
+        // Basically as a result just perform an inflated estimate of what our
+        // command line will look like and test if it's > 8192 (we actually
+        // test against 6k to artificially inflate our estimate). If all else
+        // fails we'll fall back to the normal unix logic of testing the OS
+        // error code if we fail to spawn and automatically re-spawning the
+        // linker with smaller arguments.
+        //
+        // [1]: https://msdn.microsoft.com/en-us/library/windows/desktop/ms682425(v=vs.85).aspx
+        // [2]: https://blogs.msdn.microsoft.com/oldnewthing/20031210-00/?p=41553
+
+        let estimated_command_line_len =
+            self.args.iter().map(|a| a.len()).sum::<usize>();
+        estimated_command_line_len > 1024 * 6
+    }
+}
+
+impl fmt::Debug for Command {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        self.command().fmt(f)
+    }
+}
diff --git a/src/librustc_codegen_ssa/back/link.rs b/src/librustc_codegen_ssa/back/link.rs
new file mode 100644
index 00000000000..b0575b841d5
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/link.rs
@@ -0,0 +1,208 @@
+// Copyright 2018 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+/// For all the linkers we support, and information they might
+/// need out of the shared crate context before we get rid of it.
+
+use rustc::session::{Session, config};
+use rustc::session::search_paths::PathKind;
+use rustc::middle::dependency_format::Linkage;
+use rustc::middle::cstore::LibSource;
+use rustc_target::spec::LinkerFlavor;
+use rustc::hir::def_id::CrateNum;
+
+use super::command::Command;
+use CrateInfo;
+
+use cc::windows_registry;
+use std::fs;
+use std::path::{Path, PathBuf};
+use std::env;
+
+pub fn remove(sess: &Session, path: &Path) {
+    if let Err(e) = fs::remove_file(path) {
+        sess.err(&format!("failed to remove {}: {}",
+                          path.display(),
+                          e));
+    }
+}
+
+// The third parameter is for env vars, used on windows to set up the
+// path for MSVC to find its DLLs, and gcc to find its bundled
+// toolchain
+pub fn get_linker(sess: &Session, linker: &Path, flavor: LinkerFlavor) -> (PathBuf, Command) {
+    let msvc_tool = windows_registry::find_tool(&sess.opts.target_triple.triple(), "link.exe");
+
+    // If our linker looks like a batch script on Windows then to execute this
+    // we'll need to spawn `cmd` explicitly. This is primarily done to handle
+    // emscripten where the linker is `emcc.bat` and needs to be spawned as
+    // `cmd /c emcc.bat ...`.
+    //
+    // This worked historically but is needed manually since #42436 (regression
+    // was tagged as #42791) and some more info can be found on #44443 for
+    // emscripten itself.
+    let mut cmd = match linker.to_str() {
+        Some(linker) if cfg!(windows) && linker.ends_with(".bat") => Command::bat_script(linker),
+        _ => match flavor {
+            LinkerFlavor::Lld(f) => Command::lld(linker, f),
+            LinkerFlavor::Msvc
+                if sess.opts.cg.linker.is_none() && sess.target.target.options.linker.is_none() =>
+            {
+                Command::new(msvc_tool.as_ref().map(|t| t.path()).unwrap_or(linker))
+            },
+            _ => Command::new(linker),
+        }
+    };
+
+    // The compiler's sysroot often has some bundled tools, so add it to the
+    // PATH for the child.
+    let mut new_path = sess.host_filesearch(PathKind::All)
+                           .get_tools_search_paths();
+    let mut msvc_changed_path = false;
+    if sess.target.target.options.is_like_msvc {
+        if let Some(ref tool) = msvc_tool {
+            cmd.args(tool.args());
+            for &(ref k, ref v) in tool.env() {
+                if k == "PATH" {
+                    new_path.extend(env::split_paths(v));
+                    msvc_changed_path = true;
+                } else {
+                    cmd.env(k, v);
+                }
+            }
+        }
+    }
+
+    if !msvc_changed_path {
+        if let Some(path) = env::var_os("PATH") {
+            new_path.extend(env::split_paths(&path));
+        }
+    }
+    cmd.env("PATH", env::join_paths(new_path).unwrap());
+
+    (linker.to_path_buf(), cmd)
+}
+
+pub fn each_linked_rlib(sess: &Session,
+                               info: &CrateInfo,
+                               f: &mut dyn FnMut(CrateNum, &Path)) -> Result<(), String> {
+    let crates = info.used_crates_static.iter();
+    let fmts = sess.dependency_formats.borrow();
+    let fmts = fmts.get(&config::CrateType::Executable)
+                   .or_else(|| fmts.get(&config::CrateType::Staticlib))
+                   .or_else(|| fmts.get(&config::CrateType::Cdylib))
+                   .or_else(|| fmts.get(&config::CrateType::ProcMacro));
+    let fmts = match fmts {
+        Some(f) => f,
+        None => return Err("could not find formats for rlibs".to_string())
+    };
+    for &(cnum, ref path) in crates {
+        match fmts.get(cnum.as_usize() - 1) {
+            Some(&Linkage::NotLinked) |
+            Some(&Linkage::IncludedFromDylib) => continue,
+            Some(_) => {}
+            None => return Err("could not find formats for rlibs".to_string())
+        }
+        let name = &info.crate_name[&cnum];
+        let path = match *path {
+            LibSource::Some(ref p) => p,
+            LibSource::MetadataOnly => {
+                return Err(format!("could not find rlib for: `{}`, found rmeta (metadata) file",
+                                   name))
+            }
+            LibSource::None => {
+                return Err(format!("could not find rlib for: `{}`", name))
+            }
+        };
+        f(cnum, &path);
+    }
+    Ok(())
+}
+
+/// Returns a boolean indicating whether the specified crate should be ignored
+/// during LTO.
+///
+/// Crates ignored during LTO are not lumped together in the "massive object
+/// file" that we create and are linked in their normal rlib states. See
+/// comments below for what crates do not participate in LTO.
+///
+/// It's unusual for a crate to not participate in LTO. Typically only
+/// compiler-specific and unstable crates have a reason to not participate in
+/// LTO.
+pub fn ignored_for_lto(sess: &Session, info: &CrateInfo, cnum: CrateNum) -> bool {
+    // If our target enables builtin function lowering in LLVM then the
+    // crates providing these functions don't participate in LTO (e.g.
+    // no_builtins or compiler builtins crates).
+    !sess.target.target.options.no_builtins &&
+        (info.compiler_builtins == Some(cnum) || info.is_no_builtins.contains(&cnum))
+}
+
+pub fn linker_and_flavor(sess: &Session) -> (PathBuf, LinkerFlavor) {
+    fn infer_from(
+        sess: &Session,
+        linker: Option<PathBuf>,
+        flavor: Option<LinkerFlavor>,
+    ) -> Option<(PathBuf, LinkerFlavor)> {
+        match (linker, flavor) {
+            (Some(linker), Some(flavor)) => Some((linker, flavor)),
+            // only the linker flavor is known; use the default linker for the selected flavor
+            (None, Some(flavor)) => Some((PathBuf::from(match flavor {
+                LinkerFlavor::Em  => if cfg!(windows) { "emcc.bat" } else { "emcc" },
+                LinkerFlavor::Gcc => "cc",
+                LinkerFlavor::Ld => "ld",
+                LinkerFlavor::Msvc => "link.exe",
+                LinkerFlavor::Lld(_) => "lld",
+            }), flavor)),
+            (Some(linker), None) => {
+                let stem = linker.file_stem().and_then(|stem| stem.to_str()).unwrap_or_else(|| {
+                    sess.fatal("couldn't extract file stem from specified linker");
+                }).to_owned();
+
+                let flavor = if stem == "emcc" {
+                    LinkerFlavor::Em
+                } else if stem == "gcc" || stem.ends_with("-gcc") {
+                    LinkerFlavor::Gcc
+                } else if stem == "ld" || stem == "ld.lld" || stem.ends_with("-ld") {
+                    LinkerFlavor::Ld
+                } else if stem == "link" || stem == "lld-link" {
+                    LinkerFlavor::Msvc
+                } else if stem == "lld" || stem == "rust-lld" {
+                    LinkerFlavor::Lld(sess.target.target.options.lld_flavor)
+                } else {
+                    // fall back to the value in the target spec
+                    sess.target.target.linker_flavor
+                };
+
+                Some((linker, flavor))
+            },
+            (None, None) => None,
+        }
+    }
+
+    // linker and linker flavor specified via command line have precedence over what the target
+    // specification specifies
+    if let Some(ret) = infer_from(
+        sess,
+        sess.opts.cg.linker.clone(),
+        sess.opts.debugging_opts.linker_flavor,
+    ) {
+        return ret;
+    }
+
+    if let Some(ret) = infer_from(
+        sess,
+        sess.target.target.options.linker.clone().map(PathBuf::from),
+        Some(sess.target.target.linker_flavor),
+    ) {
+        return ret;
+    }
+
+    bug!("Not enough information provided to determine how to invoke the linker");
+}
diff --git a/src/librustc_codegen_ssa/back/linker.rs b/src/librustc_codegen_ssa/back/linker.rs
new file mode 100644
index 00000000000..da9cfbb94d1
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/linker.rs
@@ -0,0 +1,1079 @@
+// Copyright 2015 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use super::symbol_export;
+use super::command::Command;
+use super::archive;
+
+use rustc_data_structures::fx::FxHashMap;
+use std::ffi::{OsStr, OsString};
+use std::fs::{self, File};
+use std::io::prelude::*;
+use std::io::{self, BufWriter};
+use std::path::{Path, PathBuf};
+
+use rustc::hir::def_id::{LOCAL_CRATE, CrateNum};
+use rustc::middle::dependency_format::Linkage;
+use rustc::session::Session;
+use rustc::session::config::{self, CrateType, OptLevel, DebugInfo,
+                             CrossLangLto};
+use rustc::ty::TyCtxt;
+use rustc_target::spec::{LinkerFlavor, LldFlavor};
+use serialize::{json, Encoder};
+
+/// For all the linkers we support, and information they might
+/// need out of the shared crate context before we get rid of it.
+pub struct LinkerInfo {
+    exports: FxHashMap<CrateType, Vec<String>>,
+}
+
+impl LinkerInfo {
+    pub fn new(tcx: TyCtxt) -> LinkerInfo {
+        LinkerInfo {
+            exports: tcx.sess.crate_types.borrow().iter().map(|&c| {
+                (c, exported_symbols(tcx, c))
+            }).collect(),
+        }
+    }
+
+    pub fn to_linker<'a>(
+        &'a self,
+        cmd: Command,
+        sess: &'a Session,
+        flavor: LinkerFlavor,
+        target_cpu: &'a str,
+    ) -> Box<dyn Linker+'a> {
+        match flavor {
+            LinkerFlavor::Lld(LldFlavor::Link) |
+            LinkerFlavor::Msvc => {
+                Box::new(MsvcLinker {
+                    cmd,
+                    sess,
+                    info: self
+                }) as Box<dyn Linker>
+            }
+            LinkerFlavor::Em =>  {
+                Box::new(EmLinker {
+                    cmd,
+                    sess,
+                    info: self
+                }) as Box<dyn Linker>
+            }
+            LinkerFlavor::Gcc =>  {
+                Box::new(GccLinker {
+                    cmd,
+                    sess,
+                    info: self,
+                    hinted_static: false,
+                    is_ld: false,
+                    target_cpu,
+                }) as Box<dyn Linker>
+            }
+
+            LinkerFlavor::Lld(LldFlavor::Ld) |
+            LinkerFlavor::Lld(LldFlavor::Ld64) |
+            LinkerFlavor::Ld => {
+                Box::new(GccLinker {
+                    cmd,
+                    sess,
+                    info: self,
+                    hinted_static: false,
+                    is_ld: true,
+                    target_cpu,
+                }) as Box<dyn Linker>
+            }
+
+            LinkerFlavor::Lld(LldFlavor::Wasm) => {
+                Box::new(WasmLd {
+                    cmd,
+                    sess,
+                    info: self
+                }) as Box<dyn Linker>
+            }
+        }
+    }
+}
+
+/// Linker abstraction used by back::link to build up the command to invoke a
+/// linker.
+///
+/// This trait is the total list of requirements needed by `back::link` and
+/// represents the meaning of each option being passed down. This trait is then
+/// used to dispatch on whether a GNU-like linker (generally `ld.exe`) or an
+/// MSVC linker (e.g. `link.exe`) is being used.
+pub trait Linker {
+    fn link_dylib(&mut self, lib: &str);
+    fn link_rust_dylib(&mut self, lib: &str, path: &Path);
+    fn link_framework(&mut self, framework: &str);
+    fn link_staticlib(&mut self, lib: &str);
+    fn link_rlib(&mut self, lib: &Path);
+    fn link_whole_rlib(&mut self, lib: &Path);
+    fn link_whole_staticlib(&mut self, lib: &str, search_path: &[PathBuf]);
+    fn include_path(&mut self, path: &Path);
+    fn framework_path(&mut self, path: &Path);
+    fn output_filename(&mut self, path: &Path);
+    fn add_object(&mut self, path: &Path);
+    fn gc_sections(&mut self, keep_metadata: bool);
+    fn position_independent_executable(&mut self);
+    fn no_position_independent_executable(&mut self);
+    fn full_relro(&mut self);
+    fn partial_relro(&mut self);
+    fn no_relro(&mut self);
+    fn optimize(&mut self);
+    fn pgo_gen(&mut self);
+    fn debuginfo(&mut self);
+    fn no_default_libraries(&mut self);
+    fn build_dylib(&mut self, out_filename: &Path);
+    fn build_static_executable(&mut self);
+    fn args(&mut self, args: &[String]);
+    fn export_symbols(&mut self, tmpdir: &Path, crate_type: CrateType);
+    fn subsystem(&mut self, subsystem: &str);
+    fn group_start(&mut self);
+    fn group_end(&mut self);
+    fn cross_lang_lto(&mut self);
+    // Should have been finalize(self), but we don't support self-by-value on trait objects (yet?).
+    fn finalize(&mut self) -> Command;
+}
+
+pub struct GccLinker<'a> {
+    cmd: Command,
+    sess: &'a Session,
+    info: &'a LinkerInfo,
+    hinted_static: bool, // Keeps track of the current hinting mode.
+    // Link as ld
+    is_ld: bool,
+    target_cpu: &'a str,
+}
+
+impl<'a> GccLinker<'a> {
+    /// Argument that must be passed *directly* to the linker
+    ///
+    /// These arguments need to be prepended with '-Wl,' when a gcc-style linker is used
+    fn linker_arg<S>(&mut self, arg: S) -> &mut Self
+        where S: AsRef<OsStr>
+    {
+        if !self.is_ld {
+            let mut os = OsString::from("-Wl,");
+            os.push(arg.as_ref());
+            self.cmd.arg(os);
+        } else {
+            self.cmd.arg(arg);
+        }
+        self
+    }
+
+    fn takes_hints(&self) -> bool {
+        !self.sess.target.target.options.is_like_osx
+    }
+
+    // Some platforms take hints about whether a library is static or dynamic.
+    // For those that support this, we ensure we pass the option if the library
+    // was flagged "static" (most defaults are dynamic) to ensure that if
+    // libfoo.a and libfoo.so both exist that the right one is chosen.
+    fn hint_static(&mut self) {
+        if !self.takes_hints() { return }
+        if !self.hinted_static {
+            self.linker_arg("-Bstatic");
+            self.hinted_static = true;
+        }
+    }
+
+    fn hint_dynamic(&mut self) {
+        if !self.takes_hints() { return }
+        if self.hinted_static {
+            self.linker_arg("-Bdynamic");
+            self.hinted_static = false;
+        }
+    }
+
+    fn push_cross_lang_lto_args(&mut self, plugin_path: Option<&OsStr>) {
+        if let Some(plugin_path) = plugin_path {
+            let mut arg = OsString::from("-plugin=");
+            arg.push(plugin_path);
+            self.linker_arg(&arg);
+        }
+
+        let opt_level = match self.sess.opts.optimize {
+            config::OptLevel::No => "O0",
+            config::OptLevel::Less => "O1",
+            config::OptLevel::Default => "O2",
+            config::OptLevel::Aggressive => "O3",
+            config::OptLevel::Size => "Os",
+            config::OptLevel::SizeMin => "Oz",
+        };
+
+        self.linker_arg(&format!("-plugin-opt={}", opt_level));
+        let target_cpu = self.target_cpu;
+        self.linker_arg(&format!("-plugin-opt=mcpu={}", target_cpu));
+    }
+}
+
+impl<'a> Linker for GccLinker<'a> {
+    fn link_dylib(&mut self, lib: &str) { self.hint_dynamic(); self.cmd.arg(format!("-l{}", lib)); }
+    fn link_staticlib(&mut self, lib: &str) {
+        self.hint_static(); self.cmd.arg(format!("-l{}", lib));
+    }
+    fn link_rlib(&mut self, lib: &Path) { self.hint_static(); self.cmd.arg(lib); }
+    fn include_path(&mut self, path: &Path) { self.cmd.arg("-L").arg(path); }
+    fn framework_path(&mut self, path: &Path) { self.cmd.arg("-F").arg(path); }
+    fn output_filename(&mut self, path: &Path) { self.cmd.arg("-o").arg(path); }
+    fn add_object(&mut self, path: &Path) { self.cmd.arg(path); }
+    fn position_independent_executable(&mut self) { self.cmd.arg("-pie"); }
+    fn no_position_independent_executable(&mut self) { self.cmd.arg("-no-pie"); }
+    fn full_relro(&mut self) { self.linker_arg("-zrelro"); self.linker_arg("-znow"); }
+    fn partial_relro(&mut self) { self.linker_arg("-zrelro"); }
+    fn no_relro(&mut self) { self.linker_arg("-znorelro"); }
+    fn build_static_executable(&mut self) { self.cmd.arg("-static"); }
+    fn args(&mut self, args: &[String]) { self.cmd.args(args); }
+
+    fn link_rust_dylib(&mut self, lib: &str, _path: &Path) {
+        self.hint_dynamic();
+        self.cmd.arg(format!("-l{}", lib));
+    }
+
+    fn link_framework(&mut self, framework: &str) {
+        self.hint_dynamic();
+        self.cmd.arg("-framework").arg(framework);
+    }
+
+    // Here we explicitly ask that the entire archive is included into the
+    // result artifact. For more details see #15460, but the gist is that
+    // the linker will strip away any unused objects in the archive if we
+    // don't otherwise explicitly reference them. This can occur for
+    // libraries which are just providing bindings, libraries with generic
+    // functions, etc.
+    fn link_whole_staticlib(&mut self, lib: &str, search_path: &[PathBuf]) {
+        self.hint_static();
+        let target = &self.sess.target.target;
+        if !target.options.is_like_osx {
+            self.linker_arg("--whole-archive").cmd.arg(format!("-l{}", lib));
+            self.linker_arg("--no-whole-archive");
+        } else {
+            // -force_load is the macOS equivalent of --whole-archive, but it
+            // involves passing the full path to the library to link.
+            self.linker_arg("-force_load");
+            let lib = archive::find_library(lib, search_path, &self.sess);
+            self.linker_arg(&lib);
+        }
+    }
+
+    fn link_whole_rlib(&mut self, lib: &Path) {
+        self.hint_static();
+        if self.sess.target.target.options.is_like_osx {
+            self.linker_arg("-force_load");
+            self.linker_arg(&lib);
+        } else {
+            self.linker_arg("--whole-archive").cmd.arg(lib);
+            self.linker_arg("--no-whole-archive");
+        }
+    }
+
+    fn gc_sections(&mut self, keep_metadata: bool) {
+        // The dead_strip option to the linker specifies that functions and data
+        // unreachable by the entry point will be removed. This is quite useful
+        // with Rust's compilation model of compiling libraries at a time into
+        // one object file. For example, this brings hello world from 1.7MB to
+        // 458K.
+        //
+        // Note that this is done for both executables and dynamic libraries. We
+        // won't get much benefit from dylibs because LLVM will have already
+        // stripped away as much as it could. This has not been seen to impact
+        // link times negatively.
+        //
+        // -dead_strip can't be part of the pre_link_args because it's also used
+        // for partial linking when using multiple codegen units (-r).  So we
+        // insert it here.
+        if self.sess.target.target.options.is_like_osx {
+            self.linker_arg("-dead_strip");
+        } else if self.sess.target.target.options.is_like_solaris {
+            self.linker_arg("-zignore");
+
+        // If we're building a dylib, we don't use --gc-sections because LLVM
+        // has already done the best it can do, and we also don't want to
+        // eliminate the metadata. If we're building an executable, however,
+        // --gc-sections drops the size of hello world from 1.8MB to 597K, a 67%
