1 use crate::{ModuleCodegen, ModuleKind, CachedModuleCodegen, CompiledModule, CrateInfo,
2 CodegenResults, RLIB_BYTECODE_EXTENSION};
3 use super::linker::LinkerInfo;
4 use super::lto::{self, SerializedModule};
5 use super::link::{self, remove, get_linker};
6 use super::command::Command;
7 use super::symbol_export::ExportedSymbols;
10 use rustc_incremental::{copy_cgu_workproducts_to_incr_comp_cache_dir,
11 in_incr_comp_dir, in_incr_comp_dir_sess};
12 use rustc::dep_graph::{WorkProduct, WorkProductId, WorkProductFileKind};
13 use rustc::dep_graph::cgu_reuse_tracker::CguReuseTracker;
14 use rustc::middle::cstore::EncodedMetadata;
15 use rustc::session::config::{self, OutputFilenames, OutputType, Passes, Sanitizer, Lto};
16 use rustc::session::Session;
17 use rustc::util::nodemap::FxHashMap;
18 use rustc::util::time_graph::{self, TimeGraph, Timeline};
19 use rustc::hir::def_id::{CrateNum, LOCAL_CRATE};
20 use rustc::ty::TyCtxt;
21 use rustc::util::common::{time_depth, set_time_depth, print_time_passes_entry};
22 use rustc_fs_util::link_or_copy;
23 use rustc_data_structures::svh::Svh;
24 use rustc_errors::{Handler, Level, DiagnosticBuilder, FatalError, DiagnosticId};
25 use rustc_errors::emitter::{Emitter};
26 use rustc_target::spec::MergeFunctions;
28 use syntax::ext::hygiene::Mark;
29 use syntax_pos::MultiSpan;
30 use syntax_pos::symbol::Symbol;
31 use jobserver::{Client, Acquired};
37 use std::path::{Path, PathBuf};
40 use std::sync::mpsc::{channel, Sender, Receiver};
41 use std::time::Instant;
44 const PRE_THIN_LTO_BC_EXT: &str = "pre-thin-lto.bc";
46 /// Module-specific configuration for `optimize_and_codegen`.
47 pub struct ModuleConfig {
48 /// Names of additional optimization passes to run.
49 pub passes: Vec<String>,
50 /// Some(level) to optimize at a certain level, or None to run
51 /// absolutely no optimizations (used for the metadata module).
52 pub opt_level: Option<config::OptLevel>,
54 /// Some(level) to optimize binary size, or None to not affect program size.
55 pub opt_size: Option<config::OptLevel>,
57 pub pgo_gen: Option<String>,
60 // Flags indicating which outputs to produce.
61 pub emit_pre_thin_lto_bc: bool,
62 pub emit_no_opt_bc: bool,
64 pub emit_bc_compressed: bool,
65 pub emit_lto_bc: bool,
69 // Miscellaneous flags. These are mostly copied from command-line
71 pub verify_llvm_ir: bool,
72 pub no_prepopulate_passes: bool,
73 pub no_builtins: bool,
74 pub time_passes: bool,
75 pub vectorize_loop: bool,
76 pub vectorize_slp: bool,
77 pub merge_functions: bool,
78 pub inline_threshold: Option<usize>,
79 // Instead of creating an object file by doing LLVM codegen, just
80 // make the object file bitcode. Provides easy compatibility with
81 // emscripten's ecc compiler, when used as the linker.
82 pub obj_is_bitcode: bool,
83 pub no_integrated_as: bool,
84 pub embed_bitcode: bool,
85 pub embed_bitcode_marker: bool,
89 fn new(passes: Vec<String>) -> ModuleConfig {
96 pgo_use: String::new(),
98 emit_no_opt_bc: false,
99 emit_pre_thin_lto_bc: false,
101 emit_bc_compressed: false,
106 obj_is_bitcode: false,
107 embed_bitcode: false,
108 embed_bitcode_marker: false,
109 no_integrated_as: false,
111 verify_llvm_ir: false,
112 no_prepopulate_passes: false,
115 vectorize_loop: false,
116 vectorize_slp: false,
117 merge_functions: false,
118 inline_threshold: None
122 fn set_flags(&mut self, sess: &Session, no_builtins: bool) {
123 self.verify_llvm_ir = sess.verify_llvm_ir();
124 self.no_prepopulate_passes = sess.opts.cg.no_prepopulate_passes;
125 self.no_builtins = no_builtins || sess.target.target.options.no_builtins;
126 self.time_passes = sess.time_passes();
127 self.inline_threshold = sess.opts.cg.inline_threshold;
128 self.obj_is_bitcode = sess.target.target.options.obj_is_bitcode ||
129 sess.opts.cg.linker_plugin_lto.enabled();
130 let embed_bitcode = sess.target.target.options.embed_bitcode ||
131 sess.opts.debugging_opts.embed_bitcode;
133 match sess.opts.optimize {
134 config::OptLevel::No |
135 config::OptLevel::Less => {
136 self.embed_bitcode_marker = embed_bitcode;
138 _ => self.embed_bitcode = embed_bitcode,
142 // Copy what clang does by turning on loop vectorization at O2 and
143 // slp vectorization at O3. Otherwise configure other optimization aspects
144 // of this pass manager builder.
145 // Turn off vectorization for emscripten, as it's not very well supported.
146 self.vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
147 (sess.opts.optimize == config::OptLevel::Default ||
148 sess.opts.optimize == config::OptLevel::Aggressive) &&
149 !sess.target.target.options.is_like_emscripten;
151 self.vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
152 sess.opts.optimize == config::OptLevel::Aggressive &&
153 !sess.target.target.options.is_like_emscripten;
155 // Some targets (namely, NVPTX) interact badly with the MergeFunctions
156 // pass. This is because MergeFunctions can generate new function calls
157 // which may interfere with the target calling convention; e.g. for the
158 // NVPTX target, PTX kernels should not call other PTX kernels.
159 // MergeFunctions can also be configured to generate aliases instead,
160 // but aliases are not supported by some backends (again, NVPTX).
161 // Therefore, allow targets to opt out of the MergeFunctions pass,
162 // but otherwise keep the pass enabled (at O2 and O3) since it can be
163 // useful for reducing code size.
164 self.merge_functions = match sess.opts.debugging_opts.merge_functions
165 .unwrap_or(sess.target.target.options.merge_functions) {
166 MergeFunctions::Disabled => false,
167 MergeFunctions::Trampolines |
168 MergeFunctions::Aliases => {
169 sess.opts.optimize == config::OptLevel::Default ||
170 sess.opts.optimize == config::OptLevel::Aggressive
175 pub fn bitcode_needed(&self) -> bool {
176 self.emit_bc || self.obj_is_bitcode
177 || self.emit_bc_compressed || self.embed_bitcode
181 /// Assembler name and command used by codegen when no_integrated_as is enabled
182 pub struct AssemblerCommand {
187 // HACK(eddyb) work around `#[derive]` producing wrong bounds for `Clone`.
188 pub struct TargetMachineFactory<B: WriteBackendMethods>(
189 pub Arc<dyn Fn() -> Result<B::TargetMachine, String> + Send + Sync>,
192 impl<B: WriteBackendMethods> Clone for TargetMachineFactory<B> {
193 fn clone(&self) -> Self {
194 TargetMachineFactory(self.0.clone())
198 /// Additional resources used by optimize_and_codegen (not module specific)
200 pub struct CodegenContext<B: WriteBackendMethods> {
201 // Resources needed when running LTO
203 pub time_passes: bool,
205 pub no_landing_pads: bool,
206 pub save_temps: bool,
207 pub fewer_names: bool,
208 pub exported_symbols: Option<Arc<ExportedSymbols>>,
209 pub opts: Arc<config::Options>,
210 pub crate_types: Vec<config::CrateType>,
211 pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>,
212 pub output_filenames: Arc<OutputFilenames>,
213 pub regular_module_config: Arc<ModuleConfig>,
214 pub metadata_module_config: Arc<ModuleConfig>,
215 pub allocator_module_config: Arc<ModuleConfig>,
216 pub tm_factory: TargetMachineFactory<B>,
217 pub msvc_imps_needed: bool,
218 pub target_pointer_width: String,
219 pub debuginfo: config::DebugInfo,
221 // Number of cgus excluding the allocator/metadata modules
222 pub total_cgus: usize,
223 // Handler to use for diagnostics produced during codegen.
