]>
Commit | Line | Data |
---|---|---|
6a06907d | 1 | use super::link::{self, ensure_removed}; |
a1dfa0c6 | 2 | use super::lto::{self, SerializedModule}; |
dfeec247 XL |
3 | use super::symbol_export::symbol_name_for_instance_in_crate; |
4 | ||
5 | use crate::{ | |
6 | CachedModuleCodegen, CodegenResults, CompiledModule, CrateInfo, ModuleCodegen, ModuleKind, | |
dfeec247 | 7 | }; |
a1dfa0c6 | 8 | |
9fa01778 | 9 | use crate::traits::*; |
dfeec247 | 10 | use jobserver::{Acquired, Client}; |
dfeec247 | 11 | use rustc_data_structures::fx::FxHashMap; |
cdc7bbd5 | 12 | use rustc_data_structures::memmap::Mmap; |
60c5eb7d | 13 | use rustc_data_structures::profiling::SelfProfilerRef; |
74b04a01 | 14 | use rustc_data_structures::profiling::TimingGuard; |
dfeec247 | 15 | use rustc_data_structures::profiling::VerboseTimingGuard; |
e74abb32 | 16 | use rustc_data_structures::sync::Lrc; |
dfeec247 XL |
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::{ | |
f9f354fc | 22 | copy_cgu_workproduct_to_incr_comp_cache_dir, in_incr_comp_dir, in_incr_comp_dir_sess, |
dfeec247 | 23 | }; |
c295e0f8 | 24 | use rustc_metadata::EncodedMetadata; |
f9f354fc | 25 | use rustc_middle::dep_graph::{WorkProduct, WorkProductId}; |
ba9703b0 XL |
26 | use rustc_middle::middle::exported_symbols::SymbolExportLevel; |
27 | use rustc_middle::ty::TyCtxt; | |
dfeec247 | 28 | use rustc_session::cgu_reuse_tracker::CguReuseTracker; |
f9f354fc | 29 | use rustc_session::config::{self, CrateType, Lto, OutputFilenames, OutputType}; |
cdc7bbd5 | 30 | use rustc_session::config::{Passes, SwitchWithOptPath}; |
ba9703b0 | 31 | use rustc_session::Session; |
dfeec247 | 32 | use rustc_span::source_map::SourceMap; |
17df50a5 | 33 | use rustc_span::symbol::sym; |
f9f354fc | 34 | use rustc_span::{BytePos, FileName, InnerSpan, Pos, Span}; |
5099ac24 | 35 | use rustc_target::spec::{MergeFunctions, SanitizerSet}; |
a1dfa0c6 XL |
36 | |
37 | use std::any::Any; | |
38 | use std::fs; | |
39 | use std::io; | |
40 | use std::mem; | |
41 | use std::path::{Path, PathBuf}; | |
42 | use std::str; | |
dfeec247 | 43 | use std::sync::mpsc::{channel, Receiver, Sender}; |
a1dfa0c6 | 44 | use std::sync::Arc; |
a1dfa0c6 XL |
45 | use std::thread; |
46 | ||
9fa01778 | 47 | const PRE_LTO_BC_EXT: &str = "pre-lto.bc"; |
a1dfa0c6 | 48 | |
ba9703b0 XL |
49 | /// What kind of object file to emit. |
50 | #[derive(Clone, Copy, PartialEq)] | |
51 | pub enum EmitObj { | |
52 | // No object file. | |
53 | None, | |
54 | ||
55 | // Just uncompressed llvm bitcode. Provides easy compatibility with | |
56 | // emscripten's ecc compiler, when used as the linker. | |
57 | Bitcode, | |
58 | ||
59 | // Object code, possibly augmented with a bitcode section. | |
60 | ObjectCode(BitcodeSection), | |
61 | } | |
62 | ||
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. | |
67 | None, | |
68 | ||
ba9703b0 XL |
69 | // A full, uncompressed bitcode section. |
70 | Full, | |
71 | } | |
72 | ||
a1dfa0c6 XL |
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>, | |
80 | ||
81 | /// Some(level) to optimize binary size, or None to not affect program size. | |
82 | pub opt_size: Option<config::OptLevel>, | |
83 | ||
dc9dc135 XL |
84 | pub pgo_gen: SwitchWithOptPath, |
85 | pub pgo_use: Option<PathBuf>, | |
c295e0f8 XL |
86 | pub pgo_sample_use: Option<PathBuf>, |
87 | pub debug_info_for_profiling: bool, | |
17df50a5 XL |
88 | pub instrument_coverage: bool, |
89 | pub instrument_gcov: bool, | |
a1dfa0c6 | 90 | |
f035d41b XL |
91 | pub sanitizer: SanitizerSet, |
92 | pub sanitizer_recover: SanitizerSet, | |
60c5eb7d XL |
93 | pub sanitizer_memory_track_origins: usize, |
94 | ||
a1dfa0c6 | 95 | // Flags indicating which outputs to produce. |
9fa01778 | 96 | pub emit_pre_lto_bc: bool, |
a1dfa0c6 XL |
97 | pub emit_no_opt_bc: bool, |
98 | pub emit_bc: bool, | |
a1dfa0c6 XL |
99 | pub emit_ir: bool, |
100 | pub emit_asm: bool, | |
ba9703b0 | 101 | pub emit_obj: EmitObj, |
f9f354fc | 102 | pub bc_cmdline: String, |
ba9703b0 | 103 | |
a1dfa0c6 XL |
104 | // Miscellaneous flags. These are mostly copied from command-line |
105 | // options. | |
106 | pub verify_llvm_ir: bool, | |
107 | pub no_prepopulate_passes: bool, | |
108 | pub no_builtins: bool, | |
dfeec247 | 109 | pub time_module: bool, |
a1dfa0c6 XL |
110 | pub vectorize_loop: bool, |
111 | pub vectorize_slp: bool, | |
112 | pub merge_functions: bool, | |
cdc7bbd5 | 113 | pub inline_threshold: Option<u32>, |
17df50a5 | 114 | pub new_llvm_pass_manager: Option<bool>, |
f9f354fc | 115 | pub emit_lifetime_markers: bool, |
a2a8927a | 116 | pub llvm_plugins: Vec<String>, |
a1dfa0c6 XL |
117 | } |
118 | ||
119 | impl ModuleConfig { | |
f9f354fc XL |
120 | fn new( |
121 | kind: ModuleKind, | |
122 | sess: &Session, | |
123 | no_builtins: bool, | |
124 | is_compiler_builtins: bool, | |
125 | ) -> ModuleConfig { | |
ba9703b0 XL |
126 | // If it's a regular module, use `$regular`, otherwise use `$other`. |
127 | // `$regular` and `$other` are evaluated lazily. | |
128 | macro_rules! if_regular { | |
129 | ($regular: expr, $other: expr) => { | |
130 | if let ModuleKind::Regular = kind { $regular } else { $other } | |
131 | }; | |
a1dfa0c6 | 132 | } |
a1dfa0c6 | 133 | |
ba9703b0 XL |
134 | let opt_level_and_size = if_regular!(Some(sess.opts.optimize), None); |
135 | ||
136 | let save_temps = sess.opts.cg.save_temps; | |
137 | ||
138 | let should_emit_obj = sess.opts.output_types.contains_key(&OutputType::Exe) | |
139 | || match kind { | |
140 | ModuleKind::Regular => sess.opts.output_types.contains_key(&OutputType::Object), | |
141 | ModuleKind::Allocator => false, | |
142 | ModuleKind::Metadata => sess.opts.output_types.contains_key(&OutputType::Metadata), | |
143 | }; | |
144 | ||
145 | let emit_obj = if !should_emit_obj { | |
146 | EmitObj::None | |
29967ef6 | 147 | } else if sess.target.obj_is_bitcode |
f9f354fc | 148 | || (sess.opts.cg.linker_plugin_lto.enabled() && !no_builtins) |
ba9703b0 | 149 | { |
f9f354fc XL |
150 | // This case is selected if the target uses objects as bitcode, or |
151 | // if linker plugin LTO is enabled. In the linker plugin LTO case | |
152 | // the assumption is that the final link-step will read the bitcode | |
153 | // and convert it to object code. This may be done by either the | |
154 | // native linker or rustc itself. | |
155 | // | |
156 | // Note, however, that the linker-plugin-lto requested here is | |
157 | // explicitly ignored for `#![no_builtins]` crates. These crates are | |
158 | // specifically ignored by rustc's LTO passes and wouldn't work if | |
159 | // loaded into the linker. These crates define symbols that LLVM | |
160 | // lowers intrinsics to, and these symbol dependencies aren't known | |
161 | // until after codegen. As a result any crate marked | |
162 | // `#![no_builtins]` is assumed to not participate in LTO and | |
163 | // instead goes on to generate object code. | |
ba9703b0 | 164 | EmitObj::Bitcode |
f9f354fc XL |
165 | } else if need_bitcode_in_object(sess) { |
166 | EmitObj::ObjectCode(BitcodeSection::Full) | |
ba9703b0 XL |
167 | } else { |
168 | EmitObj::ObjectCode(BitcodeSection::None) | |
9fa01778 | 169 | }; |
ba9703b0 XL |
170 | |
171 | ModuleConfig { | |
17df50a5 | 172 | passes: if_regular!(sess.opts.cg.passes.clone(), vec![]), |
ba9703b0 XL |
173 | |
174 | opt_level: opt_level_and_size, | |
175 | opt_size: opt_level_and_size, | |
176 | ||
177 | pgo_gen: if_regular!( | |
178 | sess.opts.cg.profile_generate.clone(), | |
179 | SwitchWithOptPath::Disabled | |
180 | ), | |
181 | pgo_use: if_regular!(sess.opts.cg.profile_use.clone(), None), | |
c295e0f8 XL |
182 | pgo_sample_use: if_regular!(sess.opts.debugging_opts.profile_sample_use.clone(), None), |
183 | debug_info_for_profiling: sess.opts.debugging_opts.debug_info_for_profiling, | |
17df50a5 XL |
184 | instrument_coverage: if_regular!(sess.instrument_coverage(), false), |
185 | instrument_gcov: if_regular!( | |
186 | // compiler_builtins overrides the codegen-units settings, | |
187 | // which is incompatible with -Zprofile which requires that | |
188 | // only a single codegen unit is used per crate. | |
189 | sess.opts.debugging_opts.profile && !is_compiler_builtins, | |
190 | false | |
191 | ), | |
ba9703b0 | 192 | |
f035d41b | 193 | sanitizer: if_regular!(sess.opts.debugging_opts.sanitizer, SanitizerSet::empty()), |
ba9703b0 | 194 | sanitizer_recover: if_regular!( |
f035d41b XL |
195 | sess.opts.debugging_opts.sanitizer_recover, |
196 | SanitizerSet::empty() | |
ba9703b0 XL |
197 | ), |
198 | sanitizer_memory_track_origins: if_regular!( | |
199 | sess.opts.debugging_opts.sanitizer_memory_track_origins, | |
200 | 0 | |
201 | ), | |
202 | ||
203 | emit_pre_lto_bc: if_regular!( | |
204 | save_temps || need_pre_lto_bitcode_for_incr_comp(sess), | |
205 | false | |
206 | ), | |
207 | emit_no_opt_bc: if_regular!(save_temps, false), | |
208 | emit_bc: if_regular!( | |
209 | save_temps || sess.opts.output_types.contains_key(&OutputType::Bitcode), | |
210 | save_temps | |
211 | ), | |
ba9703b0 XL |
212 | emit_ir: if_regular!( |
213 | sess.opts.output_types.contains_key(&OutputType::LlvmAssembly), | |
214 | false | |
215 | ), | |
216 | emit_asm: if_regular!( | |
217 | sess.opts.output_types.contains_key(&OutputType::Assembly), | |
218 | false | |
219 | ), | |
220 | emit_obj, | |
29967ef6 | 221 | bc_cmdline: sess.target.bitcode_llvm_cmdline.clone(), |
ba9703b0 XL |
222 | |
223 | verify_llvm_ir: sess.verify_llvm_ir(), | |
224 | no_prepopulate_passes: sess.opts.cg.no_prepopulate_passes, | |
29967ef6 | 225 | no_builtins: no_builtins || sess.target.no_builtins, |
ba9703b0 XL |
226 | |
227 | // Exclude metadata and allocator modules from time_passes output, | |
228 | // since they throw off the "LLVM passes" measurement. | |
229 | time_module: if_regular!(true, false), | |
230 | ||
231 | // Copy what clang does by turning on loop vectorization at O2 and | |
232 | // slp vectorization at O3. | |
233 | vectorize_loop: !sess.opts.cg.no_vectorize_loops | |
234 | && (sess.opts.optimize == config::OptLevel::Default | |
235 | || sess.opts.optimize == config::OptLevel::Aggressive), | |
236 | vectorize_slp: !sess.opts.cg.no_vectorize_slp | |
237 | && sess.opts.optimize == config::OptLevel::Aggressive, | |
238 | ||
239 | // Some targets (namely, NVPTX) interact badly with the | |
240 | // MergeFunctions pass. This is because MergeFunctions can generate | |
241 | // new function calls which may interfere with the target calling | |
242 | // convention; e.g. for the NVPTX target, PTX kernels should not | |
243 | // call other PTX kernels. MergeFunctions can also be configured to | |
244 | // generate aliases instead, but aliases are not supported by some | |
