1 // Copyright 2013-2015 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
12 use back::link::{get_linker, remove};
13 use rustc_incremental::{save_trans_partition, in_incr_comp_dir};
14 use session::config::{OutputFilenames, OutputTypes, Passes, SomePasses, AllPasses};
16 use session::config::{self, OutputType};
18 use llvm::{ModuleRef, TargetMachineRef, PassManagerRef, DiagnosticInfoRef, ContextRef};
19 use llvm::SMDiagnosticRef;
20 use {CrateTranslation, ModuleLlvm, ModuleSource, ModuleTranslation};
21 use util::common::time;
22 use util::common::path2cstr;
23 use util::fs::link_or_copy;
24 use errors::{self, Handler, Level, DiagnosticBuilder};
25 use errors::emitter::Emitter;
26 use syntax_pos::MultiSpan;
27 use context::{is_pie_binary, get_reloc_model};
29 use std::ffi::CString;
31 use std::path::{Path, PathBuf};
33 use std::sync::{Arc, Mutex};
34 use std::sync::mpsc::channel;
36 use libc::{c_uint, c_void};
38 pub const RELOC_MODEL_ARGS : [(&'static str, llvm::RelocMode); 4] = [
39 ("pic", llvm::RelocMode::PIC),
40 ("static", llvm::RelocMode::Static),
41 ("default", llvm::RelocMode::Default),
42 ("dynamic-no-pic", llvm::RelocMode::DynamicNoPic),
45 pub const CODE_GEN_MODEL_ARGS : [(&'static str, llvm::CodeModel); 5] = [
46 ("default", llvm::CodeModel::Default),
47 ("small", llvm::CodeModel::Small),
48 ("kernel", llvm::CodeModel::Kernel),
49 ("medium", llvm::CodeModel::Medium),
50 ("large", llvm::CodeModel::Large),
53 pub fn llvm_err(handler: &errors::Handler, msg: String) -> ! {
54 match llvm::last_error() {
55 Some(err) => panic!(handler.fatal(&format!("{}: {}", msg, err))),
56 None => panic!(handler.fatal(&msg)),
60 pub fn write_output_file(
61 handler: &errors::Handler,
62 target: llvm::TargetMachineRef,
63 pm: llvm::PassManagerRef,
66 file_type: llvm::FileType) {
68 let output_c = path2cstr(output);
69 let result = llvm::LLVMRustWriteOutputFile(
70 target, pm, m, output_c.as_ptr(), file_type);
71 if result.into_result().is_err() {
72 llvm_err(handler, format!("could not write output to {}", output.display()));
84 // We use an Arc instead of just returning a list of diagnostics from the
85 // child thread because we need to make sure that the messages are seen even
86 // if the child thread panics (for example, when `fatal` is called).
88 struct SharedEmitter {
89 buffer: Arc<Mutex<Vec<Diagnostic>>>,
93 fn new() -> SharedEmitter {
95 buffer: Arc::new(Mutex::new(Vec::new())),
99 fn dump(&mut self, handler: &Handler) {
100 let mut buffer = self.buffer.lock().unwrap();
101 for diag in &*buffer {
104 handler.emit_with_code(&MultiSpan::new(),
110 handler.emit(&MultiSpan::new(),
120 impl Emitter for SharedEmitter {
121 fn emit(&mut self, db: &DiagnosticBuilder) {
122 self.buffer.lock().unwrap().push(Diagnostic {
123 msg: db.message.to_string(),
124 code: db.code.clone(),
127 for child in &db.children {
128 self.buffer.lock().unwrap().push(Diagnostic {
129 msg: child.message.to_string(),
137 // On android, we by default compile for armv7 processors. This enables
138 // things like double word CAS instructions (rather than emulating them)
139 // which are *far* more efficient. This is obviously undesirable in some
140 // cases, so if any sort of target feature is specified we don't append v7
141 // to the feature list.
143 // On iOS only armv7 and newer are supported. So it is useful to
144 // get all hardware potential via VFP3 (hardware floating point)
145 // and NEON (SIMD) instructions supported by LLVM.
