src/librustc_trans/back/write.rs

   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.
   4 //
   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.
  10
  11 use back::lto;
  12 use back::link::{get_cc_prog, remove};
  13 use session::config::{OutputFilenames, NoDebugInfo, Passes, SomePasses, AllPasses};
  14 use session::Session;
  15 use session::config;
  16 use llvm;
  17 use llvm::{ModuleRef, TargetMachineRef, PassManagerRef, DiagnosticInfoRef, ContextRef};
  18 use llvm::SMDiagnosticRef;
  19 use trans::{CrateTranslation, ModuleTranslation};
  20 use util::common::time;
  21 use syntax::codemap;
  22 use syntax::diagnostic;
  23 use syntax::diagnostic::{Emitter, Handler, Level, mk_handler};
  24
  25 use std::ffi::{self, CString};
  26 use std::io::Command;
  27 use std::io::fs;
  28 use std::iter::Unfold;
  29 use std::ptr;
  30 use std::str;
  31 use std::mem;
  32 use std::sync::{Arc, Mutex};
  33 use std::sync::mpsc::channel;
  34 use std::thread;
  35 use libc::{self, c_uint, c_int, c_void};
  36
  37 #[derive(Clone, Copy, PartialEq, PartialOrd, Ord, Eq)]
  38 pub enum OutputType {
  39     OutputTypeBitcode,
  40     OutputTypeAssembly,
  41     OutputTypeLlvmAssembly,
  42     OutputTypeObject,
  43     OutputTypeExe,
  44 }
  45
  46 pub fn llvm_err(handler: &diagnostic::Handler, msg: String) -> ! {
  47     unsafe {
  48         let cstr = llvm::LLVMRustGetLastError();
  49         if cstr == ptr::null() {
  50             handler.fatal(&msg[]);
  51         } else {
  52             let err = ffi::c_str_to_bytes(&cstr);
  53             let err = String::from_utf8_lossy(err.as_slice()).to_string();
  54             libc::free(cstr as *mut _);
  55             handler.fatal(&format!("{}: {}",
  56                                   &msg[],
  57                                   &err[])[]);
  58         }
  59     }
  60 }
  61
  62 pub fn write_output_file(
  63         handler: &diagnostic::Handler,
  64         target: llvm::TargetMachineRef,
  65         pm: llvm::PassManagerRef,
  66         m: ModuleRef,
  67         output: &Path,
  68         file_type: llvm::FileType) {
  69     unsafe {
  70         let output = CString::from_slice(output.as_vec());
  71         let result = llvm::LLVMRustWriteOutputFile(
  72                 target, pm, m, output.as_ptr(), file_type);
  73         if !result {
  74             llvm_err(handler, "could not write output".to_string());
  75         }
  76     }
  77 }
  78
  79
  80 struct Diagnostic {
  81     msg: String,
  82     code: Option<String>,
  83     lvl: Level,
  84 }
  85
  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).
  89 #[derive(Clone)]
  90 struct SharedEmitter {
  91     buffer: Arc<Mutex<Vec<Diagnostic>>>,
  92 }
  93
  94 impl SharedEmitter {
  95     fn new() -> SharedEmitter {
  96         SharedEmitter {
  97             buffer: Arc::new(Mutex::new(Vec::new())),
  98         }
  99     }
 100
 101     fn dump(&mut self, handler: &Handler) {
 102         let mut buffer = self.buffer.lock().unwrap();
 103         for diag in buffer.iter() {
 104             match diag.code {
 105                 Some(ref code) => {
 106                     handler.emit_with_code(None,
 107                                            &diag.msg[],
 108                                            &code[],
 109                                            diag.lvl);
 110                 },
 111                 None => {
 112                     handler.emit(None,
 113                                  &diag.msg[],
 114                                  diag.lvl);
 115                 },
 116             }
 117         }
 118         buffer.clear();
 119     }
 120 }
 121
 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");
 126
 127         self.buffer.lock().unwrap().push(Diagnostic {
 128             msg: msg.to_string(),
 129             code: code.map(|s| s.to_string()),
 130             lvl: lvl,
 131         });
 132     }
 133
 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");
 137     }
 138 }
 139
 140
 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.