+        // reduction.
+        } else if !keep_metadata {
+            self.linker_arg("--gc-sections");
+        }
+    }
+
+    fn optimize(&mut self) {
+        if !self.sess.target.target.options.linker_is_gnu { return }
+
+        // GNU-style linkers support optimization with -O. GNU ld doesn't
+        // need a numeric argument, but other linkers do.
+        if self.sess.opts.optimize == config::OptLevel::Default ||
+           self.sess.opts.optimize == config::OptLevel::Aggressive {
+            self.linker_arg("-O1");
+        }
+    }
+
+    fn pgo_gen(&mut self) {
+        if !self.sess.target.target.options.linker_is_gnu { return }
+
+        // If we're doing PGO generation stuff and on a GNU-like linker, use the
+        // "-u" flag to properly pull in the profiler runtime bits.
+        //
+        // This is because LLVM otherwise won't add the needed initialization
+        // for us on Linux (though the extra flag should be harmless if it
+        // does).
+        //
+        // See https://reviews.llvm.org/D14033 and https://reviews.llvm.org/D14030.
+        //
+        // Though it may be worth to try to revert those changes upstream, since
+        // the overhead of the initialization should be minor.
+        self.cmd.arg("-u");
+        self.cmd.arg("__llvm_profile_runtime");
+    }
+
+    fn debuginfo(&mut self) {
+        if let DebugInfo::None = self.sess.opts.debuginfo {
+            // If we are building without debuginfo enabled and we were called with
+            // `-Zstrip-debuginfo-if-disabled=yes`, tell the linker to strip any debuginfo
+            // found when linking to get rid of symbols from libstd.
+            if let Some(true) = self.sess.opts.debugging_opts.strip_debuginfo_if_disabled {
+                self.linker_arg("-S");
+            }
+        };
+    }
+
+    fn no_default_libraries(&mut self) {
+        if !self.is_ld {
+            self.cmd.arg("-nodefaultlibs");
+        }
+    }
+
+    fn build_dylib(&mut self, out_filename: &Path) {
+        // On mac we need to tell the linker to let this library be rpathed
+        if self.sess.target.target.options.is_like_osx {
+            self.cmd.arg("-dynamiclib");
+            self.linker_arg("-dylib");
+
+            // Note that the `osx_rpath_install_name` option here is a hack
+            // purely to support rustbuild right now, we should get a more
+            // principled solution at some point to force the compiler to pass
+            // the right `-Wl,-install_name` with an `@rpath` in it.
+            if self.sess.opts.cg.rpath || self.sess.opts.debugging_opts.osx_rpath_install_name {
+                self.linker_arg("-install_name");
+                let mut v = OsString::from("@rpath/");
+                v.push(out_filename.file_name().unwrap());
+                self.linker_arg(&v);
+            }
+        } else {
+            self.cmd.arg("-shared");
+        }
+    }
+
+    fn export_symbols(&mut self, tmpdir: &Path, crate_type: CrateType) {
+        // If we're compiling a dylib, then we let symbol visibility in object
+        // files to take care of whether they're exported or not.
+        //
+        // If we're compiling a cdylib, however, we manually create a list of
+        // exported symbols to ensure we don't expose any more. The object files
+        // have far more public symbols than we actually want to export, so we
+        // hide them all here.
+        if crate_type == CrateType::Dylib ||
+           crate_type == CrateType::ProcMacro {
+            return
+        }
+
+        let mut arg = OsString::new();
+        let path = tmpdir.join("list");
+
+        debug!("EXPORTED SYMBOLS:");
+
+        if self.sess.target.target.options.is_like_osx {
+            // Write a plain, newline-separated list of symbols
+            let res = (|| -> io::Result<()> {
+                let mut f = BufWriter::new(File::create(&path)?);
+                for sym in self.info.exports[&crate_type].iter() {
+                    debug!("  _{}", sym);
+                    writeln!(f, "_{}", sym)?;
+                }
+                Ok(())
+            })();
+            if let Err(e) = res {
+                self.sess.fatal(&format!("failed to write lib.def file: {}", e));
+            }
+        } else {
+            // Write an LD version script
+            let res = (|| -> io::Result<()> {
+                let mut f = BufWriter::new(File::create(&path)?);
+                writeln!(f, "{{\n  global:")?;
+                for sym in self.info.exports[&crate_type].iter() {
+                    debug!("    {};", sym);
+                    writeln!(f, "    {};", sym)?;
+                }
+                writeln!(f, "\n  local:\n    *;\n}};")?;
+                Ok(())
+            })();
+            if let Err(e) = res {
+                self.sess.fatal(&format!("failed to write version script: {}", e));
+            }
+        }
+
+        if self.sess.target.target.options.is_like_osx {
+            if !self.is_ld {
+                arg.push("-Wl,")
+            }
+            arg.push("-exported_symbols_list,");
+        } else if self.sess.target.target.options.is_like_solaris {
+            if !self.is_ld {
+                arg.push("-Wl,")
+            }
+            arg.push("-M,");
+        } else {
+            if !self.is_ld {
+                arg.push("-Wl,")
+            }
+            arg.push("--version-script=");
+        }
+
+        arg.push(&path);
+        self.cmd.arg(arg);
+    }
+
+    fn subsystem(&mut self, subsystem: &str) {
+        self.linker_arg("--subsystem");
+        self.linker_arg(&subsystem);
+    }
+
+    fn finalize(&mut self) -> Command {
+        self.hint_dynamic(); // Reset to default before returning the composed command line.
+
+        ::std::mem::replace(&mut self.cmd, Command::new(""))
+    }
+
+    fn group_start(&mut self) {
+        if !self.sess.target.target.options.is_like_osx {
+            self.linker_arg("--start-group");
+        }
+    }
+
+    fn group_end(&mut self) {
+        if !self.sess.target.target.options.is_like_osx {
+            self.linker_arg("--end-group");
+        }
+    }
+
+    fn cross_lang_lto(&mut self) {
+        match self.sess.opts.debugging_opts.cross_lang_lto {
+            CrossLangLto::Disabled => {
+                // Nothing to do
+            }
+            CrossLangLto::LinkerPluginAuto => {
+                self.push_cross_lang_lto_args(None);
+            }
+            CrossLangLto::LinkerPlugin(ref path) => {
+                self.push_cross_lang_lto_args(Some(path.as_os_str()));
+            }
+        }
+    }
+}
+
+pub struct MsvcLinker<'a> {
+    cmd: Command,
+    sess: &'a Session,
+    info: &'a LinkerInfo
+}
+
+impl<'a> Linker for MsvcLinker<'a> {
+    fn link_rlib(&mut self, lib: &Path) { self.cmd.arg(lib); }
+    fn add_object(&mut self, path: &Path) { self.cmd.arg(path); }
+    fn args(&mut self, args: &[String]) { self.cmd.args(args); }
+
+    fn build_dylib(&mut self, out_filename: &Path) {
+        self.cmd.arg("/DLL");
+        let mut arg: OsString = "/IMPLIB:".into();
+        arg.push(out_filename.with_extension("dll.lib"));
+        self.cmd.arg(arg);
+    }
+
+    fn build_static_executable(&mut self) {
+        // noop
+    }
+
+    fn gc_sections(&mut self, _keep_metadata: bool) {
+        // MSVC's ICF (Identical COMDAT Folding) link optimization is
+        // slow for Rust and thus we disable it by default when not in
+        // optimization build.
+        if self.sess.opts.optimize != config::OptLevel::No {
+            self.cmd.arg("/OPT:REF,ICF");
+        } else {
+            // It is necessary to specify NOICF here, because /OPT:REF
+            // implies ICF by default.
+            self.cmd.arg("/OPT:REF,NOICF");
+        }
+    }
+
+    fn link_dylib(&mut self, lib: &str) {
+        self.cmd.arg(&format!("{}.lib", lib));
+    }
+
+    fn link_rust_dylib(&mut self, lib: &str, path: &Path) {
+        // When producing a dll, the MSVC linker may not actually emit a
+        // `foo.lib` file if the dll doesn't actually export any symbols, so we
+        // check to see if the file is there and just omit linking to it if it's
+        // not present.
+        let name = format!("{}.dll.lib", lib);
+        if fs::metadata(&path.join(&name)).is_ok() {
+            self.cmd.arg(name);
+        }
+    }
+
+    fn link_staticlib(&mut self, lib: &str) {
+        self.cmd.arg(&format!("{}.lib", lib));
+    }
+
+    fn position_independent_executable(&mut self) {
+        // noop
+    }
+
+    fn no_position_independent_executable(&mut self) {
+        // noop
+    }
+
+    fn full_relro(&mut self) {
+        // noop
+    }
+
+    fn partial_relro(&mut self) {
+        // noop
+    }
+
+    fn no_relro(&mut self) {
+        // noop
+    }
+
+    fn no_default_libraries(&mut self) {
+        // Currently we don't pass the /NODEFAULTLIB flag to the linker on MSVC
+        // as there's been trouble in the past of linking the C++ standard
+        // library required by LLVM. This likely needs to happen one day, but
+        // in general Windows is also a more controlled environment than
+        // Unix, so it's not necessarily as critical that this be implemented.
+        //
+        // Note that there are also some licensing worries about statically
+        // linking some libraries which require a specific agreement, so it may
+        // not ever be possible for us to pass this flag.
+    }
+
+    fn include_path(&mut self, path: &Path) {
+        let mut arg = OsString::from("/LIBPATH:");
+        arg.push(path);
+        self.cmd.arg(&arg);
+    }
+
+    fn output_filename(&mut self, path: &Path) {
+        let mut arg = OsString::from("/OUT:");
+        arg.push(path);
+        self.cmd.arg(&arg);
+    }
+
+    fn framework_path(&mut self, _path: &Path) {
+        bug!("frameworks are not supported on windows")
+    }
+    fn link_framework(&mut self, _framework: &str) {
+        bug!("frameworks are not supported on windows")
+    }
+
+    fn link_whole_staticlib(&mut self, lib: &str, _search_path: &[PathBuf]) {
+        // not supported?
+        self.link_staticlib(lib);
+    }
+    fn link_whole_rlib(&mut self, path: &Path) {
+        // not supported?
+        self.link_rlib(path);
+    }
+    fn optimize(&mut self) {
+        // Needs more investigation of `/OPT` arguments
+    }
+
+    fn pgo_gen(&mut self) {
+        // Nothing needed here.
+    }
+
+    fn debuginfo(&mut self) {
+        // This will cause the Microsoft linker to generate a PDB file
+        // from the CodeView line tables in the object files.
+        self.cmd.arg("/DEBUG");
+
+        // This will cause the Microsoft linker to embed .natvis info into the PDB file
+        let sysroot = self.sess.sysroot();
+        let natvis_dir_path = sysroot.join("lib\\rustlib\\etc");
+        if let Ok(natvis_dir) = fs::read_dir(&natvis_dir_path) {
+            // LLVM 5.0.0's lld-link frontend doesn't yet recognize, and chokes
+            // on, the /NATVIS:... flags.  LLVM 6 (or earlier) should at worst ignore
+            // them, eventually mooting this workaround, per this landed patch:
+            // https://github.com/llvm-mirror/lld/commit/27b9c4285364d8d76bb43839daa100
+            if let Some(ref linker_path) = self.sess.opts.cg.linker {
+                if let Some(linker_name) = Path::new(&linker_path).file_stem() {
+                    if linker_name.to_str().unwrap().to_lowercase() == "lld-link" {
+                        self.sess.warn("not embedding natvis: lld-link may not support the flag");
+                        return;
+                    }
+                }
+            }
+            for entry in natvis_dir {
+                match entry {
+                    Ok(entry) => {
+                        let path = entry.path();
+                        if path.extension() == Some("natvis".as_ref()) {
+                            let mut arg = OsString::from("/NATVIS:");
+                            arg.push(path);
+                            self.cmd.arg(arg);
+                        }
+                    },
+                    Err(err) => {
+                        self.sess.warn(&format!("error enumerating natvis directory: {}", err));
+                    },
+                }
+            }
+        }
+    }
+
+    // Currently the compiler doesn't use `dllexport` (an LLVM attribute) to
+    // export symbols from a dynamic library. When building a dynamic library,
+    // however, we're going to want some symbols exported, so this function
+    // generates a DEF file which lists all the symbols.
+    //
+    // The linker will read this `*.def` file and export all the symbols from
+    // the dynamic library. Note that this is not as simple as just exporting
+    // all the symbols in the current crate (as specified by `codegen.reachable`)
+    // but rather we also need to possibly export the symbols of upstream
+    // crates. Upstream rlibs may be linked statically to this dynamic library,
+    // in which case they may continue to transitively be used and hence need
+    // their symbols exported.
+    fn export_symbols(&mut self,
+                      tmpdir: &Path,
+                      crate_type: CrateType) {
+        let path = tmpdir.join("lib.def");
+        let res = (|| -> io::Result<()> {
+            let mut f = BufWriter::new(File::create(&path)?);
+
+            // Start off with the standard module name header and then go
+            // straight to exports.
+            writeln!(f, "LIBRARY")?;
+            writeln!(f, "EXPORTS")?;
+            for symbol in self.info.exports[&crate_type].iter() {
+                debug!("  _{}", symbol);
+                writeln!(f, "  {}", symbol)?;
+            }
+            Ok(())
+        })();
+        if let Err(e) = res {
+            self.sess.fatal(&format!("failed to write lib.def file: {}", e));
+        }
+        let mut arg = OsString::from("/DEF:");
+        arg.push(path);
+        self.cmd.arg(&arg);
+    }
+
+    fn subsystem(&mut self, subsystem: &str) {
+        // Note that previous passes of the compiler validated this subsystem,
+        // so we just blindly pass it to the linker.
+        self.cmd.arg(&format!("/SUBSYSTEM:{}", subsystem));
+
+        // Windows has two subsystems we're interested in right now, the console
+        // and windows subsystems. These both implicitly have different entry
+        // points (starting symbols). The console entry point starts with
+        // `mainCRTStartup` and the windows entry point starts with
+        // `WinMainCRTStartup`. These entry points, defined in system libraries,
+        // will then later probe for either `main` or `WinMain`, respectively to
+        // start the application.
+        //
+        // In Rust we just always generate a `main` function so we want control
+        // to always start there, so we force the entry point on the windows
+        // subsystem to be `mainCRTStartup` to get everything booted up
+        // correctly.
+        //
+        // For more information see RFC #1665
+        if subsystem == "windows" {
+            self.cmd.arg("/ENTRY:mainCRTStartup");
+        }
+    }
+
+    fn finalize(&mut self) -> Command {
+        ::std::mem::replace(&mut self.cmd, Command::new(""))
+    }
+
+    // MSVC doesn't need group indicators
+    fn group_start(&mut self) {}
+    fn group_end(&mut self) {}
+
+    fn cross_lang_lto(&mut self) {
+        // Do nothing
+    }
+}
+
+pub struct EmLinker<'a> {
+    cmd: Command,
+    sess: &'a Session,
+    info: &'a LinkerInfo
+}
+
+impl<'a> Linker for EmLinker<'a> {
+    fn include_path(&mut self, path: &Path) {
+        self.cmd.arg("-L").arg(path);
+    }
+
+    fn link_staticlib(&mut self, lib: &str) {
+        self.cmd.arg("-l").arg(lib);
+    }
+
+    fn output_filename(&mut self, path: &Path) {
+        self.cmd.arg("-o").arg(path);
+    }
+
+    fn add_object(&mut self, path: &Path) {
+        self.cmd.arg(path);
+    }
+
+    fn link_dylib(&mut self, lib: &str) {
+        // Emscripten always links statically
+        self.link_staticlib(lib);
+    }
+
+    fn link_whole_staticlib(&mut self, lib: &str, _search_path: &[PathBuf]) {
+        // not supported?
+        self.link_staticlib(lib);
+    }
+
+    fn link_whole_rlib(&mut self, lib: &Path) {
+        // not supported?
+        self.link_rlib(lib);
+    }
+
+    fn link_rust_dylib(&mut self, lib: &str, _path: &Path) {
+        self.link_dylib(lib);
+    }
+
+    fn link_rlib(&mut self, lib: &Path) {
+        self.add_object(lib);
+    }
+
+    fn position_independent_executable(&mut self) {
+        // noop
+    }
+
+    fn no_position_independent_executable(&mut self) {
+        // noop
+    }
+
+    fn full_relro(&mut self) {
+        // noop
+    }
+
+    fn partial_relro(&mut self) {
+        // noop
+    }
+
+    fn no_relro(&mut self) {
+        // noop
+    }
+
+    fn args(&mut self, args: &[String]) {
+        self.cmd.args(args);
+    }
+
+    fn framework_path(&mut self, _path: &Path) {
+        bug!("frameworks are not supported on Emscripten")
+    }
+
+    fn link_framework(&mut self, _framework: &str) {
+        bug!("frameworks are not supported on Emscripten")
+    }
+
+    fn gc_sections(&mut self, _keep_metadata: bool) {
+        // noop
+    }
+
+    fn optimize(&mut self) {
+        // Emscripten performs own optimizations
+        self.cmd.arg(match self.sess.opts.optimize {
+            OptLevel::No => "-O0",
+            OptLevel::Less => "-O1",
+            OptLevel::Default => "-O2",
+            OptLevel::Aggressive => "-O3",
+            OptLevel::Size => "-Os",
+            OptLevel::SizeMin => "-Oz"
+        });
+        // Unusable until https://github.com/rust-lang/rust/issues/38454 is resolved
+        self.cmd.args(&["--memory-init-file", "0"]);
+    }
+
+    fn pgo_gen(&mut self) {
+        // noop, but maybe we need something like the gnu linker?
+    }
+
+    fn debuginfo(&mut self) {
+        // Preserve names or generate source maps depending on debug info
+        self.cmd.arg(match self.sess.opts.debuginfo {
+            DebugInfo::None => "-g0",
+            DebugInfo::Limited => "-g3",
+            DebugInfo::Full => "-g4"
+        });
+    }
+
+    fn no_default_libraries(&mut self) {
+        self.cmd.args(&["-s", "DEFAULT_LIBRARY_FUNCS_TO_INCLUDE=[]"]);
+    }
+
+    fn build_dylib(&mut self, _out_filename: &Path) {
+        bug!("building dynamic library is unsupported on Emscripten")
+    }
+
+    fn build_static_executable(&mut self) {
+        // noop
+    }
+
+    fn export_symbols(&mut self, _tmpdir: &Path, crate_type: CrateType) {
+        let symbols = &self.info.exports[&crate_type];
+
+        debug!("EXPORTED SYMBOLS:");
+
+        self.cmd.arg("-s");
+
+        let mut arg = OsString::from("EXPORTED_FUNCTIONS=");
+        let mut encoded = String::new();
+
+        {
+            let mut encoder = json::Encoder::new(&mut encoded);
+            let res = encoder.emit_seq(symbols.len(), |encoder| {
+                for (i, sym) in symbols.iter().enumerate() {
+                    encoder.emit_seq_elt(i, |encoder| {
+                        encoder.emit_str(&("_".to_owned() + sym))
+                    })?;
+                }
+                Ok(())
+            });
+            if let Err(e) = res {
+                self.sess.fatal(&format!("failed to encode exported symbols: {}", e));
+            }
+        }
+        debug!("{}", encoded);
+        arg.push(encoded);
+
+        self.cmd.arg(arg);
+    }
+
+    fn subsystem(&mut self, _subsystem: &str) {
+        // noop
+    }
+
+    fn finalize(&mut self) -> Command {
+        ::std::mem::replace(&mut self.cmd, Command::new(""))
+    }
+
+    // Appears not necessary on Emscripten
+    fn group_start(&mut self) {}
+    fn group_end(&mut self) {}
+
+    fn cross_lang_lto(&mut self) {
+        // Do nothing
+    }
+}
+
+pub struct WasmLd<'a> {
+    cmd: Command,
+    sess: &'a Session,
+    info: &'a LinkerInfo,
+}
+
+impl<'a> Linker for WasmLd<'a> {
+    fn link_dylib(&mut self, lib: &str) {
+        self.cmd.arg("-l").arg(lib);
+    }
+
+    fn link_staticlib(&mut self, lib: &str) {
+        self.cmd.arg("-l").arg(lib);
+    }
+
+    fn link_rlib(&mut self, lib: &Path) {
+        self.cmd.arg(lib);
+    }
+
+    fn include_path(&mut self, path: &Path) {
+        self.cmd.arg("-L").arg(path);
+    }
+
+    fn framework_path(&mut self, _path: &Path) {
+        panic!("frameworks not supported")
+    }
+
+    fn output_filename(&mut self, path: &Path) {
+        self.cmd.arg("-o").arg(path);
+    }