146 // Note that without those flags various linking errors might
147 // arise as some of intrinsics are converted into function calls
148 // and nobody provides implementations those functions
149 fn target_feature(sess: &Session) -> String {
150 let rustc_features = [
153 let requested_features = sess.opts.cg.target_feature.split(',');
154 let llvm_features = requested_features.filter(|f| {
155 !rustc_features.iter().any(|s| f.contains(s))
158 sess.target.target.options.features,
159 llvm_features.collect::<Vec<_>>().join(","))
162 fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel {
164 config::OptLevel::No => llvm::CodeGenOptLevel::None,
165 config::OptLevel::Less => llvm::CodeGenOptLevel::Less,
166 config::OptLevel::Default => llvm::CodeGenOptLevel::Default,
167 config::OptLevel::Aggressive => llvm::CodeGenOptLevel::Aggressive,
168 _ => llvm::CodeGenOptLevel::Default,
172 fn get_llvm_opt_size(optimize: config::OptLevel) -> llvm::CodeGenOptSize {
174 config::OptLevel::Size => llvm::CodeGenOptSizeDefault,
175 config::OptLevel::SizeMin => llvm::CodeGenOptSizeAggressive,
176 _ => llvm::CodeGenOptSizeNone,
180 pub fn create_target_machine(sess: &Session) -> TargetMachineRef {
181 let reloc_model = get_reloc_model(sess);
183 let opt_level = get_llvm_opt_level(sess.opts.optimize);
184 let use_softfp = sess.opts.cg.soft_float;
186 let ffunction_sections = sess.target.target.options.function_sections;
187 let fdata_sections = ffunction_sections;
189 let code_model_arg = match sess.opts.cg.code_model {
190 Some(ref s) => &s[..],
191 None => &sess.target.target.options.code_model[..],
194 let code_model = match CODE_GEN_MODEL_ARGS.iter().find(
195 |&&arg| arg.0 == code_model_arg) {
198 sess.err(&format!("{:?} is not a valid code model",
202 sess.abort_if_errors();
207 let triple = &sess.target.target.llvm_target;
210 let triple = CString::new(triple.as_bytes()).unwrap();
211 let cpu = match sess.opts.cg.target_cpu {
213 None => &*sess.target.target.options.cpu
215 let cpu = CString::new(cpu.as_bytes()).unwrap();
216 let features = CString::new(target_feature(sess).as_bytes()).unwrap();
217 llvm::LLVMRustCreateTargetMachine(
218 triple.as_ptr(), cpu.as_ptr(), features.as_ptr(),
230 llvm_err(sess.diagnostic(),
231 format!("Could not create LLVM TargetMachine for triple: {}",
232 triple).to_string());
239 /// Module-specific configuration for `optimize_and_codegen`.
241 pub struct ModuleConfig {
242 /// LLVM TargetMachine to use for codegen.
243 tm: TargetMachineRef,
244 /// Names of additional optimization passes to run.
246 /// Some(level) to optimize at a certain level, or None to run
247 /// absolutely no optimizations (used for the metadata module).
248 opt_level: Option<llvm::CodeGenOptLevel>,
250 /// Some(level) to optimize binary size, or None to not affect program size.
251 opt_size: Option<llvm::CodeGenOptSize>,
253 // Flags indicating which outputs to produce.
254 emit_no_opt_bc: bool,
260 // Miscellaneous flags. These are mostly copied from command-line
263 no_prepopulate_passes: bool,
266 vectorize_loop: bool,
268 merge_functions: bool,
269 inline_threshold: Option<usize>,
270 // Instead of creating an object file by doing LLVM codegen, just
271 // make the object file bitcode. Provides easy compatibility with
272 // emscripten's ecc compiler, when used as the linker.
273 obj_is_bitcode: bool,
276 unsafe impl Send for ModuleConfig { }
279 fn new(tm: TargetMachineRef, passes: Vec<String>) -> ModuleConfig {
286 emit_no_opt_bc: false,
292 obj_is_bitcode: false,
295 no_prepopulate_passes: false,
298 vectorize_loop: false,
299 vectorize_slp: false,
300 merge_functions: false,
301 inline_threshold: None
305 fn set_flags(&mut self, sess: &Session, trans: &CrateTranslation) {
306 self.no_verify = sess.no_verify();
307 self.no_prepopulate_passes = sess.opts.cg.no_prepopulate_passes;
308 self.no_builtins = trans.no_builtins;
309 self.time_passes = sess.time_passes();
310 self.inline_threshold = sess.opts.cg.inline_threshold;
311 self.obj_is_bitcode = sess.target.target.options.obj_is_bitcode;
313 // Copy what clang does by turning on loop vectorization at O2 and
314 // slp vectorization at O3. Otherwise configure other optimization aspects
315 // of this pass manager builder.
