1 use super::link::{self, ensure_removed};
2 use super::linker::LinkerInfo;
3 use super::lto::{self, SerializedModule};
4 use super::symbol_export::symbol_name_for_instance_in_crate;
7 CachedModuleCodegen, CodegenResults, CompiledModule, CrateInfo, ModuleCodegen, ModuleKind,
11 use jobserver::{Acquired, Client};
12 use rustc_data_structures::fx::FxHashMap;
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_middle::dep_graph::{WorkProduct, WorkProductId};
25 use rustc_middle::middle::cstore::EncodedMetadata;
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, SanitizerSet, SwitchWithOptPath};
31 use rustc_session::Session;
32 use rustc_span::source_map::SourceMap;
33 use rustc_span::symbol::{sym, Symbol};
34 use rustc_span::{BytePos, FileName, InnerSpan, Pos, Span};
35 use rustc_target::spec::{MergeFunctions, PanicStrategy};
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>,
87 pub sanitizer: SanitizerSet,
88 pub sanitizer_recover: SanitizerSet,
89 pub sanitizer_memory_track_origins: usize,
91 // Flags indicating which outputs to produce.
92 pub emit_pre_lto_bc: bool,
93 pub emit_no_opt_bc: bool,
97 pub emit_obj: EmitObj,
98 pub bc_cmdline: String,
100 // Miscellaneous flags. These are mostly copied from command-line
102 pub verify_llvm_ir: bool,
103 pub no_prepopulate_passes: bool,
104 pub no_builtins: bool,
105 pub time_module: bool,
106 pub vectorize_loop: bool,
107 pub vectorize_slp: bool,
108 pub merge_functions: bool,
109 pub inline_threshold: Option<usize>,
110 pub new_llvm_pass_manager: bool,
111 pub emit_lifetime_markers: bool,
119 is_compiler_builtins: bool,
121 // If it's a regular module, use `$regular`, otherwise use `$other`.
122 // `$regular` and `$other` are evaluated lazily.
123 macro_rules! if_regular {
124 ($regular: expr, $other: expr) => {
125 if let ModuleKind::Regular = kind { $regular } else { $other }
129 let opt_level_and_size = if_regular!(Some(sess.opts.optimize), None);
131 let save_temps = sess.opts.cg.save_temps;
133 let should_emit_obj = sess.opts.output_types.contains_key(&OutputType::Exe)
135 ModuleKind::Regular => sess.opts.output_types.contains_key(&OutputType::Object),
136 ModuleKind::Allocator => false,
137 ModuleKind::Metadata => sess.opts.output_types.contains_key(&OutputType::Metadata),
140 let emit_obj = if !should_emit_obj {
142 } else if sess.target.obj_is_bitcode
143 || (sess.opts.cg.linker_plugin_lto.enabled() && !no_builtins)
145 // This case is selected if the target uses objects as bitcode, or
146 // if linker plugin LTO is enabled. In the linker plugin LTO case
147 // the assumption is that the final link-step will read the bitcode
148 // and convert it to object code. This may be done by either the
149 // native linker or rustc itself.
151 // Note, however, that the linker-plugin-lto requested here is
152 // explicitly ignored for `#![no_builtins]` crates. These crates are
153 // specifically ignored by rustc's LTO passes and wouldn't work if
154 // loaded into the linker. These crates define symbols that LLVM
155 // lowers intrinsics to, and these symbol dependencies aren't known
156 // until after codegen. As a result any crate marked
157 // `#![no_builtins]` is assumed to not participate in LTO and
158 // instead goes on to generate object code.
160 } else if need_bitcode_in_object(sess) {
161 EmitObj::ObjectCode(BitcodeSection::Full)
163 EmitObj::ObjectCode(BitcodeSection::None)
169 let mut passes = sess.opts.cg.passes.clone();
170 // compiler_builtins overrides the codegen-units settings,
171 // which is incompatible with -Zprofile which requires that
172 // only a single codegen unit is used per crate.
173 if sess.opts.debugging_opts.profile && !is_compiler_builtins {
174 passes.push("insert-gcov-profiling".to_owned());
177 // The rustc option `-Zinstrument_coverage` injects intrinsic calls to
178 // `llvm.instrprof.increment()`, which requires the LLVM `instrprof` pass.
179 if sess.opts.debugging_opts.instrument_coverage {
180 passes.push("instrprof".to_owned());
187 opt_level: opt_level_and_size,
188 opt_size: opt_level_and_size,
190 pgo_gen: if_regular!(
191 sess.opts.cg.profile_generate.clone(),
192 SwitchWithOptPath::Disabled
194 pgo_use: if_regular!(sess.opts.cg.profile_use.clone(), None),
196 sanitizer: if_regular!(sess.opts.debugging_opts.sanitizer, SanitizerSet::empty()),
197 sanitizer_recover: if_regular!(
198 sess.opts.debugging_opts.sanitizer_recover,
199 SanitizerSet::empty()
201 sanitizer_memory_track_origins: if_regular!(
202 sess.opts.debugging_opts.sanitizer_memory_track_origins,
206 emit_pre_lto_bc: if_regular!(
207 save_temps || need_pre_lto_bitcode_for_incr_comp(sess),
210 emit_no_opt_bc: if_regular!(save_temps, false),
211 emit_bc: if_regular!(
212 save_temps || sess.opts.output_types.contains_key(&OutputType::Bitcode),
215 emit_ir: if_regular!(
216 sess.opts.output_types.contains_key(&OutputType::LlvmAssembly),
219 emit_asm: if_regular!(
220 sess.opts.output_types.contains_key(&OutputType::Assembly),
224 bc_cmdline: sess.target.bitcode_llvm_cmdline.clone(),
226 verify_llvm_ir: sess.verify_llvm_ir(),
227 no_prepopulate_passes: sess.opts.cg.no_prepopulate_passes,
228 no_builtins: no_builtins || sess.target.no_builtins,
230 // Exclude metadata and allocator modules from time_passes output,
231 // since they throw off the "LLVM passes" measurement.
232 time_module: if_regular!(true, false),
234 // Copy what clang does by turning on loop vectorization at O2 and
235 // slp vectorization at O3.
236 vectorize_loop: !sess.opts.cg.no_vectorize_loops
237 && (sess.opts.optimize == config::OptLevel::Default
238 || sess.opts.optimize == config::OptLevel::Aggressive),
239 vectorize_slp: !sess.opts.cg.no_vectorize_slp
240 && sess.opts.optimize == config::OptLevel::Aggressive,
242 // Some targets (namely, NVPTX) interact badly with the
243 // MergeFunctions pass. This is because MergeFunctions can generate
244 // new function calls which may interfere with the target calling
245 // convention; e.g. for the NVPTX target, PTX kernels should not
246 // call other PTX kernels. MergeFunctions can also be configured to
247 // generate aliases instead, but aliases are not supported by some
248 // backends (again, NVPTX). Therefore, allow targets to opt out of
249 // the MergeFunctions pass, but otherwise keep the pass enabled (at
250 // O2 and O3) since it can be useful for reducing code size.
251 merge_functions: match sess
255 .unwrap_or(sess.target.merge_functions)
257 MergeFunctions::Disabled => false,
258 MergeFunctions::Trampolines | MergeFunctions::Aliases => {
259 sess.opts.optimize == config::OptLevel::Default
260 || sess.opts.optimize == config::OptLevel::Aggressive
264 inline_threshold: sess.opts.cg.inline_threshold,
265 new_llvm_pass_manager: sess.opts.debugging_opts.new_llvm_pass_manager,
266 emit_lifetime_markers: sess.emit_lifetime_markers(),
270 pub fn bitcode_needed(&self) -> bool {
272 || self.emit_obj == EmitObj::Bitcode
273 || self.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full)
277 /// Configuration passed to the function returned by the `target_machine_factory`.
278 pub struct TargetMachineFactoryConfig {
279 /// Split DWARF is enabled in LLVM by checking that `TM.MCOptions.SplitDwarfFile` isn't empty,
280 /// so the path to the dwarf object has to be provided when we create the target machine.
281 /// This can be ignored by backends which do not need it for their Split DWARF support.