+
+    fn add_object(&mut self, path: &Path) {
+        self.cmd.arg(path);
+    }
+
+    fn position_independent_executable(&mut self) {
+    }
+
+    fn full_relro(&mut self) {
+    }
+
+    fn partial_relro(&mut self) {
+    }
+
+    fn no_relro(&mut self) {
+    }
+
+    fn build_static_executable(&mut self) {
+    }
+
+    fn args(&mut self, args: &[String]) {
+        self.cmd.args(args);
+    }
+
+    fn link_rust_dylib(&mut self, lib: &str, _path: &Path) {
+        self.cmd.arg("-l").arg(lib);
+    }
+
+    fn link_framework(&mut self, _framework: &str) {
+        panic!("frameworks not supported")
+    }
+
+    fn link_whole_staticlib(&mut self, lib: &str, _search_path: &[PathBuf]) {
+        self.cmd.arg("-l").arg(lib);
+    }
+
+    fn link_whole_rlib(&mut self, lib: &Path) {
+        self.cmd.arg(lib);
+    }
+
+    fn gc_sections(&mut self, _keep_metadata: bool) {
+        self.cmd.arg("--gc-sections");
+    }
+
+    fn optimize(&mut self) {
+        self.cmd.arg(match self.sess.opts.optimize {
+            OptLevel::No => "-O0",
+            OptLevel::Less => "-O1",
+            OptLevel::Default => "-O2",
+            OptLevel::Aggressive => "-O3",
+            // Currently LLD doesn't support `Os` and `Oz`, so pass through `O2`
+            // instead.
+            OptLevel::Size => "-O2",
+            OptLevel::SizeMin => "-O2"
+        });
+    }
+
+    fn pgo_gen(&mut self) {
+    }
+
+    fn debuginfo(&mut self) {
+    }
+
+    fn no_default_libraries(&mut self) {
+    }
+
+    fn build_dylib(&mut self, _out_filename: &Path) {
+    }
+
+    fn export_symbols(&mut self, _tmpdir: &Path, crate_type: CrateType) {
+        for sym in self.info.exports[&crate_type].iter() {
+            self.cmd.arg("--export").arg(&sym);
+        }
+    }
+
+    fn subsystem(&mut self, _subsystem: &str) {
+    }
+
+    fn no_position_independent_executable(&mut self) {
+    }
+
+    fn finalize(&mut self) -> Command {
+        // There have been reports in the wild (rustwasm/wasm-bindgen#119) of
+        // using threads causing weird hangs and bugs. Disable it entirely as
+        // this isn't yet the bottleneck of compilation at all anyway.
+        self.cmd.arg("--no-threads");
+
+        // By default LLD only gives us one page of stack (64k) which is a
+        // little small. Default to a larger stack closer to other PC platforms
+        // (1MB) and users can always inject their own link-args to override this.
+        self.cmd.arg("-z").arg("stack-size=1048576");
+
+        // By default LLD's memory layout is:
+        //
+        // 1. First, a blank page
+        // 2. Next, all static data
+        // 3. Finally, the main stack (which grows down)
+        //
+        // This has the unfortunate consequence that on stack overflows you
+        // corrupt static data and can cause some exceedingly weird bugs. To
+        // help detect this a little sooner we instead request that the stack is
+        // placed before static data.
+        //
+        // This means that we'll generate slightly larger binaries as references
+        // to static data will take more bytes in the ULEB128 encoding, but
+        // stack overflow will be guaranteed to trap as it underflows instead of
+        // corrupting static data.
+        self.cmd.arg("--stack-first");
+
+        // FIXME we probably shouldn't pass this but instead pass an explicit
+        // whitelist of symbols we'll allow to be undefined. Unfortunately
+        // though we can't handle symbols like `log10` that LLVM injects at a
+        // super late date without actually parsing object files. For now let's
+        // stick to this and hopefully fix it before stabilization happens.
+        self.cmd.arg("--allow-undefined");
+
+        // For now we just never have an entry symbol
+        self.cmd.arg("--no-entry");
+
+        // Make the default table accessible
+        self.cmd.arg("--export-table");
+
+        // Rust code should never have warnings, and warnings are often
+        // indicative of bugs, let's prevent them.
+        self.cmd.arg("--fatal-warnings");
+
+        ::std::mem::replace(&mut self.cmd, Command::new(""))
+    }
+
+    // Not needed for now with LLD
+    fn group_start(&mut self) {}
+    fn group_end(&mut self) {}
+
+    fn cross_lang_lto(&mut self) {
+        // Do nothing for now
+    }
+}
+
+fn exported_symbols(tcx: TyCtxt, crate_type: CrateType) -> Vec<String> {
+    let mut symbols = Vec::new();
+
+    let export_threshold = symbol_export::crates_export_threshold(&[crate_type]);
+    for &(symbol, level) in tcx.exported_symbols(LOCAL_CRATE).iter() {
+        if level.is_below_threshold(export_threshold) {
+            symbols.push(symbol.symbol_name(tcx).to_string());
+        }
+    }
+
+    let formats = tcx.sess.dependency_formats.borrow();
+    let deps = formats[&crate_type].iter();
+
+    for (index, dep_format) in deps.enumerate() {
+        let cnum = CrateNum::new(index + 1);
+        // For each dependency that we are linking to statically ...
+        if *dep_format == Linkage::Static {
+            // ... we add its symbol list to our export list.
+            for &(symbol, level) in tcx.exported_symbols(cnum).iter() {
+                if level.is_below_threshold(export_threshold) {
+                    symbols.push(symbol.symbol_name(tcx).to_string());
+                }
+            }
+        }
+    }
+
+    symbols
+}
diff --git a/src/librustc_codegen_ssa/back/lto.rs b/src/librustc_codegen_ssa/back/lto.rs
new file mode 100644
index 00000000000..f68a82d8780
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/lto.rs
@@ -0,0 +1,122 @@
+// Copyright 2018 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use super::write::CodegenContext;
+use interfaces::*;
+use ModuleCodegen;
+
+use rustc::util::time_graph::Timeline;
+use rustc_errors::FatalError;
+
+use std::sync::Arc;
+use std::ffi::CString;
+
+pub struct ThinModule<B: WriteBackendMethods> {
+    pub shared: Arc<ThinShared<B>>,
+    pub idx: usize,
+}
+
+impl<B: WriteBackendMethods> ThinModule<B> {
+    pub fn name(&self) -> &str {
+        self.shared.module_names[self.idx].to_str().unwrap()
+    }
+
+    pub fn cost(&self) -> u64 {
+        // Yes, that's correct, we're using the size of the bytecode as an
+        // indicator for how costly this codegen unit is.
+        self.data().len() as u64
+    }
+
+    pub fn data(&self) -> &[u8] {
+        let a = self.shared.thin_buffers.get(self.idx).map(|b| b.data());
+        a.unwrap_or_else(|| {
+            let len = self.shared.thin_buffers.len();
+            self.shared.serialized_modules[self.idx - len].data()
+        })
+    }
+}
+
+pub struct ThinShared<B: WriteBackendMethods> {
+    pub data: B::ThinData,
+    pub thin_buffers: Vec<B::ThinBuffer>,
+    pub serialized_modules: Vec<SerializedModule<B::ModuleBuffer>>,
+    pub module_names: Vec<CString>,
+}
+
+
+pub enum LtoModuleCodegen<B: WriteBackendMethods> {
+    Fat {
+        module: Option<ModuleCodegen<B::Module>>,
+        _serialized_bitcode: Vec<SerializedModule<B::ModuleBuffer>>,
+    },
+
+    Thin(ThinModule<B>),
+}
+
+impl<B: WriteBackendMethods> LtoModuleCodegen<B> {
+    pub fn name(&self) -> &str {
+        match *self {
+            LtoModuleCodegen::Fat { .. } => "everything",
+            LtoModuleCodegen::Thin(ref m) => m.name(),
+        }
+    }
+
+    /// Optimize this module within the given codegen context.
+    ///
+    /// This function is unsafe as it'll return a `ModuleCodegen` still
+    /// points to LLVM data structures owned by this `LtoModuleCodegen`.
+    /// It's intended that the module returned is immediately code generated and
+    /// dropped, and then this LTO module is dropped.
+    pub unsafe fn optimize(
+        &mut self,
+        cgcx: &CodegenContext<B>,
+        timeline: &mut Timeline
+    ) -> Result<ModuleCodegen<B::Module>, FatalError> {
+        match *self {
+            LtoModuleCodegen::Fat { ref mut module, .. } => {
+                let module = module.take().unwrap();
+                {
+                    let config = cgcx.config(module.kind);
+                    B::run_lto_pass_manager(cgcx, &module, config, false);
+                    timeline.record("fat-done");
+                }
+                Ok(module)
+            }
+            LtoModuleCodegen::Thin(ref mut thin) => B::optimize_thin(cgcx, thin, timeline),
+        }
+    }
+
+    /// A "gauge" of how costly it is to optimize this module, used to sort
+    /// biggest modules first.
+    pub fn cost(&self) -> u64 {
+        match *self {
+            // Only one module with fat LTO, so the cost doesn't matter.
+            LtoModuleCodegen::Fat { .. } => 0,
+            LtoModuleCodegen::Thin(ref m) => m.cost(),
+        }
+    }
+}
+
+
+pub enum SerializedModule<M: ModuleBufferMethods> {
+    Local(M),
+    FromRlib(Vec<u8>),
+    FromUncompressedFile(memmap::Mmap),
+}
+
+impl<M: ModuleBufferMethods> SerializedModule<M> {
+    pub fn data(&self) -> &[u8] {
+        match *self {
+            SerializedModule::Local(ref m) => m.data(),
+            SerializedModule::FromRlib(ref m) => m,
+            SerializedModule::FromUncompressedFile(ref m) => m,
+        }
+    }
+}
diff --git a/src/librustc_codegen_ssa/back/mod.rs b/src/librustc_codegen_ssa/back/mod.rs
new file mode 100644
index 00000000000..3d7ead74d1c
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/mod.rs
@@ -0,0 +1,17 @@
+// Copyright 2018 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+pub mod write;
+pub mod linker;
+pub mod lto;
+pub mod link;
+pub mod command;
+pub mod symbol_export;
+pub mod archive;
diff --git a/src/librustc_codegen_ssa/back/symbol_export.rs b/src/librustc_codegen_ssa/back/symbol_export.rs
new file mode 100644
index 00000000000..dff7e518630
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/symbol_export.rs
@@ -0,0 +1,404 @@
+// Copyright 2016 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use rustc_data_structures::sync::Lrc;
+use std::sync::Arc;
+
+use rustc::ty::Instance;
+use rustc::hir;
+use rustc::hir::Node;
+use rustc::hir::CodegenFnAttrFlags;
+use rustc::hir::def_id::{CrateNum, DefId, LOCAL_CRATE, CRATE_DEF_INDEX};
+use rustc_data_structures::fingerprint::Fingerprint;
+use rustc::middle::exported_symbols::{SymbolExportLevel, ExportedSymbol, metadata_symbol_name};
+use rustc::session::config;
+use rustc::ty::{TyCtxt, SymbolName};
+use rustc::ty::query::Providers;
+use rustc::ty::subst::Substs;
+use rustc::util::nodemap::{FxHashMap, DefIdMap};
+use rustc_allocator::ALLOCATOR_METHODS;
+use rustc_data_structures::indexed_vec::IndexVec;
+use std::collections::hash_map::Entry::*;
+
+pub type ExportedSymbols = FxHashMap<
+    CrateNum,
+    Arc<Vec<(String, SymbolExportLevel)>>,
+>;
+
+pub fn threshold(tcx: TyCtxt) -> SymbolExportLevel {
+    crates_export_threshold(&tcx.sess.crate_types.borrow())
+}
+
+fn crate_export_threshold(crate_type: config::CrateType) -> SymbolExportLevel {
+    match crate_type {
+        config::CrateType::Executable |
+        config::CrateType::Staticlib  |
+        config::CrateType::ProcMacro  |
+        config::CrateType::Cdylib     => SymbolExportLevel::C,
+        config::CrateType::Rlib       |
+        config::CrateType::Dylib      => SymbolExportLevel::Rust,
+    }
+}
+
+pub fn crates_export_threshold(crate_types: &[config::CrateType]) -> SymbolExportLevel {
+    if crate_types.iter().any(|&crate_type|
+        crate_export_threshold(crate_type) == SymbolExportLevel::Rust)
+    {
+        SymbolExportLevel::Rust
+    } else {
+        SymbolExportLevel::C
+    }
+}
+
+fn reachable_non_generics_provider<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
+                                             cnum: CrateNum)
+                                             -> Lrc<DefIdMap<SymbolExportLevel>>
+{
+    assert_eq!(cnum, LOCAL_CRATE);
+
+    if !tcx.sess.opts.output_types.should_codegen() {
+        return Lrc::new(DefIdMap())
+    }
+
+    // Check to see if this crate is a "special runtime crate". These
+    // crates, implementation details of the standard library, typically
+    // have a bunch of `pub extern` and `#[no_mangle]` functions as the
+    // ABI between them. We don't want their symbols to have a `C`
+    // export level, however, as they're just implementation details.
+    // Down below we'll hardwire all of the symbols to the `Rust` export
+    // level instead.
+    let special_runtime_crate = tcx.is_panic_runtime(LOCAL_CRATE) ||
+        tcx.is_compiler_builtins(LOCAL_CRATE);
+
+    let mut reachable_non_generics: DefIdMap<_> = tcx.reachable_set(LOCAL_CRATE).0
+        .iter()
+        .filter_map(|&node_id| {
+            // We want to ignore some FFI functions that are not exposed from
+            // this crate. Reachable FFI functions can be lumped into two
+            // categories:
+            //
+            // 1. Those that are included statically via a static library
+            // 2. Those included otherwise (e.g. dynamically or via a framework)
+            //
+            // Although our LLVM module is not literally emitting code for the
+            // statically included symbols, it's an export of our library which
+            // needs to be passed on to the linker and encoded in the metadata.
+            //
+            // As a result, if this id is an FFI item (foreign item) then we only
+            // let it through if it's included statically.
+            match tcx.hir.get(node_id) {
+                Node::ForeignItem(..) => {
+                    let def_id = tcx.hir.local_def_id(node_id);
+                    if tcx.is_statically_included_foreign_item(def_id) {
+                        Some(def_id)
+                    } else {
+                        None
+                    }
+                }
+
+                // Only consider nodes that actually have exported symbols.
+                Node::Item(&hir::Item {
+                    node: hir::ItemKind::Static(..),
+                    ..
+                }) |
+                Node::Item(&hir::Item {
+                    node: hir::ItemKind::Fn(..), ..
+                }) |
+                Node::ImplItem(&hir::ImplItem {
+                    node: hir::ImplItemKind::Method(..),
+                    ..
+                }) => {
+                    let def_id = tcx.hir.local_def_id(node_id);
+                    let generics = tcx.generics_of(def_id);
+                    if !generics.requires_monomorphization(tcx) &&
+                        // Functions marked with #[inline] are only ever codegened
+                        // with "internal" linkage and are never exported.
+                        !Instance::mono(tcx, def_id).def.requires_local(tcx) {
+                        Some(def_id)
+                    } else {
+                        None
+                    }
+                }
+
+                _ => None
+            }
+        })
+        .map(|def_id| {
+            let export_level = if special_runtime_crate {
+                let name = tcx.symbol_name(Instance::mono(tcx, def_id)).as_str();
+                // We can probably do better here by just ensuring that
+                // it has hidden visibility rather than public
+                // visibility, as this is primarily here to ensure it's
+                // not stripped during LTO.
+                //
+                // In general though we won't link right if these
+                // symbols are stripped, and LTO currently strips them.
+                if &*name == "rust_eh_personality" ||
+                   &*name == "rust_eh_register_frames" ||
+                   &*name == "rust_eh_unregister_frames" {
+                    SymbolExportLevel::C
+                } else {
+                    SymbolExportLevel::Rust
+                }
+            } else {
+                symbol_export_level(tcx, def_id)
+            };
+            debug!("EXPORTED SYMBOL (local): {} ({:?})",
+                   tcx.symbol_name(Instance::mono(tcx, def_id)),
+                   export_level);
+            (def_id, export_level)
+        })
+        .collect();
+
+    if let Some(id) = *tcx.sess.derive_registrar_fn.get() {
+        let def_id = tcx.hir.local_def_id(id);
+        reachable_non_generics.insert(def_id, SymbolExportLevel::C);
+    }
+
+    if let Some(id) = *tcx.sess.plugin_registrar_fn.get() {
+        let def_id = tcx.hir.local_def_id(id);
+        reachable_non_generics.insert(def_id, SymbolExportLevel::C);
+    }
+
+    Lrc::new(reachable_non_generics)
+}
+
+fn is_reachable_non_generic_provider_local<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
+                                                     def_id: DefId)
+                                                     -> bool {
+    let export_threshold = threshold(tcx);
+
+    if let Some(&level) = tcx.reachable_non_generics(def_id.krate).get(&def_id) {
+        level.is_below_threshold(export_threshold)
+    } else {
+        false
+    }
+}
+
+fn is_reachable_non_generic_provider_extern<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
+                                                      def_id: DefId)
+                                                      -> bool {
+    tcx.reachable_non_generics(def_id.krate).contains_key(&def_id)
+}
+
+fn exported_symbols_provider_local<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
+                                             cnum: CrateNum)
+                                             -> Arc<Vec<(ExportedSymbol<'tcx>,
+                                                         SymbolExportLevel)>>
+{
+    assert_eq!(cnum, LOCAL_CRATE);
+
+    if !tcx.sess.opts.output_types.should_codegen() {
+        return Arc::new(vec![])
+    }
+
+    let mut symbols: Vec<_> = tcx.reachable_non_generics(LOCAL_CRATE)
+                                 .iter()
+                                 .map(|(&def_id, &level)| {
+                                    (ExportedSymbol::NonGeneric(def_id), level)
+                                 })
+                                 .collect();
+
+    if tcx.sess.entry_fn.borrow().is_some() {
+        let exported_symbol = ExportedSymbol::NoDefId(SymbolName::new("main"));
+
+        symbols.push((exported_symbol, SymbolExportLevel::C));
+    }
+
+    if tcx.sess.allocator_kind.get().is_some() {
+        for method in ALLOCATOR_METHODS {
+            let symbol_name = format!("__rust_{}", method.name);
+            let exported_symbol = ExportedSymbol::NoDefId(SymbolName::new(&symbol_name));
+
+            symbols.push((exported_symbol, SymbolExportLevel::Rust));
+        }
+    }
+
+    if tcx.sess.opts.debugging_opts.pgo_gen.is_some() {
+        // These are weak symbols that point to the profile version and the
+        // profile name, which need to be treated as exported so LTO doesn't nix
+        // them.
+        const PROFILER_WEAK_SYMBOLS: [&'static str; 2] = [
+            "__llvm_profile_raw_version",
+            "__llvm_profile_filename",
+        ];
+        for sym in &PROFILER_WEAK_SYMBOLS {
+            let exported_symbol = ExportedSymbol::NoDefId(SymbolName::new(sym));
+            symbols.push((exported_symbol, SymbolExportLevel::C));
+        }
+    }
+
+    if tcx.sess.crate_types.borrow().contains(&config::CrateType::Dylib) {
+        let symbol_name = metadata_symbol_name(tcx);
+        let exported_symbol = ExportedSymbol::NoDefId(SymbolName::new(&symbol_name));
+
+        symbols.push((exported_symbol, SymbolExportLevel::Rust));
+    }
+
+    if tcx.sess.opts.share_generics() && tcx.local_crate_exports_generics() {
+        use rustc::mir::mono::{Linkage, Visibility, MonoItem};
+        use rustc::ty::InstanceDef;
+
+        // Normally, we require that shared monomorphizations are not hidden,
+        // because if we want to re-use a monomorphization from a Rust dylib, it
+        // needs to be exported.
+        // However, on platforms that don't allow for Rust dylibs, having
+        // external linkage is enough for monomorphization to be linked to.
+        let need_visibility = tcx.sess.target.target.options.dynamic_linking &&
+                              !tcx.sess.target.target.options.only_cdylib;
+
+        let (_, cgus) = tcx.collect_and_partition_mono_items(LOCAL_CRATE);
+
+        for (mono_item, &(linkage, visibility)) in cgus.iter()
+                                                       .flat_map(|cgu| cgu.items().iter()) {
+            if linkage != Linkage::External {
+                // We can only re-use things with external linkage, otherwise
+                // we'll get a linker error
+                continue
+            }
+