224 pub diag_emitter: SharedEmitter,
225 // LLVM passes added by plugins.
226 pub plugin_passes: Vec<String>,
227 // LLVM optimizations for which we want to print remarks.
229 // Worker thread number
231 // The incremental compilation session directory, or None if we are not
232 // compiling incrementally
233 pub incr_comp_session_dir: Option<PathBuf>,
234 // Used to update CGU re-use information during the thinlto phase.
235 pub cgu_reuse_tracker: CguReuseTracker,
236 // Channel back to the main control thread to send messages to
237 pub coordinator_send: Sender<Box<dyn Any + Send>>,
238 // A reference to the TimeGraph so we can register timings. None means that
239 // measuring is disabled.
240 pub time_graph: Option<TimeGraph>,
241 // The assembler command if no_integrated_as option is enabled, None otherwise
242 pub assembler_cmd: Option<Arc<AssemblerCommand>>
245 impl<B: WriteBackendMethods> CodegenContext<B> {
246 pub fn create_diag_handler(&self) -> Handler {
247 Handler::with_emitter(true, false, Box::new(self.diag_emitter.clone()))
250 pub fn config(&self, kind: ModuleKind) -> &ModuleConfig {
252 ModuleKind::Regular => &self.regular_module_config,
253 ModuleKind::Metadata => &self.metadata_module_config,
254 ModuleKind::Allocator => &self.allocator_module_config,
259 fn generate_lto_work<B: ExtraBackendMethods>(
260 cgcx: &CodegenContext<B>,
261 needs_fat_lto: Vec<ModuleCodegen<B::Module>>,
262 needs_thin_lto: Vec<(String, B::ThinBuffer)>,
263 import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>
264 ) -> Vec<(WorkItem<B>, u64)> {
265 let mut timeline = cgcx.time_graph.as_ref().map(|tg| {
266 tg.start(CODEGEN_WORKER_TIMELINE,
267 CODEGEN_WORK_PACKAGE_KIND,
269 }).unwrap_or(Timeline::noop());
271 let (lto_modules, copy_jobs) = if !needs_fat_lto.is_empty() {
272 assert!(needs_thin_lto.is_empty());
273 assert!(import_only_modules.is_empty());
274 let lto_module = B::run_fat_lto(cgcx, needs_fat_lto, &mut timeline)
275 .unwrap_or_else(|e| e.raise());
276 (vec![lto_module], vec![])
278 assert!(needs_fat_lto.is_empty());
279 B::run_thin_lto(cgcx, needs_thin_lto, import_only_modules, &mut timeline)
280 .unwrap_or_else(|e| e.raise())
283 lto_modules.into_iter().map(|module| {
284 let cost = module.cost();
285 (WorkItem::LTO(module), cost)
286 }).chain(copy_jobs.into_iter().map(|wp| {
287 (WorkItem::CopyPostLtoArtifacts(CachedModuleCodegen {
288 name: wp.cgu_name.clone(),
294 pub struct CompiledModules {
295 pub modules: Vec<CompiledModule>,
296 pub metadata_module: CompiledModule,
297 pub allocator_module: Option<CompiledModule>,
300 fn need_crate_bitcode_for_rlib(sess: &Session) -> bool {
301 sess.crate_types.borrow().contains(&config::CrateType::Rlib) &&
302 sess.opts.output_types.contains_key(&OutputType::Exe)
305 fn need_pre_thin_lto_bitcode_for_incr_comp(sess: &Session) -> bool {
306 if sess.opts.incremental.is_none() {
314 Lto::ThinLocal => true,
318 pub fn start_async_codegen<B: ExtraBackendMethods>(
321 time_graph: Option<TimeGraph>,
322 metadata: EncodedMetadata,
323 coordinator_receive: Receiver<Box<dyn Any + Send>>,
325 ) -> OngoingCodegen<B> {
327 let crate_name = tcx.crate_name(LOCAL_CRATE);
328 let crate_hash = tcx.crate_hash(LOCAL_CRATE);
329 let no_builtins = attr::contains_name(&tcx.hir().krate().attrs, "no_builtins");
330 let subsystem = attr::first_attr_value_str_by_name(&tcx.hir().krate().attrs,
331 "windows_subsystem");
332 let windows_subsystem = subsystem.map(|subsystem| {
333 if subsystem != "windows" && subsystem != "console" {
334 tcx.sess.fatal(&format!("invalid windows subsystem `{}`, only \
335 `windows` and `console` are allowed",
338 subsystem.to_string()
341 let linker_info = LinkerInfo::new(tcx);
342 let crate_info = CrateInfo::new(tcx);
344 // Figure out what we actually need to build.
345 let mut modules_config = ModuleConfig::new(sess.opts.cg.passes.clone());
346 let mut metadata_config = ModuleConfig::new(vec![]);
347 let mut allocator_config = ModuleConfig::new(vec![]);
349 if let Some(ref sanitizer) = sess.opts.debugging_opts.sanitizer {
351 Sanitizer::Address => {
352 modules_config.passes.push("asan".to_owned());
353 modules_config.passes.push("asan-module".to_owned());
355 Sanitizer::Memory => {
356 modules_config.passes.push("msan".to_owned())
358 Sanitizer::Thread => {
359 modules_config.passes.push("tsan".to_owned())
365 if sess.opts.debugging_opts.profile {
366 modules_config.passes.push("insert-gcov-profiling".to_owned())
369 modules_config.pgo_gen = sess.opts.debugging_opts.pgo_gen.clone();
370 modules_config.pgo_use = sess.opts.debugging_opts.pgo_use.clone();
372 modules_config.opt_level = Some(sess.opts.optimize);
373 modules_config.opt_size = Some(sess.opts.optimize);
375 // Save all versions of the bytecode if we're saving our temporaries.
376 if sess.opts.cg.save_temps {
377 modules_config.emit_no_opt_bc = true;
378 modules_config.emit_pre_thin_lto_bc = true;
379 modules_config.emit_bc = true;
380 modules_config.emit_lto_bc = true;
381 metadata_config.emit_bc = true;
382 allocator_config.emit_bc = true;
385 // Emit compressed bitcode files for the crate if we're emitting an rlib.
386 // Whenever an rlib is created, the bitcode is inserted into the archive in
387 // order to allow LTO against it.
388 if need_crate_bitcode_for_rlib(sess) {
389 modules_config.emit_bc_compressed = true;
390 allocator_config.emit_bc_compressed = true;
393 modules_config.emit_pre_thin_lto_bc =
394 need_pre_thin_lto_bitcode_for_incr_comp(sess);
396 modules_config.no_integrated_as = tcx.sess.opts.cg.no_integrated_as ||
397 tcx.sess.target.target.options.no_integrated_as;
399 for output_type in sess.opts.output_types.keys() {
401 OutputType::Bitcode => { modules_config.emit_bc = true; }
402 OutputType::LlvmAssembly => { modules_config.emit_ir = true; }
403 OutputType::Assembly => {
404 modules_config.emit_asm = true;
405 // If we're not using the LLVM assembler, this function
406 // could be invoked specially with output_type_assembly, so
407 // in this case we still want the metadata object file.
408 if !sess.opts.output_types.contains_key(&OutputType::Assembly) {
409 metadata_config.emit_obj = true;
410 allocator_config.emit_obj = true;
413 OutputType::Object => { modules_config.emit_obj = true; }
414 OutputType::Metadata => { metadata_config.emit_obj = true; }
416 modules_config.emit_obj = true;
417 metadata_config.emit_obj = true;
418 allocator_config.emit_obj = true;
420 OutputType::Mir => {}
421 OutputType::DepInfo => {}
425 modules_config.set_flags(sess, no_builtins);
426 metadata_config.set_flags(sess, no_builtins);
427 allocator_config.set_flags(sess, no_builtins);
429 // Exclude metadata and allocator modules from time_passes output, since
430 // they throw off the "LLVM passes" measurement.
