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 session::config::{OutputFilenames, Passes, SomePasses, AllPasses};
15 use session::config::{self, OutputType};
17 use llvm::{ModuleRef, TargetMachineRef, PassManagerRef, DiagnosticInfoRef, ContextRef};
18 use llvm::SMDiagnosticRef;
19 use trans::{CrateTranslation, ModuleTranslation};
20 use util::common::time;
21 use util::common::path2cstr;
23 use syntax::errors::{self, Handler, Level};
24 use syntax::errors::emitter::Emitter;
26 use std::collections::HashMap;
27 use std::ffi::{CStr, CString};
29 use std::path::{Path, PathBuf};
32 use std::sync::{Arc, Mutex};
33 use std::sync::mpsc::channel;
35 use libc::{self, c_uint, c_int, c_void};
37 pub fn llvm_err(handler: &errors::Handler, msg: String) -> ! {
39 let cstr = llvm::LLVMRustGetLastError();
40 if cstr == ptr::null() {
41 panic!(handler.fatal(&msg[..]));
43 let err = CStr::from_ptr(cstr).to_bytes();
44 let err = String::from_utf8_lossy(err).to_string();
45 libc::free(cstr as *mut _);
46 panic!(handler.fatal(&format!("{}: {}", &msg[..], &err[..])));
51 pub fn write_output_file(
52 handler: &errors::Handler,
53 target: llvm::TargetMachineRef,
54 pm: llvm::PassManagerRef,
57 file_type: llvm::FileType) {
59 let output_c = path2cstr(output);
60 let result = llvm::LLVMRustWriteOutputFile(
61 target, pm, m, output_c.as_ptr(), file_type);
63 llvm_err(handler, format!("could not write output to {}", output.display()));
75 // We use an Arc instead of just returning a list of diagnostics from the
76 // child thread because we need to make sure that the messages are seen even
77 // if the child thread panics (for example, when `fatal` is called).
79 struct SharedEmitter {
80 buffer: Arc<Mutex<Vec<Diagnostic>>>,
84 fn new() -> SharedEmitter {
86 buffer: Arc::new(Mutex::new(Vec::new())),
90 fn dump(&mut self, handler: &Handler) {
91 let mut buffer = self.buffer.lock().unwrap();
92 for diag in &*buffer {
95 handler.emit_with_code(None,
111 impl Emitter for SharedEmitter {
112 fn emit(&mut self, sp: Option<codemap::Span>,
113 msg: &str, code: Option<&str>, lvl: Level) {
114 assert!(sp.is_none(), "SharedEmitter doesn't support spans");
116 self.buffer.lock().unwrap().push(Diagnostic {
117 msg: msg.to_string(),
118 code: code.map(|s| s.to_string()),
123 fn custom_emit(&mut self, _sp: errors::RenderSpan, _msg: &str, _lvl: Level) {
124 panic!("SharedEmitter doesn't support custom_emit");
129 // On android, we by default compile for armv7 processors. This enables
130 // things like double word CAS instructions (rather than emulating them)
131 // which are *far* more efficient. This is obviously undesirable in some
132 // cases, so if any sort of target feature is specified we don't append v7
133 // to the feature list.
135 // On iOS only armv7 and newer are supported. So it is useful to
136 // get all hardware potential via VFP3 (hardware floating point)
137 // and NEON (SIMD) instructions supported by LLVM.