282 pub split_dwarf_file: Option<PathBuf>,
285 impl TargetMachineFactoryConfig {
287 cgcx: &CodegenContext<impl WriteBackendMethods>,
289 ) -> TargetMachineFactoryConfig {
290 let split_dwarf_file = if cgcx.target_can_use_split_dwarf {
291 cgcx.output_filenames.split_dwarf_path(cgcx.split_debuginfo, Some(module_name))
295 TargetMachineFactoryConfig { split_dwarf_file }
299 pub type TargetMachineFactoryFn<B> = Arc<
300 dyn Fn(TargetMachineFactoryConfig) -> Result<<B as WriteBackendMethods>::TargetMachine, String>
305 pub type ExportedSymbols = FxHashMap<CrateNum, Arc<Vec<(String, SymbolExportLevel)>>>;
307 /// Additional resources used by optimize_and_codegen (not module specific)
309 pub struct CodegenContext<B: WriteBackendMethods> {
310 // Resources needed when running LTO
312 pub prof: SelfProfilerRef,
314 pub no_landing_pads: bool,
315 pub save_temps: bool,
316 pub fewer_names: bool,
317 pub exported_symbols: Option<Arc<ExportedSymbols>>,
318 pub opts: Arc<config::Options>,
319 pub crate_types: Vec<CrateType>,
320 pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>,
321 pub output_filenames: Arc<OutputFilenames>,
322 pub regular_module_config: Arc<ModuleConfig>,
323 pub metadata_module_config: Arc<ModuleConfig>,
324 pub allocator_module_config: Arc<ModuleConfig>,
325 pub tm_factory: TargetMachineFactoryFn<B>,
326 pub msvc_imps_needed: bool,
327 pub is_pe_coff: bool,
328 pub target_can_use_split_dwarf: bool,
329 pub target_pointer_width: u32,
330 pub target_arch: String,
331 pub debuginfo: config::DebugInfo,
332 pub split_debuginfo: rustc_target::spec::SplitDebuginfo,
334 // Number of cgus excluding the allocator/metadata modules
335 pub total_cgus: usize,
336 // Handler to use for diagnostics produced during codegen.
337 pub diag_emitter: SharedEmitter,
338 // LLVM optimizations for which we want to print remarks.
340 // Worker thread number
342 // The incremental compilation session directory, or None if we are not
343 // compiling incrementally
344 pub incr_comp_session_dir: Option<PathBuf>,
345 // Used to update CGU re-use information during the thinlto phase.
346 pub cgu_reuse_tracker: CguReuseTracker,
347 // Channel back to the main control thread to send messages to
348 pub coordinator_send: Sender<Box<dyn Any + Send>>,
351 impl<B: WriteBackendMethods> CodegenContext<B> {
352 pub fn create_diag_handler(&self) -> Handler {
353 Handler::with_emitter(true, None, Box::new(self.diag_emitter.clone()))
356 pub fn config(&self, kind: ModuleKind) -> &ModuleConfig {
358 ModuleKind::Regular => &self.regular_module_config,
359 ModuleKind::Metadata => &self.metadata_module_config,
360 ModuleKind::Allocator => &self.allocator_module_config,
365 fn generate_lto_work<B: ExtraBackendMethods>(
366 cgcx: &CodegenContext<B>,
367 needs_fat_lto: Vec<FatLTOInput<B>>,
368 needs_thin_lto: Vec<(String, B::ThinBuffer)>,
369 import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>,
370 ) -> Vec<(WorkItem<B>, u64)> {
371 let _prof_timer = cgcx.prof.generic_activity("codegen_generate_lto_work");
373 let (lto_modules, copy_jobs) = if !needs_fat_lto.is_empty() {
374 assert!(needs_thin_lto.is_empty());
376 B::run_fat_lto(cgcx, needs_fat_lto, import_only_modules).unwrap_or_else(|e| e.raise());
377 (vec![lto_module], vec![])
379 assert!(needs_fat_lto.is_empty());
380 B::run_thin_lto(cgcx, needs_thin_lto, import_only_modules).unwrap_or_else(|e| e.raise())
386 let cost = module.cost();
387 (WorkItem::LTO(module), cost)
389 .chain(copy_jobs.into_iter().map(|wp| {
391 WorkItem::CopyPostLtoArtifacts(CachedModuleCodegen {
392 name: wp.cgu_name.clone(),
401 pub struct CompiledModules {
402 pub modules: Vec<CompiledModule>,
403 pub metadata_module: Option<CompiledModule>,
404 pub allocator_module: Option<CompiledModule>,
407 fn need_bitcode_in_object(sess: &Session) -> bool {
408 let requested_for_rlib = sess.opts.cg.embed_bitcode
409 && sess.crate_types().contains(&CrateType::Rlib)
410 && sess.opts.output_types.contains_key(&OutputType::Exe);
411 let forced_by_target = sess.target.forces_embed_bitcode;
412 requested_for_rlib || forced_by_target
415 fn need_pre_lto_bitcode_for_incr_comp(sess: &Session) -> bool {
416 if sess.opts.incremental.is_none() {
422 Lto::Fat | Lto::Thin | Lto::ThinLocal => true,
426 pub fn start_async_codegen<B: ExtraBackendMethods>(
429 metadata: EncodedMetadata,
431 ) -> OngoingCodegen<B> {
432 let (coordinator_send, coordinator_receive) = channel();
435 let crate_name = tcx.crate_name(LOCAL_CRATE);
436 let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID);
437 let no_builtins = tcx.sess.contains_name(crate_attrs, sym::no_builtins);
438 let is_compiler_builtins = tcx.sess.contains_name(crate_attrs, sym::compiler_builtins);
439 let subsystem = tcx.sess.first_attr_value_str_by_name(crate_attrs, sym::windows_subsystem);
440 let windows_subsystem = subsystem.map(|subsystem| {
441 if subsystem != sym::windows && subsystem != sym::console {
442 tcx.sess.fatal(&format!(
443 "invalid windows subsystem `{}`, only \
444 `windows` and `console` are allowed",
448 subsystem.to_string()
451 let linker_info = LinkerInfo::new(tcx);
452 let crate_info = CrateInfo::new(tcx);
455 ModuleConfig::new(ModuleKind::Regular, sess, no_builtins, is_compiler_builtins);
456 let metadata_config =
457 ModuleConfig::new(ModuleKind::Metadata, sess, no_builtins, is_compiler_builtins);
458 let allocator_config =
459 ModuleConfig::new(ModuleKind::Allocator, sess, no_builtins, is_compiler_builtins);
461 let (shared_emitter, shared_emitter_main) = SharedEmitter::new();
462 let (codegen_worker_send, codegen_worker_receive) = channel();
464 let coordinator_thread = start_executing_work(
472 sess.jobserver.clone(),
473 Arc::new(regular_config),
474 Arc::new(metadata_config),
475 Arc::new(allocator_config),
476 coordinator_send.clone(),
488 codegen_worker_receive,
490 future: coordinator_thread,
491 output_filenames: tcx.output_filenames(LOCAL_CRATE),
495 fn copy_all_cgu_workproducts_to_incr_comp_cache_dir(
497 compiled_modules: &CompiledModules,
498 ) -> FxHashMap<WorkProductId, WorkProduct> {
499 let mut work_products = FxHashMap::default();
501 if sess.opts.incremental.is_none() {
502 return work_products;
505 let _timer = sess.timer("copy_all_cgu_workproducts_to_incr_comp_cache_dir");
507 for module in compiled_modules.modules.iter().filter(|m| m.kind == ModuleKind::Regular) {
508 let path = module.object.as_ref().cloned();
510 if let Some((id, product)) =
511 copy_cgu_workproduct_to_incr_comp_cache_dir(sess, &module.name, &path)
513 work_products.insert(id, product);
520 fn produce_final_output_artifacts(
522 compiled_modules: &CompiledModules,
523 crate_output: &OutputFilenames,
525 let mut user_wants_bitcode = false;
526 let mut user_wants_objects = false;
528 // Produce final compile outputs.
529 let copy_gracefully = |from: &Path, to: &Path| {
530 if let Err(e) = fs::copy(from, to) {
531 sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
535 let copy_if_one_unit = |output_type: OutputType, keep_numbered: bool| {
536 if compiled_modules.modules.len() == 1 {
537 // 1) Only one codegen unit. In this case it's no difficulty
538 // to copy `foo.0.x` to `foo.x`.
539 let module_name = Some(&compiled_modules.modules[0].name[..]);
540 let path = crate_output.temp_path(output_type, module_name);
541 copy_gracefully(&path, &crate_output.path(output_type));
542 if !sess.opts.cg.save_temps && !keep_numbered {
543 // The user just wants `foo.x`, not `foo.#module-name#.x`.
