diff options
| author | Irina Popa <irinagpopa@gmail.com> | 2018-05-08 16:10:16 +0300 |
|---|---|---|
| committer | Irina Popa <irinagpopa@gmail.com> | 2018-05-17 15:08:30 +0300 |
| commit | b63d7e2b1c4019e40051036bcb1fd5f254a8f6e2 (patch) | |
| tree | 314792e2f467d17181d29d4988550058197ac029 /src/librustc_codegen_llvm/back | |
| parent | e3150564f889a3bad01795d9fcb31d4f14d58a99 (diff) | |
| download | rust-b63d7e2b1c4019e40051036bcb1fd5f254a8f6e2.tar.gz rust-b63d7e2b1c4019e40051036bcb1fd5f254a8f6e2.zip | |
Rename trans to codegen everywhere.
Diffstat (limited to 'src/librustc_codegen_llvm/back')
| -rw-r--r-- | src/librustc_codegen_llvm/back/archive.rs | 325 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/bytecode.rs | 160 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/command.rs | 175 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/link.rs | 1630 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/linker.rs | 1037 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/lto.rs | 773 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/rpath.rs | 282 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/symbol_export.rs | 396 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/wasm.rs | 261 | ||||
| -rw-r--r-- | src/librustc_codegen_llvm/back/write.rs | 2390 |
10 files changed, 7429 insertions, 0 deletions
diff --git a/src/librustc_codegen_llvm/back/archive.rs b/src/librustc_codegen_llvm/back/archive.rs new file mode 100644 index 00000000000..609629bffb9 --- /dev/null +++ b/src/librustc_codegen_llvm/back/archive.rs @@ -0,0 +1,325 @@ +// Copyright 2013-2014 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 helper class for dealing with static archives + +use std::ffi::{CString, CStr}; +use std::io; +use std::mem; +use std::path::{Path, PathBuf}; +use std::ptr; +use std::str; + +use back::bytecode::RLIB_BYTECODE_EXTENSION; +use libc; +use llvm::archive_ro::{ArchiveRO, Child}; +use llvm::{self, ArchiveKind}; +use metadata::METADATA_FILENAME; +use rustc::session::Session; + +pub struct ArchiveConfig<'a> { + pub sess: &'a Session, + pub dst: PathBuf, + pub src: Option<PathBuf>, + pub lib_search_paths: Vec<PathBuf>, +} + +/// Helper for adding many files to an archive. +#[must_use = "must call build() to finish building the archive"] +pub struct ArchiveBuilder<'a> { + config: ArchiveConfig<'a>, + removals: Vec<String>, + additions: Vec<Addition>, + should_update_symbols: bool, + src_archive: Option<Option<ArchiveRO>>, +} + +enum Addition { + File { + path: PathBuf, + name_in_archive: String, + }, + Archive { + archive: ArchiveRO, + skip: Box<FnMut(&str) -> bool>, + }, +} + +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)); +} + +fn is_relevant_child(c: &Child) -> bool { + match c.name() { + Some(name) => !name.contains("SYMDEF"), + None => false, + } +} + +impl<'a> ArchiveBuilder<'a> { + /// Create a new static archive, ready for modifying the archive specified + /// by `config`. + pub fn new(config: ArchiveConfig<'a>) -> ArchiveBuilder<'a> { + ArchiveBuilder { + config, + removals: Vec::new(), + additions: Vec::new(), + should_update_symbols: false, + src_archive: None, + } + } + + /// Removes a file from this archive + pub fn remove_file(&mut self, file: &str) { + self.removals.push(file.to_string()); + } + + /// Lists all files in an archive + pub fn src_files(&mut self) -> Vec<String> { + if self.src_archive().is_none() { + return Vec::new() + } + let archive = self.src_archive.as_ref().unwrap().as_ref().unwrap(); + let ret = archive.iter() + .filter_map(|child| child.ok()) + .filter(is_relevant_child) + .filter_map(|child| child.name()) + .filter(|name| !self.removals.iter().any(|x| x == name)) + .map(|name| name.to_string()) + .collect(); + return ret; + } + + fn src_archive(&mut self) -> Option<&ArchiveRO> { + if let Some(ref a) = self.src_archive { + return a.as_ref() + } + let src = self.config.src.as_ref()?; + self.src_archive = Some(ArchiveRO::open(src).ok()); + self.src_archive.as_ref().unwrap().as_ref() + } + + /// Adds all of the contents of a native library to this archive. This will + /// search in the relevant locations for a library named `name`. + pub fn add_native_library(&mut self, name: &str) { + let location = find_library(name, &self.config.lib_search_paths, + self.config.sess); + self.add_archive(&location, |_| false).unwrap_or_else(|e| { + self.config.sess.fatal(&format!("failed to add native library {}: {}", + location.to_string_lossy(), e)); + }); + } + + /// Adds all of the contents of the rlib at the specified path to this + /// archive. + /// + /// This ignores adding the bytecode from the rlib, and if LTO is enabled + /// then the object file also isn't added. + pub fn add_rlib(&mut self, + rlib: &Path, + name: &str, + lto: bool, + skip_objects: bool) -> io::Result<()> { + // Ignoring obj file starting with the crate name + // as simple comparison is not enough - there + // might be also an extra name suffix + let obj_start = format!("{}", name); + + self.add_archive(rlib, move |fname: &str| { + // Ignore bytecode/metadata files, no matter the name. + if fname.ends_with(RLIB_BYTECODE_EXTENSION) || fname == METADATA_FILENAME { + return true + } + + // Don't include Rust objects if LTO is enabled + if lto && fname.starts_with(&obj_start) && fname.ends_with(".o") { + return true + } + + // Otherwise if this is *not* a rust object and we're skipping + // objects then skip this file + if skip_objects && (!fname.starts_with(&obj_start) || !fname.ends_with(".o")) { + return true + } + + // ok, don't skip this + return false + }) + } + + fn add_archive<F>(&mut self, archive: &Path, skip: F) + -> io::Result<()> + where F: FnMut(&str) -> bool + 'static + { + let archive = match ArchiveRO::open(archive) { + Ok(ar) => ar, + Err(e) => return Err(io::Error::new(io::ErrorKind::Other, e)), + }; + self.additions.push(Addition::Archive { + archive, + skip: Box::new(skip), + }); + Ok(()) + } + + /// Adds an arbitrary file to this archive + pub fn add_file(&mut self, file: &Path) { + let name = file.file_name().unwrap().to_str().unwrap(); + self.additions.push(Addition::File { + path: file.to_path_buf(), + name_in_archive: name.to_string(), + }); + } + + /// Indicate that the next call to `build` should update all symbols in + /// the archive (equivalent to running 'ar s' over it). + pub fn update_symbols(&mut self) { + self.should_update_symbols = true; + } + + /// Combine the provided files, rlibs, and native libraries into a single + /// `Archive`. + pub fn build(&mut self) { + let kind = match self.llvm_archive_kind() { + Ok(kind) => kind, + Err(kind) => { + self.config.sess.fatal(&format!("Don't know how to build archive of type: {}", + kind)); + } + }; + + if let Err(e) = self.build_with_llvm(kind) { + self.config.sess.fatal(&format!("failed to build archive: {}", e)); + } + + } + + fn llvm_archive_kind(&self) -> Result<ArchiveKind, &str> { + let kind = &*self.config.sess.target.target.options.archive_format; + kind.parse().map_err(|_| kind) + } + + fn build_with_llvm(&mut self, kind: ArchiveKind) -> io::Result<()> { + let mut archives = Vec::new(); + let mut strings = Vec::new(); + let mut members = Vec::new(); + let removals = mem::replace(&mut self.removals, Vec::new()); + + unsafe { + if let Some(archive) = self.src_archive() { + for child in archive.iter() { + let child = child.map_err(string_to_io_error)?; + let child_name = match child.name() { + Some(s) => s, + None => continue, + }; + if removals.iter().any(|r| r == child_name) { + continue + } + + let name = CString::new(child_name)?; + members.push(llvm::LLVMRustArchiveMemberNew(ptr::null(), + name.as_ptr(), + child.raw())); + strings.push(name); + } + } + for addition in mem::replace(&mut self.additions, Vec::new()) { + match addition { + Addition::File { path, name_in_archive } => { + let path = CString::new(path.to_str().unwrap())?; + let name = CString::new(name_in_archive)?; + members.push(llvm::LLVMRustArchiveMemberNew(path.as_ptr(), + name.as_ptr(), + ptr::null_mut())); + strings.push(path); + strings.push(name); + } + Addition::Archive { archive, mut skip } => { + for child in archive.iter() { + let child = child.map_err(string_to_io_error)?; + if !is_relevant_child(&child) { + continue + } + let child_name = child.name().unwrap(); + if skip(child_name) { + continue + } + + // It appears that LLVM's archive writer is a little + // buggy if the name we pass down isn't just the + // filename component, so chop that off here and + // pass it in. + // + // See LLVM bug 25877 for more info. + let child_name = Path::new(child_name) + .file_name().unwrap() + .to_str().unwrap(); + let name = CString::new(child_name)?; + let m = llvm::LLVMRustArchiveMemberNew(ptr::null(), + name.as_ptr(), + child.raw()); + members.push(m); + strings.push(name); + } + archives.push(archive); + } + } + } + + let dst = self.config.dst.to_str().unwrap().as_bytes(); + let dst = CString::new(dst)?; + let r = llvm::LLVMRustWriteArchive(dst.as_ptr(), + members.len() as libc::size_t, + members.as_ptr(), + self.should_update_symbols, + kind); + let ret = if r.into_result().is_err() { + let err = llvm::LLVMRustGetLastError(); + let msg = if err.is_null() { + "failed to write archive".to_string() + } else { + String::from_utf8_lossy(CStr::from_ptr(err).to_bytes()) + .into_owned() + }; + Err(io::Error::new(io::ErrorKind::Other, msg)) + } else { + Ok(()) + }; + for member in members { + llvm::LLVMRustArchiveMemberFree(member); + } + return ret + } + } +} + +fn string_to_io_error(s: String) -> io::Error { + io::Error::new(io::ErrorKind::Other, format!("bad archive: {}", s)) +} diff --git a/src/librustc_codegen_llvm/back/bytecode.rs b/src/librustc_codegen_llvm/back/bytecode.rs new file mode 100644 index 00000000000..212d1aaf055 --- /dev/null +++ b/src/librustc_codegen_llvm/back/bytecode.rs @@ -0,0 +1,160 @@ +// 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. + +//! Management of the encoding of LLVM bytecode into rlibs +//! +//! This module contains the management of encoding LLVM bytecode into rlibs, +//! primarily for the usage in LTO situations. Currently the compiler will +//! unconditionally encode LLVM-IR into rlibs regardless of what's happening +//! elsewhere, so we currently compress the bytecode via deflate to avoid taking +//! up too much space on disk. +//! +//! After compressing the bytecode we then have the rest of the format to +//! basically deal with various bugs in various archive implementations. The +//! format currently is: +//! +//! RLIB LLVM-BYTECODE OBJECT LAYOUT +//! Version 2 +//! Bytes Data +//! 0..10 "RUST_OBJECT" encoded in ASCII +//! 11..14 format version as little-endian u32 +//! 15..19 the length of the module identifier string +//! 20..n the module identifier string +//! n..n+8 size in bytes of deflate compressed LLVM bitcode as +//! little-endian u64 +//! n+9.. compressed LLVM bitcode +//! ? maybe a byte to make this whole thing even length + +use std::io::{Read, Write}; +use std::ptr; +use std::str; + +use flate2::Compression; +use flate2::read::DeflateDecoder; +use flate2::write::DeflateEncoder; + +// This is the "magic number" expected at the beginning of a LLVM bytecode +// object in an rlib. +pub const RLIB_BYTECODE_OBJECT_MAGIC: &'static [u8] = b"RUST_OBJECT"; + +// The version number this compiler will write to bytecode objects in rlibs +pub const RLIB_BYTECODE_OBJECT_VERSION: u8 = 2; + +pub const RLIB_BYTECODE_EXTENSION: &str = "bc.z"; + +pub fn encode(identifier: &str, bytecode: &[u8]) -> Vec<u8> { + let mut encoded = Vec::new(); + + // Start off with the magic string + encoded.extend_from_slice(RLIB_BYTECODE_OBJECT_MAGIC); + + // Next up is the version + encoded.extend_from_slice(&[RLIB_BYTECODE_OBJECT_VERSION, 0, 0, 0]); + + // Next is the LLVM module identifier length + contents + let identifier_len = identifier.len(); + encoded.extend_from_slice(&[ + (identifier_len >> 0) as u8, + (identifier_len >> 8) as u8, + (identifier_len >> 16) as u8, + (identifier_len >> 24) as u8, + ]); + encoded.extend_from_slice(identifier.as_bytes()); + + // Next is the LLVM module deflate compressed, prefixed with its length. We + // don't know its length yet, so fill in 0s + let deflated_size_pos = encoded.len(); + encoded.extend_from_slice(&[0, 0, 0, 0, 0, 0, 0, 0]); + + let before = encoded.len(); + DeflateEncoder::new(&mut encoded, Compression::fast()) + .write_all(bytecode) + .unwrap(); + let after = encoded.len(); + + // Fill in the length we reserved space for before + let bytecode_len = (after - before) as u64; + encoded[deflated_size_pos + 0] = (bytecode_len >> 0) as u8; + encoded[deflated_size_pos + 1] = (bytecode_len >> 8) as u8; + encoded[deflated_size_pos + 2] = (bytecode_len >> 16) as u8; + encoded[deflated_size_pos + 3] = (bytecode_len >> 24) as u8; + encoded[deflated_size_pos + 4] = (bytecode_len >> 32) as u8; + encoded[deflated_size_pos + 5] = (bytecode_len >> 40) as u8; + encoded[deflated_size_pos + 6] = (bytecode_len >> 48) as u8; + encoded[deflated_size_pos + 7] = (bytecode_len >> 56) as u8; + + // If the number of bytes written to the object so far is odd, add a + // padding byte to make it even. This works around a crash bug in LLDB + // (see issue #15950) + if encoded.len() % 2 == 1 { + encoded.push(0); + } + + return encoded +} + +pub struct DecodedBytecode<'a> { + identifier: &'a str, + encoded_bytecode: &'a [u8], +} + +impl<'a> DecodedBytecode<'a> { + pub fn new(data: &'a [u8]) -> Result<DecodedBytecode<'a>, String> { + if !data.starts_with(RLIB_BYTECODE_OBJECT_MAGIC) { + return Err(format!("magic bytecode prefix not found")) + } + let data = &data[RLIB_BYTECODE_OBJECT_MAGIC.len()..]; + if !data.starts_with(&[RLIB_BYTECODE_OBJECT_VERSION, 0, 0, 0]) { + return Err(format!("wrong version prefix found in bytecode")) + } + let data = &data[4..]; + if data.len() < 4 { + return Err(format!("bytecode corrupted")) + } + let identifier_len = unsafe { + u32::from_le(ptr::read_unaligned(data.as_ptr() as *const u32)) as usize + }; + let data = &data[4..]; + if data.len() < identifier_len { + return Err(format!("bytecode corrupted")) + } + let identifier = match str::from_utf8(&data[..identifier_len]) { + Ok(s) => s, + Err(_) => return Err(format!("bytecode corrupted")) + }; + let data = &data[identifier_len..]; + if data.len() < 8 { + return Err(format!("bytecode corrupted")) + } + let bytecode_len = unsafe { + u64::from_le(ptr::read_unaligned(data.as_ptr() as *const u64)) as usize + }; + let data = &data[8..]; + if data.len() < bytecode_len { + return Err(format!("bytecode corrupted")) + } + let encoded_bytecode = &data[..bytecode_len]; + + Ok(DecodedBytecode { + identifier, + encoded_bytecode, + }) + } + + pub fn bytecode(&self) -> Vec<u8> { + let mut data = Vec::new(); + DeflateDecoder::new(self.encoded_bytecode).read_to_end(&mut data).unwrap(); + return data + } + + pub fn identifier(&self) -> &'a str { + self.identifier + } +} diff --git a/src/librustc_codegen_llvm/back/command.rs b/src/librustc_codegen_llvm/back/command.rs new file mode 100644 index 00000000000..9ebbdd7c3c9 --- /dev/null +++ b/src/librustc_codegen_llvm/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_llvm/back/link.rs b/src/librustc_codegen_llvm/back/link.rs new file mode 100644 index 00000000000..dbfd430a3e2 --- /dev/null +++ b/src/librustc_codegen_llvm/back/link.rs @@ -0,0 +1,1630 @@ +// Copyright 2012-2014 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::wasm; +use cc::windows_registry; +use super::archive::{ArchiveBuilder, ArchiveConfig}; +use super::bytecode::RLIB_BYTECODE_EXTENSION; +use super::linker::Linker; +use super::command::Command; +use super::rpath::RPathConfig; +use super::rpath; +use metadata::METADATA_FILENAME; +use rustc::session::config::{self, NoDebugInfo, OutputFilenames, OutputType, PrintRequest}; +use rustc::session::config::{RUST_CGU_EXT, Lto}; +use rustc::session::filesearch; +use rustc::session::search_paths::PathKind; +use rustc::session::Session; +use rustc::middle::cstore::{NativeLibrary, LibSource, NativeLibraryKind}; +use rustc::middle::dependency_format::Linkage; +use {CodegenResults, CrateInfo}; +use rustc::util::common::time; +use rustc::util::fs::fix_windows_verbatim_for_gcc; +use rustc::hir::def_id::CrateNum; +use tempdir::TempDir; +use rustc_target::spec::{PanicStrategy, RelroLevel, LinkerFlavor, TargetTriple}; +use rustc_data_structures::fx::FxHashSet; +use context::get_reloc_model; +use llvm; + +use std::ascii; +use std::char; +use std::env; +use std::fmt; +use std::fs; +use std::io; +use std::path::{Path, PathBuf}; +use std::process::{Output, Stdio}; +use std::str; +use syntax::attr; + +/// The LLVM module name containing crate-metadata. This includes a `.` on +/// purpose, so it cannot clash with the name of a user-defined module. +pub const METADATA_MODULE_NAME: &'static str = "crate.metadata"; + +// same as for metadata above, but for allocator shim +pub const ALLOCATOR_MODULE_NAME: &'static str = "crate.allocator"; + +pub use rustc_codegen_utils::link::{find_crate_name, filename_for_input, default_output_for_target, + invalid_output_for_target, build_link_meta, out_filename, + check_file_is_writeable}; + +// 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) -> (PathBuf, Command) { + // 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 cmd = |linker: &Path| { + if let Some(linker) = linker.to_str() { + if cfg!(windows) && linker.ends_with(".bat") { + return Command::bat_script(linker) + } + } + match sess.linker_flavor() { + LinkerFlavor::Lld(f) => Command::lld(linker, f), + _ => Command::new(linker), + + } + }; + + let msvc_tool = windows_registry::find_tool(&sess.opts.target_triple.triple(), "link.exe"); + + let linker_path = sess.opts.cg.linker.as_ref().map(|s| &**s) + .or(sess.target.target.options.linker.as_ref().map(|s| s.as_ref())) + .unwrap_or(match sess.linker_flavor() { + LinkerFlavor::Msvc => { + msvc_tool.as_ref().map(|t| t.path()).unwrap_or("link.exe".as_ref()) + } + LinkerFlavor::Em if cfg!(windows) => "emcc.bat".as_ref(), + LinkerFlavor::Em => "emcc".as_ref(), + LinkerFlavor::Gcc => "cc".as_ref(), + LinkerFlavor::Ld => "ld".as_ref(), + LinkerFlavor::Lld(_) => "lld".as_ref(), + }); + + let mut cmd = cmd(linker_path); + + // 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_path.to_path_buf(), cmd) +} + +pub fn remove(sess: &Session, path: &Path) { + match fs::remove_file(path) { + Ok(..) => {} + Err(e) => { + sess.err(&format!("failed to remove {}: {}", + path.display(), + e)); + } + } +} + +/// Perform the linkage portion of the compilation phase. This will generate all +/// of the requested outputs for this compilation session. +pub(crate) fn link_binary(sess: &Session, + codegen_results: &CodegenResults, + outputs: &OutputFilenames, + crate_name: &str) -> Vec<PathBuf> { + let mut out_filenames = Vec::new(); + for &crate_type in sess.crate_types.borrow().iter() { + // Ignore executable crates if we have -Z no-codegen, as they will error. + let output_metadata = sess.opts.output_types.contains_key(&OutputType::Metadata); + if (sess.opts.debugging_opts.no_codegen || !sess.opts.output_types.should_codegen()) && + !output_metadata && + crate_type == config::CrateTypeExecutable { + continue; + } + + if invalid_output_for_target(sess, crate_type) { + bug!("invalid output type `{:?