316 self.vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
317 (sess.opts.optimize == config::OptLevel::Default ||
318 sess.opts.optimize == config::OptLevel::Aggressive);
319 self.vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
320 sess.opts.optimize == config::OptLevel::Aggressive;
322 self.merge_functions = sess.opts.optimize == config::OptLevel::Default ||
323 sess.opts.optimize == config::OptLevel::Aggressive;
327 /// Additional resources used by optimize_and_codegen (not module specific)
328 struct CodegenContext<'a> {
329 // Extra resources used for LTO: (sess, reachable). This will be `None`
330 // when running in a worker thread.
331 lto_ctxt: Option<(&'a Session, &'a [String])>,
332 // Handler to use for diagnostics produced during codegen.
333 handler: &'a Handler,
334 // LLVM passes added by plugins.
335 plugin_passes: Vec<String>,
336 // LLVM optimizations for which we want to print remarks.
338 // Worker thread number
340 // The incremental compilation session directory, or None if we are not
341 // compiling incrementally
342 incr_comp_session_dir: Option<PathBuf>
345 impl<'a> CodegenContext<'a> {
346 fn new_with_session(sess: &'a Session, exported_symbols: &'a [String]) -> CodegenContext<'a> {
348 lto_ctxt: Some((sess, exported_symbols)),
349 handler: sess.diagnostic(),
350 plugin_passes: sess.plugin_llvm_passes.borrow().clone(),
351 remark: sess.opts.cg.remark.clone(),
353 incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone())
358 struct HandlerFreeVars<'a> {
360 cgcx: &'a CodegenContext<'a>,
363 unsafe extern "C" fn report_inline_asm<'a, 'b>(cgcx: &'a CodegenContext<'a>,
366 use syntax_pos::ExpnId;
368 match cgcx.lto_ctxt {
370 sess.codemap().with_expn_info(ExpnId::from_u32(cookie), |info| match info {
371 Some(ei) => sess.span_err(ei.call_site, msg),
372 None => sess.err(msg),
377 cgcx.handler.struct_err(msg)
378 .note("build without -C codegen-units for more exact errors")
384 unsafe extern "C" fn inline_asm_handler(diag: SMDiagnosticRef,
387 let HandlerFreeVars { cgcx, .. } = *(user as *const HandlerFreeVars);
389 let msg = llvm::build_string(|s| llvm::LLVMRustWriteSMDiagnosticToString(diag, s))
390 .expect("non-UTF8 SMDiagnostic");
392 report_inline_asm(cgcx, &msg[..], cookie);
395 unsafe extern "C" fn diagnostic_handler(info: DiagnosticInfoRef, user: *mut c_void) {
396 let HandlerFreeVars { llcx, cgcx } = *(user as *const HandlerFreeVars);
398 match llvm::diagnostic::Diagnostic::unpack(info) {
399 llvm::diagnostic::InlineAsm(inline) => {
400 report_inline_asm(cgcx,
401 &llvm::twine_to_string(inline.message),
405 llvm::diagnostic::Optimization(opt) => {
406 let enabled = match cgcx.remark {
408 SomePasses(ref v) => v.iter().any(|s| *s == opt.pass_name),
412 let loc = llvm::debug_loc_to_string(llcx, opt.debug_loc);
413 cgcx.handler.note_without_error(&format!("optimization {} for {} at {}: {}",
416 if loc.is_empty() { "[unknown]" } else { &*loc },
425 // Unsafe due to LLVM calls.
426 unsafe fn optimize_and_codegen(cgcx: &CodegenContext,
427 mtrans: ModuleTranslation,
429 config: ModuleConfig,
430 output_names: OutputFilenames) {
431 let llmod = mllvm.llmod;
432 let llcx = mllvm.llcx;
435 // llcx doesn't outlive this function, so we can put this on the stack.
436 let fv = HandlerFreeVars {
440 let fv = &fv as *const HandlerFreeVars as *mut c_void;
442 llvm::LLVMRustSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, fv);
443 llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, fv);
445 let module_name = Some(&mtrans.name[..]);
447 if config.emit_no_opt_bc {
448 let out = output_names.temp_path_ext("no-opt.bc", module_name);
449 let out = path2cstr(&out);
450 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
453 if config.opt_level.is_some() {
454 // Create the two optimizing pass managers. These mirror what clang
455 // does, and are by populated by LLVM's default PassManagerBuilder.
456 // Each manager has a different set of passes, but they also share
457 // some common passes.