544 ensure_removed(sess.diagnostic(), &path);
547 let ext = crate_output
548 .temp_path(output_type, None)
555 if crate_output.outputs.contains_key(&output_type) {
556 // 2) Multiple codegen units, with `--emit foo=some_name`. We have
557 // no good solution for this case, so warn the user.
559 "ignoring emit path because multiple .{} files \
563 } else if crate_output.single_output_file.is_some() {
564 // 3) Multiple codegen units, with `-o some_name`. We have
565 // no good solution for this case, so warn the user.
567 "ignoring -o because multiple .{} files \
572 // 4) Multiple codegen units, but no explicit name. We
573 // just leave the `foo.0.x` files in place.
574 // (We don't have to do any work in this case.)
579 // Flag to indicate whether the user explicitly requested bitcode.
580 // Otherwise, we produced it only as a temporary output, and will need
582 for output_type in crate_output.outputs.keys() {
584 OutputType::Bitcode => {
585 user_wants_bitcode = true;
586 // Copy to .bc, but always keep the .0.bc. There is a later
587 // check to figure out if we should delete .0.bc files, or keep
588 // them for making an rlib.
589 copy_if_one_unit(OutputType::Bitcode, true);
591 OutputType::LlvmAssembly => {
592 copy_if_one_unit(OutputType::LlvmAssembly, false);
594 OutputType::Assembly => {
595 copy_if_one_unit(OutputType::Assembly, false);
597 OutputType::Object => {
598 user_wants_objects = true;
599 copy_if_one_unit(OutputType::Object, true);
601 OutputType::Mir | OutputType::Metadata | OutputType::Exe | OutputType::DepInfo => {}
605 // Clean up unwanted temporary files.
607 // We create the following files by default:
608 // - #crate#.#module-name#.bc
609 // - #crate#.#module-name#.o
610 // - #crate#.crate.metadata.bc
611 // - #crate#.crate.metadata.o
612 // - #crate#.o (linked from crate.##.o)
613 // - #crate#.bc (copied from crate.##.bc)
614 // We may create additional files if requested by the user (through
615 // `-C save-temps` or `--emit=` flags).
617 if !sess.opts.cg.save_temps {
618 // Remove the temporary .#module-name#.o objects. If the user didn't
619 // explicitly request bitcode (with --emit=bc), and the bitcode is not
620 // needed for building an rlib, then we must remove .#module-name#.bc as
623 // Specific rules for keeping .#module-name#.bc:
624 // - If the user requested bitcode (`user_wants_bitcode`), and
625 // codegen_units > 1, then keep it.
626 // - If the user requested bitcode but codegen_units == 1, then we
627 // can toss .#module-name#.bc because we copied it to .bc earlier.
628 // - If we're not building an rlib and the user didn't request
629 // bitcode, then delete .#module-name#.bc.
630 // If you change how this works, also update back::link::link_rlib,
631 // where .#module-name#.bc files are (maybe) deleted after making an
633 let needs_crate_object = crate_output.outputs.contains_key(&OutputType::Exe);
635 let keep_numbered_bitcode = user_wants_bitcode && sess.codegen_units() > 1;
637 let keep_numbered_objects =
638 needs_crate_object || (user_wants_objects && sess.codegen_units() > 1);
640 for module in compiled_modules.modules.iter() {
641 if let Some(ref path) = module.object {
642 if !keep_numbered_objects {
643 ensure_removed(sess.diagnostic(), path);
647 if let Some(ref path) = module.dwarf_object {
648 if !keep_numbered_objects {
649 ensure_removed(sess.diagnostic(), path);
653 if let Some(ref path) = module.bytecode {
654 if !keep_numbered_bitcode {
655 ensure_removed(sess.diagnostic(), path);
660 if !user_wants_bitcode {
661 if let Some(ref metadata_module) = compiled_modules.metadata_module {
662 if let Some(ref path) = metadata_module.bytecode {
663 ensure_removed(sess.diagnostic(), &path);
667 if let Some(ref allocator_module) = compiled_modules.allocator_module {
668 if let Some(ref path) = allocator_module.bytecode {
669 ensure_removed(sess.diagnostic(), path);
675 // We leave the following files around by default:
677 // - #crate#.crate.metadata.o
679 // These are used in linking steps and will be cleaned up afterward.
682 pub enum WorkItem<B: WriteBackendMethods> {
683 /// Optimize a newly codegened, totally unoptimized module.
684 Optimize(ModuleCodegen<B::Module>),
685 /// Copy the post-LTO artifacts from the incremental cache to the output
687 CopyPostLtoArtifacts(CachedModuleCodegen),
688 /// Performs (Thin)LTO on the given module.
689 LTO(lto::LtoModuleCodegen<B>),
692 impl<B: WriteBackendMethods> WorkItem<B> {
693 pub fn module_kind(&self) -> ModuleKind {
695 WorkItem::Optimize(ref m) => m.kind,
696 WorkItem::CopyPostLtoArtifacts(_) | WorkItem::LTO(_) => ModuleKind::Regular,
700 fn start_profiling<'a>(&self, cgcx: &'a CodegenContext<B>) -> TimingGuard<'a> {
702 WorkItem::Optimize(ref m) => {
703 cgcx.prof.generic_activity_with_arg("codegen_module_optimize", &m.name[..])
705 WorkItem::CopyPostLtoArtifacts(ref m) => cgcx
707 .generic_activity_with_arg("codegen_copy_artifacts_from_incr_cache", &m.name[..]),
708 WorkItem::LTO(ref m) => {
709 cgcx.prof.generic_activity_with_arg("codegen_module_perform_lto", m.name())
714 /// Generate a short description of this work item suitable for use as a thread name.
715 fn short_description(&self) -> String {
716 // `pthread_setname()` on *nix is limited to 15 characters and longer names are ignored.
717 // Use very short descriptions in this case to maximize the space available for the module name.
718 // Windows does not have that limitation so use slightly more descriptive names there.
720 WorkItem::Optimize(m) => {
722 return format!("optimize module {}", m.name);
724 return format!("opt {}", m.name);
726 WorkItem::CopyPostLtoArtifacts(m) => {
728 return format!("copy LTO artifacts for {}", m.name);
730 return format!("copy {}", m.name);
732 WorkItem::LTO(m) => {
734 return format!("LTO module {}", m.name());
736 return format!("LTO {}", m.name());
742 enum WorkItemResult<B: WriteBackendMethods> {
743 Compiled(CompiledModule),
744 NeedsLink(ModuleCodegen<B::Module>),
745 NeedsFatLTO(FatLTOInput<B>),
746 NeedsThinLTO(String, B::ThinBuffer),
749 pub enum FatLTOInput<B: WriteBackendMethods> {
750 Serialized { name: String, buffer: B::ModuleBuffer },
751 InMemory(ModuleCodegen<B::Module>),
754 fn execute_work_item<B: ExtraBackendMethods>(
755 cgcx: &CodegenContext<B>,
756 work_item: WorkItem<B>,
757 ) -> Result<WorkItemResult<B>, FatalError> {
758 let module_config = cgcx.config(work_item.module_kind());
761 WorkItem::Optimize(module) => execute_optimize_work_item(cgcx, module, module_config),
762 WorkItem::CopyPostLtoArtifacts(module) => {
763 Ok(execute_copy_from_cache_work_item(cgcx, module, module_config))
765 WorkItem::LTO(module) => execute_lto_work_item(cgcx, module, module_config),
769 // Actual LTO type we end up choosing based on multiple factors.
770 pub enum ComputedLtoType {
776 pub fn compute_per_cgu_lto_type(
778 opts: &config::Options,
779 sess_crate_types: &[CrateType],
780 module_kind: ModuleKind,
781 ) -> ComputedLtoType {
782 // Metadata modules never participate in LTO regardless of the lto
784 if module_kind == ModuleKind::Metadata {
785 return ComputedLtoType::No;
788 // If the linker does LTO, we don't have to do it. Note that we
789 // keep doing full LTO, if it is requested, as not to break the
790 // assumption that the output will be a single module.
791 let linker_does_lto = opts.cg.linker_plugin_lto.enabled();
793 // When we're automatically doing ThinLTO for multi-codegen-unit
794 // builds we don't actually want to LTO the allocator modules if
795 // it shows up. This is due to various linker shenanigans that
796 // we'll encounter later.
797 let is_allocator = module_kind == ModuleKind::Allocator;
799 // We ignore a request for full crate grath LTO if the cate type
800 // is only an rlib, as there is no full crate graph to process,
801 // that'll happen later.
