1 // Copyright 2013-2015 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
12 use back::bytecode::{self, RLIB_BYTECODE_EXTENSION};
13 use back::lto::{self, ThinBuffer, SerializedModule};
14 use back::link::{self, get_linker, remove};
18 use rustc_incremental::{copy_cgu_workproducts_to_incr_comp_cache_dir,
19 in_incr_comp_dir, in_incr_comp_dir_sess};
20 use rustc::dep_graph::{WorkProduct, WorkProductId, WorkProductFileKind};
21 use rustc::dep_graph::cgu_reuse_tracker::CguReuseTracker;
22 use rustc::middle::cstore::EncodedMetadata;
23 use rustc::session::config::{self, OutputFilenames, OutputType, Passes, Sanitizer, Lto};
24 use rustc::session::Session;
25 use rustc::util::nodemap::FxHashMap;
26 use time_graph::{self, TimeGraph, Timeline};
27 use llvm::{self, DiagnosticInfo, PassManager, SMDiagnostic, BasicBlock};
29 use {CodegenResults, ModuleCodegen, CompiledModule, ModuleKind, // ModuleLlvm,
32 use rustc::hir::def_id::{CrateNum, LOCAL_CRATE};
33 use rustc::ty::TyCtxt;
34 use rustc::util::common::{time_ext, time_depth, set_time_depth, print_time_passes_entry};
35 use rustc_fs_util::{path2cstr, link_or_copy};
36 use rustc_data_structures::small_c_str::SmallCStr;
37 use rustc_data_structures::svh::Svh;
38 use rustc_codegen_utils::command::Command;
39 use rustc_codegen_utils::linker::LinkerInfo;
40 use rustc_codegen_utils::symbol_export::ExportedSymbols;
41 use errors::{self, Handler, Level, DiagnosticBuilder, FatalError, DiagnosticId};
42 use errors::emitter::{Emitter};
44 use syntax::ext::hygiene::Mark;
45 use syntax_pos::MultiSpan;
46 use syntax_pos::symbol::Symbol;
48 use context::{is_pie_binary, get_reloc_model};
49 use interfaces::{Backend, CommonWriteMethods};
51 use jobserver::{Client, Acquired};
54 use std::marker::PhantomData;
57 use std::ffi::{CString, CStr};
59 use std::io::{self, Write};
61 use std::path::{Path, PathBuf};
64 use std::sync::mpsc::{channel, Sender, Receiver};
66 use std::time::Instant;
68 use libc::{c_uint, c_void, c_char, size_t};
70 pub const RELOC_MODEL_ARGS : [(&str, llvm::RelocMode); 7] = [
71 ("pic", llvm::RelocMode::PIC),
72 ("static", llvm::RelocMode::Static),
73 ("default", llvm::RelocMode::Default),
74 ("dynamic-no-pic", llvm::RelocMode::DynamicNoPic),
75 ("ropi", llvm::RelocMode::ROPI),
76 ("rwpi", llvm::RelocMode::RWPI),
77 ("ropi-rwpi", llvm::RelocMode::ROPI_RWPI),
80 pub const CODE_GEN_MODEL_ARGS: &[(&str, llvm::CodeModel)] = &[
81 ("small", llvm::CodeModel::Small),
82 ("kernel", llvm::CodeModel::Kernel),
83 ("medium", llvm::CodeModel::Medium),
84 ("large", llvm::CodeModel::Large),
87 pub const TLS_MODEL_ARGS : [(&str, llvm::ThreadLocalMode); 4] = [
88 ("global-dynamic", llvm::ThreadLocalMode::GeneralDynamic),
89 ("local-dynamic", llvm::ThreadLocalMode::LocalDynamic),
90 ("initial-exec", llvm::ThreadLocalMode::InitialExec),
91 ("local-exec", llvm::ThreadLocalMode::LocalExec),
94 const PRE_THIN_LTO_BC_EXT: &str = "pre-thin-lto.bc";
96 pub fn llvm_err(handler: &errors::Handler, msg: &str) -> FatalError {
97 match llvm::last_error() {
98 Some(err) => handler.fatal(&format!("{}: {}", msg, err)),
99 None => handler.fatal(&msg),
103 pub fn write_output_file(
104 handler: &errors::Handler,
105 target: &'ll llvm::TargetMachine,
106 pm: &llvm::PassManager<'ll>,
107 m: &'ll llvm::Module,
109 file_type: llvm::FileType) -> Result<(), FatalError> {
111 let output_c = path2cstr(output);
112 let result = llvm::LLVMRustWriteOutputFile(target, pm, m, output_c.as_ptr(), file_type);
113 if result.into_result().is_err() {
114 let msg = format!("could not write output to {}", output.display());
115 Err(llvm_err(handler, &msg))
122 fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel {
124 config::OptLevel::No => llvm::CodeGenOptLevel::None,
125 config::OptLevel::Less => llvm::CodeGenOptLevel::Less,
126 config::OptLevel::Default => llvm::CodeGenOptLevel::Default,
127 config::OptLevel::Aggressive => llvm::CodeGenOptLevel::Aggressive,
128 _ => llvm::CodeGenOptLevel::Default,
132 fn get_llvm_opt_size(optimize: config::OptLevel) -> llvm::CodeGenOptSize {
134 config::OptLevel::Size => llvm::CodeGenOptSizeDefault,
135 config::OptLevel::SizeMin => llvm::CodeGenOptSizeAggressive,
136 _ => llvm::CodeGenOptSizeNone,
140 pub fn create_target_machine(
143 ) -> &'static mut llvm::TargetMachine {
144 target_machine_factory(sess, find_features)().unwrap_or_else(|err| {
145 llvm_err(sess.diagnostic(), &err).raise()
149 // If find_features is true this won't access `sess.crate_types` by assuming
150 // that `is_pie_binary` is false. When we discover LLVM target features
151 // `sess.crate_types` is uninitialized so we cannot access it.
152 pub fn target_machine_factory(sess: &Session, find_features: bool)
153 -> Arc<dyn Fn() -> Result<&'static mut llvm::TargetMachine, String> + Send + Sync>
155 let reloc_model = get_reloc_model(sess);
157 let opt_level = get_llvm_opt_level(sess.opts.optimize);
158 let use_softfp = sess.opts.cg.soft_float;
160 let ffunction_sections = sess.target.target.options.function_sections;
161 let fdata_sections = ffunction_sections;
163 let code_model_arg = sess.opts.cg.code_model.as_ref().or(
164 sess.target.target.options.code_model.as_ref(),
167 let code_model = match code_model_arg {
169 match CODE_GEN_MODEL_ARGS.iter().find(|arg| arg.0 == s) {
172 sess.err(&format!("{:?} is not a valid code model",
174 sess.abort_if_errors();
179 None => llvm::CodeModel::None,
182 let features = attributes::llvm_target_features(sess).collect::<Vec<_>>();
183 let mut singlethread = sess.target.target.options.singlethread;
185 // On the wasm target once the `atomics` feature is enabled that means that
186 // we're no longer single-threaded, or otherwise we don't want LLVM to
187 // lower atomic operations to single-threaded operations.
189 sess.target.target.llvm_target.contains("wasm32") &&
190 features.iter().any(|s| *s == "+atomics")
192 singlethread = false;
195 let triple = SmallCStr::new(&sess.target.target.llvm_target);
196 let cpu = SmallCStr::new(llvm_util::target_cpu(sess));
197 let features = features.join(",");
198 let features = CString::new(features).unwrap();
199 let is_pie_binary = !find_features && is_pie_binary(sess);
200 let trap_unreachable = sess.target.target.options.trap_unreachable;
201 let emit_stack_size_section = sess.opts.debugging_opts.emit_stack_sizes;
203 let asm_comments = sess.asm_comments();
207 llvm::LLVMRustCreateTargetMachine(
208 triple.as_ptr(), cpu.as_ptr(), features.as_ptr(),
219 emit_stack_size_section,
224 format!("Could not create LLVM TargetMachine for triple: {}",
225 triple.to_str().unwrap())
230 /// Module-specific configuration for `optimize_and_codegen`.
231 pub struct ModuleConfig {
232 /// Names of additional optimization passes to run.
234 /// Some(level) to optimize at a certain level, or None to run
235 /// absolutely no optimizations (used for the metadata module).
236 pub opt_level: Option<llvm::CodeGenOptLevel>,
238 /// Some(level) to optimize binary size, or None to not affect program size.
239 opt_size: Option<llvm::CodeGenOptSize>,
241 pgo_gen: Option<String>,
244 // Flags indicating which outputs to produce.
245 pub emit_pre_thin_lto_bc: bool,
246 emit_no_opt_bc: bool,
248 emit_bc_compressed: bool,
253 // Miscellaneous flags. These are mostly copied from command-line
255 pub verify_llvm_ir: bool,
256 no_prepopulate_passes: bool,
259 vectorize_loop: bool,
261 merge_functions: bool,
262 inline_threshold: Option<usize>,
263 // Instead of creating an object file by doing LLVM codegen, just
264 // make the object file bitcode. Provides easy compatibility with
265 // emscripten's ecc compiler, when used as the linker.
266 obj_is_bitcode: bool,
267 no_integrated_as: bool,
269 embed_bitcode_marker: bool,
273 fn new(passes: Vec<String>) -> ModuleConfig {
280 pgo_use: String::new(),
282 emit_no_opt_bc: false,
283 emit_pre_thin_lto_bc: false,
285 emit_bc_compressed: false,
290 obj_is_bitcode: false,
291 embed_bitcode: false,
292 embed_bitcode_marker: false,
293 no_integrated_as: false,
295 verify_llvm_ir: false,
296 no_prepopulate_passes: false,
299 vectorize_loop: false,
300 vectorize_slp: false,
301 merge_functions: false,
302 inline_threshold: None
306 fn set_flags(&mut self, sess: &Session, no_builtins: bool) {
307 self.verify_llvm_ir = sess.verify_llvm_ir();
308 self.no_prepopulate_passes = sess.opts.cg.no_prepopulate_passes;
309 self.no_builtins = no_builtins || sess.target.target.options.no_builtins;
310 self.time_passes = sess.time_passes();
311 self.inline_threshold = sess.opts.cg.inline_threshold;
312 self.obj_is_bitcode = sess.target.target.options.obj_is_bitcode ||
313 sess.opts.debugging_opts.cross_lang_lto.enabled();
314 let embed_bitcode = sess.target.target.options.embed_bitcode ||
315 sess.opts.debugging_opts.embed_bitcode;
317 match sess.opts.optimize {
318 config::OptLevel::No |
319 config::OptLevel::Less => {
320 self.embed_bitcode_marker = embed_bitcode;
322 _ => self.embed_bitcode = embed_bitcode,
326 // Copy what clang does by turning on loop vectorization at O2 and
327 // slp vectorization at O3. Otherwise configure other optimization aspects
328 // of this pass manager builder.
329 // Turn off vectorization for emscripten, as it's not very well supported.
