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_cc_prog, remove};
13 use session::config::{OutputFilenames, NoDebugInfo, Passes, SomePasses, AllPasses};
17 use llvm::{ModuleRef, TargetMachineRef, PassManagerRef, DiagnosticInfoRef, ContextRef};
18 use llvm::SMDiagnosticRef;
19 use trans::{CrateTranslation, ModuleTranslation};
20 use util::common::time;
22 use syntax::diagnostic;
23 use syntax::diagnostic::{Emitter, Handler, Level, mk_handler};
25 use std::ffi::{CStr, CString};
26 use std::old_io::Command;
28 use std::iter::Unfold;
32 use std::sync::{Arc, Mutex};
33 use std::sync::mpsc::channel;
35 use libc::{self, c_uint, c_int, c_void};
37 #[derive(Clone, Copy, PartialEq, PartialOrd, Ord, Eq)]
41 OutputTypeLlvmAssembly,
46 pub fn llvm_err(handler: &diagnostic::Handler, msg: String) -> ! {
48 let cstr = llvm::LLVMRustGetLastError();
49 if cstr == ptr::null() {
50 handler.fatal(&msg[..]);
52 let err = CStr::from_ptr(cstr).to_bytes();
53 let err = String::from_utf8_lossy(err).to_string();
54 libc::free(cstr as *mut _);
55 handler.fatal(&format!("{}: {}",
62 pub fn write_output_file(
63 handler: &diagnostic::Handler,
64 target: llvm::TargetMachineRef,
65 pm: llvm::PassManagerRef,
68 file_type: llvm::FileType) {
70 let output_c = CString::new(output.as_vec()).unwrap();
71 let result = llvm::LLVMRustWriteOutputFile(
72 target, pm, m, output_c.as_ptr(), file_type);
74 llvm_err(handler, format!("could not write output to {}", output.display()));
86 // We use an Arc instead of just returning a list of diagnostics from the
87 // child task because we need to make sure that the messages are seen even
88 // if the child task panics (for example, when `fatal` is called).
90 struct SharedEmitter {
91 buffer: Arc<Mutex<Vec<Diagnostic>>>,
95 fn new() -> SharedEmitter {
97 buffer: Arc::new(Mutex::new(Vec::new())),
101 fn dump(&mut self, handler: &Handler) {
102 let mut buffer = self.buffer.lock().unwrap();
103 for diag in &*buffer {
106 handler.emit_with_code(None,
122 impl Emitter for SharedEmitter {
123 fn emit(&mut self, cmsp: Option<(&codemap::CodeMap, codemap::Span)>,
124 msg: &str, code: Option<&str>, lvl: Level) {
125 assert!(cmsp.is_none(), "SharedEmitter doesn't support spans");
127 self.buffer.lock().unwrap().push(Diagnostic {
128 msg: msg.to_string(),
129 code: code.map(|s| s.to_string()),
134 fn custom_emit(&mut self, _cm: &codemap::CodeMap,
135 _sp: diagnostic::RenderSpan, _msg: &str, _lvl: Level) {
136 panic!("SharedEmitter doesn't support custom_emit");
141 // On android, we by default compile for armv7 processors. This enables
142 // things like double word CAS instructions (rather than emulating them)
143 // which are *far* more efficient. This is obviously undesirable in some
144 // cases, so if any sort of target feature is specified we don't append v7
145 // to the feature list.
147 // On iOS only armv7 and newer are supported. So it is useful to
148 // get all hardware potential via VFP3 (hardware floating point)
149 // and NEON (SIMD) instructions supported by LLVM.