803 // This use case currently comes up primarily for targets that
804 // require LTO so the request for LTO is always unconditionally
805 // passed down to the backend, but we don't actually want to do
806 // anything about it yet until we've got a final product.
807 let is_rlib = sess_crate_types.len() == 1 && sess_crate_types[0] == CrateType::Rlib;
810 Lto::ThinLocal if !linker_does_lto && !is_allocator => ComputedLtoType::Thin,
811 Lto::Thin if !linker_does_lto && !is_rlib => ComputedLtoType::Thin,
812 Lto::Fat if !is_rlib => ComputedLtoType::Fat,
813 _ => ComputedLtoType::No,
817 fn execute_optimize_work_item<B: ExtraBackendMethods>(
818 cgcx: &CodegenContext<B>,
819 module: ModuleCodegen<B::Module>,
820 module_config: &ModuleConfig,
821 ) -> Result<WorkItemResult<B>, FatalError> {
822 let diag_handler = cgcx.create_diag_handler();
825 B::optimize(cgcx, &diag_handler, &module, module_config)?;
828 // After we've done the initial round of optimizations we need to
829 // decide whether to synchronously codegen this module or ship it
830 // back to the coordinator thread for further LTO processing (which
831 // has to wait for all the initial modules to be optimized).
833 let lto_type = compute_per_cgu_lto_type(&cgcx.lto, &cgcx.opts, &cgcx.crate_types, module.kind);
835 // If we're doing some form of incremental LTO then we need to be sure to
836 // save our module to disk first.
837 let bitcode = if cgcx.config(module.kind).emit_pre_lto_bc {
838 let filename = pre_lto_bitcode_filename(&module.name);
839 cgcx.incr_comp_session_dir.as_ref().map(|path| path.join(&filename))
845 ComputedLtoType::No => finish_intra_module_work(cgcx, module, module_config),
846 ComputedLtoType::Thin => {
847 let (name, thin_buffer) = B::prepare_thin(module);
848 if let Some(path) = bitcode {
849 fs::write(&path, thin_buffer.data()).unwrap_or_else(|e| {
850 panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
853 Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))
855 ComputedLtoType::Fat => match bitcode {
857 let (name, buffer) = B::serialize_module(module);
858 fs::write(&path, buffer.data()).unwrap_or_else(|e| {
859 panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e);
861 Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::Serialized { name, buffer }))
863 None => Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::InMemory(module))),
868 fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>(
869 cgcx: &CodegenContext<B>,
870 module: CachedModuleCodegen,
871 module_config: &ModuleConfig,
872 ) -> WorkItemResult<B> {
873 let incr_comp_session_dir = cgcx.incr_comp_session_dir.as_ref().unwrap();
874 let mut object = None;
875 if let Some(saved_file) = module.source.saved_file {
876 let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, Some(&module.name));
877 object = Some(obj_out.clone());
878 let source_file = in_incr_comp_dir(&incr_comp_session_dir, &saved_file);
880 "copying pre-existing module `{}` from {:?} to {}",
885 if let Err(err) = link_or_copy(&source_file, &obj_out) {
886 let diag_handler = cgcx.create_diag_handler();
887 diag_handler.err(&format!(
888 "unable to copy {} to {}: {}",
889 source_file.display(),
896 assert_eq!(object.is_some(), module_config.emit_obj != EmitObj::None);
898 WorkItemResult::Compiled(CompiledModule {
900 kind: ModuleKind::Regular,
907 fn execute_lto_work_item<B: ExtraBackendMethods>(
908 cgcx: &CodegenContext<B>,
909 mut module: lto::LtoModuleCodegen<B>,
910 module_config: &ModuleConfig,
911 ) -> Result<WorkItemResult<B>, FatalError> {
912 let module = unsafe { module.optimize(cgcx)? };
913 finish_intra_module_work(cgcx, module, module_config)
916 fn finish_intra_module_work<B: ExtraBackendMethods>(
917 cgcx: &CodegenContext<B>,
918 module: ModuleCodegen<B::Module>,
919 module_config: &ModuleConfig,
920 ) -> Result<WorkItemResult<B>, FatalError> {
921 let diag_handler = cgcx.create_diag_handler();
923 if !cgcx.opts.debugging_opts.combine_cgu
924 || module.kind == ModuleKind::Metadata
925 || module.kind == ModuleKind::Allocator
927 let module = unsafe { B::codegen(cgcx, &diag_handler, module, module_config)? };
928 Ok(WorkItemResult::Compiled(module))
930 Ok(WorkItemResult::NeedsLink(module))
934 pub enum Message<B: WriteBackendMethods> {
935 Token(io::Result<Acquired>),
937 result: FatLTOInput<B>,
942 thin_buffer: B::ThinBuffer,
946 module: ModuleCodegen<B::Module>,
950 result: Result<CompiledModule, Option<WorkerFatalError>>,
954 llvm_work_item: WorkItem<B>,
957 AddImportOnlyModule {
958 module_data: SerializedModule<B::ModuleBuffer>,
959 work_product: WorkProduct,
968 code: Option<DiagnosticId>,
972 #[derive(PartialEq, Clone, Copy, Debug)]
973 enum MainThreadWorkerState {
979 fn start_executing_work<B: ExtraBackendMethods>(
982 crate_info: &CrateInfo,
983 shared_emitter: SharedEmitter,
984 codegen_worker_send: Sender<Message<B>>,
985 coordinator_receive: Receiver<Box<dyn Any + Send>>,
988 regular_config: Arc<ModuleConfig>,
989 metadata_config: Arc<ModuleConfig>,
990 allocator_config: Arc<ModuleConfig>,
991 tx_to_llvm_workers: Sender<Box<dyn Any + Send>>,
992 ) -> thread::JoinHandle<Result<CompiledModules, ()>> {
993 let coordinator_send = tx_to_llvm_workers;
996 // Compute the set of symbols we need to retain when doing LTO (if we need to)
997 let exported_symbols = {
998 let mut exported_symbols = FxHashMap::default();
1000 let copy_symbols = |cnum| {
1002 .exported_symbols(cnum)
1004 .map(|&(s, lvl)| (symbol_name_for_instance_in_crate(tcx, s, cnum), lvl))
1012 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
1013 Some(Arc::new(exported_symbols))
1015 Lto::Fat | Lto::Thin => {
1016 exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE));
1017 for &cnum in tcx.crates().iter() {
1018 exported_symbols.insert(cnum, copy_symbols(cnum));
1020 Some(Arc::new(exported_symbols))
1025 // First up, convert our jobserver into a helper thread so we can use normal
1026 // mpsc channels to manage our messages and such.
1027 // After we've requested tokens then we'll, when we can,
1028 // get tokens on `coordinator_receive` which will
1029 // get managed in the main loop below.
1030 let coordinator_send2 = coordinator_send.clone();
1031 let helper = jobserver
1032 .into_helper_thread(move |token| {
1033 drop(coordinator_send2.send(Box::new(Message::Token::<B>(token))));
1035 .expect("failed to spawn helper thread");
1037 let mut each_linked_rlib_for_lto = Vec::new();
1038 drop(link::each_linked_rlib(crate_info, &mut |cnum, path| {
1039 if link::ignored_for_lto(sess, crate_info, cnum) {
1042 each_linked_rlib_for_lto.push((cnum, path.to_path_buf()));
1045 let ol = if tcx.sess.opts.debugging_opts.no_codegen
1046 || !tcx.sess.opts.output_types.should_codegen()
1048 // If we know that we won’t be doing codegen, create target machines without optimisation.
1049 config::OptLevel::No
1051 tcx.backend_optimization_level(LOCAL_CRATE)
1053 let cgcx = CodegenContext::<B> {
1054 backend: backend.clone(),
1055 crate_types: sess.crate_types().to_vec(),
1056 each_linked_rlib_for_lto,
1058 no_landing_pads: sess.panic_strategy() == PanicStrategy::Abort,
1059 fewer_names: sess.fewer_names(),
1060 save_temps: sess.opts.cg.save_temps,
1061 opts: Arc::new(sess.opts.clone()),
1062 prof: sess.prof.clone(),
1064 remark: sess.opts.cg.remark.clone(),
1066 incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
1067 cgu_reuse_tracker: sess.cgu_reuse_tracker.clone(),
1069 diag_emitter: shared_emitter.clone(),
1070 output_filenames: tcx.output_filenames(LOCAL_CRATE),
1071 regular_module_config: regular_config,
1072 metadata_module_config: metadata_config,
1073 allocator_module_config: allocator_config,
1074 tm_factory: backend.target_machine_factory(tcx.sess, ol),
1076 msvc_imps_needed: msvc_imps_needed(tcx),
1077 is_pe_coff: tcx.sess.target.is_like_windows,
1078 target_can_use_split_dwarf: tcx.sess.target_can_use_split_dwarf(),
1079 target_pointer_width: tcx.sess.target.pointer_width,
1080 target_arch: tcx.sess.target.arch.clone(),
1081 debuginfo: tcx.sess.opts.debuginfo,
1082 split_debuginfo: tcx.sess.split_debuginfo(),
1085 // This is the "main loop" of parallel work happening for parallel codegen.