 146 //
 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)
 155 }
 156
 157 fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel {
 158     match optimize {
 159       config::No => llvm::CodeGenLevelNone,
 160       config::Less => llvm::CodeGenLevelLess,
 161       config::Default => llvm::CodeGenLevelDefault,
 162       config::Aggressive => llvm::CodeGenLevelAggressive,
 163     }
 164 }
 165
 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[]
 170     };
 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,
 176         _ => {
 177             sess.err(&format!("{:?} is not a valid relocation mode",
 178                              sess.opts
 179                                  .cg
 180                                  .relocation_model)[]);
 181             sess.abort_if_errors();
 182             unreachable!();
 183         }
 184     };
 185
 186     let opt_level = get_llvm_opt_level(sess.opts.optimize);
 187     let use_softfp = sess.opts.cg.soft_float;
 188
 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;
 192
 193     let any_library = sess.crate_types.borrow().iter().any(|ty| {
 194         *ty != config::CrateTypeExecutable
 195     });
 196
 197     let ffunction_sections = sess.target.target.options.function_sections;
 198     let fdata_sections = ffunction_sections;
 199
 200     let code_model_arg = match sess.opts.cg.code_model {
 201         Some(ref s) => &s[],
 202         None => &sess.target.target.options.code_model[]
 203     };
 204
 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,
 211         _ => {
 212             sess.err(&format!("{:?} is not a valid code model",
 213                              sess.opts
 214                                  .cg
 215                                  .code_model)[]);
 216             sess.abort_if_errors();
 217             unreachable!();
 218         }
 219     };
 220
 221     let triple = &sess.target.target.llvm_target[];
 222
 223     let tm = unsafe {
 224         let triple = CString::from_slice(triple.as_bytes());
 225         let cpu = match sess.opts.cg.target_cpu {
 226             Some(ref s) => s.as_slice(),
 227             None => sess.target.target.options.cpu.as_slice()
 228         };
 229         let cpu = CString::from_slice(cpu.as_bytes());
 230         let features = CString::from_slice(target_feature(sess).as_bytes());
 231         llvm::LLVMRustCreateTargetMachine(
 232             triple.as_ptr(), cpu.as_ptr(), features.as_ptr(),
 233             code_model,
 234             reloc_model,
 235             opt_level,
 236             true /* EnableSegstk */,
 237             use_softfp,
 238             no_fp_elim,
 239             !any_library && reloc_model == llvm::RelocPIC,
 240             ffunction_sections,
 241             fdata_sections,
 242         )
 243     };
 244
 245     if tm.is_null() {
 246         llvm_err(sess.diagnostic().handler(),
 247                  format!("Could not create LLVM TargetMachine for triple: {}",
 248                          triple).to_string());
 249     } else {
 250         return tm;
 251     };
 252 }
 253
 254
 255 /// Module-specific configuration for `optimize_and_codegen`.
 256 #[derive(Clone)]
 257 struct ModuleConfig {
 258     /// LLVM TargetMachine to use for codegen.
 259     tm: TargetMachineRef,
 260     /// Names of additional optimization passes to run.
 261     passes: Vec<String>,
 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>,
 265
 266     // Flags indicating which outputs to produce.
 267     emit_no_opt_bc: bool,
 268     emit_bc: bool,
 269     emit_lto_bc: bool,
 270     emit_ir: bool,
 271     emit_asm: bool,
 272     emit_obj: bool,
 273
 274     // Miscellaneous flags.  These are mostly copied from command-line
 275     // options.
 276     no_verify: bool,
 277     no_prepopulate_passes: bool,
 278     no_builtins: bool,
 279     time_passes: bool,
 280 }
 281
 282 unsafe impl Send for ModuleConfig { }
 283
 284 impl ModuleConfig {
 285     fn new(tm: TargetMachineRef, passes: Vec<String>) -> ModuleConfig {
 286         ModuleConfig {
 287             tm: tm,
 288             passes: passes,
 289             opt_level: None,
 290
 291             emit_no_opt_bc: false,
 292             emit_bc: false,
 293             emit_lto_bc: false,
 294             emit_ir: false,
 295             emit_asm: false,
 296             emit_obj: false,
 297
 298             no_verify: false,
 299             no_prepopulate_passes: false,
 300             no_builtins: false,
 301             time_passes: false,
 302         }
 303     }
 304
 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     }
 311 }
 312
 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.