431 metadata_config.time_passes = false;
432 allocator_config.time_passes = false;
434 let (shared_emitter, shared_emitter_main) = SharedEmitter::new();
435 let (codegen_worker_send, codegen_worker_receive) = channel();
437 let coordinator_thread = start_executing_work(backend.clone(),
444 sess.jobserver.clone(),
446 Arc::new(modules_config),
447 Arc::new(metadata_config),
448 Arc::new(allocator_config));
460 coordinator_send: tcx.tx_to_llvm_workers.lock().clone(),
461 codegen_worker_receive,
463 future: coordinator_thread,
464 output_filenames: tcx.output_filenames(LOCAL_CRATE),
468 fn copy_all_cgu_workproducts_to_incr_comp_cache_dir(
470 compiled_modules: &CompiledModules,
471 ) -> FxHashMap<WorkProductId, WorkProduct> {
472 let mut work_products = FxHashMap::default();
474 if sess.opts.incremental.is_none() {
475 return work_products;
478 for module in compiled_modules.modules.iter().filter(|m| m.kind == ModuleKind::Regular) {
479 let mut files = vec![];
481 if let Some(ref path) = module.object {
482 files.push((WorkProductFileKind::Object, path.clone()));
484 if let Some(ref path) = module.bytecode {
485 files.push((WorkProductFileKind::Bytecode, path.clone()));
487 if let Some(ref path) = module.bytecode_compressed {
488 files.push((WorkProductFileKind::BytecodeCompressed, path.clone()));
491 if let Some((id, product)) =
492 copy_cgu_workproducts_to_incr_comp_cache_dir(sess, &module.name, &files) {
493 work_products.insert(id, product);
500 fn produce_final_output_artifacts(sess: &Session,
501 compiled_modules: &CompiledModules,
502 crate_output: &OutputFilenames) {
503 let mut user_wants_bitcode = false;
504 let mut user_wants_objects = false;
506 // Produce final compile outputs.
507 let copy_gracefully = |from: &Path, to: &Path| {
508 if let Err(e) = fs::copy(from, to) {
509 sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
513 let copy_if_one_unit = |output_type: OutputType,
514 keep_numbered: bool| {
515 if compiled_modules.modules.len() == 1 {
516 // 1) Only one codegen unit. In this case it's no difficulty
517 // to copy `foo.0.x` to `foo.x`.
518 let module_name = Some(&compiled_modules.modules[0].name[..]);
519 let path = crate_output.temp_path(output_type, module_name);
520 copy_gracefully(&path,
521 &crate_output.path(output_type));
522 if !sess.opts.cg.save_temps && !keep_numbered {
523 // The user just wants `foo.x`, not `foo.#module-name#.x`.
527 let ext = crate_output.temp_path(output_type, None)
534 if crate_output.outputs.contains_key(&output_type) {
535 // 2) Multiple codegen units, with `--emit foo=some_name`. We have
536 // no good solution for this case, so warn the user.
537 sess.warn(&format!("ignoring emit path because multiple .{} files \
538 were produced", ext));
539 } else if crate_output.single_output_file.is_some() {
540 // 3) Multiple codegen units, with `-o some_name`. We have
541 // no good solution for this case, so warn the user.
542 sess.warn(&format!("ignoring -o because multiple .{} files \
543 were produced", ext));
545 // 4) Multiple codegen units, but no explicit name. We
546 // just leave the `foo.0.x` files in place.
547 // (We don't have to do any work in this case.)
552 // Flag to indicate whether the user explicitly requested bitcode.
553 // Otherwise, we produced it only as a temporary output, and will need
555 for output_type in crate_output.outputs.keys() {
557 OutputType::Bitcode => {
558 user_wants_bitcode = true;
559 // Copy to .bc, but always keep the .0.bc. There is a later
560 // check to figure out if we should delete .0.bc files, or keep
561 // them for making an rlib.
562 copy_if_one_unit(OutputType::Bitcode, true);
564 OutputType::LlvmAssembly => {
565 copy_if_one_unit(OutputType::LlvmAssembly, false);
567 OutputType::Assembly => {
568 copy_if_one_unit(OutputType::Assembly, false);
570 OutputType::Object => {
571 user_wants_objects = true;
572 copy_if_one_unit(OutputType::Object, true);
575 OutputType::Metadata |
577 OutputType::DepInfo => {}
581 // Clean up unwanted temporary files.
583 // We create the following files by default:
584 // - #crate#.#module-name#.bc
585 // - #crate#.#module-name#.o
586 // - #crate#.crate.metadata.bc
587 // - #crate#.crate.metadata.o
588 // - #crate#.o (linked from crate.##.o)
589 // - #crate#.bc (copied from crate.##.bc)
590 // We may create additional files if requested by the user (through
591 // `-C save-temps` or `--emit=` flags).
593 if !sess.opts.cg.save_temps {
594 // Remove the temporary .#module-name#.o objects. If the user didn't
595 // explicitly request bitcode (with --emit=bc), and the bitcode is not
596 // needed for building an rlib, then we must remove .#module-name#.bc as
599 // Specific rules for keeping .#module-name#.bc:
600 // - If the user requested bitcode (`user_wants_bitcode`), and
601 // codegen_units > 1, then keep it.
602 // - If the user requested bitcode but codegen_units == 1, then we
603 // can toss .#module-name#.bc because we copied it to .bc earlier.
604 // - If we're not building an rlib and the user didn't request
605 // bitcode, then delete .#module-name#.bc.
606 // If you change how this works, also update back::link::link_rlib,
607 // where .#module-name#.bc files are (maybe) deleted after making an
609 let needs_crate_object = crate_output.outputs.contains_key(&OutputType::Exe);
611 let keep_numbered_bitcode = user_wants_bitcode && sess.codegen_units() > 1;
613 let keep_numbered_objects = needs_crate_object ||
614 (user_wants_objects && sess.codegen_units() > 1);
616 for module in compiled_modules.modules.iter() {
617 if let Some(ref path) = module.object {
618 if !keep_numbered_objects {
623 if let Some(ref path) = module.bytecode {
624 if !keep_numbered_bitcode {
630 if !user_wants_bitcode {
631 if let Some(ref path) = compiled_modules.metadata_module.bytecode {
635 if let Some(ref allocator_module) = compiled_modules.allocator_module {
636 if let Some(ref path) = allocator_module.bytecode {
643 // We leave the following files around by default:
645 // - #crate#.crate.metadata.o
647 // These are used in linking steps and will be cleaned up afterward.
650 pub fn dump_incremental_data(_codegen_results: &CodegenResults) {
651 // FIXME(mw): This does not work at the moment because the situation has
652 // become more complicated due to incremental LTO. Now a CGU
653 // can have more than two caching states.
654 // println!("[incremental] Re-using {} out of {} modules",
655 // codegen_results.modules.iter().filter(|m| m.pre_existing).count(),
656 // codegen_results.modules.len());
659 pub enum WorkItem<B: WriteBackendMethods> {
660 /// Optimize a newly codegened, totally unoptimized module.
661 Optimize(ModuleCodegen<B::Module>),
662 /// Copy the post-LTO artifacts from the incremental cache to the output
664 CopyPostLtoArtifacts(CachedModuleCodegen),
665 /// Performs (Thin)LTO on the given module.
666 LTO(lto::LtoModuleCodegen<B>),
669 impl<B: WriteBackendMethods> WorkItem<B> {
670 pub fn module_kind(&self) -> ModuleKind {
672 WorkItem::Optimize(ref m) => m.kind,
673 WorkItem::CopyPostLtoArtifacts(_) |
674 WorkItem::LTO(_) => ModuleKind::Regular,
678 pub fn name(&self) -> String {
680 WorkItem::Optimize(ref m) => format!("optimize: {}", m.name),
681 WorkItem::CopyPostLtoArtifacts(ref m) => format!("copy post LTO artifacts: {}", m.name),
682 WorkItem::LTO(ref m) => format!("lto: {}", m.name()),
687 enum WorkItemResult<B: WriteBackendMethods> {
688 Compiled(CompiledModule),
689 NeedsFatLTO(ModuleCodegen<B::Module>),
690 NeedsThinLTO(String, B::ThinBuffer),
693 fn execute_work_item<B: ExtraBackendMethods>(
694 cgcx: &CodegenContext<B>,
695 work_item: WorkItem<B>,
696 timeline: &mut Timeline
697 ) -> Result<WorkItemResult<B>, FatalError> {
698 let module_config = cgcx.config(work_item.module_kind());
701 WorkItem::Optimize(module) => {
702 execute_optimize_work_item(cgcx, module, module_config, timeline)
704 WorkItem::CopyPostLtoArtifacts(module) => {
705 execute_copy_from_cache_work_item(cgcx, module, module_config, timeline)
707 WorkItem::LTO(module) => {
708 execute_lto_work_item(cgcx, module, module_config, timeline)
713 // Actual LTO type we end up chosing based on multiple factors.