}` for target os `{}`", + crate_type, sess.opts.target_triple); + } + let mut out_files = link_binary_output(sess, + codegen_results, + crate_type, + outputs, + crate_name); + out_filenames.append(&mut out_files); + } + + // Remove the temporary object file and metadata if we aren't saving temps + if !sess.opts.cg.save_temps { + if sess.opts.output_types.should_codegen() && + !preserve_objects_for_their_debuginfo(sess) + { + for obj in codegen_results.modules.iter().filter_map(|m| m.object.as_ref()) { + remove(sess, obj); + } + } + for obj in codegen_results.modules.iter().filter_map(|m| m.bytecode_compressed.as_ref()) { + remove(sess, obj); + } + if let Some(ref obj) = codegen_results.metadata_module.object { + remove(sess, obj); + } + if let Some(ref allocator) = codegen_results.allocator_module { + if let Some(ref obj) = allocator.object { + remove(sess, obj); + } + if let Some(ref bc) = allocator.bytecode_compressed { + remove(sess, bc); + } + } + } + + out_filenames +} + +/// Returns a boolean indicating whether we should preserve the object files on +/// the filesystem for their debug information. This is often useful with +/// split-dwarf like schemes. +fn preserve_objects_for_their_debuginfo(sess: &Session) -> bool { + // If the objects don't have debuginfo there's nothing to preserve. + if sess.opts.debuginfo == NoDebugInfo { + return false + } + + // If we're only producing artifacts that are archives, no need to preserve + // the objects as they're losslessly contained inside the archives. + let output_linked = sess.crate_types.borrow() + .iter() + .any(|x| *x != config::CrateTypeRlib && *x != config::CrateTypeStaticlib); + if !output_linked { + return false + } + + // If we're on OSX then the equivalent of split dwarf is turned on by + // default. The final executable won't actually have any debug information + // except it'll have pointers to elsewhere. Historically we've always run + // `dsymutil` to "link all the dwarf together" but this is actually sort of + // a bummer for incremental compilation! (the whole point of split dwarf is + // that you don't do this sort of dwarf link). + // + // Basically as a result this just means that if we're on OSX and we're + // *not* running dsymutil then the object files are the only source of truth + // for debug information, so we must preserve them. + if sess.target.target.options.is_like_osx { + match sess.opts.debugging_opts.run_dsymutil { + // dsymutil is not being run, preserve objects + Some(false) => return true, + + // dsymutil is being run, no need to preserve the objects + Some(true) => return false, + + // The default historical behavior was to always run dsymutil, so + // we're preserving that temporarily, but we're likely to switch the + // default soon. + None => return false, + } + } + + false +} + +fn filename_for_metadata(sess: &Session, crate_name: &str, outputs: &OutputFilenames) -> PathBuf { + let out_filename = outputs.single_output_file.clone() + .unwrap_or(outputs + .out_directory + .join(&format!("lib{}{}.rmeta", crate_name, sess.opts.cg.extra_filename))); + check_file_is_writeable(&out_filename, sess); + out_filename +} + +pub(crate) fn each_linked_rlib(sess: &Session, + info: &CrateInfo, + f: &mut FnMut(CrateNum, &Path)) -> Result<(), String> { + let crates = info.used_crates_static.iter(); + let fmts = sess.dependency_formats.borrow(); + let fmts = fmts.get(&config::CrateTypeExecutable) + .or_else(|| fmts.get(&config::CrateTypeStaticlib)) + .or_else(|| fmts.get(&config::CrateTypeCdylib)) + .or_else(|| fmts.get(&config::CrateTypeProcMacro)); + let fmts = match fmts { + Some(f) => f, + None => return Err(format!("could not find formats for rlibs")) + }; + for &(cnum, ref path) in crates { + match fmts.get(cnum.as_usize() - 1) { + Some(&Linkage::NotLinked) | + Some(&Linkage::IncludedFromDylib) => continue, + Some(_) => {} + None => return Err(format!("could not find formats for rlibs")) + } + 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(crate) 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.is_no_builtins.contains(&cnum) || info.compiler_builtins == Some(cnum)) +} + +fn link_binary_output(sess: &Session, + codegen_results: &CodegenResults, + crate_type: config::CrateType, + outputs: &OutputFilenames, + crate_name: &str) -> Vec<PathBuf> { + for obj in codegen_results.modules.iter().filter_map(|m| m.object.as_ref()) { + check_file_is_writeable(obj, sess); + } + + let mut out_filenames = vec![]; + + if outputs.outputs.contains_key(&OutputType::Metadata) { + let out_filename = filename_for_metadata(sess, crate_name, outputs); + // To avoid races with another rustc process scanning the output directory, + // we need to write the file somewhere else and atomically move it to its + // final destination, with a `fs::rename` call. In order for the rename to + // always succeed, the temporary file needs to be on the same filesystem, + // which is why we create it inside the output directory specifically. + let metadata_tmpdir = match TempDir::new_in(out_filename.parent().unwrap(), "rmeta") { + Ok(tmpdir) => tmpdir, + Err(err) => sess.fatal(&format!("couldn't create a temp dir: {}", err)), + }; + let metadata = emit_metadata(sess, codegen_results, &metadata_tmpdir); + if let Err(e) = fs::rename(metadata, &out_filename) { + sess.fatal(&format!("failed to write {}: {}", out_filename.display(), e)); + } + out_filenames.push(out_filename); + } + + let tmpdir = match TempDir::new("rustc") { + Ok(tmpdir) => tmpdir, + Err(err) => sess.fatal(&format!("couldn't create a temp dir: {}", err)), + }; + + if outputs.outputs.should_codegen() { + let out_filename = out_filename(sess, crate_type, outputs, crate_name); + match crate_type { + config::CrateTypeRlib => { + link_rlib(sess, + codegen_results, + RlibFlavor::Normal, + &out_filename, + &tmpdir).build(); + } + config::CrateTypeStaticlib => { + link_staticlib(sess, codegen_results, &out_filename, &tmpdir); + } + _ => { + link_natively(sess, crate_type, &out_filename, codegen_results, tmpdir.path()); + } + } + out_filenames.push(out_filename); + } + + if sess.opts.cg.save_temps { + let _ = tmpdir.into_path(); + } + + out_filenames +} + +fn archive_search_paths(sess: &Session) -> Vec<PathBuf> { + let mut search = Vec::new(); + sess.target_filesearch(PathKind::Native).for_each_lib_search_path(|path, _| { + search.push(path.to_path_buf()); + }); + return search; +} + +fn archive_config<'a>(sess: &'a Session, + output: &Path, + input: Option<&Path>) -> ArchiveConfig<'a> { + ArchiveConfig { + sess, + dst: output.to_path_buf(), + src: input.map(|p| p.to_path_buf()), + lib_search_paths: archive_search_paths(sess), + } +} + +/// We use a temp directory here to avoid races between concurrent rustc processes, +/// such as builds in the same directory using the same filename for metadata while +/// building an `.rlib` (stomping over one another), or writing an `.rmeta` into a +/// directory being searched for `extern crate` (observing an incomplete file). +/// The returned path is the temporary file containing the complete metadata. +fn emit_metadata<'a>(sess: &'a Session, codegen_results: &CodegenResults, tmpdir: &TempDir) + -> PathBuf { + let out_filename = tmpdir.path().join(METADATA_FILENAME); + let result = fs::write(&out_filename, &codegen_results.metadata.raw_data); + + if let Err(e) = result { + sess.fatal(&format!("failed to write {}: {}", out_filename.display(), e)); + } + + out_filename +} + +enum RlibFlavor { + Normal, + StaticlibBase, +} + +// Create an 'rlib' +// +// An rlib in its current incarnation is essentially a renamed .a file. The +// rlib primarily contains the object file of the crate, but it also contains +// all of the object files from native libraries. This is done by unzipping +// native libraries and inserting all of the contents into this archive. +fn link_rlib<'a>(sess: &'a Session, + codegen_results: &CodegenResults, + flavor: RlibFlavor, + out_filename: &Path, + tmpdir: &TempDir) -> ArchiveBuilder<'a> { + info!("preparing rlib to {:?}", out_filename); + let mut ab = ArchiveBuilder::new(archive_config(sess, out_filename, None)); + + for obj in codegen_results.modules.iter().filter_map(|m| m.object.as_ref()) { + ab.add_file(obj); + } + + // Note that in this loop we are ignoring the value of `lib.cfg`. That is, + // we may not be configured to actually include a static library if we're + // adding it here. That's because later when we consume this rlib we'll + // decide whether we actually needed the static library or not. + // + // To do this "correctly" we'd need to keep track of which libraries added + // which object files to the archive. We don't do that here, however. The + // #[link(cfg(..))] feature is unstable, though, and only intended to get + // liblibc working. In that sense the check below just indicates that if + // there are any libraries we want to omit object files for at link time we + // just exclude all custom object files. + // + // Eventually if we want to stabilize or flesh out the #[link(cfg(..))] + // feature then we'll need to figure out how to record what objects were + // loaded from the libraries found here and then encode that into the + // metadata of the rlib we're generating somehow. + for lib in codegen_results.crate_info.used_libraries.iter() { + match lib.kind { + NativeLibraryKind::NativeStatic => {} + NativeLibraryKind::NativeStaticNobundle | + NativeLibraryKind::NativeFramework | + NativeLibraryKind::NativeUnknown => continue, + } + ab.add_native_library(&lib.name.as_str()); + } + + // After adding all files to the archive, we need to update the + // symbol table of the archive. + ab.update_symbols(); + + // Note that it is important that we add all of our non-object "magical + // files" *after* all of the object files in the archive. The reason for + // this is as follows: + // + // * When performing LTO, this archive will be modified to remove + // objects from above. The reason for this is described below. + // + // * When the system linker looks at an archive, it will attempt to + // determine the architecture of the archive in order to see whether its + // linkable. + // + // The algorithm for this detection is: iterate over the files in the + // archive. Skip magical SYMDEF names. Interpret the first file as an + // object file. Read architecture from the object file. + // + // * As one can probably see, if "metadata" and "foo.bc" were placed + // before all of the objects, then the architecture of this archive would + // not be correctly inferred once 'foo.o' is removed. + // + // Basically, all this means is that this code should not move above the + // code above. + match flavor { + RlibFlavor::Normal => { + // Instead of putting the metadata in an object file section, rlibs + // contain the metadata in a separate file. + ab.add_file(&emit_metadata(sess, codegen_results, tmpdir)); + + // For LTO purposes, the bytecode of this library is also inserted + // into the archive. + for bytecode in codegen_results + .modules + .iter() + .filter_map(|m| m.bytecode_compressed.as_ref()) + { + ab.add_file(bytecode); + } + + // After adding all files to the archive, we need to update the + // symbol table of the archive. This currently dies on macOS (see + // #11162), and isn't necessary there anyway + if !sess.target.target.options.is_like_osx { + ab.update_symbols(); + } + } + + RlibFlavor::StaticlibBase => { + let obj = codegen_results.allocator_module + .as_ref() + .and_then(|m| m.object.as_ref()); + if let Some(obj) = obj { + ab.add_file(obj); + } + } + } + + ab +} + +// Create a static archive +// +// This is essentially the same thing as an rlib, but it also involves adding +// all of the upstream crates' objects into the archive. This will slurp in +// all of the native libraries of upstream dependencies as well. +// +// Additionally, there's no way for us to link dynamic libraries, so we warn +// about all dynamic library dependencies that they're not linked in. +// +// There's no need to include metadata in a static archive, so ensure to not +// link in the metadata object file (and also don't prepare the archive with a +// metadata file). +fn link_staticlib(sess: &Session, + codegen_results: &CodegenResults, + out_filename: &Path, + tempdir: &TempDir) { + let mut ab = link_rlib(sess, + codegen_results, + RlibFlavor::StaticlibBase, + out_filename, + tempdir); + let mut all_native_libs = vec![]; + + let res = each_linked_rlib(sess, &codegen_results.crate_info, &mut |cnum, path| { + let name = &codegen_results.crate_info.crate_name[&cnum]; + let native_libs = &codegen_results.crate_info.native_libraries[&cnum]; + + // Here when we include the rlib into our staticlib we need to make a + // decision whether to include the extra object files along the way. + // These extra object files come from statically included native + // libraries, but they may be cfg'd away with #[link(cfg(..))]. + // + // This unstable feature, though, only needs liblibc to work. The only + // use case there is where musl is statically included in liblibc.rlib, + // so if we don't want the included version we just need to skip it. As + // a result the logic here is that if *any* linked library is cfg'd away + // we just skip all object files. + // + // Clearly this is not sufficient for a general purpose feature, and + // we'd want to read from the library's metadata to determine which + // object files come from where and selectively skip them. + let skip_object_files = native_libs.iter().any(|lib| { + lib.kind == NativeLibraryKind::NativeStatic && !relevant_lib(sess, lib) + }); + ab.add_rlib(path, + &name.as_str(), + is_full_lto_enabled(sess) && + !ignored_for_lto(sess, &codegen_results.crate_info, cnum), + skip_object_files).unwrap(); + + all_native_libs.extend(codegen_results.crate_info.native_libraries[&cnum].iter().cloned()); + }); + if let Err(e) = res { + sess.fatal(&e); + } + + ab.update_symbols(); + ab.build(); + + if !all_native_libs.is_empty() { + if sess.opts.prints.contains(&PrintRequest::NativeStaticLibs) { + print_native_static_libs(sess, &all_native_libs); + } + } +} + +fn print_native_static_libs(sess: &Session, all_native_libs: &[NativeLibrary]) { + let lib_args: Vec<_> = all_native_libs.iter() + .filter(|l| relevant_lib(sess, l)) + .filter_map(|lib| match lib.kind { + NativeLibraryKind::NativeStaticNobundle | + NativeLibraryKind::NativeUnknown => { + if sess.target.target.options.is_like_msvc { + Some(format!("{}.lib", lib.name)) + } else { + Some(format!("-l{}", lib.name)) + } + }, + NativeLibraryKind::NativeFramework => { + // ld-only syntax, since there are no frameworks in MSVC + Some(format!("-framework {}", lib.name)) + }, + // These are included, no need to print them + NativeLibraryKind::NativeStatic => None, + }) + .collect(); + if !lib_args.is_empty() { + sess.note_without_error("Link against the following native artifacts when linking \ + against this static library. The order and any duplication \ + can be significant on some platforms."); + // Prefix for greppability + sess.note_without_error(&format!("native-static-libs: {}", &lib_args.join(" "))); + } +} + +// Create a dynamic library or executable +// +// This will invoke the system linker/cc to create the resulting file. This +// links to all upstream files as well. +fn link_natively(sess: &Session, + crate_type: config::CrateType, + out_filename: &Path, + codegen_results: &CodegenResults, + tmpdir: &Path) { + info!("preparing {:?} to {:?}", crate_type, out_filename); + let flavor = sess.linker_flavor(); + + // The invocations of cc share some flags across platforms + let (pname, mut cmd) = get_linker(sess); + + let root = sess.target_filesearch(PathKind::Native).get_lib_path(); + if let Some(args) = sess.target.target.options.pre_link_args.get(&flavor) { + cmd.args(args); + } + if let Some(args) = sess.target.target.options.pre_link_args_crt.get(&flavor) { + if sess.crt_static() { + cmd.args(args); + } + } + if let Some(ref args) = sess.opts.debugging_opts.pre_link_args { + cmd.args(args); + } + cmd.args(&sess.opts.debugging_opts.pre_link_arg); + + let pre_link_objects = if crate_type == config::CrateTypeExecutable { + &sess.target.target.options.pre_link_objects_exe + } else { + &sess.target.target.options.pre_link_objects_dll + }; + for obj in pre_link_objects { + cmd.arg(root.join(obj)); + } + + if crate_type == config::CrateTypeExecutable && sess.crt_static() { + for obj in &sess.target.target.options.pre_link_objects_exe_crt { + cmd.arg(root.join(obj)); + } + + for obj in &sess.target.target.options.pre_link_objects_exe_crt_sys { + if flavor == LinkerFlavor::Gcc { + cmd.arg(format!("-l:{}", obj)); + } + } + } + + if sess.target.target.options.is_like_emscripten { + cmd.arg("-s"); + cmd.arg(if sess.panic_strategy() == PanicStrategy::Abort { + "DISABLE_EXCEPTION_CATCHING=1" + } else { + "DISABLE_EXCEPTION_CATCHING=0" + }); + } + + { + let mut linker = codegen_results.linker_info.to_linker(cmd, &sess); + link_args(&mut *linker, sess, crate_type, tmpdir, + out_filename, codegen_results); + cmd = linker.finalize(); + } + if let Some(args) = sess.target.target.options.late_link_args.get(&flavor) { + cmd.args(args); + } + for obj in &sess.target.target.options.post_link_objects { + cmd.arg(root.join(obj)); + } + if sess.crt_static() { + for obj in &sess.target.target.options.post_link_objects_crt_sys { + if flavor == LinkerFlavor::Gcc { + cmd.arg(format!("-l:{}", obj)); + } + } + for obj in &sess.target.target.options.post_link_objects_crt { + cmd.arg(root.join(obj)); + } + } + if let Some(args) = sess.target.target.options.post_link_args.get(&flavor) { + cmd.args(args); + } + for &(ref k, ref v) in &sess.target.target.options.link_env { + cmd.env(k, v); + } + + if sess.opts.debugging_opts.print_link_args { + println!("{:?}", &cmd); + } + + // May have not found libraries in the right formats. + sess.abort_if_errors(); + + // Invoke the system linker + // + // Note that there's a terribly awful hack that really shouldn't be present + // in any compiler. Here an environment variable is supported to + // automatically retry the linker invocation if the linker looks like it + // segfaulted. + // + // Gee that seems odd, normally segfaults are things we want to know about! + // Unfortunately though in rust-lang/rust#38878 we're experiencing the + // linker segfaulting on Travis quite a bit which is causing quite a bit of + // pain to land PRs when they spuriously fail due to a segfault. + // + // The issue #38878 has some more debugging information on it as well, but + // this unfortunately looks like it's just a race condition in macOS's linker + // with some thread pool working in the background. It seems that no one + // currently knows a fix for this so in the meantime we're left with this... + info!("{:?}", &cmd); + let retry_on_segfault = env::var("RUSTC_RETRY_LINKER_ON_SEGFAULT").is_ok(); + let mut prog; + let mut i = 0; + loop { + i += 1; + prog = time(sess, "running linker", || { + exec_linker(sess, &mut cmd, out_filename, tmpdir) + }); + let output = match prog { + Ok(ref output) => output, + Err(_) => break, + }; + if output.status.success() { + break + } + let mut out = output.stderr.clone(); + out.extend(&output.stdout); + let out = String::from_utf8_lossy(&out); + + // Check to see if the link failed with "unrecognized command line option: + // '-no-pie'" for gcc or "unknown argument: '-no-pie'" for clang. If so, + // reperform the link step without the -no-pie option. This is safe because + // if the linker doesn't support -no-pie then it should not default to + // linking executables as pie. Different versions of gcc seem to use + // different quotes in the error message so don't check for them. + if sess.target.target.options.linker_is_gnu && + sess.linker_flavor() != LinkerFlavor::Ld && + (out.contains("unrecognized command line option") || + out.contains("unknown argument")) && + out.contains("-no-pie") && + cmd.get_args().iter().any(|e| e.to_string_lossy() == "-no-pie") { + info!