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_linker, remove};
  13 use session::config::{OutputFilenames, Passes, SomePasses, AllPasses};
  14 use session::Session;
  15 use session::config::{self, OutputType};
  16 use llvm;
  17 use llvm::{ModuleRef, TargetMachineRef, PassManagerRef, DiagnosticInfoRef, ContextRef};
  18 use llvm::SMDiagnosticRef;
  19 use {CrateTranslation, ModuleTranslation};
  20 use util::common::time;
  21 use util::common::path2cstr;
  22 use syntax::codemap;
  23 use syntax::errors::{self, Handler, Level};
  24 use syntax::errors::emitter::Emitter;
  25
  26 use std::collections::HashMap;
  27 use std::ffi::{CStr, CString};
  28 use std::fs;
  29 use std::path::{Path, PathBuf};
  30 use std::str;
  31 use std::sync::{Arc, Mutex};
  32 use std::sync::mpsc::channel;
  33 use std::thread;
  34 use libc::{c_uint, c_int, c_void};
  35
  36 pub fn llvm_err(handler: &errors::Handler, msg: String) -> ! {
  37     match llvm::last_error() {
  38         Some(err) => panic!(handler.fatal(&format!("{}: {}", msg, err))),
  39         None => panic!(handler.fatal(&msg)),
  40     }
  41 }
  42
  43 pub fn write_output_file(
  44         handler: &errors::Handler,
  45         target: llvm::TargetMachineRef,
  46         pm: llvm::PassManagerRef,
  47         m: ModuleRef,
  48         output: &Path,
  49         file_type: llvm::FileType) {
  50     unsafe {
  51         let output_c = path2cstr(output);
  52         let result = llvm::LLVMRustWriteOutputFile(
  53                 target, pm, m, output_c.as_ptr(), file_type);
  54         if !result {
  55             llvm_err(handler, format!("could not write output to {}", output.display()));
  56         }
  57     }
  58 }
  59
  60
  61 struct Diagnostic {
  62     msg: String,
  63     code: Option<String>,
  64     lvl: Level,
  65 }
  66
  67 // We use an Arc instead of just returning a list of diagnostics from the
  68 // child thread because we need to make sure that the messages are seen even
  69 // if the child thread panics (for example, when `fatal` is called).
  70 #[derive(Clone)]
  71 struct SharedEmitter {
  72     buffer: Arc<Mutex<Vec<Diagnostic>>>,
  73 }
  74
  75 impl SharedEmitter {
  76     fn new() -> SharedEmitter {
  77         SharedEmitter {
  78             buffer: Arc::new(Mutex::new(Vec::new())),
  79         }
  80     }
  81
  82     fn dump(&mut self, handler: &Handler) {
  83         let mut buffer = self.buffer.lock().unwrap();
  84         for diag in &*buffer {
  85             match diag.code {
  86                 Some(ref code) => {
  87                     handler.emit_with_code(None,
  88                                            &diag.msg,
  89                                            &code[..],
  90                                            diag.lvl);
  91                 },
  92                 None => {
  93                     handler.emit(None,
  94                                  &diag.msg,
  95                                  diag.lvl);
  96                 },
  97             }
  98         }
  99         buffer.clear();
 100     }
 101 }
 102
 103 impl Emitter for SharedEmitter {
 104     fn emit(&mut self, sp: Option<&codemap::MultiSpan>,
 105             msg: &str, code: Option<&str>, lvl: Level) {
 106         assert!(sp.is_none(), "SharedEmitter doesn't support spans");
 107
 108         self.buffer.lock().unwrap().push(Diagnostic {
 109             msg: msg.to_string(),
 110             code: code.map(|s| s.to_string()),
 111             lvl: lvl,
 112         });
 113     }
 114
 115     fn custom_emit(&mut self, _sp: &errors::RenderSpan, _msg: &str, _lvl: Level) {
 116         bug!("SharedEmitter doesn't support custom_emit");
 117     }
 118 }
 119
 120
 121 // On android, we by default compile for armv7 processors. This enables
 122 // things like double word CAS instructions (rather than emulating them)
 123 // which are *far* more efficient. This is obviously undesirable in some
 124 // cases, so if any sort of target feature is specified we don't append v7
 125 // to the feature list.