330 self.vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
331 (sess.opts.optimize == config::OptLevel::Default ||
332 sess.opts.optimize == config::OptLevel::Aggressive) &&
333 !sess.target.target.options.is_like_emscripten;
335 self.vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
336 sess.opts.optimize == config::OptLevel::Aggressive &&
337 !sess.target.target.options.is_like_emscripten;
339 self.merge_functions = sess.opts.optimize == config::OptLevel::Default ||
340 sess.opts.optimize == config::OptLevel::Aggressive;
343 pub fn bitcode_needed(&self) -> bool {
344 self.emit_bc || self.obj_is_bitcode
345 || self.emit_bc_compressed || self.embed_bitcode
349 /// Assembler name and command used by codegen when no_integrated_as is enabled
350 struct AssemblerCommand {
355 /// Additional resources used by optimize_and_codegen (not module specific)
357 pub struct CodegenContext<'ll> {
358 // Resources needed when running LTO
359 pub time_passes: bool,
361 pub no_landing_pads: bool,
362 pub save_temps: bool,
363 pub fewer_names: bool,
364 pub exported_symbols: Option<Arc<ExportedSymbols>>,
365 pub opts: Arc<config::Options>,
366 pub crate_types: Vec<config::CrateType>,
367 pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>,
368 output_filenames: Arc<OutputFilenames>,
369 regular_module_config: Arc<ModuleConfig>,
370 metadata_module_config: Arc<ModuleConfig>,
371 allocator_module_config: Arc<ModuleConfig>,
372 pub tm_factory: Arc<dyn Fn() -> Result<&'static mut llvm::TargetMachine, String> + Send + Sync>,
373 pub msvc_imps_needed: bool,
374 pub target_pointer_width: String,
375 debuginfo: config::DebugInfo,
377 // Number of cgus excluding the allocator/metadata modules
378 pub total_cgus: usize,
379 // Handler to use for diagnostics produced during codegen.
380 pub diag_emitter: SharedEmitter,
381 // LLVM passes added by plugins.
382 pub plugin_passes: Vec<String>,
383 // LLVM optimizations for which we want to print remarks.
385 // Worker thread number
387 // The incremental compilation session directory, or None if we are not
388 // compiling incrementally
389 pub incr_comp_session_dir: Option<PathBuf>,
390 // Used to update CGU re-use information during the thinlto phase.
391 pub cgu_reuse_tracker: CguReuseTracker,
392 // Channel back to the main control thread to send messages to
393 coordinator_send: Sender<Box<dyn Any + Send>>,
394 // A reference to the TimeGraph so we can register timings. None means that
395 // measuring is disabled.
396 time_graph: Option<TimeGraph>,
397 // The assembler command if no_integrated_as option is enabled, None otherwise
398 assembler_cmd: Option<Arc<AssemblerCommand>>,
399 // This field is used to give a lifetime parameter to the struct so that it can implement
400 // the Backend trait.
401 phantom: PhantomData<&'ll ()>
404 impl CodegenContext<'ll> {
405 pub fn create_diag_handler(&self) -> Handler {
406 Handler::with_emitter(true, false, Box::new(self.diag_emitter.clone()))
409 pub(crate) fn config(&self, kind: ModuleKind) -> &ModuleConfig {
411 ModuleKind::Regular => &self.regular_module_config,
412 ModuleKind::Metadata => &self.metadata_module_config,
413 ModuleKind::Allocator => &self.allocator_module_config,
417 pub(crate) fn save_temp_bitcode(&self, module: &ModuleCodegen, name: &str) {
418 if !self.save_temps {
422 let ext = format!("{}.bc", name);
423 let cgu = Some(&module.name[..]);
424 let path = self.output_filenames.temp_path_ext(&ext, cgu);
425 let cstr = path2cstr(&path);
426 let llmod = module.module_llvm.llmod();
427 llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr());
432 impl<'ll> Backend for CodegenContext<'ll> {
433 type Value = &'ll Value;
434 type BasicBlock = &'ll BasicBlock;
435 type Type = &'ll Type;
436 type Context = &'ll llvm::Context;
439 impl CommonWriteMethods for CodegenContext<'ll> {
440 fn val_ty(&self, v: &'ll Value) -> &'ll Type {
444 fn c_bytes_in_context(&self, llcx: &'ll llvm::Context, bytes: &[u8]) -> &'ll Value {
445 common::c_bytes_in_context(llcx, bytes)
448 fn c_struct_in_context(
450 llcx: &'a llvm::Context,
454 common::c_struct_in_context(llcx, elts, packed)
459 pub struct DiagnosticHandlers<'a> {
460 data: *mut (&'a CodegenContext<'a>, &'a Handler),
461 llcx: &'a llvm::Context,
464 impl<'a> DiagnosticHandlers<'a> {
465 pub fn new(cgcx: &'a CodegenContext<'a>,
466 handler: &'a Handler,
467 llcx: &'a llvm::Context) -> Self {
468 let data = Box::into_raw(Box::new((cgcx, handler)));
470 llvm::LLVMRustSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, data as *mut _);
471 llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, data as *mut _);
473 DiagnosticHandlers { data, llcx }
477 impl<'a> Drop for DiagnosticHandlers<'a> {
479 use std::ptr::null_mut;
481 llvm::LLVMRustSetInlineAsmDiagnosticHandler(self.llcx, inline_asm_handler, null_mut());
482 llvm::LLVMContextSetDiagnosticHandler(self.llcx, diagnostic_handler, null_mut());
483 drop(Box::from_raw(self.data));
488 unsafe extern "C" fn report_inline_asm<'a, 'b>(cgcx: &'a CodegenContext,
491 cgcx.diag_emitter.inline_asm_error(cookie as u32, msg.to_owned());
494 unsafe extern "C" fn inline_asm_handler(diag: &SMDiagnostic,
500 let (cgcx, _) = *(user as *const (&CodegenContext, &Handler));
502 let msg = llvm::build_string(|s| llvm::LLVMRustWriteSMDiagnosticToString(diag, s))
503 .expect("non-UTF8 SMDiagnostic");
505 report_inline_asm(cgcx, &msg, cookie);
508 unsafe extern "C" fn diagnostic_handler(info: &DiagnosticInfo, user: *mut c_void) {
512 let (cgcx, diag_handler) = *(user as *const (&CodegenContext, &Handler));
514 match llvm::diagnostic::Diagnostic::unpack(info) {
515 llvm::diagnostic::InlineAsm(inline) => {
516 report_inline_asm(cgcx,
517 &llvm::twine_to_string(inline.message),
521 llvm::diagnostic::Optimization(opt) => {
522 let enabled = match cgcx.remark {
524 Passes::Some(ref v) => v.iter().any(|s| *s == opt.pass_name),
528 diag_handler.note_without_error(&format!("optimization {} for {} at {}:{}:{}: {}",
537 llvm::diagnostic::PGO(diagnostic_ref) |
538 llvm::diagnostic::Linker(diagnostic_ref) => {
539 let msg = llvm::build_string(|s| {
540 llvm::LLVMRustWriteDiagnosticInfoToString(diagnostic_ref, s)
541 }).expect("non-UTF8 diagnostic");
542 diag_handler.warn(&msg);
544 llvm::diagnostic::UnknownDiagnostic(..) => {},
548 // Unsafe due to LLVM calls.
549 unsafe fn optimize(cgcx: &CodegenContext,
550 diag_handler: &Handler,
551 module: &ModuleCodegen,
552 config: &ModuleConfig,
553 timeline: &mut Timeline)
554 -> Result<(), FatalError>
556 let llmod = module.module_llvm.llmod();
557 let llcx = &*module.module_llvm.llcx;
558 let tm = &*module.module_llvm.tm;
559 let _handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
561 let module_name = module.name.clone();
562 let module_name = Some(&module_name[..]);
564 if config.emit_no_opt_bc {
565 let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name);
566 let out = path2cstr(&out);
567 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
570 if config.opt_level.is_some() {
571 // Create the two optimizing pass managers. These mirror what clang
572 // does, and are by populated by LLVM's default PassManagerBuilder.
573 // Each manager has a different set of passes, but they also share
574 // some common passes.
575 let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
576 let mpm = llvm::LLVMCreatePassManager();
579 // If we're verifying or linting, add them to the function pass
581 let addpass = |pass_name: &str| {
582 let pass_name = SmallCStr::new(pass_name);
583 let pass = match llvm::LLVMRustFindAndCreatePass(pass_name.as_ptr()) {
585 None => return false,
587 let pass_manager = match llvm::LLVMRustPassKind(pass) {
588 llvm::PassKind::Function => &*fpm,
589 llvm::PassKind::Module => &*mpm,
590 llvm::PassKind::Other => {
591 diag_handler.err("Encountered LLVM pass kind we can't handle");
595 llvm::LLVMRustAddPass(pass_manager, pass);
599 if config.verify_llvm_ir { assert!(addpass("verify")); }
601 // Some options cause LLVM bitcode to be emitted, which uses ThinLTOBuffers, so we need
602 // to make sure we run LLVM's NameAnonGlobals pass when emitting bitcode; otherwise
603 // we'll get errors in LLVM.