138 // Note that without those flags various linking errors might
139 // arise as some of intrinsics are converted into function calls
140 // and nobody provides implementations those functions
141 fn target_feature(sess: &Session) -> String {
142 format!("{},{}", sess.target.target.options.features, sess.opts.cg.target_feature)
145 fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel {
147 config::No => llvm::CodeGenLevelNone,
148 config::Less => llvm::CodeGenLevelLess,
149 config::Default => llvm::CodeGenLevelDefault,
150 config::Aggressive => llvm::CodeGenLevelAggressive,
154 pub fn create_target_machine(sess: &Session) -> TargetMachineRef {
155 let reloc_model_arg = match sess.opts.cg.relocation_model {
156 Some(ref s) => &s[..],
157 None => &sess.target.target.options.relocation_model[..],
159 let reloc_model = match reloc_model_arg {
160 "pic" => llvm::RelocPIC,
161 "static" => llvm::RelocStatic,
162 "default" => llvm::RelocDefault,
163 "dynamic-no-pic" => llvm::RelocDynamicNoPic,
165 sess.err(&format!("{:?} is not a valid relocation mode",
169 sess.abort_if_errors();
174 let opt_level = get_llvm_opt_level(sess.opts.optimize);
175 let use_softfp = sess.opts.cg.soft_float;
177 let any_library = sess.crate_types.borrow().iter().any(|ty| {
178 *ty != config::CrateTypeExecutable
181 let ffunction_sections = sess.target.target.options.function_sections;
182 let fdata_sections = ffunction_sections;
184 let code_model_arg = match sess.opts.cg.code_model {
185 Some(ref s) => &s[..],
186 None => &sess.target.target.options.code_model[..],
189 let code_model = match code_model_arg {
190 "default" => llvm::CodeModelDefault,
191 "small" => llvm::CodeModelSmall,
192 "kernel" => llvm::CodeModelKernel,
193 "medium" => llvm::CodeModelMedium,
194 "large" => llvm::CodeModelLarge,
196 sess.err(&format!("{:?} is not a valid code model",
200 sess.abort_if_errors();
205 let triple = &sess.target.target.llvm_target;
208 let triple = CString::new(triple.as_bytes()).unwrap();
209 let cpu = match sess.opts.cg.target_cpu {
211 None => &*sess.target.target.options.cpu
213 let cpu = CString::new(cpu.as_bytes()).unwrap();
214 let features = CString::new(target_feature(sess).as_bytes()).unwrap();
215 llvm::LLVMRustCreateTargetMachine(
216 triple.as_ptr(), cpu.as_ptr(), features.as_ptr(),
221 !any_library && reloc_model == llvm::RelocPIC,
228 llvm_err(sess.diagnostic(),
229 format!("Could not create LLVM TargetMachine for triple: {}",
230 triple).to_string());
237 /// Module-specific configuration for `optimize_and_codegen`.
239 pub struct ModuleConfig {
240 /// LLVM TargetMachine to use for codegen.
241 tm: TargetMachineRef,
242 /// Names of additional optimization passes to run.
244 /// Some(level) to optimize at a certain level, or None to run
245 /// absolutely no optimizations (used for the metadata module).
246 opt_level: Option<llvm::CodeGenOptLevel>,
248 // Flags indicating which outputs to produce.
249 emit_no_opt_bc: bool,
256 // Miscellaneous flags. These are mostly copied from command-line
259 no_prepopulate_passes: bool,
262 vectorize_loop: bool,
264 merge_functions: bool,
265 inline_threshold: Option<usize>
268 unsafe impl Send for ModuleConfig { }
271 fn new(tm: TargetMachineRef, passes: Vec<String>) -> ModuleConfig {
277 emit_no_opt_bc: false,
285 no_prepopulate_passes: false,
288 vectorize_loop: false,
289 vectorize_slp: false,
290 merge_functions: false,
291 inline_threshold: None
295 fn set_flags(&mut self, sess: &Session, trans: &CrateTranslation) {
296 self.no_verify = sess.no_verify();
297 self.no_prepopulate_passes = sess.opts.cg.no_prepopulate_passes;
298 self.no_builtins = trans.no_builtins;
299 self.time_passes = sess.time_passes();
300 self.inline_threshold = sess.opts.cg.inline_threshold;
302 // Copy what clang does by turning on loop vectorization at O2 and
303 // slp vectorization at O3. Otherwise configure other optimization aspects
304 // of this pass manager builder.