245 | // backends (again, NVPTX). Therefore, allow targets to opt out of | |
246 | // the MergeFunctions pass, but otherwise keep the pass enabled (at | |
247 | // O2 and O3) since it can be useful for reducing code size. | |
248 | merge_functions: match sess | |
249 | .opts | |
250 | .debugging_opts | |
251 | .merge_functions | |
29967ef6 | 252 | .unwrap_or(sess.target.merge_functions) |
ba9703b0 XL |
253 | { |
254 | MergeFunctions::Disabled => false, | |
255 | MergeFunctions::Trampolines | MergeFunctions::Aliases => { | |
256 | sess.opts.optimize == config::OptLevel::Default | |
257 | || sess.opts.optimize == config::OptLevel::Aggressive | |
258 | } | |
259 | }, | |
260 | ||
261 | inline_threshold: sess.opts.cg.inline_threshold, | |
262 | new_llvm_pass_manager: sess.opts.debugging_opts.new_llvm_pass_manager, | |
f9f354fc | 263 | emit_lifetime_markers: sess.emit_lifetime_markers(), |
a2a8927a | 264 | llvm_plugins: if_regular!(sess.opts.debugging_opts.llvm_plugins.clone(), vec![]), |
ba9703b0 | 265 | } |
a1dfa0c6 XL |
266 | } |
267 | ||
268 | pub fn bitcode_needed(&self) -> bool { | |
ba9703b0 | 269 | self.emit_bc |
ba9703b0 XL |
270 | || self.emit_obj == EmitObj::Bitcode |
271 | || self.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) | |
a1dfa0c6 XL |
272 | } |
273 | } | |
274 | ||
fc512014 XL |
275 | /// Configuration passed to the function returned by the `target_machine_factory`. |
276 | pub struct TargetMachineFactoryConfig { | |
277 | /// Split DWARF is enabled in LLVM by checking that `TM.MCOptions.SplitDwarfFile` isn't empty, | |
278 | /// so the path to the dwarf object has to be provided when we create the target machine. | |
279 | /// This can be ignored by backends which do not need it for their Split DWARF support. | |
280 | pub split_dwarf_file: Option<PathBuf>, | |
a1dfa0c6 XL |
281 | } |
282 | ||
5869c6ff XL |
283 | impl TargetMachineFactoryConfig { |
284 | pub fn new( | |
285 | cgcx: &CodegenContext<impl WriteBackendMethods>, | |
286 | module_name: &str, | |
287 | ) -> TargetMachineFactoryConfig { | |
288 | let split_dwarf_file = if cgcx.target_can_use_split_dwarf { | |
a2a8927a XL |
289 | cgcx.output_filenames.split_dwarf_path( |
290 | cgcx.split_debuginfo, | |
291 | cgcx.split_dwarf_kind, | |
292 | Some(module_name), | |
293 | ) | |
5869c6ff XL |
294 | } else { |
295 | None | |
296 | }; | |
297 | TargetMachineFactoryConfig { split_dwarf_file } | |
298 | } | |
299 | } | |
300 | ||
fc512014 XL |
301 | pub type TargetMachineFactoryFn<B> = Arc< |
302 | dyn Fn(TargetMachineFactoryConfig) -> Result<<B as WriteBackendMethods>::TargetMachine, String> | |
303 | + Send | |
304 | + Sync, | |
305 | >; | |
306 | ||
dfeec247 XL |
307 | pub type ExportedSymbols = FxHashMap<CrateNum, Arc<Vec<(String, SymbolExportLevel)>>>; |
308 | ||
a1dfa0c6 XL |
309 | /// Additional resources used by optimize_and_codegen (not module specific) |
310 | #[derive(Clone)] | |
311 | pub struct CodegenContext<B: WriteBackendMethods> { | |
312 | // Resources needed when running LTO | |
313 | pub backend: B, | |
e74abb32 | 314 | pub prof: SelfProfilerRef, |
a1dfa0c6 | 315 | pub lto: Lto, |
a1dfa0c6 XL |
316 | pub save_temps: bool, |
317 | pub fewer_names: bool, | |
3c0e092e | 318 | pub time_trace: bool, |
a1dfa0c6 XL |
319 | pub exported_symbols: Option<Arc<ExportedSymbols>>, |
320 | pub opts: Arc<config::Options>, | |
f9f354fc | 321 | pub crate_types: Vec<CrateType>, |
a1dfa0c6 XL |
322 | pub each_linked_rlib_for_lto: Vec<(CrateNum, PathBuf)>, |
323 | pub output_filenames: Arc<OutputFilenames>, | |
324 | pub regular_module_config: Arc<ModuleConfig>, | |
325 | pub metadata_module_config: Arc<ModuleConfig>, | |
326 | pub allocator_module_config: Arc<ModuleConfig>, | |
fc512014 | 327 | pub tm_factory: TargetMachineFactoryFn<B>, |
a1dfa0c6 | 328 | pub msvc_imps_needed: bool, |
fc512014 | 329 | pub is_pe_coff: bool, |
5869c6ff | 330 | pub target_can_use_split_dwarf: bool, |
29967ef6 | 331 | pub target_pointer_width: u32, |
48663c56 | 332 | pub target_arch: String, |
a1dfa0c6 | 333 | pub debuginfo: config::DebugInfo, |
5869c6ff | 334 | pub split_debuginfo: rustc_target::spec::SplitDebuginfo, |
a2a8927a | 335 | pub split_dwarf_kind: rustc_session::config::SplitDwarfKind, |
a1dfa0c6 XL |
336 | |
337 | // Number of cgus excluding the allocator/metadata modules | |
338 | pub total_cgus: usize, | |
339 | // Handler to use for diagnostics produced during codegen. | |
340 | pub diag_emitter: SharedEmitter, | |
a1dfa0c6 XL |
341 | // LLVM optimizations for which we want to print remarks. |
342 | pub remark: Passes, | |
343 | // Worker thread number | |
344 | pub worker: usize, | |
345 | // The incremental compilation session directory, or None if we are not | |
346 | // compiling incrementally | |
347 | pub incr_comp_session_dir: Option<PathBuf>, | |
348 | // Used to update CGU re-use information during the thinlto phase. | |
349 | pub cgu_reuse_tracker: CguReuseTracker, | |
350 | // Channel back to the main control thread to send messages to | |
351 | pub coordinator_send: Sender<Box<dyn Any + Send>>, | |
a1dfa0c6 XL |
352 | } |
353 | ||
354 | impl<B: WriteBackendMethods> CodegenContext<B> { | |
355 | pub fn create_diag_handler(&self) -> Handler { | |
532ac7d7 | 356 | Handler::with_emitter(true, None, Box::new(self.diag_emitter.clone())) |
a1dfa0c6 XL |
357 | } |
358 | ||
359 | pub fn config(&self, kind: ModuleKind) -> &ModuleConfig { | |
360 | match kind { | |
361 | ModuleKind::Regular => &self.regular_module_config, | |
362 | ModuleKind::Metadata => &self.metadata_module_config, | |
363 | ModuleKind::Allocator => &self.allocator_module_config, | |
364 | } | |
365 | } | |
366 | } | |
367 | ||
368 | fn generate_lto_work<B: ExtraBackendMethods>( | |
369 | cgcx: &CodegenContext<B>, | |
9fa01778 | 370 | needs_fat_lto: Vec<FatLTOInput<B>>, |
0731742a | 371 | needs_thin_lto: Vec<(String, B::ThinBuffer)>, |
dfeec247 | 372 | import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>, |
a1dfa0c6 | 373 | ) -> Vec<(WorkItem<B>, u64)> { |
e74abb32 | 374 | let _prof_timer = cgcx.prof.generic_activity("codegen_generate_lto_work"); |
a1dfa0c6 | 375 | |
0731742a XL |
376 | let (lto_modules, copy_jobs) = if !needs_fat_lto.is_empty() { |
377 | assert!(needs_thin_lto.is_empty()); | |
dfeec247 XL |
378 | let lto_module = |
379 | B::run_fat_lto(cgcx, needs_fat_lto, import_only_modules).unwrap_or_else(|e| e.raise()); | |
0731742a XL |
380 | (vec![lto_module], vec![]) |
381 | } else { | |
382 | assert!(needs_fat_lto.is_empty()); | |
dfeec247 | 383 | B::run_thin_lto(cgcx, needs_thin_lto, import_only_modules).unwrap_or_else(|e| e.raise()) |
0731742a XL |
384 | }; |
385 | ||
ba9703b0 | 386 | lto_modules |
dfeec247 XL |
387 | .into_iter() |
388 | .map(|module| { | |
389 | let cost = module.cost(); | |
390 | (WorkItem::LTO(module), cost) | |
391 | }) | |
392 | .chain(copy_jobs.into_iter().map(|wp| { | |
393 | ( | |
394 | WorkItem::CopyPostLtoArtifacts(CachedModuleCodegen { | |
395 | name: wp.cgu_name.clone(), | |
396 | source: wp, | |
397 | }), | |
398 | 0, | |
399 | ) | |
400 | })) | |
ba9703b0 | 401 | .collect() |
a1dfa0c6 XL |
402 | } |
403 | ||
404 | pub struct CompiledModules { | |
405 | pub modules: Vec<CompiledModule>, | |
a1dfa0c6 XL |
406 | pub allocator_module: Option<CompiledModule>, |
407 | } | |
408 | ||
f9f354fc XL |
409 | fn need_bitcode_in_object(sess: &Session) -> bool { |
410 | let requested_for_rlib = sess.opts.cg.embed_bitcode | |
411 | && sess.crate_types().contains(&CrateType::Rlib) | |
412 | && sess.opts.output_types.contains_key(&OutputType::Exe); | |
29967ef6 | 413 | let forced_by_target = sess.target.forces_embed_bitcode; |
f9f354fc | 414 | requested_for_rlib || forced_by_target |
a1dfa0c6 XL |
415 | } |
416 | ||
9fa01778 | 417 | fn need_pre_lto_bitcode_for_incr_comp(sess: &Session) -> bool { |
a1dfa0c6 | 418 | if sess.opts.incremental.is_none() { |
dfeec247 | 419 | return false; |
a1dfa0c6 XL |
420 | } |
421 | ||
422 | match sess.lto() { | |
a1dfa0c6 | 423 | Lto::No => false, |
dfeec247 | 424 | Lto::Fat | Lto::Thin | Lto::ThinLocal => true, |
a1dfa0c6 XL |
425 | } |
426 | } | |
427 | ||
428 | pub fn start_async_codegen<B: ExtraBackendMethods>( | |
429 | backend: B, | |
dc9dc135 | 430 | tcx: TyCtxt<'_>, |
17df50a5 | 431 | target_cpu: String, |
a1dfa0c6 | 432 | metadata: EncodedMetadata, |
a2a8927a | 433 | metadata_module: Option<CompiledModule>, |
dc9dc135 | 434 | total_cgus: usize, |
a1dfa0c6 | 435 | ) -> OngoingCodegen<B> { |
e74abb32 | 436 | let (coordinator_send, coordinator_receive) = channel(); |
a1dfa0c6 | 437 | let sess = tcx.sess; |
e74abb32 | 438 | |
6a06907d XL |
439 | let crate_attrs = tcx.hir().attrs(rustc_hir::CRATE_HIR_ID); |
440 | let no_builtins = tcx.sess.contains_name(crate_attrs, sym::no_builtins); | |
441 | let is_compiler_builtins = tcx.sess.contains_name(crate_attrs, sym::compiler_builtins); | |
a1dfa0c6 | 442 | |
136023e0 | 443 | let crate_info = CrateInfo::new(tcx, target_cpu); |
a1dfa0c6 | 444 | |
f9f354fc XL |
445 | let regular_config = |
446 | ModuleConfig::new(ModuleKind::Regular, sess, no_builtins, is_compiler_builtins); | |
447 | let metadata_config = | |
448 | ModuleConfig::new(ModuleKind::Metadata, sess, no_builtins, is_compiler_builtins); | |
449 | let allocator_config = | |
450 | ModuleConfig::new(ModuleKind::Allocator, sess, no_builtins, is_compiler_builtins); | |
a1dfa0c6 XL |
451 | |
452 | let (shared_emitter, shared_emitter_main) = SharedEmitter::new(); | |
453 | let (codegen_worker_send, codegen_worker_receive) = channel(); | |
454 | ||
dfeec247 XL |
455 | let coordinator_thread = start_executing_work( |
456 | backend.clone(), | |
457 | tcx, | |
458 | &crate_info, | |
459 | shared_emitter, | |
460 | codegen_worker_send, | |
461 | coordinator_receive, | |
462 | total_cgus, | |
463 | sess.jobserver.clone(), | |
ba9703b0 | 464 | Arc::new(regular_config), |
dfeec247 XL |
465 | Arc::new(metadata_config), |
466 | Arc::new(allocator_config), | |
467 | coordinator_send.clone(), | |
468 | ); | |
a1dfa0c6 XL |
469 | |
470 | OngoingCodegen { | |
471 | backend, | |
a1dfa0c6 | 472 | metadata, |
a2a8927a | 473 | metadata_module, |
a1dfa0c6 XL |
474 | crate_info, |
475 | ||
e74abb32 | 476 | coordinator_send, |
a1dfa0c6 XL |
477 | codegen_worker_receive, |
478 | shared_emitter_main, | |
479 | future: coordinator_thread, | |
5099ac24 | 480 | output_filenames: tcx.output_filenames(()).clone(), |
a1dfa0c6 XL |
481 | } |
482 | } | |
483 | ||
484 | fn copy_all_cgu_workproducts_to_incr_comp_cache_dir( | |
485 | sess: &Session, | |
486 | compiled_modules: &CompiledModules, | |
487 | ) -> FxHashMap<WorkProductId, WorkProduct> { | |
488 | let mut work_products = FxHashMap::default(); | |
489 | ||