 321     remark: Passes,
 322 }
 323
 324 impl<'a> CodegenContext<'a> {
 325     fn new_with_session(sess: &'a Session, reachable: &'a [String]) -> CodegenContext<'a> {
 326         CodegenContext {
 327             lto_ctxt: Some((sess, reachable)),
 328             handler: sess.diagnostic().handler(),
 329             remark: sess.opts.cg.remark.clone(),
 330         }
 331     }
 332 }
 333
 334 struct HandlerFreeVars<'a> {
 335     llcx: ContextRef,
 336     cgcx: &'a CodegenContext<'a>,
 337 }
 338
 339 unsafe extern "C" fn inline_asm_handler(diag: SMDiagnosticRef,
 340                                         user: *const c_void,
 341                                         cookie: c_uint) {
 342     use syntax::codemap::ExpnId;
 343
 344     let HandlerFreeVars { cgcx, .. }
 345         = *mem::transmute::<_, *const HandlerFreeVars>(user);
 346
 347     let msg = llvm::build_string(|s| llvm::LLVMWriteSMDiagnosticToString(diag, s))
 348         .expect("non-UTF8 SMDiagnostic");
 349
 350     match cgcx.lto_ctxt {
 351         Some((sess, _)) => {
 352             sess.codemap().with_expn_info(ExpnId::from_llvm_cookie(cookie), |info| match info {
 353                 Some(ei) => sess.span_err(ei.call_site, &msg[]),
 354                 None     => sess.err(&msg[]),
 355             });
 356         }
 357
 358         None => {
 359             cgcx.handler.err(&msg[]);
 360             cgcx.handler.note("build without -C codegen-units for more exact errors");
 361         }
 362     }
 363 }
 364
 365 unsafe extern "C" fn diagnostic_handler(info: DiagnosticInfoRef, user: *mut c_void) {
 366     let HandlerFreeVars { llcx, cgcx }
 367         = *mem::transmute::<_, *const HandlerFreeVars>(user);
 368
 369     match llvm::diagnostic::Diagnostic::unpack(info) {
 370         llvm::diagnostic::Optimization(opt) => {
 371             let pass_name = str::from_utf8(ffi::c_str_to_bytes(&opt.pass_name))
 372                                 .ok()
 373                                 .expect("got a non-UTF8 pass name from LLVM");
 374             let enabled = match cgcx.remark {
 375                 AllPasses => true,
 376                 SomePasses(ref v) => v.iter().any(|s| *s == pass_name),
 377             };
 378
 379             if enabled {
 380                 let loc = llvm::debug_loc_to_string(llcx, opt.debug_loc);
 381                 cgcx.handler.note(format!("optimization {} for {} at {}: {}",
 382                                           opt.kind.describe(),
 383                                           pass_name,
 384                                           if loc.is_empty() { "[unknown]" } else { loc.as_slice() },
 385                                           llvm::twine_to_string(opt.message)).as_slice());
 386             }
 387         }
 388
 389         _ => (),
 390     }
 391 }
 392
 393 // Unsafe due to LLVM calls.
 394 unsafe fn optimize_and_codegen(cgcx: &CodegenContext,
 395                                mtrans: ModuleTranslation,
 396                                config: ModuleConfig,
 397                                name_extra: String,
 398                                output_names: OutputFilenames) {
 399     let ModuleTranslation { llmod, llcx } = mtrans;
 400     let tm = config.tm;
 401
 402     // llcx doesn't outlive this function, so we can put this on the stack.
 403     let fv = HandlerFreeVars {
 404         llcx: llcx,
 405         cgcx: cgcx,
 406     };
 407     let fv = &fv as *const HandlerFreeVars as *mut c_void;
 408
 409     llvm::LLVMSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, fv);
 410
 411     if !cgcx.remark.is_empty() {
 412         llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, fv);
 413     }
 414
 415     if config.emit_no_opt_bc {
 416         let ext = format!("{}.no-opt.bc", name_extra);
 417         let out = output_names.with_extension(ext.as_slice());
 418         let out = CString::from_slice(out.as_vec());
 419         llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
 420     }
 421
 422     match config.opt_level {
 423         Some(opt_level) => {
 424             // Create the two optimizing pass managers. These mirror what clang
 425             // does, and are by populated by LLVM's default PassManagerBuilder.
 426             // Each manager has a different set of passes, but they also share
 427             // some common passes.
 428             let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
 429             let mpm = llvm::LLVMCreatePassManager();
 430
 431             // If we're verifying or linting, add them to the function pass
 432             // manager.
 433             let addpass = |&: pass: &str| {
 434                 let pass = CString::from_slice(pass.as_bytes());
 435                 llvm::LLVMRustAddPass(fpm, pass.as_ptr())
 436             };
 437             if !config.no_verify { assert!(addpass("verify")); }
 438
 439             if !config.no_prepopulate_passes {
 440                 llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod);
 441                 llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod);
 442                 populate_llvm_passes(fpm, mpm, llmod, opt_level,
 443                                      config.no_builtins);
 444             }
 445
 446             for pass in config.passes.iter() {
 447                 let pass = CString::from_slice(pass.as_bytes());
 448                 if !llvm::LLVMRustAddPass(mpm, pass.as_ptr()) {
 449                     cgcx.handler.warn(format!("unknown pass {:?}, ignoring",
 450                                               pass).as_slice());
 451                 }
 452             }
 453
 454             // Finally, run the actual optimization passes
 455             time(config.time_passes, "llvm function passes", (), |()|
 456                  llvm::LLVMRustRunFunctionPassManager(fpm, llmod));
 457             time(config.time_passes, "llvm module passes", (), |()|
 458                  llvm::LLVMRunPassManager(mpm, llmod));
 459
 460             // Deallocate managers that we're now done with
 461             llvm::LLVMDisposePassManager(fpm);
 462             llvm::LLVMDisposePassManager(mpm);
 463
 464             match cgcx.lto_ctxt {
 465                 Some((sess, reachable)) if sess.lto() =>  {
 466                     time(sess.time_passes(), "all lto passes", (), |()|
 467                          lto::run(sess, llmod, tm, reachable));
 468
 469                     if config.emit_lto_bc {
 470                         let name = format!("{}.lto.bc", name_extra);
 471                         let out = output_names.with_extension(name.as_slice());
 472                         let out = CString::from_slice(out.as_vec());
 473                         llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
 474                     }
 475                 },
 476                 _ => {},
 477             }
 478         },
 479         None => {},
 480     }
 481
 482     // A codegen-specific pass manager is used to generate object
 483     // files for an LLVM module.