150 // Note that without those flags various linking errors might
151 // arise as some of intrinsics are converted into function calls
152 // and nobody provides implementations those functions
153 fn target_feature(sess: &Session) -> String {
154 format!("{},{}", sess.target.target.options.features, sess.opts.cg.target_feature)
157 fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel {
159 config::No => llvm::CodeGenLevelNone,
160 config::Less => llvm::CodeGenLevelLess,
161 config::Default => llvm::CodeGenLevelDefault,
162 config::Aggressive => llvm::CodeGenLevelAggressive,
166 fn create_target_machine(sess: &Session) -> TargetMachineRef {
167 let reloc_model_arg = match sess.opts.cg.relocation_model {
168 Some(ref s) => &s[..],
169 None => &sess.target.target.options.relocation_model[]
171 let reloc_model = match reloc_model_arg {
172 "pic" => llvm::RelocPIC,
173 "static" => llvm::RelocStatic,
174 "default" => llvm::RelocDefault,
175 "dynamic-no-pic" => llvm::RelocDynamicNoPic,
177 sess.err(&format!("{:?} is not a valid relocation mode",
180 .relocation_model)[]);
181 sess.abort_if_errors();
186 let opt_level = get_llvm_opt_level(sess.opts.optimize);
187 let use_softfp = sess.opts.cg.soft_float;
189 // FIXME: #11906: Omitting frame pointers breaks retrieving the value of a parameter.
190 let no_fp_elim = (sess.opts.debuginfo != NoDebugInfo) ||
191 !sess.target.target.options.eliminate_frame_pointer;
193 let any_library = sess.crate_types.borrow().iter().any(|ty| {
194 *ty != config::CrateTypeExecutable
197 let ffunction_sections = sess.target.target.options.function_sections;
198 let fdata_sections = ffunction_sections;
200 let code_model_arg = match sess.opts.cg.code_model {
201 Some(ref s) => &s[..],
202 None => &sess.target.target.options.code_model[]
205 let code_model = match code_model_arg {
206 "default" => llvm::CodeModelDefault,
207 "small" => llvm::CodeModelSmall,
208 "kernel" => llvm::CodeModelKernel,
209 "medium" => llvm::CodeModelMedium,
210 "large" => llvm::CodeModelLarge,
212 sess.err(&format!("{:?} is not a valid code model",
216 sess.abort_if_errors();
221 let triple = &sess.target.target.llvm_target[];
224 let triple = CString::new(triple.as_bytes()).unwrap();
225 let cpu = match sess.opts.cg.target_cpu {
227 None => &*sess.target.target.options.cpu
229 let cpu = CString::new(cpu.as_bytes()).unwrap();
230 let features = CString::new(target_feature(sess).as_bytes()).unwrap();
231 llvm::LLVMRustCreateTargetMachine(
232 triple.as_ptr(), cpu.as_ptr(), features.as_ptr(),
236 true /* EnableSegstk */,
239 !any_library && reloc_model == llvm::RelocPIC,
246 llvm_err(sess.diagnostic().handler(),
247 format!("Could not create LLVM TargetMachine for triple: {}",
248 triple).to_string());
255 /// Module-specific configuration for `optimize_and_codegen`.
257 struct ModuleConfig {
258 /// LLVM TargetMachine to use for codegen.
259 tm: TargetMachineRef,
260 /// Names of additional optimization passes to run.
262 /// Some(level) to optimize at a certain level, or None to run
263 /// absolutely no optimizations (used for the metadata module).
264 opt_level: Option<llvm::CodeGenOptLevel>,
266 // Flags indicating which outputs to produce.
267 emit_no_opt_bc: bool,
274 // Miscellaneous flags. These are mostly copied from command-line
277 no_prepopulate_passes: bool,
282 unsafe impl Send for ModuleConfig { }
285 fn new(tm: TargetMachineRef, passes: Vec<String>) -> ModuleConfig {
291 emit_no_opt_bc: false,
299 no_prepopulate_passes: false,
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();
313 /// Additional resources used by optimize_and_codegen (not module specific)
314 struct CodegenContext<'a> {
315 // Extra resources used for LTO: (sess, reachable). This will be `None`
316 // when running in a worker thread.
317 lto_ctxt: Option<(&'a Session, &'a [String])>,
318 // Handler to use for diagnostics produced during codegen.
319 handler: &'a Handler,
320 // LLVM optimizations for which we want to print remarks.