604 let using_thin_buffers = config.bitcode_needed();
605 let mut have_name_anon_globals_pass = false;
606 if !config.no_prepopulate_passes {
607 llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod);
608 llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod);
609 let opt_level = config.opt_level.unwrap_or(llvm::CodeGenOptLevel::None);
610 let prepare_for_thin_lto = cgcx.lto == Lto::Thin || cgcx.lto == Lto::ThinLocal ||
611 (cgcx.lto != Lto::Fat && cgcx.opts.debugging_opts.cross_lang_lto.enabled());
612 have_name_anon_globals_pass = have_name_anon_globals_pass || prepare_for_thin_lto;
613 if using_thin_buffers && !prepare_for_thin_lto {
614 assert!(addpass("name-anon-globals"));
615 have_name_anon_globals_pass = true;
617 with_llvm_pmb(llmod, &config, opt_level, prepare_for_thin_lto, &mut |b| {
618 llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(b, fpm);
619 llvm::LLVMPassManagerBuilderPopulateModulePassManager(b, mpm);
623 for pass in &config.passes {
625 diag_handler.warn(&format!("unknown pass `{}`, ignoring", pass));
627 if pass == "name-anon-globals" {
628 have_name_anon_globals_pass = true;
632 for pass in &cgcx.plugin_passes {
634 diag_handler.err(&format!("a plugin asked for LLVM pass \
635 `{}` but LLVM does not \
636 recognize it", pass));
638 if pass == "name-anon-globals" {
639 have_name_anon_globals_pass = true;
643 if using_thin_buffers && !have_name_anon_globals_pass {
644 // As described above, this will probably cause an error in LLVM
645 if config.no_prepopulate_passes {
646 diag_handler.err("The current compilation is going to use thin LTO buffers \
647 without running LLVM's NameAnonGlobals pass. \
648 This will likely cause errors in LLVM. Consider adding \
649 -C passes=name-anon-globals to the compiler command line.");
651 bug!("We are using thin LTO buffers without running the NameAnonGlobals pass. \
652 This will likely cause errors in LLVM and should never happen.");
657 diag_handler.abort_if_errors();
659 // Finally, run the actual optimization passes
660 time_ext(config.time_passes,
662 &format!("llvm function passes [{}]", module_name.unwrap()),
664 llvm::LLVMRustRunFunctionPassManager(fpm, llmod)
666 timeline.record("fpm");
667 time_ext(config.time_passes,
669 &format!("llvm module passes [{}]", module_name.unwrap()),
671 llvm::LLVMRunPassManager(mpm, llmod)
674 // Deallocate managers that we're now done with
675 llvm::LLVMDisposePassManager(fpm);
676 llvm::LLVMDisposePassManager(mpm);
681 fn generate_lto_work(cgcx: &CodegenContext,
682 modules: Vec<ModuleCodegen>,
683 import_only_modules: Vec<(SerializedModule, WorkProduct)>)
684 -> Vec<(WorkItem, u64)>
686 let mut timeline = cgcx.time_graph.as_ref().map(|tg| {
687 tg.start(CODEGEN_WORKER_TIMELINE,
688 CODEGEN_WORK_PACKAGE_KIND,
690 }).unwrap_or(Timeline::noop());
691 let (lto_modules, copy_jobs) = lto::run(cgcx, modules, import_only_modules, &mut timeline)
692 .unwrap_or_else(|e| e.raise());
694 let lto_modules = lto_modules.into_iter().map(|module| {
695 let cost = module.cost();
696 (WorkItem::LTO(module), cost)
699 let copy_jobs = copy_jobs.into_iter().map(|wp| {
700 (WorkItem::CopyPostLtoArtifacts(CachedModuleCodegen {
701 name: wp.cgu_name.clone(),
706 lto_modules.chain(copy_jobs).collect()
709 unsafe fn codegen(cgcx: &CodegenContext,
710 diag_handler: &Handler,
711 module: ModuleCodegen,
712 config: &ModuleConfig,
713 timeline: &mut Timeline)
714 -> Result<CompiledModule, FatalError>
716 timeline.record("codegen");
718 let llmod = module.module_llvm.llmod();
719 let llcx = &*module.module_llvm.llcx;
720 let tm = &*module.module_llvm.tm;
721 let module_name = module.name.clone();
722 let module_name = Some(&module_name[..]);
723 let handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
725 if cgcx.msvc_imps_needed {
726 create_msvc_imps(cgcx, llcx, llmod);
729 // A codegen-specific pass manager is used to generate object
730 // files for an LLVM module.
732 // Apparently each of these pass managers is a one-shot kind of
733 // thing, so we create a new one for each type of output. The
734 // pass manager passed to the closure should be ensured to not
735 // escape the closure itself, and the manager should only be
737 unsafe fn with_codegen<'ll, F, R>(tm: &'ll llvm::TargetMachine,
738 llmod: &'ll llvm::Module,
741 where F: FnOnce(&'ll mut PassManager<'ll>) -> R,
743 let cpm = llvm::LLVMCreatePassManager();
744 llvm::LLVMRustAddAnalysisPasses(tm, cpm, llmod);
745 llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
749 // If we don't have the integrated assembler, then we need to emit asm
750 // from LLVM and use `gcc` to create the object file.
751 let asm_to_obj = config.emit_obj && config.no_integrated_as;
753 // Change what we write and cleanup based on whether obj files are
754 // just llvm bitcode. In that case write bitcode, and possibly
755 // delete the bitcode if it wasn't requested. Don't generate the
756 // machine code, instead copy the .o file from the .bc
757 let write_bc = config.emit_bc || config.obj_is_bitcode;
758 let rm_bc = !config.emit_bc && config.obj_is_bitcode;
759 let write_obj = config.emit_obj && !config.obj_is_bitcode && !asm_to_obj;
760 let copy_bc_to_obj = config.emit_obj && config.obj_is_bitcode;
762 let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name);
763 let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name);
766 if write_bc || config.emit_bc_compressed || config.embed_bitcode {
767 let thin = ThinBuffer::new(llmod);
768 let data = thin.data();
769 timeline.record("make-bc");
772 if let Err(e) = fs::write(&bc_out, data) {
773 diag_handler.err(&format!("failed to write bytecode: {}", e));
775 timeline.record("write-bc");
778 if config.embed_bitcode {
779 embed_bitcode(cgcx, llcx, llmod, Some(data));
780 timeline.record("embed-bc");
783 if config.emit_bc_compressed {
784 let dst = bc_out.with_extension(RLIB_BYTECODE_EXTENSION);
785 let data = bytecode::encode(&module.name, data);
786 if let Err(e) = fs::write(&dst, data) {
787 diag_handler.err(&format!("failed to write bytecode: {}", e));
789 timeline.record("compress-bc");
791 } else if config.embed_bitcode_marker {
792 embed_bitcode(cgcx, llcx, llmod, None);
795 time_ext(config.time_passes, None, &format!("codegen passes [{}]", module_name.unwrap()),
796 || -> Result<(), FatalError> {
798 let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name);
799 let out = path2cstr(&out);
801 extern "C" fn demangle_callback(input_ptr: *const c_char,
803 output_ptr: *mut c_char,
804 output_len: size_t) -> size_t {
806 slice::from_raw_parts(input_ptr as *const u8, input_len as usize)
809 let input = match str::from_utf8(input) {
814 let output = unsafe {
815 slice::from_raw_parts_mut(output_ptr as *mut u8, output_len as usize)
817 let mut cursor = io::Cursor::new(output);
819 let demangled = match rustc_demangle::try_demangle(input) {
824 if let Err(_) = write!(cursor, "{:#}", demangled) {
825 // Possible only if provided buffer is not big enough
829 cursor.position() as size_t
832 with_codegen(tm, llmod, config.no_builtins, |cpm| {
833 llvm::LLVMRustPrintModule(cpm, llmod, out.as_ptr(), demangle_callback);
834 llvm::LLVMDisposePassManager(cpm);
836 timeline.record("ir");
839 if config.emit_asm || asm_to_obj {
840 let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
842 // We can't use the same module for asm and binary output, because that triggers
843 // various errors like invalid IR or broken binaries, so we might have to clone the
844 // module to produce the asm output
845 let llmod = if config.emit_obj {
846 llvm::LLVMCloneModule(llmod)
850 with_codegen(tm, llmod, config.no_builtins, |cpm| {
851 write_output_file(diag_handler, tm, cpm, llmod, &path,
852 llvm::FileType::AssemblyFile)
854 timeline.record("asm");
858 with_codegen(tm, llmod, config.no_builtins, |cpm| {
859 write_output_file(diag_handler, tm, cpm, llmod, &obj_out,
860 llvm::FileType::ObjectFile)
862 timeline.record("obj");
863 } else if asm_to_obj {
864 let assembly = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name);
865 run_assembler(cgcx, diag_handler, &assembly, &obj_out);
866 timeline.record("asm_to_obj");
868 if !config.emit_asm && !cgcx.save_temps {
869 drop(fs::remove_file(&assembly));
877 debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out);
878 if let Err(e) = link_or_copy(&bc_out, &obj_out) {
879 diag_handler.err(&format!("failed to copy bitcode to object file: {}", e));
884 debug!("removing_bitcode {:?}", bc_out);
885 if let Err(e) = fs::remove_file(&bc_out) {
886 diag_handler.err(&format!("failed to remove bitcode: {}", e));
892 Ok(module.into_compiled_module(config.emit_obj,
894 config.emit_bc_compressed,
895 &cgcx.output_filenames))
898 /// Embed the bitcode of an LLVM module in the LLVM module itself.
900 /// This is done primarily for iOS where it appears to be standard to compile C
901 /// code at least with `-fembed-bitcode` which creates two sections in the
904 /// * __LLVM,__bitcode
905 /// * __LLVM,__cmdline
907 /// It appears *both* of these sections are necessary to get the linker to
908 /// recognize what's going on. For us though we just always throw in an empty
911 /// Furthermore debug/O1 builds don't actually embed bitcode but rather just
912 /// embed an empty section.
914 /// Basically all of this is us attempting to follow in the footsteps of clang
915 /// on iOS. See #35968 for lots more info.
916 unsafe fn embed_bitcode(cgcx: &CodegenContext,
917 llcx: &llvm::Context,
918 llmod: &llvm::Module,
919 bitcode: Option<&[u8]>) {
920 let llconst = cgcx.c_bytes_in_context(llcx, bitcode.unwrap_or(&[]));
921 let llglobal = llvm::LLVMAddGlobal(
923 cgcx.val_ty(llconst),
924 "rustc.embedded.module\0".as_ptr() as *const _,
926 llvm::LLVMSetInitializer(llglobal, llconst);
928 let is_apple = cgcx.opts.target_triple.triple().contains("-ios") ||
929 cgcx.opts.target_triple.triple().contains("-darwin");
931 let section = if is_apple {
936 llvm::LLVMSetSection(llglobal, section.as_ptr() as *const _);
937 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
938 llvm::LLVMSetGlobalConstant(llglobal, llvm::True);
940 let llconst = cgcx.c_bytes_in_context(llcx, &[]);
941 let llglobal = llvm::LLVMAddGlobal(
943 cgcx.val_ty(llconst),
944 "rustc.embedded.cmdline\0".as_ptr() as *const _,
946 llvm::LLVMSetInitializer(llglobal, llconst);
947 let section = if is_apple {
952 llvm::LLVMSetSection(llglobal, section.as_ptr() as *const _);
953 llvm::LLVMRustSetLinkage(llglobal, llvm::Linkage::PrivateLinkage);
956 pub(crate) struct CompiledModules {
957 pub modules: Vec<CompiledModule>,
958 pub metadata_module: CompiledModule,
959 pub allocator_module: Option<CompiledModule>,
962 fn need_crate_bitcode_for_rlib(sess: &Session) -> bool {
963 sess.crate_types.borrow().contains(&config::CrateType::Rlib) &&
964 sess.opts.output_types.contains_key(&OutputType::Exe)
967 fn need_pre_thin_lto_bitcode_for_incr_comp(sess: &Session) -> bool {
968 if sess.opts.incremental.is_none() {
976 Lto::ThinLocal => true,
980 pub fn start_async_codegen(tcx: TyCtxt,
981 time_graph: Option<TimeGraph>,
982 metadata: EncodedMetadata,
983 coordinator_receive: Receiver<Box<dyn Any + Send>>,
987 let crate_name = tcx.crate_name(LOCAL_CRATE);
988 let crate_hash = tcx.crate_hash(LOCAL_CRATE);
989 let no_builtins = attr::contains_name(&tcx.hir.krate().attrs, "no_builtins");
990 let subsystem = attr::first_attr_value_str_by_name(&tcx.hir.krate().attrs,
991 "windows_subsystem");
992 let windows_subsystem = subsystem.map(|subsystem| {
993 if subsystem != "windows" && subsystem != "console" {
994 tcx.sess.fatal(&format!("invalid windows subsystem `{}`, only \
995 `windows` and `console` are allowed",
998 subsystem.to_string()
1001 let linker_info = LinkerInfo::new(tcx);
1002 let crate_info = CrateInfo::new(tcx);
1004 // Figure out what we actually need to build.