 484     //
 485     // Apparently each of these pass managers is a one-shot kind of
 486     // thing, so we create a new one for each type of output. The
 487     // pass manager passed to the closure should be ensured to not
 488     // escape the closure itself, and the manager should only be
 489     // used once.
 490     unsafe fn with_codegen<F>(tm: TargetMachineRef,
 491                               llmod: ModuleRef,
 492                               no_builtins: bool,
 493                               f: F) where
 494         F: FnOnce(PassManagerRef),
 495     {
 496         let cpm = llvm::LLVMCreatePassManager();
 497         llvm::LLVMRustAddAnalysisPasses(tm, cpm, llmod);
 498         llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
 499         f(cpm);
 500         llvm::LLVMDisposePassManager(cpm);
 501     }
 502
 503     if config.emit_bc {
 504         let ext = format!("{}.bc", name_extra);
 505         let out = output_names.with_extension(ext.as_slice());
 506         let out = CString::from_slice(out.as_vec());
 507         llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
 508     }
 509
 510     time(config.time_passes, "codegen passes", (), |()| {
 511         if config.emit_ir {
 512             let ext = format!("{}.ll", name_extra);
 513             let out = output_names.with_extension(ext.as_slice());
 514             let out = CString::from_slice(out.as_vec());
 515             with_codegen(tm, llmod, config.no_builtins, |cpm| {
 516                 llvm::LLVMRustPrintModule(cpm, llmod, out.as_ptr());
 517             })
 518         }
 519
 520         if config.emit_asm {
 521             let path = output_names.with_extension(&format!("{}.s", name_extra)[]);
 522             with_codegen(tm, llmod, config.no_builtins, |cpm| {
 523                 write_output_file(cgcx.handler, tm, cpm, llmod, &path, llvm::AssemblyFileType);
 524             });
 525         }
 526
 527         if config.emit_obj {
 528             let path = output_names.with_extension(&format!("{}.o", name_extra)[]);
 529             with_codegen(tm, llmod, config.no_builtins, |cpm| {
 530                 write_output_file(cgcx.handler, tm, cpm, llmod, &path, llvm::ObjectFileType);
 531             });
 532         }
 533     });
 534
 535     llvm::LLVMDisposeModule(llmod);
 536     llvm::LLVMContextDispose(llcx);
 537     llvm::LLVMRustDisposeTargetMachine(tm);
 538 }
 539
 540 pub fn run_passes(sess: &Session,
 541                   trans: &CrateTranslation,
 542                   output_types: &[config::OutputType],
 543                   crate_output: &OutputFilenames) {
 544     // It's possible that we have `codegen_units > 1` but only one item in
 545     // `trans.modules`.  We could theoretically proceed and do LTO in that
 546     // case, but it would be confusing to have the validity of
 547     // `-Z lto -C codegen-units=2` depend on details of the crate being
 548     // compiled, so we complain regardless.
 549     if sess.lto() && sess.opts.cg.codegen_units > 1 {
 550         // This case is impossible to handle because LTO expects to be able
 551         // to combine the entire crate and all its dependencies into a
 552         // single compilation unit, but each codegen unit is in a separate
 553         // LLVM context, so they can't easily be combined.
 554         sess.fatal("can't perform LTO when using multiple codegen units");
 555     }
 556
 557     // Sanity check
 558     assert!(trans.modules.len() == sess.opts.cg.codegen_units);
 559
 560     unsafe {
 561         configure_llvm(sess);
 562     }
 563
 564     let tm = create_target_machine(sess);
 565
 566     // Figure out what we actually need to build.
 567
 568     let mut modules_config = ModuleConfig::new(tm, sess.opts.cg.passes.clone());
 569     let mut metadata_config = ModuleConfig::new(tm, vec!());
 570
 571     modules_config.opt_level = Some(get_llvm_opt_level(sess.opts.optimize));
 572
 573     // Save all versions of the bytecode if we're saving our temporaries.
 574     if sess.opts.cg.save_temps {
 575         modules_config.emit_no_opt_bc = true;
 576         modules_config.emit_bc = true;
 577         modules_config.emit_lto_bc = true;
 578         metadata_config.emit_bc = true;
 579     }
 580
 581     // Emit bitcode files for the crate if we're emitting an rlib.
 582     // Whenever an rlib is created, the bitcode is inserted into the
 583     // archive in order to allow LTO against it.
 584     let needs_crate_bitcode =
 585             sess.crate_types.borrow().contains(&config::CrateTypeRlib) &&
 586             sess.opts.output_types.contains(&config::OutputTypeExe);
 587     if needs_crate_bitcode {
 588         modules_config.emit_bc = true;
 589     }
 590
 591     for output_type in output_types.iter() {
 592         match *output_type {
 593             config::OutputTypeBitcode => { modules_config.emit_bc = true; },
 594             config::OutputTypeLlvmAssembly => { modules_config.emit_ir = true; },
 595             config::OutputTypeAssembly => {
 596                 modules_config.emit_asm = true;
 597                 // If we're not using the LLVM assembler, this function
 598                 // could be invoked specially with output_type_assembly, so
 599                 // in this case we still want the metadata object file.