("linker output: {:?}", out); + warn!("Linker does not support -no-pie command line option. Retrying without."); + for arg in cmd.take_args() { + if arg.to_string_lossy() != "-no-pie" { + cmd.arg(arg); + } + } + info!("{:?}", &cmd); + continue; + } + if !retry_on_segfault || i > 3 { + break + } + let msg_segv = "clang: error: unable to execute command: Segmentation fault: 11"; + let msg_bus = "clang: error: unable to execute command: Bus error: 10"; + if !(out.contains(msg_segv) || out.contains(msg_bus)) { + break + } + + warn!( + "looks like the linker segfaulted when we tried to call it, \ + automatically retrying again. cmd = {:?}, out = {}.", + cmd, + out, + ); + } + + match prog { + Ok(prog) => { + fn escape_string(s: &[u8]) -> String { + str::from_utf8(s).map(|s| s.to_owned()) + .unwrap_or_else(|_| { + let mut x = "Non-UTF-8 output: ".to_string(); + x.extend(s.iter() + .flat_map(|&b| ascii::escape_default(b)) + .map(|b| char::from_u32(b as u32).unwrap())); + x + }) + } + if !prog.status.success() { + let mut output = prog.stderr.clone(); + output.extend_from_slice(&prog.stdout); + sess.struct_err(&format!("linking with `{}` failed: {}", + pname.display(), + prog.status)) + .note(&format!("{:?}", &cmd)) + .note(&escape_string(&output)) + .emit(); + sess.abort_if_errors(); + } + info!("linker stderr:\n{}", escape_string(&prog.stderr)); + info!("linker stdout:\n{}", escape_string(&prog.stdout)); + }, + Err(e) => { + let linker_not_found = e.kind() == io::ErrorKind::NotFound; + + let mut linker_error = { + if linker_not_found { + sess.struct_err(&format!("linker `{}` not found", pname.display())) + } else { + sess.struct_err(&format!("could not exec the linker `{}`", pname.display())) + } + }; + + linker_error.note(&format!("{}", e)); + + if !linker_not_found { + linker_error.note(&format!("{:?}", &cmd)); + } + + linker_error.emit(); + + if sess.target.target.options.is_like_msvc && linker_not_found { + sess.note_without_error("the msvc targets depend on the msvc linker \ + but `link.exe` was not found"); + sess.note_without_error("please ensure that VS 2013 or VS 2015 was installed \ + with the Visual C++ option"); + } + sess.abort_if_errors(); + } + } + + + // On macOS, debuggers need this utility to get run to do some munging of + // the symbols. Note, though, that if the object files are being preserved + // for their debug information there's no need for us to run dsymutil. + if sess.target.target.options.is_like_osx && + sess.opts.debuginfo != NoDebugInfo && + !preserve_objects_for_their_debuginfo(sess) + { + match Command::new("dsymutil").arg(out_filename).output() { + Ok(..) => {} + Err(e) => sess.fatal(&format!("failed to run dsymutil: {}", e)), + } + } + + if sess.opts.target_triple == TargetTriple::from_triple("wasm32-unknown-unknown") { + wasm::rewrite_imports(&out_filename, &codegen_results.crate_info.wasm_imports); + wasm::add_custom_sections(&out_filename, + &codegen_results.crate_info.wasm_custom_sections); + } +} + +fn exec_linker(sess: &Session, cmd: &mut Command, out_filename: &Path, tmpdir: &Path) + -> io::Result<Output> +{ + // When attempting to spawn the linker we run a risk of blowing out the + // size limits for spawning a new process with respect to the arguments + // we pass on the command line. + // + // Here we attempt to handle errors from the OS saying "your list of + // arguments is too big" by reinvoking the linker again with an `@`-file + // that contains all the arguments. The theory is that this is then + // accepted on all linkers and the linker will read all its options out of + // there instead of looking at the command line. + if !cmd.very_likely_to_exceed_some_spawn_limit() { + match cmd.command().stdout(Stdio::piped()).stderr(Stdio::piped()).spawn() { + Ok(child) => { + let output = child.wait_with_output(); + flush_linked_file(&output, out_filename)?; + return output; + } + Err(ref e) if command_line_too_big(e) => { + info!("command line to linker was too big: {}", e); + } + Err(e) => return Err(e) + } + } + + info!("falling back to passing arguments to linker via an @-file"); + let mut cmd2 = cmd.clone(); + let mut args = String::new(); + for arg in cmd2.take_args() { + args.push_str(&Escape { + arg: arg.to_str().unwrap(), + is_like_msvc: sess.target.target.options.is_like_msvc, + }.to_string()); + args.push_str("\n"); + } + let file = tmpdir.join("linker-arguments"); + let bytes = if sess.target.target.options.is_like_msvc { + let mut out = vec![]; + // start the stream with a UTF-16 BOM + for c in vec![0xFEFF].into_iter().chain(args.encode_utf16()) { + // encode in little endian + out.push(c as u8); + out.push((c >> 8) as u8); + } + out + } else { + args.into_bytes() + }; + fs::write(&file, &bytes)?; + cmd2.arg(format!("@{}", file.display())); + info!("invoking linker {:?}", cmd2); + let output = cmd2.output(); + flush_linked_file(&output, out_filename)?; + return output; + + #[cfg(unix)] + fn flush_linked_file(_: &io::Result<Output>, _: &Path) -> io::Result<()> { + Ok(()) + } + + #[cfg(windows)] + fn flush_linked_file(command_output: &io::Result<Output>, out_filename: &Path) + -> io::Result<()> + { + // On Windows, under high I/O load, output buffers are sometimes not flushed, + // even long after process exit, causing nasty, non-reproducible output bugs. + // + // File::sync_all() calls FlushFileBuffers() down the line, which solves the problem. + // + // А full writeup of the original Chrome bug can be found at + // randomascii.wordpress.com/2018/02/25/compiler-bug-linker-bug-windows-kernel-bug/amp + + if let &Ok(ref out) = command_output { + if out.status.success() { + if let Ok(of) = fs::OpenOptions::new().write(true).open(out_filename) { + of.sync_all()?; + } + } + } + + Ok(()) + } + + #[cfg(unix)] + fn command_line_too_big(err: &io::Error) -> bool { + err.raw_os_error() == Some(::libc::E2BIG) + } + + #[cfg(windows)] + fn command_line_too_big(err: &io::Error) -> bool { + const ERROR_FILENAME_EXCED_RANGE: i32 = 206; + err.raw_os_error() == Some(ERROR_FILENAME_EXCED_RANGE) + } + + struct Escape<'a> { + arg: &'a str, + is_like_msvc: bool, + } + + impl<'a> fmt::Display for Escape<'a> { + fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { + if self.is_like_msvc { + // This is "documented" at + // https://msdn.microsoft.com/en-us/library/4xdcbak7.aspx + // + // Unfortunately there's not a great specification of the + // syntax I could find online (at least) but some local + // testing showed that this seemed sufficient-ish to catch + // at least a few edge cases. + write!(f, "\"")?; + for c in self.arg.chars() { + match c { + '"' => write!(f, "\\{}", c)?, + c => write!(f, "{}", c)?, + } + } + write!(f, "\"")?; + } else { + // This is documented at https://linux.die.net/man/1/ld, namely: + // + // > Options in file are separated by whitespace. A whitespace + // > character may be included in an option by surrounding the + // > entire option in either single or double quotes. Any + // > character (including a backslash) may be included by + // > prefixing the character to be included with a backslash. + // + // We put an argument on each line, so all we need to do is + // ensure the line is interpreted as one whole argument. + for c in self.arg.chars() { + match c { + '\\' | + ' ' => write!(f, "\\{}", c)?, + c => write!(f, "{}", c)?, + } + } + } + Ok(()) + } + } +} + +fn link_args(cmd: &mut Linker, + sess: &Session, + crate_type: config::CrateType, + tmpdir: &Path, + out_filename: &Path, + codegen_results: &CodegenResults) { + + // Linker plugins should be specified early in the list of arguments + cmd.cross_lang_lto(); + + // The default library location, we need this to find the runtime. + // The location of crates will be determined as needed. + let lib_path = sess.target_filesearch(PathKind::All).get_lib_path(); + + // target descriptor + let t = &sess.target.target; + + cmd.include_path(&fix_windows_verbatim_for_gcc(&lib_path)); + for obj in codegen_results.modules.iter().filter_map(|m| m.object.as_ref()) { + cmd.add_object(obj); + } + cmd.output_filename(out_filename); + + if crate_type == config::CrateTypeExecutable && + sess.target.target.options.is_like_windows { + if let Some(ref s) = codegen_results.windows_subsystem { + cmd.subsystem(s); + } + } + + // If we're building a dynamic library then some platforms need to make sure + // that all symbols are exported correctly from the dynamic library. + if crate_type != config::CrateTypeExecutable || + sess.target.target.options.is_like_emscripten { + cmd.export_symbols(tmpdir, crate_type); + } + + // When linking a dynamic library, we put the metadata into a section of the + // executable. This metadata is in a separate object file from the main + // object file, so we link that in here. + if crate_type == config::CrateTypeDylib || + crate_type == config::CrateTypeProcMacro { + if let Some(obj) = codegen_results.metadata_module.object.as_ref() { + cmd.add_object(obj); + } + } + + let obj = codegen_results.allocator_module + .as_ref() + .and_then(|m| m.object.as_ref()); + if let Some(obj) = obj { + cmd.add_object(obj); + } + + // Try to strip as much out of the generated object by removing unused + // sections if possible. See more comments in linker.rs + if !sess.opts.cg.link_dead_code { + let keep_metadata = crate_type == config::CrateTypeDylib; + cmd.gc_sections(keep_metadata); + } + + let used_link_args = &codegen_results.crate_info.link_args; + + if crate_type == config::CrateTypeExecutable { + let mut position_independent_executable = false; + + if t.options.position_independent_executables { + let empty_vec = Vec::new(); + let args = sess.opts.cg.link_args.as_ref().unwrap_or(&empty_vec); + let more_args = &sess.opts.cg.link_arg; + let mut args = args.iter().chain(more_args.iter()).chain(used_link_args.iter()); + + if get_reloc_model(sess) == llvm::RelocMode::PIC + && !sess.crt_static() && !args.any(|x| *x == "-static") { + position_independent_executable = true; + } + } + + if position_independent_executable { + cmd.position_independent_executable(); + } else { + // recent versions of gcc can be configured to generate position + // independent executables by default. We have to pass -no-pie to + // explicitly turn that off. Not applicable to ld. + if sess.target.target.options.linker_is_gnu + && sess.linker_flavor() != LinkerFlavor::Ld { + cmd.no_position_independent_executable(); + } + } + } + + let relro_level = match sess.opts.debugging_opts.relro_level { + Some(level) => level, + None => t.options.relro_level, + }; + match relro_level { + RelroLevel::Full => { + cmd.full_relro(); + }, + RelroLevel::Partial => { + cmd.partial_relro(); + }, + RelroLevel::Off => { + cmd.no_relro(); + }, + RelroLevel::None => { + }, + } + + // Pass optimization flags down to the linker. + cmd.optimize(); + + // Pass debuginfo flags down to the linker. + cmd.debuginfo(); + + // We want to prevent the compiler from accidentally leaking in any system + // libraries, so we explicitly ask gcc to not link to any libraries by + // default. Note that this does not happen for windows because windows pulls + // in some large number of libraries and I couldn't quite figure out which + // subset we wanted. + if t.options.no_default_libraries { + cmd.no_default_libraries(); + } + + // Take careful note of the ordering of the arguments we pass to the linker + // here. Linkers will assume that things on the left depend on things to the + // right. Things on the right cannot depend on things on the left. This is + // all formally implemented in terms of resolving symbols (libs on the right + // resolve unknown symbols of libs on the left, but not vice versa). + // + // For this reason, we have organized the arguments we pass to the linker as + // such: + // + // 1. The local object that LLVM just generated + // 2. Local native libraries + // 3. Upstream rust libraries + // 4. Upstream native libraries + // + // The rationale behind this ordering is that those items lower down in the + // list can't depend on items higher up in the list. For example nothing can + // depend on what we just generated (e.g. that'd be a circular dependency). + // Upstream rust libraries are not allowed to depend on our local native + // libraries as that would violate the structure of the DAG, in that + // scenario they are required to link to them as well in a shared fashion. + // + // Note that upstream rust libraries may contain native dependencies as + // well, but they also can't depend on what we just started to add to the + // link line. And finally upstream native libraries can't depend on anything + // in this DAG so far because they're only dylibs and dylibs can only depend + // on other dylibs (e.g. other native deps). + add_local_native_libraries(cmd, sess, codegen_results); + add_upstream_rust_crates(cmd, sess, codegen_results, crate_type, tmpdir); + add_upstream_native_libraries(cmd, sess, codegen_results, crate_type); + + // Tell the linker what we're doing. + if crate_type != config::CrateTypeExecutable { + cmd.build_dylib(out_filename); + } + if crate_type == config::CrateTypeExecutable && sess.crt_static() { + cmd.build_static_executable(); + } + + if sess.opts.debugging_opts.pgo_gen.is_some() { + cmd.pgo_gen(); + } + + // FIXME (#2397): At some point we want to rpath our guesses as to + // where extern libraries might live, based on the + // addl_lib_search_paths + if sess.opts.cg.rpath { + let sysroot = sess.sysroot(); + let target_triple = sess.opts.target_triple.triple(); + let mut get_install_prefix_lib_path = || { + let install_prefix = option_env!("CFG_PREFIX").expect("CFG_PREFIX"); + let tlib = filesearch::relative_target_lib_path(sysroot, target_triple); + let mut path = PathBuf::from(install_prefix); + path.push(&tlib); + + path + }; + let mut rpath_config = RPathConfig { + used_crates: &codegen_results.crate_info.used_crates_dynamic, + out_filename: out_filename.to_path_buf(), + has_rpath: sess.target.target.options.has_rpath, + is_like_osx: sess.target.target.options.is_like_osx, + linker_is_gnu: sess.target.target.options.linker_is_gnu, + get_install_prefix_lib_path: &mut get_install_prefix_lib_path, + }; + cmd.args(&rpath::get_rpath_flags(&mut rpath_config)); + } + + // Finally add all the linker arguments provided on the command line along + // with any #[link_args] attributes found inside the crate + if let Some(ref args) = sess.opts.cg.link_args { + cmd.args(args); + } + cmd.args(&sess.opts.cg.link_arg); + cmd.args(&used_link_args); +} + +// # Native library linking +// +// User-supplied library search paths (-L on the command line). These are +// the same paths used to find Rust crates, so some of them may have been +// added already by the previous crate linking code. This only allows them +// to be found at compile time so it is still entirely up to outside +// forces to make sure that library can be found at runtime. +// +// Also note that the native libraries linked here are only the ones located +// in the current crate. Upstream crates with native library dependencies +// may have their native library pulled in above. +fn add_local_native_libraries(cmd: &mut Linker, + sess: &Session, + codegen_results: &CodegenResults) { + sess.target_filesearch(PathKind::All).for_each_lib_search_path(|path, k| { + match k { + PathKind::Framework => { cmd.framework_path(path); } + _ => { cmd.include_path(&fix_windows_verbatim_for_gcc(path)); } + } + }); + + let relevant_libs = codegen_results.crate_info.used_libraries.iter().filter(|l| { + relevant_lib(sess, l) + }); + + let search_path = archive_search_paths(sess); + for lib in relevant_libs { + match lib.kind { + NativeLibraryKind::NativeUnknown => cmd.link_dylib(&lib.name.as_str()), + NativeLibraryKind::NativeFramework => cmd.link_framework(&lib.name.as_str()), + NativeLibraryKind::NativeStaticNobundle => cmd.link_staticlib(&lib.name.as_str()), + NativeLibraryKind::NativeStatic => cmd.link_whole_staticlib(&lib.name.as_str(), + &search_path) + } + } +} + +// # Rust Crate linking +// +// Rust crates are not considered at all when creating an rlib output. All +// dependencies will be linked when producing the final output (instead of +// the intermediate rlib version) +fn add_upstream_rust_crates(cmd: &mut Linker, + sess: &Session, + codegen_results: &CodegenResults, + crate_type: config::CrateType, + tmpdir: &Path) { + // All of the heavy lifting has previously been accomplished by the + // dependency_format module of the compiler. This is just crawling the + // output of that module, adding crates as necessary. + // + // Linking to a rlib involves just passing it to the linker (the linker + // will slurp up the object files inside), and linking to a dynamic library + // involves just passing the right -l flag. + + let formats = sess.dependency_formats.borrow(); + let data = formats.get(&crate_type).unwrap(); + + // Invoke get_used_crates to ensure that we get a topological sorting of + // crates. + let deps = &codegen_results.crate_info.used_crates_dynamic; + + // There's a few internal crates in the standard library (aka libcore and + // libstd) which actually have a circular dependence upon one another. This + // currently arises through "weak lang items" where libcore requires things + // like `rust_begin_unwind` but libstd ends up defining it. To get this + // circular dependence to work correctly in all situations we'll need to be + // sure to correctly apply the `--start-group` and `--end-group` options to + // GNU linkers, otherwise if we don't use any other symbol from the standard + // library it'll get discarded and the whole application won't link. + // + // In this loop we're calculating the `group_end`, after which crate to + // pass `--end-group` and `group_start`, before which crate to pass + // `--start-group`. We currently do this by passing `--end-group` after + // the first crate (when iterating backwards) that requires a lang item + // defined somewhere else. Once that's set then when we've defined all the + // necessary lang items we'll pass `--start-group`. + // + // Note that this isn't amazing logic for now but it should do the trick + // for the current implementation of the standard library. + let mut group_end = None; + let mut group_start = None; + let mut end_with = FxHashSet(); + let info = &codegen_results.crate_info; + for &(cnum, _) in deps.iter().rev() { + if let Some(missing) = info.missing_lang_items.get(&cnum) { + end_with.extend(missing.iter().cloned()); + if end_with.len() > 0 && group_end.is_none() { + group_end = Some(cnum); + } + } + end_with.retain(|item| info.lang_item_to_crate.get(item) != Some(&cnum)); + if end_with.len() == 0 && group_end.is_some() { + group_start = Some(cnum); + break + } + } + + // If we didn't end up filling in all lang items from upstream crates then + // we'll be filling it in with our crate. This probably means we're the + // standard library itself, so skip this for now. + if group_end.is_some() && group_start.is_none() { + group_end = None; + } + + let mut compiler_builtins = None; + + for &(cnum, _) in deps.iter() { + if group_start == Some(cnum) { + cmd.group_start(); + } + + // We may not pass all crates through to the linker. Some crates may + // appear statically in an existing dylib, meaning we'll pick up all the + // symbols from the dylib. + let src = &codegen_results.crate_info.used_crate_source[&cnum]; + match data[cnum.as_usize() - 1] { + _ if codegen_results.crate_info.profiler_runtime == Some(cnum) => { + add_static_crate(cmd, sess, codegen_results, tmpdir, crate_type, cnum); + } + _ if codegen_results.crate_info.sanitizer_runtime == Some(cnum) => { + link_sanitizer_runtime(cmd, sess, codegen_results, tmpdir, cnum); + } + // compiler-builtins are always placed last to ensure that they're + // linked correctly. + _ if codegen_results.crate_info.compiler_builtins == Some(cnum) => { + assert!