305 self.vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
306 (sess.opts.optimize == config::Default ||
307 sess.opts.optimize == config::Aggressive);
308 self.vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
309 sess.opts.optimize == config::Aggressive;
311 self.merge_functions = sess.opts.optimize == config::Default ||
312 sess.opts.optimize == config::Aggressive;
316 /// Additional resources used by optimize_and_codegen (not module specific)
317 struct CodegenContext<'a> {
318 // Extra resources used for LTO: (sess, reachable). This will be `None`
319 // when running in a worker thread.
320 lto_ctxt: Option<(&'a Session, &'a [String])>,
321 // Handler to use for diagnostics produced during codegen.
322 handler: &'a Handler,
323 // LLVM passes added by plugins.
324 plugin_passes: Vec<String>,
325 // LLVM optimizations for which we want to print remarks.
327 // Worker thread number
331 impl<'a> CodegenContext<'a> {
332 fn new_with_session(sess: &'a Session, reachable: &'a [String]) -> CodegenContext<'a> {
334 lto_ctxt: Some((sess, reachable)),
335 handler: sess.diagnostic(),
336 plugin_passes: sess.plugin_llvm_passes.borrow().clone(),
337 remark: sess.opts.cg.remark.clone(),
343 struct HandlerFreeVars<'a> {
345 cgcx: &'a CodegenContext<'a>,
348 unsafe extern "C" fn report_inline_asm<'a, 'b>(cgcx: &'a CodegenContext<'a>,
351 use syntax::codemap::ExpnId;
353 match cgcx.lto_ctxt {
355 sess.codemap().with_expn_info(ExpnId::from_u32(cookie), |info| match info {
356 Some(ei) => sess.span_err(ei.call_site, msg),
357 None => sess.err(msg),
362 cgcx.handler.err(msg);
363 cgcx.handler.note("build without -C codegen-units for more exact errors");
368 unsafe extern "C" fn inline_asm_handler(diag: SMDiagnosticRef,
371 let HandlerFreeVars { cgcx, .. } = *(user as *const HandlerFreeVars);
373 let msg = llvm::build_string(|s| llvm::LLVMWriteSMDiagnosticToString(diag, s))
374 .expect("non-UTF8 SMDiagnostic");
376 report_inline_asm(cgcx, &msg[..], cookie);
379 unsafe extern "C" fn diagnostic_handler(info: DiagnosticInfoRef, user: *mut c_void) {
380 let HandlerFreeVars { llcx, cgcx } = *(user as *const HandlerFreeVars);
382 match llvm::diagnostic::Diagnostic::unpack(info) {
383 llvm::diagnostic::InlineAsm(inline) => {
384 report_inline_asm(cgcx,
385 &*llvm::twine_to_string(inline.message),
389 llvm::diagnostic::Optimization(opt) => {
390 let pass_name = str::from_utf8(CStr::from_ptr(opt.pass_name).to_bytes())
392 .expect("got a non-UTF8 pass name from LLVM");
393 let enabled = match cgcx.remark {
395 SomePasses(ref v) => v.iter().any(|s| *s == pass_name),
399 let loc = llvm::debug_loc_to_string(llcx, opt.debug_loc);
400 cgcx.handler.note(&format!("optimization {} for {} at {}: {}",
403 if loc.is_empty() { "[unknown]" } else { &*loc },
404 llvm::twine_to_string(opt.message)));
412 // Unsafe due to LLVM calls.
413 unsafe fn optimize_and_codegen(cgcx: &CodegenContext,
414 mtrans: ModuleTranslation,
415 config: ModuleConfig,
417 output_names: OutputFilenames) {
418 let ModuleTranslation { llmod, llcx } = mtrans;
421 // llcx doesn't outlive this function, so we can put this on the stack.
