1 use super::link::{self, ensure_removed};
2 use super::lto::{self, SerializedModule};
3 use super::symbol_export::symbol_name_for_instance_in_crate;
6 CachedModuleCodegen, CodegenResults, CompiledModule, CrateInfo, ModuleCodegen, ModuleKind,
10 use jobserver::{Acquired, Client};
11 use rustc_data_structures::fx::FxHashMap;
12 use rustc_data_structures::memmap::Mmap;
13 use rustc_data_structures::profiling::SelfProfilerRef;
14 use rustc_data_structures::profiling::TimingGuard;
15 use rustc_data_structures::profiling::VerboseTimingGuard;
16 use rustc_data_structures::sync::Lrc;
17 use rustc_errors::emitter::Emitter;
18 use rustc_errors::{DiagnosticId, FatalError, Handler, Level};
19 use rustc_fs_util::link_or_copy;
20 use rustc_hir::def_id::{CrateNum, LOCAL_CRATE};
21 use rustc_incremental::{
22 copy_cgu_workproduct_to_incr_comp_cache_dir, in_incr_comp_dir, in_incr_comp_dir_sess,
24 use rustc_metadata::EncodedMetadata;
25 use rustc_middle::dep_graph::{WorkProduct, WorkProductId};
26 use rustc_middle::middle::exported_symbols::SymbolExportLevel;
27 use rustc_middle::ty::TyCtxt;
28 use rustc_session::cgu_reuse_tracker::CguReuseTracker;
29 use rustc_session::config::{self, CrateType, Lto, OutputFilenames, OutputType};
30 use rustc_session::config::{Passes, SwitchWithOptPath};
31 use rustc_session::Session;
32 use rustc_span::source_map::SourceMap;
33 use rustc_span::symbol::sym;
34 use rustc_span::{BytePos, FileName, InnerSpan, Pos, Span};
35 use rustc_target::spec::{MergeFunctions, PanicStrategy, SanitizerSet};
41 use std::path::{Path, PathBuf};
43 use std::sync::mpsc::{channel, Receiver, Sender};
47 const PRE_LTO_BC_EXT: &str = "pre-lto.bc";
49 /// What kind of object file to emit.
50 #[derive(Clone, Copy, PartialEq)]
55 // Just uncompressed llvm bitcode. Provides easy compatibility with
56 // emscripten's ecc compiler, when used as the linker.
59 // Object code, possibly augmented with a bitcode section.
60 ObjectCode(BitcodeSection),
63 /// What kind of llvm bitcode section to embed in an object file.
64 #[derive(Clone, Copy, PartialEq)]
65 pub enum BitcodeSection {
66 // No bitcode section.
69 // A full, uncompressed bitcode section.
73 /// Module-specific configuration for `optimize_and_codegen`.
74 pub struct ModuleConfig {
75 /// Names of additional optimization passes to run.
76 pub passes: Vec<String>,
77 /// Some(level) to optimize at a certain level, or None to run
78 /// absolutely no optimizations (used for the metadata module).
79 pub opt_level: Option<config::OptLevel>,
81 /// Some(level) to optimize binary size, or None to not affect program size.
82 pub opt_size: Option<config::OptLevel>,
84 pub pgo_gen: SwitchWithOptPath,
85 pub pgo_use: Option<PathBuf>,
86 pub pgo_sample_use: Option<PathBuf>,
87 pub debug_info_for_profiling: bool,
88 pub instrument_coverage: bool,
89 pub instrument_gcov: bool,
91 pub sanitizer: SanitizerSet,
92 pub sanitizer_recover: SanitizerSet,
93 pub sanitizer_memory_track_origins: usize,
95 // Flags indicating which outputs to produce.
96 pub emit_pre_lto_bc: bool,
97 pub emit_no_opt_bc: bool,
101 pub emit_obj: EmitObj,
102 pub bc_cmdline: String,
104 // Miscellaneous flags. These are mostly copied from command-line
106 pub verify_llvm_ir: bool,
107 pub no_prepopulate_passes: bool,
108 pub no_builtins: bool,
109 pub time_module: bool,
110 pub vectorize_loop: bool,
111 pub vectorize_slp: bool,
112 pub merge_functions: bool,
113 pub inline_threshold: Option<u32>,
114 pub new_llvm_pass_manager: Option<bool>,
115 pub emit_lifetime_markers: bool,
123 is_compiler_builtins: bool,
125 // If it's a regular module, use `$regular`, otherwise use `$other`.
126 // `$regular` and `$other` are evaluated lazily.
127 macro_rules! if_regular {
128 ($regular: expr, $other: expr) => {
129 if let ModuleKind::Regular = kind { $regular } else { $other }
133 let opt_level_and_size = if_regular!(Some(sess.opts.optimize), None);
135 let save_temps = sess.opts.cg.save_temps;
137 let should_emit_obj = sess.opts.output_types.contains_key(&OutputType::Exe)
139 ModuleKind::Regular => sess.opts.output_types.contains_key(&OutputType::Object),
140 ModuleKind::Allocator => false,
141 ModuleKind::Metadata => sess.opts.output_types.contains_key(&OutputType::Metadata),
144 let emit_obj = if !should_emit_obj {
146 } else if sess.target.obj_is_bitcode
147 || (sess.opts.cg.linker_plugin_lto.enabled() && !no_builtins)
149 // This case is selected if the target uses objects as bitcode, or
150 // if linker plugin LTO is enabled. In the linker plugin LTO case
151 // the assumption is that the final link-step will read the bitcode
152 // and convert it to object code. This may be done by either the
153 // native linker or rustc itself.
155 // Note, however, that the linker-plugin-lto requested here is
156 // explicitly ignored for `#![no_builtins]` crates. These crates are
157 // specifically ignored by rustc's LTO passes and wouldn't work if
158 // loaded into the linker. These crates define symbols that LLVM
159 // lowers intrinsics to, and these symbol dependencies aren't known
160 // until after codegen. As a result any crate marked
161 // `#![no_builtins]` is assumed to not participate in LTO and
162 // instead goes on to generate object code.
164 } else if need_bitcode_in_object(sess) {
165 EmitObj::ObjectCode(BitcodeSection::Full)
167 EmitObj::ObjectCode(BitcodeSection::None)
171 passes: if_regular!(sess.opts.cg.passes.clone(), vec![]),
173 opt_level: opt_level_and_size,
174 opt_size: opt_level_and_size,
176 pgo_gen: if_regular!(
177 sess.opts.cg.profile_generate.clone(),
178 SwitchWithOptPath::Disabled
180 pgo_use: if_regular!(sess.opts.cg.profile_use.clone(), None),
181 pgo_sample_use: if_regular!(sess.opts.debugging_opts.profile_sample_use.clone(), None),
182 debug_info_for_profiling: sess.opts.debugging_opts.debug_info_for_profiling,
183 instrument_coverage: if_regular!(sess.instrument_coverage(), false),
184 instrument_gcov: if_regular!(
185 // compiler_builtins overrides the codegen-units settings,
186 // which is incompatible with -Zprofile which requires that
187 // only a single codegen unit is used per crate.
188 sess.opts.debugging_opts.profile && !is_compiler_builtins,
192 sanitizer: if_regular!(sess.opts.debugging_opts.sanitizer, SanitizerSet::empty()),
193 sanitizer_recover: if_regular!(
194 sess.opts.debugging_opts.sanitizer_recover,
195 SanitizerSet::empty()
197 sanitizer_memory_track_origins: if_regular!(
198 sess.opts.debugging_opts.sanitizer_memory_track_origins,
202 emit_pre_lto_bc: if_regular!(
203 save_temps || need_pre_lto_bitcode_for_incr_comp(sess),
206 emit_no_opt_bc: if_regular!(save_temps, false),
207 emit_bc: if_regular!(
208 save_temps || sess.opts.output_types.contains_key(&OutputType::Bitcode),
211 emit_ir: if_regular!(
212 sess.opts.output_types.contains_key(&OutputType::LlvmAssembly),
215 emit_asm: if_regular!(
216 sess.opts.output_types.contains_key(&OutputType::Assembly),
220 bc_cmdline: sess.target.bitcode_llvm_cmdline.clone(),
222 verify_llvm_ir: sess.verify_llvm_ir(),
223 no_prepopulate_passes: sess.opts.cg.no_prepopulate_passes,
224 no_builtins: no_builtins || sess.target.no_builtins,
226 // Exclude metadata and allocator modules from time_passes output,
227 // since they throw off the "LLVM passes" measurement.
228 time_module: if_regular!(true, false),
230 // Copy what clang does by turning on loop vectorization at O2 and
231 // slp vectorization at O3.