324 impl<'a> CodegenContext<'a> {
325 fn new_with_session(sess: &'a Session, reachable: &'a [String]) -> CodegenContext<'a> {
327 lto_ctxt: Some((sess, reachable)),
328 handler: sess.diagnostic().handler(),
329 remark: sess.opts.cg.remark.clone(),
334 struct HandlerFreeVars<'a> {
336 cgcx: &'a CodegenContext<'a>,
339 unsafe extern "C" fn report_inline_asm<'a, 'b>(cgcx: &'a CodegenContext<'a>,
342 use syntax::codemap::ExpnId;
344 match cgcx.lto_ctxt {
346 sess.codemap().with_expn_info(ExpnId::from_llvm_cookie(cookie), |info| match info {
347 Some(ei) => sess.span_err(ei.call_site, msg),
348 None => sess.err(msg),
353 cgcx.handler.err(msg);
354 cgcx.handler.note("build without -C codegen-units for more exact errors");
359 unsafe extern "C" fn inline_asm_handler(diag: SMDiagnosticRef,
362 let HandlerFreeVars { cgcx, .. }
363 = *mem::transmute::<_, *const HandlerFreeVars>(user);
365 let msg = llvm::build_string(|s| llvm::LLVMWriteSMDiagnosticToString(diag, s))
366 .expect("non-UTF8 SMDiagnostic");
368 report_inline_asm(cgcx, &msg[..], cookie);
371 unsafe extern "C" fn diagnostic_handler(info: DiagnosticInfoRef, user: *mut c_void) {
372 let HandlerFreeVars { llcx, cgcx }
373 = *mem::transmute::<_, *const HandlerFreeVars>(user);
375 match llvm::diagnostic::Diagnostic::unpack(info) {
376 llvm::diagnostic::InlineAsm(inline) => {
377 report_inline_asm(cgcx,
378 &*llvm::twine_to_string(inline.message),
382 llvm::diagnostic::Optimization(opt) => {
383 let pass_name = str::from_utf8(CStr::from_ptr(opt.pass_name).to_bytes())
385 .expect("got a non-UTF8 pass name from LLVM");
386 let enabled = match cgcx.remark {
388 SomePasses(ref v) => v.iter().any(|s| *s == pass_name),
392 let loc = llvm::debug_loc_to_string(llcx, opt.debug_loc);
393 cgcx.handler.note(&format!("optimization {} for {} at {}: {}",
396 if loc.is_empty() { "[unknown]" } else { &*loc },
397 llvm::twine_to_string(opt.message)));
405 // Unsafe due to LLVM calls.
406 unsafe fn optimize_and_codegen(cgcx: &CodegenContext,
407 mtrans: ModuleTranslation,
408 config: ModuleConfig,
410 output_names: OutputFilenames) {
411 let ModuleTranslation { llmod, llcx } = mtrans;
414 // llcx doesn't outlive this function, so we can put this on the stack.
415 let fv = HandlerFreeVars {
419 let fv = &fv as *const HandlerFreeVars as *mut c_void;
421 llvm::LLVMSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, fv);
422 llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, fv);
424 if config.emit_no_opt_bc {
425 let ext = format!("{}.no-opt.bc", name_extra);
426 let out = output_names.with_extension(&ext);
427 let out = CString::new(out.as_vec()).unwrap();
428 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
431 match config.opt_level {
433 // Create the two optimizing pass managers. These mirror what clang
434 // does, and are by populated by LLVM's default PassManagerBuilder.
435 // Each manager has a different set of passes, but they also share
436 // some common passes.
437 let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
438 let mpm = llvm::LLVMCreatePassManager();
440 // If we're verifying or linting, add them to the function pass
442 let addpass = |pass: &str| {
443 let pass = CString::new(pass).unwrap();
444 llvm::LLVMRustAddPass(fpm, pass.as_ptr())
446 if !config.no_verify { assert!(addpass("verify")); }
448 if !config.no_prepopulate_passes {
449 llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod);
450 llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod);
451 populate_llvm_passes(fpm, mpm, llmod, opt_level,
455 for pass in &config.passes {
456 let pass = CString::new(pass.clone()).unwrap();
457 if !llvm::LLVMRustAddPass(mpm, pass.as_ptr()) {
458 cgcx.handler.warn(&format!("unknown pass {:?}, ignoring", pass));
462 // Finally, run the actual optimization passes
463 time(config.time_passes, "llvm function passes", (), |()|
464 llvm::LLVMRustRunFunctionPassManager(fpm, llmod));
465 time(config.time_passes, "llvm module passes", (), |()|
466 llvm::LLVMRunPassManager(mpm, llmod));
468 // Deallocate managers that we're now done with
469 llvm::LLVMDisposePassManager(fpm);
470 llvm::LLVMDisposePassManager(mpm);
472 match cgcx.lto_ctxt {
473 Some((sess, reachable)) if sess.lto() => {
474 time(sess.time_passes(), "all lto passes", (), |()|
475 lto::run(sess, llmod, tm, reachable));
477 if config.emit_lto_bc {
478 let name = format!("{}.lto.bc", name_extra);
479 let out = output_names.with_extension(&name);
480 let out = CString::new(out.as_vec()).unwrap();
481 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
490 // A codegen-specific pass manager is used to generate object
491 // files for an LLVM module.