422 let fv = HandlerFreeVars {
426 let fv = &fv as *const HandlerFreeVars as *mut c_void;
428 llvm::LLVMSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, fv);
429 llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, fv);
431 if config.emit_no_opt_bc {
432 let ext = format!("{}.no-opt.bc", name_extra);
433 let out = output_names.with_extension(&ext);
434 let out = path2cstr(&out);
435 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
438 if config.opt_level.is_some() {
439 // Create the two optimizing pass managers. These mirror what clang
440 // does, and are by populated by LLVM's default PassManagerBuilder.
441 // Each manager has a different set of passes, but they also share
442 // some common passes.
443 let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
444 let mpm = llvm::LLVMCreatePassManager();
446 // If we're verifying or linting, add them to the function pass
448 let addpass = |pass: &str| {
449 let pass = CString::new(pass).unwrap();
450 llvm::LLVMRustAddPass(fpm, pass.as_ptr())
453 if !config.no_verify { assert!(addpass("verify")); }
454 if !config.no_prepopulate_passes {
455 llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod);
456 llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod);
457 with_llvm_pmb(llmod, &config, &mut |b| {
458 llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(b, fpm);
459 llvm::LLVMPassManagerBuilderPopulateModulePassManager(b, mpm);
463 for pass in &config.passes {
465 cgcx.handler.warn(&format!("unknown pass `{}`, ignoring",
470 for pass in &cgcx.plugin_passes {
472 cgcx.handler.err(&format!("a plugin asked for LLVM pass \
473 `{}` but LLVM does not \
474 recognize it", pass));
478 cgcx.handler.abort_if_errors();
480 // Finally, run the actual optimization passes
481 time(config.time_passes, &format!("llvm function passes [{}]", cgcx.worker), ||
482 llvm::LLVMRustRunFunctionPassManager(fpm, llmod));
483 time(config.time_passes, &format!("llvm module passes [{}]", cgcx.worker), ||
484 llvm::LLVMRunPassManager(mpm, llmod));
486 // Deallocate managers that we're now done with
487 llvm::LLVMDisposePassManager(fpm);
488 llvm::LLVMDisposePassManager(mpm);
490 match cgcx.lto_ctxt {
491 Some((sess, reachable)) if sess.lto() => {
492 time(sess.time_passes(), "all lto passes", ||
493 lto::run(sess, llmod, tm, reachable, &config,
494 &name_extra, &output_names));
496 if config.emit_lto_bc {
497 let name = format!("{}.lto.bc", name_extra);
498 let out = output_names.with_extension(&name);
499 let out = path2cstr(&out);
500 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
507 // A codegen-specific pass manager is used to generate object
508 // files for an LLVM module.
510 // Apparently each of these pass managers is a one-shot kind of
511 // thing, so we create a new one for each type of output. The
512 // pass manager passed to the closure should be ensured to not
513 // escape the closure itself, and the manager should only be
515 unsafe fn with_codegen<F>(tm: TargetMachineRef,
519 F: FnOnce(PassManagerRef),
521 let cpm = llvm::LLVMCreatePassManager();
522 llvm::LLVMRustAddAnalysisPasses(tm, cpm, llmod);
523 llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
528 let ext = format!("{}.bc", name_extra);
529 let out = output_names.with_extension(&ext);
530 let out = path2cstr(&out);
531 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
534 time(config.time_passes, &format!("codegen passes [{}]", cgcx.worker), || {
536 let ext = format!("{}.ll", name_extra);
537 let out = output_names.with_extension(&ext);
538 let out = path2cstr(&out);
539 with_codegen(tm, llmod, config.no_builtins, |cpm| {
540 llvm::LLVMRustPrintModule(cpm, llmod, out.as_ptr());
541 llvm::LLVMDisposePassManager(cpm);
546 let path = output_names.with_extension(&format!("{}.s", name_extra));