 126 //
 127 // On iOS only armv7 and newer are supported. So it is useful to
 128 // get all hardware potential via VFP3 (hardware floating point)
 129 // and NEON (SIMD) instructions supported by LLVM.
 130 // Note that without those flags various linking errors might
 131 // arise as some of intrinsics are converted into function calls
 132 // and nobody provides implementations those functions
 133 fn target_feature(sess: &Session) -> String {
 134     format!("{},{}", sess.target.target.options.features, sess.opts.cg.target_feature)
 135 }
 136
 137 fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel {
 138     match optimize {
 139       config::OptLevel::No => llvm::CodeGenLevelNone,
 140       config::OptLevel::Less => llvm::CodeGenLevelLess,
 141       config::OptLevel::Default => llvm::CodeGenLevelDefault,
 142       config::OptLevel::Aggressive => llvm::CodeGenLevelAggressive,
 143     }
 144 }
 145
 146 pub fn create_target_machine(sess: &Session) -> TargetMachineRef {
 147     let reloc_model_arg = match sess.opts.cg.relocation_model {
 148         Some(ref s) => &s[..],
 149         None => &sess.target.target.options.relocation_model[..],
 150     };
 151     let reloc_model = match reloc_model_arg {
 152         "pic" => llvm::RelocPIC,
 153         "static" => llvm::RelocStatic,
 154         "default" => llvm::RelocDefault,
 155         "dynamic-no-pic" => llvm::RelocDynamicNoPic,
 156         _ => {
 157             sess.err(&format!("{:?} is not a valid relocation mode",
 158                              sess.opts
 159                                  .cg
 160                                  .relocation_model));
 161             sess.abort_if_errors();
 162             bug!();
 163         }
 164     };
 165
 166     let opt_level = get_llvm_opt_level(sess.opts.optimize);
 167     let use_softfp = sess.opts.cg.soft_float;
 168
 169     let any_library = sess.crate_types.borrow().iter().any(|ty| {
 170         *ty != config::CrateTypeExecutable
 171     });
 172
 173     let ffunction_sections = sess.target.target.options.function_sections;
 174     let fdata_sections = ffunction_sections;
 175
 176     let code_model_arg = match sess.opts.cg.code_model {
 177         Some(ref s) => &s[..],
 178         None => &sess.target.target.options.code_model[..],
 179     };
 180
 181     let code_model = match code_model_arg {
 182         "default" => llvm::CodeModelDefault,
 183         "small" => llvm::CodeModelSmall,
 184         "kernel" => llvm::CodeModelKernel,
 185         "medium" => llvm::CodeModelMedium,
 186         "large" => llvm::CodeModelLarge,
 187         _ => {
 188             sess.err(&format!("{:?} is not a valid code model",
 189                              sess.opts
 190                                  .cg
 191                                  .code_model));
 192             sess.abort_if_errors();
 193             bug!();
 194         }
 195     };
 196
 197     let triple = &sess.target.target.llvm_target;
 198
 199     let tm = unsafe {
 200         let triple = CString::new(triple.as_bytes()).unwrap();
 201         let cpu = match sess.opts.cg.target_cpu {
 202             Some(ref s) => &**s,
 203             None => &*sess.target.target.options.cpu
 204         };
 205         let cpu = CString::new(cpu.as_bytes()).unwrap();
 206         let features = CString::new(target_feature(sess).as_bytes()).unwrap();
 207         llvm::LLVMRustCreateTargetMachine(
 208             triple.as_ptr(), cpu.as_ptr(), features.as_ptr(),
 209             code_model,
 210             reloc_model,
 211             opt_level,
 212             use_softfp,
 213             !any_library && reloc_model == llvm::RelocPIC,
 214             ffunction_sections,
 215             fdata_sections,
 216         )
 217     };
 218
 219     if tm.is_null() {
 220         llvm_err(sess.diagnostic(),
 221                  format!("Could not create LLVM TargetMachine for triple: {}",
 222                          triple).to_string());
 223     } else {
 224         return tm;
 225     };
 226 }
 227
 228
 229 /// Module-specific configuration for `optimize_and_codegen`.