714 enum ComputedLtoType {
720 fn execute_optimize_work_item<B: ExtraBackendMethods>(
721 cgcx: &CodegenContext<B>,
722 module: ModuleCodegen<B::Module>,
723 module_config: &ModuleConfig,
724 timeline: &mut Timeline
725 ) -> Result<WorkItemResult<B>, FatalError> {
726 let diag_handler = cgcx.create_diag_handler();
729 B::optimize(cgcx, &diag_handler, &module, module_config, timeline)?;
732 // After we've done the initial round of optimizations we need to
733 // decide whether to synchronously codegen this module or ship it
734 // back to the coordinator thread for further LTO processing (which
735 // has to wait for all the initial modules to be optimized).
737 // If the linker does LTO, we don't have to do it. Note that we
738 // keep doing full LTO, if it is requested, as not to break the
739 // assumption that the output will be a single module.
740 let linker_does_lto = cgcx.opts.cg.linker_plugin_lto.enabled();
742 // When we're automatically doing ThinLTO for multi-codegen-unit
743 // builds we don't actually want to LTO the allocator modules if
744 // it shows up. This is due to various linker shenanigans that
745 // we'll encounter later.
746 let is_allocator = module.kind == ModuleKind::Allocator;
748 // We ignore a request for full crate grath LTO if the cate type
749 // is only an rlib, as there is no full crate graph to process,
750 // that'll happen later.
752 // This use case currently comes up primarily for targets that
753 // require LTO so the request for LTO is always unconditionally
754 // passed down to the backend, but we don't actually want to do
755 // anything about it yet until we've got a final product.
756 let is_rlib = cgcx.crate_types.len() == 1
757 && cgcx.crate_types[0] == config::CrateType::Rlib;
759 // Metadata modules never participate in LTO regardless of the lto
761 let lto_type = if module.kind == ModuleKind::Metadata {
765 Lto::ThinLocal if !linker_does_lto && !is_allocator
766 => ComputedLtoType::Thin,
767 Lto::Thin if !linker_does_lto && !is_rlib
768 => ComputedLtoType::Thin,
769 Lto::Fat if !is_rlib => ComputedLtoType::Fat,
770 _ => ComputedLtoType::No,
775 ComputedLtoType::No => {
776 let module = unsafe {
777 B::codegen(cgcx, &diag_handler, module, module_config, timeline)?
779 WorkItemResult::Compiled(module)
781 ComputedLtoType::Thin => {
782 let (name, thin_buffer) = B::prepare_thin(cgcx, module);
783 WorkItemResult::NeedsThinLTO(name, thin_buffer)
785 ComputedLtoType::Fat => WorkItemResult::NeedsFatLTO(module),
789 fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>(
790 cgcx: &CodegenContext<B>,
791 module: CachedModuleCodegen,
792 module_config: &ModuleConfig,
794 ) -> Result<WorkItemResult<B>, FatalError> {
795 let incr_comp_session_dir = cgcx.incr_comp_session_dir
798 let mut object = None;
799 let mut bytecode = None;
800 let mut bytecode_compressed = None;
801 for (kind, saved_file) in &module.source.saved_files {
802 let obj_out = match kind {
803 WorkProductFileKind::Object => {
804 let path = cgcx.output_filenames.temp_path(OutputType::Object,
806 object = Some(path.clone());
809 WorkProductFileKind::Bytecode => {
810 let path = cgcx.output_filenames.temp_path(OutputType::Bitcode,
812 bytecode = Some(path.clone());
815 WorkProductFileKind::BytecodeCompressed => {
816 let path = cgcx.output_filenames.temp_path(OutputType::Bitcode,
818 .with_extension(RLIB_BYTECODE_EXTENSION);
819 bytecode_compressed = Some(path.clone());
823 let source_file = in_incr_comp_dir(&incr_comp_session_dir,
825 debug!("copying pre-existing module `{}` from {:?} to {}",
829 if let Err(err) = link_or_copy(&source_file, &obj_out) {
830 let diag_handler = cgcx.create_diag_handler();
831 diag_handler.err(&format!("unable to copy {} to {}: {}",
832 source_file.display(),
838 assert_eq!(object.is_some(), module_config.emit_obj);
839 assert_eq!(bytecode.is_some(), module_config.emit_bc);
840 assert_eq!(bytecode_compressed.is_some(), module_config.emit_bc_compressed);
842 Ok(WorkItemResult::Compiled(CompiledModule {
844 kind: ModuleKind::Regular,
851 fn execute_lto_work_item<B: ExtraBackendMethods>(
852 cgcx: &CodegenContext<B>,
853 mut module: lto::LtoModuleCodegen<B>,
854 module_config: &ModuleConfig,
855 timeline: &mut Timeline
856 ) -> Result<WorkItemResult<B>, FatalError> {
857 let diag_handler = cgcx.create_diag_handler();
860 let module = module.optimize(cgcx, timeline)?;
861 let module = B::codegen(cgcx, &diag_handler, module, module_config, timeline)?;
862 Ok(WorkItemResult::Compiled(module))
866 pub enum Message<B: WriteBackendMethods> {
867 Token(io::Result<Acquired>),
869 result: ModuleCodegen<B::Module>,
874 thin_buffer: B::ThinBuffer,
878 result: Result<CompiledModule, ()>,
882 llvm_work_item: WorkItem<B>,
885 AddImportOnlyModule {
886 module_data: SerializedModule<B::ModuleBuffer>,
887 work_product: WorkProduct,
896 code: Option<DiagnosticId>,
900 #[derive(PartialEq, Clone, Copy, Debug)]
901 enum MainThreadWorkerState {
907 fn start_executing_work<B: ExtraBackendMethods>(
910 crate_info: &CrateInfo,
911 shared_emitter: SharedEmitter,
912 codegen_worker_send: Sender<Message<B>>,
913 coordinator_receive: Receiver<Box<dyn Any + Send>>,
916 time_graph: Option<TimeGraph>,
917 modules_config: Arc<ModuleConfig>,
918 metadata_config: Arc<ModuleConfig>,
919 allocator_config: Arc<ModuleConfig>
920 ) -> thread::JoinHandle<Result<CompiledModules, ()>> {
921 let coordinator_send = tcx.tx_to_llvm_workers.lock().clone();
924 // Compute the set of symbols we need to retain when doing LTO (if we need to)
925 let exported_symbols = {
926 let mut exported_symbols = FxHashMap::default();
928 let copy_symbols = |cnum| {
929 let symbols = tcx.exported_symbols(cnum)
931 .map(|&(s, lvl)| (s.symbol_name(tcx).to_string(), lvl))
939 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
940 Some(Arc::new(exported_symbols))
942 Lto::Fat | Lto::Thin => {
943 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
944 for &cnum in tcx.crates().iter() {
945 exported_symbols.insert(cnum, copy_symbols(cnum));
947 Some(Arc::new(exported_symbols))
952 // First up, convert our jobserver into a helper thread so we can use normal
953 // mpsc channels to manage our messages and such.
954 // After we've requested tokens then we'll, when we can,
955 // get tokens on `coordinator_receive` which will
956 // get managed in the main loop below.