548 // We can't use the same module for asm and binary output, because that triggers
549 // various errors like invalid IR or broken binaries, so we might have to clone the
550 // module to produce the asm output
551 let llmod = if config.emit_obj {
552 llvm::LLVMCloneModule(llmod)
556 with_codegen(tm, llmod, config.no_builtins, |cpm| {
557 write_output_file(cgcx.handler, tm, cpm, llmod, &path,
558 llvm::AssemblyFileType);
561 llvm::LLVMDisposeModule(llmod);
566 let path = output_names.with_extension(&format!("{}.o", name_extra));
567 with_codegen(tm, llmod, config.no_builtins, |cpm| {
568 write_output_file(cgcx.handler, tm, cpm, llmod, &path, llvm::ObjectFileType);
573 llvm::LLVMDisposeModule(llmod);
574 llvm::LLVMContextDispose(llcx);
575 llvm::LLVMRustDisposeTargetMachine(tm);
578 pub fn run_passes(sess: &Session,
579 trans: &CrateTranslation,
580 output_types: &HashMap<OutputType, Option<PathBuf>>,
581 crate_output: &OutputFilenames) {
582 // It's possible that we have `codegen_units > 1` but only one item in
583 // `trans.modules`. We could theoretically proceed and do LTO in that
584 // case, but it would be confusing to have the validity of
585 // `-Z lto -C codegen-units=2` depend on details of the crate being
586 // compiled, so we complain regardless.
587 if sess.lto() && sess.opts.cg.codegen_units > 1 {
588 // This case is impossible to handle because LTO expects to be able
589 // to combine the entire crate and all its dependencies into a
590 // single compilation unit, but each codegen unit is in a separate
591 // LLVM context, so they can't easily be combined.
592 sess.fatal("can't perform LTO when using multiple codegen units");
596 assert!(trans.modules.len() == sess.opts.cg.codegen_units);
598 let tm = create_target_machine(sess);
600 // Figure out what we actually need to build.
602 let mut modules_config = ModuleConfig::new(tm, sess.opts.cg.passes.clone());
603 let mut metadata_config = ModuleConfig::new(tm, vec!());
605 modules_config.opt_level = Some(get_llvm_opt_level(sess.opts.optimize));
607 // Save all versions of the bytecode if we're saving our temporaries.
608 if sess.opts.cg.save_temps {
609 modules_config.emit_no_opt_bc = true;
610 modules_config.emit_bc = true;
611 modules_config.emit_lto_bc = true;
612 metadata_config.emit_bc = true;
615 // Emit bitcode files for the crate if we're emitting an rlib.
616 // Whenever an rlib is created, the bitcode is inserted into the
617 // archive in order to allow LTO against it.
618 let needs_crate_bitcode =
619 sess.crate_types.borrow().contains(&config::CrateTypeRlib) &&
620 sess.opts.output_types.contains_key(&OutputType::Exe);
621 let needs_crate_object =
622 sess.opts.output_types.contains_key(&OutputType::Exe);
623 if needs_crate_bitcode {
624 modules_config.emit_bc = true;
627 for output_type in output_types.keys() {
629 OutputType::Bitcode => { modules_config.emit_bc = true; },
630 OutputType::LlvmAssembly => { modules_config.emit_ir = true; },
631 OutputType::Assembly => {
632 modules_config.emit_asm = true;
633 // If we're not using the LLVM assembler, this function
634 // could be invoked specially with output_type_assembly, so
635 // in this case we still want the metadata object file.
636 if !sess.opts.output_types.contains_key(&OutputType::Assembly) {
637 metadata_config.emit_obj = true;
640 OutputType::Object => { modules_config.emit_obj = true; },
642 modules_config.emit_obj = true;
643 metadata_config.emit_obj = true;
645 OutputType::DepInfo => {}
649 modules_config.set_flags(sess, trans);
650 metadata_config.set_flags(sess, trans);
653 // Populate a buffer with a list of codegen threads. Items are processed in
654 // LIFO order, just because it's a tiny bit simpler that way. (The order
655 // doesn't actually matter.)