 230 #[derive(Clone)]
 231 pub struct ModuleConfig {
 232     /// LLVM TargetMachine to use for codegen.
 233     tm: TargetMachineRef,
 234     /// Names of additional optimization passes to run.
 235     passes: Vec<String>,
 236     /// Some(level) to optimize at a certain level, or None to run
 237     /// absolutely no optimizations (used for the metadata module).
 238     opt_level: Option<llvm::CodeGenOptLevel>,
 239
 240     // Flags indicating which outputs to produce.
 241     emit_no_opt_bc: bool,
 242     emit_bc: bool,
 243     emit_lto_bc: bool,
 244     emit_ir: bool,
 245     emit_asm: bool,
 246     emit_obj: bool,
 247     // Miscellaneous flags.  These are mostly copied from command-line
 248     // options.
 249     no_verify: bool,
 250     no_prepopulate_passes: bool,
 251     no_builtins: bool,
 252     time_passes: bool,
 253     vectorize_loop: bool,
 254     vectorize_slp: bool,
 255     merge_functions: bool,
 256     inline_threshold: Option<usize>,
 257     // Instead of creating an object file by doing LLVM codegen, just
 258     // make the object file bitcode. Provides easy compatibility with
 259     // emscripten's ecc compiler, when used as the linker.
 260     obj_is_bitcode: bool,
 261 }
 262
 263 unsafe impl Send for ModuleConfig { }
 264
 265 impl ModuleConfig {
 266     fn new(tm: TargetMachineRef, passes: Vec<String>) -> ModuleConfig {
 267         ModuleConfig {
 268             tm: tm,
 269             passes: passes,
 270             opt_level: None,
 271
 272             emit_no_opt_bc: false,
 273             emit_bc: false,
 274             emit_lto_bc: false,
 275             emit_ir: false,
 276             emit_asm: false,
 277             emit_obj: false,
 278             obj_is_bitcode: false,
 279
 280             no_verify: false,
 281             no_prepopulate_passes: false,
 282             no_builtins: false,
 283             time_passes: false,
 284             vectorize_loop: false,
 285             vectorize_slp: false,
 286             merge_functions: false,
 287             inline_threshold: None
 288         }
 289     }
 290
 291     fn set_flags(&mut self, sess: &Session, trans: &CrateTranslation) {
 292         self.no_verify = sess.no_verify();
 293         self.no_prepopulate_passes = sess.opts.cg.no_prepopulate_passes;
 294         self.no_builtins = trans.no_builtins;
 295         self.time_passes = sess.time_passes();
 296         self.inline_threshold = sess.opts.cg.inline_threshold;
 297         self.obj_is_bitcode = sess.target.target.options.obj_is_bitcode;
 298
 299         // Copy what clang does by turning on loop vectorization at O2 and
 300         // slp vectorization at O3. Otherwise configure other optimization aspects
 301         // of this pass manager builder.
 302         self.vectorize_loop = !sess.opts.cg.no_vectorize_loops &&
 303                              (sess.opts.optimize == config::OptLevel::Default ||
 304                               sess.opts.optimize == config::OptLevel::Aggressive);
 305         self.vectorize_slp = !sess.opts.cg.no_vectorize_slp &&
 306                             sess.opts.optimize == config::OptLevel::Aggressive;
 307
 308         self.merge_functions = sess.opts.optimize == config::OptLevel::Default ||
 309                                sess.opts.optimize == config::OptLevel::Aggressive;
 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 passes added by plugins.