957 let coordinator_send2 = coordinator_send.clone();
958 let helper = jobserver.into_helper_thread(move |token| {
959 drop(coordinator_send2.send(Box::new(Message::Token::<B>(token))));
960 }).expect("failed to spawn helper thread");
962 let mut each_linked_rlib_for_lto = Vec::new();
963 drop(link::each_linked_rlib(sess, crate_info, &mut |cnum, path| {
964 if link::ignored_for_lto(sess, crate_info, cnum) {
967 each_linked_rlib_for_lto.push((cnum, path.to_path_buf()));
970 let assembler_cmd = if modules_config.no_integrated_as {
971 // HACK: currently we use linker (gcc) as our assembler
972 let (linker, flavor) = link::linker_and_flavor(sess);
974 let (name, mut cmd) = get_linker(sess, &linker, flavor);
975 cmd.args(&sess.target.target.options.asm_args);
976 Some(Arc::new(AssemblerCommand {
984 let ol = tcx.backend_optimization_level(LOCAL_CRATE);
985 let cgcx = CodegenContext::<B> {
986 backend: backend.clone(),
987 crate_types: sess.crate_types.borrow().clone(),
988 each_linked_rlib_for_lto,
990 no_landing_pads: sess.no_landing_pads(),
991 fewer_names: sess.fewer_names(),
992 save_temps: sess.opts.cg.save_temps,
993 opts: Arc::new(sess.opts.clone()),
994 time_passes: sess.time_passes(),
996 plugin_passes: sess.plugin_llvm_passes.borrow().clone(),
997 remark: sess.opts.cg.remark.clone(),
999 incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
1000 cgu_reuse_tracker: sess.cgu_reuse_tracker.clone(),
1002 diag_emitter: shared_emitter.clone(),
1004 output_filenames: tcx.output_filenames(LOCAL_CRATE),
1005 regular_module_config: modules_config,
1006 metadata_module_config: metadata_config,
1007 allocator_module_config: allocator_config,
1008 tm_factory: TargetMachineFactory(backend.target_machine_factory(tcx.sess, ol, false)),
1010 msvc_imps_needed: msvc_imps_needed(tcx),
1011 target_pointer_width: tcx.sess.target.target.target_pointer_width.clone(),
1012 debuginfo: tcx.sess.opts.debuginfo,
1016 // This is the "main loop" of parallel work happening for parallel codegen.
1017 // It's here that we manage parallelism, schedule work, and work with
1018 // messages coming from clients.
1020 // There are a few environmental pre-conditions that shape how the system
1023 // - Error reporting only can happen on the main thread because that's the
1024 // only place where we have access to the compiler `Session`.
1025 // - LLVM work can be done on any thread.
1026 // - Codegen can only happen on the main thread.
1027 // - Each thread doing substantial work most be in possession of a `Token`
1028 // from the `Jobserver`.
1029 // - The compiler process always holds one `Token`. Any additional `Tokens`
1030 // have to be requested from the `Jobserver`.
1034 // The error reporting restriction is handled separately from the rest: We
1035 // set up a `SharedEmitter` the holds an open channel to the main thread.
1036 // When an error occurs on any thread, the shared emitter will send the
1037 // error message to the receiver main thread (`SharedEmitterMain`). The
1038 // main thread will periodically query this error message queue and emit
1039 // any error messages it has received. It might even abort compilation if
1040 // has received a fatal error. In this case we rely on all other threads
1041 // being torn down automatically with the main thread.
1042 // Since the main thread will often be busy doing codegen work, error
1043 // reporting will be somewhat delayed, since the message queue can only be
1044 // checked in between to work packages.
1046 // Work Processing Infrastructure
1047 // ==============================
1048 // The work processing infrastructure knows three major actors:
1050 // - the coordinator thread,
1051 // - the main thread, and
1052 // - LLVM worker threads
1054 // The coordinator thread is running a message loop. It instructs the main
1055 // thread about what work to do when, and it will spawn off LLVM worker
1056 // threads as open LLVM WorkItems become available.
1058 // The job of the main thread is to codegen CGUs into LLVM work package
1059 // (since the main thread is the only thread that can do this). The main
1060 // thread will block until it receives a message from the coordinator, upon
1061 // which it will codegen one CGU, send it to the coordinator and block
1062 // again. This way the coordinator can control what the main thread is
1065 // The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is
1066 // available, it will spawn off a new LLVM worker thread and let it process
1067 // that a WorkItem. When a LLVM worker thread is done with its WorkItem,
1068 // it will just shut down, which also frees all resources associated with
1069 // the given LLVM module, and sends a message to the coordinator that the
1070 // has been completed.
1074 // The scheduler's goal is to minimize the time it takes to complete all
1075 // work there is, however, we also want to keep memory consumption low
1076 // if possible. These two goals are at odds with each other: If memory
1077 // consumption were not an issue, we could just let the main thread produce
1078 // LLVM WorkItems at full speed, assuring maximal utilization of
1079 // Tokens/LLVM worker threads. However, since codegen usual is faster
1080 // than LLVM processing, the queue of LLVM WorkItems would fill up and each
1081 // WorkItem potentially holds on to a substantial amount of memory.
1083 // So the actual goal is to always produce just enough LLVM WorkItems as
1084 // not to starve our LLVM worker threads. That means, once we have enough
1085 // WorkItems in our queue, we can block the main thread, so it does not
1086 // produce more until we need them.
1088 // Doing LLVM Work on the Main Thread
1089 // ----------------------------------
1090 // Since the main thread owns the compiler processes implicit `Token`, it is
1091 // wasteful to keep it blocked without doing any work. Therefore, what we do
1092 // in this case is: We spawn off an additional LLVM worker thread that helps
1093 // reduce the queue. The work it is doing corresponds to the implicit
1094 // `Token`. The coordinator will mark the main thread as being busy with
1095 // LLVM work. (The actual work happens on another OS thread but we just care
1096 // about `Tokens`, not actual threads).
1098 // When any LLVM worker thread finishes while the main thread is marked as
1099 // "busy with LLVM work", we can do a little switcheroo: We give the Token
1100 // of the just finished thread to the LLVM worker thread that is working on
1101 // behalf of the main thread's implicit Token, thus freeing up the main
1102 // thread again. The coordinator can then again decide what the main thread
1103 // should do. This allows the coordinator to make decisions at more points
1106 // Striking a Balance between Throughput and Memory Consumption
1107 // ------------------------------------------------------------
1108 // Since our two goals, (1) use as many Tokens as possible and (2) keep
1109 // memory consumption as low as possible, are in conflict with each other,
1110 // we have to find a trade off between them. Right now, the goal is to keep
1111 // all workers busy, which means that no worker should find the queue empty
1112 // when it is ready to start.
1113 // How do we do achieve this? Good question :) We actually never know how
1114 // many `Tokens` are potentially available so it's hard to say how much to
1115 // fill up the queue before switching the main thread to LLVM work. Also we
1116 // currently don't have a means to estimate how long a running LLVM worker
1117 // will still be busy with it's current WorkItem. However, we know the
1118 // maximal count of available Tokens that makes sense (=the number of CPU
1119 // cores), so we can take a conservative guess. The heuristic we use here
1120 // is implemented in the `queue_full_enough()` function.
1122 // Some Background on Jobservers
1123 // -----------------------------
1124 // It's worth also touching on the management of parallelism here. We don't
1125 // want to just spawn a thread per work item because while that's optimal
1126 // parallelism it may overload a system with too many threads or violate our
1127 // configuration for the maximum amount of cpu to use for this process. To
1128 // manage this we use the `jobserver` crate.
1130 // Job servers are an artifact of GNU make and are used to manage
1131 // parallelism between processes. A jobserver is a glorified IPC semaphore
1132 // basically. Whenever we want to run some work we acquire the semaphore,
1133 // and whenever we're done with that work we release the semaphore. In this
1134 // manner we can ensure that the maximum number of parallel workers is
1135 // capped at any one point in time.
1137 // LTO and the coordinator thread
1138 // ------------------------------
1140 // The final job the coordinator thread is responsible for is managing LTO
1141 // and how that works. When LTO is requested what we'll to is collect all
1142 // optimized LLVM modules into a local vector on the coordinator. Once all
1143 // modules have been codegened and optimized we hand this to the `lto`
1144 // module for further optimization. The `lto` module will return back a list
1145 // of more modules to work on, which the coordinator will continue to spawn
1148 // Each LLVM module is automatically sent back to the coordinator for LTO if
1149 // necessary. There's already optimizations in place to avoid sending work
1150 // back to the coordinator if LTO isn't requested.