232 vectorize_loop: !sess.opts.cg.no_vectorize_loops
233 && (sess.opts.optimize == config::OptLevel::Default
234 || sess.opts.optimize == config::OptLevel::Aggressive),
235 vectorize_slp: !sess.opts.cg.no_vectorize_slp
236 && sess.opts.optimize == config::OptLevel::Aggressive,
238 // Some targets (namely, NVPTX) interact badly with the
239 // MergeFunctions pass. This is because MergeFunctions can generate
240 // new function calls which may interfere with the target calling
241 // convention; e.g. for the NVPTX target, PTX kernels should not
242 // call other PTX kernels. MergeFunctions can also be configured to
243 // generate aliases instead, but aliases are not supported by some
244 // backends (again, NVPTX). Therefore, allow targets to opt out of
245 // the MergeFunctions pass, but otherwise keep the pass enabled (at
246 // O2 and O3) since it can be useful for reducing code size.
247 merge_functions: match sess
251 .unwrap_or(sess.target.merge_functions)
253 MergeFunctions::Disabled => false,
254 MergeFunctions::Trampolines | MergeFunctions::Aliases => {
255 sess.opts.optimize == config::OptLevel::Default
256 || sess.opts.optimize == config::OptLevel::Aggressive
260 inline_threshold: sess.opts.cg.inline_threshold,
261 new_llvm_pass_manager: sess.opts.debugging_opts.new_llvm_pass_manager,
262 emit_lifetime_markers: sess.emit_lifetime_markers(),
266 pub fn bitcode_needed(&self) -> bool {
268 || self.emit_obj == EmitObj::Bitcode
269 || self.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full)
273 /// Configuration passed to the function returned by the `target_machine_factory`.
274 pub struct TargetMachineFactoryConfig {
275 /// Split DWARF is enabled in LLVM by checking that `TM.MCOptions.SplitDwarfFile` isn't empty,
276 /// so the path to the dwarf object has to be provided when we create the target machine.
277 /// This can be ignored by backends which do not need it for their Split DWARF support.
278 pub split_dwarf_file: Option<PathBuf>,
281 impl TargetMachineFactoryConfig {
283 cgcx: &CodegenContext<impl WriteBackendMethods>,
285 ) -> TargetMachineFactoryConfig {
286 let split_dwarf_file = if cgcx.target_can_use_split_dwarf {
287 cgcx.output_filenames.split_dwarf_path(cgcx.split_debuginfo, Some(module_name))
291 TargetMachineFactoryConfig { split_dwarf_file }
295 pub type TargetMachineFactoryFn<B> = Arc<
296 dyn Fn(TargetMachineFactoryConfig) -> Result<<B as WriteBackendMethods>::TargetMachine, String>
301 pub type ExportedSymbols = FxHashMap<CrateNum, Arc<Vec<(String, SymbolExportLevel)>>>;
303 /// Additional resources used by optimize_and_codegen (not module specific)
305 pub struct CodegenContext<B: WriteBackendMethods> {
306 // Resources needed when running LTO
308 pub prof: SelfProfilerRef,
310 pub no_landing_pads: bool,
311 pub save_temps: bool,
312 pub fewer_names: bool,
313 pub time_trace: bool,
314 pub exported_symbols: Option<Arc<ExportedSymbols>>,
315 pub opts: Arc<config::Options>,
316 pub crate_types: Vec<CrateType>,
317 pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>,
318 pub output_filenames: Arc<OutputFilenames>,
319 pub regular_module_config: Arc<ModuleConfig>,
320 pub metadata_module_config: Arc<ModuleConfig>,
321 pub allocator_module_config: Arc<ModuleConfig>,
322 pub tm_factory: TargetMachineFactoryFn<B>,
323 pub msvc_imps_needed: bool,
324 pub is_pe_coff: bool,
325 pub target_can_use_split_dwarf: bool,
326 pub target_pointer_width: u32,
327 pub target_arch: String,
328 pub debuginfo: config::DebugInfo,
329 pub split_debuginfo: rustc_target::spec::SplitDebuginfo,
331 // Number of cgus excluding the allocator/metadata modules
332 pub total_cgus: usize,
333 // Handler to use for diagnostics produced during codegen.
334 pub diag_emitter: SharedEmitter,
335 // LLVM optimizations for which we want to print remarks.
337 // Worker thread number
339 // The incremental compilation session directory, or None if we are not
340 // compiling incrementally
341 pub incr_comp_session_dir: Option<PathBuf>,
342 // Used to update CGU re-use information during the thinlto phase.
343 pub cgu_reuse_tracker: CguReuseTracker,
344 // Channel back to the main control thread to send messages to
345 pub coordinator_send: Sender<Box<dyn Any + Send>>,
348 impl<B: WriteBackendMethods> CodegenContext<B> {
349 pub fn create_diag_handler(&self) -> Handler {
350 Handler::with_emitter(true, None, Box::new(self.diag_emitter.clone()))
353 pub fn config(&self, kind: ModuleKind) -> &ModuleConfig {
355 ModuleKind::Regular => &self.regular_module_config,
356 ModuleKind::Metadata => &self.metadata_module_config,
357 ModuleKind::Allocator => &self.allocator_module_config,
362 fn generate_lto_work<B: ExtraBackendMethods>(
363 cgcx: &CodegenContext<B>,
364 needs_fat_lto: Vec<FatLTOInput<B>>,
365 needs_thin_lto: Vec<(String, B::ThinBuffer)>,
366 import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>,
367 ) -> Vec<(WorkItem<B>, u64)> {
368 let _prof_timer = cgcx.prof.generic_activity("codegen_generate_lto_work");
370 let (lto_modules, copy_jobs) = if !needs_fat_lto.is_empty() {
371 assert!(needs_thin_lto.is_empty());
373 B::run_fat_lto(cgcx, needs_fat_lto, import_only_modules).unwrap_or_else(|e| e.raise());
374 (vec![lto_module], vec![])
376 assert!(needs_fat_lto.is_empty());
377 B::run_thin_lto(cgcx, needs_thin_lto, import_only_modules).unwrap_or_else(|e| e.raise())
383 let cost = module.cost();
384 (WorkItem::LTO(module), cost)
386 .chain(copy_jobs.into_iter().map(|wp| {
388 WorkItem::CopyPostLtoArtifacts(CachedModuleCodegen {
389 name: wp.cgu_name.clone(),
398 pub struct CompiledModules {
399 pub modules: Vec<CompiledModule>,
400 pub metadata_module: Option<CompiledModule>,
401 pub allocator_module: Option<CompiledModule>,
404 fn need_bitcode_in_object(sess: &Session) -> bool {
405 let requested_for_rlib = sess.opts.cg.embed_bitcode
406 && sess.crate_types().contains(&CrateType::Rlib)
407 && sess.opts.output_types.contains_key(&OutputType::Exe);
408 let forced_by_target = sess.target.forces_embed_bitcode;
409 requested_for_rlib || forced_by_target
412 fn need_pre_lto_bitcode_for_incr_comp(sess: &Session) -> bool {
413 if sess.opts.incremental.is_none() {
419 Lto::Fat | Lto::Thin | Lto::ThinLocal => true,
423 pub fn start_async_codegen<B: ExtraBackendMethods>(
427 metadata: EncodedMetadata,
429 ) -> OngoingCodegen<B> {
430 let (coordinator_send, coordinator_receive) = channel();
433 let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID);
434 let no_builtins = tcx.sess.contains_name(crate_attrs, sym::no_builtins);
435 let is_compiler_builtins = tcx.sess.contains_name(crate_attrs, sym::compiler_builtins);
437 let crate_info = CrateInfo::new(tcx, target_cpu);
440 ModuleConfig::new(ModuleKind::Regular, sess, no_builtins, is_compiler_builtins);
441 let metadata_config =
442 ModuleConfig::new(ModuleKind::Metadata, sess, no_builtins, is_compiler_builtins);
443 let allocator_config =
444 ModuleConfig::new(ModuleKind::Allocator, sess, no_builtins, is_compiler_builtins);
446 let (shared_emitter, shared_emitter_main) = SharedEmitter::new();
447 let (codegen_worker_send, codegen_worker_receive) = channel();
449 let coordinator_thread = start_executing_work(
457 sess.jobserver.clone(),
458 Arc::new(regular_config),
459 Arc::new(metadata_config),
460 Arc::new(allocator_config),
461 coordinator_send.clone(),
470 codegen_worker_receive,
472 future: coordinator_thread,
473 output_filenames: tcx.output_filenames(()),
477 fn copy_all_cgu_workproducts_to_incr_comp_cache_dir(
479 compiled_modules: &CompiledModules,
480 ) -> FxHashMap<WorkProductId, WorkProduct> {
481 let mut work_products = FxHashMap::default();
483 if sess.opts.incremental.is_none() {
484 return work_products;
487 let _timer = sess.timer("copy_all_cgu_workproducts_to_incr_comp_cache_dir");
489 for module in compiled_modules.modules.iter().filter(|m| m.kind == ModuleKind::Regular) {
490 let path = module.object.as_ref().cloned();
492 if let Some((id, product)) =
493 copy_cgu_workproduct_to_incr_comp_cache_dir(sess, &module.name, &path)
495 work_products.insert(id, product);
502 fn produce_final_output_artifacts(
504 compiled_modules: &CompiledModules,
505 crate_output: &OutputFilenames,
507 let mut user_wants_bitcode = false;
508 let mut user_wants_objects = false;
510 // Produce final compile outputs.