1086 // It's here that we manage parallelism, schedule work, and work with
1087 // messages coming from clients.
1089 // There are a few environmental pre-conditions that shape how the system
1092 // - Error reporting only can happen on the main thread because that's the
1093 // only place where we have access to the compiler `Session`.
1094 // - LLVM work can be done on any thread.
1095 // - Codegen can only happen on the main thread.
1096 // - Each thread doing substantial work most be in possession of a `Token`
1097 // from the `Jobserver`.
1098 // - The compiler process always holds one `Token`. Any additional `Tokens`
1099 // have to be requested from the `Jobserver`.
1103 // The error reporting restriction is handled separately from the rest: We
1104 // set up a `SharedEmitter` the holds an open channel to the main thread.
1105 // When an error occurs on any thread, the shared emitter will send the
1106 // error message to the receiver main thread (`SharedEmitterMain`). The
1107 // main thread will periodically query this error message queue and emit
1108 // any error messages it has received. It might even abort compilation if
1109 // has received a fatal error. In this case we rely on all other threads
1110 // being torn down automatically with the main thread.
1111 // Since the main thread will often be busy doing codegen work, error
1112 // reporting will be somewhat delayed, since the message queue can only be
1113 // checked in between to work packages.
1115 // Work Processing Infrastructure
1116 // ==============================
1117 // The work processing infrastructure knows three major actors:
1119 // - the coordinator thread,
1120 // - the main thread, and
1121 // - LLVM worker threads
1123 // The coordinator thread is running a message loop. It instructs the main
1124 // thread about what work to do when, and it will spawn off LLVM worker
1125 // threads as open LLVM WorkItems become available.
1127 // The job of the main thread is to codegen CGUs into LLVM work package
1128 // (since the main thread is the only thread that can do this). The main
1129 // thread will block until it receives a message from the coordinator, upon
1130 // which it will codegen one CGU, send it to the coordinator and block
1131 // again. This way the coordinator can control what the main thread is
1134 // The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is
1135 // available, it will spawn off a new LLVM worker thread and let it process
1136 // that a WorkItem. When a LLVM worker thread is done with its WorkItem,
1137 // it will just shut down, which also frees all resources associated with
1138 // the given LLVM module, and sends a message to the coordinator that the
1139 // has been completed.
1143 // The scheduler's goal is to minimize the time it takes to complete all
1144 // work there is, however, we also want to keep memory consumption low
1145 // if possible. These two goals are at odds with each other: If memory
1146 // consumption were not an issue, we could just let the main thread produce
1147 // LLVM WorkItems at full speed, assuring maximal utilization of
1148 // Tokens/LLVM worker threads. However, since codegen usual is faster
1149 // than LLVM processing, the queue of LLVM WorkItems would fill up and each
1150 // WorkItem potentially holds on to a substantial amount of memory.
1152 // So the actual goal is to always produce just enough LLVM WorkItems as
1153 // not to starve our LLVM worker threads. That means, once we have enough
1154 // WorkItems in our queue, we can block the main thread, so it does not
1155 // produce more until we need them.
1157 // Doing LLVM Work on the Main Thread
1158 // ----------------------------------
1159 // Since the main thread owns the compiler processes implicit `Token`, it is
1160 // wasteful to keep it blocked without doing any work. Therefore, what we do
1161 // in this case is: We spawn off an additional LLVM worker thread that helps
1162 // reduce the queue. The work it is doing corresponds to the implicit
1163 // `Token`. The coordinator will mark the main thread as being busy with
1164 // LLVM work. (The actual work happens on another OS thread but we just care
1165 // about `Tokens`, not actual threads).
1167 // When any LLVM worker thread finishes while the main thread is marked as
1168 // "busy with LLVM work", we can do a little switcheroo: We give the Token
1169 // of the just finished thread to the LLVM worker thread that is working on
1170 // behalf of the main thread's implicit Token, thus freeing up the main
1171 // thread again. The coordinator can then again decide what the main thread
1172 // should do. This allows the coordinator to make decisions at more points
1175 // Striking a Balance between Throughput and Memory Consumption
1176 // ------------------------------------------------------------
1177 // Since our two goals, (1) use as many Tokens as possible and (2) keep
1178 // memory consumption as low as possible, are in conflict with each other,
1179 // we have to find a trade off between them. Right now, the goal is to keep
1180 // all workers busy, which means that no worker should find the queue empty
1181 // when it is ready to start.
1182 // How do we do achieve this? Good question :) We actually never know how
1183 // many `Tokens` are potentially available so it's hard to say how much to
1184 // fill up the queue before switching the main thread to LLVM work. Also we
1185 // currently don't have a means to estimate how long a running LLVM worker
1186 // will still be busy with it's current WorkItem. However, we know the
1187 // maximal count of available Tokens that makes sense (=the number of CPU
1188 // cores), so we can take a conservative guess. The heuristic we use here
1189 // is implemented in the `queue_full_enough()` function.
1191 // Some Background on Jobservers
1192 // -----------------------------
1193 // It's worth also touching on the management of parallelism here. We don't
1194 // want to just spawn a thread per work item because while that's optimal
1195 // parallelism it may overload a system with too many threads or violate our
1196 // configuration for the maximum amount of cpu to use for this process. To
1197 // manage this we use the `jobserver` crate.
1199 // Job servers are an artifact of GNU make and are used to manage
1200 // parallelism between processes. A jobserver is a glorified IPC semaphore
1201 // basically. Whenever we want to run some work we acquire the semaphore,
1202 // and whenever we're done with that work we release the semaphore. In this
1203 // manner we can ensure that the maximum number of parallel workers is
1204 // capped at any one point in time.
1206 // LTO and the coordinator thread
1207 // ------------------------------
1209 // The final job the coordinator thread is responsible for is managing LTO
1210 // and how that works. When LTO is requested what we'll to is collect all
1211 // optimized LLVM modules into a local vector on the coordinator. Once all
1212 // modules have been codegened and optimized we hand this to the `lto`
1213 // module for further optimization. The `lto` module will return back a list
1214 // of more modules to work on, which the coordinator will continue to spawn
1217 // Each LLVM module is automatically sent back to the coordinator for LTO if
1218 // necessary. There's already optimizations in place to avoid sending work
1219 // back to the coordinator if LTO isn't requested.
1220 return thread::spawn(move || {
1221 let mut worker_id_counter = 0;
1222 let mut free_worker_ids = Vec::new();
1223 let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| {
1224 if let Some(id) = free_worker_ids.pop() {
1227 let id = worker_id_counter;
1228 worker_id_counter += 1;
1233 // This is where we collect codegen units that have gone all the way
1234 // through codegen and LLVM.
1235 let mut compiled_modules = vec![];
1236 let mut compiled_metadata_module = None;
1237 let mut compiled_allocator_module = None;
1238 let mut needs_link = Vec::new();
1239 let mut needs_fat_lto = Vec::new();
1240 let mut needs_thin_lto = Vec::new();
1241 let mut lto_import_only_modules = Vec::new();
1242 let mut started_lto = false;
1243 let mut codegen_aborted = false;
1245 // This flag tracks whether all items have gone through codegens
1246 let mut codegen_done = false;
1248 // This is the queue of LLVM work items that still need processing.
1249 let mut work_items = Vec::<(WorkItem<B>, u64)>::new();
1251 // This are the Jobserver Tokens we currently hold. Does not include
1252 // the implicit Token the compiler process owns no matter what.