1151 return thread::spawn(move || {
1152 // We pretend to be within the top-level LLVM time-passes task here:
1155 let max_workers = ::num_cpus::get();
1156 let mut worker_id_counter = 0;
1157 let mut free_worker_ids = Vec::new();
1158 let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
1159 if let Some(id) = free_worker_ids.pop() {
1162 let id = worker_id_counter;
1163 worker_id_counter += 1;
1168 // This is where we collect codegen units that have gone all the way
1169 // through codegen and LLVM.
1170 let mut compiled_modules = vec![];
1171 let mut compiled_metadata_module = None;
1172 let mut compiled_allocator_module = None;
1173 let mut needs_fat_lto = Vec::new();
1174 let mut needs_thin_lto = Vec::new();
1175 let mut lto_import_only_modules = Vec::new();
1176 let mut started_lto = false;
1177 let mut codegen_aborted = false;
1179 // This flag tracks whether all items have gone through codegens
1180 let mut codegen_done = false;
1182 // This is the queue of LLVM work items that still need processing.
1183 let mut work_items = Vec::<(WorkItem<B>, u64)>::new();
1185 // This are the Jobserver Tokens we currently hold. Does not include
1186 // the implicit Token the compiler process owns no matter what.
1187 let mut tokens = Vec::new();
1189 let mut main_thread_worker_state = MainThreadWorkerState::Idle;
1190 let mut running = 0;
1192 let mut llvm_start_time = None;
1194 // Run the message loop while there's still anything that needs message
1195 // processing. Note that as soon as codegen is aborted we simply want to
1196 // wait for all existing work to finish, so many of the conditions here
1197 // only apply if codegen hasn't been aborted as they represent pending
1199 while !codegen_done ||
1201 (!codegen_aborted && (
1202 work_items.len() > 0 ||
1203 needs_fat_lto.len() > 0 ||
1204 needs_thin_lto.len() > 0 ||
1205 lto_import_only_modules.len() > 0 ||
1206 main_thread_worker_state != MainThreadWorkerState::Idle
1210 // While there are still CGUs to be codegened, the coordinator has
1211 // to decide how to utilize the compiler processes implicit Token:
1212 // For codegenning more CGU or for running them through LLVM.
1214 if main_thread_worker_state == MainThreadWorkerState::Idle {
1215 if !queue_full_enough(work_items.len(), running, max_workers) {
1216 // The queue is not full enough, codegen more items:
1217 if let Err(_) = codegen_worker_send.send(Message::CodegenItem) {
1218 panic!("Could not send Message::CodegenItem to main thread")
1220 main_thread_worker_state = MainThreadWorkerState::Codegenning;
1222 // The queue is full enough to not let the worker
1223 // threads starve. Use the implicit Token to do some
1225 let (item, _) = work_items.pop()
1226 .expect("queue empty - queue_full_enough() broken?");
1227 let cgcx = CodegenContext {
1228 worker: get_worker_id(&mut free_worker_ids),
1231 maybe_start_llvm_timer(cgcx.config(item.module_kind()),
1232 &mut llvm_start_time);
1233 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1234 spawn_work(cgcx, item);
1237 } else if codegen_aborted {
1238 // don't queue up any more work if codegen was aborted, we're
1239 // just waiting for our existing children to finish
1241 // If we've finished everything related to normal codegen
1242 // then it must be the case that we've got some LTO work to do.
1243 // Perform the serial work here of figuring out what we're
1244 // going to LTO and then push a bunch of work items onto our
1246 if work_items.len() == 0 &&
1248 main_thread_worker_state == MainThreadWorkerState::Idle {
1249 assert!(!started_lto);
1253 mem::replace(&mut needs_fat_lto, Vec::new());
1254 let needs_thin_lto =
1255 mem::replace(&mut needs_thin_lto, Vec::new());
1256 let import_only_modules =
1257 mem::replace(&mut lto_import_only_modules, Vec::new());
1259 for (work, cost) in generate_lto_work(&cgcx, needs_fat_lto,
1260 needs_thin_lto, import_only_modules) {
1261 let insertion_index = work_items
1262 .binary_search_by_key(&cost, |&(_, cost)| cost)
1263 .unwrap_or_else(|e| e);
1264 work_items.insert(insertion_index, (work, cost));
1265 if !cgcx.opts.debugging_opts.no_parallel_llvm {
1266 helper.request_token();
1271 // In this branch, we know that everything has been codegened,
1272 // so it's just a matter of determining whether the implicit
1273 // Token is free to use for LLVM work.
1274 match main_thread_worker_state {
1275 MainThreadWorkerState::Idle => {
1276 if let Some((item, _)) = work_items.pop() {
1277 let cgcx = CodegenContext {
1278 worker: get_worker_id(&mut free_worker_ids),
1281 maybe_start_llvm_timer(cgcx.config(item.module_kind()),
1282 &mut llvm_start_time);
1283 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1284 spawn_work(cgcx, item);
1286 // There is no unstarted work, so let the main thread
1287 // take over for a running worker. Otherwise the
1288 // implicit token would just go to waste.
1289 // We reduce the `running` counter by one. The
1290 // `tokens.truncate()` below will take care of
1291 // giving the Token back.
1292 debug_assert!(running > 0);
1294 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1297 MainThreadWorkerState::Codegenning => {
1298 bug!("codegen worker should not be codegenning after \
1299 codegen was already completed")
1301 MainThreadWorkerState::LLVMing => {
1302 // Already making good use of that token
1307 // Spin up what work we can, only doing this while we've got available
1308 // parallelism slots and work left to spawn.
1309 while !codegen_aborted && work_items.len() > 0 && running < tokens.len() {
1310 let (item, _) = work_items.pop().unwrap();
1312 maybe_start_llvm_timer(cgcx.config(item.module_kind()),
1313 &mut llvm_start_time);
1315 let cgcx = CodegenContext {
1316 worker: get_worker_id(&mut free_worker_ids),
1320 spawn_work(cgcx, item);
1324 // Relinquish accidentally acquired extra tokens
1325 tokens.truncate(running);
1327 // If a thread exits successfully then we drop a token associated
1328 // with that worker and update our `running` count. We may later
1329 // re-acquire a token to continue running more work. We may also not
1330 // actually drop a token here if the worker was running with an
1331 // "ephemeral token"
1332 let mut free_worker = |worker_id| {
1333 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
1334 main_thread_worker_state = MainThreadWorkerState::Idle;
1339 free_worker_ids.push(worker_id);
1342 let msg = coordinator_receive.recv().unwrap();
1343 match *msg.downcast::<Message<B>>().ok().unwrap() {
1344 // Save the token locally and the next turn of the loop will use
1345 // this to spawn a new unit of work, or it may get dropped
1346 // immediately if we have no more work to spawn.
1347 Message::Token(token) => {
1352 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
1353 // If the main thread token is used for LLVM work
1354 // at the moment, we turn that thread into a regular
1355 // LLVM worker thread, so the main thread is free
1356 // to react to codegen demand.
1357 main_thread_worker_state = MainThreadWorkerState::Idle;
1362 let msg = &format!("failed to acquire jobserver token: {}", e);
1363 shared_emitter.fatal(msg);
1364 // Exit the coordinator thread
1370 Message::CodegenDone { llvm_work_item, cost } => {
1371 // We keep the queue sorted by estimated processing cost,
1372 // so that more expensive items are processed earlier. This
1373 // is good for throughput as it gives the main thread more
1374 // time to fill up the queue and it avoids scheduling
1375 // expensive items to the end.