511 let copy_gracefully = |from: &Path, to: &Path| {
512 if let Err(e) = fs::copy(from, to) {
513 sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
517 let copy_if_one_unit = |output_type: OutputType, keep_numbered: bool| {
518 if compiled_modules.modules.len() == 1 {
519 // 1) Only one codegen unit. In this case it's no difficulty
520 // to copy `foo.0.x` to `foo.x`.
521 let module_name = Some(&compiled_modules.modules[0].name[..]);
522 let path = crate_output.temp_path(output_type, module_name);
523 copy_gracefully(&path, &crate_output.path(output_type));
524 if !sess.opts.cg.save_temps && !keep_numbered {
525 // The user just wants `foo.x`, not `foo.#module-name#.x`.
526 ensure_removed(sess.diagnostic(), &path);
529 let ext = crate_output
530 .temp_path(output_type, None)
537 if crate_output.outputs.contains_key(&output_type) {
538 // 2) Multiple codegen units, with `--emit foo=some_name`. We have
539 // no good solution for this case, so warn the user.
541 "ignoring emit path because multiple .{} files \
545 } else if crate_output.single_output_file.is_some() {
546 // 3) Multiple codegen units, with `-o some_name`. We have
547 // no good solution for this case, so warn the user.
549 "ignoring -o because multiple .{} files \
554 // 4) Multiple codegen units, but no explicit name. We
555 // just leave the `foo.0.x` files in place.
556 // (We don't have to do any work in this case.)
561 // Flag to indicate whether the user explicitly requested bitcode.
562 // Otherwise, we produced it only as a temporary output, and will need
564 for output_type in crate_output.outputs.keys() {
566 OutputType::Bitcode => {
567 user_wants_bitcode = true;
568 // Copy to .bc, but always keep the .0.bc. There is a later
569 // check to figure out if we should delete .0.bc files, or keep
570 // them for making an rlib.
571 copy_if_one_unit(OutputType::Bitcode, true);
573 OutputType::LlvmAssembly => {
574 copy_if_one_unit(OutputType::LlvmAssembly, false);
576 OutputType::Assembly => {
577 copy_if_one_unit(OutputType::Assembly, false);
579 OutputType::Object => {
580 user_wants_objects = true;
581 copy_if_one_unit(OutputType::Object, true);
583 OutputType::Mir | OutputType::Metadata | OutputType::Exe | OutputType::DepInfo => {}
587 // Clean up unwanted temporary files.
589 // We create the following files by default:
590 // - #crate#.#module-name#.bc
591 // - #crate#.#module-name#.o
592 // - #crate#.crate.metadata.bc
593 // - #crate#.crate.metadata.o
594 // - #crate#.o (linked from crate.##.o)
595 // - #crate#.bc (copied from crate.##.bc)
596 // We may create additional files if requested by the user (through
597 // `-C save-temps` or `--emit=` flags).
599 if !sess.opts.cg.save_temps {
600 // Remove the temporary .#module-name#.o objects. If the user didn't
601 // explicitly request bitcode (with --emit=bc), and the bitcode is not
602 // needed for building an rlib, then we must remove .#module-name#.bc as
605 // Specific rules for keeping .#module-name#.bc:
606 // - If the user requested bitcode (`user_wants_bitcode`), and
607 // codegen_units > 1, then keep it.
608 // - If the user requested bitcode but codegen_units == 1, then we
609 // can toss .#module-name#.bc because we copied it to .bc earlier.
610 // - If we're not building an rlib and the user didn't request
611 // bitcode, then delete .#module-name#.bc.
612 // If you change how this works, also update back::link::link_rlib,
613 // where .#module-name#.bc files are (maybe) deleted after making an
615 let needs_crate_object = crate_output.outputs.contains_key(&OutputType::Exe);
617 let keep_numbered_bitcode = user_wants_bitcode && sess.codegen_units() > 1;
619 let keep_numbered_objects =
620 needs_crate_object || (user_wants_objects && sess.codegen_units() > 1);
622 for module in compiled_modules.modules.iter() {
623 if let Some(ref path) = module.object {
624 if !keep_numbered_objects {
625 ensure_removed(sess.diagnostic(), path);
629 if let Some(ref path) = module.dwarf_object {
630 if !keep_numbered_objects {
631 ensure_removed(sess.diagnostic(), path);
635 if let Some(ref path) = module.bytecode {
636 if !keep_numbered_bitcode {
637 ensure_removed(sess.diagnostic(), path);
642 if !user_wants_bitcode {
643 if let Some(ref metadata_module) = compiled_modules.metadata_module {
644 if let Some(ref path) = metadata_module.bytecode {
645 ensure_removed(sess.diagnostic(), &path);
649 if let Some(ref allocator_module) = compiled_modules.allocator_module {
650 if let Some(ref path) = allocator_module.bytecode {
651 ensure_removed(sess.diagnostic(), path);
657 // We leave the following files around by default:
659 // - #crate#.crate.metadata.o
661 // These are used in linking steps and will be cleaned up afterward.
664 pub enum WorkItem<B: WriteBackendMethods> {
665 /// Optimize a newly codegened, totally unoptimized module.
666 Optimize(ModuleCodegen<B::Module>),
667 /// Copy the post-LTO artifacts from the incremental cache to the output
669 CopyPostLtoArtifacts(CachedModuleCodegen),
670 /// Performs (Thin)LTO on the given module.
671 LTO(lto::LtoModuleCodegen<B>),
674 impl<B: WriteBackendMethods> WorkItem<B> {
675 pub fn module_kind(&self) -> ModuleKind {
677 WorkItem::Optimize(ref m) => m.kind,
678 WorkItem::CopyPostLtoArtifacts(_) | WorkItem::LTO(_) => ModuleKind::Regular,
682 fn start_profiling<'a>(&self, cgcx: &'a CodegenContext<B>) -> TimingGuard<'a> {
684 WorkItem::Optimize(ref m) => {
685 cgcx.prof.generic_activity_with_arg("codegen_module_optimize", &m.name[..])
687 WorkItem::CopyPostLtoArtifacts(ref m) => cgcx
689 .generic_activity_with_arg("codegen_copy_artifacts_from_incr_cache", &m.name[..]),
690 WorkItem::LTO(ref m) => {
691 cgcx.prof.generic_activity_with_arg("codegen_module_perform_lto", m.name())
696 /// Generate a short description of this work item suitable for use as a thread name.
697 fn short_description(&self) -> String {
698 // `pthread_setname()` on *nix is limited to 15 characters and longer names are ignored.
699 // Use very short descriptions in this case to maximize the space available for the module name.
700 // Windows does not have that limitation so use slightly more descriptive names there.
702 WorkItem::Optimize(m) => {
704 return format!("optimize module {}", m.name);
706 return format!("opt {}", m.name);
708 WorkItem::CopyPostLtoArtifacts(m) => {
710 return format!("copy LTO artifacts for {}", m.name);
712 return format!("copy {}", m.name);
714 WorkItem::LTO(m) => {
716 return format!("LTO module {}", m.name());
718 return format!("LTO {}", m.name());
724 enum WorkItemResult<B: WriteBackendMethods> {
725 Compiled(CompiledModule),
726 NeedsLink(ModuleCodegen<B::Module>),
727 NeedsFatLTO(FatLTOInput<B>),
728 NeedsThinLTO(String, B::ThinBuffer),
731 pub enum FatLTOInput<B: WriteBackendMethods> {
732 Serialized { name: String, buffer: B::ModuleBuffer },
733 InMemory(ModuleCodegen<B::Module>),
736 fn execute_work_item<B: ExtraBackendMethods>(
737 cgcx: &CodegenContext<B>,
738 work_item: WorkItem<B>,
739 ) -> Result<WorkItemResult<B>, FatalError> {
740 let module_config = cgcx.config(work_item.module_kind());
743 WorkItem::Optimize(module) => execute_optimize_work_item(cgcx, module, module_config),
744 WorkItem::CopyPostLtoArtifacts(module) => {
745 Ok(execute_copy_from_cache_work_item(cgcx, module, module_config))
747 WorkItem::LTO(module) => execute_lto_work_item(cgcx, module, module_config),
751 // Actual LTO type we end up choosing based on multiple factors.