 321     plugin_passes: Vec<String>,
 322     // LLVM optimizations for which we want to print remarks.
 323     remark: Passes,
 324     // Worker thread number
 325     worker: usize,
 326 }
 327
 328 impl<'a> CodegenContext<'a> {
 329     fn new_with_session(sess: &'a Session, reachable: &'a [String]) -> CodegenContext<'a> {
 330         CodegenContext {
 331             lto_ctxt: Some((sess, reachable)),
 332             handler: sess.diagnostic(),
 333             plugin_passes: sess.plugin_llvm_passes.borrow().clone(),
 334             remark: sess.opts.cg.remark.clone(),
 335             worker: 0,
 336         }
 337     }
 338 }
 339
 340 struct HandlerFreeVars<'a> {
 341     llcx: ContextRef,
 342     cgcx: &'a CodegenContext<'a>,
 343 }
 344
 345 unsafe extern "C" fn report_inline_asm<'a, 'b>(cgcx: &'a CodegenContext<'a>,
 346                                                msg: &'b str,
 347                                                cookie: c_uint) {
 348     use syntax::codemap::ExpnId;
 349
 350     match cgcx.lto_ctxt {
 351         Some((sess, _)) => {
 352             sess.codemap().with_expn_info(ExpnId::from_u32(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.struct_err(msg)
 360                         .note("build without -C codegen-units for more exact errors")
 361                         .emit();
 362         }
 363     }
 364 }
 365
 366 unsafe extern "C" fn inline_asm_handler(diag: SMDiagnosticRef,
 367                                         user: *const c_void,
 368                                         cookie: c_uint) {
 369     let HandlerFreeVars { cgcx, .. } = *(user as *const HandlerFreeVars);
 370
 371     let msg = llvm::build_string(|s| llvm::LLVMWriteSMDiagnosticToString(diag, s))
 372         .expect("non-UTF8 SMDiagnostic");
 373
 374     report_inline_asm(cgcx, &msg[..], cookie);
 375 }
 376
 377 unsafe extern "C" fn diagnostic_handler(info: DiagnosticInfoRef, user: *mut c_void) {
 378     let HandlerFreeVars { llcx, cgcx } = *(user as *const HandlerFreeVars);
 379
 380     match llvm::diagnostic::Diagnostic::unpack(info) {
 381         llvm::diagnostic::InlineAsm(inline) => {
 382             report_inline_asm(cgcx,
 383                               &llvm::twine_to_string(inline.message),
 384                               inline.cookie);
 385         }
 386
 387         llvm::diagnostic::Optimization(opt) => {
 388             let pass_name = str::from_utf8(CStr::from_ptr(opt.pass_name).to_bytes())
 389                                 .ok()
 390                                 .expect("got a non-UTF8 pass name from LLVM");
 391             let enabled = match cgcx.remark {
 392                 AllPasses => true,
 393                 SomePasses(ref v) => v.iter().any(|s| *s == pass_name),
 394             };
 395
 396             if enabled {
 397                 let loc = llvm::debug_loc_to_string(llcx, opt.debug_loc);
 398                 cgcx.handler.note_without_error(&format!("optimization {} for {} at {}: {}",
 399                                                 opt.kind.describe(),
 400                                                 pass_name,
 401                                                 if loc.is_empty() { "[unknown]" } else { &*loc },
 402                                                 llvm::twine_to_string(opt.message)));
 403             }
 404         }
 405
 406         _ => (),
 407     }
 408 }
 409
 410 // Unsafe due to LLVM calls.
 411 unsafe fn optimize_and_codegen(cgcx: &CodegenContext,
 412                                mtrans: ModuleTranslation,
 413                                config: ModuleConfig,
 414                                name_extra: String,
 415                                output_names: OutputFilenames) {
 416     let ModuleTranslation { llmod, llcx } = mtrans;
 417     let tm = config.tm;
 418
 419     // llcx doesn't outlive this function, so we can put this on the stack.