458 let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
459 let mpm = llvm::LLVMCreatePassManager();
461 // If we're verifying or linting, add them to the function pass
463 let addpass = |pass_name: &str| {
464 let pass_name = CString::new(pass_name).unwrap();
465 let pass = llvm::LLVMRustFindAndCreatePass(pass_name.as_ptr());
469 let pass_manager = match llvm::LLVMRustPassKind(pass) {
470 llvm::PassKind::Function => fpm,
471 llvm::PassKind::Module => mpm,
472 llvm::PassKind::Other => {
473 cgcx.handler.err("Encountered LLVM pass kind we can't handle");
477 llvm::LLVMRustAddPass(pass_manager, pass);
481 if !config.no_verify { assert!(addpass("verify")); }
482 if !config.no_prepopulate_passes {
483 llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod);
484 llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod);
485 with_llvm_pmb(llmod, &config, &mut |b| {
486 llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(b, fpm);
487 llvm::LLVMPassManagerBuilderPopulateModulePassManager(b, mpm);
491 for pass in &config.passes {
493 cgcx.handler.warn(&format!("unknown pass `{}`, ignoring",
498 for pass in &cgcx.plugin_passes {
500 cgcx.handler.err(&format!("a plugin asked for LLVM pass \
501 `{}` but LLVM does not \
502 recognize it", pass));
506 cgcx.handler.abort_if_errors();
508 // Finally, run the actual optimization passes
509 time(config.time_passes, &format!("llvm function passes [{}]", cgcx.worker), ||
510 llvm::LLVMRustRunFunctionPassManager(fpm, llmod));
511 time(config.time_passes, &format!("llvm module passes [{}]", cgcx.worker), ||
512 llvm::LLVMRunPassManager(mpm, llmod));
514 // Deallocate managers that we're now done with
515 llvm::LLVMDisposePassManager(fpm);
516 llvm::LLVMDisposePassManager(mpm);
518 match cgcx.lto_ctxt {
519 Some((sess, exported_symbols)) if sess.lto() => {
520 time(sess.time_passes(), "all lto passes", || {
521 let temp_no_opt_bc_filename =
522 output_names.temp_path_ext("no-opt.lto.bc", module_name);
528 &temp_no_opt_bc_filename);
530 if config.emit_lto_bc {
531 let out = output_names.temp_path_ext("lto.bc", module_name);
532 let out = path2cstr(&out);
533 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
540 // A codegen-specific pass manager is used to generate object
541 // files for an LLVM module.
543 // Apparently each of these pass managers is a one-shot kind of
544 // thing, so we create a new one for each type of output. The
545 // pass manager passed to the closure should be ensured to not
546 // escape the closure itself, and the manager should only be
548 unsafe fn with_codegen<F>(tm: TargetMachineRef,
552 F: FnOnce(PassManagerRef),
554 let cpm = llvm::LLVMCreatePassManager();
555 llvm::LLVMRustAddAnalysisPasses(tm, cpm, llmod);
556 llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
560 // Change what we write and cleanup based on whether obj files are
561 // just llvm bitcode. In that case write bitcode, and possibly
562 // delete the bitcode if it wasn't requested. Don't generate the
563 // machine code, instead copy the .o file from the .bc
564 let write_bc = config.emit_bc || config.obj_is_bitcode;
565 let rm_bc = !config.emit_bc && config.obj_is_bitcode;
566 let write_obj = config.emit_obj && !config.obj_is_bitcode;
567 let copy_bc_to_obj = config.emit_obj && config.obj_is_bitcode;
569 let bc_out = output_names.temp_path(OutputType::Bitcode, module_name);
570 let obj_out = output_names.temp_path(OutputType::Object, module_name);
573 let bc_out_c = path2cstr(&bc_out);
574 llvm::LLVMWriteBitcodeToFile(llmod, bc_out_c.as_ptr());
577 time(config.time_passes, &format!("codegen passes [{}]", cgcx.worker), || {
579 let out = output_names.temp_path(OutputType::LlvmAssembly, module_name);
580 let out = path2cstr(&out);
581 with_codegen(tm, llmod, config.no_builtins, |cpm| {
582 llvm::LLVMRustPrintModule(cpm, llmod, out.as_ptr());
583 llvm::LLVMDisposePassManager(cpm);
588 let path = output_names.temp_path(OutputType::Assembly, module_name);
590 // We can't use the same module for asm and binary output, because that triggers
591 // various errors like invalid IR or broken binaries, so we might have to clone the
592 // module to produce the asm output
593 let llmod = if config.emit_obj {
594 llvm::LLVMCloneModule(llmod)
598 with_codegen(tm, llmod, config.no_builtins, |cpm| {
599 write_output_file(cgcx.handler, tm, cpm, llmod, &path,
600 llvm::FileType::AssemblyFile);
603 llvm::LLVMDisposeModule(llmod);