490 | if sess.opts.incremental.is_none() { | |
491 | return work_products; | |
492 | } | |
493 | ||
f9f354fc | 494 | let _timer = sess.timer("copy_all_cgu_workproducts_to_incr_comp_cache_dir"); |
dfeec247 | 495 | |
a1dfa0c6 | 496 | for module in compiled_modules.modules.iter().filter(|m| m.kind == ModuleKind::Regular) { |
3dfed10e | 497 | let path = module.object.as_ref().cloned(); |
a1dfa0c6 XL |
498 | |
499 | if let Some((id, product)) = | |
f9f354fc | 500 | copy_cgu_workproduct_to_incr_comp_cache_dir(sess, &module.name, &path) |
dfeec247 | 501 | { |
a1dfa0c6 XL |
502 | work_products.insert(id, product); |
503 | } | |
504 | } | |
505 | ||
506 | work_products | |
507 | } | |
508 | ||
dfeec247 XL |
509 | fn produce_final_output_artifacts( |
510 | sess: &Session, | |
511 | compiled_modules: &CompiledModules, | |
512 | crate_output: &OutputFilenames, | |
513 | ) { | |
a1dfa0c6 XL |
514 | let mut user_wants_bitcode = false; |
515 | let mut user_wants_objects = false; | |
516 | ||
517 | // Produce final compile outputs. | |
518 | let copy_gracefully = |from: &Path, to: &Path| { | |
519 | if let Err(e) = fs::copy(from, to) { | |
520 | sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e)); | |
521 | } | |
522 | }; | |
523 | ||
dfeec247 | 524 | let copy_if_one_unit = |output_type: OutputType, keep_numbered: bool| { |
a1dfa0c6 XL |
525 | if compiled_modules.modules.len() == 1 { |
526 | // 1) Only one codegen unit. In this case it's no difficulty | |
527 | // to copy `foo.0.x` to `foo.x`. | |
528 | let module_name = Some(&compiled_modules.modules[0].name[..]); | |
529 | let path = crate_output.temp_path(output_type, module_name); | |
dfeec247 | 530 | copy_gracefully(&path, &crate_output.path(output_type)); |
a1dfa0c6 XL |
531 | if !sess.opts.cg.save_temps && !keep_numbered { |
532 | // The user just wants `foo.x`, not `foo.#module-name#.x`. | |
6a06907d | 533 | ensure_removed(sess.diagnostic(), &path); |
a1dfa0c6 XL |
534 | } |
535 | } else { | |
dfeec247 XL |
536 | let ext = crate_output |
537 | .temp_path(output_type, None) | |
538 | .extension() | |
539 | .unwrap() | |
540 | .to_str() | |
541 | .unwrap() | |
542 | .to_owned(); | |
a1dfa0c6 XL |
543 | |
544 | if crate_output.outputs.contains_key(&output_type) { | |
545 | // 2) Multiple codegen units, with `--emit foo=some_name`. We have | |
546 | // no good solution for this case, so warn the user. | |
dfeec247 XL |
547 | sess.warn(&format!( |
548 | "ignoring emit path because multiple .{} files \ | |
549 | were produced", | |
550 | ext | |
551 | )); | |
a1dfa0c6 XL |
552 | } else if crate_output.single_output_file.is_some() { |
553 | // 3) Multiple codegen units, with `-o some_name`. We have | |
554 | // no good solution for this case, so warn the user. | |
dfeec247 XL |
555 | sess.warn(&format!( |
556 | "ignoring -o because multiple .{} files \ | |
557 | were produced", | |
558 | ext | |
559 | )); | |
a1dfa0c6 XL |
560 | } else { |
561 | // 4) Multiple codegen units, but no explicit name. We | |
562 | // just leave the `foo.0.x` files in place. | |
563 | // (We don't have to do any work in this case.) | |
564 | } | |
565 | } | |
566 | }; | |
567 | ||
568 | // Flag to indicate whether the user explicitly requested bitcode. | |
569 | // Otherwise, we produced it only as a temporary output, and will need | |
570 | // to get rid of it. | |
571 | for output_type in crate_output.outputs.keys() { | |
572 | match *output_type { | |
573 | OutputType::Bitcode => { | |
574 | user_wants_bitcode = true; | |
575 | // Copy to .bc, but always keep the .0.bc. There is a later | |
576 | // check to figure out if we should delete .0.bc files, or keep | |
577 | // them for making an rlib. | |
578 | copy_if_one_unit(OutputType::Bitcode, true); | |
579 | } | |
580 | OutputType::LlvmAssembly => { | |
581 | copy_if_one_unit(OutputType::LlvmAssembly, false); | |
582 | } | |
583 | OutputType::Assembly => { | |
584 | copy_if_one_unit(OutputType::Assembly, false); | |
585 | } | |
586 | OutputType::Object => { | |
587 | user_wants_objects = true; | |
588 | copy_if_one_unit(OutputType::Object, true); | |
589 | } | |
dfeec247 | 590 | OutputType::Mir | OutputType::Metadata | OutputType::Exe | OutputType::DepInfo => {} |
a1dfa0c6 XL |
591 | } |
592 | } | |
593 | ||
594 | // Clean up unwanted temporary files. | |
595 | ||
596 | // We create the following files by default: | |
597 | // - #crate#.#module-name#.bc | |
598 | // - #crate#.#module-name#.o | |
599 | // - #crate#.crate.metadata.bc | |
600 | // - #crate#.crate.metadata.o | |
601 | // - #crate#.o (linked from crate.##.o) | |
602 | // - #crate#.bc (copied from crate.##.bc) | |
603 | // We may create additional files if requested by the user (through | |
604 | // `-C save-temps` or `--emit=` flags). | |
605 | ||
606 | if !sess.opts.cg.save_temps { | |
607 | // Remove the temporary .#module-name#.o objects. If the user didn't | |
608 | // explicitly request bitcode (with --emit=bc), and the bitcode is not | |
609 | // needed for building an rlib, then we must remove .#module-name#.bc as | |
610 | // well. | |
611 | ||
612 | // Specific rules for keeping .#module-name#.bc: | |
613 | // - If the user requested bitcode (`user_wants_bitcode`), and | |
614 | // codegen_units > 1, then keep it. | |
615 | // - If the user requested bitcode but codegen_units == 1, then we | |
616 | // can toss .#module-name#.bc because we copied it to .bc earlier. | |
617 | // - If we're not building an rlib and the user didn't request | |
618 | // bitcode, then delete .#module-name#.bc. | |
619 | // If you change how this works, also update back::link::link_rlib, | |
620 | // where .#module-name#.bc files are (maybe) deleted after making an | |
621 | // rlib. | |
622 | let needs_crate_object = crate_output.outputs.contains_key(&OutputType::Exe); | |
623 | ||
624 | let keep_numbered_bitcode = user_wants_bitcode && sess.codegen_units() > 1; | |
625 | ||
dfeec247 XL |
626 | let keep_numbered_objects = |
627 | needs_crate_object || (user_wants_objects && sess.codegen_units() > 1); | |
a1dfa0c6 XL |
628 | |
629 | for module in compiled_modules.modules.iter() { | |
630 | if let Some(ref path) = module.object { | |
631 | if !keep_numbered_objects { | |
6a06907d | 632 | ensure_removed(sess.diagnostic(), path); |
a1dfa0c6 XL |
633 | } |
634 | } | |
635 | ||
fc512014 XL |
636 | if let Some(ref path) = module.dwarf_object { |
637 | if !keep_numbered_objects { | |
6a06907d | 638 | ensure_removed(sess.diagnostic(), path); |
fc512014 XL |
639 | } |
640 | } | |
641 | ||
a1dfa0c6 XL |
642 | if let Some(ref path) = module.bytecode { |
643 | if !keep_numbered_bitcode { | |
6a06907d | 644 | ensure_removed(sess.diagnostic(), path); |
a1dfa0c6 XL |
645 | } |
646 | } | |
647 | } | |
648 | ||
649 | if !user_wants_bitcode { | |
a1dfa0c6 XL |
650 | if let Some(ref allocator_module) = compiled_modules.allocator_module { |
651 | if let Some(ref path) = allocator_module.bytecode { | |
6a06907d | 652 | ensure_removed(sess.diagnostic(), path); |
a1dfa0c6 XL |
653 | } |
654 | } | |
655 | } | |
656 | } | |
657 | ||
658 | // We leave the following files around by default: | |
659 | // - #crate#.o | |
660 | // - #crate#.crate.metadata.o | |
661 | // - #crate#.bc | |
662 | // These are used in linking steps and will be cleaned up afterward. | |
663 | } | |
664 | ||
a1dfa0c6 XL |
665 | pub enum WorkItem<B: WriteBackendMethods> { |
666 | /// Optimize a newly codegened, totally unoptimized module. | |
667 | Optimize(ModuleCodegen<B::Module>), | |
668 | /// Copy the post-LTO artifacts from the incremental cache to the output | |
669 | /// directory. | |
670 | CopyPostLtoArtifacts(CachedModuleCodegen), | |
9fa01778 | 671 | /// Performs (Thin)LTO on the given module. |
a1dfa0c6 XL |
672 | LTO(lto::LtoModuleCodegen<B>), |
673 | } | |
674 | ||
675 | impl<B: WriteBackendMethods> WorkItem<B> { | |
676 | pub fn module_kind(&self) -> ModuleKind { | |
677 | match *self { | |
678 | WorkItem::Optimize(ref m) => m.kind, | |
dfeec247 | 679 | WorkItem::CopyPostLtoArtifacts(_) | WorkItem::LTO(_) => ModuleKind::Regular, |
a1dfa0c6 XL |
680 | } |
681 | } | |
682 | ||
74b04a01 | 683 | fn start_profiling<'a>(&self, cgcx: &'a CodegenContext<B>) -> TimingGuard<'a> { |
a1dfa0c6 | 684 | match *self { |
74b04a01 | 685 | WorkItem::Optimize(ref m) => { |
a2a8927a | 686 | cgcx.prof.generic_activity_with_arg("codegen_module_optimize", &*m.name) |
74b04a01 XL |
687 | } |
688 | WorkItem::CopyPostLtoArtifacts(ref m) => cgcx | |
689 | .prof | |
a2a8927a | 690 | .generic_activity_with_arg("codegen_copy_artifacts_from_incr_cache", &*m.name), |
74b04a01 XL |
691 | WorkItem::LTO(ref m) => { |
692 | cgcx.prof.generic_activity_with_arg("codegen_module_perform_lto", m.name()) | |
693 | } | |
a1dfa0c6 XL |
694 | } |
695 | } | |
6a06907d XL |
696 | |
697 | /// Generate a short description of this work item suitable for use as a thread name. | |
698 | fn short_description(&self) -> String { | |
699 | // `pthread_setname()` on *nix is limited to 15 characters and longer names are ignored. | |
700 | // Use very short descriptions in this case to maximize the space available for the module name. | |
701 | // Windows does not have that limitation so use slightly more descriptive names there. | |
702 | match self { | |
703 | WorkItem::Optimize(m) => { | |
704 | #[cfg(windows)] | |
705 | return format!("optimize module {}", m.name); | |
706 | #[cfg(not(windows))] | |
707 | return format!("opt {}", m.name); | |
708 | } | |
709 | WorkItem::CopyPostLtoArtifacts(m) => { | |
710 | #[cfg(windows)] | |
711 | return format!("copy LTO artifacts for {}", m.name); | |
712 | #[cfg(not(windows))] | |
713 | return format!("copy {}", m.name); | |
714 | } | |
715 | WorkItem::LTO(m) => { | |
716 | #[cfg(windows)] | |
717 | return format!("LTO module {}", m.name()); | |
718 | #[cfg(not(windows))] | |
719 | return format!("LTO {}", m.name()); | |
720 | } | |
721 | } | |
722 | } | |
a1dfa0c6 XL |
723 | } |
724 | ||
0731742a | 725 | enum WorkItemResult<B: WriteBackendMethods> { |
a1dfa0c6 | 726 | Compiled(CompiledModule), |
1b1a35ee | 727 | NeedsLink(ModuleCodegen<B::Module>), |
9fa01778 | 728 | NeedsFatLTO(FatLTOInput<B>), |
0731742a | 729 | NeedsThinLTO(String, B::ThinBuffer), |
a1dfa0c6 XL |
730 | } |
731 | ||
9fa01778 | 732 | pub enum FatLTOInput<B: WriteBackendMethods> { |
dfeec247 | 733 | Serialized { name: String, buffer: B::ModuleBuffer }, |
9fa01778 XL |
734 | InMemory(ModuleCodegen<B::Module>), |
735 | } | |
736 | ||
a1dfa0c6 XL |
737 | fn execute_work_item<B: ExtraBackendMethods>( |
738 | cgcx: &CodegenContext<B>, | |
739 | work_item: WorkItem<B>, | |
0731742a | 740 | ) -> Result<WorkItemResult<B>, FatalError> { |
a1dfa0c6 XL |