 600                 if !sess.opts.output_types.contains(&config::OutputTypeAssembly) {
 601                     metadata_config.emit_obj = true;
 602                 }
 603             },
 604             config::OutputTypeObject => { modules_config.emit_obj = true; },
 605             config::OutputTypeExe => {
 606                 modules_config.emit_obj = true;
 607                 metadata_config.emit_obj = true;
 608             },
 609             config::OutputTypeDepInfo => {}
 610         }
 611     }
 612
 613     modules_config.set_flags(sess, trans);
 614     metadata_config.set_flags(sess, trans);
 615
 616
 617     // Populate a buffer with a list of codegen tasks.  Items are processed in
 618     // LIFO order, just because it's a tiny bit simpler that way.  (The order
 619     // doesn't actually matter.)
 620     let mut work_items = Vec::with_capacity(1 + trans.modules.len());
 621
 622     {
 623         let work = build_work_item(sess,
 624                                    trans.metadata_module,
 625                                    metadata_config.clone(),
 626                                    crate_output.clone(),
 627                                    "metadata".to_string());
 628         work_items.push(work);
 629     }
 630
 631     for (index, mtrans) in trans.modules.iter().enumerate() {
 632         let work = build_work_item(sess,
 633                                    *mtrans,
 634                                    modules_config.clone(),
 635                                    crate_output.clone(),
 636                                    format!("{}", index));
 637         work_items.push(work);
 638     }
 639
 640     // Process the work items, optionally using worker threads.
 641     if sess.opts.cg.codegen_units == 1 {
 642         run_work_singlethreaded(sess, &trans.reachable[], work_items);
 643     } else {
 644         run_work_multithreaded(sess, work_items, sess.opts.cg.codegen_units);
 645     }
 646
 647     // All codegen is finished.
 648     unsafe {
 649         llvm::LLVMRustDisposeTargetMachine(tm);
 650     }
 651
 652     // Produce final compile outputs.
 653
 654     let copy_if_one_unit = |&: ext: &str, output_type: config::OutputType, keep_numbered: bool| {
 655         // Three cases:
 656         if sess.opts.cg.codegen_units == 1 {
 657             // 1) Only one codegen unit.  In this case it's no difficulty
 658             //    to copy `foo.0.x` to `foo.x`.
 659             fs::copy(&crate_output.with_extension(ext),
 660                      &crate_output.path(output_type)).unwrap();
 661             if !sess.opts.cg.save_temps && !keep_numbered {
 662                 // The user just wants `foo.x`, not `foo.0.x`.
 663                 remove(sess, &crate_output.with_extension(ext));
 664             }
 665         } else {
 666             if crate_output.single_output_file.is_some() {
 667                 // 2) Multiple codegen units, with `-o some_name`.  We have
 668                 //    no good solution for this case, so warn the user.
 669                 sess.warn(&format!("ignoring -o because multiple .{} files were produced",
 670                                   ext)[]);
 671             } else {
 672                 // 3) Multiple codegen units, but no `-o some_name`.  We
 673                 //    just leave the `foo.0.x` files in place.
 674                 // (We don't have to do any work in this case.)
 675             }
 676         }
 677     };
 678
 679     let link_obj = |&: output_path: &Path| {
 680         // Running `ld -r` on a single input is kind of pointless.
 681         if sess.opts.cg.codegen_units == 1 {
 682             fs::copy(&crate_output.with_extension("0.o"),
 683                      output_path).unwrap();
 684             // Leave the .0.o file around, to mimic the behavior of the normal
 685             // code path.
 686             return;
 687         }
 688
 689         // Some builds of MinGW GCC will pass --force-exe-suffix to ld, which
 690         // will automatically add a .exe extension if the extension is not
 691         // already .exe or .dll.  To ensure consistent behavior on Windows, we
 692         // add the .exe suffix explicitly and then rename the output file to
 693         // the desired path.  This will give the correct behavior whether or
 694         // not GCC adds --force-exe-suffix.