608 with_codegen(tm, llmod, config.no_builtins, |cpm| {
609 write_output_file(cgcx.handler, tm, cpm, llmod, &obj_out,
610 llvm::FileType::ObjectFile);
616 debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out);
617 if let Err(e) = link_or_copy(&bc_out, &obj_out) {
618 cgcx.handler.err(&format!("failed to copy bitcode to object file: {}", e));
623 debug!("removing_bitcode {:?}", bc_out);
624 if let Err(e) = fs::remove_file(&bc_out) {
625 cgcx.handler.err(&format!("failed to remove bitcode: {}", e));
629 llvm::LLVMRustDisposeTargetMachine(tm);
633 pub fn cleanup_llvm(trans: &CrateTranslation) {
634 for module in trans.modules.iter() {
636 match module.source {
637 ModuleSource::Translated(llvm) => {
638 llvm::LLVMDisposeModule(llvm.llmod);
639 llvm::LLVMContextDispose(llvm.llcx);
641 ModuleSource::Preexisting(_) => {
648 pub fn run_passes(sess: &Session,
649 trans: &CrateTranslation,
650 output_types: &OutputTypes,
651 crate_output: &OutputFilenames) {
652 // It's possible that we have `codegen_units > 1` but only one item in
653 // `trans.modules`. We could theoretically proceed and do LTO in that
654 // case, but it would be confusing to have the validity of
655 // `-Z lto -C codegen-units=2` depend on details of the crate being
656 // compiled, so we complain regardless.
657 if sess.lto() && sess.opts.cg.codegen_units > 1 {
658 // This case is impossible to handle because LTO expects to be able
659 // to combine the entire crate and all its dependencies into a
660 // single compilation unit, but each codegen unit is in a separate
661 // LLVM context, so they can't easily be combined.
662 sess.fatal("can't perform LTO when using multiple codegen units");
666 assert!(trans.modules.len() == sess.opts.cg.codegen_units ||
667 sess.opts.debugging_opts.incremental.is_some());
669 let tm = create_target_machine(sess);
671 // Figure out what we actually need to build.
673 let mut modules_config = ModuleConfig::new(tm, sess.opts.cg.passes.clone());
674 let mut metadata_config = ModuleConfig::new(tm, vec![]);
676 modules_config.opt_level = Some(get_llvm_opt_level(sess.opts.optimize));
677 modules_config.opt_size = Some(get_llvm_opt_size(sess.opts.optimize));
679 // Save all versions of the bytecode if we're saving our temporaries.
680 if sess.opts.cg.save_temps {
681 modules_config.emit_no_opt_bc = true;
682 modules_config.emit_bc = true;
683 modules_config.emit_lto_bc = true;
684 metadata_config.emit_bc = true;
687 // Emit bitcode files for the crate if we're emitting an rlib.
688 // Whenever an rlib is created, the bitcode is inserted into the
689 // archive in order to allow LTO against it.
690 let needs_crate_bitcode =
691 sess.crate_types.borrow().contains(&config::CrateTypeRlib) &&
692 sess.opts.output_types.contains_key(&OutputType::Exe);
693 let needs_crate_object =
694 sess.opts.output_types.contains_key(&OutputType::Exe);
695 if needs_crate_bitcode {
696 modules_config.emit_bc = true;
699 for output_type in output_types.keys() {
701 OutputType::Bitcode => { modules_config.emit_bc = true; },
702 OutputType::LlvmAssembly => { modules_config.emit_ir = true; },
703 OutputType::Assembly => {
704 modules_config.emit_asm = true;
705 // If we're not using the LLVM assembler, this function
706 // could be invoked specially with output_type_assembly, so
707 // in this case we still want the metadata object file.
708 if !sess.opts.output_types.contains_key(&OutputType::Assembly) {
709 metadata_config.emit_obj = true;
712 OutputType::Object => { modules_config.emit_obj = true; },
714 modules_config.emit_obj = true;
715 metadata_config.emit_obj = true;
717 OutputType::DepInfo => {}
721 modules_config.set_flags(sess, trans);
722 metadata_config.set_flags(sess, trans);
725 // Populate a buffer with a list of codegen threads. Items are processed in
726 // LIFO order, just because it's a tiny bit simpler that way. (The order
727 // doesn't actually matter.)
728 let mut work_items = Vec::with_capacity(1 + trans.modules.len());
731 let work = build_work_item(sess,
732 trans.metadata_module.clone(),
733 metadata_config.clone(),
734 crate_output.clone());
735 work_items.push(work);
738 for mtrans in trans.modules.iter() {
739 let work = build_work_item(sess,
741 modules_config.clone(),
742 crate_output.clone());
743 work_items.push(work);
746 if sess.opts.debugging_opts.incremental_info {
747 dump_incremental_data(&trans);
750 // Process the work items, optionally using worker threads.