1253 let mut tokens = Vec::new();
1255 let mut main_thread_worker_state = MainThreadWorkerState::Idle;
1256 let mut running = 0;
1258 let prof = &cgcx.prof;
1259 let mut llvm_start_time: Option<VerboseTimingGuard<'_>> = None;
1261 // Run the message loop while there's still anything that needs message
1262 // processing. Note that as soon as codegen is aborted we simply want to
1263 // wait for all existing work to finish, so many of the conditions here
1264 // only apply if codegen hasn't been aborted as they represent pending
1268 || (!codegen_aborted
1269 && !(work_items.is_empty()
1270 && needs_fat_lto.is_empty()
1271 && needs_thin_lto.is_empty()
1272 && lto_import_only_modules.is_empty()
1273 && main_thread_worker_state == MainThreadWorkerState::Idle))
1275 // While there are still CGUs to be codegened, the coordinator has
1276 // to decide how to utilize the compiler processes implicit Token:
1277 // For codegenning more CGU or for running them through LLVM.
1279 if main_thread_worker_state == MainThreadWorkerState::Idle {
1280 // Compute the number of workers that will be running once we've taken as many
1281 // items from the work queue as we can, plus one for the main thread. It's not
1282 // critically important that we use this instead of just `running`, but it
1283 // prevents the `queue_full_enough` heuristic from fluctuating just because a
1284 // worker finished up and we decreased the `running` count, even though we're
1285 // just going to increase it right after this when we put a new worker to work.
1286 let extra_tokens = tokens.len().checked_sub(running).unwrap();
1287 let additional_running = std::cmp::min(extra_tokens, work_items.len());
1288 let anticipated_running = running + additional_running + 1;
1290 if !queue_full_enough(work_items.len(), anticipated_running) {
1291 // The queue is not full enough, codegen more items:
1292 if codegen_worker_send.send(Message::CodegenItem).is_err() {
1293 panic!("Could not send Message::CodegenItem to main thread")
1295 main_thread_worker_state = MainThreadWorkerState::Codegenning;
1297 // The queue is full enough to not let the worker
1298 // threads starve. Use the implicit Token to do some
1301 work_items.pop().expect("queue empty - queue_full_enough() broken?");
1302 let cgcx = CodegenContext {
1303 worker: get_worker_id(&mut free_worker_ids),
1306 maybe_start_llvm_timer(
1308 cgcx.config(item.module_kind()),
1309 &mut llvm_start_time,
1311 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1312 spawn_work(cgcx, item);
1315 } else if codegen_aborted {
1316 // don't queue up any more work if codegen was aborted, we're
1317 // just waiting for our existing children to finish
1319 // If we've finished everything related to normal codegen
1320 // then it must be the case that we've got some LTO work to do.
1321 // Perform the serial work here of figuring out what we're
1322 // going to LTO and then push a bunch of work items onto our
1324 if work_items.is_empty()
1326 && main_thread_worker_state == MainThreadWorkerState::Idle
1328 assert!(!started_lto);
1331 let needs_fat_lto = mem::take(&mut needs_fat_lto);
1332 let needs_thin_lto = mem::take(&mut needs_thin_lto);
1333 let import_only_modules = mem::take(&mut lto_import_only_modules);
1336 generate_lto_work(&cgcx, needs_fat_lto, needs_thin_lto, import_only_modules)
1338 let insertion_index = work_items
1339 .binary_search_by_key(&cost, |&(_, cost)| cost)
1340 .unwrap_or_else(|e| e);
1341 work_items.insert(insertion_index, (work, cost));
1342 if !cgcx.opts.debugging_opts.no_parallel_llvm {
1343 helper.request_token();
1348 // In this branch, we know that everything has been codegened,
1349 // so it's just a matter of determining whether the implicit
1350 // Token is free to use for LLVM work.
1351 match main_thread_worker_state {
1352 MainThreadWorkerState::Idle => {
1353 if let Some((item, _)) = work_items.pop() {
1354 let cgcx = CodegenContext {
1355 worker: get_worker_id(&mut free_worker_ids),
1358 maybe_start_llvm_timer(
1360 cgcx.config(item.module_kind()),
1361 &mut llvm_start_time,
1363 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1364 spawn_work(cgcx, item);
1366 // There is no unstarted work, so let the main thread
1367 // take over for a running worker. Otherwise the
1368 // implicit token would just go to waste.
1369 // We reduce the `running` counter by one. The
1370 // `tokens.truncate()` below will take care of
1371 // giving the Token back.
1372 debug_assert!(running > 0);
1374 main_thread_worker_state = MainThreadWorkerState::LLVMing;
1377 MainThreadWorkerState::Codegenning => bug!(
1378 "codegen worker should not be codegenning after \
1379 codegen was already completed"
1381 MainThreadWorkerState::LLVMing => {
1382 // Already making good use of that token
1387 // Spin up what work we can, only doing this while we've got available
1388 // parallelism slots and work left to spawn.
1389 while !codegen_aborted && !work_items.is_empty() && running < tokens.len() {
1390 let (item, _) = work_items.pop().unwrap();
1392 maybe_start_llvm_timer(prof, cgcx.config(item.module_kind()), &mut llvm_start_time);
1395 CodegenContext { worker: get_worker_id(&mut free_worker_ids), ..cgcx.clone() };
1397 spawn_work(cgcx, item);
1401 // Relinquish accidentally acquired extra tokens
1402 tokens.truncate(running);
1404 // If a thread exits successfully then we drop a token associated
1405 // with that worker and update our `running` count. We may later
1406 // re-acquire a token to continue running more work. We may also not
1407 // actually drop a token here if the worker was running with an
1408 // "ephemeral token"
1409 let mut free_worker = |worker_id| {
1410 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
1411 main_thread_worker_state = MainThreadWorkerState::Idle;
1416 free_worker_ids.push(worker_id);
1419 let msg = coordinator_receive.recv().unwrap();
1420 match *msg.downcast::<Message<B>>().ok().unwrap() {
1421 // Save the token locally and the next turn of the loop will use
1422 // this to spawn a new unit of work, or it may get dropped
1423 // immediately if we have no more work to spawn.
1424 Message::Token(token) => {
1429 if main_thread_worker_state == MainThreadWorkerState::LLVMing {
1430 // If the main thread token is used for LLVM work
1431 // at the moment, we turn that thread into a regular
1432 // LLVM worker thread, so the main thread is free
1433 // to react to codegen demand.
1434 main_thread_worker_state = MainThreadWorkerState::Idle;
1439 let msg = &format!("failed to acquire jobserver token: {}", e);
1440 shared_emitter.fatal(msg);
1441 // Exit the coordinator thread
1447 Message::CodegenDone { llvm_work_item, cost } => {
1448 // We keep the queue sorted by estimated processing cost,
1449 // so that more expensive items are processed earlier. This
1450 // is good for throughput as it gives the main thread more
1451 // time to fill up the queue and it avoids scheduling
1452 // expensive items to the end.
1453 // Note, however, that this is not ideal for memory
1454 // consumption, as LLVM module sizes are not evenly
1456 let insertion_index = work_items.binary_search_by_key(&cost, |&(_, cost)| cost);
1457 let insertion_index = match insertion_index {
1458 Ok(idx) | Err(idx) => idx,
1460 work_items.insert(insertion_index, (llvm_work_item, cost));
1462 if !cgcx.opts.debugging_opts.no_parallel_llvm {
1463 helper.request_token();
1465 assert!(!codegen_aborted);
1466 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1467 main_thread_worker_state = MainThreadWorkerState::Idle;
1470 Message::CodegenComplete => {
1471 codegen_done = true;
1472 assert!(!codegen_aborted);
1473 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1474 main_thread_worker_state = MainThreadWorkerState::Idle;
1477 // If codegen is aborted that means translation was aborted due
1478 // to some normal-ish compiler error. In this situation we want
1479 // to exit as soon as possible, but we want to make sure all
1480 // existing work has finished. Flag codegen as being done, and
1481 // then conditions above will ensure no more work is spawned but
1482 // we'll keep executing this loop until `running` hits 0.
1483 Message::CodegenAborted => {
1484 assert!(!codegen_aborted);
1485 codegen_done = true;
1486 codegen_aborted = true;
1487 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1489 Message::Done { result: Ok(compiled_module), worker_id } => {
1490 free_worker(worker_id);
1491 match compiled_module.kind {
1492 ModuleKind::Regular => {
1493 compiled_modules.push(compiled_module);
1495 ModuleKind::Metadata => {
1496 assert!(compiled_metadata_module.is_none());
1497 compiled_metadata_module = Some(compiled_module);
1499 ModuleKind::Allocator => {
1500 assert!(compiled_allocator_module.is_none());
1501 compiled_allocator_module = Some(compiled_module);
1505 Message::NeedsLink { module, worker_id } => {
1506 free_worker(worker_id);
1507 needs_link.push(module);
1509 Message::NeedsFatLTO { result, worker_id } => {
1510 assert!(!started_lto);
1511 free_worker(worker_id);
1512 needs_fat_lto.push(result);
1514 Message::NeedsThinLTO { name, thin_buffer, worker_id } => {
1515 assert!(!started_lto);
1516 free_worker(worker_id);
1517 needs_thin_lto.push((name, thin_buffer));
1519 Message::AddImportOnlyModule { module_data, work_product } => {
1520 assert!(!started_lto);
1521 assert!(!codegen_done);
1522 assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning);
1523 lto_import_only_modules.push((module_data, work_product));
1524 main_thread_worker_state = MainThreadWorkerState::Idle;
1526 // If the thread failed that means it panicked, so we abort immediately.