 420     let fv = HandlerFreeVars {
 421         llcx: llcx,
 422         cgcx: cgcx,
 423     };
 424     let fv = &fv as *const HandlerFreeVars as *mut c_void;
 425
 426     llvm::LLVMSetInlineAsmDiagnosticHandler(llcx, inline_asm_handler, fv);
 427     llvm::LLVMContextSetDiagnosticHandler(llcx, diagnostic_handler, fv);
 428
 429     if config.emit_no_opt_bc {
 430         let ext = format!("{}.no-opt.bc", name_extra);
 431         let out = output_names.with_extension(&ext);
 432         let out = path2cstr(&out);
 433         llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
 434     }
 435
 436     if config.opt_level.is_some() {
 437         // Create the two optimizing pass managers. These mirror what clang
 438         // does, and are by populated by LLVM's default PassManagerBuilder.
 439         // Each manager has a different set of passes, but they also share
 440         // some common passes.
 441         let fpm = llvm::LLVMCreateFunctionPassManagerForModule(llmod);
 442         let mpm = llvm::LLVMCreatePassManager();
 443
 444         // If we're verifying or linting, add them to the function pass
 445         // manager.
 446         let addpass = |pass_name: &str| {
 447             let pass_name = CString::new(pass_name).unwrap();
 448             let pass = llvm::LLVMRustFindAndCreatePass(pass_name.as_ptr());
 449             if pass.is_null() {
 450                 return false;
 451             }
 452             let pass_manager = match llvm::LLVMRustPassKind(pass) {
 453                 llvm::SupportedPassKind::Function => fpm,
 454                 llvm::SupportedPassKind::Module => mpm,
 455                 llvm::SupportedPassKind::Unsupported => {
 456                     cgcx.handler.err("Encountered LLVM pass kind we can't handle");
 457                     return true
 458                 },
 459             };
 460             llvm::LLVMRustAddPass(pass_manager, pass);
 461             true
 462         };
 463
 464         if !config.no_verify { assert!(addpass("verify")); }
 465         if !config.no_prepopulate_passes {
 466             llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod);
 467             llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod);
 468             with_llvm_pmb(llmod, &config, &mut |b| {
 469                 llvm::LLVMPassManagerBuilderPopulateFunctionPassManager(b, fpm);
 470                 llvm::LLVMPassManagerBuilderPopulateModulePassManager(b, mpm);
 471             })
 472         }
 473
 474         for pass in &config.passes {
 475             if !addpass(pass) {
 476                 cgcx.handler.warn(&format!("unknown pass `{}`, ignoring",
 477                                            pass));
 478             }
 479         }
 480
 481         for pass in &cgcx.plugin_passes {
 482             if !addpass(pass) {
 483                 cgcx.handler.err(&format!("a plugin asked for LLVM pass \
 484                                            `{}` but LLVM does not \
 485                                            recognize it", pass));
 486             }
 487         }
 488
 489         cgcx.handler.abort_if_errors();
 490
 491         // Finally, run the actual optimization passes
 492         time(config.time_passes, &format!("llvm function passes [{}]", cgcx.worker), ||
 493              llvm::LLVMRustRunFunctionPassManager(fpm, llmod));
 494         time(config.time_passes, &format!("llvm module passes [{}]", cgcx.worker), ||
 495              llvm::LLVMRunPassManager(mpm, llmod));
 496
 497         // Deallocate managers that we're now done with
 498         llvm::LLVMDisposePassManager(fpm);
 499         llvm::LLVMDisposePassManager(mpm);
 500
 501         match cgcx.lto_ctxt {
 502             Some((sess, reachable)) if sess.lto() =>  {
 503                 time(sess.time_passes(), "all lto passes", ||
 504                      lto::run(sess, llmod, tm, reachable, &config,
 505                               &name_extra, &output_names));
 506
 507                 if config.emit_lto_bc {
 508                     let name = format!("{}.lto.bc", name_extra);
 509                     let out = output_names.with_extension(&name);
 510                     let out = path2cstr(&out);
 511                     llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr());
 512                 }
 513             },
 514             _ => {},
 515         }
 516     }
 517
 518     // A codegen-specific pass manager is used to generate object
 519     // files for an LLVM module.