(compiler_builtins.is_none()); + compiler_builtins = Some(cnum); + } + Linkage::NotLinked | + Linkage::IncludedFromDylib => {} + Linkage::Static => { + add_static_crate(cmd, sess, codegen_results, tmpdir, crate_type, cnum); + } + Linkage::Dynamic => { + add_dynamic_crate(cmd, sess, &src.dylib.as_ref().unwrap().0) + } + } + + if group_end == Some(cnum) { + cmd.group_end(); + } + } + + // compiler-builtins are always placed last to ensure that they're + // linked correctly. + // We must always link the `compiler_builtins` crate statically. Even if it + // was already "included" in a dylib (e.g. `libstd` when `-C prefer-dynamic` + // is used) + if let Some(cnum) = compiler_builtins { + add_static_crate(cmd, sess, codegen_results, tmpdir, crate_type, cnum); + } + + // Converts a library file-stem into a cc -l argument + fn unlib<'a>(config: &config::Config, stem: &'a str) -> &'a str { + if stem.starts_with("lib") && !config.target.options.is_like_windows { + &stem[3..] + } else { + stem + } + } + + // We must link the sanitizer runtime using -Wl,--whole-archive but since + // it's packed in a .rlib, it contains stuff that are not objects that will + // make the linker error. So we must remove those bits from the .rlib before + // linking it. + fn link_sanitizer_runtime(cmd: &mut Linker, + sess: &Session, + codegen_results: &CodegenResults, + tmpdir: &Path, + cnum: CrateNum) { + let src = &codegen_results.crate_info.used_crate_source[&cnum]; + let cratepath = &src.rlib.as_ref().unwrap().0; + + if sess.target.target.options.is_like_osx { + // On Apple platforms, the sanitizer is always built as a dylib, and + // LLVM will link to `@rpath/*.dylib`, so we need to specify an + // rpath to the library as well (the rpath should be absolute, see + // PR #41352 for details). + // + // FIXME: Remove this logic into librustc_*san once Cargo supports it + let rpath = cratepath.parent().unwrap(); + let rpath = rpath.to_str().expect("non-utf8 component in path"); + cmd.args(&["-Wl,-rpath".into(), "-Xlinker".into(), rpath.into()]); + } + + let dst = tmpdir.join(cratepath.file_name().unwrap()); + let cfg = archive_config(sess, &dst, Some(cratepath)); + let mut archive = ArchiveBuilder::new(cfg); + archive.update_symbols(); + + for f in archive.src_files() { + if f.ends_with(RLIB_BYTECODE_EXTENSION) || f == METADATA_FILENAME { + archive.remove_file(&f); + continue + } + } + + archive.build(); + + cmd.link_whole_rlib(&dst); + } + + // Adds the static "rlib" versions of all crates to the command line. + // There's a bit of magic which happens here specifically related to LTO and + // dynamic libraries. Specifically: + // + // * For LTO, we remove upstream object files. + // * For dylibs we remove metadata and bytecode from upstream rlibs + // + // When performing LTO, almost(*) all of the bytecode from the upstream + // libraries has already been included in our object file output. As a + // result we need to remove the object files in the upstream libraries so + // the linker doesn't try to include them twice (or whine about duplicate + // symbols). We must continue to include the rest of the rlib, however, as + // it may contain static native libraries which must be linked in. + // + // (*) Crates marked with `#![no_builtins]` don't participate in LTO and + // their bytecode wasn't included. The object files in those libraries must + // still be passed to the linker. + // + // When making a dynamic library, linkers by default don't include any + // object files in an archive if they're not necessary to resolve the link. + // We basically want to convert the archive (rlib) to a dylib, though, so we + // *do* want everything included in the output, regardless of whether the + // linker thinks it's needed or not. As a result we must use the + // --whole-archive option (or the platform equivalent). When using this + // option the linker will fail if there are non-objects in the archive (such + // as our own metadata and/or bytecode). All in all, for rlibs to be + // entirely included in dylibs, we need to remove all non-object files. + // + // Note, however, that if we're not doing LTO or we're not producing a dylib + // (aka we're making an executable), we can just pass the rlib blindly to + // the linker (fast) because it's fine if it's not actually included as + // we're at the end of the dependency chain. + fn add_static_crate(cmd: &mut Linker, + sess: &Session, + codegen_results: &CodegenResults, + tmpdir: &Path, + crate_type: config::CrateType, + cnum: CrateNum) { + let src = &codegen_results.crate_info.used_crate_source[&cnum]; + let cratepath = &src.rlib.as_ref().unwrap().0; + + // See the comment above in `link_staticlib` and `link_rlib` for why if + // there's a static library that's not relevant we skip all object + // files. + let native_libs = &codegen_results.crate_info.native_libraries[&cnum]; + let skip_native = native_libs.iter().any(|lib| { + lib.kind == NativeLibraryKind::NativeStatic && !relevant_lib(sess, lib) + }); + + if (!is_full_lto_enabled(sess) || + ignored_for_lto(sess, &codegen_results.crate_info, cnum)) && + crate_type != config::CrateTypeDylib && + !skip_native { + cmd.link_rlib(&fix_windows_verbatim_for_gcc(cratepath)); + return + } + + let dst = tmpdir.join(cratepath.file_name().unwrap()); + let name = cratepath.file_name().unwrap().to_str().unwrap(); + let name = &name[3..name.len() - 5]; // chop off lib/.rlib + + time(sess, &format!("altering {}.rlib", name), || { + let cfg = archive_config(sess, &dst, Some(cratepath)); + let mut archive = ArchiveBuilder::new(cfg); + archive.update_symbols(); + + let mut any_objects = false; + for f in archive.src_files() { + if f.ends_with(RLIB_BYTECODE_EXTENSION) || f == METADATA_FILENAME { + archive.remove_file(&f); + continue + } + + let canonical = f.replace("-", "_"); + let canonical_name = name.replace("-", "_"); + + // Look for `.rcgu.o` at the end of the filename to conclude + // that this is a Rust-related object file. + fn looks_like_rust(s: &str) -> bool { + let path = Path::new(s); + let ext = path.extension().and_then(|s| s.to_str()); + if ext != Some(OutputType::Object.extension()) { + return false + } + let ext2 = path.file_stem() + .and_then(|s| Path::new(s).extension()) + .and_then(|s| s.to_str()); + ext2 == Some(RUST_CGU_EXT) + } + + let is_rust_object = + canonical.starts_with(&canonical_name) && + looks_like_rust(&f); + + // If we've been requested to skip all native object files + // (those not generated by the rust compiler) then we can skip + // this file. See above for why we may want to do this. + let skip_because_cfg_say_so = skip_native && !is_rust_object; + + // If we're performing LTO and this is a rust-generated object + // file, then we don't need the object file as it's part of the + // LTO module. Note that `#![no_builtins]` is excluded from LTO, + // though, so we let that object file slide. + let skip_because_lto = is_full_lto_enabled(sess) && + is_rust_object && + (sess.target.target.options.no_builtins || + !codegen_results.crate_info.is_no_builtins.contains(&cnum)); + + if skip_because_cfg_say_so || skip_because_lto { + archive.remove_file(&f); + } else { + any_objects = true; + } + } + + if !any_objects { + return + } + archive.build(); + + // If we're creating a dylib, then we need to include the + // whole of each object in our archive into that artifact. This is + // because a `dylib` can be reused as an intermediate artifact. + // + // Note, though, that we don't want to include the whole of a + // compiler-builtins crate (e.g. compiler-rt) because it'll get + // repeatedly linked anyway. + if crate_type == config::CrateTypeDylib && + codegen_results.crate_info.compiler_builtins != Some(cnum) { + cmd.link_whole_rlib(&fix_windows_verbatim_for_gcc(&dst)); + } else { + cmd.link_rlib(&fix_windows_verbatim_for_gcc(&dst)); + } + }); + } + + // Same thing as above, but for dynamic crates instead of static crates. + fn add_dynamic_crate(cmd: &mut Linker, sess: &Session, cratepath: &Path) { + // If we're performing LTO, then it should have been previously required + // that all upstream rust dependencies were available in an rlib format. + assert!(!is_full_lto_enabled(sess)); + + // Just need to tell the linker about where the library lives and + // what its name is + let parent = cratepath.parent(); + if let Some(dir) = parent { + cmd.include_path(&fix_windows_verbatim_for_gcc(dir)); + } + let filestem = cratepath.file_stem().unwrap().to_str().unwrap(); + cmd.link_rust_dylib(&unlib(&sess.target, filestem), + parent.unwrap_or(Path::new(""))); + } +} + +// Link in all of our upstream crates' native dependencies. Remember that +// all of these upstream native dependencies are all non-static +// dependencies. We've got two cases then: +// +// 1. The upstream crate is an rlib. In this case we *must* link in the +// native dependency because the rlib is just an archive. +// +// 2. The upstream crate is a dylib. In order to use the dylib, we have to +// have the dependency present on the system somewhere. Thus, we don't +// gain a whole lot from not linking in the dynamic dependency to this +// crate as well. +// +// The use case for this is a little subtle. In theory the native +// dependencies of a crate are purely an implementation detail of the crate +// itself, but the problem arises with generic and inlined functions. If a +// generic function calls a native function, then the generic function must +// be instantiated in the target crate, meaning that the native symbol must +// also be resolved in the target crate. +fn add_upstream_native_libraries(cmd: &mut Linker, + sess: &Session, + codegen_results: &CodegenResults, + crate_type: config::CrateType) { + // Be sure to use a topological sorting of crates because there may be + // interdependencies between native libraries. When passing -nodefaultlibs, + // for example, almost all native libraries depend on libc, so we have to + // make sure that's all the way at the right (liblibc is near the base of + // the dependency chain). + // + // This passes RequireStatic, but the actual requirement doesn't matter, + // we're just getting an ordering of crate numbers, we're not worried about + // the paths. + let formats = sess.dependency_formats.borrow(); + let data = formats.get(&crate_type).unwrap(); + + let crates = &codegen_results.crate_info.used_crates_static; + for &(cnum, _) in crates { + for lib in codegen_results.crate_info.native_libraries[&cnum].iter() { + if !relevant_lib(sess, &lib) { + continue + } + match lib.kind { + NativeLibraryKind::NativeUnknown => cmd.link_dylib(&lib.name.as_str()), + NativeLibraryKind::NativeFramework => cmd.link_framework(&lib.name.as_str()), + NativeLibraryKind::NativeStaticNobundle => { + // Link "static-nobundle" native libs only if the crate they originate from + // is being linked statically to the current crate. If it's linked dynamically + // or is an rlib already included via some other dylib crate, the symbols from + // native libs will have already been included in that dylib. + if data[cnum.as_usize() - 1] == Linkage::Static { + cmd.link_staticlib(&lib.name.as_str()) + } + }, + // ignore statically included native libraries here as we've + // already included them when we included the rust library + // previously + NativeLibraryKind::NativeStatic => {} + } + } + } +} + +fn relevant_lib(sess: &Session, lib: &NativeLibrary) -> bool { + match lib.cfg { + Some(ref cfg) => attr::cfg_matches(cfg, &sess.parse_sess, None), + None => true, + } +} + +fn is_full_lto_enabled(sess: &Session) -> bool { + match sess.lto() { + Lto::Yes | + Lto::Thin | + Lto::Fat => true, + Lto::No | + Lto::ThinLocal => false, + } +} diff --git a/src/librustc_codegen_llvm/back/linker.rs b/src/librustc_codegen_llvm/back/linker.rs new file mode 100644 index 00000000000..dd1983bdc17 --- /dev/null +++ b/src/librustc_codegen_llvm/back/linker.rs @@ -0,0 +1,1037 @@ +// 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 std::collections::HashMap; +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 back::archive; +use back::command::Command; +use back::symbol_export; +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, DebugInfoLevel, + 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: HashMap<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) -> Box<Linker+'a> { + match sess.linker_flavor() { + LinkerFlavor::Lld(LldFlavor::Link) | + LinkerFlavor::Msvc => { + Box::new(MsvcLinker { + cmd, + sess, + info: self + }) as Box<Linker> + } + LinkerFlavor::Em => { + Box::new(EmLinker { + cmd, + sess, + info: self + }) as Box<Linker> + } + LinkerFlavor::Gcc => { + Box::new(GccLinker { + cmd, + sess, + info: self, + hinted_static: false, + is_ld: false, + }) as Box<Linker> + } + + LinkerFlavor::Lld(LldFlavor::Ld) | + LinkerFlavor::Lld(LldFlavor::Ld64) | + LinkerFlavor::Ld => { + Box::new(GccLinker { + cmd, + sess, + info: self, + hinted_static: false, + is_ld: true, + }) as Box<Linker> + } + + LinkerFlavor::Lld(LldFlavor::Wasm) => { + Box::new(WasmLd { + cmd, + }) as Box<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, +} + +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; + } + } +} + +impl<'a> Linker for GccLinker<'a> { + fn link_dylib(&mut self, lib: &str) { self.hint_dynamic(); self.cmd.arg("-l").arg(lib); } + fn link_staticlib(&mut self, lib: &str) { self.hint_static(); self.cmd.arg("-l").arg(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("-z,relro,-z,now"); } + fn partial_relro(&mut self) { self.linker_arg("-z,relro"); } + fn no_relro(&mut self) { self.linker_arg("-z,norelro"); } + 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("-l").arg(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("-l").arg(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. + let mut v = OsString::from("-force_load,"); + v.push(&archive::find_library(lib, search_path, &self.sess)); + self.linker_arg(&v); + } + } + + fn link_whole_rlib(&mut self, lib: &Path) { + self.hint_static(); + if self.sess.target.target.options.is_like_osx { + let mut v = OsString::from("-force_load,"); + v.push(lib); + self.linker_arg(&v); + } 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("-z"); + self.linker_arg("ignore"); + + // 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) { + match self.sess.opts.debuginfo { + DebugInfoLevel::NoDebugInfo => { + // 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. + match self.sess.opts.debugging_opts.strip_debuginfo_if_disabled { + Some(true) => { 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 { + let mut v = OsString::from("-install_name,@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::CrateTypeDylib || + crate_type == CrateType::CrateTypeProcMacro { + 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(&format!("--subsystem,{}", subsystem)); + } + + fn finalize(&mut self) -> Command { + self.hint_dynamic(); // Reset to default before returning the composed command line. + let mut cmd = Command::new(""); + ::std::mem::swap(&mut cmd, &mut self.cmd); + cmd + } + + 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 | + CrossLangLto::NoLink => { + // Nothing to do + } + CrossLangLto::LinkerPlugin(ref path) => { + self.linker_arg(&format!("-plugin={}", path.display())); + + 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)); + self.linker_arg(&format!("-plugin-opt=mcpu={}", self.sess.target_cpu())); + + match self.sess.opts.cg.lto { + config::Lto::Thin | + config::Lto::ThinLocal => { + self.linker_arg(&format!("-plugin-opt=thin")); + } + config::Lto::Fat | + config::Lto::Yes | + config::Lto::No => { + // default to regular LTO + } + } + } + } + } +} + +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 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 { + let mut cmd = Command::new(""); + ::std::mem::swap(&mut cmd, &mut self.cmd); + cmd + } + + // 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 { + DebugInfoLevel::NoDebugInfo => "-g0", + DebugInfoLevel::LimitedDebugInfo => "-g3", + DebugInfoLevel::FullDebugInfo => "-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_string() + 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 { + let mut cmd = Command::new(""); + ::std::mem::swap(&mut cmd, &mut self.cmd); + cmd + } + + // Appears not necessary on Emscripten + fn group_start(&mut self) {} + fn group_end(&mut self) {} + + fn cross_lang_lto(&mut self) { + // Do nothing + } +} + +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 +} + +pub struct WasmLd { + cmd: Command, +} + +impl Linker for WasmLd { + 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) { + } + + 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 export_symbols(&mut self, _tmpdir: &Path, _crate_type: CrateType) { + } + + 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"); + + self.cmd.arg("-z").arg("stack-size=1048576"); + + // 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"); + + let mut cmd = Command::new(""); + ::std::mem::swap(&mut cmd, &mut self.cmd); + cmd + } + + // 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 + } +} diff --git a/src/librustc_codegen_llvm/back/lto.rs b/src/librustc_codegen_llvm/back/lto.rs new file mode 100644 index 00000000000..96eda50d788 --- /dev/null +++ b/src/librustc_codegen_llvm/back/lto.rs @@ -0,0 +1,773 @@ +// Copyright 2013 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::bytecode::{DecodedBytecode, RLIB_BYTECODE_EXTENSION}; +use back::symbol_export; +use back::write::{ModuleConfig, with_llvm_pmb, CodegenContext}; +use back::write; +use errors::{FatalError, Handler}; +use llvm::archive_ro::ArchiveRO; +use llvm::{ModuleRef, TargetMachineRef, True, False}; +use llvm; +use rustc::hir::def_id::LOCAL_CRATE; +use rustc::middle::exported_symbols::SymbolExportLevel; +use rustc::session::config::{self, Lto}; +use rustc::util::common::time_ext; +use time_graph::Timeline; +use {ModuleCodegen, ModuleLlvm, ModuleKind, ModuleSource}; + +use libc; + +use std::ffi::CString; +use std::ptr; +use std::slice; +use std::sync::Arc; + +pub fn crate_type_allows_lto(crate_type: config::CrateType) -> bool { + match crate_type { + config::CrateTypeExecutable | + config::CrateTypeStaticlib | + config::CrateTypeCdylib => true, + + config::CrateTypeDylib | + config::CrateTypeRlib | + config::CrateTypeProcMacro => false, + } +} + +pub(crate) enum LtoModuleCodegen { + Fat { + module: Option<ModuleCodegen>, + _serialized_bitcode: Vec<SerializedModule>, + }, + + Thin(ThinModule), +} + +impl LtoModuleCodegen { + 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(crate) unsafe fn optimize(&mut self, + cgcx: &CodegenContext, + timeline: &mut Timeline) + -> Result<ModuleCodegen, FatalError> + { + match *self { + LtoModuleCodegen::Fat { ref mut module, .. } => { + let module = module.take().unwrap(); + let config = cgcx.config(module.kind); + let llmod = module.llvm().unwrap().llmod; + let tm = module.llvm().unwrap().tm; + run_pass_manager(cgcx, tm, llmod, config, false); + timeline.record("fat-done"); + Ok(module) + } + LtoModuleCodegen::Thin(ref mut thin) => thin.optimize(cgcx, 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(crate) fn run(cgcx: &CodegenContext, + modules: Vec<ModuleCodegen>, + timeline: &mut Timeline) + -> Result<Vec<LtoModuleCodegen>, FatalError> +{ + let diag_handler = cgcx.create_diag_handler(); + let export_threshold = match cgcx.lto { + // We're just doing LTO for our one crate + Lto::ThinLocal => SymbolExportLevel::Rust, + + // We're doing LTO for the entire crate graph + Lto::Yes | Lto::Fat | Lto::Thin => { + symbol_export::crates_export_threshold(&cgcx.crate_types) + } + + Lto::No => panic!