1527 Message::Done { result: Err(None), worker_id: _ } => {
1528 bug!("worker thread panicked");
1530 Message::Done { result: Err(Some(WorkerFatalError)), worker_id: _ } => {
1533 Message::CodegenItem => bug!("the coordinator should not receive codegen requests"),
1537 let needs_link = mem::take(&mut needs_link);
1538 if !needs_link.is_empty() {
1539 assert!(compiled_modules.is_empty());
1540 let diag_handler = cgcx.create_diag_handler();
1541 let module = B::run_link(&cgcx, &diag_handler, needs_link).map_err(|_| ())?;
1542 let module = unsafe {
1543 B::codegen(&cgcx, &diag_handler, module, cgcx.config(ModuleKind::Regular))
1546 compiled_modules.push(module);
1549 // Drop to print timings
1550 drop(llvm_start_time);
1552 // Regardless of what order these modules completed in, report them to
1553 // the backend in the same order every time to ensure that we're handing
1554 // out deterministic results.
1555 compiled_modules.sort_by(|a, b| a.name.cmp(&b.name));
1557 Ok(CompiledModules {
1558 modules: compiled_modules,
1559 metadata_module: compiled_metadata_module,
1560 allocator_module: compiled_allocator_module,
1564 // A heuristic that determines if we have enough LLVM WorkItems in the
1565 // queue so that the main thread can do LLVM work instead of codegen
1566 fn queue_full_enough(items_in_queue: usize, workers_running: usize) -> bool {
1567 // This heuristic scales ahead-of-time codegen according to available
1568 // concurrency, as measured by `workers_running`. The idea is that the
1569 // more concurrency we have available, the more demand there will be for
1570 // work items, and the fuller the queue should be kept to meet demand.
1571 // An important property of this approach is that we codegen ahead of
1572 // time only as much as necessary, so as to keep fewer LLVM modules in
1573 // memory at once, thereby reducing memory consumption.
1575 // When the number of workers running is less than the max concurrency
1576 // available to us, this heuristic can cause us to instruct the main
1577 // thread to work on an LLVM item (that is, tell it to "LLVM") instead
1578 // of codegen, even though it seems like it *should* be codegenning so
1579 // that we can create more work items and spawn more LLVM workers.
1581 // But this is not a problem. When the main thread is told to LLVM,
1582 // according to this heuristic and how work is scheduled, there is
1583 // always at least one item in the queue, and therefore at least one
1584 // pending jobserver token request. If there *is* more concurrency
1585 // available, we will immediately receive a token, which will upgrade
1586 // the main thread's LLVM worker to a real one (conceptually), and free
1587 // up the main thread to codegen if necessary. On the other hand, if
1588 // there isn't more concurrency, then the main thread working on an LLVM
1589 // item is appropriate, as long as the queue is full enough for demand.
1591 // Speaking of which, how full should we keep the queue? Probably less
1592 // full than you'd think. A lot has to go wrong for the queue not to be
1593 // full enough and for that to have a negative effect on compile times.
1595 // Workers are unlikely to finish at exactly the same time, so when one
1596 // finishes and takes another work item off the queue, we often have
1597 // ample time to codegen at that point before the next worker finishes.
1598 // But suppose that codegen takes so long that the workers exhaust the
1599 // queue, and we have one or more workers that have nothing to work on.
1600 // Well, it might not be so bad. Of all the LLVM modules we create and
1601 // optimize, one has to finish last. It's not necessarily the case that
1602 // by losing some concurrency for a moment, we delay the point at which
1603 // that last LLVM module is finished and the rest of compilation can
1604 // proceed. Also, when we can't take advantage of some concurrency, we
1605 // give tokens back to the job server. That enables some other rustc to
1606 // potentially make use of the available concurrency. That could even
1607 // *decrease* overall compile time if we're lucky. But yes, if no other
1608 // rustc can make use of the concurrency, then we've squandered it.
1610 // However, keeping the queue full is also beneficial when we have a
1611 // surge in available concurrency. Then items can be taken from the
1612 // queue immediately, without having to wait for codegen.
1614 // So, the heuristic below tries to keep one item in the queue for every
1615 // four running workers. Based on limited benchmarking, this appears to
1616 // be more than sufficient to avoid increasing compilation times.
1617 let quarter_of_workers = workers_running - 3 * workers_running / 4;
1618 items_in_queue > 0 && items_in_queue >= quarter_of_workers
1621 fn maybe_start_llvm_timer<'a>(
1622 prof: &'a SelfProfilerRef,
1623 config: &ModuleConfig,
1624 llvm_start_time: &mut Option<VerboseTimingGuard<'a>>,
1626 if config.time_module && llvm_start_time.is_none() {
1627 *llvm_start_time = Some(prof.extra_verbose_generic_activity("LLVM_passes", "crate"));
1632 /// `FatalError` is explicitly not `Send`.
1634 pub struct WorkerFatalError;
1636 fn spawn_work<B: ExtraBackendMethods>(cgcx: CodegenContext<B>, work: WorkItem<B>) {
1637 let builder = thread::Builder::new().name(work.short_description());
1640 // Set up a destructor which will fire off a message that we're done as
1642 struct Bomb<B: ExtraBackendMethods> {
1643 coordinator_send: Sender<Box<dyn Any + Send>>,
1644 result: Option<Result<WorkItemResult<B>, FatalError>>,
1647 impl<B: ExtraBackendMethods> Drop for Bomb<B> {
1648 fn drop(&mut self) {
1649 let worker_id = self.worker_id;
1650 let msg = match self.result.take() {
1651 Some(Ok(WorkItemResult::Compiled(m))) => {
1652 Message::Done::<B> { result: Ok(m), worker_id }
1654 Some(Ok(WorkItemResult::NeedsLink(m))) => {
1655 Message::NeedsLink::<B> { module: m, worker_id }
1657 Some(Ok(WorkItemResult::NeedsFatLTO(m))) => {
1658 Message::NeedsFatLTO::<B> { result: m, worker_id }
1660 Some(Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))) => {
1661 Message::NeedsThinLTO::<B> { name, thin_buffer, worker_id }
1663 Some(Err(FatalError)) => {
1664 Message::Done::<B> { result: Err(Some(WorkerFatalError)), worker_id }
1666 None => Message::Done::<B> { result: Err(None), worker_id },
1668 drop(self.coordinator_send.send(Box::new(msg)));
1672 let mut bomb = Bomb::<B> {
1673 coordinator_send: cgcx.coordinator_send.clone(),
1675 worker_id: cgcx.worker,
1678 // Execute the work itself, and if it finishes successfully then flag
1679 // ourselves as a success as well.
1681 // Note that we ignore any `FatalError` coming out of `execute_work_item`,
1682 // as a diagnostic was already sent off to the main thread - just
1683 // surface that there was an error in this worker.