493 // Apparently each of these pass managers is a one-shot kind of
494 // thing, so we create a new one for each type of output. The
495 // pass manager passed to the closure should be ensured to not
496 // escape the closure itself, and the manager should only be
498 unsafe fn with_codegen<F>(tm: TargetMachineRef,
502 F: FnOnce(PassManagerRef),
504 let cpm = llvm::LLVMCreatePassManager();
505 llvm::LLVMRustAddAnalysisPasses(tm, cpm, llmod);
506 llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
508 llvm::LLVMDisposePassManager(cpm);
512 let ext = format!("{}.bc", name_extra);
513 let out = output_names.with_extension(&ext);
514 let out = CString::new(out.as_vec()).unwrap();
515 llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
518 time(config.time_passes, "codegen passes", (), |()| {
520 let ext = format!("{}.ll", name_extra);
521 let out = output_names.with_extension(&ext);
522 let out = CString::new(out.as_vec()).unwrap();
523 with_codegen(tm, llmod, config.no_builtins, |cpm| {
524 llvm::LLVMRustPrintModule(cpm, llmod, out.as_ptr());
529 let path = output_names.with_extension(&format!("{}.s", name_extra)[]);
530 with_codegen(tm, llmod, config.no_builtins, |cpm| {
531 write_output_file(cgcx.handler, tm, cpm, llmod, &path, llvm::AssemblyFileType);
536 let path = output_names.with_extension(&format!("{}.o", name_extra)[]);
537 with_codegen(tm, llmod, config.no_builtins, |cpm| {
538 write_output_file(cgcx.handler, tm, cpm, llmod, &path, llvm::ObjectFileType);
543 llvm::LLVMDisposeModule(llmod);
544 llvm::LLVMContextDispose(llcx);
545 llvm::LLVMRustDisposeTargetMachine(tm);
548 pub fn run_passes(sess: &Session,
549 trans: &CrateTranslation,
550 output_types: &[config::OutputType],
551 crate_output: &OutputFilenames) {
552 // It's possible that we have `codegen_units > 1` but only one item in
553 // `trans.modules`. We could theoretically proceed and do LTO in that
554 // case, but it would be confusing to have the validity of
555 // `-Z lto -C codegen-units=2` depend on details of the crate being
556 // compiled, so we complain regardless.
557 if sess.lto() && sess.opts.cg.codegen_units > 1 {
558 // This case is impossible to handle because LTO expects to be able
559 // to combine the entire crate and all its dependencies into a
560 // single compilation unit, but each codegen unit is in a separate
561 // LLVM context, so they can't easily be combined.
562 sess.fatal("can't perform LTO when using multiple codegen units");
566 assert!(trans.modules.len() == sess.opts.cg.codegen_units);
569 configure_llvm(sess);
572 let tm = create_target_machine(sess);
574 // Figure out what we actually need to build.
576 let mut modules_config = ModuleConfig::new(tm, sess.opts.cg.passes.clone());
577 let mut metadata_config = ModuleConfig::new(tm, vec!());
579 modules_config.opt_level = Some(get_llvm_opt_level(sess.opts.optimize));
581 // Save all versions of the bytecode if we're saving our temporaries.
582 if sess.opts.cg.save_temps {
583 modules_config.emit_no_opt_bc = true;
584 modules_config.emit_bc = true;
585 modules_config.emit_lto_bc = true;
586 metadata_config.emit_bc = true;
589 // Emit bitcode files for the crate if we're emitting an rlib.
590 // Whenever an rlib is created, the bitcode is inserted into the
591 // archive in order to allow LTO against it.