("didn't request LTO but we're doing LTO"), + }; + + let symbol_filter = &|&(ref name, level): &(String, SymbolExportLevel)| { + if level.is_below_threshold(export_threshold) { + let mut bytes = Vec::with_capacity(name.len() + 1); + bytes.extend(name.bytes()); + Some(CString::new(bytes).unwrap()) + } else { + None + } + }; + let exported_symbols = cgcx.exported_symbols + .as_ref().expect("needs exported symbols for LTO"); + let mut symbol_white_list = exported_symbols[&LOCAL_CRATE] + .iter() + .filter_map(symbol_filter) + .collect::<Vec<CString>>(); + timeline.record("whitelist"); + info!("{} symbols to preserve in this crate", symbol_white_list.len()); + + // If we're performing LTO for the entire crate graph, then for each of our + // upstream dependencies, find the corresponding rlib and load the bitcode + // from the archive. + // + // We save off all the bytecode and LLVM module ids for later processing + // with either fat or thin LTO + let mut upstream_modules = Vec::new(); + if cgcx.lto != Lto::ThinLocal { + if cgcx.opts.cg.prefer_dynamic { + diag_handler.struct_err("cannot prefer dynamic linking when performing LTO") + .note("only 'staticlib', 'bin', and 'cdylib' outputs are \ + supported with LTO") + .emit(); + return Err(FatalError) + } + + // Make sure we actually can run LTO + for crate_type in cgcx.crate_types.iter() { + if !crate_type_allows_lto(*crate_type) { + let e = diag_handler.fatal("lto can only be run for executables, cdylibs and \ + static library outputs"); + return Err(e) + } + } + + for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() { + let exported_symbols = cgcx.exported_symbols + .as_ref().expect("needs exported symbols for LTO"); + symbol_white_list.extend( + exported_symbols[&cnum] + .iter() + .filter_map(symbol_filter)); + + let archive = ArchiveRO::open(&path).expect("wanted an rlib"); + let bytecodes = archive.iter().filter_map(|child| { + child.ok().and_then(|c| c.name().map(|name| (name, c))) + }).filter(|&(name, _)| name.ends_with(RLIB_BYTECODE_EXTENSION)); + for (name, data) in bytecodes { + info!("adding bytecode {}", name); + let bc_encoded = data.data(); + + let (bc, id) = time_ext(cgcx.time_passes, None, &format!("decode {}", name), || { + match DecodedBytecode::new(bc_encoded) { + Ok(b) => Ok((b.bytecode(), b.identifier().to_string())), + Err(e) => Err(diag_handler.fatal(&e)), + } + })?; + let bc = SerializedModule::FromRlib(bc); + upstream_modules.push((bc, CString::new(id).unwrap())); + } + timeline.record(&format!("load: {}", path.display())); + } + } + + let arr = symbol_white_list.iter().map(|c| c.as_ptr()).collect::<Vec<_>>(); + match cgcx.lto { + Lto::Yes | // `-C lto` == fat LTO by default + Lto::Fat => { + fat_lto(cgcx, &diag_handler, modules, upstream_modules, &arr, timeline) + } + Lto::Thin | + Lto::ThinLocal => { + thin_lto(&diag_handler, modules, upstream_modules, &arr, timeline) + } + Lto::No => unreachable!(), + } +} + +fn fat_lto(cgcx: &CodegenContext, + diag_handler: &Handler, + mut modules: Vec<ModuleCodegen>, + mut serialized_modules: Vec<(SerializedModule, CString)>, + symbol_white_list: &[*const libc::c_char], + timeline: &mut Timeline) + -> Result<Vec<LtoModuleCodegen>, FatalError> +{ + info!("going for a fat lto"); + + // Find the "costliest" module and merge everything into that codegen unit. + // All the other modules will be serialized and reparsed into the new + // context, so this hopefully avoids serializing and parsing the largest + // codegen unit. + // + // Additionally use a regular module as the base here to ensure that various + // file copy operations in the backend work correctly. The only other kind + // of module here should be an allocator one, and if your crate is smaller + // than the allocator module then the size doesn't really matter anyway. + let (_, costliest_module) = modules.iter() + .enumerate() + .filter(|&(_, module)| module.kind == ModuleKind::Regular) + .map(|(i, module)| { + let cost = unsafe { + llvm::LLVMRustModuleCost(module.llvm().unwrap().llmod) + }; + (cost, i) + }) + .max() + .expect("must be codegen'ing at least one module"); + let module = modules.remove(costliest_module); + let llmod = module.llvm().expect("can't lto pre-codegened modules").llmod; + info!("using {:?} as a base module", module.llmod_id); + + // For all other modules we codegened we'll need to link them into our own + // bitcode. All modules were codegened in their own LLVM context, however, + // and we want to move everything to the same LLVM context. Currently the + // way we know of to do that is to serialize them to a string and them parse + // them later. Not great but hey, that's why it's "fat" LTO, right? + for module in modules { + let llvm = module.llvm().expect("can't lto pre-codegened modules"); + let buffer = ModuleBuffer::new(llvm.llmod); + let llmod_id = CString::new(&module.llmod_id[..]).unwrap(); + serialized_modules.push((SerializedModule::Local(buffer), llmod_id)); + } + + // For all serialized bitcode files we parse them and link them in as we did + // above, this is all mostly handled in C++. Like above, though, we don't + // know much about the memory management here so we err on the side of being + // save and persist everything with the original module. + let mut serialized_bitcode = Vec::new(); + let mut linker = Linker::new(llmod); + for (bc_decoded, name) in serialized_modules { + info!("linking {:?}", name); + time_ext(cgcx.time_passes, None, &format!("ll link {:?}", name), || { + let data = bc_decoded.data(); + linker.add(&data).map_err(|()| { + let msg = format!("failed to load bc of {:?}", name); + write::llvm_err(&diag_handler, msg) + }) + })?; + timeline.record(&format!("link {:?}", name)); + serialized_bitcode.push(bc_decoded); + } + drop(linker); + cgcx.save_temp_bitcode(&module, "lto.input"); + + // Internalize everything that *isn't* in our whitelist to help strip out + // more modules and such + unsafe { + let ptr = symbol_white_list.as_ptr(); + llvm::LLVMRustRunRestrictionPass(llmod, + ptr as *const *const libc::c_char, + symbol_white_list.len() as libc::size_t); + cgcx.save_temp_bitcode(&module, "lto.after-restriction"); + } + + if cgcx.no_landing_pads { + unsafe { + llvm::LLVMRustMarkAllFunctionsNounwind(llmod); + } + cgcx.save_temp_bitcode(&module, "lto.after-nounwind"); + } + timeline.record("passes"); + + Ok(vec![LtoModuleCodegen::Fat { + module: Some(module), + _serialized_bitcode: serialized_bitcode, + }]) +} + +struct Linker(llvm::LinkerRef); + +impl Linker { + fn new(llmod: ModuleRef) -> Linker { + unsafe { Linker(llvm::LLVMRustLinkerNew(llmod)) } + } + + fn add(&mut self, bytecode: &[u8]) -> Result<(), ()> { + unsafe { + if llvm::LLVMRustLinkerAdd(self.0, + bytecode.as_ptr() as *const libc::c_char, + bytecode.len()) { + Ok(()) + } else { + Err(()) + } + } + } +} + +impl Drop for Linker { + fn drop(&mut self) { + unsafe { llvm::LLVMRustLinkerFree(self.0); } + } +} + +/// Prepare "thin" LTO to get run on these modules. +/// +/// The general structure of ThinLTO is quite different from the structure of +/// "fat" LTO above. With "fat" LTO all LLVM modules in question are merged into +/// one giant LLVM module, and then we run more optimization passes over this +/// big module after internalizing most symbols. Thin LTO, on the other hand, +/// avoid this large bottleneck through more targeted optimization. +/// +/// At a high level Thin LTO looks like: +/// +/// 1. Prepare a "summary" of each LLVM module in question which describes +/// the values inside, cost of the values, etc. +/// 2. Merge the summaries of all modules in question into one "index" +/// 3. Perform some global analysis on this index +/// 4. For each module, use the index and analysis calculated previously to +/// perform local transformations on the module, for example inlining +/// small functions from other modules. +/// 5. Run thin-specific optimization passes over each module, and then code +/// generate everything at the end. +/// +/// The summary for each module is intended to be quite cheap, and the global +/// index is relatively quite cheap to create as well. As a result, the goal of +/// ThinLTO is to reduce the bottleneck on LTO and enable LTO to be used in more +/// situations. For example one cheap optimization is that we can parallelize +/// all codegen modules, easily making use of all the cores on a machine. +/// +/// With all that in mind, the function here is designed at specifically just +/// calculating the *index* for ThinLTO. This index will then be shared amongst +/// all of the `LtoModuleCodegen` units returned below and destroyed once +/// they all go out of scope. +fn thin_lto(diag_handler: &Handler, + modules: Vec<ModuleCodegen>, + serialized_modules: Vec<(SerializedModule, CString)>, + symbol_white_list: &[*const libc::c_char], + timeline: &mut Timeline) + -> Result<Vec<LtoModuleCodegen>, FatalError> +{ + unsafe { + info!("going for that thin, thin LTO"); + + let mut thin_buffers = Vec::new(); + let mut module_names = Vec::new(); + let mut thin_modules = Vec::new(); + + // FIXME: right now, like with fat LTO, we serialize all in-memory + // modules before working with them and ThinLTO. We really + // shouldn't do this, however, and instead figure out how to + // extract a summary from an in-memory module and then merge that + // into the global index. It turns out that this loop is by far + // the most expensive portion of this small bit of global + // analysis! + for (i, module) in modules.iter().enumerate() { + info!("local module: {} - {}", i, module.llmod_id); + let llvm = module.llvm().expect("can't lto precodegened module"); + let name = CString::new(module.llmod_id.clone()).unwrap(); + let buffer = ThinBuffer::new(llvm.llmod); + thin_modules.push(llvm::ThinLTOModule { + identifier: name.as_ptr(), + data: buffer.data().as_ptr(), + len: buffer.data().len(), + }); + thin_buffers.push(buffer); + module_names.push(name); + timeline.record(&module.llmod_id); + } + + // FIXME: All upstream crates are deserialized internally in the + // function below to extract their summary and modules. Note that + // unlike the loop above we *must* decode and/or read something + // here as these are all just serialized files on disk. An + // improvement, however, to make here would be to store the + // module summary separately from the actual module itself. Right + // now this is store in one large bitcode file, and the entire + // file is deflate-compressed. We could try to bypass some of the + // decompression by storing the index uncompressed and only + // lazily decompressing the bytecode if necessary. + // + // Note that truly taking advantage of this optimization will + // likely be further down the road. We'd have to implement + // incremental ThinLTO first where we could actually avoid + // looking at upstream modules entirely sometimes (the contents, + // we must always unconditionally look at the index). + let mut serialized = Vec::new(); + for (module, name) in serialized_modules { + info!("foreign module {:?}", name); + thin_modules.push(llvm::ThinLTOModule { + identifier: name.as_ptr(), + data: module.data().as_ptr(), + len: module.data().len(), + }); + serialized.push(module); + module_names.push(name); + } + + // Delegate to the C++ bindings to create some data here. Once this is a + // tried-and-true interface we may wish to try to upstream some of this + // to LLVM itself, right now we reimplement a lot of what they do + // upstream... + let data = llvm::LLVMRustCreateThinLTOData( + thin_modules.as_ptr(), + thin_modules.len() as u32, + symbol_white_list.as_ptr(), + symbol_white_list.len() as u32, + ); + if data.is_null() { + let msg = format!("failed to prepare thin LTO context"); + return Err(write::llvm_err(&diag_handler, msg)) + } + let data = ThinData(data); + info!("thin LTO data created"); + timeline.record("data"); + + // Throw our data in an `Arc` as we'll be sharing it across threads. We + // also put all memory referenced by the C++ data (buffers, ids, etc) + // into the arc as well. After this we'll create a thin module + // codegen per module in this data. + let shared = Arc::new(ThinShared { + data, + thin_buffers, + serialized_modules: serialized, + module_names, + }); + Ok((0..shared.module_names.len()).map(|i| { + LtoModuleCodegen::Thin(ThinModule { + shared: shared.clone(), + idx: i, + }) + }).collect()) + } +} + +fn run_pass_manager(cgcx: &CodegenContext, + tm: TargetMachineRef, + llmod: ModuleRef, + config: &ModuleConfig, + thin: bool) { + // Now we have one massive module inside of llmod. Time to run the + // LTO-specific optimization passes that LLVM provides. + // + // This code is based off the code found in llvm's LTO code generator: + // tools/lto/LTOCodeGenerator.cpp + debug!("running the pass manager"); + unsafe { + let pm = llvm::LLVMCreatePassManager(); + llvm::LLVMRustAddAnalysisPasses(tm, pm, llmod); + let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr() as *const _); + assert!(!pass.is_null()); + llvm::LLVMRustAddPass(pm, pass); + + // When optimizing for LTO we don't actually pass in `-O0`, but we force + // it to always happen at least with `-O1`. + // + // With ThinLTO we mess around a lot with symbol visibility in a way + // that will actually cause linking failures if we optimize at O0 which + // notable is lacking in dead code elimination. To ensure we at least + // get some optimizations and correctly link we forcibly switch to `-O1` + // to get dead code elimination. + // + // Note that in general this shouldn't matter too much as you typically + // only turn on ThinLTO when you're compiling with optimizations + // otherwise. + let opt_level = config.opt_level.unwrap_or(llvm::CodeGenOptLevel::None); + let opt_level = match opt_level { + llvm::CodeGenOptLevel::None => llvm::CodeGenOptLevel::Less, + level => level, + }; + with_llvm_pmb(llmod, config, opt_level, false, &mut |b| { + if thin { + if !llvm::LLVMRustPassManagerBuilderPopulateThinLTOPassManager(b, pm) { + panic!("this version of LLVM does not support ThinLTO"); + } + } else { + llvm::LLVMPassManagerBuilderPopulateLTOPassManager(b, pm, + /* Internalize = */ False, + /* RunInliner = */ True); + } + }); + + let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr() as *const _); + assert!(!pass.is_null()); + llvm::LLVMRustAddPass(pm, pass); + + time_ext(cgcx.time_passes, None, "LTO passes", || + llvm::LLVMRunPassManager(pm, llmod)); + + llvm::LLVMDisposePassManager(pm); + } + debug!("lto done"); +} + +pub enum SerializedModule { + Local(ModuleBuffer), + FromRlib(Vec<u8>), +} + +impl SerializedModule { + fn data(&self) -> &[u8] { + match *self { + SerializedModule::Local(ref m) => m.data(), + SerializedModule::FromRlib(ref m) => m, + } + } +} + +pub struct ModuleBuffer(*mut llvm::ModuleBuffer); + +unsafe impl Send for ModuleBuffer {} +unsafe impl Sync for ModuleBuffer {} + +impl ModuleBuffer { + pub fn new(m: ModuleRef) -> ModuleBuffer { + ModuleBuffer(unsafe { + llvm::LLVMRustModuleBufferCreate(m) + }) + } + + pub fn data(&self) -> &[u8] { + unsafe { + let ptr = llvm::LLVMRustModuleBufferPtr(self.0); + let len = llvm::LLVMRustModuleBufferLen(self.0); + slice::from_raw_parts(ptr, len) + } + } +} + +impl Drop for ModuleBuffer { + fn drop(&mut self) { + unsafe { llvm::LLVMRustModuleBufferFree(self.0); } + } +} + +pub struct ThinModule { + shared: Arc<ThinShared>, + idx: usize, +} + +struct ThinShared { + data: ThinData, + thin_buffers: Vec<ThinBuffer>, + serialized_modules: Vec<SerializedModule>, + module_names: Vec<CString>, +} + +struct ThinData(*mut llvm::ThinLTOData); + +unsafe impl Send for ThinData {} +unsafe impl Sync for ThinData {} + +impl Drop for ThinData { + fn drop(&mut self) { + unsafe { + llvm::LLVMRustFreeThinLTOData(self.0); + } + } +} + +pub struct ThinBuffer(*mut llvm::ThinLTOBuffer); + +unsafe impl Send for ThinBuffer {} +unsafe impl Sync for ThinBuffer {} + +impl ThinBuffer { + pub fn new(m: ModuleRef) -> ThinBuffer { + unsafe { + let buffer = llvm::LLVMRustThinLTOBufferCreate(m); + ThinBuffer(buffer) + } + } + + pub fn data(&self) -> &[u8] { + unsafe { + let ptr = llvm::LLVMRustThinLTOBufferPtr(self.0) as *const _; + let len = llvm::LLVMRustThinLTOBufferLen(self.0); + slice::from_raw_parts(ptr, len) + } + } +} + +impl Drop for ThinBuffer { + fn drop(&mut self) { + unsafe { + llvm::LLVMRustThinLTOBufferFree(self.0); + } + } +} + +impl ThinModule { + fn name(&self) -> &str { + self.shared.module_names[self.idx].to_str().unwrap() + } + + 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 + } + + 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() + }) + } + + unsafe fn optimize(&mut self, cgcx: &CodegenContext, timeline: &mut Timeline) + -> Result<ModuleCodegen, FatalError> + { + let diag_handler = cgcx.create_diag_handler(); + let tm = (cgcx.tm_factory)().map_err(|e| { + write::llvm_err(&diag_handler, e) + })?; + + // Right now the implementation we've got only works over serialized + // modules, so we create a fresh new LLVM context and parse the module + // into that context. One day, however, we may do this for upstream + // crates but for locally codegened modules we may be able to reuse + // that LLVM Context and Module. + let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names); + let llmod = llvm::LLVMRustParseBitcodeForThinLTO( + llcx, + self.data().as_ptr(), + self.data().len(), + self.shared.module_names[self.idx].as_ptr(), + ); + if llmod.is_null() { + let msg = format!("failed to parse bitcode for thin LTO module"); + return Err(write::llvm_err(&diag_handler, msg)); + } + let module = ModuleCodegen { + source: ModuleSource::Codegened(ModuleLlvm { + llmod, + llcx, + tm, + }), + llmod_id: self.name().to_string(), + name: self.name().to_string(), + kind: ModuleKind::Regular, + }; + cgcx.save_temp_bitcode(&module, "thin-lto-input"); + + // Before we do much else find the "main" `DICompileUnit` that we'll be + // using below. If we find more than one though then rustc has changed + // in a way we're not ready for, so generate an ICE by returning + // an error. + let mut cu1 = ptr::null_mut(); + let mut cu2 = ptr::null_mut(); + llvm::LLVMRustThinLTOGetDICompileUnit(llmod, &mut cu1, &mut cu2); + if !cu2.is_null() { + let msg = format!("multiple source DICompileUnits found"); + return Err(write::llvm_err(&diag_handler, msg)) + } + + // Like with "fat" LTO, get some better optimizations if landing pads + // are disabled by removing all landing pads. + if cgcx.no_landing_pads { + llvm::LLVMRustMarkAllFunctionsNounwind(llmod); + cgcx.save_temp_bitcode(&module, "thin-lto-after-nounwind"); + timeline.record("nounwind"); + } + + // Up next comes the per-module local analyses that we do for Thin LTO. + // Each of these functions is basically copied from the LLVM + // implementation and then tailored to suit this implementation. Ideally + // each of these would be supported by upstream LLVM but that's perhaps + // a patch for another day! + // + // You can find some more comments about these functions in the LLVM + // bindings we've got (currently `PassWrapper.cpp`) + if !llvm::LLVMRustPrepareThinLTORename(self.shared.data.0, llmod) { + let msg = format!("failed to prepare thin LTO module"); + return Err(write::llvm_err(&diag_handler, msg)) + } + cgcx.save_temp_bitcode(&module, "thin-lto-after-rename"); + timeline.record("rename"); + if !llvm::LLVMRustPrepareThinLTOResolveWeak(self.shared.data.0, llmod) { + let msg = format!