1376 // Note, however, that this is not ideal for memory
1377 // consumption, as LLVM module sizes are not evenly
1379 let insertion_index =
1380 work_items.binary_search_by_key(&cost, |&(_, cost)| cost);
1381 let insertion_index = match insertion_index {
1382 Ok(idx) | Err(idx) => idx
1384 work_items.insert(insertion_index, (llvm_work_item, cost));
1386 if !cgcx.opts.debugging_opts.no_parallel_llvm {
1387 helper.request_token();
1389 assert!(!codegen_aborted);
1390 assert_eq!(main_thread_worker_state,
1391 MainThreadWorkerState::Codegenning);
1392 main_thread_worker_state = MainThreadWorkerState::Idle;
1395 Message::CodegenComplete => {
1396 codegen_done = true;
1397 assert!(!codegen_aborted);
1398 assert_eq!(main_thread_worker_state,
1399 MainThreadWorkerState::Codegenning);
1400 main_thread_worker_state = MainThreadWorkerState::Idle;
1403 // If codegen is aborted that means translation was aborted due
1404 // to some normal-ish compiler error. In this situation we want
1405 // to exit as soon as possible, but we want to make sure all
1406 // existing work has finished. Flag codegen as being done, and
1407 // then conditions above will ensure no more work is spawned but
1408 // we'll keep executing this loop until `running` hits 0.
1409 Message::CodegenAborted => {
1410 assert!(!codegen_aborted);
1411 codegen_done = true;
1412 codegen_aborted = true;
1413 assert_eq!(main_thread_worker_state,
1414 MainThreadWorkerState::Codegenning);
1416 Message::Done { result: Ok(compiled_module), worker_id } => {
1417 free_worker(worker_id);
1418 match compiled_module.kind {
1419 ModuleKind::Regular => {
1420 compiled_modules.push(compiled_module);
1422 ModuleKind::Metadata => {
1423 assert!(compiled_metadata_module.is_none());
1424 compiled_metadata_module = Some(compiled_module);
1426 ModuleKind::Allocator => {
1427 assert!(compiled_allocator_module.is_none());
1428 compiled_allocator_module = Some(compiled_module);
1432 Message::NeedsFatLTO { result, worker_id } => {
1433 assert!(!started_lto);
1434 free_worker(worker_id);
1435 needs_fat_lto.push(result);
1437 Message::NeedsThinLTO { name, thin_buffer, worker_id } => {
1438 assert!(!started_lto);
1439 free_worker(worker_id);
1440 needs_thin_lto.push((name, thin_buffer));
1442 Message::AddImportOnlyModule { module_data, work_product } => {
1443 assert!(!started_lto);
1444 assert!(!codegen_done);
1445 assert_eq!(main_thread_worker_state,
1446 MainThreadWorkerState::Codegenning);
1447 lto_import_only_modules.push((module_data, work_product));
1448 main_thread_worker_state = MainThreadWorkerState::Idle;
1450 // If the thread failed that means it panicked, so we abort immediately.
1451 Message::Done { result: Err(()), worker_id: _ } => {
1452 bug!("worker thread panicked");
1454 Message::CodegenItem => {
1455 bug!("the coordinator should not receive codegen requests")
1460 if let Some(llvm_start_time) = llvm_start_time {
1461 let total_llvm_time = Instant::now().duration_since(llvm_start_time);
1462 // This is the top-level timing for all of LLVM, set the time-depth
1465 print_time_passes_entry(cgcx.time_passes,
1470 // Regardless of what order these modules completed in, report them to
1471 // the backend in the same order every time to ensure that we're handing
1472 // out deterministic results.
1473 compiled_modules.sort_by(|a, b| a.name.cmp(&b.name));
1475 let compiled_metadata_module = compiled_metadata_module
1476 .expect("Metadata module not compiled?");
1478 Ok(CompiledModules {
1479 modules: compiled_modules,
1480 metadata_module: compiled_metadata_module,
1481 allocator_module: compiled_allocator_module,
1485 // A heuristic that determines if we have enough LLVM WorkItems in the
1486 // queue so that the main thread can do LLVM work instead of codegen
1487 fn queue_full_enough(items_in_queue: usize,
1488 workers_running: usize,
1489 max_workers: usize) -> bool {
1491 items_in_queue > 0 &&
1492 items_in_queue >= max_workers.saturating_sub(workers_running / 2)
1495 fn maybe_start_llvm_timer(config: &ModuleConfig,
1496 llvm_start_time: &mut Option<Instant>) {
1497 // We keep track of the -Ztime-passes output manually,
1498 // since the closure-based interface does not fit well here.
1499 if config.time_passes {
1500 if llvm_start_time.is_none() {
1501 *llvm_start_time = Some(Instant::now());
1507 pub const CODEGEN_WORKER_ID: usize = ::std::usize::MAX;
1508 pub const CODEGEN_WORKER_TIMELINE: time_graph::TimelineId =
1509 time_graph::TimelineId(CODEGEN_WORKER_ID);
1510 pub const CODEGEN_WORK_PACKAGE_KIND: time_graph::WorkPackageKind =
1511 time_graph::WorkPackageKind(&["#DE9597", "#FED1D3", "#FDC5C7", "#B46668", "#88494B"]);
1512 const LLVM_WORK_PACKAGE_KIND: time_graph::WorkPackageKind =
1513 time_graph::WorkPackageKind(&["#7DB67A", "#C6EEC4", "#ACDAAA", "#579354", "#3E6F3C"]);
1515 fn spawn_work<B: ExtraBackendMethods>(
1516 cgcx: CodegenContext<B>,
1519 let depth = time_depth();
1521 thread::spawn(move || {
1522 set_time_depth(depth);
1524 // Set up a destructor which will fire off a message that we're done as
1526 struct Bomb<B: ExtraBackendMethods> {
1527 coordinator_send: Sender<Box<dyn Any + Send>>,
1528 result: Option<WorkItemResult<B>>,
1531 impl<B: ExtraBackendMethods> Drop for Bomb<B> {
1532 fn drop(&mut self) {
1533 let worker_id = self.worker_id;
1534 let msg = match self.result.take() {
1535 Some(WorkItemResult::Compiled(m)) => {
1536 Message::Done::<B> { result: Ok(m), worker_id }
1538 Some(WorkItemResult::NeedsFatLTO(m)) => {
1539 Message::NeedsFatLTO::<B> { result: m, worker_id }
1541 Some(WorkItemResult::NeedsThinLTO(name, thin_buffer)) => {
1542 Message::NeedsThinLTO::<B> { name, thin_buffer, worker_id }
1544 None => Message::Done::<B> { result: Err(()), worker_id }
1546 drop(self.coordinator_send.send(Box::new(msg)));
1550 let mut bomb = Bomb::<B> {
1551 coordinator_send: cgcx.coordinator_send.clone(),
1553 worker_id: cgcx.worker,
1556 // Execute the work itself, and if it finishes successfully then flag
1557 // ourselves as a success as well.
1559 // Note that we ignore any `FatalError` coming out of `execute_work_item`,
1560 // as a diagnostic was already sent off to the main thread - just
1561 // surface that there was an error in this worker.