592 let needs_crate_bitcode =
593 sess.crate_types.borrow().contains(&config::CrateTypeRlib) &&
594 sess.opts.output_types.contains(&config::OutputTypeExe);
595 if needs_crate_bitcode {
596 modules_config.emit_bc = true;
599 for output_type in output_types {
601 config::OutputTypeBitcode => { modules_config.emit_bc = true; },
602 config::OutputTypeLlvmAssembly => { modules_config.emit_ir = true; },
603 config::OutputTypeAssembly => {
604 modules_config.emit_asm = true;
605 // If we're not using the LLVM assembler, this function
606 // could be invoked specially with output_type_assembly, so
607 // in this case we still want the metadata object file.
608 if !sess.opts.output_types.contains(&config::OutputTypeAssembly) {
609 metadata_config.emit_obj = true;
612 config::OutputTypeObject => { modules_config.emit_obj = true; },
613 config::OutputTypeExe => {
614 modules_config.emit_obj = true;
615 metadata_config.emit_obj = true;
617 config::OutputTypeDepInfo => {}
621 modules_config.set_flags(sess, trans);
622 metadata_config.set_flags(sess, trans);
625 // Populate a buffer with a list of codegen tasks. Items are processed in
626 // LIFO order, just because it's a tiny bit simpler that way. (The order
627 // doesn't actually matter.)
628 let mut work_items = Vec::with_capacity(1 + trans.modules.len());
631 let work = build_work_item(sess,
632 trans.metadata_module,
633 metadata_config.clone(),
634 crate_output.clone(),
635 "metadata".to_string());
636 work_items.push(work);
639 for (index, mtrans) in trans.modules.iter().enumerate() {
640 let work = build_work_item(sess,
642 modules_config.clone(),
643 crate_output.clone(),
644 format!("{}", index));
645 work_items.push(work);
648 // Process the work items, optionally using worker threads.
649 if sess.opts.cg.codegen_units == 1 {
650 run_work_singlethreaded(sess, &trans.reachable[], work_items);
652 run_work_multithreaded(sess, work_items, sess.opts.cg.codegen_units);
655 // All codegen is finished.
657 llvm::LLVMRustDisposeTargetMachine(tm);
660 // Produce final compile outputs.
661 let copy_gracefully = |from: &Path, to: &Path| {
662 if let Err(e) = fs::copy(from, to) {
663 sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
667 let copy_if_one_unit = |ext: &str, output_type: config::OutputType, keep_numbered: bool| {
669 if sess.opts.cg.codegen_units == 1 {
670 // 1) Only one codegen unit. In this case it's no difficulty
671 // to copy `foo.0.x` to `foo.x`.
672 copy_gracefully(&crate_output.with_extension(ext), &crate_output.path(output_type));
673 if !sess.opts.cg.save_temps && !keep_numbered {
674 // The user just wants `foo.x`, not `foo.0.x`.
675 remove(sess, &crate_output.with_extension(ext));
678 if crate_output.single_output_file.is_some() {
679 // 2) Multiple codegen units, with `-o some_name`. We have
680 // no good solution for this case, so warn the user.
681 sess.warn(&format!("ignoring -o because multiple .{} files were produced",
684 // 3) Multiple codegen units, but no `-o some_name`. We
685 // just leave the `foo.0.x` files in place.
686 // (We don't have to do any work in this case.)
691 let link_obj = |output_path: &Path| {
692 // Running `ld -r` on a single input is kind of pointless.
693 if sess.opts.cg.codegen_units == 1 {
694 copy_gracefully(&crate_output.with_extension("0.o"), output_path);
695 // Leave the .0.o file around, to mimic the behavior of the normal
700 // Some builds of MinGW GCC will pass --force-exe-suffix to ld, which
701 // will automatically add a .exe extension if the extension is not
702 // already .exe or .dll. To ensure consistent behavior on Windows, we
703 // add the .exe suffix explicitly and then rename the output file to
704 // the desired path. This will give the correct behavior whether or
705 // not GCC adds --force-exe-suffix.