752 pub enum ComputedLtoType {
758 pub fn compute_per_cgu_lto_type(
760 opts: &config::Options,
761 sess_crate_types: &[CrateType],
762 module_kind: ModuleKind,
763 ) -> ComputedLtoType {
764 // Metadata modules never participate in LTO regardless of the lto
766 if module_kind == ModuleKind::Metadata {
767 return ComputedLtoType::No;
770 // If the linker does LTO, we don't have to do it. Note that we
771 // keep doing full LTO, if it is requested, as not to break the
772 // assumption that the output will be a single module.
773 let linker_does_lto = opts.cg.linker_plugin_lto.enabled();
775 // When we're automatically doing ThinLTO for multi-codegen-unit
776 // builds we don't actually want to LTO the allocator modules if
777 // it shows up. This is due to various linker shenanigans that
778 // we'll encounter later.
779 let is_allocator = module_kind == ModuleKind::Allocator;
781 // We ignore a request for full crate grath LTO if the cate type
782 // is only an rlib, as there is no full crate graph to process,
783 // that'll happen later.
785 // This use case currently comes up primarily for targets that
786 // require LTO so the request for LTO is always unconditionally
787 // passed down to the backend, but we don't actually want to do
788 // anything about it yet until we've got a final product.
789 let is_rlib = sess_crate_types.len() == 1 && sess_crate_types[0] == CrateType::Rlib;
792 Lto::ThinLocal if !linker_does_lto && !is_allocator => ComputedLtoType::Thin,
793 Lto::Thin if !linker_does_lto && !is_rlib => ComputedLtoType::Thin,
794 Lto::Fat if !is_rlib => ComputedLtoType::Fat,
795 _ => ComputedLtoType::No,
799 fn execute_optimize_work_item<B: ExtraBackendMethods>(
800 cgcx: &CodegenContext<B>,
801 module: ModuleCodegen<B::Module>,
802 module_config: &ModuleConfig,
803 ) -> Result<WorkItemResult<B>, FatalError> {
804 let diag_handler = cgcx.create_diag_handler();
807 B::optimize(cgcx, &diag_handler, &module, module_config)?;
810 // After we've done the initial round of optimizations we need to
811 // decide whether to synchronously codegen this module or ship it
812 // back to the coordinator thread for further LTO processing (which
813 // has to wait for all the initial modules to be optimized).
815 let lto_type = compute_per_cgu_lto_type(&cgcx.lto, &cgcx.opts, &cgcx.crate_types, module.kind);
817 // If we're doing some form of incremental LTO then we need to be sure to
818 // save our module to disk first.
819 let bitcode = if cgcx.config(module.kind).emit_pre_lto_bc {
820 let filename = pre_lto_bitcode_filename(&module.name);
821 cgcx.incr_comp_session_dir.as_ref().map(|path| path.join(&filename))
827 ComputedLtoType::No => finish_intra_module_work(cgcx, module, module_config),
828 ComputedLtoType::Thin => {
829 let (name, thin_buffer) = B::prepare_thin(module);
830 if let Some(path) = bitcode {
831 fs::write(&path, thin_buffer.data()).unwrap_or_else(|e| {
832 panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
835 Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))
837 ComputedLtoType::Fat => match bitcode {
839 let (name, buffer) = B::serialize_module(module);
840 fs::write(&path, buffer.data()).unwrap_or_else(|e| {
841 panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
843 Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::Serialized { name, buffer }))
845 None => Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::InMemory(module))),
850 fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>(
851 cgcx: &CodegenContext<B>,
852 module: CachedModuleCodegen,
853 module_config: &ModuleConfig,
854 ) -> WorkItemResult<B> {
855 let incr_comp_session_dir = cgcx.incr_comp_session_dir.as_ref().unwrap();
856 let mut object = None;
857 if let Some(saved_file) = module.source.saved_file {
858 let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, Some(&module.name));
859 object = Some(obj_out.clone());
860 let source_file = in_incr_comp_dir(&incr_comp_session_dir, &saved_file);
862 "copying pre-existing module `{}` from {:?} to {}",
867 if let Err(err) = link_or_copy(&source_file, &obj_out) {
868 let diag_handler = cgcx.create_diag_handler();
869 diag_handler.err(&format!(
870 "unable to copy {} to {}: {}",
871 source_file.display(),
878 assert_eq!(object.is_some(), module_config.emit_obj != EmitObj::None);
880 WorkItemResult::Compiled(CompiledModule {
882 kind: ModuleKind::Regular,
889 fn execute_lto_work_item<B: ExtraBackendMethods>(
890 cgcx: &CodegenContext<B>,
891 mut module: lto::LtoModuleCodegen<B>,
892 module_config: &ModuleConfig,
893 ) -> Result<WorkItemResult<B>, FatalError> {
894 let module = unsafe { module.optimize(cgcx)? };
895 finish_intra_module_work(cgcx, module, module_config)
898 fn finish_intra_module_work<B: ExtraBackendMethods>(
899 cgcx: &CodegenContext<B>,
900 module: ModuleCodegen<B::Module>,
901 module_config: &ModuleConfig,
902 ) -> Result<WorkItemResult<B>, FatalError> {
903 let diag_handler = cgcx.create_diag_handler();
905 if !cgcx.opts.debugging_opts.combine_cgu
906 || module.kind == ModuleKind::Metadata
907 || module.kind == ModuleKind::Allocator
909 let module = unsafe { B::codegen(cgcx, &diag_handler, module, module_config)? };
910 Ok(WorkItemResult::Compiled(module))
912 Ok(WorkItemResult::NeedsLink(module))
916 pub enum Message<B: WriteBackendMethods> {
917 Token(io::Result<Acquired>),
919 result: FatLTOInput<B>,
924 thin_buffer: B::ThinBuffer,
928 module: ModuleCodegen<B::Module>,
932 result: Result<CompiledModule, Option<WorkerFatalError>>,
936 llvm_work_item: WorkItem<B>,
939 AddImportOnlyModule {
940 module_data: SerializedModule<B::ModuleBuffer>,
941 work_product: WorkProduct,
950 code: Option<DiagnosticId>,
954 #[derive(PartialEq, Clone, Copy, Debug)]
955 enum MainThreadWorkerState {
961 fn start_executing_work<B: ExtraBackendMethods>(
964 crate_info: &CrateInfo,
965 shared_emitter: SharedEmitter,
966 codegen_worker_send: Sender<Message<B>>,
967 coordinator_receive: Receiver<Box<dyn Any + Send>>,
970 regular_config: Arc<ModuleConfig>,
971 metadata_config: Arc<ModuleConfig>,
972 allocator_config: Arc<ModuleConfig>,
973 tx_to_llvm_workers: Sender<Box<dyn Any + Send>>,
974 ) -> thread::JoinHandle<Result<CompiledModules, ()>> {
975 let coordinator_send = tx_to_llvm_workers;
978 // Compute the set of symbols we need to retain when doing LTO (if we need to)
979 let exported_symbols = {
980 let mut exported_symbols = FxHashMap::default();
982 let copy_symbols = |cnum| {
984 .exported_symbols(cnum)
986 .map(|&(s, lvl)| (symbol_name_for_instance_in_crate(tcx, s, cnum), lvl))
994 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
995 Some(Arc::new(exported_symbols))
997 Lto::Fat | Lto::Thin => {
998 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
999 for &cnum in tcx.crates(()).iter() {
1000 exported_symbols.insert(cnum, copy_symbols(cnum));
1002 Some(Arc::new(exported_symbols))
1007 // First up, convert our jobserver into a helper thread so we can use normal
1008 // mpsc channels to manage our messages and such.
1009 // After we've requested tokens then we'll, when we can,
1010 // get tokens on `coordinator_receive` which will
1011 // get managed in the main loop below.
1012 let coordinator_send2 = coordinator_send.clone();
1013 let helper = jobserver
1014 .into_helper_thread(move |token| {
1015 drop(coordinator_send2.send(Box::new(Message::Token::<B>(token))));
1017 .expect("failed to spawn helper thread");
1019 let mut each_linked_rlib_for_lto = Vec::new();
1020 drop(link::each_linked_rlib(crate_info, &mut |cnum, path| {
1021 if link::ignored_for_lto(sess, crate_info, cnum) {
1024 each_linked_rlib_for_lto.push((cnum, path.to_path_buf()));
1027 let ol = if tcx.sess.opts.debugging_opts.no_codegen
1028 || !tcx.sess.opts.output_types.should_codegen()
1030 // If we know that we won’t be doing codegen, create target machines without optimisation.