1005 let mut modules_config = ModuleConfig::new(sess.opts.cg.passes.clone());
1006 let mut metadata_config = ModuleConfig::new(vec![]);
1007 let mut allocator_config = ModuleConfig::new(vec![]);
1009 if let Some(ref sanitizer) = sess.opts.debugging_opts.sanitizer {
1011 Sanitizer::Address => {
1012 modules_config.passes.push("asan".to_owned());
1013 modules_config.passes.push("asan-module".to_owned());
1015 Sanitizer::Memory => {
1016 modules_config.passes.push("msan".to_owned())
1018 Sanitizer::Thread => {
1019 modules_config.passes.push("tsan".to_owned())
1025 if sess.opts.debugging_opts.profile {
1026 modules_config.passes.push("insert-gcov-profiling".to_owned())
1029 modules_config.pgo_gen = sess.opts.debugging_opts.pgo_gen.clone();
1030 modules_config.pgo_use = sess.opts.debugging_opts.pgo_use.clone();
1032 modules_config.opt_level = Some(get_llvm_opt_level(sess.opts.optimize));
1033 modules_config.opt_size = Some(get_llvm_opt_size(sess.opts.optimize));
1035 // Save all versions of the bytecode if we're saving our temporaries.
1036 if sess.opts.cg.save_temps {
1037 modules_config.emit_no_opt_bc = true;
1038 modules_config.emit_pre_thin_lto_bc = true;
1039 modules_config.emit_bc = true;
1040 modules_config.emit_lto_bc = true;
1041 metadata_config.emit_bc = true;
1042 allocator_config.emit_bc = true;
1045 // Emit compressed bitcode files for the crate if we're emitting an rlib.
1046 // Whenever an rlib is created, the bitcode is inserted into the archive in
1047 // order to allow LTO against it.
1048 if need_crate_bitcode_for_rlib(sess) {
1049 modules_config.emit_bc_compressed = true;
1050 allocator_config.emit_bc_compressed = true;
1053 modules_config.emit_pre_thin_lto_bc =
1054 need_pre_thin_lto_bitcode_for_incr_comp(sess);
1056 modules_config.no_integrated_as = tcx.sess.opts.cg.no_integrated_as ||
1057 tcx.sess.target.target.options.no_integrated_as;
1059 for output_type in sess.opts.output_types.keys() {
1060 match *output_type {
1061 OutputType::Bitcode => { modules_config.emit_bc = true; }
1062 OutputType::LlvmAssembly => { modules_config.emit_ir = true; }
1063 OutputType::Assembly => {
1064 modules_config.emit_asm = true;
1065 // If we're not using the LLVM assembler, this function
1066 // could be invoked specially with output_type_assembly, so
1067 // in this case we still want the metadata object file.
1068 if !sess.opts.output_types.contains_key(&OutputType::Assembly) {
1069 metadata_config.emit_obj = true;
1070 allocator_config.emit_obj = true;
1073 OutputType::Object => { modules_config.emit_obj = true; }
1074 OutputType::Metadata => { metadata_config.emit_obj = true; }
1075 OutputType::Exe => {
1076 modules_config.emit_obj = true;
1077 metadata_config.emit_obj = true;
1078 allocator_config.emit_obj = true;
1080 OutputType::Mir => {}
1081 OutputType::DepInfo => {}
1085 modules_config.set_flags(sess, no_builtins);
1086 metadata_config.set_flags(sess, no_builtins);
1087 allocator_config.set_flags(sess, no_builtins);
1089 // Exclude metadata and allocator modules from time_passes output, since
1090 // they throw off the "LLVM passes" measurement.
1091 metadata_config.time_passes = false;
1092 allocator_config.time_passes = false;
1094 let (shared_emitter, shared_emitter_main) = SharedEmitter::new();
1095 let (codegen_worker_send, codegen_worker_receive) = channel();
1097 let coordinator_thread = start_executing_work(tcx,
1100 codegen_worker_send,
1101 coordinator_receive,
1103 sess.jobserver.clone(),
1105 Arc::new(modules_config),
1106 Arc::new(metadata_config),
1107 Arc::new(allocator_config));
1118 coordinator_send: tcx.tx_to_llvm_workers.lock().clone(),
1119 codegen_worker_receive,
1120 shared_emitter_main,
1121 future: coordinator_thread,
1122 output_filenames: tcx.output_filenames(LOCAL_CRATE),
1126 fn copy_all_cgu_workproducts_to_incr_comp_cache_dir(
1128 compiled_modules: &CompiledModules,
1129 ) -> FxHashMap<WorkProductId, WorkProduct> {
1130 let mut work_products = FxHashMap::default();
1132 if sess.opts.incremental.is_none() {
1133 return work_products;
1136 for module in compiled_modules.modules.iter().filter(|m| m.kind == ModuleKind::Regular) {
1137 let mut files = vec![];
1139 if let Some(ref path) = module.object {
1140 files.push((WorkProductFileKind::Object, path.clone()));
1142 if let Some(ref path) = module.bytecode {
1143 files.push((WorkProductFileKind::Bytecode, path.clone()));
1145 if let Some(ref path) = module.bytecode_compressed {
1146 files.push((WorkProductFileKind::BytecodeCompressed, path.clone()));
1149 if let Some((id, product)) =
1150 copy_cgu_workproducts_to_incr_comp_cache_dir(sess, &module.name, &files)
1152 work_products.insert(id, product);
1159 fn produce_final_output_artifacts(sess: &Session,
1160 compiled_modules: &CompiledModules,
1161 crate_output: &OutputFilenames) {
1162 let mut user_wants_bitcode = false;
1163 let mut user_wants_objects = false;
1165 // Produce final compile outputs.
1166 let copy_gracefully = |from: &Path, to: &Path| {
1167 if let Err(e) = fs::copy(from, to) {
1168 sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
1172 let copy_if_one_unit = |output_type: OutputType,
1173 keep_numbered: bool| {
1174 if compiled_modules.modules.len() == 1 {
1175 // 1) Only one codegen unit. In this case it's no difficulty
1176 // to copy `foo.0.x` to `foo.x`.
1177 let module_name = Some(&compiled_modules.modules[0].name[..]);
1178 let path = crate_output.temp_path(output_type, module_name);
1179 copy_gracefully(&path,
1180 &crate_output.path(output_type));
1181 if !sess.opts.cg.save_temps && !keep_numbered {
1182 // The user just wants `foo.x`, not `foo.#module-name#.x`.
1183 remove(sess, &path);
1186 let ext = crate_output.temp_path(output_type, None)
1193 if crate_output.outputs.contains_key(&output_type) {
1194 // 2) Multiple codegen units, with `--emit foo=some_name`. We have
1195 // no good solution for this case, so warn the user.
1196 sess.warn(&format!("ignoring emit path because multiple .{} files \
1197 were produced", ext));
1198 } else if crate_output.single_output_file.is_some() {
1199 // 3) Multiple codegen units, with `-o some_name`. We have
1200 // no good solution for this case, so warn the user.
1201 sess.warn(&format!("ignoring -o because multiple .{} files \
1202 were produced", ext));
1204 // 4) Multiple codegen units, but no explicit name. We
1205 // just leave the `foo.0.x` files in place.
1206 // (We don't have to do any work in this case.)
1211 // Flag to indicate whether the user explicitly requested bitcode.
1212 // Otherwise, we produced it only as a temporary output, and will need
1213 // to get rid of it.
1214 for output_type in crate_output.outputs.keys() {
1215 match *output_type {
1216 OutputType::Bitcode => {
1217 user_wants_bitcode = true;
1218 // Copy to .bc, but always keep the .0.bc. There is a later
1219 // check to figure out if we should delete .0.bc files, or keep
1220 // them for making an rlib.
1221 copy_if_one_unit(OutputType::Bitcode, true);
1223 OutputType::LlvmAssembly => {
1224 copy_if_one_unit(OutputType::LlvmAssembly, false);
1226 OutputType::Assembly => {
1227 copy_if_one_unit(OutputType::Assembly, false);
1229 OutputType::Object => {
1230 user_wants_objects = true;
1231 copy_if_one_unit(OutputType::Object, true);
1234 OutputType::Metadata |
1236 OutputType::DepInfo => {}
1240 // Clean up unwanted temporary files.
1242 // We create the following files by default:
1243 // - #crate#.#module-name#.bc
1244 // - #crate#.#module-name#.o
1245 // - #crate#.crate.metadata.bc
1246 // - #crate#.crate.metadata.o
1247 // - #crate#.o (linked from crate.##.o)
1248 // - #crate#.bc (copied from crate.##.bc)
1249 // We may create additional files if requested by the user (through
1250 // `-C save-temps` or `--emit=` flags).
1252 if !sess.opts.cg.save_temps {
1253 // Remove the temporary .#module-name#.o objects. If the user didn't
1254 // explicitly request bitcode (with --emit=bc), and the bitcode is not
1255 // needed for building an rlib, then we must remove .#module-name#.bc as
1258 // Specific rules for keeping .#module-name#.bc:
1259 // - If the user requested bitcode (`user_wants_bitcode`), and
1260 // codegen_units > 1, then keep it.
1261 // - If the user requested bitcode but codegen_units == 1, then we
1262 // can toss .#module-name#.bc because we copied it to .bc earlier.
1263 // - If we're not building an rlib and the user didn't request
1264 // bitcode, then delete .#module-name#.bc.
1265 // If you change how this works, also update back::link::link_rlib,
1266 // where .#module-name#.bc files are (maybe) deleted after making an
1268 let needs_crate_object = crate_output.outputs.contains_key(&OutputType::Exe);
1270 let keep_numbered_bitcode = user_wants_bitcode && sess.codegen_units() > 1;
1272 let keep_numbered_objects = needs_crate_object ||
1273 (user_wants_objects && sess.codegen_units() > 1);
1275 for module in compiled_modules.modules.iter() {
1276 if let Some(ref path) = module.object {
1277 if !keep_numbered_objects {
1282 if let Some(ref path) = module.bytecode {
1283 if !keep_numbered_bitcode {
1289 if !user_wants_bitcode {
1290 if let Some(ref path) = compiled_modules.metadata_module.bytecode {
1291 remove(sess, &path);
1294 if let Some(ref allocator_module) = compiled_modules.allocator_module {
1295 if let Some(ref path) = allocator_module.bytecode {
1302 // We leave the following files around by default:
1304 // - #crate#.crate.metadata.o
1306 // These are used in linking steps and will be cleaned up afterward.
1309 pub(crate) fn dump_incremental_data(_codegen_results: &CodegenResults) {
1310 // FIXME(mw): This does not work at the moment because the situation has
1311 // become more complicated due to incremental LTO. Now a CGU
1312 // can have more than two caching states.
1313 // println!("[incremental] Re-using {} out of {} modules",
1314 // codegen_results.modules.iter().filter(|m| m.pre_existing).count(),
1315 // codegen_results.modules.len());
1319 /// Optimize a newly codegened, totally unoptimized module.
1320 Optimize(ModuleCodegen),
1321 /// Copy the post-LTO artifacts from the incremental cache to the output
1323 CopyPostLtoArtifacts(CachedModuleCodegen),
1324 /// Perform (Thin)LTO on the given module.