706 let windows_output_path =
707 if sess.target.target.options.is_like_windows {
708 Some(output_path.with_extension("o.exe"))
713 let pname = get_cc_prog(sess);
714 let mut cmd = Command::new(&pname[..]);
716 cmd.args(&sess.target.target.options.pre_link_args[]);
717 cmd.arg("-nostdlib");
719 for index in 0..trans.modules.len() {
720 cmd.arg(crate_output.with_extension(&format!("{}.o", index)[]));
725 .arg(windows_output_path.as_ref().unwrap_or(output_path));
727 cmd.args(&sess.target.target.options.post_link_args[]);
729 if sess.opts.debugging_opts.print_link_args {
730 println!("{:?}", &cmd);
733 cmd.stdin(::std::old_io::process::Ignored)
734 .stdout(::std::old_io::process::InheritFd(1))
735 .stderr(::std::old_io::process::InheritFd(2));
738 if !status.success() {
739 sess.err(&format!("linking of {} with `{:?}` failed",
740 output_path.display(), cmd)[]);
741 sess.abort_if_errors();
745 sess.err(&format!("could not exec the linker `{}`: {}",
748 sess.abort_if_errors();
752 match windows_output_path {
753 Some(ref windows_path) => {
754 fs::rename(windows_path, output_path).unwrap();
757 // The file is already named according to `output_path`.
762 // Flag to indicate whether the user explicitly requested bitcode.
763 // Otherwise, we produced it only as a temporary output, and will need
765 let mut user_wants_bitcode = false;
766 for output_type in output_types {
768 config::OutputTypeBitcode => {
769 user_wants_bitcode = true;
770 // Copy to .bc, but always keep the .0.bc. There is a later
771 // check to figure out if we should delete .0.bc files, or keep
772 // them for making an rlib.
773 copy_if_one_unit("0.bc", config::OutputTypeBitcode, true);
775 config::OutputTypeLlvmAssembly => {
776 copy_if_one_unit("0.ll", config::OutputTypeLlvmAssembly, false);
778 config::OutputTypeAssembly => {
779 copy_if_one_unit("0.s", config::OutputTypeAssembly, false);
781 config::OutputTypeObject => {
782 link_obj(&crate_output.path(config::OutputTypeObject));
784 config::OutputTypeExe => {
785 // If config::OutputTypeObject is already in the list, then
786 // `crate.o` will be handled by the config::OutputTypeObject case.
787 // Otherwise, we need to create the temporary object so we
788 // can run the linker.
789 if !sess.opts.output_types.contains(&config::OutputTypeObject) {
790 link_obj(&crate_output.temp_path(config::OutputTypeObject));
793 config::OutputTypeDepInfo => {}
796 let user_wants_bitcode = user_wants_bitcode;
798 // Clean up unwanted temporary files.
800 // We create the following files by default:
803 // - crate.metadata.bc
804 // - crate.metadata.o
805 // - crate.o (linked from crate.##.o)
806 // - crate.bc (copied from crate.0.bc)
807 // We may create additional files if requested by the user (through
808 // `-C save-temps` or `--emit=` flags).
810 if !sess.opts.cg.save_temps {
811 // Remove the temporary .0.o objects. If the user didn't
812 // explicitly request bitcode (with --emit=bc), and the bitcode is not
813 // needed for building an rlib, then we must remove .0.bc as well.
815 // Specific rules for keeping .0.bc:
816 // - If we're building an rlib (`needs_crate_bitcode`), then keep
818 // - If the user requested bitcode (`user_wants_bitcode`), and
819 // codegen_units > 1, then keep it.
820 // - If the user requested bitcode but codegen_units == 1, then we
821 // can toss .0.bc because we copied it to .bc earlier.
822 // - If we're not building an rlib and the user didn't request
823 // bitcode, then delete .0.bc.
824 // If you change how this works, also update back::link::link_rlib,
825 // where .0.bc files are (maybe) deleted after making an rlib.
826 let keep_numbered_bitcode = needs_crate_bitcode ||
827 (user_wants_bitcode && sess.opts.cg.codegen_units > 1);
829 for i in 0..trans.modules.len() {
830 if modules_config.emit_obj {
831 let ext = format!("{}.o", i);
832 remove(sess, &crate_output.with_extension(&ext[..]));
835 if modules_config.emit_bc && !keep_numbered_bitcode {
836 let ext = format!("{}.bc", i);
837 remove(sess, &crate_output.with_extension(&ext[..]));
841 if metadata_config.emit_bc && !user_wants_bitcode {
842 remove(sess, &crate_output.with_extension("metadata.bc"));
846 // We leave the following files around by default:
848 // - crate.metadata.o
850 // These are used in linking steps and will be cleaned up afterward.