1031 config::OptLevel::No
1033 tcx.backend_optimization_level(())
1035 let cgcx = CodegenContext::<B> {
1036 backend: backend.clone(),
1037 crate_types: sess.crate_types().to_vec(),
1038 each_linked_rlib_for_lto,
1040 no_landing_pads: sess.panic_strategy() == PanicStrategy::Abort,
1041 fewer_names: sess.fewer_names(),
1042 save_temps: sess.opts.cg.save_temps,
1043 time_trace: sess.opts.debugging_opts.llvm_time_trace,
1044 opts: Arc::new(sess.opts.clone()),
1045 prof: sess.prof.clone(),
1047 remark: sess.opts.cg.remark.clone(),
1049 incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
1050 cgu_reuse_tracker: sess.cgu_reuse_tracker.clone(),
1052 diag_emitter: shared_emitter.clone(),
1053 output_filenames: tcx.output_filenames(()),
1054 regular_module_config: regular_config,
1055 metadata_module_config: metadata_config,
1056 allocator_module_config: allocator_config,
1057 tm_factory: backend.target_machine_factory(tcx.sess, ol),
1059 msvc_imps_needed: msvc_imps_needed(tcx),
1060 is_pe_coff: tcx.sess.target.is_like_windows,
1061 target_can_use_split_dwarf: tcx.sess.target_can_use_split_dwarf(),
1062 target_pointer_width: tcx.sess.target.pointer_width,
1063 target_arch: tcx.sess.target.arch.clone(),
1064 debuginfo: tcx.sess.opts.debuginfo,
1065 split_debuginfo: tcx.sess.split_debuginfo(),
1068 // This is the "main loop" of parallel work happening for parallel codegen.
1069 // It's here that we manage parallelism, schedule work, and work with
1070 // messages coming from clients.
1072 // There are a few environmental pre-conditions that shape how the system
1075 // - Error reporting only can happen on the main thread because that's the
1076 // only place where we have access to the compiler `Session`.
1077 // - LLVM work can be done on any thread.
1078 // - Codegen can only happen on the main thread.
1079 // - Each thread doing substantial work must be in possession of a `Token`
1080 // from the `Jobserver`.
1081 // - The compiler process always holds one `Token`. Any additional `Tokens`
1082 // have to be requested from the `Jobserver`.
1086 // The error reporting restriction is handled separately from the rest: We
1087 // set up a `SharedEmitter` the holds an open channel to the main thread.
1088 // When an error occurs on any thread, the shared emitter will send the
1089 // error message to the receiver main thread (`SharedEmitterMain`). The
1090 // main thread will periodically query this error message queue and emit
1091 // any error messages it has received. It might even abort compilation if
1092 // has received a fatal error. In this case we rely on all other threads
1093 // being torn down automatically with the main thread.
1094 // Since the main thread will often be busy doing codegen work, error
1095 // reporting will be somewhat delayed, since the message queue can only be
1096 // checked in between to work packages.
1098 // Work Processing Infrastructure
1099 // ==============================
1100 // The work processing infrastructure knows three major actors:
1102 // - the coordinator thread,
1103 // - the main thread, and
1104 // - LLVM worker threads
1106 // The coordinator thread is running a message loop. It instructs the main
1107 // thread about what work to do when, and it will spawn off LLVM worker
1108 // threads as open LLVM WorkItems become available.
1110 // The job of the main thread is to codegen CGUs into LLVM work package
1111 // (since the main thread is the only thread that can do this). The main
1112 // thread will block until it receives a message from the coordinator, upon
1113 // which it will codegen one CGU, send it to the coordinator and block
1114 // again. This way the coordinator can control what the main thread is
1117 // The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is
1118 // available, it will spawn off a new LLVM worker thread and let it process
1119 // that a WorkItem. When a LLVM worker thread is done with its WorkItem,
1120 // it will just shut down, which also frees all resources associated with
1121 // the given LLVM module, and sends a message to the coordinator that the
1122 // has been completed.
1126 // The scheduler's goal is to minimize the time it takes to complete all
1127 // work there is, however, we also want to keep memory consumption low
1128 // if possible. These two goals are at odds with each other: If memory
1129 // consumption were not an issue, we could just let the main thread produce
1130 // LLVM WorkItems at full speed, assuring maximal utilization of
1131 // Tokens/LLVM worker threads. However, since codegen is usually faster
1132 // than LLVM processing, the queue of LLVM WorkItems would fill up and each
1133 // WorkItem potentially holds on to a substantial amount of memory.
1135 // So the actual goal is to always produce just enough LLVM WorkItems as
1136 // not to starve our LLVM worker threads. That means, once we have enough
1137 // WorkItems in our queue, we can block the main thread, so it does not
1138 // produce more until we need them.
1140 // Doing LLVM Work on the Main Thread
1141 // ----------------------------------
1142 // Since the main thread owns the compiler processes implicit `Token`, it is
1143 // wasteful to keep it blocked without doing any work. Therefore, what we do
1144 // in this case is: We spawn off an additional LLVM worker thread that helps
1145 // reduce the queue. The work it is doing corresponds to the implicit
1146 // `Token`. The coordinator will mark the main thread as being busy with
1147 // LLVM work. (The actual work happens on another OS thread but we just care
1148 // about `Tokens`, not actual threads).
1150 // When any LLVM worker thread finishes while the main thread is marked as
1151 // "busy with LLVM work", we can do a little switcheroo: We give the Token
1152 // of the just finished thread to the LLVM worker thread that is working on
1153 // behalf of the main thread's implicit Token, thus freeing up the main
1154 // thread again. The coordinator can then again decide what the main thread
1155 // should do. This allows the coordinator to make decisions at more points
1158 // Striking a Balance between Throughput and Memory Consumption
1159 // ------------------------------------------------------------
1160 // Since our two goals, (1) use as many Tokens as possible and (2) keep
1161 // memory consumption as low as possible, are in conflict with each other,
1162 // we have to find a trade off between them. Right now, the goal is to keep
1163 // all workers busy, which means that no worker should find the queue empty
1164 // when it is ready to start.
1165 // How do we do achieve this? Good question :) We actually never know how
1166 // many `Tokens` are potentially available so it's hard to say how much to
1167 // fill up the queue before switching the main thread to LLVM work. Also we
1168 // currently don't have a means to estimate how long a running LLVM worker
1169 // will still be busy with it's current WorkItem. However, we know the
1170 // maximal count of available Tokens that makes sense (=the number of CPU
1171 // cores), so we can take a conservative guess. The heuristic we use here
1172 // is implemented in the `queue_full_enough()` function.
1174 // Some Background on Jobservers
1175 // -----------------------------
1176 // It's worth also touching on the management of parallelism here. We don't
1177 // want to just spawn a thread per work item because while that's optimal
1178 // parallelism it may overload a system with too many threads or violate our
1179 // configuration for the maximum amount of cpu to use for this process. To
1180 // manage this we use the `jobserver` crate.
1182 // Job servers are an artifact of GNU make and are used to manage
1183 // parallelism between processes. A jobserver is a glorified IPC semaphore
1184 // basically. Whenever we want to run some work we acquire the semaphore,
1185 // and whenever we're done with that work we release the semaphore. In this
1186 // manner we can ensure that the maximum number of parallel workers is
1187 // capped at any one point in time.
1189 // LTO and the coordinator thread
1190 // ------------------------------
1192 // The final job the coordinator thread is responsible for is managing LTO
1193 // and how that works. When LTO is requested what we'll to is collect all
1194 // optimized LLVM modules into a local vector on the coordinator. Once all
1195 // modules have been codegened and optimized we hand this to the `lto`
1196 // module for further optimization. The `lto` module will return back a list
1197 // of more modules to work on, which the coordinator will continue to spawn
1200 // Each LLVM module is automatically sent back to the coordinator for LTO if
1201 // necessary. There's already optimizations in place to avoid sending work
1202 // back to the coordinator if LTO isn't requested.
1203 return B::spawn_thread(cgcx.time_trace, move || {
1204 let mut worker_id_counter = 0;
1205 let mut free_worker_ids = Vec::new();
1206 let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
1207 if let Some(id) = free_worker_ids.pop() {
1210 let id = worker_id_counter;
1211 worker_id_counter += 1;
1216 // This is where we collect codegen units that have gone all the way
1217 // through codegen and LLVM.