1325 LTO(lto::LtoModuleCodegen),
1329 fn module_kind(&self) -> ModuleKind {
1331 WorkItem::Optimize(ref m) => m.kind,
1332 WorkItem::CopyPostLtoArtifacts(_) |
1333 WorkItem::LTO(_) => ModuleKind::Regular,
1337 fn name(&self) -> String {
1339 WorkItem::Optimize(ref m) => format!("optimize: {}", m.name),
1340 WorkItem::CopyPostLtoArtifacts(ref m) => format!("copy post LTO artifacts: {}", m.name),
1341 WorkItem::LTO(ref m) => format!("lto: {}", m.name()),
1346 enum WorkItemResult {
1347 Compiled(CompiledModule),
1348 NeedsLTO(ModuleCodegen),
1351 fn execute_work_item(cgcx: &CodegenContext,
1352 work_item: WorkItem,
1353 timeline: &mut Timeline)
1354 -> Result<WorkItemResult, FatalError>
1356 let module_config = cgcx.config(work_item.module_kind());
1359 WorkItem::Optimize(module) => {
1360 execute_optimize_work_item(cgcx, module, module_config, timeline)
1362 WorkItem::CopyPostLtoArtifacts(module) => {
1363 execute_copy_from_cache_work_item(cgcx, module, module_config, timeline)
1365 WorkItem::LTO(module) => {
1366 execute_lto_work_item(cgcx, module, module_config, timeline)
1371 fn execute_optimize_work_item(cgcx: &CodegenContext,
1372 module: ModuleCodegen,
1373 module_config: &ModuleConfig,
1374 timeline: &mut Timeline)
1375 -> Result<WorkItemResult, FatalError>
1377 let diag_handler = cgcx.create_diag_handler();
1380 optimize(cgcx, &diag_handler, &module, module_config, timeline)?;
1383 let linker_does_lto = cgcx.opts.debugging_opts.cross_lang_lto.enabled();
1385 // After we've done the initial round of optimizations we need to
1386 // decide whether to synchronously codegen this module or ship it
1387 // back to the coordinator thread for further LTO processing (which
1388 // has to wait for all the initial modules to be optimized).
1390 // Here we dispatch based on the `cgcx.lto` and kind of module we're
1392 let needs_lto = match cgcx.lto {
1395 // If the linker does LTO, we don't have to do it. Note that we
1396 // keep doing full LTO, if it is requested, as not to break the
1397 // assumption that the output will be a single module.
1398 Lto::Thin | Lto::ThinLocal if linker_does_lto => false,
1400 // Here we've got a full crate graph LTO requested. We ignore
1401 // this, however, if the crate type is only an rlib as there's
1402 // no full crate graph to process, that'll happen later.
1404 // This use case currently comes up primarily for targets that
1405 // require LTO so the request for LTO is always unconditionally
1406 // passed down to the backend, but we don't actually want to do
1407 // anything about it yet until we've got a final product.
1408 Lto::Fat | Lto::Thin => {
1409 cgcx.crate_types.len() != 1 ||
1410 cgcx.crate_types[0] != config::CrateType::Rlib
1413 // When we're automatically doing ThinLTO for multi-codegen-unit
1414 // builds we don't actually want to LTO the allocator modules if
1415 // it shows up. This is due to various linker shenanigans that
1416 // we'll encounter later.
1418 module.kind != ModuleKind::Allocator
1422 // Metadata modules never participate in LTO regardless of the lto
1424 let needs_lto = needs_lto && module.kind != ModuleKind::Metadata;
1427 Ok(WorkItemResult::NeedsLTO(module))
1429 let module = unsafe {
1430 codegen(cgcx, &diag_handler, module, module_config, timeline)?
1432 Ok(WorkItemResult::Compiled(module))
1436 fn execute_copy_from_cache_work_item(cgcx: &CodegenContext,
1437 module: CachedModuleCodegen,
1438 module_config: &ModuleConfig,
1440 -> Result<WorkItemResult, FatalError>
1442 let incr_comp_session_dir = cgcx.incr_comp_session_dir
1445 let mut object = None;
1446 let mut bytecode = None;
1447 let mut bytecode_compressed = None;
1448 for (kind, saved_file) in &module.source.saved_files {
1449 let obj_out = match kind {
1450 WorkProductFileKind::Object => {
1451 let path = cgcx.output_filenames.temp_path(OutputType::Object,
1452 Some(&module.name));
1453 object = Some(path.clone());
1456 WorkProductFileKind::Bytecode => {
1457 let path = cgcx.output_filenames.temp_path(OutputType::Bitcode,
1458 Some(&module.name));
1459 bytecode = Some(path.clone());
1462 WorkProductFileKind::BytecodeCompressed => {
1463 let path = cgcx.output_filenames.temp_path(OutputType::Bitcode,
1465 .with_extension(RLIB_BYTECODE_EXTENSION);
1466 bytecode_compressed = Some(path.clone());
1470 let source_file = in_incr_comp_dir(&incr_comp_session_dir,
1472 debug!("copying pre-existing module `{}` from {:?} to {}",
1476 if let Err(err) = link_or_copy(&source_file, &obj_out) {
1477 let diag_handler = cgcx.create_diag_handler();
1478 diag_handler.err(&format!("unable to copy {} to {}: {}",
1479 source_file.display(),
1485 assert_eq!(object.is_some(), module_config.emit_obj);
1486 assert_eq!(bytecode.is_some(), module_config.emit_bc);
1487 assert_eq!(bytecode_compressed.is_some(), module_config.emit_bc_compressed);
1489 Ok(WorkItemResult::Compiled(CompiledModule {
1491 kind: ModuleKind::Regular,
1494 bytecode_compressed,
1498 fn execute_lto_work_item(cgcx: &CodegenContext,
1499 mut module: lto::LtoModuleCodegen,
1500 module_config: &ModuleConfig,
1501 timeline: &mut Timeline)
1502 -> Result<WorkItemResult, FatalError>
1504 let diag_handler = cgcx.create_diag_handler();
1507 let module = module.optimize(cgcx, timeline)?;
1508 let module = codegen(cgcx, &diag_handler, module, module_config, timeline)?;
1509 Ok(WorkItemResult::Compiled(module))
1514 Token(io::Result<Acquired>),
1516 result: ModuleCodegen,
1520 result: Result<CompiledModule, ()>,
1524 llvm_work_item: WorkItem,
1527 AddImportOnlyModule {
1528 module_data: SerializedModule,
1529 work_product: WorkProduct,
1538 code: Option<DiagnosticId>,
1542 #[derive(PartialEq, Clone, Copy, Debug)]
1543 enum MainThreadWorkerState {
1549 fn start_executing_work(tcx: TyCtxt,
1550 crate_info: &CrateInfo,
1551 shared_emitter: SharedEmitter,
1552 codegen_worker_send: Sender<Message>,
1553 coordinator_receive: Receiver<Box<dyn Any + Send>>,
1556 time_graph: Option<TimeGraph>,
1557 modules_config: Arc<ModuleConfig>,
1558 metadata_config: Arc<ModuleConfig>,
1559 allocator_config: Arc<ModuleConfig>)
1560 -> thread::JoinHandle<Result<CompiledModules, ()>> {
1561 let coordinator_send = tcx.tx_to_llvm_workers.lock().clone();
1562 let sess = tcx.sess;
1564 // Compute the set of symbols we need to retain when doing LTO (if we need to)
1565 let exported_symbols = {
1566 let mut exported_symbols = FxHashMap::default();
1568 let copy_symbols = |cnum| {
1569 let symbols = tcx.exported_symbols(cnum)
1571 .map(|&(s, lvl)| (s.symbol_name(tcx).to_string(), lvl))
1579 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
1580 Some(Arc::new(exported_symbols))
1582 Lto::Fat | Lto::Thin => {
1583 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
1584 for &cnum in tcx.crates().iter() {
1585 exported_symbols.insert(cnum, copy_symbols(cnum));
1587 Some(Arc::new(exported_symbols))
1592 // First up, convert our jobserver into a helper thread so we can use normal
1593 // mpsc channels to manage our messages and such.
1594 // After we've requested tokens then we'll, when we can,
1595 // get tokens on `coordinator_receive` which will
1596 // get managed in the main loop below.
1597 let coordinator_send2 = coordinator_send.clone();
1598 let helper = jobserver.into_helper_thread(move |token| {
1599 drop(coordinator_send2.send(Box::new(Message::Token(token))));
1600 }).expect("failed to spawn helper thread");
1602 let mut each_linked_rlib_for_lto = Vec::new();
1603 drop(link::each_linked_rlib(sess, crate_info, &mut |cnum, path| {
1604 if link::ignored_for_lto(sess, crate_info, cnum) {
1607 each_linked_rlib_for_lto.push((cnum, path.to_path_buf()));
1610 let assembler_cmd = if modules_config.no_integrated_as {
1611 // HACK: currently we use linker (gcc) as our assembler
1612 let (linker, flavor) = link::linker_and_flavor(sess);
1614 let (name, mut cmd) = get_linker(sess, &linker, flavor);
1615 cmd.args(&sess.target.target.options.asm_args);
1617 Some(Arc::new(AssemblerCommand { name, cmd }))
1622 let cgcx = CodegenContext {
1623 crate_types: sess.crate_types.borrow().clone(),
1624 each_linked_rlib_for_lto,
1626 no_landing_pads: sess.no_landing_pads(),
1627 fewer_names: sess.fewer_names(),
1628 save_temps: sess.opts.cg.save_temps,
1629 opts: Arc::new(sess.opts.clone()),
1630 time_passes: sess.time_passes(),
1632 plugin_passes: sess.plugin_llvm_passes.borrow().clone(),
1633 remark: sess.opts.cg.remark.clone(),
1635 incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
1636 cgu_reuse_tracker: sess.cgu_reuse_tracker.clone(),
1638 diag_emitter: shared_emitter.clone(),
1640 output_filenames: tcx.output_filenames(LOCAL_CRATE),
1641 regular_module_config: modules_config,
1642 metadata_module_config: metadata_config,
1643 allocator_module_config: allocator_config,
1644 tm_factory: target_machine_factory(tcx.sess, false),
1646 msvc_imps_needed: msvc_imps_needed(tcx),
1647 target_pointer_width: tcx.sess.target.target.target_pointer_width.clone(),
1648 debuginfo: tcx.sess.opts.debuginfo,
1650 phantom: PhantomData
1653 // This is the "main loop" of parallel work happening for parallel codegen.
1654 // It's here that we manage parallelism, schedule work, and work with
1655 // messages coming from clients.
1657 // There are a few environmental pre-conditions that shape how the system
1660 // - Error reporting only can happen on the main thread because that's the
1661 // only place where we have access to the compiler `Session`.
1662 // - LLVM work can be done on any thread.
1663 // - Codegen can only happen on the main thread.
1664 // - Each thread doing substantial work most be in possession of a `Token`
1665 // from the `Jobserver`.
1666 // - The compiler process always holds one `Token`. Any additional `Tokens`
1667 // have to be requested from the `Jobserver`.
1671 // The error reporting restriction is handled separately from the rest: We
1672 // set up a `SharedEmitter` the holds an open channel to the main thread.