751 // NOTE: This code is not really adapted to incremental compilation where
752 // the compiler decides the number of codegen units (and will
753 // potentially create hundreds of them).
754 let num_workers = work_items.len() - 1;
755 if num_workers == 1 {
756 run_work_singlethreaded(sess, &trans.exported_symbols, work_items);
758 run_work_multithreaded(sess, work_items, num_workers);
761 // If in incr. comp. mode, preserve the `.o` files for potential re-use
762 for mtrans in trans.modules.iter() {
763 let mut files = vec![];
765 if modules_config.emit_obj {
766 let path = crate_output.temp_path(OutputType::Object, Some(&mtrans.name));
767 files.push((OutputType::Object, path));
770 if modules_config.emit_bc {
771 let path = crate_output.temp_path(OutputType::Bitcode, Some(&mtrans.name));
772 files.push((OutputType::Bitcode, path));
775 save_trans_partition(sess, &mtrans.name, mtrans.symbol_name_hash, &files);
778 // All codegen is finished.
780 llvm::LLVMRustDisposeTargetMachine(tm);
783 // Produce final compile outputs.
784 let copy_gracefully = |from: &Path, to: &Path| {
785 if let Err(e) = fs::copy(from, to) {
786 sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
790 let copy_if_one_unit = |output_type: OutputType,
791 keep_numbered: bool| {
792 if trans.modules.len() == 1 {
793 // 1) Only one codegen unit. In this case it's no difficulty
794 // to copy `foo.0.x` to `foo.x`.
795 let module_name = Some(&(trans.modules[0].name)[..]);
796 let path = crate_output.temp_path(output_type, module_name);
797 copy_gracefully(&path,
798 &crate_output.path(output_type));
799 if !sess.opts.cg.save_temps && !keep_numbered {
800 // The user just wants `foo.x`, not `foo.#module-name#.x`.
804 let ext = crate_output.temp_path(output_type, None)
811 if crate_output.outputs.contains_key(&output_type) {
812 // 2) Multiple codegen units, with `--emit foo=some_name`. We have
813 // no good solution for this case, so warn the user.
814 sess.warn(&format!("ignoring emit path because multiple .{} files \
815 were produced", ext));
816 } else if crate_output.single_output_file.is_some() {
817 // 3) Multiple codegen units, with `-o some_name`. We have
818 // no good solution for this case, so warn the user.
819 sess.warn(&format!("ignoring -o because multiple .{} files \
820 were produced", ext));
822 // 4) Multiple codegen units, but no explicit name. We
823 // just leave the `foo.0.x` files in place.
824 // (We don't have to do any work in this case.)
829 // Flag to indicate whether the user explicitly requested bitcode.
830 // Otherwise, we produced it only as a temporary output, and will need
832 let mut user_wants_bitcode = false;
833 let mut user_wants_objects = false;
834 for output_type in output_types.keys() {
836 OutputType::Bitcode => {
837 user_wants_bitcode = true;
838 // Copy to .bc, but always keep the .0.bc. There is a later
839 // check to figure out if we should delete .0.bc files, or keep
840 // them for making an rlib.
841 copy_if_one_unit(OutputType::Bitcode, true);
843 OutputType::LlvmAssembly => {
844 copy_if_one_unit(OutputType::LlvmAssembly, false);
846 OutputType::Assembly => {
847 copy_if_one_unit(OutputType::Assembly, false);
849 OutputType::Object => {
850 user_wants_objects = true;
851 copy_if_one_unit(OutputType::Object, true);
854 OutputType::DepInfo => {}
857 let user_wants_bitcode = user_wants_bitcode;
859 // Clean up unwanted temporary files.
861 // We create the following files by default:
862 // - crate.#module-name#.bc
863 // - crate.#module-name#.o
864 // - crate.metadata.bc
865 // - crate.metadata.o
866 // - crate.o (linked from crate.##.o)
867 // - crate.bc (copied from crate.##.bc)
868 // We may create additional files if requested by the user (through
869 // `-C save-temps` or `--emit=` flags).
871 if !sess.opts.cg.save_temps {
872 // Remove the temporary .#module-name#.o objects. If the user didn't
873 // explicitly request bitcode (with --emit=bc), and the bitcode is not
874 // needed for building an rlib, then we must remove .#module-name#.bc as
877 // Specific rules for keeping .#module-name#.bc:
878 // - If we're building an rlib (`needs_crate_bitcode`), then keep
880 // - If the user requested bitcode (`user_wants_bitcode`), and
881 // codegen_units > 1, then keep it.