 695         let windows_output_path =
 696             if sess.target.target.options.is_like_windows {
 697                 Some(output_path.with_extension("o.exe"))
 698             } else {
 699                 None
 700             };
 701
 702         let pname = get_cc_prog(sess);
 703         let mut cmd = Command::new(&pname[]);
 704
 705         cmd.args(&sess.target.target.options.pre_link_args[]);
 706         cmd.arg("-nostdlib");
 707
 708         for index in range(0, trans.modules.len()) {
 709             cmd.arg(crate_output.with_extension(&format!("{}.o", index)[]));
 710         }
 711
 712         cmd.arg("-r")
 713            .arg("-o")
 714            .arg(windows_output_path.as_ref().unwrap_or(output_path));
 715
 716         cmd.args(&sess.target.target.options.post_link_args[]);
 717
 718         if sess.opts.debugging_opts.print_link_args {
 719             println!("{}", &cmd);
 720         }
 721
 722         cmd.stdin(::std::io::process::Ignored)
 723            .stdout(::std::io::process::InheritFd(1))
 724            .stderr(::std::io::process::InheritFd(2));
 725         match cmd.status() {
 726             Ok(status) => {
 727                 if !status.success() {
 728                     sess.err(&format!("linking of {} with `{}` failed",
 729                                      output_path.display(), cmd)[]);
 730                     sess.abort_if_errors();
 731                 }
 732             },
 733             Err(e) => {
 734                 sess.err(&format!("could not exec the linker `{}`: {}",
 735                                  pname,
 736                                  e)[]);
 737                 sess.abort_if_errors();
 738             },
 739         }
 740
 741         match windows_output_path {
 742             Some(ref windows_path) => {
 743                 fs::rename(windows_path, output_path).unwrap();
 744             },
 745             None => {
 746                 // The file is already named according to `output_path`.
 747             }
 748         }
 749     };
 750
 751     // Flag to indicate whether the user explicitly requested bitcode.
 752     // Otherwise, we produced it only as a temporary output, and will need
 753     // to get rid of it.
 754     let mut user_wants_bitcode = false;
 755     for output_type in output_types.iter() {
 756         match *output_type {
 757             config::OutputTypeBitcode => {
 758                 user_wants_bitcode = true;
 759                 // Copy to .bc, but always keep the .0.bc.  There is a later
 760                 // check to figure out if we should delete .0.bc files, or keep
 761                 // them for making an rlib.
 762                 copy_if_one_unit("0.bc", config::OutputTypeBitcode, true);
 763             }
 764             config::OutputTypeLlvmAssembly => {
 765                 copy_if_one_unit("0.ll", config::OutputTypeLlvmAssembly, false);
 766             }
 767             config::OutputTypeAssembly => {
 768                 copy_if_one_unit("0.s", config::OutputTypeAssembly, false);
 769             }
 770             config::OutputTypeObject => {
 771                 link_obj(&crate_output.path(config::OutputTypeObject));
 772             }
 773             config::OutputTypeExe => {
 774                 // If config::OutputTypeObject is already in the list, then
 775                 // `crate.o` will be handled by the config::OutputTypeObject case.
 776                 // Otherwise, we need to create the temporary object so we
 777                 // can run the linker.
 778                 if !sess.opts.output_types.contains(&config::OutputTypeObject) {
 779                     link_obj(&crate_output.temp_path(config::OutputTypeObject));
 780                 }
 781             }
 782             config::OutputTypeDepInfo => {}
 783         }
 784     }
 785     let user_wants_bitcode = user_wants_bitcode;
 786
 787     // Clean up unwanted temporary files.
 788
 789     // We create the following files by default:
 790     //  - crate.0.bc
 791     //  - crate.0.o
 792     //  - crate.metadata.bc
 793     //  - crate.metadata.o
 794     //  - crate.o (linked from crate.##.o)
 795     //  - crate.bc (copied from crate.0.bc)
 796     // We may create additional files if requested by the user (through
 797     // `-C save-temps` or `--emit=` flags).
 798
 799     if !sess.opts.cg.save_temps {
 800         // Remove the temporary .0.o objects.  If the user didn't
 801         // explicitly request bitcode (with --emit=bc), and the bitcode is not
 802         // needed for building an rlib, then we must remove .0.bc as well.
 803
 804         // Specific rules for keeping .0.bc:
 805         //  - If we're building an rlib (`needs_crate_bitcode`), then keep
 806         //    it.
 807         //  - If the user requested bitcode (`user_wants_bitcode`), and
 808         //    codegen_units > 1, then keep it.
 809         //  - If the user requested bitcode but codegen_units == 1, then we
 810         //    can toss .0.bc because we copied it to .bc earlier.
 811         //  - If we're not building an rlib and the user didn't request
 812         //    bitcode, then delete .0.bc.
 813         // If you change how this works, also update back::link::link_rlib,
 814         // where .0.bc files are (maybe) deleted after making an rlib.
 815         let keep_numbered_bitcode = needs_crate_bitcode ||
 816                 (user_wants_bitcode && sess.opts.cg.codegen_units > 1);
 817
 818         for i in range(0, trans.modules.len()) {
 819             if modules_config.emit_obj {
 820                 let ext = format!("{}.o", i);
 821                 remove(sess, &crate_output.with_extension(&ext[]));
 822             }
 823
 824             if modules_config.emit_bc && !keep_numbered_bitcode {
 825                 let ext = format!("{}.bc", i);
 826                 remove(sess, &crate_output.with_extension(&ext[]));
 827             }
 828         }
 829
 830         if metadata_config.emit_bc && !user_wants_bitcode {
 831             remove(sess, &crate_output.with_extension("metadata.bc"));
 832         }
 833     }
 834
 835     // We leave the following files around by default:
 836     //  - crate.o
 837     //  - crate.metadata.o
 838     //  - crate.bc
 839     // These are used in linking steps and will be cleaned up afterward.