1673 // When an error occurs on any thread, the shared emitter will send the
1674 // error message to the receiver main thread (`SharedEmitterMain`). The
1675 // main thread will periodically query this error message queue and emit
1676 // any error messages it has received. It might even abort compilation if
1677 // has received a fatal error. In this case we rely on all other threads
1678 // being torn down automatically with the main thread.
1679 // Since the main thread will often be busy doing codegen work, error
1680 // reporting will be somewhat delayed, since the message queue can only be
1681 // checked in between to work packages.
1683 // Work Processing Infrastructure
1684 // ==============================
1685 // The work processing infrastructure knows three major actors:
1687 // - the coordinator thread,
1688 // - the main thread, and
1689 // - LLVM worker threads
1691 // The coordinator thread is running a message loop. It instructs the main
1692 // thread about what work to do when, and it will spawn off LLVM worker
1693 // threads as open LLVM WorkItems become available.
1695 // The job of the main thread is to codegen CGUs into LLVM work package
1696 // (since the main thread is the only thread that can do this). The main
1697 // thread will block until it receives a message from the coordinator, upon
1698 // which it will codegen one CGU, send it to the coordinator and block
1699 // again. This way the coordinator can control what the main thread is
1702 // The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is
1703 // available, it will spawn off a new LLVM worker thread and let it process
1704 // that a WorkItem. When a LLVM worker thread is done with its WorkItem,
1705 // it will just shut down, which also frees all resources associated with
1706 // the given LLVM module, and sends a message to the coordinator that the
1707 // has been completed.
1711 // The scheduler's goal is to minimize the time it takes to complete all
1712 // work there is, however, we also want to keep memory consumption low
1713 // if possible. These two goals are at odds with each other: If memory
1714 // consumption were not an issue, we could just let the main thread produce
1715 // LLVM WorkItems at full speed, assuring maximal utilization of
1716 // Tokens/LLVM worker threads. However, since codegen usual is faster
1717 // than LLVM processing, the queue of LLVM WorkItems would fill up and each
1718 // WorkItem potentially holds on to a substantial amount of memory.
1720 // So the actual goal is to always produce just enough LLVM WorkItems as
1721 // not to starve our LLVM worker threads. That means, once we have enough
1722 // WorkItems in our queue, we can block the main thread, so it does not
1723 // produce more until we need them.
1725 // Doing LLVM Work on the Main Thread
1726 // ----------------------------------
1727 // Since the main thread owns the compiler processes implicit `Token`, it is
1728 // wasteful to keep it blocked without doing any work. Therefore, what we do
1729 // in this case is: We spawn off an additional LLVM worker thread that helps
1730 // reduce the queue. The work it is doing corresponds to the implicit
1731 // `Token`. The coordinator will mark the main thread as being busy with
1732 // LLVM work. (The actual work happens on another OS thread but we just care
1733 // about `Tokens`, not actual threads).
1735 // When any LLVM worker thread finishes while the main thread is marked as
1736 // "busy with LLVM work", we can do a little switcheroo: We give the Token
1737 // of the just finished thread to the LLVM worker thread that is working on
1738 // behalf of the main thread's implicit Token, thus freeing up the main
1739 // thread again. The coordinator can then again decide what the main thread
1740 // should do. This allows the coordinator to make decisions at more points
1743 // Striking a Balance between Throughput and Memory Consumption
1744 // ------------------------------------------------------------
1745 // Since our two goals, (1) use as many Tokens as possible and (2) keep
1746 // memory consumption as low as possible, are in conflict with each other,
1747 // we have to find a trade off between them. Right now, the goal is to keep
1748 // all workers busy, which means that no worker should find the queue empty
1749 // when it is ready to start.
1750 // How do we do achieve this? Good question :) We actually never know how
1751 // many `Tokens` are potentially available so it's hard to say how much to
1752 // fill up the queue before switching the main thread to LLVM work. Also we
1753 // currently don't have a means to estimate how long a running LLVM worker
1754 // will still be busy with it's current WorkItem. However, we know the
1755 // maximal count of available Tokens that makes sense (=the number of CPU
1756 // cores), so we can take a conservative guess. The heuristic we use here
1757 // is implemented in the `queue_full_enough()` function.
1759 // Some Background on Jobservers
1760 // -----------------------------
1761 // It's worth also touching on the management of parallelism here. We don't
1762 // want to just spawn a thread per work item because while that's optimal
1763 // parallelism it may overload a system with too many threads or violate our
1764 // configuration for the maximum amount of cpu to use for this process. To
1765 // manage this we use the `jobserver` crate.
1767 // Job servers are an artifact of GNU make and are used to manage
1768 // parallelism between processes. A jobserver is a glorified IPC semaphore
1769 // basically. Whenever we want to run some work we acquire the semaphore,
1770 // and whenever we're done with that work we release the semaphore. In this
1771 // manner we can ensure that the maximum number of parallel workers is
1772 // capped at any one point in time.
1774 // LTO and the coordinator thread
1775 // ------------------------------
1777 // The final job the coordinator thread is responsible for is managing LTO
1778 // and how that works. When LTO is requested what we'll to is collect all
1779 // optimized LLVM modules into a local vector on the coordinator. Once all
1780 // modules have been codegened and optimized we hand this to the `lto`
1781 // module for further optimization. The `lto` module will return back a list
1782 // of more modules to work on, which the coordinator will continue to spawn
1785 // Each LLVM module is automatically sent back to the coordinator for LTO if
1786 // necessary. There's already optimizations in place to avoid sending work
1787 // back to the coordinator if LTO isn't requested.
1788 return thread::spawn(move || {
1789 // We pretend to be within the top-level LLVM time-passes task here:
1792 let max_workers = ::num_cpus::get();
1793 let mut worker_id_counter = 0;
1794 let mut free_worker_ids = Vec::new();
1795 let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
1796 if let Some(id) = free_worker_ids.pop() {
1799 let id = worker_id_counter;
1800 worker_id_counter += 1;
1805 // This is where we collect codegen units that have gone all the way
1806 // through codegen and LLVM.
1807 let mut compiled_modules = vec![];
1808 let mut compiled_metadata_module = None;
1809 let mut compiled_allocator_module = None;
1810 let mut needs_lto = Vec::new();
1811 let mut lto_import_only_modules = Vec::new();
1812 let mut started_lto = false;
1813 let mut codegen_aborted = false;
1815 // This flag tracks whether all items have gone through codegens
1816 let mut codegen_done = false;
1818 // This is the queue of LLVM work items that still need processing.
1819 let mut work_items = Vec::<(WorkItem, u64)>::new();
1821 // This are the Jobserver Tokens we currently hold. Does not include
1822 // the implicit Token the compiler process owns no matter what.
1823 let mut tokens = Vec::new();
1825 let mut main_thread_worker_state = MainThreadWorkerState::Idle;
1826 let mut running = 0;
1828 let mut llvm_start_time = None;
1830 // Run the message loop while there's still anything that needs message
1831 // processing. Note that as soon as codegen is aborted we simply want to
1832 // wait for all existing work to finish, so many of the conditions here
1833 // only apply if codegen hasn't been aborted as they represent pending
1835 while !codegen_done ||
1837 (!codegen_aborted && (
1838 work_items.len() > 0 ||
1839 needs_lto.len() > 0 ||
1840 lto_import_only_modules.len() > 0 ||
1841 main_thread_worker_state != MainThreadWorkerState::Idle
1845 // While there are still CGUs to be codegened, the coordinator has
1846 // to decide how to utilize the compiler processes implicit Token:
1847 // For codegenning more CGU or for running them through LLVM.
1849 if main_thread_worker_state == MainThreadWorkerState::Idle {
1850 if !queue_full_enough(work_items.len(), running, max_workers) {
1851 // The queue is not full enough, codegen more items:
1852 if let Err(_) = codegen_worker_send.send(Message::CodegenItem) {
1853 panic!("Could not send Message::CodegenItem to main thread")
1855 main_thread_worker_state = MainThreadWorkerState::Codegenning;
1857 // The queue is full enough to not let the worker
1858 // threads starve. Use the implicit Token to do some
1860 let (item, _) = work_items.pop()
1861 .expect("queue empty - queue_full_enough() broken?");
1862 let cgcx = CodegenContext {
1863 worker: get_worker_id(&mut free_worker_ids),
1866 maybe_start_llvm_timer(cgcx.config(item.module_kind()),
1867 &mut llvm_start_time);
1868 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1869 spawn_work(cgcx, item);
1872 } else if codegen_aborted {
1873 // don't queue up any more work if codegen was aborted, we're
1874 // just waiting for our existing children to finish
1876 // If we've finished everything related to normal codegen
1877 // then it must be the case that we've got some LTO work to do.
1878 // Perform the serial work here of figuring out what we're
1879 // going to LTO and then push a bunch of work items onto our
1881 if work_items.len() == 0 &&
1883 main_thread_worker_state == MainThreadWorkerState::Idle {
1884 assert!(!started_lto);
1885 assert!(needs_lto.len() + lto_import_only_modules.len() > 0);
1887 let modules = mem::replace(&mut needs_lto, Vec::new());
1888 let import_only_modules =
1889 mem::replace(&mut lto_import_only_modules, Vec::new());
1890 for (work, cost) in generate_lto_work(&cgcx, modules, import_only_modules) {
1891 let insertion_index = work_items
1892 .binary_search_by_key(&cost, |&(_, cost)| cost)
1893 .unwrap_or_else(|e| e);
1894 work_items.insert(insertion_index, (work, cost));
1895 if !cgcx.opts.debugging_opts.no_parallel_llvm {
1896 helper.request_token();
1901 // In this branch, we know that everything has been codegened,
1902 // so it's just a matter of determining whether the implicit
1903 // Token is free to use for LLVM work.
1904 match main_thread_worker_state {
1905 MainThreadWorkerState::Idle => {
1906 if let Some((item, _)) = work_items.pop() {
1907 let cgcx = CodegenContext {
1908 worker: get_worker_id(&mut free_worker_ids),
1911 maybe_start_llvm_timer(cgcx.config(item.module_kind()),
1912 &mut llvm_start_time);
1913 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1914 spawn_work(cgcx, item);
1916 // There is no unstarted work, so let the main thread
1917 // take over for a running worker. Otherwise the
1918 // implicit token would just go to waste.
1919 // We reduce the `running` counter by one. The
1920 // `tokens.truncate()` below will take care of
1921 // giving the Token back.
1922 debug_assert!(running > 0);
1924 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1927 MainThreadWorkerState::Codegenning => {
1928 bug!("codegen worker should not be codegenning after \
1929 codegen was already completed")
1931 MainThreadWorkerState::LLVMing => {
1932 // Already making good use of that token
1937 // Spin up what work we can, only doing this while we've got available
1938 // parallelism slots and work left to spawn.