882 // - If the user requested bitcode but codegen_units == 1, then we
883 // can toss .#module-name#.bc because we copied it to .bc earlier.
884 // - If we're not building an rlib and the user didn't request
885 // bitcode, then delete .#module-name#.bc.
886 // If you change how this works, also update back::link::link_rlib,
887 // where .#module-name#.bc files are (maybe) deleted after making an
889 let keep_numbered_bitcode = needs_crate_bitcode ||
890 (user_wants_bitcode && sess.opts.cg.codegen_units > 1);
892 let keep_numbered_objects = needs_crate_object ||
893 (user_wants_objects && sess.opts.cg.codegen_units > 1);
895 for module_name in trans.modules.iter().map(|m| Some(&m.name[..])) {
896 if modules_config.emit_obj && !keep_numbered_objects {
897 let path = crate_output.temp_path(OutputType::Object, module_name);
901 if modules_config.emit_bc && !keep_numbered_bitcode {
902 let path = crate_output.temp_path(OutputType::Bitcode, module_name);
907 if metadata_config.emit_bc && !user_wants_bitcode {
908 let path = crate_output.temp_path(OutputType::Bitcode,
909 Some(&trans.metadata_module.name[..]));
914 // We leave the following files around by default:
916 // - crate.metadata.o
918 // These are used in linking steps and will be cleaned up afterward.
920 // FIXME: time_llvm_passes support - does this use a global context or
922 if sess.opts.cg.codegen_units == 1 && sess.time_llvm_passes() {
923 unsafe { llvm::LLVMRustPrintPassTimings(); }
927 fn dump_incremental_data(trans: &CrateTranslation) {
929 for mtrans in trans.modules.iter() {
930 match mtrans.source {
931 ModuleSource::Preexisting(..) => reuse += 1,
932 ModuleSource::Translated(..) => (),
935 println!("incremental: re-using {} out of {} modules", reuse, trans.modules.len());
939 mtrans: ModuleTranslation,
940 config: ModuleConfig,
941 output_names: OutputFilenames
944 fn build_work_item(sess: &Session,
945 mtrans: ModuleTranslation,
946 config: ModuleConfig,
947 output_names: OutputFilenames)
950 let mut config = config;
951 config.tm = create_target_machine(sess);
955 output_names: output_names
959 fn execute_work_item(cgcx: &CodegenContext,
960 work_item: WorkItem) {
962 match work_item.mtrans.source {
963 ModuleSource::Translated(mllvm) => {
964 debug!("llvm-optimizing {:?}", work_item.mtrans.name);
965 optimize_and_codegen(cgcx,
969 work_item.output_names);
971 ModuleSource::Preexisting(wp) => {
972 let incr_comp_session_dir = cgcx.incr_comp_session_dir
975 let name = &work_item.mtrans.name;
976 for (kind, saved_file) in wp.saved_files {
977 let obj_out = work_item.output_names.temp_path(kind, Some(name));
978 let source_file = in_incr_comp_dir(&incr_comp_session_dir,
980 debug!("copying pre-existing module `{}` from {:?} to {}",
981 work_item.mtrans.name,
984 match link_or_copy(&source_file, &obj_out) {
987 cgcx.handler.err(&format!("unable to copy {} to {}: {}",
988 source_file.display(),
999 fn run_work_singlethreaded(sess: &Session,
1000 exported_symbols: &[String],
1001 work_items: Vec<WorkItem>) {
1002 let cgcx = CodegenContext::new_with_session(sess, exported_symbols);
1004 // Since we're running single-threaded, we can pass the session to
1005 // the proc, allowing `optimize_and_codegen` to perform LTO.
1006 for work in work_items.into_iter().rev() {
1007 execute_work_item(&cgcx, work);
1011 fn run_work_multithreaded(sess: &Session,
1012 work_items: Vec<WorkItem>,
1013 num_workers: usize) {
1014 assert!(num_workers > 0);
1016 // Run some workers to process the work items.