+            if need_visibility && visibility == Visibility::Hidden {
+                // If we potentially share things from Rust dylibs, they must
+                // not be hidden
+                continue
+            }
+
+            if let &MonoItem::Fn(Instance {
+                def: InstanceDef::Item(def_id),
+                substs,
+            }) = mono_item {
+                if substs.types().next().is_some() {
+                    symbols.push((ExportedSymbol::Generic(def_id, substs),
+                                  SymbolExportLevel::Rust));
+                }
+            }
+        }
+    }
+
+    // Sort so we get a stable incr. comp. hash.
+    symbols.sort_unstable_by(|&(ref symbol1, ..), &(ref symbol2, ..)| {
+        symbol1.compare_stable(tcx, symbol2)
+    });
+
+    Arc::new(symbols)
+}
+
+fn upstream_monomorphizations_provider<'a, 'tcx>(
+    tcx: TyCtxt<'a, 'tcx, 'tcx>,
+    cnum: CrateNum)
+    -> Lrc<DefIdMap<Lrc<FxHashMap<&'tcx Substs<'tcx>, CrateNum>>>>
+{
+    debug_assert!(cnum == LOCAL_CRATE);
+
+    let cnums = tcx.all_crate_nums(LOCAL_CRATE);
+
+    let mut instances: DefIdMap<FxHashMap<_, _>> = DefIdMap();
+
+    let cnum_stable_ids: IndexVec<CrateNum, Fingerprint> = {
+        let mut cnum_stable_ids = IndexVec::from_elem_n(Fingerprint::ZERO,
+                                                        cnums.len() + 1);
+
+        for &cnum in cnums.iter() {
+            cnum_stable_ids[cnum] = tcx.def_path_hash(DefId {
+                krate: cnum,
+                index: CRATE_DEF_INDEX,
+            }).0;
+        }
+
+        cnum_stable_ids
+    };
+
+    for &cnum in cnums.iter() {
+        for &(ref exported_symbol, _) in tcx.exported_symbols(cnum).iter() {
+            if let &ExportedSymbol::Generic(def_id, substs) = exported_symbol {
+                let substs_map = instances.entry(def_id).or_default();
+
+                match substs_map.entry(substs) {
+                    Occupied(mut e) => {
+                        // If there are multiple monomorphizations available,
+                        // we select one deterministically.
+                        let other_cnum = *e.get();
+                        if cnum_stable_ids[other_cnum] > cnum_stable_ids[cnum] {
+                            e.insert(cnum);
+                        }
+                    }
+                    Vacant(e) => {
+                        e.insert(cnum);
+                    }
+                }
+            }
+        }
+    }
+
+    Lrc::new(instances.into_iter()
+                      .map(|(key, value)| (key, Lrc::new(value)))
+                      .collect())
+}
+
+fn upstream_monomorphizations_for_provider<'a, 'tcx>(
+    tcx: TyCtxt<'a, 'tcx, 'tcx>,
+    def_id: DefId)
+    -> Option<Lrc<FxHashMap<&'tcx Substs<'tcx>, CrateNum>>>
+{
+    debug_assert!(!def_id.is_local());
+    tcx.upstream_monomorphizations(LOCAL_CRATE)
+       .get(&def_id)
+       .cloned()
+}
+
+fn is_unreachable_local_definition_provider(tcx: TyCtxt, def_id: DefId) -> bool {
+    if let Some(node_id) = tcx.hir.as_local_node_id(def_id) {
+        !tcx.reachable_set(LOCAL_CRATE).0.contains(&node_id)
+    } else {
+        bug!("is_unreachable_local_definition called with non-local DefId: {:?}",
+             def_id)
+    }
+}
+
+pub fn provide(providers: &mut Providers) {
+    providers.reachable_non_generics = reachable_non_generics_provider;
+    providers.is_reachable_non_generic = is_reachable_non_generic_provider_local;
+    providers.exported_symbols = exported_symbols_provider_local;
+    providers.upstream_monomorphizations = upstream_monomorphizations_provider;
+    providers.is_unreachable_local_definition = is_unreachable_local_definition_provider;
+}
+
+pub fn provide_extern(providers: &mut Providers) {
+    providers.is_reachable_non_generic = is_reachable_non_generic_provider_extern;
+    providers.upstream_monomorphizations_for = upstream_monomorphizations_for_provider;
+}
+
+fn symbol_export_level(tcx: TyCtxt, sym_def_id: DefId) -> SymbolExportLevel {
+    // We export anything that's not mangled at the "C" layer as it probably has
+    // to do with ABI concerns. We do not, however, apply such treatment to
+    // special symbols in the standard library for various plumbing between
+    // core/std/allocators/etc. For example symbols used to hook up allocation
+    // are not considered for export
+    let codegen_fn_attrs = tcx.codegen_fn_attrs(sym_def_id);
+    let is_extern = codegen_fn_attrs.contains_extern_indicator();
+    let std_internal =
+        codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::RUSTC_STD_INTERNAL_SYMBOL);
+
+    if is_extern && !std_internal {
+        // Emscripten cannot export statics, so reduce their export level here
+        if tcx.sess.target.target.options.is_like_emscripten {
+            if let Some(Node::Item(&hir::Item {
+                node: hir::ItemKind::Static(..),
+                ..
+            })) = tcx.hir.get_if_local(sym_def_id) {
+                return SymbolExportLevel::Rust;
+            }
+        }
+
+        SymbolExportLevel::C
+    } else {
+        SymbolExportLevel::Rust
+    }
+}
diff --git a/src/librustc_codegen_ssa/back/write.rs b/src/librustc_codegen_ssa/back/write.rs
new file mode 100644
index 00000000000..e958b5441f2
--- /dev/null
+++ b/src/librustc_codegen_ssa/back/write.rs
@@ -0,0 +1,1843 @@
+// Copyright 2013-2015 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use {ModuleCodegen, ModuleKind, CachedModuleCodegen, CompiledModule, CrateInfo, CodegenResults,
+    RLIB_BYTECODE_EXTENSION};
+use super::linker::LinkerInfo;
+use super::lto::{self, SerializedModule};
+use super::link::{self, remove, get_linker};
+use super::command::Command;
+use super::symbol_export::ExportedSymbols;
+
+use memmap;
+use rustc_incremental::{copy_cgu_workproducts_to_incr_comp_cache_dir,
+                        in_incr_comp_dir, in_incr_comp_dir_sess};
+use rustc::dep_graph::{WorkProduct, WorkProductId, WorkProductFileKind};
+use rustc::dep_graph::cgu_reuse_tracker::CguReuseTracker;
+use rustc::middle::cstore::EncodedMetadata;
+use rustc::session::config::{self, OutputFilenames, OutputType, Passes, Sanitizer, Lto};
+use rustc::session::Session;
+use rustc::util::nodemap::FxHashMap;
+use rustc::util::time_graph::{self, TimeGraph, Timeline};
+use interfaces::*;
+use rustc::hir::def_id::{CrateNum, LOCAL_CRATE};
+use rustc::ty::TyCtxt;
+use rustc::util::common::{time_depth, set_time_depth, print_time_passes_entry};
+use rustc_fs_util::link_or_copy;
+use rustc_data_structures::svh::Svh;
+use rustc_errors::{Handler, Level, DiagnosticBuilder, FatalError, DiagnosticId};
+use rustc_errors::emitter::{Emitter};
+use syntax::attr;
+use syntax::ext::hygiene::Mark;
+use syntax_pos::MultiSpan;
+use syntax_pos::symbol::Symbol;
+use jobserver::{Client, Acquired};
+
+use std::any::Any;
+use std::fs;
+use std::io;
+use std::mem;
+use std::path::{Path, PathBuf};
+use std::str;
+use std::sync::Arc;
+use std::sync::mpsc::{channel, Sender, Receiver};
+use std::time::Instant;
+use std::thread;
+
+const PRE_THIN_LTO_BC_EXT: &str = "pre-thin-lto.bc";
+
+/// Module-specific configuration for `optimize_and_codegen`.
+pub struct ModuleConfig {
+    /// Names of additional optimization passes to run.
+    pub passes: Vec<String>,
+    /// Some(level) to optimize at a certain level, or None to run
+    /// absolutely no optimizations (used for the metadata module).
+    pub opt_level: Option<config::OptLevel>,
+
+    /// Some(level) to optimize binary size, or None to not affect program size.
+    pub opt_size: Option<config::OptLevel>,
+
+    pub pgo_gen: Option<String>,
+    pub pgo_use: String,
+
+    // Flags indicating which outputs to produce.
+    pub emit_pre_thin_lto_bc: bool,
+    pub emit_no_opt_bc: bool,
+    pub emit_bc: bool,
+    pub emit_bc_compressed: bool,
+    pub emit_lto_bc: bool,
+    pub emit_ir: bool,
+    pub emit_asm: bool,
+    pub emit_obj: bool,
+    // Miscellaneous flags.  These are mostly copied from command-line
+    // options.
+    pub verify_llvm_ir: bool,
+    pub no_prepopulate_passes: bool,
+    pub no_builtins: bool,
+    pub time_passes: bool,
+    pub vectorize_loop: bool,
+    pub vectorize_slp: bool,
+    pub merge_functions: bool,
+    pub inline_threshold: Option<usize>,
+    // Instead of creating an object file by doing LLVM codegen, just
+    // make the object file bitcode. Provides easy compatibility with
+    // emscripten's ecc compiler, when used as the linker.
+    pub obj_is_bitcode: bool,
+    pub no_integrated_as: bool,
+    pub embed_bitcode: bool,
+    pub embed_bitcode_marker: bool,
+}
+
+impl ModuleConfig {
+    fn new(passes: Vec<String>) -> ModuleConfig {
+        ModuleConfig {
+            passes,
+            opt_level: None,
+            opt_size: None,
+
+            pgo_gen: None,
+            pgo_use: String::new(),
+
+            emit_no_opt_bc: false,
+            emit_pre_thin_lto_bc: false,
+            emit_bc: false,
+            emit_bc_compressed: false,
+            emit_lto_bc: false,
+            emit_ir: false,
+            emit_asm: false,
+            emit_obj: false,
+            obj_is_bitcode: false,
+            embed_bitcode: false,
+            embed_bitcode_marker: false,
+            no_integrated_as: false,
+
+            verify_llvm_ir: false,
+            no_prepopulate_passes: false,
+            no_builtins: false,
+            time_passes: false,
+            vectorize_loop: false,
+            vectorize_slp: false,
+            merge_functions: false,
+            inline_threshold: None
+        }
+    }
+
+    fn set_flags(&mut self, sess: &Session, no_builtins: bool) {
+        self.verify_llvm_ir = sess.verify_llvm_ir();
+        self.no_prepopulate_passes = sess.opts.cg.no_prepopulate_passes;
+        self.no_builtins = no_builtins || sess.target.target.options.no_builtins;
+        self.time_passes = sess.time_passes();
+        self.inline_threshold = sess.opts.cg.inline_threshold;
+        self.obj_is_bitcode = sess.target.target.options.obj_is_bitcode ||
+                              sess.opts.debugging_opts.cross_lang_lto.enabled();
+        let embed_bitcode = sess.target.target.options.embed_bitcode ||
+                            sess.opts.debugging_opts.embed_bitcode;
+        if embed_bitcode {
+            match sess.opts.optimize {
+                config::OptLevel::No |
+                config::OptLevel::Less => {
+                    self.embed_bitcode_marker = embed_bitcode;
+                }
+                _ => self.embed_bitcode = embed_bitcode,
+            }
+        }
+
+        // Copy what clang does by turning on loop vectorization at O2 and
+        // slp vectorization at O3. Otherwise configure other optimization aspects
+        // of this pass manager builder.
+        // Turn off vectorization for emscripten, as it's not very well supported.
+        self.vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
+                             (sess.opts.optimize == config::OptLevel::Default ||
+                              sess.opts.optimize == config::OptLevel::Aggressive) &&
+                             !sess.target.target.options.is_like_emscripten;
+
+        self.vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
+                            sess.opts.optimize == config::OptLevel::Aggressive &&
+                            !sess.target.target.options.is_like_emscripten;
+
+        self.merge_functions = sess.opts.optimize == config::OptLevel::Default ||
+                               sess.opts.optimize == config::OptLevel::Aggressive;
+    }
+
+    pub fn bitcode_needed(&self) -> bool {
+        self.emit_bc || self.obj_is_bitcode
+            || self.emit_bc_compressed || self.embed_bitcode
+    }
+}
+
+/// Assembler name and command used by codegen when no_integrated_as is enabled
+pub struct AssemblerCommand {
+    name: PathBuf,
+    cmd: Command,
+}
+
+/// Additional resources used by optimize_and_codegen (not module specific)
+#[derive(Clone)]
+pub struct CodegenContext<B: WriteBackendMethods> {
+    // Resources needed when running LTO
+    pub backend: B,
+    pub time_passes: bool,
+    pub lto: Lto,
+    pub no_landing_pads: bool,
+    pub save_temps: bool,
+    pub fewer_names: bool,
+    pub exported_symbols: Option<Arc<ExportedSymbols>>,
+    pub opts: Arc<config::Options>,
+    pub crate_types: Vec<config::CrateType>,
+    pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>,
+    pub output_filenames: Arc<OutputFilenames>,
+    pub regular_module_config: Arc<ModuleConfig>,
+    pub metadata_module_config: Arc<ModuleConfig>,
+    pub allocator_module_config: Arc<ModuleConfig>,
+    pub tm_factory: Arc<dyn Fn()
+        -> Result<B::TargetMachine, String> + Send + Sync>,
+    pub msvc_imps_needed: bool,
+    pub target_pointer_width: String,
+    pub debuginfo: config::DebugInfo,
+
+    // Number of cgus excluding the allocator/metadata modules
+    pub total_cgus: usize,
+    // Handler to use for diagnostics produced during codegen.
+    pub diag_emitter: SharedEmitter,
+    // LLVM passes added by plugins.
+    pub plugin_passes: Vec<String>,
+    // LLVM optimizations for which we want to print remarks.
+    pub remark: Passes,
+    // Worker thread number
+    pub worker: usize,
+    // The incremental compilation session directory, or None if we are not
+    // compiling incrementally
+    pub incr_comp_session_dir: Option<PathBuf>,
+    // Used to update CGU re-use information during the thinlto phase.
+    pub cgu_reuse_tracker: CguReuseTracker,
+    // Channel back to the main control thread to send messages to
+    pub coordinator_send: Sender<Box<dyn Any + Send>>,
+    // A reference to the TimeGraph so we can register timings. None means that
+    // measuring is disabled.
+    pub time_graph: Option<TimeGraph>,
+    // The assembler command if no_integrated_as option is enabled, None otherwise
+    pub assembler_cmd: Option<Arc<AssemblerCommand>>
+}
+
+impl<B: WriteBackendMethods> CodegenContext<B> {
+    pub fn create_diag_handler(&self) -> Handler {
+        Handler::with_emitter(true, false, Box::new(self.diag_emitter.clone()))
+    }
+
+    pub fn config(&self, kind: ModuleKind) -> &ModuleConfig {
+        match kind {
+            ModuleKind::Regular => &self.regular_module_config,
+            ModuleKind::Metadata => &self.metadata_module_config,
+            ModuleKind::Allocator => &self.allocator_module_config,
+        }
+    }
+}
+
+fn generate_lto_work<B: ExtraBackendMethods>(
+    cgcx: &CodegenContext<B>,
+    modules: Vec<ModuleCodegen<B::Module>>,
+    import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>
+) -> Vec<(WorkItem<B>, u64)> {
+    let mut timeline = cgcx.time_graph.as_ref().map(|tg| {
+        tg.start(CODEGEN_WORKER_TIMELINE,
+                 CODEGEN_WORK_PACKAGE_KIND,
+                 "generate lto")
+    }).unwrap_or(Timeline::noop());
+    let (lto_modules, copy_jobs) = B::run_lto(cgcx, modules, import_only_modules, &mut timeline)
+        .unwrap_or_else(|e| e.raise());
+
+    let lto_modules = lto_modules.into_iter().map(|module| {
+        let cost = module.cost();
+        (WorkItem::LTO(module), cost)
+    });
+
+    let copy_jobs = copy_jobs.into_iter().map(|wp| {
+        (WorkItem::CopyPostLtoArtifacts(CachedModuleCodegen {
+            name: wp.cgu_name.clone(),
+            source: wp,
+        }), 0)
+    });
+
+    lto_modules.chain(copy_jobs).collect()
+}
+
+pub struct CompiledModules {
+    pub modules: Vec<CompiledModule>,
+    pub metadata_module: CompiledModule,
+    pub allocator_module: Option<CompiledModule>,
+}
+
+fn need_crate_bitcode_for_rlib(sess: &Session) -> bool {
+    sess.crate_types.borrow().contains(&config::CrateType::Rlib) &&
+    sess.opts.output_types.contains_key(&OutputType::Exe)
+}
+
+fn need_pre_thin_lto_bitcode_for_incr_comp(sess: &Session) -> bool {
+    if sess.opts.incremental.is_none() {
+        return false
+    }
+
+    match sess.lto() {
+        Lto::Fat |
+        Lto::No => false,
+        Lto::Thin |
+        Lto::ThinLocal => true,
+    }
+}
+
+pub fn start_async_codegen<B: ExtraBackendMethods>(
+    backend: B,
+    tcx: TyCtxt,
+    time_graph: Option<TimeGraph>,
+    metadata: EncodedMetadata,
+    coordinator_receive: Receiver<Box<dyn Any + Send>>,
+    total_cgus: usize
+) -> OngoingCodegen<B> {
+    let sess = tcx.sess;
+    let crate_name = tcx.crate_name(LOCAL_CRATE);
+    let crate_hash = tcx.crate_hash(LOCAL_CRATE);
+    let no_builtins = attr::contains_name(&tcx.hir.krate().attrs, "no_builtins");
+    let subsystem = attr::first_attr_value_str_by_name(&tcx.hir.krate().attrs,
+                                                       "windows_subsystem");
+    let windows_subsystem = subsystem.map(|subsystem| {
+        if subsystem != "windows" && subsystem != "console" {
+            tcx.sess.fatal(&format!("invalid windows subsystem `{}`, only \
+                                     `windows` and `console` are allowed",
+                                    subsystem));
+        }
+        subsystem.to_string()
+    });
+
+    let linker_info = LinkerInfo::new(tcx);
+    let crate_info = CrateInfo::new(tcx);
+
+    // Figure out what we actually need to build.
+    let mut modules_config = ModuleConfig::new(sess.opts.cg.passes.clone());
+    let mut metadata_config = ModuleConfig::new(vec![]);
+    let mut allocator_config = ModuleConfig::new(vec![]);
+
+    if let Some(ref sanitizer) = sess.opts.debugging_opts.sanitizer {
+        match *sanitizer {
+            Sanitizer::Address => {
+                modules_config.passes.push("asan".to_owned());
+                modules_config.passes.push("asan-module".to_owned());
+            }
+            Sanitizer::Memory => {
+                modules_config.passes.push("msan".to_owned())
+            }
+            Sanitizer::Thread => {
+                modules_config.passes.push("tsan".to_owned())
+            }
+            _ => {}
+        }
+    }
+
+    if sess.opts.debugging_opts.profile {
+        modules_config.passes.push("insert-gcov-profiling".to_owned())
+    }
+
+    modules_config.pgo_gen = sess.opts.debugging_opts.pgo_gen.clone();
+    modules_config.pgo_use = sess.opts.debugging_opts.pgo_use.clone();
+
+    modules_config.opt_level = Some(sess.opts.optimize);
+    modules_config.opt_size = Some(sess.opts.optimize);
+
+    // Save all versions of the bytecode if we're saving our temporaries.
+    if sess.opts.cg.save_temps {
+        modules_config.emit_no_opt_bc = true;
+        modules_config.emit_pre_thin_lto_bc = true;
+        modules_config.emit_bc = true;
+        modules_config.emit_lto_bc = true;
+        metadata_config.emit_bc = true;
+        allocator_config.emit_bc = true;
+    }
+
+    // Emit compressed bitcode files for the crate if we're emitting an rlib.
+    // Whenever an rlib is created, the bitcode is inserted into the archive in
+    // order to allow LTO against it.
+    if need_crate_bitcode_for_rlib(sess) {
+        modules_config.emit_bc_compressed = true;
+        allocator_config.emit_bc_compressed = true;
+    }
+
+    modules_config.emit_pre_thin_lto_bc =
+        need_pre_thin_lto_bitcode_for_incr_comp(sess);
+
+    modules_config.no_integrated_as = tcx.sess.opts.cg.no_integrated_as ||
+        tcx.sess.target.target.options.no_integrated_as;
+
+    for output_type in sess.opts.output_types.keys() {
+        match *output_type {
+            OutputType::Bitcode => { modules_config.emit_bc = true; }
+            OutputType::LlvmAssembly => { modules_config.emit_ir = true; }
+            OutputType::Assembly => {
+                modules_config.emit_asm = true;
+                // If we're not using the LLVM assembler, this function
+                // could be invoked specially with output_type_assembly, so
+                // in this case we still want the metadata object file.
+                if !sess.opts.output_types.contains_key(&OutputType::Assembly) {
+                    metadata_config.emit_obj = true;
+                    allocator_config.emit_obj = true;
+                }
+            }
+            OutputType::Object => { modules_config.emit_obj = true; }
+            OutputType::Metadata => { metadata_config.emit_obj = true; }
+            OutputType::Exe => {
+                modules_config.emit_obj = true;
+                metadata_config.emit_obj = true;
+                allocator_config.emit_obj = true;
+            },
+            OutputType::Mir => {}
+            OutputType::DepInfo => {}
+        }
+    }
+
+    modules_config.set_flags(sess, no_builtins);
+    metadata_config.set_flags(sess, no_builtins);
+    allocator_config.set_flags(sess, no_builtins);