("failed to prepare thin LTO module"); + return Err(write::llvm_err(&diag_handler, msg)) + } + cgcx.save_temp_bitcode(&module, "thin-lto-after-resolve"); + timeline.record("resolve"); + if !llvm::LLVMRustPrepareThinLTOInternalize(self.shared.data.0, llmod) { + let msg = format!("failed to prepare thin LTO module"); + return Err(write::llvm_err(&diag_handler, msg)) + } + cgcx.save_temp_bitcode(&module, "thin-lto-after-internalize"); + timeline.record("internalize"); + if !llvm::LLVMRustPrepareThinLTOImport(self.shared.data.0, llmod) { + let msg = format!("failed to prepare thin LTO module"); + return Err(write::llvm_err(&diag_handler, msg)) + } + cgcx.save_temp_bitcode(&module, "thin-lto-after-import"); + timeline.record("import"); + + // Ok now this is a bit unfortunate. This is also something you won't + // find upstream in LLVM's ThinLTO passes! This is a hack for now to + // work around bugs in LLVM. + // + // First discovered in #45511 it was found that as part of ThinLTO + // importing passes LLVM will import `DICompileUnit` metadata + // information across modules. This means that we'll be working with one + // LLVM module that has multiple `DICompileUnit` instances in it (a + // bunch of `llvm.dbg.cu` members). Unfortunately there's a number of + // bugs in LLVM's backend which generates invalid DWARF in a situation + // like this: + // + // https://bugs.llvm.org/show_bug.cgi?id=35212 + // https://bugs.llvm.org/show_bug.cgi?id=35562 + // + // While the first bug there is fixed the second ended up causing #46346 + // which was basically a resurgence of #45511 after LLVM's bug 35212 was + // fixed. + // + // This function below is a huge hack around this problem. The function + // below is defined in `PassWrapper.cpp` and will basically "merge" + // all `DICompileUnit` instances in a module. Basically it'll take all + // the objects, rewrite all pointers of `DISubprogram` to point to the + // first `DICompileUnit`, and then delete all the other units. + // + // This is probably mangling to the debug info slightly (but hopefully + // not too much) but for now at least gets LLVM to emit valid DWARF (or + // so it appears). Hopefully we can remove this once upstream bugs are + // fixed in LLVM. + llvm::LLVMRustThinLTOPatchDICompileUnit(llmod, cu1); + cgcx.save_temp_bitcode(&module, "thin-lto-after-patch"); + timeline.record("patch"); + + // Alright now that we've done everything related to the ThinLTO + // analysis it's time to run some optimizations! Here we use the same + // `run_pass_manager` as the "fat" LTO above except that we tell it to + // populate a thin-specific pass manager, which presumably LLVM treats a + // little differently. + info!("running thin lto passes over {}", module.name); + let config = cgcx.config(module.kind); + run_pass_manager(cgcx, tm, llmod, config, true); + cgcx.save_temp_bitcode(&module, "thin-lto-after-pm"); + timeline.record("thin-done"); + + // FIXME: this is a hack around a bug in LLVM right now. Discovered in + // #46910 it was found out that on 32-bit MSVC LLVM will hit a codegen + // error if there's an available_externally function in the LLVM module. + // Typically we don't actually use these functions but ThinLTO makes + // heavy use of them when inlining across modules. + // + // Tracked upstream at https://bugs.llvm.org/show_bug.cgi?id=35736 this + // function call (and its definition on the C++ side of things) + // shouldn't be necessary eventually and we can safetly delete these few + // lines. + llvm::LLVMRustThinLTORemoveAvailableExternally(llmod); + cgcx.save_temp_bitcode(&module, "thin-lto-after-rm-ae"); + timeline.record("no-ae"); + + Ok(module) + } +} diff --git a/src/librustc_codegen_llvm/back/rpath.rs b/src/librustc_codegen_llvm/back/rpath.rs new file mode 100644 index 00000000000..8e5e7d37648 --- /dev/null +++ b/src/librustc_codegen_llvm/back/rpath.rs @@ -0,0 +1,282 @@ +// Copyright 2012-2015 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 std::collections::HashSet; +use std::env; +use std::path::{Path, PathBuf}; +use std::fs; + +use rustc::hir::def_id::CrateNum; +use rustc::middle::cstore::LibSource; + +pub struct RPathConfig<'a> { + pub used_crates: &'a [(CrateNum, LibSource)], + pub out_filename: PathBuf, + pub is_like_osx: bool, + pub has_rpath: bool, + pub linker_is_gnu: bool, + pub get_install_prefix_lib_path: &'a mut FnMut() -> PathBuf, +} + +pub fn get_rpath_flags(config: &mut RPathConfig) -> Vec<String> { + // No rpath on windows + if !config.has_rpath { + return Vec::new(); + } + + let mut flags = Vec::new(); + + debug!("preparing the RPATH!"); + + let libs = config.used_crates.clone(); + let libs = libs.iter().filter_map(|&(_, ref l)| l.option()).collect::<Vec<_>>(); + let rpaths = get_rpaths(config, &libs); + flags.extend_from_slice(&rpaths_to_flags(&rpaths)); + + // Use DT_RUNPATH instead of DT_RPATH if available + if config.linker_is_gnu { + flags.push("-Wl,--enable-new-dtags".to_string()); + } + + flags +} + +fn rpaths_to_flags(rpaths: &[String]) -> Vec<String> { + let mut ret = Vec::new(); + for rpath in rpaths { + if rpath.contains(',') { + ret.push("-Wl,-rpath".into()); + ret.push("-Xlinker".into()); + ret.push(rpath.clone()); + } else { + ret.push(format!("-Wl,-rpath,{}", &(*rpath))); + } + } + return ret; +} + +fn get_rpaths(config: &mut RPathConfig, libs: &[PathBuf]) -> Vec<String> { + debug!("output: {:?}", config.out_filename.display()); + debug!("libs:"); + for libpath in libs { + debug!(" {:?}", libpath.display()); + } + + // Use relative paths to the libraries. Binaries can be moved + // as long as they maintain the relative relationship to the + // crates they depend on. + let rel_rpaths = get_rpaths_relative_to_output(config, libs); + + // And a final backup rpath to the global library location. + let fallback_rpaths = vec![get_install_prefix_rpath(config)]; + + fn log_rpaths(desc: &str, rpaths: &[String]) { + debug!("{} rpaths:", desc); + for rpath in rpaths { + debug!(" {}", *rpath); + } + } + + log_rpaths("relative", &rel_rpaths); + log_rpaths("fallback", &fallback_rpaths); + + let mut rpaths = rel_rpaths; + rpaths.extend_from_slice(&fallback_rpaths); + + // Remove duplicates + let rpaths = minimize_rpaths(&rpaths); + return rpaths; +} + +fn get_rpaths_relative_to_output(config: &mut RPathConfig, + libs: &[PathBuf]) -> Vec<String> { + libs.iter().map(|a| get_rpath_relative_to_output(config, a)).collect() +} + +fn get_rpath_relative_to_output(config: &mut RPathConfig, lib: &Path) -> String { + // Mac doesn't appear to support $ORIGIN + let prefix = if config.is_like_osx { + "@loader_path" + } else { + "$ORIGIN" + }; + + let cwd = env::current_dir().unwrap(); + let mut lib = fs::canonicalize(&cwd.join(lib)).unwrap_or(cwd.join(lib)); + lib.pop(); + let mut output = cwd.join(&config.out_filename); + output.pop(); + let output = fs::canonicalize(&output).unwrap_or(output); + let relative = path_relative_from(&lib, &output) + .expect(&format!("couldn't create relative path from {:?} to {:?}", output, lib)); + // FIXME (#9639): This needs to handle non-utf8 paths + format!("{}/{}", prefix, + relative.to_str().expect("non-utf8 component in path")) +} + +// This routine is adapted from the *old* Path's `path_relative_from` +// function, which works differently from the new `relative_from` function. +// In particular, this handles the case on unix where both paths are +// absolute but with only the root as the common directory. +fn path_relative_from(path: &Path, base: &Path) -> Option<PathBuf> { + use std::path::Component; + + if path.is_absolute() != base.is_absolute() { + if path.is_absolute() { + Some(PathBuf::from(path)) + } else { + None + } + } else { + let mut ita = path.components(); + let mut itb = base.components(); + let mut comps: Vec<Component> = vec![]; + loop { + match (ita.next(), itb.next()) { + (None, None) => break, + (Some(a), None) => { + comps.push(a); + comps.extend(ita.by_ref()); + break; + } + (None, _) => comps.push(Component::ParentDir), + (Some(a), Some(b)) if comps.is_empty() && a == b => (), + (Some(a), Some(b)) if b == Component::CurDir => comps.push(a), + (Some(_), Some(b)) if b == Component::ParentDir => return None, + (Some(a), Some(_)) => { + comps.push(Component::ParentDir); + for _ in itb { + comps.push(Component::ParentDir); + } + comps.push(a); + comps.extend(ita.by_ref()); + break; + } + } + } + Some(comps.iter().map(|c| c.as_os_str()).collect()) + } +} + + +fn get_install_prefix_rpath(config: &mut RPathConfig) -> String { + let path = (config.get_install_prefix_lib_path)(); + let path = env::current_dir().unwrap().join(&path); + // FIXME (#9639): This needs to handle non-utf8 paths + path.to_str().expect("non-utf8 component in rpath").to_string() +} + +fn minimize_rpaths(rpaths: &[String]) -> Vec<String> { + let mut set = HashSet::new(); + let mut minimized = Vec::new(); + for rpath in rpaths { + if set.insert(rpath) { + minimized.push(rpath.clone()); + } + } + minimized +} + +#[cfg(all(unix, test))] +mod tests { + use super::{RPathConfig}; + use super::{minimize_rpaths, rpaths_to_flags, get_rpath_relative_to_output}; + use std::path::{Path, PathBuf}; + + #[test] + fn test_rpaths_to_flags() { + let flags = rpaths_to_flags(&[ + "path1".to_string(), + "path2".to_string() + ]); + assert_eq!(flags, + ["-Wl,-rpath,path1", + "-Wl,-rpath,path2"]); + } + + #[test] + fn test_minimize1() { + let res = minimize_rpaths(&[ + "rpath1".to_string(), + "rpath2".to_string(), + "rpath1".to_string() + ]); + assert!(res == [ + "rpath1", + "rpath2", + ]); + } + + #[test] + fn test_minimize2() { + let res = minimize_rpaths(&[ + "1a".to_string(), + "2".to_string(), + "2".to_string(), + "1a".to_string(), + "4a".to_string(), + "1a".to_string(), + "2".to_string(), + "3".to_string(), + "4a".to_string(), + "3".to_string() + ]); + assert!(res == [ + "1a", + "2", + "4a", + "3", + ]); + } + + #[test] + fn test_rpath_relative() { + if cfg!(target_os = "macos") { + let config = &mut RPathConfig { + used_crates: Vec::new(), + has_rpath: true, + is_like_osx: true, + linker_is_gnu: false, + out_filename: PathBuf::from("bin/rustc"), + get_install_prefix_lib_path: &mut || panic!(), + }; + let res = get_rpath_relative_to_output(config, + Path::new("lib/libstd.so")); + assert_eq!(res, "@loader_path/../lib"); + } else { + let config = &mut RPathConfig { + used_crates: Vec::new(), + out_filename: PathBuf::from("bin/rustc"), + get_install_prefix_lib_path: &mut || panic!(), + has_rpath: true, + is_like_osx: false, + linker_is_gnu: true, + }; + let res = get_rpath_relative_to_output(config, + Path::new("lib/libstd.so")); + assert_eq!(res, "$ORIGIN/../lib"); + } + } + + #[test] + fn test_xlinker() { + let args = rpaths_to_flags(&[ + "a/normal/path".to_string(), + "a,comma,path".to_string() + ]); + + assert_eq!(args, vec![ + "-Wl,-rpath,a/normal/path".to_string(), + "-Wl,-rpath".to_string(), + "-Xlinker".to_string(), + "a,comma,path".to_string() + ]); + } +} diff --git a/src/librustc_codegen_llvm/back/symbol_export.rs b/src/librustc_codegen_llvm/back/symbol_export.rs new file mode 100644 index 00000000000..81ac684aee2 --- /dev/null +++ b/src/librustc_codegen_llvm/back/symbol_export.rs @@ -0,0 +1,396 @@ +// 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 monomorphize::Instance; +use rustc::hir; +use rustc::hir::CodegenFnAttrFlags; +use rustc::hir::def_id::{CrateNum, DefId, LOCAL_CRATE, CRATE_DEF_INDEX}; +use rustc::ich::Fingerprint; +use rustc::middle::exported_symbols::{SymbolExportLevel, ExportedSymbol, metadata_symbol_name}; +use rustc::session::config; +use rustc::ty::{TyCtxt, SymbolName}; +use rustc::ty::maps::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::CrateTypeExecutable | + config::CrateTypeStaticlib | + config::CrateTypeProcMacro | + config::CrateTypeCdylib => SymbolExportLevel::C, + config::CrateTypeRlib | + config::CrateTypeDylib => 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) { + hir::map::NodeForeignItem(..) => { + 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. + hir::map::NodeItem(&hir::Item { + node: hir::ItemStatic(..), + .. + }) | + hir::map::NodeItem(&hir::Item { + node: hir::ItemFn(..), .. + }) | + hir::map::NodeImplItem(&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 let Some(_) = *tcx.sess.entry_fn.borrow() { + let symbol_name = "main".to_string(); + let exported_symbol = ExportedSymbol::NoDefId(SymbolName::new(&symbol_name)); + + 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::CrateTypeDylib) { + let symbol_name = metadata_symbol_name(tcx); + let exported_symbol = ExportedSymbol::NoDefId(SymbolName::new(&symbol_name)); + + symbols.push((exported_symbol, SymbolExportLevel::Rust)); + } + + if tcx.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(); + + 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_insert_with(|| FxHashMap()); + + 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 { + SymbolExportLevel::C + } else { + SymbolExportLevel::Rust + } +} diff --git a/src/librustc_codegen_llvm/back/wasm.rs b/src/librustc_codegen_llvm/back/wasm.rs new file mode 100644 index 00000000000..d6d386c9fbe --- /dev/null +++ b/src/librustc_codegen_llvm/back/wasm.rs @@ -0,0 +1,261 @@ +// 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 std::collections::BTreeMap; +use std::fs; +use std::path::Path; +use std::str; + +use rustc_data_structures::fx::FxHashMap; +use serialize::leb128; + +// https://webassembly.github.io/spec/core/binary/modules.html#binary-importsec +const WASM_IMPORT_SECTION_ID: u8 = 2; + +const WASM_EXTERNAL_KIND_FUNCTION: u8 = 0; +const WASM_EXTERNAL_KIND_TABLE: u8 = 1; +const WASM_EXTERNAL_KIND_MEMORY: u8 = 2; +const WASM_EXTERNAL_KIND_GLOBAL: u8 = 3; + +/// Append all the custom sections listed in `sections` to the wasm binary +/// specified at `path`. +/// +/// LLVM 6 which we're using right now doesn't have the ability to create custom +/// sections in wasm files nor does LLD have the ability to merge these sections +/// into one larger section when linking. It's expected that this will +/// eventually get implemented, however! +/// +/// Until that time though this is a custom implementation in rustc to append +/// all sections to a wasm file to the finished product that LLD produces. +/// +/// Support for this is landing in LLVM in https://reviews.llvm.org/D43097, +/// although after that support will need to be in LLD as well. +pub fn add_custom_sections(path: &Path, sections: &BTreeMap<String, Vec<u8>>) { + if sections.len() == 0 { + return + } + + let wasm = fs::read(path).expect("failed to read wasm output"); + + // see https://webassembly.github.io/spec/core/binary/modules.html#custom-section + let mut wasm = WasmEncoder { data: wasm }; + for (section, bytes) in sections { + // write the `id` identifier, 0 for a custom section + wasm.byte(0); + + // figure out how long our name descriptor will be + let mut name = WasmEncoder::new(); + name.str(section); + + // write the length of the payload followed by all its contents + wasm.u32((bytes.len() + name.data.len()) as u32); + wasm.data.extend_from_slice(&name.data); + wasm.data.extend_from_slice(bytes); + } + + fs::write(path, &wasm.data).expect("failed to write wasm output"); +} + +/// Rewrite the module imports are listed from in a wasm module given the field +/// name to module name mapping in `import_map`. +/// +/// LLVM 6 which we're using right now doesn't have the ability to configure the +/// module a wasm symbol is import from. Rather all imported symbols come from +/// the bland `"env"` module unconditionally. Furthermore we'd *also* need +/// support in LLD for preserving these import modules, which it unfortunately +/// currently does not. +/// +/// This function is intended as a hack for now where we manually rewrite the +/// wasm output by LLVM to have the correct import modules listed. The +/// `#[wasm_import_module]` attribute in Rust translates to the module that each +/// symbol is imported from, so here we manually go through the wasm file, +/// decode it, rewrite imports, and then rewrite the wasm module. +/// +/// Support for this was added to LLVM in +/// https://github.com/llvm-mirror/llvm/commit/0f32e1365, although support still +/// needs to be added (AFAIK at the time of this writing) to LLD +pub fn rewrite_imports(path: &Path, import_map: &FxHashMap<String, String>) { + if import_map.len() == 0 { + return + } + + let wasm = fs::read(path).expect("failed to read wasm output"); + let mut ret = WasmEncoder::new(); + ret.data.extend(&wasm[..8]); + + // skip the 8 byte wasm/version header + for (id, raw) in WasmSections(WasmDecoder::new(&wasm[8..])) { + ret.byte(id); + if id == WASM_IMPORT_SECTION_ID { + info!("rewriting import section"); + let data = rewrite_import_section( + &mut WasmDecoder::new(raw), + import_map, + ); + ret.bytes(&data); + } else { + info!("carry forward section {}, {} bytes long", id, raw.len()); + ret.bytes(raw); + } + } + + fs::write(path, &ret.data).expect("failed to write wasm output"); + + fn rewrite_import_section( + wasm: &mut WasmDecoder, + import_map: &FxHashMap<String, String>, + ) + -> Vec<u8> + { + let mut dst = WasmEncoder::new(); + let n = wasm.u32(); + dst.u32(n); + info!("rewriting {} imports", n); + for _ in 0..n { + rewrite_import_entry(wasm, &mut dst, import_map); + } + return dst.data + } + + fn rewrite_import_entry(wasm: &mut WasmDecoder, + dst: &mut WasmEncoder, + import_map: &FxHashMap<String, String>) { + // More info about the binary format here is available at: + // https://webassembly.github.io/spec/core/binary/modules.html#import-section + // + // Note that you can also find the whole point of existence of this + // function here, where we map the `module` name to a different one if + // we've got one listed. + let module = wasm.str(); + let field = wasm.str(); + let new_module = if module == "env" { + import_map.get(field).map(|s| &**s).unwrap_or(module) + } else { + module + }; + info!("import rewrite ({} => {}) / {}", module, new_module, field); + dst.str(new_module); + dst.str(field); + let kind = wasm.byte(); + dst.byte(kind); + match kind { + WASM_EXTERNAL_KIND_FUNCTION => dst.u32(wasm.u32()), + WASM_EXTERNAL_KIND_TABLE => { + dst.byte(wasm.byte()); // element_type + dst.limits(wasm.limits()); + } + WASM_EXTERNAL_KIND_MEMORY => dst.limits(wasm.limits()), + WASM_EXTERNAL_KIND_GLOBAL => { + dst.byte(wasm.byte()); // content_type + dst.bool(wasm.bool()); // mutable + } + b => panic!("unknown kind: {}", b), + } + } +} + +struct WasmSections<'a>(WasmDecoder<'a>); + +impl<'a> Iterator for WasmSections<'a> { + type Item = (u8, &'a [u8]); + + fn next(&mut self) -> Option<(u8, &'a [u8])> { + if self.0.data.len() == 0 { + return None + } + + // see https://webassembly.github.io/spec/core/binary/modules.html#sections + let id = self.0.byte(); + let section_len = self.0.u32(); + info!