741 | let module_config = cgcx.config(work_item.module_kind()); |
742 | ||
743 | match work_item { | |
dfeec247 | 744 | WorkItem::Optimize(module) => execute_optimize_work_item(cgcx, module, module_config), |
a1dfa0c6 | 745 | WorkItem::CopyPostLtoArtifacts(module) => { |
6a06907d | 746 | Ok(execute_copy_from_cache_work_item(cgcx, module, module_config)) |
a1dfa0c6 | 747 | } |
dfeec247 | 748 | WorkItem::LTO(module) => execute_lto_work_item(cgcx, module, module_config), |
a1dfa0c6 XL |
749 | } |
750 | } | |
751 | ||
74b04a01 | 752 | // Actual LTO type we end up choosing based on multiple factors. |
f9f354fc | 753 | pub enum ComputedLtoType { |
0731742a XL |
754 | No, |
755 | Thin, | |
756 | Fat, | |
757 | } | |
758 | ||
f9f354fc XL |
759 | pub fn compute_per_cgu_lto_type( |
760 | sess_lto: &Lto, | |
761 | opts: &config::Options, | |
762 | sess_crate_types: &[CrateType], | |
763 | module_kind: ModuleKind, | |
764 | ) -> ComputedLtoType { | |
765 | // Metadata modules never participate in LTO regardless of the lto | |
766 | // settings. | |
767 | if module_kind == ModuleKind::Metadata { | |
768 | return ComputedLtoType::No; | |
a1dfa0c6 XL |
769 | } |
770 | ||
0731742a XL |
771 | // If the linker does LTO, we don't have to do it. Note that we |
772 | // keep doing full LTO, if it is requested, as not to break the | |
773 | // assumption that the output will be a single module. | |
f9f354fc | 774 | let linker_does_lto = opts.cg.linker_plugin_lto.enabled(); |
a1dfa0c6 | 775 | |
0731742a XL |
776 | // When we're automatically doing ThinLTO for multi-codegen-unit |
777 | // builds we don't actually want to LTO the allocator modules if | |
778 | // it shows up. This is due to various linker shenanigans that | |
779 | // we'll encounter later. | |
f9f354fc | 780 | let is_allocator = module_kind == ModuleKind::Allocator; |
a1dfa0c6 | 781 | |
0731742a XL |
782 | // We ignore a request for full crate grath LTO if the cate type |
783 | // is only an rlib, as there is no full crate graph to process, | |
784 | // that'll happen later. | |
785 | // | |
786 | // This use case currently comes up primarily for targets that | |
787 | // require LTO so the request for LTO is always unconditionally | |
788 | // passed down to the backend, but we don't actually want to do | |
789 | // anything about it yet until we've got a final product. | |
f9f354fc | 790 | let is_rlib = sess_crate_types.len() == 1 && sess_crate_types[0] == CrateType::Rlib; |
a1dfa0c6 | 791 | |
f9f354fc XL |
792 | match sess_lto { |
793 | Lto::ThinLocal if !linker_does_lto && !is_allocator => ComputedLtoType::Thin, | |
794 | Lto::Thin if !linker_does_lto && !is_rlib => ComputedLtoType::Thin, | |
795 | Lto::Fat if !is_rlib => ComputedLtoType::Fat, | |
796 | _ => ComputedLtoType::No, | |
797 | } | |
798 | } | |
799 | ||
800 | fn execute_optimize_work_item<B: ExtraBackendMethods>( | |
801 | cgcx: &CodegenContext<B>, | |
802 | module: ModuleCodegen<B::Module>, | |
803 | module_config: &ModuleConfig, | |
804 | ) -> Result<WorkItemResult<B>, FatalError> { | |
805 | let diag_handler = cgcx.create_diag_handler(); | |
806 | ||
807 | unsafe { | |
808 | B::optimize(cgcx, &diag_handler, &module, module_config)?; | |
809 | } | |
810 | ||
811 | // After we've done the initial round of optimizations we need to | |
812 | // decide whether to synchronously codegen this module or ship it | |
813 | // back to the coordinator thread for further LTO processing (which | |
814 | // has to wait for all the initial modules to be optimized). | |
815 | ||
816 | let lto_type = compute_per_cgu_lto_type(&cgcx.lto, &cgcx.opts, &cgcx.crate_types, module.kind); | |
0731742a | 817 | |
9fa01778 XL |
818 | // If we're doing some form of incremental LTO then we need to be sure to |
819 | // save our module to disk first. | |
820 | let bitcode = if cgcx.config(module.kind).emit_pre_lto_bc { | |
821 | let filename = pre_lto_bitcode_filename(&module.name); | |
822 | cgcx.incr_comp_session_dir.as_ref().map(|path| path.join(&filename)) | |
823 | } else { | |
824 | None | |
825 | }; | |
826 | ||
1b1a35ee XL |
827 | match lto_type { |
828 | ComputedLtoType::No => finish_intra_module_work(cgcx, module, module_config), | |
0731742a | 829 | ComputedLtoType::Thin => { |
9fa01778 XL |
830 | let (name, thin_buffer) = B::prepare_thin(module); |
831 | if let Some(path) = bitcode { | |
832 | fs::write(&path, thin_buffer.data()).unwrap_or_else(|e| { | |
dfeec247 | 833 | panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e); |
9fa01778 XL |
834 | }); |
835 | } | |
1b1a35ee | 836 | Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer)) |
0731742a | 837 | } |
dfeec247 XL |
838 | ComputedLtoType::Fat => match bitcode { |
839 | Some(path) => { | |
840 | let (name, buffer) = B::serialize_module(module); | |
841 | fs::write(&path, buffer.data()).unwrap_or_else(|e| { | |
842 | panic!("Error writing pre-lto-bitcode file `{}`: {}", path.display(), e); | |
843 | }); | |
1b1a35ee | 844 | Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::Serialized { name, buffer })) |
9fa01778 | 845 | } |
1b1a35ee | 846 | None => Ok(WorkItemResult::NeedsFatLTO(FatLTOInput::InMemory(module))), |
dfeec247 | 847 | }, |
1b1a35ee | 848 | } |
a1dfa0c6 XL |
849 | } |
850 | ||
851 | fn execute_copy_from_cache_work_item<B: ExtraBackendMethods>( | |
852 | cgcx: &CodegenContext<B>, | |
853 | module: CachedModuleCodegen, | |
854 | module_config: &ModuleConfig, | |
6a06907d | 855 | ) -> WorkItemResult<B> { |
dfeec247 | 856 | let incr_comp_session_dir = cgcx.incr_comp_session_dir.as_ref().unwrap(); |
a1dfa0c6 | 857 | let mut object = None; |
f9f354fc XL |
858 | if let Some(saved_file) = module.source.saved_file { |
859 | let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, Some(&module.name)); | |
860 | object = Some(obj_out.clone()); | |
dfeec247 XL |
861 | let source_file = in_incr_comp_dir(&incr_comp_session_dir, &saved_file); |
862 | debug!( | |
863 | "copying pre-existing module `{}` from {:?} to {}", | |
864 | module.name, | |
865 | source_file, | |
866 | obj_out.display() | |
867 | ); | |
a1dfa0c6 XL |
868 | if let Err(err) = link_or_copy(&source_file, &obj_out) { |
869 | let diag_handler = cgcx.create_diag_handler(); | |
dfeec247 XL |
870 | diag_handler.err(&format!( |
871 | "unable to copy {} to {}: {}", | |
872 | source_file.display(), | |
873 | obj_out.display(), | |
874 | err | |
875 | )); | |
a1dfa0c6 XL |
876 | } |
877 | } | |
878 | ||
ba9703b0 | 879 | assert_eq!(object.is_some(), module_config.emit_obj != EmitObj::None); |
a1dfa0c6 | 880 | |
6a06907d | 881 | WorkItemResult::Compiled(CompiledModule { |
a1dfa0c6 XL |
882 | name: module.name, |
883 | kind: ModuleKind::Regular, | |
884 | object, | |
fc512014 | 885 | dwarf_object: None, |
f9f354fc | 886 | bytecode: None, |
6a06907d | 887 | }) |
a1dfa0c6 XL |
888 | } |
889 | ||
890 | fn execute_lto_work_item<B: ExtraBackendMethods>( | |
891 | cgcx: &CodegenContext<B>, | |
892 | mut module: lto::LtoModuleCodegen<B>, | |
893 | module_config: &ModuleConfig, | |
1b1a35ee XL |
894 | ) -> Result<WorkItemResult<B>, FatalError> { |
895 | let module = unsafe { module.optimize(cgcx)? }; | |
896 | finish_intra_module_work(cgcx, module, module_config) | |
897 | } | |
898 | ||
899 | fn finish_intra_module_work<B: ExtraBackendMethods>( | |
900 | cgcx: &CodegenContext<B>, | |
901 | module: ModuleCodegen<B::Module>, | |
902 | module_config: &ModuleConfig, | |
0731742a | 903 | ) -> Result<WorkItemResult<B>, FatalError> { |
a1dfa0c6 XL |
904 | let diag_handler = cgcx.create_diag_handler(); |
905 | ||
1b1a35ee XL |
906 | if !cgcx.opts.debugging_opts.combine_cgu |
907 | || module.kind == ModuleKind::Metadata | |
908 | || module.kind == ModuleKind::Allocator | |
909 | { | |
910 | let module = unsafe { B::codegen(cgcx, &diag_handler, module, module_config)? }; | |
a1dfa0c6 | 911 | Ok(WorkItemResult::Compiled(module)) |
1b1a35ee XL |
912 | } else { |
913 | Ok(WorkItemResult::NeedsLink(module)) | |
a1dfa0c6 XL |
914 | } |
915 | } | |
916 | ||
917 | pub enum Message<B: WriteBackendMethods> { | |
918 | Token(io::Result<Acquired>), | |
0731742a | 919 | NeedsFatLTO { |
9fa01778 | 920 | result: FatLTOInput<B>, |
a1dfa0c6 XL |
921 | worker_id: usize, |
922 | }, | |
0731742a XL |
923 | NeedsThinLTO { |
924 | name: String, | |
925 | thin_buffer: B::ThinBuffer, | |
926 | worker_id: usize, | |
927 | }, | |
1b1a35ee XL |
928 | NeedsLink { |
929 | module: ModuleCodegen<B::Module>, | |
930 | worker_id: usize, | |
931 | }, | |
a1dfa0c6 | 932 | Done { |
dfeec247 | 933 | result: Result<CompiledModule, Option<WorkerFatalError>>, |
a1dfa0c6 XL |
934 | worker_id: usize, |
935 | }, | |
936 | CodegenDone { | |
937 | llvm_work_item: WorkItem<B>, | |
938 | cost: u64, | |
939 | }, | |
940 | AddImportOnlyModule { | |
941 | module_data: SerializedModule<B::ModuleBuffer>, | |
942 | work_product: WorkProduct, | |
943 | }, | |
944 | CodegenComplete, | |
945 | CodegenItem, | |
946 | CodegenAborted, | |
947 | } | |
948 | ||
949 | struct Diagnostic { | |
950 | msg: String, | |
951 | code: Option<DiagnosticId>, | |
952 | lvl: Level, | |
953 | } | |
954 | ||
955 | #[derive(PartialEq, Clone, Copy, Debug)] | |
956 | enum MainThreadWorkerState { | |
957 | Idle, | |
958 | Codegenning, | |
959 | LLVMing, | |
960 | } | |
961 | ||
962 | fn start_executing_work<B: ExtraBackendMethods>( | |
963 | backend: B, | |
dc9dc135 | 964 | tcx: TyCtxt<'_>, |
a1dfa0c6 XL |
965 | crate_info: &CrateInfo, |
966 | shared_emitter: SharedEmitter, | |
967 | codegen_worker_send: Sender<Message<B>>, | |
968 | coordinator_receive: Receiver<Box<dyn Any + Send>>, | |
969 | total_cgus: usize, | |
970 | jobserver: Client, | |
ba9703b0 | 971 | regular_config: Arc<ModuleConfig>, |
a1dfa0c6 | 972 | metadata_config: Arc<ModuleConfig>, |
dc9dc135 | 973 | allocator_config: Arc<ModuleConfig>, |
e74abb32 | 974 | tx_to_llvm_workers: Sender<Box<dyn Any + Send>>, |
a1dfa0c6 | 975 | ) -> thread::JoinHandle<Result<CompiledModules, ()>> { |
e74abb32 | 976 | let coordinator_send = tx_to_llvm_workers; |
a1dfa0c6 XL |
977 | let sess = tcx.sess; |
978 | ||
979 | // Compute the set of symbols we need to retain when doing LTO (if we need to) | |
980 | let exported_symbols = { | |
981 | let mut exported_symbols = FxHashMap::default(); | |
982 | ||
983 | let copy_symbols = |cnum| { | |
dfeec247 XL |
984 | let symbols = tcx |
985 | .exported_symbols(cnum) | |
986 | .iter() | |
987 | .map(|&(s, lvl)| (symbol_name_for_instance_in_crate(tcx, s, cnum), lvl)) | |
988 | .collect(); | |
a1dfa0c6 XL |
989 | Arc::new(symbols) |
990 | }; | |
991 | ||
992 | match sess.lto() { | |
993 | Lto::No => None, | |
994 | Lto::ThinLocal => { | |
995 | exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE)); | |