852 // FIXME: time_llvm_passes support - does this use a global context or
854 //if sess.time_llvm_passes() { llvm::LLVMRustPrintPassTimings(); }
858 mtrans: ModuleTranslation,
859 config: ModuleConfig,
860 output_names: OutputFilenames,
864 fn build_work_item(sess: &Session,
865 mtrans: ModuleTranslation,
866 config: ModuleConfig,
867 output_names: OutputFilenames,
871 let mut config = config;
872 config.tm = create_target_machine(sess);
873 WorkItem { mtrans: mtrans, config: config, output_names: output_names,
874 name_extra: name_extra }
877 fn execute_work_item(cgcx: &CodegenContext,
878 work_item: WorkItem) {
880 optimize_and_codegen(cgcx, work_item.mtrans, work_item.config,
881 work_item.name_extra, work_item.output_names);
885 fn run_work_singlethreaded(sess: &Session,
886 reachable: &[String],
887 work_items: Vec<WorkItem>) {
888 let cgcx = CodegenContext::new_with_session(sess, reachable);
889 let mut work_items = work_items;
891 // Since we're running single-threaded, we can pass the session to
892 // the proc, allowing `optimize_and_codegen` to perform LTO.
893 for work in Unfold::new((), |_| work_items.pop()) {
894 execute_work_item(&cgcx, work);
898 fn run_work_multithreaded(sess: &Session,
899 work_items: Vec<WorkItem>,
901 // Run some workers to process the work items.
902 let work_items_arc = Arc::new(Mutex::new(work_items));
903 let mut diag_emitter = SharedEmitter::new();
904 let mut futures = Vec::with_capacity(num_workers);
906 for i in 0..num_workers {
907 let work_items_arc = work_items_arc.clone();
908 let diag_emitter = diag_emitter.clone();
909 let remark = sess.opts.cg.remark.clone();
911 let (tx, rx) = channel();
912 let mut tx = Some(tx);
915 thread::Builder::new().name(format!("codegen-{}", i)).spawn(move || {
916 let diag_handler = mk_handler(true, box diag_emitter);
918 // Must construct cgcx inside the proc because it has non-Send
920 let cgcx = CodegenContext {
922 handler: &diag_handler,
927 // Avoid holding the lock for the entire duration of the match.
928 let maybe_work = work_items_arc.lock().unwrap().pop();
931 execute_work_item(&cgcx, work);
933 // Make sure to fail the worker so the main thread can
934 // tell that there were errors.
935 cgcx.handler.abort_if_errors();
941 tx.take().unwrap().send(()).unwrap();
945 let mut panicked = false;
953 // Display any new diagnostics.
954 diag_emitter.dump(sess.diagnostic().handler());
957 sess.fatal("aborting due to worker thread panic");
961 pub fn run_assembler(sess: &Session, outputs: &OutputFilenames) {
962 let pname = get_cc_prog(sess);
963 let mut cmd = Command::new(&pname[..]);
965 cmd.arg("-c").arg("-o").arg(outputs.path(config::OutputTypeObject))
966 .arg(outputs.temp_path(config::OutputTypeAssembly));
967 debug!("{:?}", &cmd);
971 if !prog.status.success() {
972 sess.err(&format!("linking with `{}` failed: {}",
975 sess.note(&format!("{:?}", &cmd)[]);
976 let mut note = prog.error.clone();
977 note.push_all(&prog.output[]);
978 sess.note(str::from_utf8(¬e[..]).unwrap());
979 sess.abort_if_errors();
983 sess.err(&format!("could not exec the linker `{}`: {}",
986 sess.abort_if_errors();
991 unsafe fn configure_llvm(sess: &Session) {
992 use std::sync::{Once, ONCE_INIT};
993 static INIT: Once = ONCE_INIT;
995 // Copy what clang does by turning on loop vectorization at O2 and
996 // slp vectorization at O3
997 let vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
998 (sess.opts.optimize == config::Default ||
999 sess.opts.optimize == config::Aggressive);
1000 let vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
1001 sess.opts.optimize == config::Aggressive;
1003 let mut llvm_c_strs = Vec::new();
1004 let mut llvm_args = Vec::new();
1006 let mut add = |arg: &str| {
1007 let s = CString::new(arg).unwrap();
1008 llvm_args.push(s.as_ptr());
1009 llvm_c_strs.push(s);
1011 add("rustc"); // fake program name
1012 if vectorize_loop { add("-vectorize-loops"); }
1013 if vectorize_slp { add("-vectorize-slp"); }
1014 if sess.time_llvm_passes() { add("-time-passes"); }
1015 if sess.print_llvm_passes() { add("-debug-pass=Structure"); }
1017 // FIXME #21627 disable faulty FastISel on AArch64 (even for -O0)
1018 if sess.target.target.arch == "aarch64" { add("-fast-isel=0"); }
1020 for arg in &sess.opts.cg.llvm_args {
1026 llvm::LLVMInitializePasses();
1028 // Only initialize the platforms supported by Rust here, because
1029 // using --llvm-root will have multiple platforms that rustllvm
1030 // doesn't actually link to and it's pointless to put target info
1031 // into the registry that Rust cannot generate machine code for.