+
+    // Exclude metadata and allocator modules from time_passes output, since
+    // they throw off the "LLVM passes" measurement.
+    metadata_config.time_passes = false;
+    allocator_config.time_passes = false;
+
+    let (shared_emitter, shared_emitter_main) = SharedEmitter::new();
+    let (codegen_worker_send, codegen_worker_receive) = channel();
+
+    let coordinator_thread = start_executing_work(backend.clone(),
+                                                  tcx,
+                                                  &crate_info,
+                                                  shared_emitter,
+                                                  codegen_worker_send,
+                                                  coordinator_receive,
+                                                  total_cgus,
+                                                  sess.jobserver.clone(),
+                                                  time_graph.clone(),
+                                                  Arc::new(modules_config),
+                                                  Arc::new(metadata_config),
+                                                  Arc::new(allocator_config));
+
+    OngoingCodegen {
+        backend,
+        crate_name,
+        crate_hash,
+        metadata,
+        windows_subsystem,
+        linker_info,
+        crate_info,
+
+        time_graph,
+        coordinator_send: tcx.tx_to_llvm_workers.lock().clone(),
+        codegen_worker_receive,
+        shared_emitter_main,
+        future: coordinator_thread,
+        output_filenames: tcx.output_filenames(LOCAL_CRATE),
+    }
+}
+
+fn copy_all_cgu_workproducts_to_incr_comp_cache_dir(
+    sess: &Session,
+    compiled_modules: &CompiledModules,
+) -> FxHashMap<WorkProductId, WorkProduct> {
+    let mut work_products = FxHashMap::default();
+
+    if sess.opts.incremental.is_none() {
+        return work_products;
+    }
+
+    for module in compiled_modules.modules.iter().filter(|m| m.kind == ModuleKind::Regular) {
+        let mut files = vec![];
+
+        if let Some(ref path) = module.object {
+            files.push((WorkProductFileKind::Object, path.clone()));
+        }
+        if let Some(ref path) = module.bytecode {
+            files.push((WorkProductFileKind::Bytecode, path.clone()));
+        }
+        if let Some(ref path) = module.bytecode_compressed {
+            files.push((WorkProductFileKind::BytecodeCompressed, path.clone()));
+        }
+
+        if let Some((id, product)) =
+                copy_cgu_workproducts_to_incr_comp_cache_dir(sess, &module.name, &files) {
+            work_products.insert(id, product);
+        }
+    }
+
+    work_products
+}
+
+fn produce_final_output_artifacts(sess: &Session,
+                                  compiled_modules: &CompiledModules,
+                                  crate_output: &OutputFilenames) {
+    let mut user_wants_bitcode = false;
+    let mut user_wants_objects = false;
+
+    // Produce final compile outputs.
+    let copy_gracefully = |from: &Path, to: &Path| {
+        if let Err(e) = fs::copy(from, to) {
+            sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
+        }
+    };
+
+    let copy_if_one_unit = |output_type: OutputType,
+                            keep_numbered: bool| {
+        if compiled_modules.modules.len() == 1 {
+            // 1) Only one codegen unit.  In this case it's no difficulty
+            //    to copy `foo.0.x` to `foo.x`.
+            let module_name = Some(&compiled_modules.modules[0].name[..]);
+            let path = crate_output.temp_path(output_type, module_name);
+            copy_gracefully(&path,
+                            &crate_output.path(output_type));
+            if !sess.opts.cg.save_temps && !keep_numbered {
+                // The user just wants `foo.x`, not `foo.#module-name#.x`.
+                remove(sess, &path);
+            }
+        } else {
+            let ext = crate_output.temp_path(output_type, None)
+                                  .extension()
+                                  .unwrap()
+                                  .to_str()
+                                  .unwrap()
+                                  .to_owned();
+
+            if crate_output.outputs.contains_key(&output_type) {
+                // 2) Multiple codegen units, with `--emit foo=some_name`.  We have
+                //    no good solution for this case, so warn the user.
+                sess.warn(&format!("ignoring emit path because multiple .{} files \
+                                    were produced", ext));
+            } else if crate_output.single_output_file.is_some() {
+                // 3) Multiple codegen units, with `-o some_name`.  We have
+                //    no good solution for this case, so warn the user.
+                sess.warn(&format!("ignoring -o because multiple .{} files \
+                                    were produced", ext));
+            } else {
+                // 4) Multiple codegen units, but no explicit name.  We
+                //    just leave the `foo.0.x` files in place.
+                // (We don't have to do any work in this case.)
+            }
+        }
+    };
+
+    // Flag to indicate whether the user explicitly requested bitcode.
+    // Otherwise, we produced it only as a temporary output, and will need
+    // to get rid of it.
+    for output_type in crate_output.outputs.keys() {
+        match *output_type {
+            OutputType::Bitcode => {
+                user_wants_bitcode = true;
+                // Copy to .bc, but always keep the .0.bc.  There is a later
+                // check to figure out if we should delete .0.bc files, or keep
+                // them for making an rlib.
+                copy_if_one_unit(OutputType::Bitcode, true);
+            }
+            OutputType::LlvmAssembly => {
+                copy_if_one_unit(OutputType::LlvmAssembly, false);
+            }
+            OutputType::Assembly => {
+                copy_if_one_unit(OutputType::Assembly, false);
+            }
+            OutputType::Object => {
+                user_wants_objects = true;
+                copy_if_one_unit(OutputType::Object, true);
+            }
+            OutputType::Mir |
+            OutputType::Metadata |
+            OutputType::Exe |
+            OutputType::DepInfo => {}
+        }
+    }
+
+    // Clean up unwanted temporary files.
+
+    // We create the following files by default:
+    //  - #crate#.#module-name#.bc
+    //  - #crate#.#module-name#.o
+    //  - #crate#.crate.metadata.bc
+    //  - #crate#.crate.metadata.o
+    //  - #crate#.o (linked from crate.##.o)
+    //  - #crate#.bc (copied from crate.##.bc)
+    // We may create additional files if requested by the user (through
+    // `-C save-temps` or `--emit=` flags).
+
+    if !sess.opts.cg.save_temps {
+        // Remove the temporary .#module-name#.o objects.  If the user didn't
+        // explicitly request bitcode (with --emit=bc), and the bitcode is not
+        // needed for building an rlib, then we must remove .#module-name#.bc as
+        // well.
+
+        // Specific rules for keeping .#module-name#.bc:
+        //  - If the user requested bitcode (`user_wants_bitcode`), and
+        //    codegen_units > 1, then keep it.
+        //  - If the user requested bitcode but codegen_units == 1, then we
+        //    can toss .#module-name#.bc because we copied it to .bc earlier.
+        //  - If we're not building an rlib and the user didn't request
+        //    bitcode, then delete .#module-name#.bc.
+        // If you change how this works, also update back::link::link_rlib,
+        // where .#module-name#.bc files are (maybe) deleted after making an
+        // rlib.
+        let needs_crate_object = crate_output.outputs.contains_key(&OutputType::Exe);
+
+        let keep_numbered_bitcode = user_wants_bitcode && sess.codegen_units() > 1;
+
+        let keep_numbered_objects = needs_crate_object ||
+                (user_wants_objects && sess.codegen_units() > 1);
+
+        for module in compiled_modules.modules.iter() {
+            if let Some(ref path) = module.object {
+                if !keep_numbered_objects {
+                    remove(sess, path);
+                }
+            }
+
+            if let Some(ref path) = module.bytecode {
+                if !keep_numbered_bitcode {
+                    remove(sess, path);
+                }
+            }
+        }
+
+        if !user_wants_bitcode {
+            if let Some(ref path) = compiled_modules.metadata_module.bytecode {
+                remove(sess, &path);
+            }
+
+            if let Some(ref allocator_module) = compiled_modules.allocator_module {
+                if let Some(ref path) = allocator_module.bytecode {
+                    remove(sess, path);
+                }
+            }
+        }
+    }
+
+    // We leave the following files around by default:
+    //  - #crate#.o
+    //  - #crate#.crate.metadata.o
+    //  - #crate#.bc
+    // These are used in linking steps and will be cleaned up afterward.
+}
+
+pub fn dump_incremental_data(_codegen_results: &CodegenResults) {
+    // FIXME(mw): This does not work at the moment because the situation has
+    //            become more complicated due to incremental LTO. Now a CGU
+    //            can have more than two caching states.
+    // println!("[incremental] Re-using {} out of {} modules",
+    //           codegen_results.modules.iter().filter(|m| m.pre_existing).count(),
+    //           codegen_results.modules.len());
+}
+
+pub enum WorkItem<B: WriteBackendMethods> {
+    /// Optimize a newly codegened, totally unoptimized module.
+    Optimize(ModuleCodegen<B::Module>),
+    /// Copy the post-LTO artifacts from the incremental cache to the output
+    /// directory.
+    CopyPostLtoArtifacts(CachedModuleCodegen),
+    /// Perform (Thin)LTO on the given module.
+    LTO(lto::LtoModuleCodegen<B>),
+}
+
+impl<B: WriteBackendMethods> WorkItem<B> {
+    pub fn module_kind(&self) -> ModuleKind {
+        match *self {
+            WorkItem::Optimize(ref m) => m.kind,
+            WorkItem::CopyPostLtoArtifacts(_) |
+            WorkItem::LTO(_) => ModuleKind::Regular,
+        }
+    }
+
+    pub fn name(&self) -> String {
+        match *self {
+            WorkItem::Optimize(ref m) => format!("optimize: {}", m.name),
+            WorkItem::CopyPostLtoArtifacts(ref m) => format!("copy post LTO artifacts: {}", m.name),
+            WorkItem::LTO(ref m) => format!("lto: {}", m.name()),
+        }
+    }
+}
+
+enum WorkItemResult<M> {
+    Compiled(CompiledModule),
+    NeedsLTO(ModuleCodegen<M>),
+}
+
+fn execute_work_item<B: ExtraBackendMethods>(
+    cgcx: &CodegenContext<B>,
+    work_item: WorkItem<B>,
+    timeline: &mut Timeline
+) -> Result<WorkItemResult<B::Module>, FatalError> {
+    let module_config = cgcx.config(work_item.module_kind());
+
+    match work_item {
+        WorkItem::Optimize(module) => {
+            execute_optimize_work_item(cgcx, module, module_config, timeline)
+        }
+        WorkItem::CopyPostLtoArtifacts(module) => {
+            execute_copy_from_cache_work_item(cgcx, module, module_config, timeline)
+        }
+        WorkItem::LTO(module) => {
+            execute_lto_work_item(cgcx, module, module_config, timeline)
+        }
+    }
+}
+
+fn execute_optimize_work_item<B: ExtraBackendMethods>(
+    cgcx: &CodegenContext<B>,
+    module: ModuleCodegen<B::Module>,
+    module_config: &ModuleConfig,
+    timeline: &mut Timeline
+) -> Result<WorkItemResult<B::Module>, FatalError> {
+    let diag_handler = cgcx.create_diag_handler();
+
+    unsafe {
+        B::optimize(cgcx, &diag_handler, &module, module_config, timeline)?;
+    }
+
+    let linker_does_lto = cgcx.opts.debugging_opts.cross_lang_lto.enabled();
+
+    // After we've done the initial round of optimizations we need to
+    // decide whether to synchronously codegen this module or ship it
+    // back to the coordinator thread for further LTO processing (which
+    // has to wait for all the initial modules to be optimized).
+    //
+    // Here we dispatch based on the `cgcx.lto` and kind of module we're
+    // codegenning...
+    let needs_lto = match cgcx.lto {
+        Lto::No => false,
+
+        // If the linker does LTO, we don't have to do it. Note that we
+        // keep doing full LTO, if it is requested, as not to break the
+        // assumption that the output will be a single module.
+        Lto::Thin | Lto::ThinLocal if linker_does_lto => false,
+
+        // Here we've got a full crate graph LTO requested. We ignore
+        // this, however, if the crate type is only an rlib as there's
+        // no full crate graph to process, that'll happen later.
+        //
+        // This use case currently comes up primarily for targets that
+        // require LTO so the request for LTO is always unconditionally
+        // passed down to the backend, but we don't actually want to do
+        // anything about it yet until we've got a final product.
+        Lto::Fat | Lto::Thin => {
+            cgcx.crate_types.len() != 1 ||
+                cgcx.crate_types[0] != config::CrateType::Rlib
+        }
+
+        // When we're automatically doing ThinLTO for multi-codegen-unit
+        // builds we don't actually want to LTO the allocator modules if
+        // it shows up. This is due to various linker shenanigans that
+        // we'll encounter later.
+        Lto::ThinLocal => {
+            module.kind != ModuleKind::Allocator
+        }
+    };
+
+    // Metadata modules never participate in LTO regardless of the lto
+    // settings.
+    let needs_lto = needs_lto && module.kind != ModuleKind::Metadata;
+
+    if needs_lto {
+        Ok(WorkItemResult::NeedsLTO(module))
+    } else {
+        let module = unsafe { B::codegen(cgcx, &diag_handler, module, module_config, timeline)? };
+        Ok(WorkItemResult::Compiled(module))
+    }
+}
+
+fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>(
+    cgcx: &CodegenContext<B>,
+    module: CachedModuleCodegen,
+    module_config: &ModuleConfig,
+    _: &mut Timeline
+) -> Result<WorkItemResult<B::Module>, FatalError> {
+    let incr_comp_session_dir = cgcx.incr_comp_session_dir
+                                    .as_ref()
+                                    .unwrap();
+    let mut object = None;
+    let mut bytecode = None;
+    let mut bytecode_compressed = None;
+    for (kind, saved_file) in &module.source.saved_files {
+        let obj_out = match kind {
+            WorkProductFileKind::Object => {
+                let path = cgcx.output_filenames.temp_path(OutputType::Object,
+                                                           Some(&module.name));
+                object = Some(path.clone());
+                path
+            }
+            WorkProductFileKind::Bytecode => {
+                let path = cgcx.output_filenames.temp_path(OutputType::Bitcode,
+                                                           Some(&module.name));
+                bytecode = Some(path.clone());
+                path
+            }
+            WorkProductFileKind::BytecodeCompressed => {
+                let path = cgcx.output_filenames.temp_path(OutputType::Bitcode,
+                                                           Some(&module.name))
+                    .with_extension(RLIB_BYTECODE_EXTENSION);
+                bytecode_compressed = Some(path.clone());
+                path
+            }
+        };
+        let source_file = in_incr_comp_dir(&incr_comp_session_dir,
+                                           &saved_file);
+        debug!("copying pre-existing module `{}` from {:?} to {}",
+               module.name,
+               source_file,
+               obj_out.display());
+        if let Err(err) = link_or_copy(&source_file, &obj_out) {
+            let diag_handler = cgcx.create_diag_handler();
+            diag_handler.err(&format!("unable to copy {} to {}: {}",
+                                      source_file.display(),
+                                      obj_out.display(),
+                                      err));
+        }
+    }
+
+    assert_eq!(object.is_some(), module_config.emit_obj);
+    assert_eq!(bytecode.is_some(), module_config.emit_bc);
+    assert_eq!(bytecode_compressed.is_some(), module_config.emit_bc_compressed);
+
+    Ok(WorkItemResult::Compiled(CompiledModule {
+        name: module.name,
+        kind: ModuleKind::Regular,
+        object,
+        bytecode,
+        bytecode_compressed,
+    }))
+}
+
+fn execute_lto_work_item<B: ExtraBackendMethods>(
+    cgcx: &CodegenContext<B>,
+    mut module: lto::LtoModuleCodegen<B>,
+    module_config: &ModuleConfig,
+    timeline: &mut Timeline
+) -> Result<WorkItemResult<B::Module>, FatalError> {
+    let diag_handler = cgcx.create_diag_handler();
+
+    unsafe {
+        let module = module.optimize(cgcx, timeline)?;
+        let module = B::codegen(cgcx, &diag_handler, module, module_config, timeline)?;
+        Ok(WorkItemResult::Compiled(module))
+    }
+}
+
+pub enum Message<B: WriteBackendMethods> {
+    Token(io::Result<Acquired>),
+    NeedsLTO {
+        result: ModuleCodegen<B::Module>,
+        worker_id: usize,
+    },
+    Done {
+        result: Result<CompiledModule, ()>,
+        worker_id: usize,
+    },
+    CodegenDone {
+        llvm_work_item: WorkItem<B>,
+        cost: u64,
+    },
+    AddImportOnlyModule {
+        module_data: SerializedModule<B::ModuleBuffer>,
+        work_product: WorkProduct,
+    },
+    CodegenComplete,
+    CodegenItem,
+    CodegenAborted,
+}
+
+struct Diagnostic {
+    msg: String,
+    code: Option<DiagnosticId>,
+    lvl: Level,
+}
+
+#[derive(PartialEq, Clone, Copy, Debug)]
+enum MainThreadWorkerState {
+    Idle,
+    Codegenning,
+    LLVMing,
+}
+
+fn start_executing_work<B: ExtraBackendMethods>(
+    backend: B,
+    tcx: TyCtxt,
+    crate_info: &CrateInfo,
+    shared_emitter: SharedEmitter,
+    codegen_worker_send: Sender<Message<B>>,
+    coordinator_receive: Receiver<Box<dyn Any + Send>>,
+    total_cgus: usize,
+    jobserver: Client,
+    time_graph: Option<TimeGraph>,
+    modules_config: Arc<ModuleConfig>,
+    metadata_config: Arc<ModuleConfig>,
+    allocator_config: Arc<ModuleConfig>
+) -> thread::JoinHandle<Result<CompiledModules, ()>> {
+    let coordinator_send = tcx.tx_to_llvm_workers.lock().clone();
+    let sess = tcx.sess;
+
+    // Compute the set of symbols we need to retain when doing LTO (if we need to)
+    let exported_symbols = {
+        let mut exported_symbols = FxHashMap::default();
+
+        let copy_symbols = |cnum| {
+            let symbols = tcx.exported_symbols(cnum)
+                             .iter()
+                             .map(|&(s, lvl)| (s.symbol_name(tcx).to_string(), lvl))
+                             .collect();
+            Arc::new(symbols)
+        };
+
+        match sess.lto() {
+            Lto::No => None,
+            Lto::ThinLocal => {
+                exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
+                Some(Arc::new(exported_symbols))
+            }
+            Lto::Fat | Lto::Thin => {
+                exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
+                for &cnum in tcx.crates().iter() {
+                    exported_symbols.insert(cnum, copy_symbols(cnum));
+                }
+                Some(Arc::new(exported_symbols))
+            }
+        }
+    };
+
+    // First up, convert our jobserver into a helper thread so we can use normal
+    // mpsc channels to manage our messages and such.
+    // After we've requested tokens then we'll, when we can,
+    // get tokens on `coordinator_receive` which will
+    // get managed in the main loop below.
+    let coordinator_send2 = coordinator_send.clone();
+    let helper = jobserver.into_helper_thread(move |token| {
+        drop(coordinator_send2.send(Box::new(Message::Token::<B>(token))));
+    }).expect("failed to spawn helper thread");
+
+    let mut each_linked_rlib_for_lto = Vec::new();
+    drop(link::each_linked_rlib(sess, crate_info, &mut |cnum, path| {
+        if link::ignored_for_lto(sess, crate_info, cnum) {
+            return
+        }
+        each_linked_rlib_for_lto.push((cnum, path.to_path_buf()));
+    }));
+
+    let assembler_cmd = if modules_config.no_integrated_as {
+        // HACK: currently we use linker (gcc) as our assembler
+        let (linker, flavor) = link::linker_and_flavor(sess);
+
+        let (name, mut cmd) = get_linker(sess, &linker, flavor);