996 | Some(Arc::new(exported_symbols)) | |
997 | } | |
998 | Lto::Fat | Lto::Thin => { | |
999 | exported_symbols.insert(LOCAL_CRATE, copy_symbols(LOCAL_CRATE)); | |
136023e0 | 1000 | for &cnum in tcx.crates(()).iter() { |
a1dfa0c6 XL |
1001 | exported_symbols.insert(cnum, copy_symbols(cnum)); |
1002 | } | |
1003 | Some(Arc::new(exported_symbols)) | |
1004 | } | |
1005 | } | |
1006 | }; | |
1007 | ||
1008 | // First up, convert our jobserver into a helper thread so we can use normal | |
1009 | // mpsc channels to manage our messages and such. | |
1010 | // After we've requested tokens then we'll, when we can, | |
1011 | // get tokens on `coordinator_receive` which will | |
1012 | // get managed in the main loop below. | |
1013 | let coordinator_send2 = coordinator_send.clone(); | |
dfeec247 XL |
1014 | let helper = jobserver |
1015 | .into_helper_thread(move |token| { | |
1016 | drop(coordinator_send2.send(Box::new(Message::Token::<B>(token)))); | |
1017 | }) | |
1018 | .expect("failed to spawn helper thread"); | |
a1dfa0c6 XL |
1019 | |
1020 | let mut each_linked_rlib_for_lto = Vec::new(); | |
e74abb32 | 1021 | drop(link::each_linked_rlib(crate_info, &mut |cnum, path| { |
a1dfa0c6 | 1022 | if link::ignored_for_lto(sess, crate_info, cnum) { |
dfeec247 | 1023 | return; |
a1dfa0c6 XL |
1024 | } |
1025 | each_linked_rlib_for_lto.push((cnum, path.to_path_buf())); | |
1026 | })); | |
1027 | ||
532ac7d7 | 1028 | let ol = if tcx.sess.opts.debugging_opts.no_codegen |
dfeec247 XL |
1029 | || !tcx.sess.opts.output_types.should_codegen() |
1030 | { | |
532ac7d7 XL |
1031 | // If we know that we won’t be doing codegen, create target machines without optimisation. |
1032 | config::OptLevel::No | |
1033 | } else { | |
17df50a5 | 1034 | tcx.backend_optimization_level(()) |
532ac7d7 | 1035 | }; |
a1dfa0c6 XL |
1036 | let cgcx = CodegenContext::<B> { |
1037 | backend: backend.clone(), | |
f9f354fc | 1038 | crate_types: sess.crate_types().to_vec(), |
a1dfa0c6 XL |
1039 | each_linked_rlib_for_lto, |
1040 | lto: sess.lto(), | |
a1dfa0c6 XL |
1041 | fewer_names: sess.fewer_names(), |
1042 | save_temps: sess.opts.cg.save_temps, | |
3c0e092e | 1043 | time_trace: sess.opts.debugging_opts.llvm_time_trace, |
a1dfa0c6 | 1044 | opts: Arc::new(sess.opts.clone()), |
e74abb32 | 1045 | prof: sess.prof.clone(), |
a1dfa0c6 | 1046 | exported_symbols, |
a1dfa0c6 XL |
1047 | remark: sess.opts.cg.remark.clone(), |
1048 | worker: 0, | |
1049 | incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()), | |
1050 | cgu_reuse_tracker: sess.cgu_reuse_tracker.clone(), | |
1051 | coordinator_send, | |
1052 | diag_emitter: shared_emitter.clone(), | |
5099ac24 | 1053 | output_filenames: tcx.output_filenames(()).clone(), |
ba9703b0 | 1054 | regular_module_config: regular_config, |
a1dfa0c6 XL |
1055 | metadata_module_config: metadata_config, |
1056 | allocator_module_config: allocator_config, | |
fc512014 | 1057 | tm_factory: backend.target_machine_factory(tcx.sess, ol), |
a1dfa0c6 XL |
1058 | total_cgus, |
1059 | msvc_imps_needed: msvc_imps_needed(tcx), | |
fc512014 | 1060 | is_pe_coff: tcx.sess.target.is_like_windows, |
5869c6ff | 1061 | target_can_use_split_dwarf: tcx.sess.target_can_use_split_dwarf(), |
29967ef6 XL |
1062 | target_pointer_width: tcx.sess.target.pointer_width, |
1063 | target_arch: tcx.sess.target.arch.clone(), | |
a1dfa0c6 | 1064 | debuginfo: tcx.sess.opts.debuginfo, |
5869c6ff | 1065 | split_debuginfo: tcx.sess.split_debuginfo(), |
a2a8927a | 1066 | split_dwarf_kind: tcx.sess.opts.debugging_opts.split_dwarf_kind, |
a1dfa0c6 XL |
1067 | }; |
1068 | ||
1069 | // This is the "main loop" of parallel work happening for parallel codegen. | |
1070 | // It's here that we manage parallelism, schedule work, and work with | |
1071 | // messages coming from clients. | |
1072 | // | |
1073 | // There are a few environmental pre-conditions that shape how the system | |
1074 | // is set up: | |
1075 | // | |
1076 | // - Error reporting only can happen on the main thread because that's the | |
1077 | // only place where we have access to the compiler `Session`. | |
1078 | // - LLVM work can be done on any thread. | |
1079 | // - Codegen can only happen on the main thread. | |
cdc7bbd5 | 1080 | // - Each thread doing substantial work must be in possession of a `Token` |
a1dfa0c6 XL |
1081 | // from the `Jobserver`. |
1082 | // - The compiler process always holds one `Token`. Any additional `Tokens` | |
1083 | // have to be requested from the `Jobserver`. | |
1084 | // | |
1085 | // Error Reporting | |
1086 | // =============== | |
1087 | // The error reporting restriction is handled separately from the rest: We | |
1088 | // set up a `SharedEmitter` the holds an open channel to the main thread. | |
1089 | // When an error occurs on any thread, the shared emitter will send the | |
1090 | // error message to the receiver main thread (`SharedEmitterMain`). The | |
1091 | // main thread will periodically query this error message queue and emit | |
1092 | // any error messages it has received. It might even abort compilation if | |
1093 | // has received a fatal error. In this case we rely on all other threads | |
1094 | // being torn down automatically with the main thread. | |
1095 | // Since the main thread will often be busy doing codegen work, error | |
1096 | // reporting will be somewhat delayed, since the message queue can only be | |
1097 | // checked in between to work packages. | |
1098 | // | |
1099 | // Work Processing Infrastructure | |
1100 | // ============================== | |
1101 | // The work processing infrastructure knows three major actors: | |
1102 | // | |
1103 | // - the coordinator thread, | |
1104 | // - the main thread, and | |
1105 | // - LLVM worker threads | |
1106 | // | |
1107 | // The coordinator thread is running a message loop. It instructs the main | |
1108 | // thread about what work to do when, and it will spawn off LLVM worker | |
1109 | // threads as open LLVM WorkItems become available. | |
1110 | // | |
1111 | // The job of the main thread is to codegen CGUs into LLVM work package | |
1112 | // (since the main thread is the only thread that can do this). The main | |
1113 | // thread will block until it receives a message from the coordinator, upon | |
1114 | // which it will codegen one CGU, send it to the coordinator and block | |
1115 | // again. This way the coordinator can control what the main thread is | |
1116 | // doing. | |
1117 | // | |
1118 | // The coordinator keeps a queue of LLVM WorkItems, and when a `Token` is | |
1119 | // available, it will spawn off a new LLVM worker thread and let it process | |
1120 | // that a WorkItem. When a LLVM worker thread is done with its WorkItem, | |
1121 | // it will just shut down, which also frees all resources associated with | |
1122 | // the given LLVM module, and sends a message to the coordinator that the | |
1123 | // has been completed. | |
1124 | // | |
1125 | // Work Scheduling | |
1126 | // =============== | |
1127 | // The scheduler's goal is to minimize the time it takes to complete all | |
1128 | // work there is, however, we also want to keep memory consumption low | |
1129 | // if possible. These two goals are at odds with each other: If memory | |
1130 | // consumption were not an issue, we could just let the main thread produce | |
1131 | // LLVM WorkItems at full speed, assuring maximal utilization of | |
cdc7bbd5 | 1132 | // Tokens/LLVM worker threads. However, since codegen is usually faster |
a1dfa0c6 XL |
1133 | // than LLVM processing, the queue of LLVM WorkItems would fill up and each |
1134 | // WorkItem potentially holds on to a substantial amount of memory. | |
1135 | // | |
1136 | // So the actual goal is to always produce just enough LLVM WorkItems as | |
1137 | // not to starve our LLVM worker threads. That means, once we have enough | |
1138 | // WorkItems in our queue, we can block the main thread, so it does not | |
1139 | // produce more until we need them. | |
1140 | // | |
1141 | // Doing LLVM Work on the Main Thread | |
1142 | // ---------------------------------- | |
1143 | // Since the main thread owns the compiler processes implicit `Token`, it is | |
1144 | // wasteful to keep it blocked without doing any work. Therefore, what we do | |
1145 | // in this case is: We spawn off an additional LLVM worker thread that helps | |
1146 | // reduce the queue. The work it is doing corresponds to the implicit | |
1147 | // `Token`. The coordinator will mark the main thread as being busy with | |
1148 | // LLVM work. (The actual work happens on another OS thread but we just care | |
1149 | // about `Tokens`, not actual threads). | |
1150 | // | |
1151 | // When any LLVM worker thread finishes while the main thread is marked as | |
1152 | // "busy with LLVM work", we can do a little switcheroo: We give the Token | |
1153 | // of the just finished thread to the LLVM worker thread that is working on | |
1154 | // behalf of the main thread's implicit Token, thus freeing up the main | |
1155 | // thread again. The coordinator can then again decide what the main thread | |
1156 | // should do. This allows the coordinator to make decisions at more points | |
1157 | // in time. | |
1158 | // | |
1159 | // Striking a Balance between Throughput and Memory Consumption | |
1160 | // ------------------------------------------------------------ | |
1161 | // Since our two goals, (1) use as many Tokens as possible and (2) keep | |
1162 | // memory consumption as low as possible, are in conflict with each other, | |
1163 | // we have to find a trade off between them. Right now, the goal is to keep | |
1164 | // all workers busy, which means that no worker should find the queue empty | |
1165 | // when it is ready to start. | |
1166 | // How do we do achieve this? Good question :) We actually never know how | |
1167 | // many `Tokens` are potentially available so it's hard to say how much to | |
1168 | // fill up the queue before switching the main thread to LLVM work. Also we | |
1169 | // currently don't have a means to estimate how long a running LLVM worker | |
1170 | // will still be busy with it's current WorkItem. However, we know the | |
1171 | // maximal count of available Tokens that makes sense (=the number of CPU | |
1172 | // cores), so we can take a conservative guess. The heuristic we use here | |
1173 | // is implemented in the `queue_full_enough()` function. | |
1174 | // | |
1175 | // Some Background on Jobservers | |
1176 | // ----------------------------- | |
1177 | // It's worth also touching on the management of parallelism here. We don't | |
1178 | // want to just spawn a thread per work item because while that's optimal | |