1218 let mut compiled_modules = vec![];
1219 let mut compiled_metadata_module = None;
1220 let mut compiled_allocator_module = None;
1221 let mut needs_link = Vec::new();
1222 let mut needs_fat_lto = Vec::new();
1223 let mut needs_thin_lto = Vec::new();
1224 let mut lto_import_only_modules = Vec::new();
1225 let mut started_lto = false;
1226 let mut codegen_aborted = false;
1228 // This flag tracks whether all items have gone through codegens
1229 let mut codegen_done = false;
1231 // This is the queue of LLVM work items that still need processing.
1232 let mut work_items = Vec::<(WorkItem<B>, u64)>::new();
1234 // This are the Jobserver Tokens we currently hold. Does not include
1235 // the implicit Token the compiler process owns no matter what.
1236 let mut tokens = Vec::new();
1238 let mut main_thread_worker_state = MainThreadWorkerState::Idle;
1239 let mut running = 0;
1241 let prof = &cgcx.prof;
1242 let mut llvm_start_time: Option<VerboseTimingGuard<'_>> = None;
1244 // Run the message loop while there's still anything that needs message
1245 // processing. Note that as soon as codegen is aborted we simply want to
1246 // wait for all existing work to finish, so many of the conditions here
1247 // only apply if codegen hasn't been aborted as they represent pending
1251 || (!codegen_aborted
1252 && !(work_items.is_empty()
1253 && needs_fat_lto.is_empty()
1254 && needs_thin_lto.is_empty()
1255 && lto_import_only_modules.is_empty()
1256 && main_thread_worker_state == MainThreadWorkerState::Idle))
1258 // While there are still CGUs to be codegened, the coordinator has
1259 // to decide how to utilize the compiler processes implicit Token:
1260 // For codegenning more CGU or for running them through LLVM.
1262 if main_thread_worker_state == MainThreadWorkerState::Idle {
1263 // Compute the number of workers that will be running once we've taken as many
1264 // items from the work queue as we can, plus one for the main thread. It's not
1265 // critically important that we use this instead of just `running`, but it
1266 // prevents the `queue_full_enough` heuristic from fluctuating just because a
1267 // worker finished up and we decreased the `running` count, even though we're
1268 // just going to increase it right after this when we put a new worker to work.
1269 let extra_tokens = tokens.len().checked_sub(running).unwrap();
1270 let additional_running = std::cmp::min(extra_tokens, work_items.len());
1271 let anticipated_running = running + additional_running + 1;
1273 if !queue_full_enough(work_items.len(), anticipated_running) {
1274 // The queue is not full enough, codegen more items:
1275 if codegen_worker_send.send(Message::CodegenItem).is_err() {
1276 panic!("Could not send Message::CodegenItem to main thread")
1278 main_thread_worker_state = MainThreadWorkerState::Codegenning;
1280 // The queue is full enough to not let the worker
1281 // threads starve. Use the implicit Token to do some
1284 work_items.pop().expect("queue empty - queue_full_enough() broken?");
1285 let cgcx = CodegenContext {
1286 worker: get_worker_id(&mut free_worker_ids),
1289 maybe_start_llvm_timer(
1291 cgcx.config(item.module_kind()),
1292 &mut llvm_start_time,
1294 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1295 spawn_work(cgcx, item);
1298 } else if codegen_aborted {
1299 // don't queue up any more work if codegen was aborted, we're
1300 // just waiting for our existing children to finish
1302 // If we've finished everything related to normal codegen
1303 // then it must be the case that we've got some LTO work to do.
1304 // Perform the serial work here of figuring out what we're
1305 // going to LTO and then push a bunch of work items onto our
1307 if work_items.is_empty()
1309 && main_thread_worker_state == MainThreadWorkerState::Idle
1311 assert!(!started_lto);
1314 let needs_fat_lto = mem::take(&mut needs_fat_lto);
1315 let needs_thin_lto = mem::take(&mut needs_thin_lto);
1316 let import_only_modules = mem::take(&mut lto_import_only_modules);
1319 generate_lto_work(&cgcx, needs_fat_lto, needs_thin_lto, import_only_modules)
1321 let insertion_index = work_items
1322 .binary_search_by_key(&cost, |&(_, cost)| cost)
1323 .unwrap_or_else(|e| e);
1324 work_items.insert(insertion_index, (work, cost));
1325 if !cgcx.opts.debugging_opts.no_parallel_llvm {
1326 helper.request_token();
1331 // In this branch, we know that everything has been codegened,
1332 // so it's just a matter of determining whether the implicit
1333 // Token is free to use for LLVM work.
1334 match main_thread_worker_state {
1335 MainThreadWorkerState::Idle => {
1336 if let Some((item, _)) = work_items.pop() {
1337 let cgcx = CodegenContext {
1338 worker: get_worker_id(&mut free_worker_ids),
1341 maybe_start_llvm_timer(
1343 cgcx.config(item.module_kind()),
1344 &mut llvm_start_time,
1346 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1347 spawn_work(cgcx, item);
1349 // There is no unstarted work, so let the main thread
1350 // take over for a running worker. Otherwise the
1351 // implicit token would just go to waste.
1352 // We reduce the `running` counter by one. The
1353 // `tokens.truncate()` below will take care of
1354 // giving the Token back.
1355 debug_assert!(running > 0);
1357 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1360 MainThreadWorkerState::Codegenning => bug!(
1361 "codegen worker should not be codegenning after \
1362 codegen was already completed"
1364 MainThreadWorkerState::LLVMing => {
1365 // Already making good use of that token
1370 // Spin up what work we can, only doing this while we've got available
1371 // parallelism slots and work left to spawn.
1372 while !codegen_aborted && !work_items.is_empty() && running < tokens.len() {
1373 let (item, _) = work_items.pop().unwrap();
1375 maybe_start_llvm_timer(prof, cgcx.config(item.module_kind()), &mut llvm_start_time);
1378 CodegenContext { worker: get_worker_id(&mut free_worker_ids), ..cgcx.clone() };
1380 spawn_work(cgcx, item);
1384 // Relinquish accidentally acquired extra tokens
1385 tokens.truncate(running);
1387 // If a thread exits successfully then we drop a token associated
1388 // with that worker and update our `running` count. We may later
1389 // re-acquire a token to continue running more work. We may also not
1390 // actually drop a token here if the worker was running with an
1391 // "ephemeral token"
1392 let mut free_worker = |worker_id| {
1393 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
1394 main_thread_worker_state = MainThreadWorkerState::Idle;
1399 free_worker_ids.push(worker_id);
1402 let msg = coordinator_receive.recv().unwrap();
1403 match *msg.downcast::<Message<B>>().ok().unwrap() {
1404 // Save the token locally and the next turn of the loop will use
1405 // this to spawn a new unit of work, or it may get dropped
1406 // immediately if we have no more work to spawn.
1407 Message::Token(token) => {
1412 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
1413 // If the main thread token is used for LLVM work
1414 // at the moment, we turn that thread into a regular
1415 // LLVM worker thread, so the main thread is free
1416 // to react to codegen demand.
1417 main_thread_worker_state = MainThreadWorkerState::Idle;
1422 let msg = &format!("failed to acquire jobserver token: {}", e);
1423 shared_emitter.fatal(msg);
1424 // Exit the coordinator thread
1430 Message::CodegenDone { llvm_work_item, cost } => {
1431 // We keep the queue sorted by estimated processing cost,
1432 // so that more expensive items are processed earlier. This
1433 // is good for throughput as it gives the main thread more
1434 // time to fill up the queue and it avoids scheduling
1435 // expensive items to the end.
1436 // Note, however, that this is not ideal for memory
1437 // consumption, as LLVM module sizes are not evenly
1439 let insertion_index = work_items.binary_search_by_key(&cost, |&(_, cost)| cost);
1440 let insertion_index = match insertion_index {
1441 Ok(idx) | Err(idx) => idx,
1443 work_items.insert(insertion_index, (llvm_work_item, cost));
1445 if !cgcx.opts.debugging_opts.no_parallel_llvm {
1446 helper.request_token();
1448 assert!(!codegen_aborted);
1449 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1450 main_thread_worker_state = MainThreadWorkerState::Idle;
1453 Message::CodegenComplete => {
1454 codegen_done = true;
1455 assert!(!codegen_aborted);
1456 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1457 main_thread_worker_state = MainThreadWorkerState::Idle;
1460 // If codegen is aborted that means translation was aborted due
1461 // to some normal-ish compiler error. In this situation we want
1462 // to exit as soon as possible, but we want to make sure all
1463 // existing work has finished. Flag codegen as being done, and
1464 // then conditions above will ensure no more work is spawned but
1465 // we'll keep executing this loop until `running` hits 0.