 840
 841     // FIXME: time_llvm_passes support - does this use a global context or
 842     // something?
 843     //if sess.time_llvm_passes() { llvm::LLVMRustPrintPassTimings(); }
 844 }
 845
 846 struct WorkItem {
 847     mtrans: ModuleTranslation,
 848     config: ModuleConfig,
 849     output_names: OutputFilenames,
 850     name_extra: String
 851 }
 852
 853 fn build_work_item(sess: &Session,
 854                    mtrans: ModuleTranslation,
 855                    config: ModuleConfig,
 856                    output_names: OutputFilenames,
 857                    name_extra: String)
 858                    -> WorkItem
 859 {
 860     let mut config = config;
 861     config.tm = create_target_machine(sess);
 862     WorkItem { mtrans: mtrans, config: config, output_names: output_names,
 863                name_extra: name_extra }
 864 }
 865
 866 fn execute_work_item(cgcx: &CodegenContext,
 867                      work_item: WorkItem) {
 868     unsafe {
 869         optimize_and_codegen(cgcx, work_item.mtrans, work_item.config,
 870                              work_item.name_extra, work_item.output_names);
 871     }
 872 }
 873
 874 fn run_work_singlethreaded(sess: &Session,
 875                            reachable: &[String],
 876                            work_items: Vec<WorkItem>) {
 877     let cgcx = CodegenContext::new_with_session(sess, reachable);
 878     let mut work_items = work_items;
 879
 880     // Since we're running single-threaded, we can pass the session to
 881     // the proc, allowing `optimize_and_codegen` to perform LTO.
 882     for work in Unfold::new((), |_| work_items.pop()) {
 883         execute_work_item(&cgcx, work);
 884     }
 885 }
 886
 887 fn run_work_multithreaded(sess: &Session,
 888                           work_items: Vec<WorkItem>,
 889                           num_workers: uint) {
 890     // Run some workers to process the work items.
 891     let work_items_arc = Arc::new(Mutex::new(work_items));
 892     let mut diag_emitter = SharedEmitter::new();
 893     let mut futures = Vec::with_capacity(num_workers);
 894
 895     for i in range(0, num_workers) {
 896         let work_items_arc = work_items_arc.clone();
 897         let diag_emitter = diag_emitter.clone();
 898         let remark = sess.opts.cg.remark.clone();
 899
 900         let (tx, rx) = channel();
 901         let mut tx = Some(tx);
 902         futures.push(rx);
 903
 904         thread::Builder::new().name(format!("codegen-{}", i)).spawn(move |:| {
 905             let diag_handler = mk_handler(box diag_emitter);
 906
 907             // Must construct cgcx inside the proc because it has non-Send
 908             // fields.
 909             let cgcx = CodegenContext {
 910                 lto_ctxt: None,
 911                 handler: &diag_handler,
 912                 remark: remark,
 913             };
 914
 915             loop {
 916                 // Avoid holding the lock for the entire duration of the match.
 917                 let maybe_work = work_items_arc.lock().unwrap().pop();
 918                 match maybe_work {
 919                     Some(work) => {
 920                         execute_work_item(&cgcx, work);
 921
 922                         // Make sure to fail the worker so the main thread can
 923                         // tell that there were errors.
 924                         cgcx.handler.abort_if_errors();
 925                     }
 926                     None => break,
 927                 }
 928             }
 929
 930             tx.take().unwrap().send(()).unwrap();
 931         });
 932     }
 933
 934     let mut panicked = false;
 935     for rx in futures.into_iter() {
 936         match rx.recv() {
 937             Ok(()) => {},
 938             Err(_) => {
 939                 panicked = true;
 940             },
 941         }
 942         // Display any new diagnostics.