1685 let _prof_timer = work.start_profiling(&cgcx);
1686 Some(execute_work_item(&cgcx, work))
1689 .expect("failed to spawn thread");
1692 enum SharedEmitterMessage {
1693 Diagnostic(Diagnostic),
1694 InlineAsmError(u32, String, Level, Option<(String, Vec<InnerSpan>)>),
1700 pub struct SharedEmitter {
1701 sender: Sender<SharedEmitterMessage>,
1704 pub struct SharedEmitterMain {
1705 receiver: Receiver<SharedEmitterMessage>,
1708 impl SharedEmitter {
1709 pub fn new() -> (SharedEmitter, SharedEmitterMain) {
1710 let (sender, receiver) = channel();
1712 (SharedEmitter { sender }, SharedEmitterMain { receiver })
1715 pub fn inline_asm_error(
1720 source: Option<(String, Vec<InnerSpan>)>,
1722 drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source)));
1725 pub fn fatal(&self, msg: &str) {
1726 drop(self.sender.send(SharedEmitterMessage::Fatal(msg.to_string())));
1730 impl Emitter for SharedEmitter {
1731 fn emit_diagnostic(&mut self, diag: &rustc_errors::Diagnostic) {
1732 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
1733 msg: diag.message(),
1734 code: diag.code.clone(),
1737 for child in &diag.children {
1738 drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic {
1739 msg: child.message(),
1744 drop(self.sender.send(SharedEmitterMessage::AbortIfErrors));
1746 fn source_map(&self) -> Option<&Lrc<SourceMap>> {
1751 impl SharedEmitterMain {
1752 pub fn check(&self, sess: &Session, blocking: bool) {
1754 let message = if blocking {
1755 match self.receiver.recv() {
1756 Ok(message) => Ok(message),
1760 match self.receiver.try_recv() {
1761 Ok(message) => Ok(message),
1767 Ok(SharedEmitterMessage::Diagnostic(diag)) => {
1768 let handler = sess.diagnostic();
1769 let mut d = rustc_errors::Diagnostic::new(diag.lvl, &diag.msg);
1770 if let Some(code) = diag.code {
1773 handler.emit_diagnostic(&d);
1775 Ok(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source)) => {
1776 let msg = msg.strip_prefix("error: ").unwrap_or(&msg);
1778 let mut err = match level {
1779 Level::Error => sess.struct_err(&msg),
1780 Level::Warning => sess.struct_warn(&msg),
1781 Level::Note => sess.struct_note_without_error(&msg),
1782 _ => bug!("Invalid inline asm diagnostic level"),
1785 // If the cookie is 0 then we don't have span information.
1787 let pos = BytePos::from_u32(cookie);
1788 let span = Span::with_root_ctxt(pos, pos);
1792 // Point to the generated assembly if it is available.
1793 if let Some((buffer, spans)) = source {
1796 .new_source_file(FileName::inline_asm_source_code(&buffer), buffer);
1797 let source_span = Span::with_root_ctxt(source.start_pos, source.end_pos);
1799 spans.iter().map(|sp| source_span.from_inner(*sp)).collect();
1800 err.span_note(spans, "instantiated into assembly here");
1805 Ok(SharedEmitterMessage::AbortIfErrors) => {
1806 sess.abort_if_errors();
1808 Ok(SharedEmitterMessage::Fatal(msg)) => {
1819 pub struct OngoingCodegen<B: ExtraBackendMethods> {
1821 pub crate_name: Symbol,
1822 pub metadata: EncodedMetadata,
1823 pub windows_subsystem: Option<String>,
1824 pub linker_info: LinkerInfo,
1825 pub crate_info: CrateInfo,
1826 pub coordinator_send: Sender<Box<dyn Any + Send>>,
1827 pub codegen_worker_receive: Receiver<Message<B>>,
1828 pub shared_emitter_main: SharedEmitterMain,
1829 pub future: thread::JoinHandle<Result<CompiledModules, ()>>,
1830 pub output_filenames: Arc<OutputFilenames>,
1833 impl<B: ExtraBackendMethods> OngoingCodegen<B> {
1834 pub fn join(self, sess: &Session) -> (CodegenResults, FxHashMap<WorkProductId, WorkProduct>) {
1835 let _timer = sess.timer("finish_ongoing_codegen");
1837 self.shared_emitter_main.check(sess, true);
1838 let future = self.future;
1839 let compiled_modules = sess.time("join_worker_thread", || match future.join() {
1840 Ok(Ok(compiled_modules)) => compiled_modules,
1842 sess.abort_if_errors();
1843 panic!("expected abort due to worker thread errors")
1846 bug!("panic during codegen/LLVM phase");
1850 sess.cgu_reuse_tracker.check_expected_reuse(sess.diagnostic());
1852 sess.abort_if_errors();
1855 copy_all_cgu_workproducts_to_incr_comp_cache_dir(sess, &compiled_modules);
1856 produce_final_output_artifacts(sess, &compiled_modules, &self.output_filenames);
1858 // FIXME: time_llvm_passes support - does this use a global context or
1860 if sess.codegen_units() == 1 && sess.time_llvm_passes() {
1861 self.backend.print_pass_timings()
1866 crate_name: self.crate_name,
1867 metadata: self.metadata,
1868 windows_subsystem: self.windows_subsystem,
1869 linker_info: self.linker_info,
1870 crate_info: self.crate_info,
1872 modules: compiled_modules.modules,
1873 allocator_module: compiled_modules.allocator_module,
1874 metadata_module: compiled_modules.metadata_module,
1880 pub fn submit_pre_codegened_module_to_llvm(
1883 module: ModuleCodegen<B::Module>,
1885 self.wait_for_signal_to_codegen_item();
1886 self.check_for_errors(tcx.sess);
1888 // These are generally cheap and won't throw off scheduling.
1890 submit_codegened_module_to_llvm(&self.backend, &self.coordinator_send, module, cost);
1893 pub fn codegen_finished(&self, tcx: TyCtxt<'_>) {
1894 self.wait_for_signal_to_codegen_item();
1895 self.check_for_errors(tcx.sess);
1896 drop(self.coordinator_send.send(Box::new(Message::CodegenComplete::<B>)));
1899 /// Consumes this context indicating that codegen was entirely aborted, and
1900 /// we need to exit as quickly as possible.
1902 /// This method blocks the current thread until all worker threads have
1903 /// finished, and all worker threads should have exited or be real close to
1904 /// exiting at this point.
1905 pub fn codegen_aborted(self) {
1906 // Signal to the coordinator it should spawn no more work and start
1908 drop(self.coordinator_send.send(Box::new(Message::CodegenAborted::<B>)));
1909 drop(self.future.join());
1912 pub fn check_for_errors(&self, sess: &Session) {
1913 self.shared_emitter_main.check(sess, false);
1916 pub fn wait_for_signal_to_codegen_item(&self) {
1917 match self.codegen_worker_receive.recv() {
1918 Ok(Message::CodegenItem) => {
1921 Ok(_) => panic!("unexpected message"),
1923 // One of the LLVM threads must have panicked, fall through so
1924 // error handling can be reached.
1930 pub fn submit_codegened_module_to_llvm<B: ExtraBackendMethods>(
1932 tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
1933 module: ModuleCodegen<B::Module>,
1936 let llvm_work_item = WorkItem::Optimize(module);
1937 drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost })));
1940 pub fn submit_post_lto_module_to_llvm<B: ExtraBackendMethods>(
1942 tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
1943 module: CachedModuleCodegen,
1945 let llvm_work_item = WorkItem::CopyPostLtoArtifacts(module);
1946 drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost: 0 })));
1949 pub fn submit_pre_lto_module_to_llvm<B: ExtraBackendMethods>(
1952 tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>,
1953 module: CachedModuleCodegen,
1955 let filename = pre_lto_bitcode_filename(&module.name);
1956 let bc_path = in_incr_comp_dir_sess(tcx.sess, &filename);
1957 let file = fs::File::open(&bc_path)
1958 .unwrap_or_else(|e| panic!("failed to open bitcode file `{}`: {}", bc_path.display(), e));
1961 memmap2::Mmap::map(&file).unwrap_or_else(|e| {
1962 panic!("failed to mmap bitcode file `{}`: {}", bc_path.display(), e)
1965 // Schedule the module to be loaded
1966 drop(tx_to_llvm_workers.send(Box::new(Message::AddImportOnlyModule::<B> {
1967 module_data: SerializedModule::FromUncompressedFile(mmap),
1968 work_product: module.source,
1972 pub fn pre_lto_bitcode_filename(module_name: &str) -> String {
1973 format!("{}.{}", module_name, PRE_LTO_BC_EXT)
1976 fn msvc_imps_needed(tcx: TyCtxt<'_>) -> bool {
1977 // This should never be true (because it's not supported). If it is true,
1978 // something is wrong with commandline arg validation.
1980 !(tcx.sess.opts.cg.linker_plugin_lto.enabled()
1981 && tcx.sess.target.is_like_windows
1982 && tcx.sess.opts.cg.prefer_dynamic)
1985 tcx.sess.target.is_like_windows &&
1986 tcx.sess.crate_types().iter().any(|ct| *ct == CrateType::Rlib) &&
1987 // ThinLTO can't handle this workaround in all cases, so we don't
1988 // emit the `__imp_` symbols. Instead we make them unnecessary by disallowing
1989 // dynamic linking when linker plugin LTO is enabled.
1990 !tcx.sess.opts.cg.linker_plugin_lto.enabled()