656 let mut work_items = Vec::with_capacity(1 + trans.modules.len());
659 let work = build_work_item(sess,
660 trans.metadata_module,
661 metadata_config.clone(),
662 crate_output.clone(),
663 "metadata".to_string());
664 work_items.push(work);
667 for (index, mtrans) in trans.modules.iter().enumerate() {
668 let work = build_work_item(sess,
670 modules_config.clone(),
671 crate_output.clone(),
672 format!("{}", index));
673 work_items.push(work);
676 // Process the work items, optionally using worker threads.
677 if sess.opts.cg.codegen_units == 1 {
678 run_work_singlethreaded(sess, &trans.reachable, work_items);
680 run_work_multithreaded(sess, work_items, sess.opts.cg.codegen_units);
683 // All codegen is finished.
685 llvm::LLVMRustDisposeTargetMachine(tm);
688 // Produce final compile outputs.
689 let copy_gracefully = |from: &Path, to: &Path| {
690 if let Err(e) = fs::copy(from, to) {
691 sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
695 let copy_if_one_unit = |ext: &str,
696 output_type: OutputType,
697 keep_numbered: bool| {
698 if sess.opts.cg.codegen_units == 1 {
699 // 1) Only one codegen unit. In this case it's no difficulty
700 // to copy `foo.0.x` to `foo.x`.
701 copy_gracefully(&crate_output.with_extension(ext),
702 &crate_output.path(output_type));
703 if !sess.opts.cg.save_temps && !keep_numbered {
704 // The user just wants `foo.x`, not `foo.0.x`.
705 remove(sess, &crate_output.with_extension(ext));
707 } else if crate_output.outputs.contains_key(&output_type) {
708 // 2) Multiple codegen units, with `--emit foo=some_name`. We have
709 // no good solution for this case, so warn the user.
710 sess.warn(&format!("ignoring emit path because multiple .{} files \
711 were produced", ext));
712 } else if crate_output.single_output_file.is_some() {
713 // 3) Multiple codegen units, with `-o some_name`. We have
714 // no good solution for this case, so warn the user.
715 sess.warn(&format!("ignoring -o because multiple .{} files \
716 were produced", ext));
718 // 4) Multiple codegen units, but no explicit name. We
719 // just leave the `foo.0.x` files in place.
720 // (We don't have to do any work in this case.)
724 // Flag to indicate whether the user explicitly requested bitcode.
725 // Otherwise, we produced it only as a temporary output, and will need
727 let mut user_wants_bitcode = false;
728 let mut user_wants_objects = false;
729 for output_type in output_types.keys() {
731 OutputType::Bitcode => {
732 user_wants_bitcode = true;
733 // Copy to .bc, but always keep the .0.bc. There is a later
734 // check to figure out if we should delete .0.bc files, or keep
735 // them for making an rlib.
736 copy_if_one_unit("0.bc", OutputType::Bitcode, true);
738 OutputType::LlvmAssembly => {
739 copy_if_one_unit("0.ll", OutputType::LlvmAssembly, false);
741 OutputType::Assembly => {
742 copy_if_one_unit("0.s", OutputType::Assembly, false);
744 OutputType::Object => {
745 user_wants_objects = true;
746 copy_if_one_unit("0.o", OutputType::Object, true);
749 OutputType::DepInfo => {}
752 let user_wants_bitcode = user_wants_bitcode;
754 // Clean up unwanted temporary files.
756 // We create the following files by default:
759 // - crate.metadata.bc
760 // - crate.metadata.o
761 // - crate.o (linked from crate.##.o)
762 // - crate.bc (copied from crate.0.bc)
763 // We may create additional files if requested by the user (through
764 // `-C save-temps` or `--emit=` flags).
766 if !sess.opts.cg.save_temps {
767 // Remove the temporary .0.o objects. If the user didn't
768 // explicitly request bitcode (with --emit=bc), and the bitcode is not
769 // needed for building an rlib, then we must remove .0.bc as well.