1179 | // parallelism it may overload a system with too many threads or violate our | |
1180 | // configuration for the maximum amount of cpu to use for this process. To | |
1181 | // manage this we use the `jobserver` crate. | |
1182 | // | |
1183 | // Job servers are an artifact of GNU make and are used to manage | |
1184 | // parallelism between processes. A jobserver is a glorified IPC semaphore | |
1185 | // basically. Whenever we want to run some work we acquire the semaphore, | |
1186 | // and whenever we're done with that work we release the semaphore. In this | |
1187 | // manner we can ensure that the maximum number of parallel workers is | |
1188 | // capped at any one point in time. | |
1189 | // | |
1190 | // LTO and the coordinator thread | |
1191 | // ------------------------------ | |
1192 | // | |
1193 | // The final job the coordinator thread is responsible for is managing LTO | |
1194 | // and how that works. When LTO is requested what we'll to is collect all | |
1195 | // optimized LLVM modules into a local vector on the coordinator. Once all | |
1196 | // modules have been codegened and optimized we hand this to the `lto` | |
1197 | // module for further optimization. The `lto` module will return back a list | |
1198 | // of more modules to work on, which the coordinator will continue to spawn | |
1199 | // work for. | |
1200 | // | |
1201 | // Each LLVM module is automatically sent back to the coordinator for LTO if | |
1202 | // necessary. There's already optimizations in place to avoid sending work | |
1203 | // back to the coordinator if LTO isn't requested. | |
3c0e092e | 1204 | return B::spawn_thread(cgcx.time_trace, move || { |
a1dfa0c6 XL |
1205 | let mut worker_id_counter = 0; |
1206 | let mut free_worker_ids = Vec::new(); | |
1207 | let mut get_worker_id = |free_worker_ids: &mut Vec<usize>| { | |
1208 | if let Some(id) = free_worker_ids.pop() { | |
1209 | id | |
1210 | } else { | |
1211 | let id = worker_id_counter; | |
1212 | worker_id_counter += 1; | |
1213 | id | |
1214 | } | |
1215 | }; | |
1216 | ||
1217 | // This is where we collect codegen units that have gone all the way | |
1218 | // through codegen and LLVM. | |
1219 | let mut compiled_modules = vec![]; | |
a1dfa0c6 | 1220 | let mut compiled_allocator_module = None; |
1b1a35ee | 1221 | let mut needs_link = Vec::new(); |
0731742a XL |
1222 | let mut needs_fat_lto = Vec::new(); |
1223 | let mut needs_thin_lto = Vec::new(); | |
a1dfa0c6 XL |
1224 | let mut lto_import_only_modules = Vec::new(); |
1225 | let mut started_lto = false; | |
1226 | let mut codegen_aborted = false; | |
1227 | ||
1228 | // This flag tracks whether all items have gone through codegens | |
1229 | let mut codegen_done = false; | |
1230 | ||
1231 | // This is the queue of LLVM work items that still need processing. | |
1232 | let mut work_items = Vec::<(WorkItem<B>, u64)>::new(); | |
1233 | ||
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(); | |
1237 | ||
1238 | let mut main_thread_worker_state = MainThreadWorkerState::Idle; | |
1239 | let mut running = 0; | |
1240 | ||
dfeec247 XL |
1241 | let prof = &cgcx.prof; |
1242 | let mut llvm_start_time: Option<VerboseTimingGuard<'_>> = None; | |
a1dfa0c6 XL |
1243 | |
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 | |
1248 | // work to be done. | |
dfeec247 XL |
1249 | while !codegen_done |
1250 | || running > 0 | |
1251 | || (!codegen_aborted | |
74b04a01 XL |
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)) | |
a1dfa0c6 | 1257 | { |
a1dfa0c6 XL |
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. | |
1261 | if !codegen_done { | |
1262 | if main_thread_worker_state == MainThreadWorkerState::Idle { | |
6a06907d XL |
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; | |
1272 | ||
1273 | if !queue_full_enough(work_items.len(), anticipated_running) { | |
a1dfa0c6 | 1274 | // The queue is not full enough, codegen more items: |
74b04a01 | 1275 | if codegen_worker_send.send(Message::CodegenItem).is_err() { |
a1dfa0c6 XL |
1276 | panic!("Could not send Message::CodegenItem to main thread") |
1277 | } | |
1278 | main_thread_worker_state = MainThreadWorkerState::Codegenning; | |
1279 | } else { | |
1280 | // The queue is full enough to not let the worker | |
1281 | // threads starve. Use the implicit Token to do some | |
1282 | // LLVM work too. | |
dfeec247 XL |
1283 | let (item, _) = |
1284 | work_items.pop().expect("queue empty - queue_full_enough() broken?"); | |
a1dfa0c6 XL |
1285 | let cgcx = CodegenContext { |
1286 | worker: get_worker_id(&mut free_worker_ids), | |
dfeec247 | 1287 | ..cgcx.clone() |
a1dfa0c6 | 1288 | }; |
dfeec247 XL |
1289 | maybe_start_llvm_timer( |
1290 | prof, | |
1291 | cgcx.config(item.module_kind()), | |
1292 | &mut llvm_start_time, | |
1293 | ); | |
a1dfa0c6 XL |
1294 | main_thread_worker_state = MainThreadWorkerState::LLVMing; |
1295 | spawn_work(cgcx, item); | |
1296 | } | |
1297 | } | |
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 | |
1301 | } else { | |
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 | |
1306 | // queue to do LTO | |
74b04a01 | 1307 | if work_items.is_empty() |
dfeec247 XL |
1308 | && running == 0 |
1309 | && main_thread_worker_state == MainThreadWorkerState::Idle | |
1310 | { | |
a1dfa0c6 | 1311 | assert!(!started_lto); |
a1dfa0c6 | 1312 | started_lto = true; |
0731742a | 1313 | |
416331ca XL |
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); | |
0731742a | 1317 | |
dfeec247 XL |
1318 | for (work, cost) in |
1319 | generate_lto_work(&cgcx, needs_fat_lto, needs_thin_lto, import_only_modules) | |
1320 | { | |
a1dfa0c6 XL |
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(); | |
1327 | } | |
1328 | } | |
1329 | } | |
1330 | ||
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), | |
dfeec247 | 1339 | ..cgcx.clone() |
a1dfa0c6 | 1340 | }; |
dfeec247 XL |
1341 | maybe_start_llvm_timer( |
1342 | prof, | |
1343 | cgcx.config(item.module_kind()), | |
1344 | &mut llvm_start_time, | |
1345 | ); | |
a1dfa0c6 XL |
1346 | main_thread_worker_state = MainThreadWorkerState::LLVMing; |
1347 | spawn_work(cgcx, item); | |
1348 | } else { | |
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); | |
1356 | running -= 1; | |
1357 | main_thread_worker_state = MainThreadWorkerState::LLVMing; | |
1358 | } | |
1359 | } | |
dfeec247 XL |
1360 | MainThreadWorkerState::Codegenning => bug!( |
1361 | "codegen worker should not be codegenning after \ | |
1362 | codegen was already completed" | |
1363 | ), | |
a1dfa0c6 XL |
1364 | MainThreadWorkerState::LLVMing => { |
1365 | // Already making good use of that token | |
1366 | } | |
1367 | } | |
1368 | } | |
1369 | ||
1370 | // Spin up what work we can, only doing this while we've got available | |
1371 | // parallelism slots and work left to spawn. | |
74b04a01 | 1372 | while !codegen_aborted && !work_items.is_empty() && running < tokens.len() { |
a1dfa0c6 XL |
1373 | let (item, _) = work_items.pop().unwrap(); |
1374 | ||
dfeec247 | 1375 | maybe_start_llvm_timer(prof, cgcx.config(item.module_kind()), &mut llvm_start_time); |
a1dfa0c6 | 1376 | |
dfeec247 XL |
1377 | let cgcx = |
1378 | CodegenContext { worker: get_worker_id(&mut free_worker_ids), ..cgcx.clone() }; | |
a1dfa0c6 XL |
1379 | |
1380 | spawn_work(cgcx, item); | |
1381 | running += 1; | |
1382 | } | |
1383 | ||
1384 | // Relinquish accidentally acquired extra tokens | |
1385 | tokens.truncate(running); | |
1386 | ||
0731742a XL |
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; | |
1395 | } else { | |
1396 | running -= 1; | |
1397 | } | |
1398 | ||
1399 | free_worker_ids.push(worker_id); | |
1400 | }; | |
1401 | ||
a1dfa0c6 XL |
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) => { | |
1408 | match token { | |
1409 | Ok(token) => { | |
1410 | tokens.push(token); | |
1411 | ||
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; | |
1418 | running += 1; | |
1419 | } | |
1420 | } | |
1421 | Err(e) => { | |
1422 | let msg = &format!("failed to acquire jobserver token: {}", e); | |
1423 | shared_emitter.fatal(msg); | |
1424 | // Exit the coordinator thread | |
1425 | panic!("{}", msg) | |
1426 | } | |
1427 | } | |
1428 | } | |
1429 | ||
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 | |
1438 | // distributed. | |
dfeec247 | 1439 | let insertion_index = work_items.binary_search_by_key(&cost, |&(_, cost)| cost); |
a1dfa0c6 | 1440 | let insertion_index = match insertion_index { |
dfeec247 | 1441 | Ok(idx) | Err(idx) => idx, |
a1dfa0c6 XL |
1442 | }; |
1443 | work_items.insert(insertion_index, (llvm_work_item, cost)); | |
1444 | ||
1445 | if !cgcx.opts.debugging_opts.no_parallel_llvm { | |
1446 | helper.request_token(); | |
1447 | } | |
1448 | assert!(!codegen_aborted); | |
dfeec247 | 1449 | assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning); |
a1dfa0c6 XL |
1450 | main_thread_worker_state = MainThreadWorkerState::Idle; |
1451 | } | |
1452 | ||
1453 | Message::CodegenComplete => { | |
1454 | codegen_done = true; | |
1455 | assert!(!codegen_aborted); | |
dfeec247 | 1456 | assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning); |
a1dfa0c6 XL |
1457 | main_thread_worker_state = MainThreadWorkerState::Idle; |
1458 | } | |
1459 | ||
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; | |
dfeec247 | 1470 | assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning); |
a1dfa0c6 | 1471 | } |
a1dfa0c6 | 1472 | Message::Done { result: Ok(compiled_module), worker_id } => { |
0731742a | 1473 | free_worker(worker_id); |
a1dfa0c6 XL |
1474 | match compiled_module.kind { |
1475 | ModuleKind::Regular => { | |
1476 | compiled_modules.push(compiled_module); | |
1477 | } | |
a1dfa0c6 XL |
1478 | ModuleKind::Allocator => { |
1479 | assert!(compiled_allocator_module.is_none()); | |
1480 | compiled_allocator_module = Some(compiled_module); | |
1481 | } | |
a2a8927a | 1482 | ModuleKind::Metadata => bug!("Should be handled separately"), |
a1dfa0c6 XL |
1483 | } |
1484 | } | |
1b1a35ee XL |
1485 | Message::NeedsLink { module, worker_id } => { |
1486 | free_worker(worker_id); | |
1487 | needs_link.push(module); | |
1488 | } | |
0731742a | 1489 | Message::NeedsFatLTO { result, worker_id } => { |
a1dfa0c6 | 1490 | assert!(!started_lto); |
0731742a XL |
1491 | free_worker(worker_id); |
1492 | needs_fat_lto.push(result); | |
1493 | } | |
1494 | Message::NeedsThinLTO { name, thin_buffer, worker_id } => { | |
1495 | assert!(!started_lto); | |
1496 | free_worker(worker_id); | |
1497 | needs_thin_lto.push((name, thin_buffer)); | |
a1dfa0c6 XL |
1498 | } |
1499 | Message::AddImportOnlyModule { module_data, work_product } => { | |
1500 | assert!(!started_lto); | |
1501 | assert!(!codegen_done); | |