 520     //
 521     // Apparently each of these pass managers is a one-shot kind of
 522     // thing, so we create a new one for each type of output. The
 523     // pass manager passed to the closure should be ensured to not
 524     // escape the closure itself, and the manager should only be
 525     // used once.
 526     unsafe fn with_codegen<F>(tm: TargetMachineRef,
 527                               llmod: ModuleRef,
 528                               no_builtins: bool,
 529                               f: F) where
 530         F: FnOnce(PassManagerRef),
 531     {
 532         let cpm = llvm::LLVMCreatePassManager();
 533         llvm::LLVMRustAddAnalysisPasses(tm, cpm, llmod);
 534         llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins);
 535         f(cpm);
 536     }
 537
 538     // Change what we write and cleanup based on whether obj files are
 539     // just llvm bitcode. In that case write bitcode, and possibly
 540     // delete the bitcode if it wasn't requested. Don't generate the
 541     // machine code, instead copy the .o file from the .bc
 542     let write_bc = config.emit_bc || config.obj_is_bitcode;
 543     let rm_bc = !config.emit_bc && config.obj_is_bitcode;
 544     let write_obj = config.emit_obj && !config.obj_is_bitcode;
 545     let copy_bc_to_obj = config.emit_obj && config.obj_is_bitcode;
 546
 547     let bc_out = output_names.with_extension(&format!("{}.bc", name_extra));
 548     let obj_out = output_names.with_extension(&format!("{}.o", name_extra));
 549
 550     if write_bc {
 551         let bc_out_c = path2cstr(&bc_out);
 552         llvm::LLVMWriteBitcodeToFile(llmod, bc_out_c.as_ptr());
 553     }
 554
 555     time(config.time_passes, &format!("codegen passes [{}]", cgcx.worker), || {
 556         if config.emit_ir {
 557             let ext = format!("{}.ll", name_extra);
 558             let out = output_names.with_extension(&ext);
 559             let out = path2cstr(&out);
 560             with_codegen(tm, llmod, config.no_builtins, |cpm| {
 561                 llvm::LLVMRustPrintModule(cpm, llmod, out.as_ptr());
 562                 llvm::LLVMDisposePassManager(cpm);
 563             })
 564         }
 565
 566         if config.emit_asm {
 567             let path = output_names.with_extension(&format!("{}.s", name_extra));
 568
 569             // We can't use the same module for asm and binary output, because that triggers
 570             // various errors like invalid IR or broken binaries, so we might have to clone the
 571             // module to produce the asm output
 572             let llmod = if config.emit_obj {
 573                 llvm::LLVMCloneModule(llmod)
 574             } else {
 575                 llmod
 576             };
 577             with_codegen(tm, llmod, config.no_builtins, |cpm| {
 578                 write_output_file(cgcx.handler, tm, cpm, llmod, &path,
 579                                   llvm::AssemblyFileType);
 580             });
 581             if config.emit_obj {
 582                 llvm::LLVMDisposeModule(llmod);
 583             }
 584         }
 585
 586         if write_obj {
 587             with_codegen(tm, llmod, config.no_builtins, |cpm| {
 588                 write_output_file(cgcx.handler, tm, cpm, llmod, &obj_out, llvm::ObjectFileType);
 589             });
 590         }
 591     });
 592
 593     if copy_bc_to_obj {
 594         debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out);
 595         if let Err(e) = fs::copy(&bc_out, &obj_out) {
 596             cgcx.handler.err(&format!("failed to copy bitcode to object file: {}", e));
 597         }
 598     }
 599
 600     if rm_bc {
 601         debug!("removing_bitcode {:?}", bc_out);
 602         if let Err(e) = fs::remove_file(&bc_out) {
 603             cgcx.handler.err(&format!("failed to remove bitcode: {}", e));
 604         }
 605     }
 606
 607     llvm::LLVMDisposeModule(llmod);
 608     llvm::LLVMContextDispose(llcx);
 609     llvm::LLVMRustDisposeTargetMachine(tm);
 610 }
 611
 612 pub fn run_passes(sess: &Session,
 613                   trans: &CrateTranslation,
 614                   output_types: &HashMap<OutputType, Option<PathBuf>>,
 615                   crate_output: &OutputFilenames) {
 616     // It's possible that we have `codegen_units > 1` but only one item in
 617     // `trans.modules`.  We could theoretically proceed and do LTO in that
 618     // case, but it would be confusing to have the validity of
 619     // `-Z lto -C codegen-units=2` depend on details of the crate being
 620     // compiled, so we complain regardless.