 943         diag_emitter.dump(sess.diagnostic().handler());
 944     }
 945     if panicked {
 946         sess.fatal("aborting due to worker thread panic");
 947     }
 948 }
 949
 950 pub fn run_assembler(sess: &Session, outputs: &OutputFilenames) {
 951     let pname = get_cc_prog(sess);
 952     let mut cmd = Command::new(&pname[]);
 953
 954     cmd.arg("-c").arg("-o").arg(outputs.path(config::OutputTypeObject))
 955                            .arg(outputs.temp_path(config::OutputTypeAssembly));
 956     debug!("{}", &cmd);
 957
 958     match cmd.output() {
 959         Ok(prog) => {
 960             if !prog.status.success() {
 961                 sess.err(&format!("linking with `{}` failed: {}",
 962                                  pname,
 963                                  prog.status)[]);
 964                 sess.note(&format!("{}", &cmd)[]);
 965                 let mut note = prog.error.clone();
 966                 note.push_all(&prog.output[]);
 967                 sess.note(str::from_utf8(&note[]).unwrap());
 968                 sess.abort_if_errors();
 969             }
 970         },
 971         Err(e) => {
 972             sess.err(&format!("could not exec the linker `{}`: {}",
 973                              pname,
 974                              e)[]);
 975             sess.abort_if_errors();
 976         }
 977     }
 978 }
 979
 980 unsafe fn configure_llvm(sess: &Session) {
 981     use std::sync::{Once, ONCE_INIT};
 982     static INIT: Once = ONCE_INIT;
 983
 984     // Copy what clang does by turning on loop vectorization at O2 and
 985     // slp vectorization at O3
 986     let vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
 987                          (sess.opts.optimize == config::Default ||
 988                           sess.opts.optimize == config::Aggressive);
 989     let vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
 990                         sess.opts.optimize == config::Aggressive;
 991
 992     let mut llvm_c_strs = Vec::new();
 993     let mut llvm_args = Vec::new();
 994     {
 995         let mut add = |&mut : arg: &str| {
 996             let s = CString::from_slice(arg.as_bytes());
 997             llvm_args.push(s.as_ptr());
 998             llvm_c_strs.push(s);
 999         };
1000         add("rustc"); // fake program name
1001         if vectorize_loop { add("-vectorize-loops"); }
1002         if vectorize_slp  { add("-vectorize-slp");   }
1003         if sess.time_llvm_passes() { add("-time-passes"); }
1004         if sess.print_llvm_passes() { add("-debug-pass=Structure"); }
1005
1006         for arg in sess.opts.cg.llvm_args.iter() {
1007             add(&(*arg)[]);
1008         }
1009     }
1010
1011     INIT.call_once(|| {
1012         llvm::LLVMInitializePasses();
1013
1014         // Only initialize the platforms supported by Rust here, because
1015         // using --llvm-root will have multiple platforms that rustllvm
1016         // doesn't actually link to and it's pointless to put target info
1017         // into the registry that Rust cannot generate machine code for.
1018         llvm::LLVMInitializeX86TargetInfo();
1019         llvm::LLVMInitializeX86Target();
1020         llvm::LLVMInitializeX86TargetMC();
1021         llvm::LLVMInitializeX86AsmPrinter();
1022         llvm::LLVMInitializeX86AsmParser();
1023
1024         llvm::LLVMInitializeARMTargetInfo();
1025         llvm::LLVMInitializeARMTarget();
1026         llvm::LLVMInitializeARMTargetMC();
1027         llvm::LLVMInitializeARMAsmPrinter();
1028         llvm::LLVMInitializeARMAsmParser();
1029
1030         llvm::LLVMInitializeAArch64TargetInfo();
1031         llvm::LLVMInitializeAArch64Target();
1032         llvm::LLVMInitializeAArch64TargetMC();
1033         llvm::LLVMInitializeAArch64AsmPrinter();
1034         llvm::LLVMInitializeAArch64AsmParser();
1035
1036         llvm::LLVMInitializeMipsTargetInfo();
1037         llvm::LLVMInitializeMipsTarget();
1038         llvm::LLVMInitializeMipsTargetMC();
1039         llvm::LLVMInitializeMipsAsmPrinter();
1040         llvm::LLVMInitializeMipsAsmParser();
1041
1042         llvm::LLVMRustSetLLVMOptions(llvm_args.len() as c_int,
1043                                      llvm_args.as_ptr());
1044     });
1045 }
1046
1047 unsafe fn populate_llvm_passes(fpm: llvm::PassManagerRef,
1048                                mpm: llvm::PassManagerRef,
1049                                llmod: ModuleRef,
1050                                opt: llvm::CodeGenOptLevel,
1051                                no_builtins: bool) {
1052     // Create the PassManagerBuilder for LLVM. We configure it with
1053     // reasonable defaults and prepare it to actually populate the pass
1054     // manager.
1055     let builder = llvm::LLVMPassManagerBuilderCreate();
1056     match opt {
1057         llvm::CodeGenLevelNone => {
1058             // Don't add lifetime intrinsics at O0
1059             llvm::LLVMRustAddAlwaysInlinePass(builder, false);
1060         }
1061         llvm::CodeGenLevelLess => {
1062             llvm::LLVMRustAddAlwaysInlinePass(builder, true);
1063         }
1064         // numeric values copied from clang
1065         llvm::CodeGenLevelDefault => {
1066             llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder,
1067                                                                 225);
1068         }
1069         llvm::CodeGenLevelAggressive => {
1070             llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder,
1071                                                                 275);
1072         }
1073     }
1074     llvm::LLVMPassManagerBuilderSetOptLevel(builder, opt as c_uint);
1075     llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, no_builtins);
1076
1077     // Use the builder to populate the function/module pass managers.
1078     llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(builder, fpm);
1079     llvm::LLVMPassManagerBuilderPopulateModulePassManager(builder, mpm);
1080     llvm::LLVMPassManagerBuilderDispose(builder);
1081
1082     match opt {
1083         llvm::CodeGenLevelDefault | llvm::CodeGenLevelAggressive => {
1084             llvm::LLVMRustAddPass(mpm, "mergefunc\0".as_ptr() as *const _);
1085         }
1086         _ => {}
1087     };
1088 }