+        cmd.args(&sess.target.target.options.asm_args);
+        Some(Arc::new(AssemblerCommand {
+            name,
+            cmd,
+        }))
+    } else {
+        None
+    };
+
+    let cgcx = CodegenContext::<B> {
+        backend: backend.clone(),
+        crate_types: sess.crate_types.borrow().clone(),
+        each_linked_rlib_for_lto,
+        lto: sess.lto(),
+        no_landing_pads: sess.no_landing_pads(),
+        fewer_names: sess.fewer_names(),
+        save_temps: sess.opts.cg.save_temps,
+        opts: Arc::new(sess.opts.clone()),
+        time_passes: sess.time_passes(),
+        exported_symbols,
+        plugin_passes: sess.plugin_llvm_passes.borrow().clone(),
+        remark: sess.opts.cg.remark.clone(),
+        worker: 0,
+        incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
+        cgu_reuse_tracker: sess.cgu_reuse_tracker.clone(),
+        coordinator_send,
+        diag_emitter: shared_emitter.clone(),
+        time_graph,
+        output_filenames: tcx.output_filenames(LOCAL_CRATE),
+        regular_module_config: modules_config,
+        metadata_module_config: metadata_config,
+        allocator_module_config: allocator_config,
+        tm_factory: backend.target_machine_factory(tcx.sess, false),
+        total_cgus,
+        msvc_imps_needed: msvc_imps_needed(tcx),
+        target_pointer_width: tcx.sess.target.target.target_pointer_width.clone(),
+        debuginfo: tcx.sess.opts.debuginfo,
+        assembler_cmd,
+    };
+
+    // This is the "main loop" of parallel work happening for parallel codegen.
+    // It's here that we manage parallelism, schedule work, and work with
+    // messages coming from clients.
+    //
+    // There are a few environmental pre-conditions that shape how the system
+    // is set up:
+    //
+    // - Error reporting only can happen on the main thread because that's the
+    //   only place where we have access to the compiler `Session`.
+    // - LLVM work can be done on any thread.
+    // - Codegen can only happen on the main thread.
+    // - Each thread doing substantial work most be in possession of a `Token`
+    //   from the `Jobserver`.
+    // - The compiler process always holds one `Token`. Any additional `Tokens`
+    //   have to be requested from the `Jobserver`.
+    //
+    // Error Reporting
+    // ===============
+    // The error reporting restriction is handled separately from the rest: We
+    // set up a `SharedEmitter` the holds an open channel to the main thread.
+    // When an error occurs on any thread, the shared emitter will send the
+    // error message to the receiver main thread (`SharedEmitterMain`). The
+    // main thread will periodically query this error message queue and emit
+    // any error messages it has received. It might even abort compilation if
+    // has received a fatal error. In this case we rely on all other threads
+    // being torn down automatically with the main thread.
+    // Since the main thread will often be busy doing codegen work, error
+    // reporting will be somewhat delayed, since the message queue can only be
+    // checked in between to work packages.
+    //
+    // Work Processing Infrastructure
+    // ==============================
+    // The work processing infrastructure knows three major actors:
+    //
+    // - the coordinator thread,
+    // - the main thread, and
+    // - LLVM worker threads
+    //
+    // The coordinator thread is running a message loop. It instructs the main
+    // thread about what work to do when, and it will spawn off LLVM worker
+    // threads as open LLVM WorkItems become available.
+    //
+    // The job of the main thread is to codegen CGUs into LLVM work package
+    // (since the main thread is the only thread that can do this). The main
+    // thread will block until it receives a message from the coordinator, upon
+    // which it will codegen one CGU, send it to the coordinator and block
+    // again. This way the coordinator can control what the main thread is
+    // doing.
+    //
+    // The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is
+    // available, it will spawn off a new LLVM worker thread and let it process
+    // that a WorkItem. When a LLVM worker thread is done with its WorkItem,
+    // it will just shut down, which also frees all resources associated with
+    // the given LLVM module, and sends a message to the coordinator that the
+    // has been completed.
+    //
+    // Work Scheduling
+    // ===============
+    // The scheduler's goal is to minimize the time it takes to complete all
+    // work there is, however, we also want to keep memory consumption low
+    // if possible. These two goals are at odds with each other: If memory
+    // consumption were not an issue, we could just let the main thread produce
+    // LLVM WorkItems at full speed, assuring maximal utilization of
+    // Tokens/LLVM worker threads. However, since codegen usual is faster
+    // than LLVM processing, the queue of LLVM WorkItems would fill up and each
+    // WorkItem potentially holds on to a substantial amount of memory.
+    //
+    // So the actual goal is to always produce just enough LLVM WorkItems as
+    // not to starve our LLVM worker threads. That means, once we have enough
+    // WorkItems in our queue, we can block the main thread, so it does not
+    // produce more until we need them.
+    //
+    // Doing LLVM Work on the Main Thread
+    // ----------------------------------
+    // Since the main thread owns the compiler processes implicit `Token`, it is
+    // wasteful to keep it blocked without doing any work. Therefore, what we do
+    // in this case is: We spawn off an additional LLVM worker thread that helps
+    // reduce the queue. The work it is doing corresponds to the implicit
+    // `Token`. The coordinator will mark the main thread as being busy with
+    // LLVM work. (The actual work happens on another OS thread but we just care
+    // about `Tokens`, not actual threads).
+    //
+    // When any LLVM worker thread finishes while the main thread is marked as
+    // "busy with LLVM work", we can do a little switcheroo: We give the Token
+    // of the just finished thread to the LLVM worker thread that is working on
+    // behalf of the main thread's implicit Token, thus freeing up the main
+    // thread again. The coordinator can then again decide what the main thread
+    // should do. This allows the coordinator to make decisions at more points
+    // in time.
+    //
+    // Striking a Balance between Throughput and Memory Consumption
+    // ------------------------------------------------------------
+    // Since our two goals, (1) use as many Tokens as possible and (2) keep
+    // memory consumption as low as possible, are in conflict with each other,
+    // we have to find a trade off between them. Right now, the goal is to keep
+    // all workers busy, which means that no worker should find the queue empty
+    // when it is ready to start.
+    // How do we do achieve this? Good question :) We actually never know how
+    // many `Tokens` are potentially available so it's hard to say how much to
+    // fill up the queue before switching the main thread to LLVM work. Also we
+    // currently don't have a means to estimate how long a running LLVM worker
+    // will still be busy with it's current WorkItem. However, we know the
+    // maximal count of available Tokens that makes sense (=the number of CPU
+    // cores), so we can take a conservative guess. The heuristic we use here
+    // is implemented in the `queue_full_enough()` function.
+    //
+    // Some Background on Jobservers
+    // -----------------------------
+    // It's worth also touching on the management of parallelism here. We don't
+    // want to just spawn a thread per work item because while that's optimal
+    // parallelism it may overload a system with too many threads or violate our
+    // configuration for the maximum amount of cpu to use for this process. To
+    // manage this we use the `jobserver` crate.
+    //
+    // Job servers are an artifact of GNU make and are used to manage
+    // parallelism between processes. A jobserver is a glorified IPC semaphore
+    // basically. Whenever we want to run some work we acquire the semaphore,
+    // and whenever we're done with that work we release the semaphore. In this
+    // manner we can ensure that the maximum number of parallel workers is
+    // capped at any one point in time.
+    //
+    // LTO and the coordinator thread
+    // ------------------------------
+    //
+    // The final job the coordinator thread is responsible for is managing LTO
+    // and how that works. When LTO is requested what we'll to is collect all
+    // optimized LLVM modules into a local vector on the coordinator. Once all
+    // modules have been codegened and optimized we hand this to the `lto`
+    // module for further optimization. The `lto` module will return back a list
+    // of more modules to work on, which the coordinator will continue to spawn
+    // work for.
+    //
+    // Each LLVM module is automatically sent back to the coordinator for LTO if
+    // necessary. There's already optimizations in place to avoid sending work
+    // back to the coordinator if LTO isn't requested.
+    return thread::spawn(move || {
+        // We pretend to be within the top-level LLVM time-passes task here:
+        set_time_depth(1);
+
+        let max_workers = ::num_cpus::get();
+        let mut worker_id_counter = 0;
+        let mut free_worker_ids = Vec::new();
+        let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
+            if let Some(id) = free_worker_ids.pop() {
+                id
+            } else {
+                let id = worker_id_counter;
+                worker_id_counter += 1;
+                id
+            }
+        };
+
+        // This is where we collect codegen units that have gone all the way
+        // through codegen and LLVM.
+        let mut compiled_modules = vec![];
+        let mut compiled_metadata_module = None;
+        let mut compiled_allocator_module = None;
+        let mut needs_lto = Vec::new();
+        let mut lto_import_only_modules = Vec::new();
+        let mut started_lto = false;
+        let mut codegen_aborted = false;
+
+        // This flag tracks whether all items have gone through codegens
+        let mut codegen_done = false;
+
+        // This is the queue of LLVM work items that still need processing.
+        let mut work_items = Vec::<(WorkItem<B>, u64)>::new();
+
+        // This are the Jobserver Tokens we currently hold. Does not include
+        // the implicit Token the compiler process owns no matter what.
+        let mut tokens = Vec::new();
+
+        let mut main_thread_worker_state = MainThreadWorkerState::Idle;
+        let mut running = 0;
+
+        let mut llvm_start_time = None;
+
+        // Run the message loop while there's still anything that needs message
+        // processing. Note that as soon as codegen is aborted we simply want to
+        // wait for all existing work to finish, so many of the conditions here
+        // only apply if codegen hasn't been aborted as they represent pending
+        // work to be done.
+        while !codegen_done ||
+              running > 0 ||
+              (!codegen_aborted && (
+                  work_items.len() > 0 ||
+                  needs_lto.len() > 0 ||
+                  lto_import_only_modules.len() > 0 ||
+                  main_thread_worker_state != MainThreadWorkerState::Idle
+              ))
+        {
+
+            // While there are still CGUs to be codegened, the coordinator has
+            // to decide how to utilize the compiler processes implicit Token:
+            // For codegenning more CGU or for running them through LLVM.
+            if !codegen_done {
+                if main_thread_worker_state == MainThreadWorkerState::Idle {
+                    if !queue_full_enough(work_items.len(), running, max_workers) {
+                        // The queue is not full enough, codegen more items:
+                        if let Err(_) = codegen_worker_send.send(Message::CodegenItem) {
+                            panic!("Could not send Message::CodegenItem to main thread")
+                        }
+                        main_thread_worker_state = MainThreadWorkerState::Codegenning;
+                    } else {
+                        // The queue is full enough to not let the worker
+                        // threads starve. Use the implicit Token to do some
+                        // LLVM work too.
+                        let (item, _) = work_items.pop()
+                            .expect("queue empty - queue_full_enough() broken?");
+                        let cgcx = CodegenContext {
+                            worker: get_worker_id(&mut free_worker_ids),
+                            .. cgcx.clone()
+                        };
+                        maybe_start_llvm_timer(cgcx.config(item.module_kind()),
+                                               &mut llvm_start_time);
+                        main_thread_worker_state = MainThreadWorkerState::LLVMing;
+                        spawn_work(cgcx, item);
+                    }
+                }
+            } else if codegen_aborted {
+                // don't queue up any more work if codegen was aborted, we're
+                // just waiting for our existing children to finish
+            } else {
+                // If we've finished everything related to normal codegen
+                // then it must be the case that we've got some LTO work to do.
+                // Perform the serial work here of figuring out what we're
+                // going to LTO and then push a bunch of work items onto our
+                // queue to do LTO
+                if work_items.len() == 0 &&
+                   running == 0 &&
+                   main_thread_worker_state == MainThreadWorkerState::Idle {
+                    assert!(!started_lto);
+                    assert!(needs_lto.len() + lto_import_only_modules.len() > 0);
+                    started_lto = true;
+                    let modules = mem::replace(&mut needs_lto, Vec::new());
+                    let import_only_modules =
+                        mem::replace(&mut lto_import_only_modules, Vec::new());
+                    for (work, cost) in generate_lto_work(&cgcx, modules, import_only_modules) {
+                        let insertion_index = work_items
+                            .binary_search_by_key(&cost, |&(_, cost)| cost)
+                            .unwrap_or_else(|e| e);
+                        work_items.insert(insertion_index, (work, cost));
+                        if !cgcx.opts.debugging_opts.no_parallel_llvm {
+                            helper.request_token();
+                        }
+                    }
+                }
+
+                // In this branch, we know that everything has been codegened,
+                // so it's just a matter of determining whether the implicit
+                // Token is free to use for LLVM work.
+                match main_thread_worker_state {
+                    MainThreadWorkerState::Idle => {
+                        if let Some((item, _)) = work_items.pop() {
+                            let cgcx = CodegenContext {
+                                worker: get_worker_id(&mut free_worker_ids),
+                                .. cgcx.clone()
+                            };
+                            maybe_start_llvm_timer(cgcx.config(item.module_kind()),
+                                                   &mut llvm_start_time);
+                            main_thread_worker_state = MainThreadWorkerState::LLVMing;
+                            spawn_work(cgcx, item);
+                        } else {
+                            // There is no unstarted work, so let the main thread
+                            // take over for a running worker. Otherwise the
+                            // implicit token would just go to waste.
+                            // We reduce the `running` counter by one. The
+                            // `tokens.truncate()` below will take care of
+                            // giving the Token back.
+                            debug_assert!(running > 0);
+                            running -= 1;
+                            main_thread_worker_state = MainThreadWorkerState::LLVMing;
+                        }
+                    }
+                    MainThreadWorkerState::Codegenning => {
+                        bug!("codegen worker should not be codegenning after \
+                              codegen was already completed")
+                    }
+                    MainThreadWorkerState::LLVMing => {
+                        // Already making good use of that token
+                    }
+                }
+            }
+
+            // Spin up what work we can, only doing this while we've got available
+            // parallelism slots and work left to spawn.
+            while !codegen_aborted && work_items.len() > 0 && running < tokens.len() {
+                let (item, _) = work_items.pop().unwrap();
+
+                maybe_start_llvm_timer(cgcx.config(item.module_kind()),
+                                       &mut llvm_start_time);
+
+                let cgcx = CodegenContext {
+                    worker: get_worker_id(&mut free_worker_ids),
+                    .. cgcx.clone()
+                };
+
+                spawn_work(cgcx, item);
+                running += 1;
+            }
+
+            // Relinquish accidentally acquired extra tokens
+            tokens.truncate(running);
+
+            let msg = coordinator_receive.recv().unwrap();
+            match *msg.downcast::<Message<B>>().ok().unwrap() {
+                // Save the token locally and the next turn of the loop will use
+                // this to spawn a new unit of work, or it may get dropped
+                // immediately if we have no more work to spawn.
+                Message::Token(token) => {
+                    match token {
+                        Ok(token) => {
+                            tokens.push(token);
+
+                            if main_thread_worker_state == MainThreadWorkerState::LLVMing {
+                                // If the main thread token is used for LLVM work
+                                // at the moment, we turn that thread into a regular
+                                // LLVM worker thread, so the main thread is free
+                                // to react to codegen demand.
+                                main_thread_worker_state = MainThreadWorkerState::Idle;
+                                running += 1;
+                            }
+                        }
+                        Err(e) => {
+                            let msg = &format!("failed to acquire jobserver token: {}", e);
+                            shared_emitter.fatal(msg);
+                            // Exit the coordinator thread
+                            panic!("{}", msg)
+                        }
+                    }
+                }
+
+                Message::CodegenDone { llvm_work_item, cost } => {
+                    // We keep the queue sorted by estimated processing cost,
+                    // so that more expensive items are processed earlier. This
+                    // is good for throughput as it gives the main thread more
+                    // time to fill up the queue and it avoids scheduling
+                    // expensive items to the end.
+                    // Note, however, that this is not ideal for memory
+                    // consumption, as LLVM module sizes are not evenly
+                    // distributed.
+                    let insertion_index =
+                        work_items.binary_search_by_key(&cost, |&(_, cost)| cost);
+                    let insertion_index = match insertion_index {
+                        Ok(idx) | Err(idx) => idx
+                    };
+                    work_items.insert(insertion_index, (llvm_work_item, cost));
+
+                    if !cgcx.opts.debugging_opts.no_parallel_llvm {
+                        helper.request_token();
+                    }
+                    assert!(!codegen_aborted);
+                    assert_eq!(main_thread_worker_state,
+                               MainThreadWorkerState::Codegenning);
+                    main_thread_worker_state = MainThreadWorkerState::Idle;
+                }
+
+                Message::CodegenComplete => {