dfeec247 | 1502 | assert_eq!(main_thread_worker_state, MainThreadWorkerState::Codegenning); |
a1dfa0c6 XL |
1503 | lto_import_only_modules.push((module_data, work_product)); |
1504 | main_thread_worker_state = MainThreadWorkerState::Idle; | |
1505 | } | |
0731742a | 1506 | // If the thread failed that means it panicked, so we abort immediately. |
dfeec247 | 1507 | Message::Done { result: Err(None), worker_id: _ } => { |
a1dfa0c6 XL |
1508 | bug!("worker thread panicked"); |
1509 | } | |
dfeec247 XL |
1510 | Message::Done { result: Err(Some(WorkerFatalError)), worker_id: _ } => { |
1511 | return Err(()); | |
a1dfa0c6 | 1512 | } |
dfeec247 | 1513 | Message::CodegenItem => bug!("the coordinator should not receive codegen requests"), |
a1dfa0c6 XL |
1514 | } |
1515 | } | |
1516 | ||
1b1a35ee XL |
1517 | let needs_link = mem::take(&mut needs_link); |
1518 | if !needs_link.is_empty() { | |
1519 | assert!(compiled_modules.is_empty()); | |
1520 | let diag_handler = cgcx.create_diag_handler(); | |
1521 | let module = B::run_link(&cgcx, &diag_handler, needs_link).map_err(|_| ())?; | |
1522 | let module = unsafe { | |
1523 | B::codegen(&cgcx, &diag_handler, module, cgcx.config(ModuleKind::Regular)) | |
1524 | .map_err(|_| ())? | |
1525 | }; | |
1526 | compiled_modules.push(module); | |
1527 | } | |
1528 | ||
dfeec247 XL |
1529 | // Drop to print timings |
1530 | drop(llvm_start_time); | |
a1dfa0c6 XL |
1531 | |
1532 | // Regardless of what order these modules completed in, report them to | |
1533 | // the backend in the same order every time to ensure that we're handing | |
1534 | // out deterministic results. | |
1535 | compiled_modules.sort_by(|a, b| a.name.cmp(&b.name)); | |
1536 | ||
a1dfa0c6 XL |
1537 | Ok(CompiledModules { |
1538 | modules: compiled_modules, | |
a1dfa0c6 XL |
1539 | allocator_module: compiled_allocator_module, |
1540 | }) | |
1541 | }); | |
1542 | ||
1543 | // A heuristic that determines if we have enough LLVM WorkItems in the | |
1544 | // queue so that the main thread can do LLVM work instead of codegen | |
6a06907d XL |
1545 | fn queue_full_enough(items_in_queue: usize, workers_running: usize) -> bool { |
1546 | // This heuristic scales ahead-of-time codegen according to available | |
1547 | // concurrency, as measured by `workers_running`. The idea is that the | |
1548 | // more concurrency we have available, the more demand there will be for | |
1549 | // work items, and the fuller the queue should be kept to meet demand. | |
1550 | // An important property of this approach is that we codegen ahead of | |
1551 | // time only as much as necessary, so as to keep fewer LLVM modules in | |
1552 | // memory at once, thereby reducing memory consumption. | |
1553 | // | |
1554 | // When the number of workers running is less than the max concurrency | |
1555 | // available to us, this heuristic can cause us to instruct the main | |
1556 | // thread to work on an LLVM item (that is, tell it to "LLVM") instead | |
1557 | // of codegen, even though it seems like it *should* be codegenning so | |
1558 | // that we can create more work items and spawn more LLVM workers. | |
1559 | // | |
1560 | // But this is not a problem. When the main thread is told to LLVM, | |
1561 | // according to this heuristic and how work is scheduled, there is | |
1562 | // always at least one item in the queue, and therefore at least one | |
1563 | // pending jobserver token request. If there *is* more concurrency | |
1564 | // available, we will immediately receive a token, which will upgrade | |
1565 | // the main thread's LLVM worker to a real one (conceptually), and free | |
1566 | // up the main thread to codegen if necessary. On the other hand, if | |
1567 | // there isn't more concurrency, then the main thread working on an LLVM | |
1568 | // item is appropriate, as long as the queue is full enough for demand. | |
1569 | // | |
1570 | // Speaking of which, how full should we keep the queue? Probably less | |
1571 | // full than you'd think. A lot has to go wrong for the queue not to be | |
1572 | // full enough and for that to have a negative effect on compile times. | |
1573 | // | |
1574 | // Workers are unlikely to finish at exactly the same time, so when one | |
1575 | // finishes and takes another work item off the queue, we often have | |
1576 | // ample time to codegen at that point before the next worker finishes. | |
1577 | // But suppose that codegen takes so long that the workers exhaust the | |
1578 | // queue, and we have one or more workers that have nothing to work on. | |
1579 | // Well, it might not be so bad. Of all the LLVM modules we create and | |
1580 | // optimize, one has to finish last. It's not necessarily the case that | |
1581 | // by losing some concurrency for a moment, we delay the point at which | |
1582 | // that last LLVM module is finished and the rest of compilation can | |
1583 | // proceed. Also, when we can't take advantage of some concurrency, we | |
1584 | // give tokens back to the job server. That enables some other rustc to | |
1585 | // potentially make use of the available concurrency. That could even | |
1586 | // *decrease* overall compile time if we're lucky. But yes, if no other | |
1587 | // rustc can make use of the concurrency, then we've squandered it. | |
1588 | // | |
1589 | // However, keeping the queue full is also beneficial when we have a | |
1590 | // surge in available concurrency. Then items can be taken from the | |
1591 | // queue immediately, without having to wait for codegen. | |
1592 | // | |
1593 | // So, the heuristic below tries to keep one item in the queue for every | |
1594 | // four running workers. Based on limited benchmarking, this appears to | |
1595 | // be more than sufficient to avoid increasing compilation times. | |
1596 | let quarter_of_workers = workers_running - 3 * workers_running / 4; | |
1597 | items_in_queue > 0 && items_in_queue >= quarter_of_workers | |
a1dfa0c6 XL |
1598 | } |
1599 | ||
dfeec247 XL |
1600 | fn maybe_start_llvm_timer<'a>( |
1601 | prof: &'a SelfProfilerRef, | |
1602 | config: &ModuleConfig, | |
1603 | llvm_start_time: &mut Option<VerboseTimingGuard<'a>>, | |
1604 | ) { | |
1605 | if config.time_module && llvm_start_time.is_none() { | |
74b04a01 | 1606 | *llvm_start_time = Some(prof.extra_verbose_generic_activity("LLVM_passes", "crate")); |
a1dfa0c6 XL |
1607 | } |
1608 | } | |
1609 | } | |
1610 | ||
dfeec247 XL |
1611 | /// `FatalError` is explicitly not `Send`. |
1612 | #[must_use] | |
1613 | pub struct WorkerFatalError; | |
a1dfa0c6 | 1614 | |
dfeec247 | 1615 | fn spawn_work<B: ExtraBackendMethods>(cgcx: CodegenContext<B>, work: WorkItem<B>) { |
3c0e092e XL |
1616 | B::spawn_named_thread(cgcx.time_trace, work.short_description(), move || { |
1617 | // Set up a destructor which will fire off a message that we're done as | |
1618 | // we exit. | |
1619 | struct Bomb<B: ExtraBackendMethods> { | |
1620 | coordinator_send: Sender<Box<dyn Any + Send>>, | |
1621 | result: Option<Result<WorkItemResult<B>, FatalError>>, | |
1622 | worker_id: usize, | |
1623 | } | |
1624 | impl<B: ExtraBackendMethods> Drop for Bomb<B> { | |
1625 | fn drop(&mut self) { | |
1626 | let worker_id = self.worker_id; | |
1627 | let msg = match self.result.take() { | |
1628 | Some(Ok(WorkItemResult::Compiled(m))) => { | |
1629 | Message::Done::<B> { result: Ok(m), worker_id } | |
1630 | } | |
1631 | Some(Ok(WorkItemResult::NeedsLink(m))) => { | |
1632 | Message::NeedsLink::<B> { module: m, worker_id } | |
1633 | } | |
1634 | Some(Ok(WorkItemResult::NeedsFatLTO(m))) => { | |
1635 | Message::NeedsFatLTO::<B> { result: m, worker_id } | |
1636 | } | |
1637 | Some(Ok(WorkItemResult::NeedsThinLTO(name, thin_buffer))) => { | |
1638 | Message::NeedsThinLTO::<B> { name, thin_buffer, worker_id } | |
1639 | } | |
1640 | Some(Err(FatalError)) => { | |
1641 | Message::Done::<B> { result: Err(Some(WorkerFatalError)), worker_id } | |
1642 | } | |
1643 | None => Message::Done::<B> { result: Err(None), worker_id }, | |
1644 | }; | |
1645 | drop(self.coordinator_send.send(Box::new(msg))); | |
a1dfa0c6 | 1646 | } |
3c0e092e | 1647 | } |
a1dfa0c6 | 1648 | |
3c0e092e XL |
1649 | let mut bomb = Bomb::<B> { |
1650 | coordinator_send: cgcx.coordinator_send.clone(), | |
1651 | result: None, | |
1652 | worker_id: cgcx.worker, | |
1653 | }; | |
a1dfa0c6 | 1654 | |
3c0e092e XL |
1655 | // Execute the work itself, and if it finishes successfully then flag |
1656 | // ourselves as a success as well. | |
1657 | // | |
1658 | // Note that we ignore any `FatalError` coming out of `execute_work_item`, | |
1659 | // as a diagnostic was already sent off to the main thread - just | |
1660 | // surface that there was an error in this worker. | |
1661 | bomb.result = { | |
1662 | let _prof_timer = work.start_profiling(&cgcx); | |
1663 | Some(execute_work_item(&cgcx, work)) | |
1664 | }; | |
1665 | }) | |
1666 | .expect("failed to spawn thread"); | |
a1dfa0c6 XL |
1667 | } |
1668 | ||
a1dfa0c6 XL |
1669 | enum SharedEmitterMessage { |
1670 | Diagnostic(Diagnostic), | |
f035d41b | 1671 | InlineAsmError(u32, String, Level, Option<(String, Vec<InnerSpan>)>), |
a1dfa0c6 XL |
1672 | AbortIfErrors, |
1673 | Fatal(String), | |
1674 | } | |
1675 | ||
1676 | #[derive(Clone)] | |
1677 | pub struct SharedEmitter { | |
1678 | sender: Sender<SharedEmitterMessage>, | |
1679 | } | |
1680 | ||
1681 | pub struct SharedEmitterMain { | |
1682 | receiver: Receiver<SharedEmitterMessage>, | |
1683 | } | |
1684 | ||
1685 | impl SharedEmitter { | |
1686 | pub fn new() -> (SharedEmitter, SharedEmitterMain) { | |
1687 | let (sender, receiver) = channel(); | |
1688 | ||
1689 | (SharedEmitter { sender }, SharedEmitterMain { receiver }) | |
1690 | } | |
1691 | ||
f9f354fc XL |
1692 | pub fn inline_asm_error( |
1693 | &self, | |
1694 | cookie: u32, | |
1695 | msg: String, | |
f035d41b | 1696 | level: Level, |
f9f354fc XL |
1697 | source: Option<(String, Vec<InnerSpan>)>, |
1698 | ) { | |
f035d41b | 1699 | drop(self.sender.send(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source))); |
a1dfa0c6 XL |
1700 | } |
1701 | ||
1702 | pub fn fatal(&self, msg: &str) { | |
1703 | drop(self.sender.send(SharedEmitterMessage::Fatal(msg.to_string()))); | |
1704 | } | |
1705 | } | |
1706 | ||
1707 | impl Emitter for SharedEmitter { | |
e74abb32 | 1708 | fn emit_diagnostic(&mut self, diag: &rustc_errors::Diagnostic) { |
a1dfa0c6 | 1709 | drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic { |
e74abb32 XL |
1710 | msg: diag.message(), |
1711 | code: diag.code.clone(), | |
1712 | lvl: diag.level, | |
a1dfa0c6 | 1713 | }))); |
e74abb32 | 1714 | for child in &diag.children { |
a1dfa0c6 XL |