("new section {} / {} bytes", id, section_len); + let section = self.0.skip(section_len as usize); + Some((id, section)) + } +} + +struct WasmDecoder<'a> { + data: &'a [u8], +} + +impl<'a> WasmDecoder<'a> { + fn new(data: &'a [u8]) -> WasmDecoder<'a> { + WasmDecoder { data } + } + + fn byte(&mut self) -> u8 { + self.skip(1)[0] + } + + fn u32(&mut self) -> u32 { + let (n, l1) = leb128::read_u32_leb128(self.data); + self.data = &self.data[l1..]; + return n + } + + fn skip(&mut self, amt: usize) -> &'a [u8] { + let (data, rest) = self.data.split_at(amt); + self.data = rest; + data + } + + fn str(&mut self) -> &'a str { + let len = self.u32(); + str::from_utf8(self.skip(len as usize)).unwrap() + } + + fn bool(&mut self) -> bool { + self.byte() == 1 + } + + fn limits(&mut self) -> (u32, Option<u32>) { + let has_max = self.bool(); + (self.u32(), if has_max { Some(self.u32()) } else { None }) + } +} + +struct WasmEncoder { + data: Vec<u8>, +} + +impl WasmEncoder { + fn new() -> WasmEncoder { + WasmEncoder { data: Vec::new() } + } + + fn u32(&mut self, val: u32) { + let at = self.data.len(); + leb128::write_u32_leb128(&mut self.data, at, val); + } + + fn byte(&mut self, val: u8) { + self.data.push(val); + } + + fn bytes(&mut self, val: &[u8]) { + self.u32(val.len() as u32); + self.data.extend_from_slice(val); + } + + fn str(&mut self, val: &str) { + self.bytes(val.as_bytes()) + } + + fn bool(&mut self, b: bool) { + self.byte(b as u8); + } + + fn limits(&mut self, limits: (u32, Option<u32>)) { + self.bool(limits.1.is_some()); + self.u32(limits.0); + if let Some(c) = limits.1 { + self.u32(c); + } + } +} diff --git a/src/librustc_codegen_llvm/back/write.rs b/src/librustc_codegen_llvm/back/write.rs new file mode 100644 index 00000000000..1151e013131 --- /dev/null +++ b/src/librustc_codegen_llvm/back/write.rs @@ -0,0 +1,2390 @@ +// 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 attributes; +use back::bytecode::{self, RLIB_BYTECODE_EXTENSION}; +use back::lto::{self, ModuleBuffer, ThinBuffer}; +use back::link::{self, get_linker, remove}; +use back::command::Command; +use back::linker::LinkerInfo; +use back::symbol_export::ExportedSymbols; +use base; +use consts; +use rustc_incremental::{copy_cgu_workproducts_to_incr_comp_cache_dir, in_incr_comp_dir}; +use rustc::dep_graph::{WorkProduct, WorkProductId, WorkProductFileKind}; +use rustc::middle::cstore::{LinkMeta, EncodedMetadata}; +use rustc::session::config::{self, OutputFilenames, OutputType, Passes, SomePasses, + AllPasses, Sanitizer, Lto}; +use rustc::session::Session; +use rustc::util::nodemap::FxHashMap; +use time_graph::{self, TimeGraph, Timeline}; +use llvm; +use llvm::{ModuleRef, TargetMachineRef, PassManagerRef, DiagnosticInfoRef}; +use llvm::{SMDiagnosticRef, ContextRef}; +use {CodegenResults, ModuleSource, ModuleCodegen, CompiledModule, ModuleKind}; +use CrateInfo; +use rustc::hir::def_id::{CrateNum, LOCAL_CRATE}; +use rustc::ty::TyCtxt; +use rustc::util::common::{time_ext, time_depth, set_time_depth, print_time_passes_entry}; +use rustc::util::common::path2cstr; +use rustc::util::fs::{link_or_copy}; +use errors::{self, Handler, Level, DiagnosticBuilder, FatalError, DiagnosticId}; +use errors::emitter::{Emitter}; +use syntax::attr; +use syntax::ext::hygiene::Mark; +use syntax_pos::MultiSpan; +use syntax_pos::symbol::Symbol; +use type_::Type; +use context::{is_pie_binary, get_reloc_model}; +use common::{C_bytes_in_context, val_ty}; +use jobserver::{Client, Acquired}; +use rustc_demangle; + +use std::any::Any; +use std::ffi::{CString, CStr}; +use std::fs; +use std::io::{self, Write}; +use std::mem; +use std::path::{Path, PathBuf}; +use std::str; +use std::sync::Arc; +use std::sync::mpsc::{channel, Sender, Receiver}; +use std::slice; +use std::time::Instant; +use std::thread; +use libc::{c_uint, c_void, c_char, size_t}; + +pub const RELOC_MODEL_ARGS : [(&'static str, llvm::RelocMode); 7] = [ + ("pic", llvm::RelocMode::PIC), + ("static", llvm::RelocMode::Static), + ("default", llvm::RelocMode::Default), + ("dynamic-no-pic", llvm::RelocMode::DynamicNoPic), + ("ropi", llvm::RelocMode::ROPI), + ("rwpi", llvm::RelocMode::RWPI), + ("ropi-rwpi", llvm::RelocMode::ROPI_RWPI), +]; + +pub const CODE_GEN_MODEL_ARGS: &[(&str, llvm::CodeModel)] = &[ + ("small", llvm::CodeModel::Small), + ("kernel", llvm::CodeModel::Kernel), + ("medium", llvm::CodeModel::Medium), + ("large", llvm::CodeModel::Large), +]; + +pub const TLS_MODEL_ARGS : [(&'static str, llvm::ThreadLocalMode); 4] = [ + ("global-dynamic", llvm::ThreadLocalMode::GeneralDynamic), + ("local-dynamic", llvm::ThreadLocalMode::LocalDynamic), + ("initial-exec", llvm::ThreadLocalMode::InitialExec), + ("local-exec", llvm::ThreadLocalMode::LocalExec), +]; + +pub fn llvm_err(handler: &errors::Handler, msg: String) -> FatalError { + match llvm::last_error() { + Some(err) => handler.fatal(&format!("{}: {}", msg, err)), + None => handler.fatal(&msg), + } +} + +pub fn write_output_file( + handler: &errors::Handler, + target: llvm::TargetMachineRef, + pm: llvm::PassManagerRef, + m: ModuleRef, + output: &Path, + file_type: llvm::FileType) -> Result<(), FatalError> { + unsafe { + let output_c = path2cstr(output); + let result = llvm::LLVMRustWriteOutputFile( + target, pm, m, output_c.as_ptr(), file_type); + if result.into_result().is_err() { + let msg = format!("could not write output to {}", output.display()); + Err(llvm_err(handler, msg)) + } else { + Ok(()) + } + } +} + +fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel { + match optimize { + config::OptLevel::No => llvm::CodeGenOptLevel::None, + config::OptLevel::Less => llvm::CodeGenOptLevel::Less, + config::OptLevel::Default => llvm::CodeGenOptLevel::Default, + config::OptLevel::Aggressive => llvm::CodeGenOptLevel::Aggressive, + _ => llvm::CodeGenOptLevel::Default, + } +} + +fn get_llvm_opt_size(optimize: config::OptLevel) -> llvm::CodeGenOptSize { + match optimize { + config::OptLevel::Size => llvm::CodeGenOptSizeDefault, + config::OptLevel::SizeMin => llvm::CodeGenOptSizeAggressive, + _ => llvm::CodeGenOptSizeNone, + } +} + +pub fn create_target_machine(sess: &Session, find_features: bool) -> TargetMachineRef { + target_machine_factory(sess, find_features)().unwrap_or_else(|err| { + llvm_err(sess.diagnostic(), err).raise() + }) +} + +// 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. +pub fn target_machine_factory(sess: &Session, find_features: bool) + -> Arc<Fn() -> Result<TargetMachineRef, String> + Send + Sync> +{ + let reloc_model = get_reloc_model(sess); + + let opt_level = get_llvm_opt_level(sess.opts.optimize); + let use_softfp = sess.opts.cg.soft_float; + + let ffunction_sections = sess.target.target.options.function_sections; + let fdata_sections = ffunction_sections; + + let code_model_arg = sess.opts.cg.code_model.as_ref().or( + sess.target.target.options.code_model.as_ref(), + ); + + let code_model = match code_model_arg { + Some(s) => { + match CODE_GEN_MODEL_ARGS.iter().find(|arg| arg.0 == s) { + Some(x) => x.1, + _ => { + sess.err(&format!("{:?} is not a valid code model", + code_model_arg)); + sess.abort_if_errors(); + bug!(); + } + } + } + None => llvm::CodeModel::None, + }; + + let singlethread = sess.target.target.options.singlethread; + + let triple = &sess.target.target.llvm_target; + + let triple = CString::new(triple.as_bytes()).unwrap(); + let cpu = sess.target_cpu(); + let cpu = CString::new(cpu.as_bytes()).unwrap(); + let features = attributes::llvm_target_features(sess) + .collect::<Vec<_>>() + .join(","); + let features = CString::new(features).unwrap(); + let is_pie_binary = !find_features && is_pie_binary(sess); + let trap_unreachable = sess.target.target.options.trap_unreachable; + + Arc::new(move || { + let tm = unsafe { + llvm::LLVMRustCreateTargetMachine( + triple.as_ptr(), cpu.as_ptr(), features.as_ptr(), + code_model, + reloc_model, + opt_level, + use_softfp, + is_pie_binary, + ffunction_sections, + fdata_sections, + trap_unreachable, + singlethread, + ) + }; + + if tm.is_null() { + Err(format!("Could not create LLVM TargetMachine for triple: {}", + triple.to_str().unwrap())) + } else { + Ok(tm) + } + }) +} + +/// Module-specific configuration for `optimize_and_codegen`. +pub struct ModuleConfig { + /// Names of additional optimization passes to run. + 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<llvm::CodeGenOptLevel>, + + /// Some(level) to optimize binary size, or None to not affect program size. + opt_size: Option<llvm::CodeGenOptSize>, + + pgo_gen: Option<String>, + pgo_use: String, + + // Flags indicating which outputs to produce. + emit_no_opt_bc: bool, + emit_bc: bool, + emit_bc_compressed: bool, + emit_lto_bc: bool, + emit_ir: bool, + emit_asm: bool, + emit_obj: bool, + // Miscellaneous flags. These are mostly copied from command-line + // options. + no_verify: bool, + no_prepopulate_passes: bool, + no_builtins: bool, + time_passes: bool, + vectorize_loop: bool, + vectorize_slp: bool, + merge_functions: bool, + 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. + obj_is_bitcode: bool, + no_integrated_as: bool, + embed_bitcode: bool, + 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_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, + + no_verify: 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.no_verify = sess.no_verify(); + 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; + let embed_bitcode = sess.target.target.options.embed_bitcode || + sess.opts.debugging_opts.embed_bitcode || + sess.opts.debugging_opts.cross_lang_lto.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; + } +} + +/// Assembler name and command used by codegen when no_integrated_as is enabled +struct AssemblerCommand { + name: PathBuf, + cmd: Command, +} + +/// Additional resources used by optimize_and_codegen (not module specific) +#[derive(Clone)] +pub struct CodegenContext { + // Resouces needed when running LTO + 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)>, + output_filenames: Arc<OutputFilenames>, + regular_module_config: Arc<ModuleConfig>, + metadata_module_config: Arc<ModuleConfig>, + allocator_module_config: Arc<ModuleConfig>, + pub tm_factory: Arc<Fn() -> Result<TargetMachineRef, String> + Send + Sync>, + pub msvc_imps_needed: bool, + pub target_pointer_width: String, + debuginfo: config::DebugInfoLevel, + + // 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>, + // Channel back to the main control thread to send messages to + coordinator_send: Sender<Box<Any + Send>>, + // A reference to the TimeGraph so we can register timings. None means that + // measuring is disabled. + time_graph: Option<TimeGraph>, + // The assembler command if no_integrated_as option is enabled, None otherwise + assembler_cmd: Option<Arc<AssemblerCommand>>, +} + +impl CodegenContext { + pub fn create_diag_handler(&self) -> Handler { + Handler::with_emitter(true, false, Box::new(self.diag_emitter.clone())) + } + + pub(crate) 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, + } + } + + pub(crate) fn save_temp_bitcode(&self, module: &ModuleCodegen, name: &str) { + if !self.save_temps { + return + } + unsafe { + let ext = format!("{}.bc", name); + let cgu = Some(&module.name[..]); + let path = self.output_filenames.temp_path_ext(&ext, cgu); + let cstr = path2cstr(&path); + let llmod = module.llvm().unwrap().llmod; + llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr()); + } + } +} + +struct DiagnosticHandlers<'a> { + inner: Box<(&'a CodegenContext, &'a Handler)>, + llcx: ContextRef, +} + +impl<'a> DiagnosticHandlers<'a> { + fn new(cgcx: &'a CodegenContext, + handler: &'a Handler, + llcx: ContextRef) -> DiagnosticHandlers<'a> { + let data = Box::new((cgcx, handler)); + unsafe { + let arg = &*data as &(_, _) as *const _ as *mut _; + llvm::LLVMRustSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, arg); + llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, arg); + } + DiagnosticHandlers { + inner: data, + llcx: llcx, + } + } +} + +impl<'a> Drop for DiagnosticHandlers<'a> { + fn drop(&mut self) { + unsafe { + llvm::LLVMRustSetInlineAsmDiagnosticHandler(self.llcx, inline_asm_handler, 0 as *mut _); + llvm::LLVMContextSetDiagnosticHandler(self.llcx, diagnostic_handler, 0 as *mut _); + } + } +} + +unsafe extern "C" fn report_inline_asm<'a, 'b>(cgcx: &'a CodegenContext, + msg: &'b str, + cookie: c_uint) { + cgcx.diag_emitter.inline_asm_error(cookie as u32, msg.to_string()); +} + +unsafe extern "C" fn inline_asm_handler(diag: SMDiagnosticRef, + user: *const c_void, + cookie: c_uint) { + if user.is_null() { + return + } + let (cgcx, _) = *(user as *const (&CodegenContext, &Handler)); + + let msg = llvm::build_string(|s| llvm::LLVMRustWriteSMDiagnosticToString(diag, s)) + .expect("non-UTF8 SMDiagnostic"); + + report_inline_asm(cgcx, &msg, cookie); +} + +unsafe extern "C" fn diagnostic_handler(info: DiagnosticInfoRef, user: *mut c_void) { + if user.is_null() { + return + } + let (cgcx, diag_handler) = *(user as *const (&CodegenContext, &Handler)); + + match llvm::diagnostic::Diagnostic::unpack(info) { + llvm::diagnostic::InlineAsm(inline) => { + report_inline_asm(cgcx, + &llvm::twine_to_string(inline.message), + inline.cookie); + } + + llvm::diagnostic::Optimization(opt) => { + let enabled = match cgcx.remark { + AllPasses => true, + SomePasses(ref v) => v.iter().any(|s| *s == opt.pass_name), + }; + + if enabled { + diag_handler.note_without_error(&format!("optimization {} for {} at {}:{}:{}: {}", + opt.kind.describe(), + opt.pass_name, + opt.filename, + opt.line, + opt.column, + opt.message)); + } + } + llvm::diagnostic::PGO(diagnostic_ref) => { + let msg = llvm::build_string(|s| { + llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s) + }).expect("non-UTF8 PGO diagnostic"); + diag_handler.warn(&msg); + } + llvm::diagnostic::UnknownDiagnostic(..) => {}, + } +} + +// Unsafe due to LLVM calls. +unsafe fn optimize(cgcx: &CodegenContext, + diag_handler: &Handler, + module: &ModuleCodegen, + config: &ModuleConfig, + timeline: &mut Timeline) + -> Result<(), FatalError> +{ + let (llmod, llcx, tm) = match module.source { + ModuleSource::Codegened(ref llvm) => (llvm.llmod, llvm.llcx, llvm.tm), + ModuleSource::Preexisting(_) => { + bug!("optimize_and_codegen: called with ModuleSource::Preexisting") + } + }; + + let _handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx); + + let module_name = module.name.clone(); + let module_name = Some(&module_name[..]); + + if config.emit_no_opt_bc { + let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name); + let out = path2cstr(&out); + llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr()); + } + + if config.opt_level.is_some() { + // Create the two optimizing pass managers. These mirror what clang + // does, and are by populated by LLVM's default PassManagerBuilder. + // Each manager has a different set of passes, but they also share + // some common passes. + let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod); + let mpm = llvm::LLVMCreatePassManager(); + + // If we're verifying or linting, add them to the function pass + // manager. + let addpass = |pass_name: &str| { + let pass_name = CString::new(pass_name).unwrap(); + let pass = llvm::LLVMRustFindAndCreatePass(pass_name.as_ptr()); + if pass.is_null() { + return false; + } + let pass_manager = match llvm::LLVMRustPassKind(pass) { + llvm::PassKind::Function => fpm, + llvm::PassKind::Module => mpm, + llvm::PassKind::Other => { + diag_handler.err("Encountered LLVM pass kind we can't handle"); + return true + }, + }; + llvm::LLVMRustAddPass(pass_manager, pass); + true + }; + + if !config.no_verify { assert!(addpass("verify")); } + if !config.no_prepopulate_passes { + llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod); + llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod); + let opt_level = config.opt_level.unwrap_or(llvm::CodeGenOptLevel::None); + let prepare_for_thin_lto = cgcx.lto == Lto::Thin || cgcx.lto == Lto::ThinLocal; + with_llvm_pmb(llmod, &config, opt_level, prepare_for_thin_lto, &mut |b| { + llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(b, fpm); + llvm::LLVMPassManagerBuilderPopulateModulePassManager(b, mpm); + }) + } + + for pass in &config.passes { + if !addpass(pass) { + diag_handler.warn(&format!("unknown pass `{}`, ignoring", + pass)); + } + } + + for pass in &cgcx.plugin_passes { + if !addpass(pass) { + diag_handler.err(&format!("a plugin asked for LLVM pass \ + `{}` but LLVM does not \ + recognize it", pass)); + } + } + + diag_handler.abort_if_errors(); + + // Finally, run the actual optimization passes + time_ext(config.time_passes, + None, + &format!("llvm function passes [{}]", module_name.unwrap()), + || { + llvm::LLVMRustRunFunctionPassManager(fpm, llmod) + }); + timeline.record("fpm"); + time_ext(config.time_passes, + None, + &format!("llvm module passes [{}]", module_name.unwrap()), + || { + llvm::LLVMRunPassManager(mpm, llmod) + }); + + // Deallocate managers that we're now done with + llvm::LLVMDisposePassManager(fpm); + llvm::LLVMDisposePassManager(mpm); + } + Ok(()) +} + +fn generate_lto_work(cgcx: &CodegenContext, + modules: Vec<ModuleCodegen>) + -> Vec<(WorkItem, 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 = lto::run(cgcx, modules, &mut timeline) + .unwrap_or_else(|e| e.raise()); + + lto_modules.into_iter().map(|module| { + let cost = module.cost(); + (WorkItem::LTO(module), cost) + }).collect() +} + +unsafe fn codegen(cgcx: &CodegenContext, + diag_handler: &Handler, + module: ModuleCodegen, + config: &ModuleConfig, + timeline: &mut Timeline) + -> Result<CompiledModule, FatalError> +{ + timeline.record("codegen"); + let (llmod, llcx, tm) = match module.source { + ModuleSource::Codegened(ref llvm) => (llvm.llmod, llvm.llcx, llvm.tm), + ModuleSource::Preexisting(_) => { + bug!("codegen: called with ModuleSource::Preexisting") + } + }; + let module_name = module.name.clone(); + let module_name = Some(&module_name[..]); + let handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx); + + if cgcx.msvc_imps_needed { + create_msvc_imps(cgcx, llcx, llmod); + } + + // A codegen-specific pass manager is used to generate object + // files for an LLVM module. + // + // Apparently each of these pass managers is a one-shot kind of + // thing, so we create a new one for each type of output. The + // pass manager passed to the closure should be ensured to not + // escape the closure itself, and the manager should only be + // used once. + unsafe fn with_codegen<F, R>(tm: TargetMachineRef, + llmod: ModuleRef, + no_builtins: bool, + f: F) -> R + where F: FnOnce(PassManagerRef) -> R, + { + let cpm = llvm::LLVMCreatePassManager(); + llvm::LLVMRustAddAnalysisPasses(tm, cpm, llmod); + llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins); + f(cpm) + } + + // If we don't have the integrated assembler, then we need to emit asm + // from LLVM and use `gcc` to create the object file. + let asm_to_obj = config.emit_obj && config.no_integrated_as; + + // Change what we write and cleanup based on whether obj files are + // just llvm bitcode. In that case write bitcode, and possibly + // delete the bitcode if it wasn't requested. Don't generate the + // machine code, instead copy the .o file from the .bc + let write_bc = config.emit_bc || config.obj_is_bitcode; + let rm_bc = !config.emit_bc && config.obj_is_bitcode; + let write_obj = config.emit_obj && !config.obj_is_bitcode && !asm_to_obj; + let copy_bc_to_obj = config.emit_obj && config.obj_is_bitcode; + + let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name); + let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name); + + + if write_bc || config.emit_bc_compressed || config.embed_bitcode { + let thin; + let old; + let data = if llvm::LLVMRustThinLTOAvailable() { + thin = ThinBuffer::new(llmod); + thin.data() + } else { + old = ModuleBuffer::new(llmod); + old.data() + }; + timeline.record("make-bc"); + + if write_bc { + if let Err(e) = fs::write(&bc_out, data) { + diag_handler.err(&format!("failed to write bytecode: {}", e)); + } + timeline.record("write-bc"); + } + + if config.embed_bitcode { + embed_bitcode(cgcx, llcx, llmod, Some(data)); + timeline.record("embed-bc"); + } + + if config.emit_bc_compressed { + let dst = bc_out.with_extension(RLIB_BYTECODE_EXTENSION); + let data = bytecode::encode(&module.llmod_id, data); + if let Err(e) = fs::write(&dst, data) { + diag_handler.err(&format!