1563 let timeline = cgcx.time_graph.as_ref().map(|tg| {
1564 tg.start(time_graph::TimelineId(cgcx.worker),
1565 LLVM_WORK_PACKAGE_KIND,
1568 let mut timeline = timeline.unwrap_or(Timeline::noop());
1569 execute_work_item(&cgcx, work, &mut timeline).ok()
1574 pub fn run_assembler<B: ExtraBackendMethods>(
1575 cgcx: &CodegenContext<B>,
1580 let assembler = cgcx.assembler_cmd
1582 .expect("cgcx.assembler_cmd is missing?");
1584 let pname = &assembler.name;
1585 let mut cmd = assembler.cmd.clone();
1586 cmd.arg("-c").arg("-o").arg(object).arg(assembly);
1587 debug!("{:?}", cmd);
1589 match cmd.output() {
1591 if !prog.status.success() {
1592 let mut note = prog.stderr.clone();
1593 note.extend_from_slice(&prog.stdout);
1595 handler.struct_err(&format!("linking with `{}` failed: {}",
1598 .note(&format!("{:?}", &cmd))
1599 .note(str::from_utf8(¬e[..]).unwrap())
1601 handler.abort_if_errors();
1605 handler.err(&format!("could not exec the linker `{}`: {}", pname.display(), e));
1606 handler.abort_if_errors();
1612 enum SharedEmitterMessage {
1613 Diagnostic(Diagnostic),
1614 InlineAsmError(u32, String),
1620 pub struct SharedEmitter {
1621 sender: Sender<SharedEmitterMessage>,
1624 pub struct SharedEmitterMain {
1625 receiver: Receiver<SharedEmitterMessage>,
1628 impl SharedEmitter {
1629 pub fn new() -> (SharedEmitter, SharedEmitterMain) {
1630 let (sender, receiver) = channel();
1632 (SharedEmitter { sender }, SharedEmitterMain { receiver })
1635 pub fn inline_asm_error(&self, cookie: u32, msg: String) {
1636 drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg)));
1639 pub fn fatal(&self, msg: &str) {
1640 drop(self.sender.send(SharedEmitterMessage::Fatal(msg.to_string())));
1644 impl Emitter for SharedEmitter {
1645 fn emit(&mut self, db: &DiagnosticBuilder) {
1646 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
1648 code: db.code.clone(),
1651 for child in &db.children {
1652 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
1653 msg: child.message(),
1658 drop(self.sender.send(SharedEmitterMessage::AbortIfErrors));
1662 impl SharedEmitterMain {
1663 pub fn check(&self, sess: &Session, blocking: bool) {
1665 let message = if blocking {
1666 match self.receiver.recv() {
1667 Ok(message) => Ok(message),
1671 match self.receiver.try_recv() {
1672 Ok(message) => Ok(message),
1678 Ok(SharedEmitterMessage::Diagnostic(diag)) => {
1679 let handler = sess.diagnostic();
1682 handler.emit_with_code(&MultiSpan::new(),
1688 handler.emit(&MultiSpan::new(),
1694 Ok(SharedEmitterMessage::InlineAsmError(cookie, msg)) => {
1695 match Mark::from_u32(cookie).expn_info() {
1696 Some(ei) => sess.span_err(ei.call_site, &msg),
1697 None => sess.err(&msg),
1700 Ok(SharedEmitterMessage::AbortIfErrors) => {
1701 sess.abort_if_errors();
1703 Ok(SharedEmitterMessage::Fatal(msg)) => {
1715 pub struct OngoingCodegen<B: ExtraBackendMethods> {
1717 pub crate_name: Symbol,
1718 pub crate_hash: Svh,
1719 pub metadata: EncodedMetadata,
1720 pub windows_subsystem: Option<String>,
1721 pub linker_info: LinkerInfo,
1722 pub crate_info: CrateInfo,
1723 pub time_graph: Option<TimeGraph>,
1724 pub coordinator_send: Sender<Box<dyn Any + Send>>,
1725 pub codegen_worker_receive: Receiver<Message<B>>,
1726 pub shared_emitter_main: SharedEmitterMain,
1727 pub future: thread::JoinHandle<Result<CompiledModules, ()>>,
1728 pub output_filenames: Arc<OutputFilenames>,
1731 impl<B: ExtraBackendMethods> OngoingCodegen<B> {
1735 ) -> (CodegenResults, FxHashMap<WorkProductId, WorkProduct>) {
1736 self.shared_emitter_main.check(sess, true);
1737 let compiled_modules = match self.future.join() {
1738 Ok(Ok(compiled_modules)) => compiled_modules,
1740 sess.abort_if_errors();
1741 panic!("expected abort due to worker thread errors")
1744 bug!("panic during codegen/LLVM phase");
1748 sess.cgu_reuse_tracker.check_expected_reuse(sess);
1750 sess.abort_if_errors();
1752 if let Some(time_graph) = self.time_graph {
1753 time_graph.dump(&format!("{}-timings", self.crate_name));
1757 copy_all_cgu_workproducts_to_incr_comp_cache_dir(sess,
1759 produce_final_output_artifacts(sess,
1761 &self.output_filenames);
1763 // FIXME: time_llvm_passes support - does this use a global context or
1765 if sess.codegen_units() == 1 && sess.time_llvm_passes() {
1766 self.backend.print_pass_timings()
1770 crate_name: self.crate_name,
1771 crate_hash: self.crate_hash,
1772 metadata: self.metadata,
1773 windows_subsystem: self.windows_subsystem,
1774 linker_info: self.linker_info,
1775 crate_info: self.crate_info,
1777 modules: compiled_modules.modules,
1778 allocator_module: compiled_modules.allocator_module,
1779 metadata_module: compiled_modules.metadata_module,
1783 pub fn submit_pre_codegened_module_to_llvm(&self,
1785 module: ModuleCodegen<B::Module>) {
1786 self.wait_for_signal_to_codegen_item();
1787 self.check_for_errors(tcx.sess);
1789 // These are generally cheap and won't through off scheduling.
1791 submit_codegened_module_to_llvm(&self.backend, tcx, module, cost);
1794 pub fn codegen_finished(&self, tcx: TyCtxt) {
1795 self.wait_for_signal_to_codegen_item();
1796 self.check_for_errors(tcx.sess);
1797 drop(self.coordinator_send.send(Box::new(Message::CodegenComplete::<B>)));
1800 /// Consumes this context indicating that codegen was entirely aborted, and
1801 /// we need to exit as quickly as possible.
1803 /// This method blocks the current thread until all worker threads have
1804 /// finished, and all worker threads should have exited or be real close to
1805 /// exiting at this point.
1806 pub fn codegen_aborted(self) {
1807 // Signal to the coordinator it should spawn no more work and start
1809 drop(self.coordinator_send.send(Box::new(Message::CodegenAborted::<B>)));
1810 drop(self.future.join());
1813 pub fn check_for_errors(&self, sess: &Session) {
1814 self.shared_emitter_main.check(sess, false);
1817 pub fn wait_for_signal_to_codegen_item(&self) {
1818 match self.codegen_worker_receive.recv() {
1819 Ok(Message::CodegenItem) => {
1822 Ok(_) => panic!("unexpected message"),
1824 // One of the LLVM threads must have panicked, fall through so
1825 // error handling can be reached.
1831 pub fn submit_codegened_module_to_llvm<B: ExtraBackendMethods>(
1834 module: ModuleCodegen<B::Module>,
1837 let llvm_work_item = WorkItem::Optimize(module);
1838 drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::CodegenDone::<B> {
1844 pub fn submit_post_lto_module_to_llvm<B: ExtraBackendMethods>(
1847 module: CachedModuleCodegen
1849 let llvm_work_item = WorkItem::CopyPostLtoArtifacts(module);
1850 drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::CodegenDone::<B> {
1856 pub fn submit_pre_lto_module_to_llvm<B: ExtraBackendMethods>(
1859 module: CachedModuleCodegen
1861 let filename = pre_lto_bitcode_filename(&module.name);
1862 let bc_path = in_incr_comp_dir_sess(tcx.sess, &filename);
1863 let file = fs::File::open(&bc_path).unwrap_or_else(|e| {
1864 panic!("failed to open bitcode file `{}`: {}", bc_path.display(), e)
1868 memmap::Mmap::map(&file).unwrap_or_else(|e| {
1869 panic!("failed to mmap bitcode file `{}`: {}", bc_path.display(), e)
1872 // Schedule the module to be loaded
1873 drop(tcx.tx_to_llvm_workers.lock().send(Box::new(Message::AddImportOnlyModule::<B> {
1874 module_data: SerializedModule::FromUncompressedFile(mmap),
1875 work_product: module.source,
1879 pub fn pre_lto_bitcode_filename(module_name: &str) -> String {
1880 format!("{}.{}", module_name, PRE_THIN_LTO_BC_EXT)
1883 fn msvc_imps_needed(tcx: TyCtxt) -> bool {
1884 // This should never be true (because it's not supported). If it is true,
1885 // something is wrong with commandline arg validation.
1886 assert!(!(tcx.sess.opts.cg.linker_plugin_lto.enabled() &&
1887 tcx.sess.target.target.options.is_like_msvc &&
1888 tcx.sess.opts.cg.prefer_dynamic));
1890 tcx.sess.target.target.options.is_like_msvc &&
1891 tcx.sess.crate_types.borrow().iter().any(|ct| *ct == config::CrateType::Rlib) &&
1892 // ThinLTO can't handle this workaround in all cases, so we don't
1893 // emit the `__imp_` symbols. Instead we make them unnecessary by disallowing
1894 // dynamic linking when linker plugin LTO is enabled.
1895 !tcx.sess.opts.cg.linker_plugin_lto.enabled()