771 // Specific rules for keeping .0.bc:
772 // - If we're building an rlib (`needs_crate_bitcode`), then keep
774 // - If the user requested bitcode (`user_wants_bitcode`), and
775 // codegen_units > 1, then keep it.
776 // - If the user requested bitcode but codegen_units == 1, then we
777 // can toss .0.bc because we copied it to .bc earlier.
778 // - If we're not building an rlib and the user didn't request
779 // bitcode, then delete .0.bc.
780 // If you change how this works, also update back::link::link_rlib,
781 // where .0.bc files are (maybe) deleted after making an rlib.
782 let keep_numbered_bitcode = needs_crate_bitcode ||
783 (user_wants_bitcode && sess.opts.cg.codegen_units > 1);
785 let keep_numbered_objects = needs_crate_object ||
786 (user_wants_objects && sess.opts.cg.codegen_units > 1);
788 for i in 0..trans.modules.len() {
789 if modules_config.emit_obj && !keep_numbered_objects {
790 let ext = format!("{}.o", i);
791 remove(sess, &crate_output.with_extension(&ext));
794 if modules_config.emit_bc && !keep_numbered_bitcode {
795 let ext = format!("{}.bc", i);
796 remove(sess, &crate_output.with_extension(&ext));
800 if metadata_config.emit_bc && !user_wants_bitcode {
801 remove(sess, &crate_output.with_extension("metadata.bc"));
805 // We leave the following files around by default:
807 // - crate.metadata.o
809 // These are used in linking steps and will be cleaned up afterward.
811 // FIXME: time_llvm_passes support - does this use a global context or
813 if sess.opts.cg.codegen_units == 1 && sess.time_llvm_passes() {
814 unsafe { llvm::LLVMRustPrintPassTimings(); }
819 mtrans: ModuleTranslation,
820 config: ModuleConfig,
821 output_names: OutputFilenames,
825 fn build_work_item(sess: &Session,
826 mtrans: ModuleTranslation,
827 config: ModuleConfig,
828 output_names: OutputFilenames,
832 let mut config = config;
833 config.tm = create_target_machine(sess);
834 WorkItem { mtrans: mtrans, config: config, output_names: output_names,
835 name_extra: name_extra }
838 fn execute_work_item(cgcx: &CodegenContext,
839 work_item: WorkItem) {
841 optimize_and_codegen(cgcx, work_item.mtrans, work_item.config,
842 work_item.name_extra, work_item.output_names);
846 fn run_work_singlethreaded(sess: &Session,
847 reachable: &[String],
848 work_items: Vec<WorkItem>) {
849 let cgcx = CodegenContext::new_with_session(sess, reachable);
851 // Since we're running single-threaded, we can pass the session to
852 // the proc, allowing `optimize_and_codegen` to perform LTO.
853 for work in work_items.into_iter().rev() {
854 execute_work_item(&cgcx, work);
858 fn run_work_multithreaded(sess: &Session,
859 work_items: Vec<WorkItem>,
860 num_workers: usize) {
861 // Run some workers to process the work items.
862 let work_items_arc = Arc::new(Mutex::new(work_items));
863 let mut diag_emitter = SharedEmitter::new();
864 let mut futures = Vec::with_capacity(num_workers);
866 for i in 0..num_workers {
867 let work_items_arc = work_items_arc.clone();
868 let diag_emitter = diag_emitter.clone();
869 let plugin_passes = sess.plugin_llvm_passes.borrow().clone();
870 let remark = sess.opts.cg.remark.clone();
872 let (tx, rx) = channel();
873 let mut tx = Some(tx);
876 thread::Builder::new().name(format!("codegen-{}", i)).spawn(move || {
877 let diag_handler = Handler::with_emitter(true, false, box diag_emitter);
879 // Must construct cgcx inside the proc because it has non-Send
881 let cgcx = CodegenContext {
883 handler: &diag_handler,
884 plugin_passes: plugin_passes,
890 // Avoid holding the lock for the entire duration of the match.