 621     if sess.lto() && sess.opts.cg.codegen_units > 1 {
 622         // This case is impossible to handle because LTO expects to be able
 623         // to combine the entire crate and all its dependencies into a
 624         // single compilation unit, but each codegen unit is in a separate
 625         // LLVM context, so they can't easily be combined.
 626         sess.fatal("can't perform LTO when using multiple codegen units");
 627     }
 628
 629     // Sanity check
 630     assert!(trans.modules.len() == sess.opts.cg.codegen_units);
 631
 632     let tm = create_target_machine(sess);
 633
 634     // Figure out what we actually need to build.
 635
 636     let mut modules_config = ModuleConfig::new(tm, sess.opts.cg.passes.clone());
 637     let mut metadata_config = ModuleConfig::new(tm, vec!());
 638
 639     modules_config.opt_level = Some(get_llvm_opt_level(sess.opts.optimize));
 640
 641     // Save all versions of the bytecode if we're saving our temporaries.
 642     if sess.opts.cg.save_temps {
 643         modules_config.emit_no_opt_bc = true;
 644         modules_config.emit_bc = true;
 645         modules_config.emit_lto_bc = true;
 646         metadata_config.emit_bc = true;
 647     }
 648
 649     // Emit bitcode files for the crate if we're emitting an rlib.
 650     // Whenever an rlib is created, the bitcode is inserted into the
 651     // archive in order to allow LTO against it.
 652     let needs_crate_bitcode =
 653             sess.crate_types.borrow().contains(&config::CrateTypeRlib) &&
 654             sess.opts.output_types.contains_key(&OutputType::Exe);
 655     let needs_crate_object =
 656             sess.opts.output_types.contains_key(&OutputType::Exe);
 657     if needs_crate_bitcode {
 658         modules_config.emit_bc = true;
 659     }
 660
 661     for output_type in output_types.keys() {
 662         match *output_type {
 663             OutputType::Bitcode => { modules_config.emit_bc = true; },
 664             OutputType::LlvmAssembly => { modules_config.emit_ir = true; },
 665             OutputType::Assembly => {
 666                 modules_config.emit_asm = true;
 667                 // If we're not using the LLVM assembler, this function
 668                 // could be invoked specially with output_type_assembly, so
 669                 // in this case we still want the metadata object file.
 670                 if !sess.opts.output_types.contains_key(&OutputType::Assembly) {
 671                     metadata_config.emit_obj = true;
 672                 }
 673             },
 674             OutputType::Object => { modules_config.emit_obj = true; },
 675             OutputType::Exe => {
 676                 modules_config.emit_obj = true;
 677                 metadata_config.emit_obj = true;
 678             },
 679             OutputType::DepInfo => {}
 680         }
 681     }
 682
 683     modules_config.set_flags(sess, trans);
 684     metadata_config.set_flags(sess, trans);
 685
 686
 687     // Populate a buffer with a list of codegen threads.  Items are processed in
 688     // LIFO order, just because it's a tiny bit simpler that way.  (The order
 689     // doesn't actually matter.)