+                    codegen_done = true;
+                    assert!(!codegen_aborted);
+                    assert_eq!(main_thread_worker_state,
+                               MainThreadWorkerState::Codegenning);
+                    main_thread_worker_state = MainThreadWorkerState::Idle;
+                }
+
+                // If codegen is aborted that means translation was aborted due
+                // to some normal-ish compiler error. In this situation we want
+                // to exit as soon as possible, but we want to make sure all
+                // existing work has finished. Flag codegen as being done, and
+                // then conditions above will ensure no more work is spawned but
+                // we'll keep executing this loop until `running` hits 0.
+                Message::CodegenAborted => {
+                    assert!(!codegen_aborted);
+                    codegen_done = true;
+                    codegen_aborted = true;
+                    assert_eq!(main_thread_worker_state,
+                               MainThreadWorkerState::Codegenning);
+                }
+
+                // If a thread exits successfully then we drop a token associated
+                // with that worker and update our `running` count. We may later
+                // re-acquire a token to continue running more work. We may also not
+                // actually drop a token here if the worker was running with an
+                // "ephemeral token"
+                //
+                // Note that if the thread failed that means it panicked, so we
+                // abort immediately.
+                Message::Done { result: Ok(compiled_module), worker_id } => {
+                    if main_thread_worker_state == MainThreadWorkerState::LLVMing {
+                        main_thread_worker_state = MainThreadWorkerState::Idle;
+                    } else {
+                        running -= 1;
+                    }
+
+                    free_worker_ids.push(worker_id);
+
+                    match compiled_module.kind {
+                        ModuleKind::Regular => {
+                            compiled_modules.push(compiled_module);
+                        }
+                        ModuleKind::Metadata => {
+                            assert!(compiled_metadata_module.is_none());
+                            compiled_metadata_module = Some(compiled_module);
+                        }
+                        ModuleKind::Allocator => {
+                            assert!(compiled_allocator_module.is_none());
+                            compiled_allocator_module = Some(compiled_module);
+                        }
+                    }
+                }
+                Message::NeedsLTO { result, worker_id } => {
+                    assert!(!started_lto);
+                    if main_thread_worker_state == MainThreadWorkerState::LLVMing {
+                        main_thread_worker_state = MainThreadWorkerState::Idle;
+                    } else {
+                        running -= 1;
+                    }
+                    free_worker_ids.push(worker_id);
+                    needs_lto.push(result);
+                }
+                Message::AddImportOnlyModule { module_data, work_product } => {
+                    assert!(!started_lto);
+                    assert!(!codegen_done);
+                    assert_eq!(main_thread_worker_state,
+                               MainThreadWorkerState::Codegenning);
+                    lto_import_only_modules.push((module_data, work_product));
+                    main_thread_worker_state = MainThreadWorkerState::Idle;
+                }
+                Message::Done { result: Err(()), worker_id: _ } => {
+                    bug!("worker thread panicked");
+                }
+                Message::CodegenItem => {
+                    bug!("the coordinator should not receive codegen requests")
+                }
+            }
+        }
+
+        if let Some(llvm_start_time) = llvm_start_time {
+            let total_llvm_time = Instant::now().duration_since(llvm_start_time);
+            // This is the top-level timing for all of LLVM, set the time-depth
+            // to zero.
+            set_time_depth(0);
+            print_time_passes_entry(cgcx.time_passes,
+                                    "LLVM passes",
+                                    total_llvm_time);
+        }
+
+        // Regardless of what order these modules completed in, report them to
+        // the backend in the same order every time to ensure that we're handing
+        // out deterministic results.
+        compiled_modules.sort_by(|a, b| a.name.cmp(&b.name));
+
+        let compiled_metadata_module = compiled_metadata_module
+            .expect("Metadata module not compiled?");
+
+        Ok(CompiledModules {
+            modules: compiled_modules,
+            metadata_module: compiled_metadata_module,
+            allocator_module: compiled_allocator_module,
+        })
+    });
+
+    // A heuristic that determines if we have enough LLVM WorkItems in the
+    // queue so that the main thread can do LLVM work instead of codegen
+    fn queue_full_enough(items_in_queue: usize,
+                         workers_running: usize,
+                         max_workers: usize) -> bool {
+        // Tune me, plz.
+        items_in_queue > 0 &&
+        items_in_queue >= max_workers.saturating_sub(workers_running / 2)
+    }
+
+    fn maybe_start_llvm_timer(config: &ModuleConfig,
+                              llvm_start_time: &mut Option<Instant>) {
+        // We keep track of the -Ztime-passes output manually,
+        // since the closure-based interface does not fit well here.
+        if config.time_passes {
+            if llvm_start_time.is_none() {
+                *llvm_start_time = Some(Instant::now());
+            }
+        }
+    }
+}
+
+pub const CODEGEN_WORKER_ID: usize = ::std::usize::MAX;
+pub const CODEGEN_WORKER_TIMELINE: time_graph::TimelineId =
+    time_graph::TimelineId(CODEGEN_WORKER_ID);
+pub const CODEGEN_WORK_PACKAGE_KIND: time_graph::WorkPackageKind =
+    time_graph::WorkPackageKind(&["#DE9597", "#FED1D3", "#FDC5C7", "#B46668", "#88494B"]);
+const LLVM_WORK_PACKAGE_KIND: time_graph::WorkPackageKind =
+    time_graph::WorkPackageKind(&["#7DB67A", "#C6EEC4", "#ACDAAA", "#579354", "#3E6F3C"]);
+
+fn spawn_work<B: ExtraBackendMethods>(
+    cgcx: CodegenContext<B>,
+    work: WorkItem<B>
+) {
+    let depth = time_depth();
+
+    thread::spawn(move || {
+        set_time_depth(depth);
+
+        // Set up a destructor which will fire off a message that we're done as
+        // we exit.
+        struct Bomb<B: ExtraBackendMethods> {
+            coordinator_send: Sender<Box<dyn Any + Send>>,
+            result: Option<WorkItemResult<B::Module>>,
+            worker_id: usize,
+        }
+        impl<B: ExtraBackendMethods> Drop for Bomb<B> {
+            fn drop(&mut self) {
+                let worker_id = self.worker_id;
+                let msg = match self.result.take() {
+                    Some(WorkItemResult::Compiled(m)) => {
+                        Message::Done::<B> { result: Ok(m), worker_id }
+                    }
+                    Some(WorkItemResult::NeedsLTO(m)) => {
+                        Message::NeedsLTO::<B> { result: m, worker_id }
+                    }
+                    None => Message::Done::<B> { result: Err(()), worker_id }
+                };
+                drop(self.coordinator_send.send(Box::new(msg)));
+            }
+        }
+
+        let mut bomb = Bomb::<B> {
+            coordinator_send: cgcx.coordinator_send.clone(),
+            result: None,
+            worker_id: cgcx.worker,
+        };
+
+        // Execute the work itself, and if it finishes successfully then flag
+        // ourselves as a success as well.
+        //
+        // Note that we ignore any `FatalError` coming out of `execute_work_item`,
+        // as a diagnostic was already sent off to the main thread - just
+        // surface that there was an error in this worker.
+        bomb.result = {
+            let timeline = cgcx.time_graph.as_ref().map(|tg| {
+                tg.start(time_graph::TimelineId(cgcx.worker),
+                         LLVM_WORK_PACKAGE_KIND,
+                         &work.name())
+            });
+            let mut timeline = timeline.unwrap_or(Timeline::noop());
+            execute_work_item(&cgcx, work, &mut timeline).ok()
+        };
+    });
+}
+
+pub fn run_assembler<B: ExtraBackendMethods>(
+    cgcx: &CodegenContext<B>,
+    handler: &Handler,
+    assembly: &Path,
+    object: &Path
+) {
+    let assembler = cgcx.assembler_cmd
+        .as_ref()
+        .expect("cgcx.assembler_cmd is missing?");
+
+    let pname = &assembler.name;
+    let mut cmd = assembler.cmd.clone();
+    cmd.arg("-c").arg("-o").arg(object).arg(assembly);
+    debug!("{:?}", cmd);
+
+    match cmd.output() {
+        Ok(prog) => {
+            if !prog.status.success() {
+                let mut note = prog.stderr.clone();
+                note.extend_from_slice(&prog.stdout);
+
+                handler.struct_err(&format!("linking with `{}` failed: {}",
+                                            pname.display(),
+                                            prog.status))
+                    .note(&format!("{:?}", &cmd))
+                    .note(str::from_utf8(&note[..]).unwrap())
+                    .emit();
+                handler.abort_if_errors();
+            }
+        },
+        Err(e) => {
+            handler.err(&format!("could not exec the linker `{}`: {}", pname.display(), e));
+            handler.abort_if_errors();
+        }
+    }
+}
+
+
+enum SharedEmitterMessage {
+    Diagnostic(Diagnostic),
+    InlineAsmError(u32, String),
+    AbortIfErrors,
+    Fatal(String),
+}
+
+#[derive(Clone)]
+pub struct SharedEmitter {
+    sender: Sender<SharedEmitterMessage>,
+}
+
+pub struct SharedEmitterMain {
+    receiver: Receiver<SharedEmitterMessage>,
+}
+
+impl SharedEmitter {
+    pub fn new() -> (SharedEmitter, SharedEmitterMain) {
+        let (sender, receiver) = channel();
+
+        (SharedEmitter { sender }, SharedEmitterMain { receiver })
+    }
+
+    pub fn inline_asm_error(&self, cookie: u32, msg: String) {
+        drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg)));
+    }
+
+    pub fn fatal(&self, msg: &str) {
+        drop(self.sender.send(SharedEmitterMessage::Fatal(msg.to_string())));
+    }
+}
+
+impl Emitter for SharedEmitter {
+    fn emit(&mut self, db: &DiagnosticBuilder) {
+        drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
+            msg: db.message(),
+            code: db.code.clone(),
+            lvl: db.level,
+        })));
+        for child in &db.children {
+            drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
+                msg: child.message(),
+                code: None,
+                lvl: child.level,
+            })));
+        }
+        drop(self.sender.send(SharedEmitterMessage::AbortIfErrors));
+    }
+}
+
+impl SharedEmitterMain {
+    pub fn check(&self, sess: &Session, blocking: bool) {
+        loop {
+            let message = if blocking {
+                match self.receiver.recv() {
+                    Ok(message) => Ok(message),
+                    Err(_) => Err(()),
+                }
+            } else {
+                match self.receiver.try_recv() {
+                    Ok(message) => Ok(message),
+                    Err(_) => Err(()),
+                }
+            };
+
+            match message {
+                Ok(SharedEmitterMessage::Diagnostic(diag)) => {
+                    let handler = sess.diagnostic();
+                    match diag.code {
+                        Some(ref code) => {
+                            handler.emit_with_code(&MultiSpan::new(),
+                                                   &diag.msg,
+                                                   code.clone(),
+                                                   diag.lvl);
+                        }
+                        None => {
+                            handler.emit(&MultiSpan::new(),
+                                         &diag.msg,
+                                         diag.lvl);
+                        }
+                    }
+                }
+                Ok(SharedEmitterMessage::InlineAsmError(cookie, msg)) => {
+                    match Mark::from_u32(cookie).expn_info() {
+                        Some(ei) => sess.span_err(ei.call_site, &msg),
+                        None     => sess.err(&msg),
+                    }
+                }
+                Ok(SharedEmitterMessage::AbortIfErrors) => {
+                    sess.abort_if_errors();
+                }
+                Ok(SharedEmitterMessage::Fatal(msg)) => {
+                    sess.fatal(&msg);
+                }
+                Err(_) => {
+                    break;
+                }
+            }
+
+        }
+    }
+}
+
+pub struct OngoingCodegen<B: ExtraBackendMethods> {
+    pub backend: B,
+    pub crate_name: Symbol,
+    pub crate_hash: Svh,
+    pub metadata: EncodedMetadata,
+    pub windows_subsystem: Option<String>,
+    pub linker_info: LinkerInfo,
+    pub crate_info: CrateInfo,
+    pub time_graph: Option<TimeGraph>,
+    pub coordinator_send: Sender<Box<dyn Any + Send>>,
+    pub codegen_worker_receive: Receiver<Message<B>>,
+    pub shared_emitter_main: SharedEmitterMain,
+    pub future: thread::JoinHandle<Result<CompiledModules, ()>>,
+    pub output_filenames: Arc<OutputFilenames>,
+}
+
+impl<B: ExtraBackendMethods> OngoingCodegen<B> {
+    pub fn join(
+        self,
+        sess: &Session
+    ) -> (CodegenResults, FxHashMap<WorkProductId, WorkProduct>) {
+        self.shared_emitter_main.check(sess, true);
+        let compiled_modules = match self.future.join() {
+            Ok(Ok(compiled_modules)) => compiled_modules,
+            Ok(Err(())) => {
+                sess.abort_if_errors();
+                panic!("expected abort due to worker thread errors")
+            },
+            Err(_) => {
+                bug!("panic during codegen/LLVM phase");
+            }
+        };
+
+        sess.cgu_reuse_tracker.check_expected_reuse(sess);
+
+        sess.abort_if_errors();
+
+        if let Some(time_graph) = self.time_graph {
+            time_graph.dump(&format!("{}-timings", self.crate_name));
+        }
+
+        let work_products =
+            copy_all_cgu_workproducts_to_incr_comp_cache_dir(sess,
+                                                             &compiled_modules);
+        produce_final_output_artifacts(sess,
+                                       &compiled_modules,
+                                       &self.output_filenames);
+
+        // FIXME: time_llvm_passes support - does this use a global context or
+        // something?
+        if sess.codegen_units() == 1 && sess.time_llvm_passes() {
+            self.backend.print_pass_timings()
+        }
+
+        (CodegenResults {
+            crate_name: self.crate_name,
+            crate_hash: self.crate_hash,
+            metadata: self.metadata,
+            windows_subsystem: self.windows_subsystem,
+            linker_info: self.linker_info,
+            crate_info: self.crate_info,
+
+            modules: compiled_modules.modules,
+            allocator_module: compiled_modules.allocator_module,
+            metadata_module: compiled_modules.metadata_module,
+        }, work_products)
+    }
+
+    pub fn submit_pre_codegened_module_to_llvm(&self,
+                                                       tcx: TyCtxt,
+                                                       module: ModuleCodegen<B::Module>) {
+        self.wait_for_signal_to_codegen_item();
+        self.check_for_errors(tcx.sess);
+
+        // These are generally cheap and won't through off scheduling.
+        let cost = 0;
+        submit_codegened_module_to_llvm(&self.backend, tcx, module, cost);
+    }
+
+    pub fn codegen_finished(&self, tcx: TyCtxt) {
+        self.wait_for_signal_to_codegen_item();
+        self.check_for_errors(tcx.sess);
+        drop(self.coordinator_send.send(Box::new(Message::CodegenComplete::<B>)));
+    }
+
+    /// Consume this context indicating that codegen was entirely aborted, and
+    /// we need to exit as quickly as possible.
+    ///
+    /// This method blocks the current thread until all worker threads have
+    /// finished, and all worker threads should have exited or be real close to
+    /// exiting at this point.
+    pub fn codegen_aborted(self) {
+        // Signal to the coordinator it should spawn no more work and start
+        // shutdown.
+        drop(self.coordinator_send.send(Box::new(Message::CodegenAborted::<B>)));
+        drop(self.future.join());
+    }
+
+    pub fn check_for_errors(&self, sess: &Session) {
+        self.shared_emitter_main.check(sess, false);
+    }
+
+    pub fn wait_for_signal_to_codegen_item(&self) {
+        match self.codegen_worker_receive.recv() {
+            Ok(Message::CodegenItem) => {
+                // Nothing to do
+            }
+            Ok(_) => panic!("unexpected message"),
+            Err(_) => {
+                // One of the LLVM threads must have panicked, fall through so
+                // error handling can be reached.
+            }
+        }
+    }
+}
+
+pub fn submit_codegened_module_to_llvm<B: ExtraBackendMethods>(
+    _backend: &B,
+    tcx: TyCtxt,
+    module: ModuleCodegen<B::Module>,
+    cost: u64
+) {
+    let llvm_work_item = WorkItem::Optimize(module);
+    drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::CodegenDone::<B> {
+        llvm_work_item,
+        cost,
+    })));
+}
+
+pub fn submit_post_lto_module_to_llvm<B: ExtraBackendMethods>(
+    _backend: &B,
+    tcx: TyCtxt,
+    module: CachedModuleCodegen
+) {
+    let llvm_work_item = WorkItem::CopyPostLtoArtifacts(module);
+    drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::CodegenDone::<B> {
+        llvm_work_item,
+        cost: 0,
+    })));
+}
+
+pub fn submit_pre_lto_module_to_llvm<B: ExtraBackendMethods>(
+    _backend: &B,
+    tcx: TyCtxt,
+    module: CachedModuleCodegen
+) {
+    let filename = pre_lto_bitcode_filename(&module.name);
+    let bc_path = in_incr_comp_dir_sess(tcx.sess, &filename);
+    let file = fs::File::open(&bc_path).unwrap_or_else(|e| {
+        panic!("failed to open bitcode file `{}`: {}", bc_path.display(), e)
+    });
+
+    let mmap = unsafe {
+        memmap::Mmap::map(&file).unwrap_or_else(|e| {
+            panic!("failed to mmap bitcode file `{}`: {}", bc_path.display(), e)
+        })
+    };
+    // Schedule the module to be loaded
+    drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::AddImportOnlyModule::<B> {
+        module_data: SerializedModule::FromUncompressedFile(mmap),
+        work_product: module.source,
+    })));
+}
+
+pub fn pre_lto_bitcode_filename(module_name: &str) -> String {
+    format!("{}.{}", module_name, PRE_THIN_LTO_BC_EXT)
+}
+
+fn msvc_imps_needed(tcx: TyCtxt) -> bool {
+    // This should never be true (because it's not supported). If it is true,
+    // something is wrong with commandline arg validation.
+    assert!(!(tcx.sess.opts.debugging_opts.cross_lang_lto.enabled() &&
+              tcx.sess.target.target.options.is_like_msvc &&
+              tcx.sess.opts.cg.prefer_dynamic));
+
+    tcx.sess.target.target.options.is_like_msvc &&
+        tcx.sess.crate_types.borrow().iter().any(|ct| *ct == config::CrateType::Rlib) &&
+    // ThinLTO can't handle this workaround in all cases, so we don't
+    // emit the `__imp_` symbols. Instead we make them unnecessary by disallowing
+    // dynamic linking when cross-language LTO is enabled.
+    !tcx.sess.opts.debugging_opts.cross_lang_lto.enabled()
+}
diff --git a/src/librustc_codegen_ssa/base.rs b/src/librustc_codegen_ssa/base.rs
index 81a2d0a5389..a590dcd3ea8 100644
--- a/src/librustc_codegen_ssa/base.rs
+++ b/src/librustc_codegen_ssa/base.rs
@@ -39,6 +39,8 @@ use rustc::util::profiling::ProfileCategory;
 use rustc::session::config::{self, EntryFnType, Lto};
 use rustc::session::Session;
 use mir::place::PlaceRef;
+use back::write::{OngoingCodegen, start_async_codegen, submit_pre_lto_module_to_llvm,
+    submit_post_lto_module_to_llvm};
 use {MemFlags, CrateInfo};
 use callee;
 use rustc_mir::monomorphize::item::DefPathBasedNames;
@@ -556,7 +558,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
     backend: B,
     tcx: TyCtxt<'a, 'tcx, 'tcx>,
     rx: mpsc::Receiver<Box<dyn Any + Send>>
-) -> B::OngoingCodegen {
+) -> OngoingCodegen<B> {
 