1939 while !codegen_aborted && work_items.len() > 0 && running < tokens.len() {
1940 let (item, _) = work_items.pop().unwrap();
1942 maybe_start_llvm_timer(cgcx.config(item.module_kind()),
1943 &mut llvm_start_time);
1945 let cgcx = CodegenContext {
1946 worker: get_worker_id(&mut free_worker_ids),
1950 spawn_work(cgcx, item);
1954 // Relinquish accidentally acquired extra tokens
1955 tokens.truncate(running);
1957 let msg = coordinator_receive.recv().unwrap();
1958 match *msg.downcast::<Message>().ok().unwrap() {
1959 // Save the token locally and the next turn of the loop will use
1960 // this to spawn a new unit of work, or it may get dropped
1961 // immediately if we have no more work to spawn.
1962 Message::Token(token) => {
1967 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
1968 // If the main thread token is used for LLVM work
1969 // at the moment, we turn that thread into a regular
1970 // LLVM worker thread, so the main thread is free
1971 // to react to codegen demand.
1972 main_thread_worker_state = MainThreadWorkerState::Idle;
1977 let msg = &format!("failed to acquire jobserver token: {}", e);
1978 shared_emitter.fatal(msg);
1979 // Exit the coordinator thread
1985 Message::CodegenDone { llvm_work_item, cost } => {
1986 // We keep the queue sorted by estimated processing cost,
1987 // so that more expensive items are processed earlier. This
1988 // is good for throughput as it gives the main thread more
1989 // time to fill up the queue and it avoids scheduling
1990 // expensive items to the end.
1991 // Note, however, that this is not ideal for memory
1992 // consumption, as LLVM module sizes are not evenly
1994 let insertion_index =
1995 work_items.binary_search_by_key(&cost, |&(_, cost)| cost);
1996 let insertion_index = match insertion_index {
1997 Ok(idx) | Err(idx) => idx
1999 work_items.insert(insertion_index, (llvm_work_item, cost));
2001 if !cgcx.opts.debugging_opts.no_parallel_llvm {
2002 helper.request_token();
2004 assert!(!codegen_aborted);
2005 assert_eq!(main_thread_worker_state,
2006 MainThreadWorkerState::Codegenning);
2007 main_thread_worker_state = MainThreadWorkerState::Idle;
2010 Message::CodegenComplete => {
2011 codegen_done = true;
2012 assert!(!codegen_aborted);
2013 assert_eq!(main_thread_worker_state,
2014 MainThreadWorkerState::Codegenning);
2015 main_thread_worker_state = MainThreadWorkerState::Idle;
2018 // If codegen is aborted that means translation was aborted due
2019 // to some normal-ish compiler error. In this situation we want
2020 // to exit as soon as possible, but we want to make sure all
2021 // existing work has finished. Flag codegen as being done, and
2022 // then conditions above will ensure no more work is spawned but
2023 // we'll keep executing this loop until `running` hits 0.
2024 Message::CodegenAborted => {
2025 assert!(!codegen_aborted);
2026 codegen_done = true;
2027 codegen_aborted = true;
2028 assert_eq!(main_thread_worker_state,
2029 MainThreadWorkerState::Codegenning);
2032 // If a thread exits successfully then we drop a token associated
2033 // with that worker and update our `running` count. We may later
2034 // re-acquire a token to continue running more work. We may also not
2035 // actually drop a token here if the worker was running with an
2036 // "ephemeral token"
2038 // Note that if the thread failed that means it panicked, so we
2039 // abort immediately.
2040 Message::Done { result: Ok(compiled_module), worker_id } => {
2041 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
2042 main_thread_worker_state = MainThreadWorkerState::Idle;
2047 free_worker_ids.push(worker_id);
2049 match compiled_module.kind {
2050 ModuleKind::Regular => {
2051 compiled_modules.push(compiled_module);
2053 ModuleKind::Metadata => {
2054 assert!(compiled_metadata_module.is_none());
2055 compiled_metadata_module = Some(compiled_module);
2057 ModuleKind::Allocator => {
2058 assert!(compiled_allocator_module.is_none());
2059 compiled_allocator_module = Some(compiled_module);
2063 Message::NeedsLTO { result, worker_id } => {
2064 assert!(!started_lto);
2065 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
2066 main_thread_worker_state = MainThreadWorkerState::Idle;
2070 free_worker_ids.push(worker_id);
2071 needs_lto.push(result);
2073 Message::AddImportOnlyModule { module_data, work_product } => {
2074 assert!(!started_lto);
2075 assert!(!codegen_done);
2076 assert_eq!(main_thread_worker_state,
2077 MainThreadWorkerState::Codegenning);
2078 lto_import_only_modules.push((module_data, work_product));
2079 main_thread_worker_state = MainThreadWorkerState::Idle;
2081 Message::Done { result: Err(()), worker_id: _ } => {
2082 bug!("worker thread panicked");
2084 Message::CodegenItem => {
2085 bug!("the coordinator should not receive codegen requests")
2090 if let Some(llvm_start_time) = llvm_start_time {
2091 let total_llvm_time = Instant::now().duration_since(llvm_start_time);
2092 // This is the top-level timing for all of LLVM, set the time-depth
2095 print_time_passes_entry(cgcx.time_passes,
2100 // Regardless of what order these modules completed in, report them to
2101 // the backend in the same order every time to ensure that we're handing
2102 // out deterministic results.
2103 compiled_modules.sort_by(|a, b| a.name.cmp(&b.name));
2105 let compiled_metadata_module = compiled_metadata_module
2106 .expect("Metadata module not compiled?");
2108 Ok(CompiledModules {
2109 modules: compiled_modules,
2110 metadata_module: compiled_metadata_module,
2111 allocator_module: compiled_allocator_module,
2115 // A heuristic that determines if we have enough LLVM WorkItems in the
2116 // queue so that the main thread can do LLVM work instead of codegen
2117 fn queue_full_enough(items_in_queue: usize,
2118 workers_running: usize,
2119 max_workers: usize) -> bool {
2121 items_in_queue > 0 &&
2122 items_in_queue >= max_workers.saturating_sub(workers_running / 2)
2125 fn maybe_start_llvm_timer(config: &ModuleConfig,
2126 llvm_start_time: &mut Option<Instant>) {
2127 // We keep track of the -Ztime-passes output manually,
2128 // since the closure-based interface does not fit well here.
2129 if config.time_passes {
2130 if llvm_start_time.is_none() {
2131 *llvm_start_time = Some(Instant::now());
2137 pub const CODEGEN_WORKER_ID: usize = ::std::usize::MAX;
2138 pub const CODEGEN_WORKER_TIMELINE: time_graph::TimelineId =
2139 time_graph::TimelineId(CODEGEN_WORKER_ID);
2140 pub const CODEGEN_WORK_PACKAGE_KIND: time_graph::WorkPackageKind =
2141 time_graph::WorkPackageKind(&["#DE9597", "#FED1D3", "#FDC5C7", "#B46668", "#88494B"]);
2142 const LLVM_WORK_PACKAGE_KIND: time_graph::WorkPackageKind =
2143 time_graph::WorkPackageKind(&["#7DB67A", "#C6EEC4", "#ACDAAA", "#579354", "#3E6F3C"]);
2145 fn spawn_work(cgcx: CodegenContext<'static>, work: WorkItem) {
2146 let depth = time_depth();
2148 thread::spawn(move || {
2149 set_time_depth(depth);
2151 // Set up a destructor which will fire off a message that we're done as
2154 coordinator_send: Sender<Box<dyn Any + Send>>,
2155 result: Option<WorkItemResult>,
2158 impl Drop for Bomb {
2159 fn drop(&mut self) {
2160 let worker_id = self.worker_id;
2161 let msg = match self.result.take() {
2162 Some(WorkItemResult::Compiled(m)) => {
2163 Message::Done { result: Ok(m), worker_id }
2165 Some(WorkItemResult::NeedsLTO(m)) => {
2166 Message::NeedsLTO { result: m, worker_id }
2168 None => Message::Done { result: Err(()), worker_id }
2170 drop(self.coordinator_send.send(Box::new(msg)));
2174 let mut bomb = Bomb {
2175 coordinator_send: cgcx.coordinator_send.clone(),
2177 worker_id: cgcx.worker,
2180 // Execute the work itself, and if it finishes successfully then flag
2181 // ourselves as a success as well.
2183 // Note that we ignore any `FatalError` coming out of `execute_work_item`,
2184 // as a diagnostic was already sent off to the main thread - just
2185 // surface that there was an error in this worker.
2187 let timeline = cgcx.time_graph.as_ref().map(|tg| {
2188 tg.start(time_graph::TimelineId(cgcx.worker),
2189 LLVM_WORK_PACKAGE_KIND,
2192 let mut timeline = timeline.unwrap_or(Timeline::noop());
2193 execute_work_item(&cgcx, work, &mut timeline).ok()
2198 pub fn run_assembler(cgcx: &CodegenContext, handler: &Handler, assembly: &Path, object: &Path) {
2199 let assembler = cgcx.assembler_cmd
2201 .expect("cgcx.assembler_cmd is missing?");
2203 let pname = &assembler.name;
2204 let mut cmd = assembler.cmd.clone();
2205 cmd.arg("-c").arg("-o").arg(object).arg(assembly);
2206 debug!("{:?}", cmd);
2208 match cmd.output() {
2210 if !prog.status.success() {
2211 let mut note = prog.stderr.clone();
2212 note.extend_from_slice(&prog.stdout);
2214 handler.struct_err(&format!("linking with `{}` failed: {}",
2217 .note(&format!("{:?}", &cmd))
2218 .note(str::from_utf8(¬e[..]).unwrap())
2220 handler.abort_if_errors();
2224 handler.err(&format!("could not exec the linker `{}`: {}", pname.display(), e));
2225 handler.abort_if_errors();
2230 pub unsafe fn with_llvm_pmb(llmod: &llvm::Module,
2231 config: &ModuleConfig,
2232 opt_level: llvm::CodeGenOptLevel,
2233 prepare_for_thin_lto: bool,
2234 f: &mut dyn FnMut(&llvm::PassManagerBuilder)) {
2237 // Create the PassManagerBuilder for LLVM. We configure it with
2238 // reasonable defaults and prepare it to actually populate the pass
2240 let builder = llvm::LLVMPassManagerBuilderCreate();
2241 let opt_size = config.opt_size.unwrap_or(llvm::CodeGenOptSizeNone);
2242 let inline_threshold = config.inline_threshold;
2244 let pgo_gen_path = config.pgo_gen.as_ref().map(|s| {
2245 let s = if s.is_empty() { "default_%m.profraw" } else { s };
2246 CString::new(s.as_bytes()).unwrap()
2249 let pgo_use_path = if config.pgo_use.is_empty() {
2252 Some(CString::new(config.pgo_use.as_bytes()).unwrap())
2255 llvm::LLVMRustConfigurePassManagerBuilder(
2258 config.merge_functions,
2259 config.vectorize_slp,
2260 config.vectorize_loop,
2261 prepare_for_thin_lto,
2262 pgo_gen_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
2263 pgo_use_path.as_ref().map_or(ptr::null(), |s| s.as_ptr()),
2266 llvm::LLVMPassManagerBuilderSetSizeLevel(builder, opt_size as u32);
2268 if opt_size != llvm::CodeGenOptSizeNone {
2269 llvm::LLVMPassManagerBuilderSetDisableUnrollLoops(builder, 1);
2272 llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins);
2274 // Here we match what clang does (kinda). For O0 we only inline
2275 // always-inline functions (but don't add lifetime intrinsics), at O1 we
2276 // inline with lifetime intrinsics, and O2+ we add an inliner with a
2277 // thresholds copied from clang.