1466 Message::CodegenAborted => {
1467 assert!(!codegen_aborted);
1468 codegen_done = true;
1469 codegen_aborted = true;
1470 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1472 Message::Done { result: Ok(compiled_module), worker_id } => {
1473 free_worker(worker_id);
1474 match compiled_module.kind {
1475 ModuleKind::Regular => {
1476 compiled_modules.push(compiled_module);
1478 ModuleKind::Metadata => {
1479 assert!(compiled_metadata_module.is_none());
1480 compiled_metadata_module = Some(compiled_module);
1482 ModuleKind::Allocator => {
1483 assert!(compiled_allocator_module.is_none());
1484 compiled_allocator_module = Some(compiled_module);
1488 Message::NeedsLink { module, worker_id } => {
1489 free_worker(worker_id);
1490 needs_link.push(module);
1492 Message::NeedsFatLTO { result, worker_id } => {
1493 assert!(!started_lto);
1494 free_worker(worker_id);
1495 needs_fat_lto.push(result);
1497 Message::NeedsThinLTO { name, thin_buffer, worker_id } => {
1498 assert!(!started_lto);
1499 free_worker(worker_id);
1500 needs_thin_lto.push((name, thin_buffer));
1502 Message::AddImportOnlyModule { module_data, work_product } => {
1503 assert!(!started_lto);
1504 assert!(!codegen_done);
1505 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1506 lto_import_only_modules.push((module_data, work_product));
1507 main_thread_worker_state = MainThreadWorkerState::Idle;
1509 // If the thread failed that means it panicked, so we abort immediately.
1510 Message::Done { result: Err(None), worker_id: _ } => {
1511 bug!("worker thread panicked");
1513 Message::Done { result: Err(Some(WorkerFatalError)), worker_id: _ } => {
1516 Message::CodegenItem => bug!("the coordinator should not receive codegen requests"),
1520 let needs_link = mem::take(&mut needs_link);
1521 if !needs_link.is_empty() {
1522 assert!(compiled_modules.is_empty());
1523 let diag_handler = cgcx.create_diag_handler();
1524 let module = B::run_link(&cgcx, &diag_handler, needs_link).map_err(|_| ())?;
1525 let module = unsafe {
1526 B::codegen(&cgcx, &diag_handler, module, cgcx.config(ModuleKind::Regular))
1529 compiled_modules.push(module);
1532 // Drop to print timings
1533 drop(llvm_start_time);
1535 // Regardless of what order these modules completed in, report them to
1536 // the backend in the same order every time to ensure that we're handing
1537 // out deterministic results.
1538 compiled_modules.sort_by(|a, b| a.name.cmp(&b.name));
1540 Ok(CompiledModules {
1541 modules: compiled_modules,
1542 metadata_module: compiled_metadata_module,
1543 allocator_module: compiled_allocator_module,
1547 // A heuristic that determines if we have enough LLVM WorkItems in the
1548 // queue so that the main thread can do LLVM work instead of codegen
1549 fn queue_full_enough(items_in_queue: usize, workers_running: usize) -> bool {
1550 // This heuristic scales ahead-of-time codegen according to available
1551 // concurrency, as measured by `workers_running`. The idea is that the
1552 // more concurrency we have available, the more demand there will be for
1553 // work items, and the fuller the queue should be kept to meet demand.
1554 // An important property of this approach is that we codegen ahead of
1555 // time only as much as necessary, so as to keep fewer LLVM modules in
1556 // memory at once, thereby reducing memory consumption.
1558 // When the number of workers running is less than the max concurrency
1559 // available to us, this heuristic can cause us to instruct the main
1560 // thread to work on an LLVM item (that is, tell it to "LLVM") instead
1561 // of codegen, even though it seems like it *should* be codegenning so
1562 // that we can create more work items and spawn more LLVM workers.
1564 // But this is not a problem. When the main thread is told to LLVM,
1565 // according to this heuristic and how work is scheduled, there is
1566 // always at least one item in the queue, and therefore at least one
1567 // pending jobserver token request. If there *is* more concurrency
1568 // available, we will immediately receive a token, which will upgrade
1569 // the main thread's LLVM worker to a real one (conceptually), and free
1570 // up the main thread to codegen if necessary. On the other hand, if
1571 // there isn't more concurrency, then the main thread working on an LLVM
1572 // item is appropriate, as long as the queue is full enough for demand.
1574 // Speaking of which, how full should we keep the queue? Probably less
1575 // full than you'd think. A lot has to go wrong for the queue not to be
1576 // full enough and for that to have a negative effect on compile times.
1578 // Workers are unlikely to finish at exactly the same time, so when one
1579 // finishes and takes another work item off the queue, we often have
1580 // ample time to codegen at that point before the next worker finishes.
1581 // But suppose that codegen takes so long that the workers exhaust the
1582 // queue, and we have one or more workers that have nothing to work on.
1583 // Well, it might not be so bad. Of all the LLVM modules we create and
1584 // optimize, one has to finish last. It's not necessarily the case that
1585 // by losing some concurrency for a moment, we delay the point at which
1586 // that last LLVM module is finished and the rest of compilation can
1587 // proceed. Also, when we can't take advantage of some concurrency, we
1588 // give tokens back to the job server. That enables some other rustc to
1589 // potentially make use of the available concurrency. That could even
1590 // *decrease* overall compile time if we're lucky. But yes, if no other
1591 // rustc can make use of the concurrency, then we've squandered it.
1593 // However, keeping the queue full is also beneficial when we have a
1594 // surge in available concurrency. Then items can be taken from the
1595 // queue immediately, without having to wait for codegen.
1597 // So, the heuristic below tries to keep one item in the queue for every
1598 // four running workers. Based on limited benchmarking, this appears to
1599 // be more than sufficient to avoid increasing compilation times.
1600 let quarter_of_workers = workers_running - 3 * workers_running / 4;
1601 items_in_queue > 0 && items_in_queue >= quarter_of_workers
1604 fn maybe_start_llvm_timer<'a>(
1605 prof: &'a SelfProfilerRef,
1606 config: &ModuleConfig,
1607 llvm_start_time: &mut Option<VerboseTimingGuard<'a>>,
1609 if config.time_module && llvm_start_time.is_none() {
1610 *llvm_start_time = Some(prof.extra_verbose_generic_activity("LLVM_passes", "crate"));
1615 /// `FatalError` is explicitly not `Send`.
1617 pub struct WorkerFatalError;
1619 fn spawn_work<B: ExtraBackendMethods>(cgcx: CodegenContext<B>, work: WorkItem<B>) {
1620 B::spawn_named_thread(cgcx.time_trace, work.short_description(), move || {
1621 // Set up a destructor which will fire off a message that we're done as
1623 struct Bomb<B: ExtraBackendMethods> {
1624 coordinator_send: Sender<Box<dyn Any + Send>>,
1625 result: Option<Result<WorkItemResult<B>, FatalError>>,
1628 impl<B: ExtraBackendMethods> Drop for Bomb<B> {
1629 fn drop(&mut self) {
1630 let worker_id = self.worker_id;
1631 let msg = match self.result.take() {
1632 Some(Ok(WorkItemResult::Compiled(m))) => {
1633 Message::Done::<B> { result: Ok(m), worker_id }
1635 Some(Ok(WorkItemResult::NeedsLink(m))) => {
1636 Message::NeedsLink::<B> { module: m, worker_id }
1638 Some(Ok(WorkItemResult::NeedsFatLTO(m))) => {
1639 Message::NeedsFatLTO::<B> { result: m, worker_id }
1641 Some(Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))) => {
1642 Message::NeedsThinLTO::<B> { name, thin_buffer, worker_id }
1644 Some(Err(FatalError)) => {
1645 Message::Done::<B> { result: Err(Some(WorkerFatalError)), worker_id }
1647 None => Message::Done::<B> { result: Err(None), worker_id },
1649 drop(self.coordinator_send.send(Box::new(msg)));
1653 let mut bomb = Bomb::<B> {
1654 coordinator_send: cgcx.coordinator_send.clone(),
1656 worker_id: cgcx.worker,
1659 // Execute the work itself, and if it finishes successfully then flag
1660 // ourselves as a success as well.
1662 // Note that we ignore any `FatalError` coming out of `execute_work_item`,
1663 // as a diagnostic was already sent off to the main thread - just
1664 // surface that there was an error in this worker.