     check_for_rustc_errors_attr(tcx);
 
@@ -590,19 +592,20 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
     // Skip crate items and just output metadata in -Z no-codegen mode.
     if tcx.sess.opts.debugging_opts.no_codegen ||
        !tcx.sess.opts.output_types.should_codegen() {
-        let ongoing_codegen = backend.start_async_codegen(
+        let ongoing_codegen = start_async_codegen(
+            backend,
             tcx,
             time_graph,
             metadata,
             rx,
             1);
 
-        backend.submit_pre_codegened_module_to_backend(&ongoing_codegen, tcx, metadata_module);
-        backend.codegen_finished(&ongoing_codegen, tcx);
+        ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
+        ongoing_codegen.codegen_finished(tcx);
 
         assert_and_save_dep_graph(tcx);
 
-        backend.check_for_errors(&ongoing_codegen, tcx.sess);
+        ongoing_codegen.check_for_errors(tcx.sess);
 
         return ongoing_codegen;
     }
@@ -623,7 +626,8 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
         }
     }
 
-    let ongoing_codegen = backend.start_async_codegen(
+    let ongoing_codegen = start_async_codegen(
+        backend.clone(),
         tcx,
         time_graph.clone(),
         metadata,
@@ -667,10 +671,10 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
     };
 
     if let Some(allocator_module) = allocator_module {
-        backend.submit_pre_codegened_module_to_backend(&ongoing_codegen, tcx, allocator_module);
+        ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, allocator_module);
     }
 
-    backend.submit_pre_codegened_module_to_backend(&ongoing_codegen, tcx, metadata_module);
+    ongoing_codegen.submit_pre_codegened_module_to_llvm(tcx, metadata_module);
 
     // We sort the codegen units by size. This way we can schedule work for LLVM
     // a bit more efficiently.
@@ -684,8 +688,8 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
     let mut all_stats = Stats::default();
 
     for cgu in codegen_units.into_iter() {
-        backend.wait_for_signal_to_codegen_item(&ongoing_codegen);
-        backend.check_for_errors(&ongoing_codegen, tcx.sess);
+        ongoing_codegen.wait_for_signal_to_codegen_item();
+        ongoing_codegen.check_for_errors(tcx.sess);
 
         let cgu_reuse = determine_cgu_reuse(tcx, &cgu);
         tcx.sess.cgu_reuse_tracker.set_actual_reuse(&cgu.name().as_str(), cgu_reuse);
@@ -704,14 +708,14 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
                 false
             }
             CguReuse::PreLto => {
-                backend.submit_pre_lto_module_to_backend(tcx, CachedModuleCodegen {
+                submit_pre_lto_module_to_llvm(&backend, tcx, CachedModuleCodegen {
                     name: cgu.name().to_string(),
                     source: cgu.work_product(tcx),
                 });
                 true
             }
             CguReuse::PostLto => {
-                backend.submit_post_lto_module_to_backend(tcx, CachedModuleCodegen {
+                submit_post_lto_module_to_llvm(&backend, tcx, CachedModuleCodegen {
                     name: cgu.name().to_string(),
                     source: cgu.work_product(tcx),
                 });
@@ -720,7 +724,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
         };
     }
 
-    backend.codegen_finished(&ongoing_codegen, tcx);
+    ongoing_codegen.codegen_finished(tcx);
 
     // Since the main thread is sometimes blocked during codegen, we keep track
     // -Ztime-passes output manually.
@@ -754,7 +758,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
         }
     }
 
-    backend.check_for_errors(&ongoing_codegen, tcx.sess);
+    ongoing_codegen.check_for_errors(tcx.sess);
 
     assert_and_save_dep_graph(tcx);
     ongoing_codegen.into_inner()
@@ -777,24 +781,24 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
 /// If you see this comment in the code, then it means that this workaround
 /// worked! We may yet one day track down the mysterious cause of that
 /// segfault...
-struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<B::OngoingCodegen>);
+struct AbortCodegenOnDrop<B: ExtraBackendMethods>(Option<OngoingCodegen<B>>);
 
 impl<B: ExtraBackendMethods> AbortCodegenOnDrop<B> {
-    fn into_inner(mut self) -> B::OngoingCodegen {
+    fn into_inner(mut self) -> OngoingCodegen<B> {
         self.0.take().unwrap()
     }
 }
 
 impl<B: ExtraBackendMethods> Deref for AbortCodegenOnDrop<B> {
-    type Target = B::OngoingCodegen;
+    type Target = OngoingCodegen<B>;
 
-    fn deref(&self) -> &B::OngoingCodegen {
+    fn deref(&self) -> &OngoingCodegen<B> {
         self.0.as_ref().unwrap()
     }
 }
 
 impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
-    fn deref_mut(&mut self) -> &mut B::OngoingCodegen {
+    fn deref_mut(&mut self) -> &mut OngoingCodegen<B> {
         self.0.as_mut().unwrap()
     }
 }
@@ -802,7 +806,7 @@ impl<B: ExtraBackendMethods> DerefMut for AbortCodegenOnDrop<B> {
 impl<B: ExtraBackendMethods> Drop for AbortCodegenOnDrop<B> {
     fn drop(&mut self) {
         if let Some(codegen) = self.0.take() {
-            B::codegen_aborted(codegen);
+            codegen.codegen_aborted();
         }
     }
 }
diff --git a/src/librustc_codegen_ssa/interfaces/backend.rs b/src/librustc_codegen_ssa/interfaces/backend.rs
index 6e8233bcb88..b4d376cf5f0 100644
--- a/src/librustc_codegen_ssa/interfaces/backend.rs
+++ b/src/librustc_codegen_ssa/interfaces/backend.rs
@@ -11,18 +11,16 @@
 use rustc::ty::layout::{HasTyCtxt, LayoutOf, TyLayout};
 use rustc::ty::Ty;
 
+use super::write::WriteBackendMethods;
 use super::CodegenObject;
 use rustc::middle::allocator::AllocatorKind;
 use rustc::middle::cstore::EncodedMetadata;
 use rustc::mir::mono::Stats;
 use rustc::session::Session;
 use rustc::ty::TyCtxt;
-use rustc::util::time_graph::TimeGraph;
 use rustc_codegen_utils::codegen_backend::CodegenBackend;
-use std::any::Any;
-use std::sync::mpsc::Receiver;
+use std::sync::Arc;
 use syntax_pos::symbol::InternedString;
-use {CachedModuleCodegen, ModuleCodegen};
 
 pub trait BackendTypes {
     type Value: CodegenObject;
@@ -43,10 +41,7 @@ impl<'tcx, T> Backend<'tcx> for T where
     Self: BackendTypes + HasTyCtxt<'tcx> + LayoutOf<Ty = Ty<'tcx>, TyLayout = TyLayout<'tcx>>
 {}
 
-pub trait ExtraBackendMethods: CodegenBackend {
-    type Module;
-    type OngoingCodegen;
-
+pub trait ExtraBackendMethods: CodegenBackend + WriteBackendMethods + Sized + Send {
     fn new_metadata(&self, sess: &Session, mod_name: &str) -> Self::Module;
     fn write_metadata<'b, 'gcx>(
         &self,
@@ -54,30 +49,18 @@ pub trait ExtraBackendMethods: CodegenBackend {
         metadata: &Self::Module,
     ) -> EncodedMetadata;
     fn codegen_allocator(&self, tcx: TyCtxt, mods: &Self::Module, kind: AllocatorKind);
-
-    fn start_async_codegen(
-        &self,
-        tcx: TyCtxt,
-        time_graph: Option<TimeGraph>,
-        metadata: EncodedMetadata,
-        coordinator_receive: Receiver<Box<dyn Any + Send>>,
-        total_cgus: usize,
-    ) -> Self::OngoingCodegen;
-    fn submit_pre_codegened_module_to_backend(
-        &self,
-        codegen: &Self::OngoingCodegen,
-        tcx: TyCtxt,
-        module: ModuleCodegen<Self::Module>,
-    );
-    fn submit_pre_lto_module_to_backend(&self, tcx: TyCtxt, module: CachedModuleCodegen);
-    fn submit_post_lto_module_to_backend(&self, tcx: TyCtxt, module: CachedModuleCodegen);
-    fn codegen_aborted(codegen: Self::OngoingCodegen);
-    fn codegen_finished(&self, codegen: &Self::OngoingCodegen, tcx: TyCtxt);
-    fn check_for_errors(&self, codegen: &Self::OngoingCodegen, sess: &Session);
-    fn wait_for_signal_to_codegen_item(&self, codegen: &Self::OngoingCodegen);
     fn compile_codegen_unit<'a, 'tcx: 'a>(
         &self,
         tcx: TyCtxt<'a, 'tcx, 'tcx>,
         cgu_name: InternedString,
     ) -> Stats;
+    // If find_features is true this won't access `sess.crate_types` by assuming
+    // that `is_pie_binary` is false. When we discover LLVM target features
+    // `sess.crate_types` is uninitialized so we cannot access it.
+    fn target_machine_factory(
+        &self,
+        sess: &Session,
+        find_features: bool,
+    ) -> Arc<dyn Fn() -> Result<Self::TargetMachine, String> + Send + Sync>;
+    fn target_cpu<'b>(&self, sess: &'b Session) -> &'b str;
 }
diff --git a/src/librustc_codegen_ssa/interfaces/mod.rs b/src/librustc_codegen_ssa/interfaces/mod.rs
index 1797060f6a4..5cff31e17b5 100644
--- a/src/librustc_codegen_ssa/interfaces/mod.rs
+++ b/src/librustc_codegen_ssa/interfaces/mod.rs
@@ -35,6 +35,7 @@ mod intrinsic;
 mod misc;
 mod statics;
 mod type_;
+mod write;
 
 pub use self::abi::{AbiBuilderMethods, AbiMethods};
 pub use self::asm::{AsmBuilderMethods, AsmMethods};
@@ -49,6 +50,7 @@ pub use self::statics::StaticMethods;
 pub use self::type_::{
     ArgTypeMethods, BaseTypeMethods, DerivedTypeMethods, LayoutTypeMethods, TypeMethods,
 };
+pub use self::write::{ModuleBufferMethods, ThinBufferMethods, WriteBackendMethods};
 
 use std::fmt;
 
diff --git a/src/librustc_codegen_ssa/interfaces/write.rs b/src/librustc_codegen_ssa/interfaces/write.rs
new file mode 100644
index 00000000000..3419e1c59ed
--- /dev/null
+++ b/src/librustc_codegen_ssa/interfaces/write.rs
@@ -0,0 +1,72 @@
+// Copyright 2018 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+use back::lto::{LtoModuleCodegen, SerializedModule, ThinModule};
+use back::write::{CodegenContext, ModuleConfig};
+use {CompiledModule, ModuleCodegen};
+
+use rustc::dep_graph::WorkProduct;
+use rustc::util::time_graph::Timeline;
+use rustc_errors::{FatalError, Handler};
+
+pub trait WriteBackendMethods: 'static + Sized + Clone {
+    type Module: Send + Sync;
+    type TargetMachine: Clone;
+    type ModuleBuffer: ModuleBufferMethods;
+    type Context: ?Sized;
+    type ThinData: Send + Sync;
+    type ThinBuffer: ThinBufferMethods;
+
+    /// Performs LTO, which in the case of full LTO means merging all modules into
+    /// a single one and returning it for further optimizing. For ThinLTO, it will
+    /// do the global analysis necessary and return two lists, one of the modules
+    /// the need optimization and another for modules that can simply be copied over
+    /// from the incr. comp. cache.
+    fn run_lto(
+        cgcx: &CodegenContext<Self>,
+        modules: Vec<ModuleCodegen<Self::Module>>,
+        cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>,
+        timeline: &mut Timeline,
+    ) -> Result<(Vec<LtoModuleCodegen<Self>>, Vec<WorkProduct>), FatalError>;
+    fn print_pass_timings(&self);
+    unsafe fn optimize(
+        cgcx: &CodegenContext<Self>,
+        diag_handler: &Handler,
+        module: &ModuleCodegen<Self::Module>,
+        config: &ModuleConfig,
+        timeline: &mut Timeline,
+    ) -> Result<(), FatalError>;
+    unsafe fn optimize_thin(
+        cgcx: &CodegenContext<Self>,
+        thin: &mut ThinModule<Self>,
+        timeline: &mut Timeline,
+    ) -> Result<ModuleCodegen<Self::Module>, FatalError>;
+    unsafe fn codegen(
+        cgcx: &CodegenContext<Self>,
+        diag_handler: &Handler,
+        module: ModuleCodegen<Self::Module>,
+        config: &ModuleConfig,
+        timeline: &mut Timeline,
+    ) -> Result<CompiledModule, FatalError>;
+    fn run_lto_pass_manager(
+        cgcx: &CodegenContext<Self>,
+        llmod: &ModuleCodegen<Self::Module>,
+        config: &ModuleConfig,
+        thin: bool,
+    );
+}
+
+pub trait ThinBufferMethods: Send + Sync {
+    fn data(&self) -> &[u8];
+}
+
+pub trait ModuleBufferMethods: Send + Sync {
+    fn data(&self) -> &[u8];
+}
diff --git a/src/librustc_codegen_ssa/lib.rs b/src/librustc_codegen_ssa/lib.rs
index 2a42ad91e3d..e779d8f1469 100644
--- a/src/librustc_codegen_ssa/lib.rs
+++ b/src/librustc_codegen_ssa/lib.rs
@@ -39,7 +39,16 @@ extern crate syntax_pos;
 extern crate rustc_incremental;
 extern crate rustc_codegen_utils;
 extern crate rustc_data_structures;
+extern crate rustc_allocator;
+extern crate rustc_fs_util;
+extern crate serialize;
+extern crate rustc_errors;
+extern crate rustc_demangle;
+extern crate cc;
 extern crate libc;
+extern crate jobserver;
+extern crate memmap;
+extern crate num_cpus;
 
 use std::path::PathBuf;
 use rustc::dep_graph::WorkProduct;
@@ -48,7 +57,9 @@ use rustc::middle::lang_items::LangItem;
 use rustc::hir::def_id::CrateNum;
 use rustc_data_structures::fx::{FxHashMap, FxHashSet};
 use rustc_data_structures::sync::Lrc;
+use rustc_data_structures::svh::Svh;
 use rustc::middle::cstore::{LibSource, CrateSource, NativeLibrary};
+use syntax_pos::symbol::Symbol;
 
 // NB: This module needs to be declared first so diagnostics are
 // registered before they are used.
@@ -63,6 +74,7 @@ pub mod callee;
 pub mod glue;
 pub mod meth;
 pub mod mono_item;
+pub mod back;
 
 pub struct ModuleCodegen<M> {
     /// The name of the module. When the crate may be saved between
@@ -159,4 +171,17 @@ pub struct CrateInfo {
     pub missing_lang_items: FxHashMap<CrateNum, Vec<LangItem>>,
 }
 
+
+pub struct CodegenResults {
+    pub crate_name: Symbol,
+    pub modules: Vec<CompiledModule>,
+    pub allocator_module: Option<CompiledModule>,
+    pub metadata_module: CompiledModule,
+    pub crate_hash: Svh,
+    pub metadata: rustc::middle::cstore::EncodedMetadata,
+    pub windows_subsystem: Option<String>,
+    pub linker_info: back::linker::LinkerInfo,
+    pub crate_info: CrateInfo,
+}
+
 __build_diagnostic_array! { librustc_codegen_ssa, DIAGNOSTICS }