2278 match (opt_level, opt_size, inline_threshold) {
2280 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, t as u32);
2282 (llvm::CodeGenOptLevel::Aggressive, ..) => {
2283 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 275);
2285 (_, llvm::CodeGenOptSizeDefault, _) => {
2286 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 75);
2288 (_, llvm::CodeGenOptSizeAggressive, _) => {
2289 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 25);
2291 (llvm::CodeGenOptLevel::None, ..) => {
2292 llvm::LLVMRustAddAlwaysInlinePass(builder, false);
2294 (llvm::CodeGenOptLevel::Less, ..) => {
2295 llvm::LLVMRustAddAlwaysInlinePass(builder, true);
2297 (llvm::CodeGenOptLevel::Default, ..) => {
2298 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 225);
2300 (llvm::CodeGenOptLevel::Other, ..) => {
2301 bug!("CodeGenOptLevel::Other selected")
2306 llvm::LLVMPassManagerBuilderDispose(builder);
2310 enum SharedEmitterMessage {
2311 Diagnostic(Diagnostic),
2312 InlineAsmError(u32, String),
2318 pub struct SharedEmitter {
2319 sender: Sender<SharedEmitterMessage>,
2322 pub struct SharedEmitterMain {
2323 receiver: Receiver<SharedEmitterMessage>,
2326 impl SharedEmitter {
2327 pub fn new() -> (SharedEmitter, SharedEmitterMain) {
2328 let (sender, receiver) = channel();
2330 (SharedEmitter { sender }, SharedEmitterMain { receiver })
2333 fn inline_asm_error(&self, cookie: u32, msg: String) {
2334 drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg)));
2337 fn fatal(&self, msg: &str) {
2338 drop(self.sender.send(SharedEmitterMessage::Fatal(msg.to_string())));
2342 impl Emitter for SharedEmitter {
2343 fn emit(&mut self, db: &DiagnosticBuilder) {
2344 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
2346 code: db.code.clone(),
2349 for child in &db.children {
2350 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
2351 msg: child.message(),
2356 drop(self.sender.send(SharedEmitterMessage::AbortIfErrors));
2360 impl SharedEmitterMain {
2361 pub fn check(&self, sess: &Session, blocking: bool) {
2363 let message = if blocking {
2364 match self.receiver.recv() {
2365 Ok(message) => Ok(message),
2369 match self.receiver.try_recv() {
2370 Ok(message) => Ok(message),
2376 Ok(SharedEmitterMessage::Diagnostic(diag)) => {
2377 let handler = sess.diagnostic();
2380 handler.emit_with_code(&MultiSpan::new(),
2386 handler.emit(&MultiSpan::new(),
2392 Ok(SharedEmitterMessage::InlineAsmError(cookie, msg)) => {
2393 match Mark::from_u32(cookie).expn_info() {
2394 Some(ei) => sess.span_err(ei.call_site, &msg),
2395 None => sess.err(&msg),
2398 Ok(SharedEmitterMessage::AbortIfErrors) => {
2399 sess.abort_if_errors();
2401 Ok(SharedEmitterMessage::Fatal(msg)) => {
2413 pub struct OngoingCodegen {
2416 metadata: EncodedMetadata,
2417 windows_subsystem: Option<String>,
2418 linker_info: LinkerInfo,
2419 crate_info: CrateInfo,
2420 time_graph: Option<TimeGraph>,
2421 coordinator_send: Sender<Box<dyn Any + Send>>,
2422 codegen_worker_receive: Receiver<Message>,
2423 shared_emitter_main: SharedEmitterMain,
2424 future: thread::JoinHandle<Result<CompiledModules, ()>>,
2425 output_filenames: Arc<OutputFilenames>,
2428 impl OngoingCodegen {
2432 ) -> (CodegenResults, FxHashMap<WorkProductId, WorkProduct>) {
2433 self.shared_emitter_main.check(sess, true);
2434 let compiled_modules = match self.future.join() {
2435 Ok(Ok(compiled_modules)) => compiled_modules,
2437 sess.abort_if_errors();
2438 panic!("expected abort due to worker thread errors")
2441 bug!("panic during codegen/LLVM phase");
2445 sess.cgu_reuse_tracker.check_expected_reuse(sess);
2447 sess.abort_if_errors();
2449 if let Some(time_graph) = self.time_graph {
2450 time_graph.dump(&format!("{}-timings", self.crate_name));
2454 copy_all_cgu_workproducts_to_incr_comp_cache_dir(sess,
2456 produce_final_output_artifacts(sess,
2458 &self.output_filenames);
2460 // FIXME: time_llvm_passes support - does this use a global context or
2462 if sess.codegen_units() == 1 && sess.time_llvm_passes() {
2463 unsafe { llvm::LLVMRustPrintPassTimings(); }
2467 crate_name: self.crate_name,
2468 crate_hash: self.crate_hash,
2469 metadata: self.metadata,
2470 windows_subsystem: self.windows_subsystem,
2471 linker_info: self.linker_info,
2472 crate_info: self.crate_info,
2474 modules: compiled_modules.modules,
2475 allocator_module: compiled_modules.allocator_module,
2476 metadata_module: compiled_modules.metadata_module,
2480 pub(crate) fn submit_pre_codegened_module_to_llvm(&self,
2482 module: ModuleCodegen) {
2483 self.wait_for_signal_to_codegen_item();
2484 self.check_for_errors(tcx.sess);
2486 // These are generally cheap and won't through off scheduling.
2488 submit_codegened_module_to_llvm(tcx, module, cost);
2491 pub fn codegen_finished(&self, tcx: TyCtxt) {
2492 self.wait_for_signal_to_codegen_item();
2493 self.check_for_errors(tcx.sess);
2494 drop(self.coordinator_send.send(Box::new(Message::CodegenComplete)));
2497 /// Consume this context indicating that codegen was entirely aborted, and
2498 /// we need to exit as quickly as possible.
2500 /// This method blocks the current thread until all worker threads have
2501 /// finished, and all worker threads should have exited or be real close to
2502 /// exiting at this point.
2503 pub fn codegen_aborted(self) {
2504 // Signal to the coordinator it should spawn no more work and start
2506 drop(self.coordinator_send.send(Box::new(Message::CodegenAborted)));
2507 drop(self.future.join());
2510 pub fn check_for_errors(&self, sess: &Session) {
2511 self.shared_emitter_main.check(sess, false);
2514 pub fn wait_for_signal_to_codegen_item(&self) {
2515 match self.codegen_worker_receive.recv() {
2516 Ok(Message::CodegenItem) => {
2519 Ok(_) => panic!("unexpected message"),
2521 // One of the LLVM threads must have panicked, fall through so
2522 // error handling can be reached.
2528 // impl Drop for OngoingCodegen {
2529 // fn drop(&mut self) {
2533 pub(crate) fn submit_codegened_module_to_llvm(tcx: TyCtxt,
2534 module: ModuleCodegen,
2536 let llvm_work_item = WorkItem::Optimize(module);
2537 drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::CodegenDone {
2543 pub(crate) fn submit_post_lto_module_to_llvm(tcx: TyCtxt,
2544 module: CachedModuleCodegen) {
2545 let llvm_work_item = WorkItem::CopyPostLtoArtifacts(module);
2546 drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::CodegenDone {
2552 pub(crate) fn submit_pre_lto_module_to_llvm(tcx: TyCtxt,
2553 module: CachedModuleCodegen) {
2554 let filename = pre_lto_bitcode_filename(&module.name);
2555 let bc_path = in_incr_comp_dir_sess(tcx.sess, &filename);
2556 let file = fs::File::open(&bc_path).unwrap_or_else(|e| {
2557 panic!("failed to open bitcode file `{}`: {}", bc_path.display(), e)
2561 memmap::Mmap::map(&file).unwrap_or_else(|e| {
2562 panic!("failed to mmap bitcode file `{}`: {}", bc_path.display(), e)
2566 // Schedule the module to be loaded
2567 drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::AddImportOnlyModule {
2568 module_data: SerializedModule::FromUncompressedFile(mmap),
2569 work_product: module.source,
2573 pub(super) fn pre_lto_bitcode_filename(module_name: &str) -> String {
2574 format!("{}.{}", module_name, PRE_THIN_LTO_BC_EXT)
2577 fn msvc_imps_needed(tcx: TyCtxt) -> bool {
2578 // This should never be true (because it's not supported). If it is true,
2579 // something is wrong with commandline arg validation.
2580 assert!(!(tcx.sess.opts.debugging_opts.cross_lang_lto.enabled() &&
2581 tcx.sess.target.target.options.is_like_msvc &&
2582 tcx.sess.opts.cg.prefer_dynamic));
2584 tcx.sess.target.target.options.is_like_msvc &&
2585 tcx.sess.crate_types.borrow().iter().any(|ct| *ct == config::CrateType::Rlib) &&
2586 // ThinLTO can't handle this workaround in all cases, so we don't
2587 // emit the `__imp_` symbols. Instead we make them unnecessary by disallowing
2588 // dynamic linking when cross-language LTO is enabled.
2589 !tcx.sess.opts.debugging_opts.cross_lang_lto.enabled()
2592 // Create a `__imp_<symbol> = &symbol` global for every public static `symbol`.
2593 // This is required to satisfy `dllimport` references to static data in .rlibs
2594 // when using MSVC linker. We do this only for data, as linker can fix up
2595 // code references on its own.
2596 // See #26591, #27438
2597 fn create_msvc_imps(cgcx: &CodegenContext, llcx: &llvm::Context, llmod: &llvm::Module) {
2598 if !cgcx.msvc_imps_needed {
2601 // The x86 ABI seems to require that leading underscores are added to symbol
2602 // names, so we need an extra underscore on 32-bit. There's also a leading
2603 // '\x01' here which disables LLVM's symbol mangling (e.g. no extra
2604 // underscores added in front).
2605 let prefix = if cgcx.target_pointer_width == "32" {
2611 let i8p_ty = Type::i8p_llcx(llcx);
2612 let globals = base::iter_globals(llmod)
2614 llvm::LLVMRustGetLinkage(val) == llvm::Linkage::ExternalLinkage &&
2615 llvm::LLVMIsDeclaration(val) == 0
2618 let name = CStr::from_ptr(llvm::LLVMGetValueName(val));
2619 let mut imp_name = prefix.as_bytes().to_vec();
2620 imp_name.extend(name.to_bytes());
2621 let imp_name = CString::new(imp_name).unwrap();
2624 .collect::<Vec<_>>();
2625 for (imp_name, val) in globals {
2626 let imp = llvm::LLVMAddGlobal(llmod,
2628 imp_name.as_ptr() as *const _);
2629 llvm::LLVMSetInitializer(imp, consts::ptrcast(val, i8p_ty));
2630 llvm::LLVMRustSetLinkage(imp, llvm::Linkage::ExternalLinkage);