891 let maybe_work = work_items_arc.lock().unwrap().pop();
894 execute_work_item(&cgcx, work);
896 // Make sure to fail the worker so the main thread can
897 // tell that there were errors.
898 cgcx.handler.abort_if_errors();
904 tx.take().unwrap().send(()).unwrap();
908 let mut panicked = false;
916 // Display any new diagnostics.
917 diag_emitter.dump(sess.diagnostic());
920 sess.fatal("aborting due to worker thread panic");
924 pub fn run_assembler(sess: &Session, outputs: &OutputFilenames) {
925 let (pname, mut cmd) = get_linker(sess);
927 cmd.arg("-c").arg("-o").arg(&outputs.path(OutputType::Object))
928 .arg(&outputs.temp_path(OutputType::Assembly));
933 if !prog.status.success() {
934 sess.err(&format!("linking with `{}` failed: {}",
937 sess.note(&format!("{:?}", &cmd));
938 let mut note = prog.stderr.clone();
939 note.extend_from_slice(&prog.stdout);
940 sess.note(str::from_utf8(¬e[..]).unwrap());
941 sess.abort_if_errors();
945 sess.err(&format!("could not exec the linker `{}`: {}", pname, e));
946 sess.abort_if_errors();
951 pub unsafe fn configure_llvm(sess: &Session) {
952 let mut llvm_c_strs = Vec::new();
953 let mut llvm_args = Vec::new();
956 let mut add = |arg: &str| {
957 let s = CString::new(arg).unwrap();
958 llvm_args.push(s.as_ptr());
961 add("rustc"); // fake program name
962 if sess.time_llvm_passes() { add("-time-passes"); }
963 if sess.print_llvm_passes() { add("-debug-pass=Structure"); }
965 // FIXME #21627 disable faulty FastISel on AArch64 (even for -O0)
966 if sess.target.target.arch == "aarch64" { add("-fast-isel=0"); }
968 for arg in &sess.opts.cg.llvm_args {
973 llvm::LLVMInitializePasses();
975 llvm::initialize_available_targets();
977 llvm::LLVMRustSetLLVMOptions(llvm_args.len() as c_int,
981 pub unsafe fn with_llvm_pmb(llmod: ModuleRef,
982 config: &ModuleConfig,
983 f: &mut FnMut(llvm::PassManagerBuilderRef)) {
984 // Create the PassManagerBuilder for LLVM. We configure it with
985 // reasonable defaults and prepare it to actually populate the pass
987 let builder = llvm::LLVMPassManagerBuilderCreate();
988 let opt = config.opt_level.unwrap_or(llvm::CodeGenLevelNone);
989 let inline_threshold = config.inline_threshold;
991 llvm::LLVMRustConfigurePassManagerBuilder(builder, opt,
992 config.merge_functions,
993 config.vectorize_slp,
994 config.vectorize_loop);
996 llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins);
998 // Here we match what clang does (kinda). For O0 we only inline
999 // always-inline functions (but don't add lifetime intrinsics), at O1 we
1000 // inline with lifetime intrinsics, and O2+ we add an inliner with a
1001 // thresholds copied from clang.
1002 match (opt, inline_threshold) {
1004 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, t as u32);
1006 (llvm::CodeGenLevelNone, _) => {
1007 llvm::LLVMRustAddAlwaysInlinePass(builder, false);
1009 (llvm::CodeGenLevelLess, _) => {
1010 llvm::LLVMRustAddAlwaysInlinePass(builder, true);
1012 (llvm::CodeGenLevelDefault, _) => {
1013 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 225);
1015 (llvm::CodeGenLevelAggressive, _) => {
1016 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 275);
1021 llvm::LLVMPassManagerBuilderDispose(builder);