1017 let work_items_arc = Arc::new(Mutex::new(work_items));
1018 let mut diag_emitter = SharedEmitter::new();
1019 let mut futures = Vec::with_capacity(num_workers);
1021 for i in 0..num_workers {
1022 let work_items_arc = work_items_arc.clone();
1023 let diag_emitter = diag_emitter.clone();
1024 let plugin_passes = sess.plugin_llvm_passes.borrow().clone();
1025 let remark = sess.opts.cg.remark.clone();
1027 let (tx, rx) = channel();
1028 let mut tx = Some(tx);
1031 let incr_comp_session_dir = sess.incr_comp_session_dir_opt().map(|r| r.clone());
1033 thread::Builder::new().name(format!("codegen-{}", i)).spawn(move || {
1034 let diag_handler = Handler::with_emitter(true, false, box diag_emitter);
1036 // Must construct cgcx inside the proc because it has non-Send
1038 let cgcx = CodegenContext {
1040 handler: &diag_handler,
1041 plugin_passes: plugin_passes,
1044 incr_comp_session_dir: incr_comp_session_dir
1048 // Avoid holding the lock for the entire duration of the match.
1049 let maybe_work = work_items_arc.lock().unwrap().pop();
1052 execute_work_item(&cgcx, work);
1054 // Make sure to fail the worker so the main thread can
1055 // tell that there were errors.
1056 cgcx.handler.abort_if_errors();
1062 tx.take().unwrap().send(()).unwrap();
1066 let mut panicked = false;
1074 // Display any new diagnostics.
1075 diag_emitter.dump(sess.diagnostic());
1078 sess.fatal("aborting due to worker thread panic");
1082 pub fn run_assembler(sess: &Session, outputs: &OutputFilenames) {
1083 let (pname, mut cmd, _) = get_linker(sess);
1085 cmd.arg("-c").arg("-o").arg(&outputs.path(OutputType::Object))
1086 .arg(&outputs.temp_path(OutputType::Assembly, None));
1087 debug!("{:?}", cmd);
1089 match cmd.output() {
1091 if !prog.status.success() {
1092 let mut note = prog.stderr.clone();
1093 note.extend_from_slice(&prog.stdout);
1095 sess.struct_err(&format!("linking with `{}` failed: {}",
1098 .note(&format!("{:?}", &cmd))
1099 .note(str::from_utf8(¬e[..]).unwrap())
1101 sess.abort_if_errors();
1105 sess.err(&format!("could not exec the linker `{}`: {}", pname, e));
1106 sess.abort_if_errors();
1111 pub unsafe fn with_llvm_pmb(llmod: ModuleRef,
1112 config: &ModuleConfig,
1113 f: &mut FnMut(llvm::PassManagerBuilderRef)) {
1114 // Create the PassManagerBuilder for LLVM. We configure it with
1115 // reasonable defaults and prepare it to actually populate the pass
1117 let builder = llvm::LLVMPassManagerBuilderCreate();
1118 let opt_level = config.opt_level.unwrap_or(llvm::CodeGenOptLevel::None);
1119 let opt_size = config.opt_size.unwrap_or(llvm::CodeGenOptSizeNone);
1120 let inline_threshold = config.inline_threshold;
1122 llvm::LLVMRustConfigurePassManagerBuilder(builder, opt_level,
1123 config.merge_functions,
1124 config.vectorize_slp,
1125 config.vectorize_loop);
1126 llvm::LLVMPassManagerBuilderSetSizeLevel(builder, opt_size as u32);
1128 if opt_size != llvm::CodeGenOptSizeNone {
1129 llvm::LLVMPassManagerBuilderSetDisableUnrollLoops(builder, 1);
1132 llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins);
1134 // Here we match what clang does (kinda). For O0 we only inline
1135 // always-inline functions (but don't add lifetime intrinsics), at O1 we
1136 // inline with lifetime intrinsics, and O2+ we add an inliner with a
1137 // thresholds copied from clang.
1138 match (opt_level, opt_size, inline_threshold) {
1140 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, t as u32);
1142 (llvm::CodeGenOptLevel::Aggressive, ..) => {
1143 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 275);
1145 (_, llvm::CodeGenOptSizeDefault, _) => {
1146 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 75);
1148 (_, llvm::CodeGenOptSizeAggressive, _) => {
1149 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 25);
1151 (llvm::CodeGenOptLevel::None, ..) => {
1152 llvm::LLVMRustAddAlwaysInlinePass(builder, false);
1154 (llvm::CodeGenOptLevel::Less, ..) => {
1155 llvm::LLVMRustAddAlwaysInlinePass(builder, true);
1157 (llvm::CodeGenOptLevel::Default, ..) => {
1158 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 225);
1160 (llvm::CodeGenOptLevel::Other, ..) => {
1161 bug!("CodeGenOptLevel::Other selected")
1166 llvm::LLVMPassManagerBuilderDispose(builder);