("failed to write bytecode: {}", e)); + } + timeline.record("compress-bc"); + } + } else if config.embed_bitcode_marker { + embed_bitcode(cgcx, llcx, llmod, None); + } + + time_ext(config.time_passes, None, &format!("codegen passes [{}]", module_name.unwrap()), + || -> Result<(), FatalError> { + if config.emit_ir { + let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name); + let out = path2cstr(&out); + + extern "C" fn demangle_callback(input_ptr: *const c_char, + input_len: size_t, + output_ptr: *mut c_char, + output_len: size_t) -> size_t { + let input = unsafe { + slice::from_raw_parts(input_ptr as *const u8, input_len as usize) + }; + + let input = match str::from_utf8(input) { + Ok(s) => s, + Err(_) => return 0, + }; + + let output = unsafe { + slice::from_raw_parts_mut(output_ptr as *mut u8, output_len as usize) + }; + let mut cursor = io::Cursor::new(output); + + let demangled = match rustc_demangle::try_demangle(input) { + Ok(d) => d, + Err(_) => return 0, + }; + + if let Err(_) = write!(cursor, "{:#}", demangled) { + // Possible only if provided buffer is not big enough + return 0; + } + + cursor.position() as size_t + } + + with_codegen(tm, llmod, config.no_builtins, |cpm| { + llvm::LLVMRustPrintModule(cpm, llmod, out.as_ptr(), demangle_callback); + llvm::LLVMDisposePassManager(cpm); + }); + timeline.record("ir"); + } + + if config.emit_asm || asm_to_obj { + let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name); + + // We can't use the same module for asm and binary output, because that triggers + // various errors like invalid IR or broken binaries, so we might have to clone the + // module to produce the asm output + let llmod = if config.emit_obj { + llvm::LLVMCloneModule(llmod) + } else { + llmod + }; + with_codegen(tm, llmod, config.no_builtins, |cpm| { + write_output_file(diag_handler, tm, cpm, llmod, &path, + llvm::FileType::AssemblyFile) + })?; + if config.emit_obj { + llvm::LLVMDisposeModule(llmod); + } + timeline.record("asm"); + } + + if write_obj { + with_codegen(tm, llmod, config.no_builtins, |cpm| { + write_output_file(diag_handler, tm, cpm, llmod, &obj_out, + llvm::FileType::ObjectFile) + })?; + timeline.record("obj"); + } else if asm_to_obj { + let assembly = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name); + run_assembler(cgcx, diag_handler, &assembly, &obj_out); + timeline.record("asm_to_obj"); + + if !config.emit_asm && !cgcx.save_temps { + drop(fs::remove_file(&assembly)); + } + } + + Ok(()) + })?; + + if copy_bc_to_obj { + debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out); + if let Err(e) = link_or_copy(&bc_out, &obj_out) { + diag_handler.err(&format!("failed to copy bitcode to object file: {}", e)); + } + } + + if rm_bc { + debug!("removing_bitcode {:?}", bc_out); + if let Err(e) = fs::remove_file(&bc_out) { + diag_handler.err(&format!("failed to remove bitcode: {}", e)); + } + } + + drop(handlers); + Ok(module.into_compiled_module(config.emit_obj, + config.emit_bc, + config.emit_bc_compressed, + &cgcx.output_filenames)) +} + +/// Embed the bitcode of an LLVM module in the LLVM module itself. +/// +/// This is done primarily for iOS where it appears to be standard to compile C +/// code at least with `-fembed-bitcode` which creates two sections in the +/// executable: +/// +/// * __LLVM,__bitcode +/// * __LLVM,__cmdline +/// +/// It appears *both* of these sections are necessary to get the linker to +/// recognize what's going on. For us though we just always throw in an empty +/// cmdline section. +/// +/// Furthermore debug/O1 builds don't actually embed bitcode but rather just +/// embed an empty section. +/// +/// Basically all of this is us attempting to follow in the footsteps of clang +/// on iOS. See #35968 for lots more info. +unsafe fn embed_bitcode(cgcx: &CodegenContext, + llcx: ContextRef, + llmod: ModuleRef, + bitcode: Option<&[u8]>) { + let llconst = C_bytes_in_context(llcx, bitcode.unwrap_or(&[])); + let llglobal = llvm::LLVMAddGlobal( + llmod, + val_ty(llconst).to_ref(), + "rustc.embedded.module\0".as_ptr() as *const _, + ); + llvm::LLVMSetInitializer(llglobal, llconst); + + let is_apple = cgcx.opts.target_triple.triple().contains("-ios") || + cgcx.opts.target_triple.triple().contains("-darwin"); + + let section = if is_apple { + "__LLVM,__bitcode\0" + } else { + ".llvmbc\0" + }; + llvm::LLVMSetSection(llglobal, section.as_ptr() as *const _); + llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage); + llvm::LLVMSetGlobalConstant(llglobal, llvm::True); + + let llconst = C_bytes_in_context(llcx, &[]); + let llglobal = llvm::LLVMAddGlobal( + llmod, + val_ty(llconst).to_ref(), + "rustc.embedded.cmdline\0".as_ptr() as *const _, + ); + llvm::LLVMSetInitializer(llglobal, llconst); + let section = if is_apple { + "__LLVM,__cmdline\0" + } else { + ".llvmcmd\0" + }; + llvm::LLVMSetSection(llglobal, section.as_ptr() as *const _); + llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage); +} + +pub(crate) 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::CrateTypeRlib) && + sess.opts.output_types.contains_key(&OutputType::Exe) +} + +pub fn start_async_codegen(tcx: TyCtxt, + time_graph: Option<TimeGraph>, + link: LinkMeta, + metadata: EncodedMetadata, + coordinator_receive: Receiver<Box<Any + Send>>, + total_cgus: usize) + -> OngoingCodegen { + let sess = tcx.sess; + let crate_name = tcx.crate_name(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(get_llvm_opt_level(sess.opts.optimize)); + modules_config.opt_size = Some(get_llvm_opt_size(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_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.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(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 { + crate_name, + link, + 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() { + 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(crate) fn dump_incremental_data(codegen_results: &CodegenResults) { + println!("[incremental] Re-using {} out of {} modules", + codegen_results.modules.iter().filter(|m| m.pre_existing).count(), + codegen_results.modules.len()); +} + +enum WorkItem { + Optimize(ModuleCodegen), + LTO(lto::LtoModuleCodegen), +} + +impl WorkItem { + fn kind(&self) -> ModuleKind { + match *self { + WorkItem::Optimize(ref m) => m.kind, + WorkItem::LTO(_) => ModuleKind::Regular, + } + } + + fn name(&self) -> String { + match *self { + WorkItem::Optimize(ref m) => format!("optimize: {}", m.name), + WorkItem::LTO(ref m) => format!("lto: {}", m.name()), + } + } +} + +enum WorkItemResult { + Compiled(CompiledModule), + NeedsLTO(ModuleCodegen), +} + +fn execute_work_item(cgcx: &CodegenContext, + work_item: WorkItem, + timeline: &mut Timeline) + -> Result<WorkItemResult, FatalError> +{ + let diag_handler = cgcx.create_diag_handler(); + let config = cgcx.config(work_item.kind()); + let module = match work_item { + WorkItem::Optimize(module) => module, + WorkItem::LTO(mut lto) => { + unsafe { + let module = lto.optimize(cgcx, timeline)?; + let module = codegen(cgcx, &diag_handler, module, config, timeline)?; + return Ok(WorkItemResult::Compiled(module)) + } + } + }; + let module_name = module.name.clone(); + + let pre_existing = match module.source { + ModuleSource::Codegened(_) => None, + ModuleSource::Preexisting(ref wp) => Some(wp.clone()), + }; + + if let Some(wp) = pre_existing { + let incr_comp_session_dir = cgcx.incr_comp_session_dir + .as_ref() + .unwrap(); + let name = &module.name; + let mut object = None; + let mut bytecode = None; + let mut bytecode_compressed = None; + for (kind, saved_file) in wp.saved_files { + let obj_out = match kind { + WorkProductFileKind::Object => { + let path = cgcx.output_filenames.temp_path(OutputType::Object, Some(name)); + object = Some(path.clone()); + path + } + WorkProductFileKind::Bytecode => { + let path = cgcx.output_filenames.temp_path(OutputType::Bitcode, Some(name)); + bytecode = Some(path.clone()); + path + } + WorkProductFileKind::BytecodeCompressed => { + let path = cgcx.output_filenames.temp_path(OutputType::Bitcode, Some(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()); + match link_or_copy(&source_file, &obj_out) { + Ok(_) => { } + Err(err) => { + diag_handler.err(&format!("unable to copy {} to {}: {}", + source_file.display(), + obj_out.display(), + err)); + } + } + } + assert_eq!(object.is_some(), config.emit_obj); + assert_eq!(bytecode.is_some(), config.emit_bc); + assert_eq!(bytecode_compressed.is_some(), config.emit_bc_compressed); + + Ok(WorkItemResult::Compiled(CompiledModule { + llmod_id: module.llmod_id.clone(), + name: module_name, + kind: ModuleKind::Regular, + pre_existing: true, + object, + bytecode, + bytecode_compressed, + })) + } else { + debug!("llvm-optimizing {:?}", module_name); + + unsafe { + optimize(cgcx, &diag_handler, &module, config, timeline)?; + + // 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, + + // 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::Yes | Lto::Fat | Lto::Thin => { + cgcx.crate_types.len() != 1 || + cgcx.crate_types[0] != config::CrateTypeRlib + } + + // 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. + // + // Additionally here's where we also factor in the current LLVM + // version. If it doesn't support ThinLTO we skip this. + Lto::ThinLocal => { + module.kind != ModuleKind::Allocator && + llvm::LLVMRustThinLTOAvailable() + } + }; + + // Metadata modules never participate in LTO regardless of the lto + // settings. + let needs_lto = needs_lto && module.kind != ModuleKind::Metadata; + + // Don't run LTO passes when cross-lang LTO is enabled. The linker + // will do that for us in this case. + let needs_lto = needs_lto && + !cgcx.opts.debugging_opts.cross_lang_lto.embed_bitcode(); + + if needs_lto { + Ok(WorkItemResult::NeedsLTO(module)) + } else { + let module = codegen(cgcx, &diag_handler, module, config, timeline)?; + Ok(WorkItemResult::Compiled(module)) + } + } + } +} + +enum Message { + Token(io::Result<Acquired>), + NeedsLTO { + result: ModuleCodegen, + worker_id: usize, + }, + Done { + result: Result<CompiledModule, ()>, + worker_id: usize, + }, + CodegenDone { + llvm_work_item: WorkItem, + cost: u64, + }, + CodegenComplete, + CodegenItem, +} + +struct Diagnostic { + msg: String, + code: Option<DiagnosticId>, + lvl: Level, +} + +#[derive(PartialEq, Clone, Copy, Debug)] +enum MainThreadWorkerState { + Idle, + Codegenning, + LLVMing, +} + +fn start_executing_work(tcx: TyCtxt, + crate_info: &CrateInfo, + shared_emitter: SharedEmitter, + codegen_worker_send: Sender<Message>, + coordinator_receive: Receiver<Box<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(); + + 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::Yes | 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(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 (name, mut cmd) = get_linker(sess); + cmd.args(&sess.target.target.options.asm_args); + Some(Arc::new(AssemblerCommand { + name, + cmd, + })) + } else { + None + }; + + let cgcx = CodegenContext { + 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()), + 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: 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 started_lto = 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, 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: + while !codegen_done || + work_items.len() > 0 || + running > 0 || + needs_lto.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.kind()), + &mut llvm_start_time); + main_thread_worker_state = MainThreadWorkerState::LLVMing; + spawn_work(cgcx, item); + } + } + } 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() > 0); + started_lto = true; + let modules = mem::replace(&mut needs_lto, Vec::new()); + for (work, cost) in generate_lto_work(&cgcx, 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)); + 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.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 work_items.len() > 0 && running < tokens.len() { + let (item, _) = work_items.pop().unwrap(); + + maybe_start_llvm_timer(cgcx.config(item.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>().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)); + + helper.request_token(); + assert_eq!(main_thread_worker_state, + MainThreadWorkerState::Codegenning); + main_thread_worker_state = MainThreadWorkerState::Idle; + } + + Message::CodegenComplete => { + codegen_done = true; + assert_eq!(main_thread_worker_state, + MainThreadWorkerState::Codegenning); + main_thread_worker_state = MainThreadWorkerState::Idle; + } + + // 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::Done { result: Err(()), worker_id: _ } => { + shared_emitter.fatal("aborting due to worker thread failure"); + // Exit the coordinator thread + return Err(()) + } + 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(cgcx: CodegenContext, work: WorkItem) { + 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 { + coordinator_send: Sender<Box<Any + Send>>, + result: Option<WorkItemResult>, + worker_id: usize, + } + impl Drop for Bomb { + fn drop(&mut self) { + let worker_id = self.worker_id; + let msg = match self.result.take() { + Some(WorkItemResult::Compiled(m)) => { + Message::Done { result: Ok(m), worker_id } + } + Some(WorkItemResult::NeedsLTO(m)) => { + Message::NeedsLTO { result: m, worker_id } + } + None => Message::Done { result: Err(()), worker_id } + }; + drop(self.coordinator_send.send(Box::new(msg))); + } + } + + let mut bomb = Bomb { + 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(cgcx: &CodegenContext, 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(¬e[..]).unwrap()) + .emit(); + handler.abort_if_errors(); + } + }, + Err(e) => { + handler.err(&format!("could not exec the linker `{}`: {}", pname.display(), e)); + handler.abort_if_errors(); + } + } +} + +pub unsafe fn with_llvm_pmb(llmod: ModuleRef, + config: &ModuleConfig, + opt_level: llvm::CodeGenOptLevel, + prepare_for_thin_lto: bool, + f: &mut FnMut(llvm::PassManagerBuilderRef)) { + use std::ptr; + + // Create the PassManagerBuilder for LLVM. We configure it with + // reasonable defaults and prepare it to actually populate the pass + // manager. + let builder = llvm::LLVMPassManagerBuilderCreate(); + let opt_size = config.opt_size.unwrap_or(llvm::CodeGenOptSizeNone); + let inline_threshold = config.inline_threshold; + + let pgo_gen_path = config.pgo_gen.as_ref().map(|s| { + let s = if s.is_empty() { "default_%m.profraw" } else { s }; + CString::new(s.as_bytes()).unwrap() + }); + + let pgo_use_path = if config.pgo_use.is_empty() { + None + } else { + Some(CString::new(config.pgo_use.as_bytes()).unwrap()) + }; + + llvm::LLVMRustConfigurePassManagerBuilder( + builder, + opt_level, + config.merge_functions, + config.vectorize_slp, + config.vectorize_loop, + prepare_for_thin_lto, + pgo_gen_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()), + pgo_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()), + ); + + llvm::LLVMPassManagerBuilderSetSizeLevel(builder, opt_size as u32); + + if opt_size != llvm::CodeGenOptSizeNone { + llvm::LLVMPassManagerBuilderSetDisableUnrollLoops(builder, 1); + } + + llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins); + + // Here we match what clang does (kinda). For O0 we only inline + // always-inline functions (but don't add lifetime intrinsics), at O1 we + // inline with lifetime intrinsics, and O2+ we add an inliner with a + // thresholds copied from clang. + match (opt_level, opt_size, inline_threshold) { + (.., Some(t)) => { + llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, t as u32); + } + (llvm::CodeGenOptLevel::Aggressive, ..) => { + llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 275); + } + (_, llvm::CodeGenOptSizeDefault, _) => { + llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 75); + } + (_, llvm::CodeGenOptSizeAggressive, _) => { + llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 25); + } + (llvm::CodeGenOptLevel::None, ..) => { + llvm::LLVMRustAddAlwaysInlinePass(builder, false); + } + (llvm::CodeGenOptLevel::Less, ..) => { + llvm::LLVMRustAddAlwaysInlinePass(builder, true); + } + (llvm::CodeGenOptLevel::Default, ..) => { + llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 225); + } + (llvm::CodeGenOptLevel::Other, ..) => { + bug!("CodeGenOptLevel::Other selected") + } + } + + f(builder); + llvm::LLVMPassManagerBuilderDispose(builder); +} + + +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 }) + } + + fn inline_asm_error(&self, cookie: u32, msg: String) { + drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg))); + } + + 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 { + crate_name: Symbol, + link: LinkMeta, + metadata: EncodedMetadata, + windows_subsystem: Option<String>, + linker_info: LinkerInfo, + crate_info: CrateInfo, + time_graph: Option<TimeGraph>, + coordinator_send: Sender<Box<Any + Send>>, + codegen_worker_receive: Receiver<Message>, + shared_emitter_main: SharedEmitterMain, + future: thread::JoinHandle<Result<CompiledModules, ()>>, + output_filenames: Arc<OutputFilenames>, +} + +impl OngoingCodegen { + pub(crate) 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(_) => { + sess.fatal("Error during codegen/LLVM phase."); + } + }; + + 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() { + unsafe { llvm::LLVMRustPrintPassTimings(); } + } + + (CodegenResults { + crate_name: self.crate_name, + link: self.link, + 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(crate) fn submit_pre_codegened_module_to_llvm(&self, + tcx: TyCtxt, + module: ModuleCodegen) { + 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(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))); + } + + 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(crate) fn submit_codegened_module_to_llvm(tcx: TyCtxt, + module: ModuleCodegen, + cost: u64) { + let llvm_work_item = WorkItem::Optimize(module); + drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::CodegenDone { + llvm_work_item, + cost, + }))); +} + +fn msvc_imps_needed(tcx: TyCtxt) -> bool { + tcx.sess.target.target.options.is_like_msvc && + tcx.sess.crate_types.borrow().iter().any(|ct| *ct == config::CrateTypeRlib) +} + +// Create a `__imp_<symbol> = &symbol` global for every public static `symbol`. +// This is required to satisfy `dllimport` references to static data in .rlibs +// when using MSVC linker. We do this only for data, as linker can fix up +// code references on its own. +// See #26591, #27438 +fn create_msvc_imps(cgcx: &CodegenContext, llcx: ContextRef, llmod: ModuleRef) { + if !cgcx.msvc_imps_needed { + return + } + // The x86 ABI seems to require that leading underscores are added to symbol + // names, so we need an extra underscore on 32-bit. There's also a leading + // '\x01' here which disables LLVM's symbol mangling (e.g. no extra + // underscores added in front). + let prefix = if cgcx.target_pointer_width == "32" { + "\x01__imp__" + } else { + "\x01__imp_" + }; + unsafe { + let i8p_ty = Type::i8p_llcx(llcx); + let globals = base::iter_globals(llmod) + .filter(|&val| { + llvm::LLVMRustGetLinkage(val) == llvm::Linkage::ExternalLinkage && + llvm::LLVMIsDeclaration(val) == 0 + }) + .map(move |val| { + let name = CStr::from_ptr(llvm::LLVMGetValueName(val)); + let mut imp_name = prefix.as_bytes().to_vec(); + imp_name.extend(name.to_bytes()); + let imp_name = CString::new(imp_name).unwrap(); + (imp_name, val) + }) + .collect::<Vec<_>>(); + for (imp_name, val) in globals { + let imp = llvm::LLVMAddGlobal(llmod, + i8p_ty.to_ref(), + imp_name.as_ptr() as *const _); + llvm::LLVMSetInitializer(imp, consts::ptrcast(val, i8p_ty)); + llvm::LLVMRustSetLinkage(imp, llvm::Linkage::ExternalLinkage); + } + } +} |