1032 llvm::LLVMInitializeX86TargetInfo();
1033 llvm::LLVMInitializeX86Target();
1034 llvm::LLVMInitializeX86TargetMC();
1035 llvm::LLVMInitializeX86AsmPrinter();
1036 llvm::LLVMInitializeX86AsmParser();
1038 llvm::LLVMInitializeARMTargetInfo();
1039 llvm::LLVMInitializeARMTarget();
1040 llvm::LLVMInitializeARMTargetMC();
1041 llvm::LLVMInitializeARMAsmPrinter();
1042 llvm::LLVMInitializeARMAsmParser();
1044 llvm::LLVMInitializeAArch64TargetInfo();
1045 llvm::LLVMInitializeAArch64Target();
1046 llvm::LLVMInitializeAArch64TargetMC();
1047 llvm::LLVMInitializeAArch64AsmPrinter();
1048 llvm::LLVMInitializeAArch64AsmParser();
1050 llvm::LLVMInitializeMipsTargetInfo();
1051 llvm::LLVMInitializeMipsTarget();
1052 llvm::LLVMInitializeMipsTargetMC();
1053 llvm::LLVMInitializeMipsAsmPrinter();
1054 llvm::LLVMInitializeMipsAsmParser();
1056 llvm::LLVMInitializePowerPCTargetInfo();
1057 llvm::LLVMInitializePowerPCTarget();
1058 llvm::LLVMInitializePowerPCTargetMC();
1059 llvm::LLVMInitializePowerPCAsmPrinter();
1060 llvm::LLVMInitializePowerPCAsmParser();
1062 llvm::LLVMRustSetLLVMOptions(llvm_args.len() as c_int,
1063 llvm_args.as_ptr());
1067 unsafe fn populate_llvm_passes(fpm: llvm::PassManagerRef,
1068 mpm: llvm::PassManagerRef,
1070 opt: llvm::CodeGenOptLevel,
1071 no_builtins: bool) {
1072 // Create the PassManagerBuilder for LLVM. We configure it with
1073 // reasonable defaults and prepare it to actually populate the pass
1075 let builder = llvm::LLVMPassManagerBuilderCreate();
1077 llvm::CodeGenLevelNone => {
1078 // Don't add lifetime intrinsics at O0
1079 llvm::LLVMRustAddAlwaysInlinePass(builder, false);
1081 llvm::CodeGenLevelLess => {
1082 llvm::LLVMRustAddAlwaysInlinePass(builder, true);
1084 // numeric values copied from clang
1085 llvm::CodeGenLevelDefault => {
1086 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder,
1089 llvm::CodeGenLevelAggressive => {
1090 llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder,
1094 llvm::LLVMPassManagerBuilderSetOptLevel(builder, opt as c_uint);
1095 llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, no_builtins);
1097 // Use the builder to populate the function/module pass managers.
1098 llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(builder, fpm);
1099 llvm::LLVMPassManagerBuilderPopulateModulePassManager(builder, mpm);
1100 llvm::LLVMPassManagerBuilderDispose(builder);
1103 llvm::CodeGenLevelDefault | llvm::CodeGenLevelAggressive => {
1104 llvm::LLVMRustAddPass(mpm, "mergefunc\0".as_ptr() as *const _);