1666 let _prof_timer = work.start_profiling(&cgcx);
1667 Some(execute_work_item(&cgcx, work))
1670 .expect("failed to spawn thread");
1673 enum SharedEmitterMessage {
1674 Diagnostic(Diagnostic),
1675 InlineAsmError(u32, String, Level, Option<(String, Vec<InnerSpan>)>),
1681 pub struct SharedEmitter {
1682 sender: Sender<SharedEmitterMessage>,
1685 pub struct SharedEmitterMain {
1686 receiver: Receiver<SharedEmitterMessage>,
1689 impl SharedEmitter {
1690 pub fn new() -> (SharedEmitter, SharedEmitterMain) {
1691 let (sender, receiver) = channel();
1693 (SharedEmitter { sender }, SharedEmitterMain { receiver })
1696 pub fn inline_asm_error(
1701 source: Option<(String, Vec<InnerSpan>)>,
1703 drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source)));
1706 pub fn fatal(&self, msg: &str) {
1707 drop(self.sender.send(SharedEmitterMessage::Fatal(msg.to_string())));
1711 impl Emitter for SharedEmitter {
1712 fn emit_diagnostic(&mut self, diag: &rustc_errors::Diagnostic) {
1713 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
1714 msg: diag.message(),
1715 code: diag.code.clone(),
1718 for child in &diag.children {
1719 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
1720 msg: child.message(),
1725 drop(self.sender.send(SharedEmitterMessage::AbortIfErrors));
1727 fn source_map(&self) -> Option<&Lrc<SourceMap>> {
1732 impl SharedEmitterMain {
1733 pub fn check(&self, sess: &Session, blocking: bool) {
1735 let message = if blocking {
1736 match self.receiver.recv() {
1737 Ok(message) => Ok(message),
1741 match self.receiver.try_recv() {
1742 Ok(message) => Ok(message),
1748 Ok(SharedEmitterMessage::Diagnostic(diag)) => {
1749 let handler = sess.diagnostic();
1750 let mut d = rustc_errors::Diagnostic::new(diag.lvl, &diag.msg);
1751 if let Some(code) = diag.code {
1754 handler.emit_diagnostic(&d);
1756 Ok(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source)) => {
1757 let msg = msg.strip_prefix("error: ").unwrap_or(&msg);
1759 let mut err = match level {
1760 Level::Error { lint: false } => sess.struct_err(&msg),
1761 Level::Warning => sess.struct_warn(&msg),
1762 Level::Note => sess.struct_note_without_error(&msg),
1763 _ => bug!("Invalid inline asm diagnostic level"),
1766 // If the cookie is 0 then we don't have span information.
1768 let pos = BytePos::from_u32(cookie);
1769 let span = Span::with_root_ctxt(pos, pos);
1773 // Point to the generated assembly if it is available.
1774 if let Some((buffer, spans)) = source {
1777 .new_source_file(FileName::inline_asm_source_code(&buffer), buffer);
1778 let source_span = Span::with_root_ctxt(source.start_pos, source.end_pos);
1780 spans.iter().map(|sp| source_span.from_inner(*sp)).collect();
1781 err.span_note(spans, "instantiated into assembly here");
1786 Ok(SharedEmitterMessage::AbortIfErrors) => {
1787 sess.abort_if_errors();
1789 Ok(SharedEmitterMessage::Fatal(msg)) => {
1800 pub struct OngoingCodegen<B: ExtraBackendMethods> {
1802 pub metadata: EncodedMetadata,
1803 pub crate_info: CrateInfo,
1804 pub coordinator_send: Sender<Box<dyn Any + Send>>,
1805 pub codegen_worker_receive: Receiver<Message<B>>,
1806 pub shared_emitter_main: SharedEmitterMain,
1807 pub future: thread::JoinHandle<Result<CompiledModules, ()>>,
1808 pub output_filenames: Arc<OutputFilenames>,
1811 impl<B: ExtraBackendMethods> OngoingCodegen<B> {
1812 pub fn join(self, sess: &Session) -> (CodegenResults, FxHashMap<WorkProductId, WorkProduct>) {
1813 let _timer = sess.timer("finish_ongoing_codegen");
1815 self.shared_emitter_main.check(sess, true);
1816 let future = self.future;
1817 let compiled_modules = sess.time("join_worker_thread", || match future.join() {
1818 Ok(Ok(compiled_modules)) => compiled_modules,
1820 sess.abort_if_errors();
1821 panic!("expected abort due to worker thread errors")
1824 bug!("panic during codegen/LLVM phase");
1828 sess.cgu_reuse_tracker.check_expected_reuse(sess.diagnostic());
1830 sess.abort_if_errors();
1833 copy_all_cgu_workproducts_to_incr_comp_cache_dir(sess, &compiled_modules);
1834 produce_final_output_artifacts(sess, &compiled_modules, &self.output_filenames);
1836 // FIXME: time_llvm_passes support - does this use a global context or
1838 if sess.codegen_units() == 1 && sess.time_llvm_passes() {
1839 self.backend.print_pass_timings()
1844 metadata: self.metadata,
1845 crate_info: self.crate_info,
1847 modules: compiled_modules.modules,
1848 allocator_module: compiled_modules.allocator_module,
1849 metadata_module: compiled_modules.metadata_module,
1855 pub fn submit_pre_codegened_module_to_llvm(
1858 module: ModuleCodegen<B::Module>,
1860 self.wait_for_signal_to_codegen_item();
1861 self.check_for_errors(tcx.sess);
1863 // These are generally cheap and won't throw off scheduling.
1865 submit_codegened_module_to_llvm(&self.backend, &self.coordinator_send, module, cost);
1868 pub fn codegen_finished(&self, tcx: TyCtxt<'_>) {
1869 self.wait_for_signal_to_codegen_item();
1870 self.check_for_errors(tcx.sess);
1871 drop(self.coordinator_send.send(Box::new(Message::CodegenComplete::<B>)));
1874 /// Consumes this context indicating that codegen was entirely aborted, and
1875 /// we need to exit as quickly as possible.
1877 /// This method blocks the current thread until all worker threads have
1878 /// finished, and all worker threads should have exited or be real close to
1879 /// exiting at this point.
1880 pub fn codegen_aborted(self) {
1881 // Signal to the coordinator it should spawn no more work and start
1883 drop(self.coordinator_send.send(Box::new(Message::CodegenAborted::<B>)));
1884 drop(self.future.join());
1887 pub fn check_for_errors(&self, sess: &Session) {
1888 self.shared_emitter_main.check(sess, false);
1891 pub fn wait_for_signal_to_codegen_item(&self) {
1892 match self.codegen_worker_receive.recv() {
1893 Ok(Message::CodegenItem) => {
1896 Ok(_) => panic!("unexpected message"),
1898 // One of the LLVM threads must have panicked, fall through so
1899 // error handling can be reached.
1905 pub fn submit_codegened_module_to_llvm<B: ExtraBackendMethods>(
1907 tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
1908 module: ModuleCodegen<B::Module>,
1911 let llvm_work_item = WorkItem::Optimize(module);
1912 drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost })));
1915 pub fn submit_post_lto_module_to_llvm<B: ExtraBackendMethods>(
1917 tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
1918 module: CachedModuleCodegen,
1920 let llvm_work_item = WorkItem::CopyPostLtoArtifacts(module);
1921 drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost: 0 })));
1924 pub fn submit_pre_lto_module_to_llvm<B: ExtraBackendMethods>(
1927 tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
1928 module: CachedModuleCodegen,
1930 let filename = pre_lto_bitcode_filename(&module.name);
1931 let bc_path = in_incr_comp_dir_sess(tcx.sess, &filename);
1932 let file = fs::File::open(&bc_path)
1933 .unwrap_or_else(|e| panic!("failed to open bitcode file `{}`: {}", bc_path.display(), e));
1936 Mmap::map(file).unwrap_or_else(|e| {
1937 panic!("failed to mmap bitcode file `{}`: {}", bc_path.display(), e)
1940 // Schedule the module to be loaded
1941 drop(tx_to_llvm_workers.send(Box::new(Message::AddImportOnlyModule::<B> {
1942 module_data: SerializedModule::FromUncompressedFile(mmap),
1943 work_product: module.source,
1947 pub fn pre_lto_bitcode_filename(module_name: &str) -> String {
1948 format!("{}.{}", module_name, PRE_LTO_BC_EXT)
1951 fn msvc_imps_needed(tcx: TyCtxt<'_>) -> bool {
1952 // This should never be true (because it's not supported). If it is true,
1953 // something is wrong with commandline arg validation.
1955 !(tcx.sess.opts.cg.linker_plugin_lto.enabled()
1956 && tcx.sess.target.is_like_windows
1957 && tcx.sess.opts.cg.prefer_dynamic)
1960 tcx.sess.target.is_like_windows &&
1961 tcx.sess.crate_types().iter().any(|ct| *ct == CrateType::Rlib) &&
1962 // ThinLTO can't handle this workaround in all cases, so we don't
1963 // emit the `__imp_` symbols. Instead we make them unnecessary by disallowing
1964 // dynamic linking when linker plugin LTO is enabled.
1965 !tcx.sess.opts.cg.linker_plugin_lto.enabled()