1715 | drop(self.sender.send(SharedEmitterMessage::Diagnostic(Diagnostic { |
1716 | msg: child.message(), | |
1717 | code: None, | |
1718 | lvl: child.level, | |
1719 | }))); | |
1720 | } | |
1721 | drop(self.sender.send(SharedEmitterMessage::AbortIfErrors)); | |
1722 | } | |
60c5eb7d | 1723 | fn source_map(&self) -> Option<&Lrc<SourceMap>> { |
e74abb32 XL |
1724 | None |
1725 | } | |
a1dfa0c6 XL |
1726 | } |
1727 | ||
1728 | impl SharedEmitterMain { | |
1729 | pub fn check(&self, sess: &Session, blocking: bool) { | |
1730 | loop { | |
1731 | let message = if blocking { | |
1732 | match self.receiver.recv() { | |
1733 | Ok(message) => Ok(message), | |
1734 | Err(_) => Err(()), | |
1735 | } | |
1736 | } else { | |
1737 | match self.receiver.try_recv() { | |
1738 | Ok(message) => Ok(message), | |
1739 | Err(_) => Err(()), | |
1740 | } | |
1741 | }; | |
1742 | ||
1743 | match message { | |
1744 | Ok(SharedEmitterMessage::Diagnostic(diag)) => { | |
1745 | let handler = sess.diagnostic(); | |
e1599b0c XL |
1746 | let mut d = rustc_errors::Diagnostic::new(diag.lvl, &diag.msg); |
1747 | if let Some(code) = diag.code { | |
1748 | d.code(code); | |
a1dfa0c6 | 1749 | } |
e1599b0c | 1750 | handler.emit_diagnostic(&d); |
a1dfa0c6 | 1751 | } |
f035d41b | 1752 | Ok(SharedEmitterMessage::InlineAsmError(cookie, msg, level, source)) => { |
f9f354fc XL |
1753 | let msg = msg.strip_prefix("error: ").unwrap_or(&msg); |
1754 | ||
f035d41b | 1755 | let mut err = match level { |
3c0e092e | 1756 | Level::Error { lint: false } => sess.struct_err(&msg), |
f035d41b XL |
1757 | Level::Warning => sess.struct_warn(&msg), |
1758 | Level::Note => sess.struct_note_without_error(&msg), | |
1759 | _ => bug!("Invalid inline asm diagnostic level"), | |
1760 | }; | |
1761 | ||
f9f354fc | 1762 | // If the cookie is 0 then we don't have span information. |
f035d41b | 1763 | if cookie != 0 { |
f9f354fc XL |
1764 | let pos = BytePos::from_u32(cookie); |
1765 | let span = Span::with_root_ctxt(pos, pos); | |
f035d41b | 1766 | err.set_span(span); |
f9f354fc XL |
1767 | }; |
1768 | ||
1769 | // Point to the generated assembly if it is available. | |
1770 | if let Some((buffer, spans)) = source { | |
1771 | let source = sess | |
1772 | .source_map() | |
1773 | .new_source_file(FileName::inline_asm_source_code(&buffer), buffer); | |
1774 | let source_span = Span::with_root_ctxt(source.start_pos, source.end_pos); | |
1775 | let spans: Vec<_> = | |
1776 | spans.iter().map(|sp| source_span.from_inner(*sp)).collect(); | |
1777 | err.span_note(spans, "instantiated into assembly here"); | |
1778 | } | |
1779 | ||
1780 | err.emit(); | |
a1dfa0c6 XL |
1781 | } |
1782 | Ok(SharedEmitterMessage::AbortIfErrors) => { | |
1783 | sess.abort_if_errors(); | |
1784 | } | |
1785 | Ok(SharedEmitterMessage::Fatal(msg)) => { | |
1786 | sess.fatal(&msg); | |
1787 | } | |
1788 | Err(_) => { | |
1789 | break; | |
1790 | } | |
1791 | } | |
a1dfa0c6 XL |
1792 | } |
1793 | } | |
1794 | } | |
1795 | ||
1796 | pub struct OngoingCodegen<B: ExtraBackendMethods> { | |
1797 | pub backend: B, | |
a1dfa0c6 | 1798 | pub metadata: EncodedMetadata, |
a2a8927a | 1799 | pub metadata_module: Option<CompiledModule>, |
a1dfa0c6 | 1800 | pub crate_info: CrateInfo, |
a1dfa0c6 XL |
1801 | pub coordinator_send: Sender<Box<dyn Any + Send>>, |
1802 | pub codegen_worker_receive: Receiver<Message<B>>, | |
1803 | pub shared_emitter_main: SharedEmitterMain, | |
1804 | pub future: thread::JoinHandle<Result<CompiledModules, ()>>, | |
1805 | pub output_filenames: Arc<OutputFilenames>, | |
1806 | } | |
1807 | ||
1808 | impl<B: ExtraBackendMethods> OngoingCodegen<B> { | |
dfeec247 XL |
1809 | pub fn join(self, sess: &Session) -> (CodegenResults, FxHashMap<WorkProductId, WorkProduct>) { |
1810 | let _timer = sess.timer("finish_ongoing_codegen"); | |
1811 | ||
a1dfa0c6 | 1812 | self.shared_emitter_main.check(sess, true); |
dfeec247 XL |
1813 | let future = self.future; |
1814 | let compiled_modules = sess.time("join_worker_thread", || match future.join() { | |
a1dfa0c6 XL |
1815 | Ok(Ok(compiled_modules)) => compiled_modules, |
1816 | Ok(Err(())) => { | |
1817 | sess.abort_if_errors(); | |
1818 | panic!("expected abort due to worker thread errors") | |
dfeec247 | 1819 | } |
a1dfa0c6 XL |
1820 | Err(_) => { |
1821 | bug!("panic during codegen/LLVM phase"); | |
1822 | } | |
dfeec247 | 1823 | }); |
a1dfa0c6 | 1824 | |
60c5eb7d | 1825 | sess.cgu_reuse_tracker.check_expected_reuse(sess.diagnostic()); |
a1dfa0c6 XL |
1826 | |
1827 | sess.abort_if_errors(); | |
1828 | ||
a1dfa0c6 | 1829 | let work_products = |
dfeec247 XL |
1830 | copy_all_cgu_workproducts_to_incr_comp_cache_dir(sess, &compiled_modules); |
1831 | produce_final_output_artifacts(sess, &compiled_modules, &self.output_filenames); | |
a1dfa0c6 XL |
1832 | |
1833 | // FIXME: time_llvm_passes support - does this use a global context or | |
1834 | // something? | |
1835 | if sess.codegen_units() == 1 && sess.time_llvm_passes() { | |
1836 | self.backend.print_pass_timings() | |
1837 | } | |
1838 | ||
dfeec247 XL |
1839 | ( |
1840 | CodegenResults { | |
dfeec247 | 1841 | metadata: self.metadata, |
dfeec247 XL |
1842 | crate_info: self.crate_info, |
1843 | ||
1844 | modules: compiled_modules.modules, | |
1845 | allocator_module: compiled_modules.allocator_module, | |
a2a8927a | 1846 | metadata_module: self.metadata_module, |
dfeec247 XL |
1847 | }, |
1848 | work_products, | |
1849 | ) | |
a1dfa0c6 XL |
1850 | } |
1851 | ||
dc9dc135 XL |
1852 | pub fn submit_pre_codegened_module_to_llvm( |
1853 | &self, | |
1854 | tcx: TyCtxt<'_>, | |
1855 | module: ModuleCodegen<B::Module>, | |
1856 | ) { | |
a1dfa0c6 XL |
1857 | self.wait_for_signal_to_codegen_item(); |
1858 | self.check_for_errors(tcx.sess); | |
1859 | ||
48663c56 | 1860 | // These are generally cheap and won't throw off scheduling. |
a1dfa0c6 | 1861 | let cost = 0; |
e74abb32 | 1862 | submit_codegened_module_to_llvm(&self.backend, &self.coordinator_send, module, cost); |
a1dfa0c6 XL |
1863 | } |
1864 | ||
dc9dc135 | 1865 | pub fn codegen_finished(&self, tcx: TyCtxt<'_>) { |
a1dfa0c6 XL |
1866 | self.wait_for_signal_to_codegen_item(); |
1867 | self.check_for_errors(tcx.sess); | |
1868 | drop(self.coordinator_send.send(Box::new(Message::CodegenComplete::<B>))); | |
1869 | } | |
1870 | ||
9fa01778 | 1871 | /// Consumes this context indicating that codegen was entirely aborted, and |
a1dfa0c6 XL |
1872 | /// we need to exit as quickly as possible. |
1873 | /// | |
1874 | /// This method blocks the current thread until all worker threads have | |
1875 | /// finished, and all worker threads should have exited or be real close to | |
1876 | /// exiting at this point. | |
1877 | pub fn codegen_aborted(self) { | |
1878 | // Signal to the coordinator it should spawn no more work and start | |
1879 | // shutdown. | |
1880 | drop(self.coordinator_send.send(Box::new(Message::CodegenAborted::<B>))); | |
1881 | drop(self.future.join()); | |
1882 | } | |
1883 | ||
1884 | pub fn check_for_errors(&self, sess: &Session) { | |
1885 | self.shared_emitter_main.check(sess, false); | |
1886 | } | |
1887 | ||
1888 | pub fn wait_for_signal_to_codegen_item(&self) { | |
1889 | match self.codegen_worker_receive.recv() { | |
1890 | Ok(Message::CodegenItem) => { | |
1891 | // Nothing to do | |
1892 | } | |
1893 | Ok(_) => panic!("unexpected message"), | |
1894 | Err(_) => { | |
1895 | // One of the LLVM threads must have panicked, fall through so | |
1896 | // error handling can be reached. | |
1897 | } | |
1898 | } | |
1899 | } | |
1900 | } | |
1901 | ||
1902 | pub fn submit_codegened_module_to_llvm<B: ExtraBackendMethods>( | |
1903 | _backend: &B, | |
e74abb32 | 1904 | tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>, |
a1dfa0c6 | 1905 | module: ModuleCodegen<B::Module>, |
dc9dc135 | 1906 | cost: u64, |
a1dfa0c6 XL |
1907 | ) { |
1908 | let llvm_work_item = WorkItem::Optimize(module); | |
dfeec247 | 1909 | drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost }))); |
a1dfa0c6 XL |
1910 | } |
1911 | ||
1912 | pub fn submit_post_lto_module_to_llvm<B: ExtraBackendMethods>( | |
1913 | _backend: &B, | |
e74abb32 | 1914 | tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>, |
dc9dc135 | 1915 | module: CachedModuleCodegen, |
a1dfa0c6 XL |
1916 | ) { |
1917 | let llvm_work_item = WorkItem::CopyPostLtoArtifacts(module); | |
dfeec247 | 1918 | drop(tx_to_llvm_workers.send(Box::new(Message::CodegenDone::<B> { llvm_work_item, cost: 0 }))); |
a1dfa0c6 XL |
1919 | } |
1920 | ||
1921 | pub fn submit_pre_lto_module_to_llvm<B: ExtraBackendMethods>( | |
1922 | _backend: &B, | |
dc9dc135 | 1923 | tcx: TyCtxt<'_>, |
e74abb32 | 1924 | tx_to_llvm_workers: &Sender<Box<dyn Any + Send>>, |
dc9dc135 | 1925 | module: CachedModuleCodegen, |
a1dfa0c6 XL |
1926 | ) { |
1927 | let filename = pre_lto_bitcode_filename(&module.name); | |
1928 | let bc_path = in_incr_comp_dir_sess(tcx.sess, &filename); | |
dfeec247 XL |
1929 | let file = fs::File::open(&bc_path) |
1930 | .unwrap_or_else(|e| panic!("failed to open bitcode file `{}`: {}", bc_path.display(), e)); | |
a1dfa0c6 XL |
1931 | |
1932 | let mmap = unsafe { | |
cdc7bbd5 | 1933 | Mmap::map(file).unwrap_or_else(|e| { |
a1dfa0c6 XL |
1934 | panic!("failed to mmap bitcode file `{}`: {}", bc_path.display(), e) |
1935 | }) | |
1936 | }; | |
1937 | // Schedule the module to be loaded | |
e74abb32 | 1938 | drop(tx_to_llvm_workers.send(Box::new(Message::AddImportOnlyModule::<B> { |
a1dfa0c6 XL |
1939 | module_data: SerializedModule::FromUncompressedFile(mmap), |
1940 | work_product: module.source, | |
1941 | }))); | |
1942 | } | |
1943 | ||
1944 | pub fn pre_lto_bitcode_filename(module_name: &str) -> String { | |
9fa01778 | 1945 | format!("{}.{}", module_name, PRE_LTO_BC_EXT) |
a1dfa0c6 XL |
1946 | } |
1947 | ||
dc9dc135 | 1948 | fn msvc_imps_needed(tcx: TyCtxt<'_>) -> bool { |
a1dfa0c6 XL |
1949 | // This should never be true (because it's not supported). If it is true, |
1950 | // something is wrong with commandline arg validation. | |
dfeec247 XL |
1951 | assert!( |
1952 | !(tcx.sess.opts.cg.linker_plugin_lto.enabled() | |
29967ef6 | 1953 | && tcx.sess.target.is_like_windows |
dfeec247 XL |
1954 | && tcx.sess.opts.cg.prefer_dynamic) |
1955 | ); | |
a1dfa0c6 | 1956 | |
29967ef6 | 1957 | tcx.sess.target.is_like_windows && |
f9f354fc | 1958 | tcx.sess.crate_types().iter().any(|ct| *ct == CrateType::Rlib) && |
a1dfa0c6 XL |
1959 | // ThinLTO can't handle this workaround in all cases, so we don't |
1960 | // emit the `__imp_` symbols. Instead we make them unnecessary by disallowing | |
9fa01778 XL |
1961 | // dynamic linking when linker plugin LTO is enabled. |
1962 | !tcx.sess.opts.cg.linker_plugin_lto.enabled() | |
a1dfa0c6 | 1963 | } |