 690     let mut work_items = Vec::with_capacity(1 + trans.modules.len());
 691
 692     {
 693         let work = build_work_item(sess,
 694                                    trans.metadata_module,
 695                                    metadata_config.clone(),
 696                                    crate_output.clone(),
 697                                    "metadata".to_string());
 698         work_items.push(work);
 699     }
 700
 701     for (index, mtrans) in trans.modules.iter().enumerate() {
 702         let work = build_work_item(sess,
 703                                    *mtrans,
 704                                    modules_config.clone(),
 705                                    crate_output.clone(),
 706                                    format!("{}", index));
 707         work_items.push(work);
 708     }
 709
 710     // Process the work items, optionally using worker threads.
 711     if sess.opts.cg.codegen_units == 1 {
 712         run_work_singlethreaded(sess, &trans.reachable, work_items);
 713     } else {
 714         run_work_multithreaded(sess, work_items, sess.opts.cg.codegen_units);
 715     }
 716
 717     // All codegen is finished.
 718     unsafe {
 719         llvm::LLVMRustDisposeTargetMachine(tm);
 720     }
 721
 722     // Produce final compile outputs.
 723     let copy_gracefully = |from: &Path, to: &Path| {
 724         if let Err(e) = fs::copy(from, to) {
 725             sess.err(&format!("could not copy {:?} to {:?}: {}", from, to, e));
 726         }
 727     };
 728
 729     let copy_if_one_unit = |ext: &str,
 730                             output_type: OutputType,
 731                             keep_numbered: bool| {
 732         if sess.opts.cg.codegen_units == 1 {
 733             // 1) Only one codegen unit.  In this case it's no difficulty
 734             //    to copy `foo.0.x` to `foo.x`.
 735             copy_gracefully(&crate_output.with_extension(ext),
 736                             &crate_output.path(output_type));
 737             if !sess.opts.cg.save_temps && !keep_numbered {
 738                 // The user just wants `foo.x`, not `foo.0.x`.
 739                 remove(sess, &crate_output.with_extension(ext));
 740             }
 741         } else if crate_output.outputs.contains_key(&output_type) {
 742             // 2) Multiple codegen units, with `--emit foo=some_name`.  We have
 743             //    no good solution for this case, so warn the user.
 744             sess.warn(&format!("ignoring emit path because multiple .{} files \
 745                                 were produced", ext));
 746         } else if crate_output.single_output_file.is_some() {
 747             // 3) Multiple codegen units, with `-o some_name`.  We have
 748             //    no good solution for this case, so warn the user.
 749             sess.warn(&format!("ignoring -o because multiple .{} files \
 750                                 were produced", ext));
 751         } else {
 752             // 4) Multiple codegen units, but no explicit name.  We
 753             //    just leave the `foo.0.x` files in place.
 754             // (We don't have to do any work in this case.)
 755         }
 756     };
 757
 758     // Flag to indicate whether the user explicitly requested bitcode.
 759     // Otherwise, we produced it only as a temporary output, and will need
 760     // to get rid of it.
 761     let mut user_wants_bitcode = false;
 762     let mut user_wants_objects = false;
 763     for output_type in output_types.keys() {
 764         match *output_type {
 765             OutputType::Bitcode => {
 766                 user_wants_bitcode = true;
 767                 // Copy to .bc, but always keep the .0.bc.  There is a later
 768                 // check to figure out if we should delete .0.bc files, or keep
 769                 // them for making an rlib.
 770                 copy_if_one_unit("0.bc", OutputType::Bitcode, true);
 771             }
 772             OutputType::LlvmAssembly => {
 773                 copy_if_one_unit("0.ll", OutputType::LlvmAssembly, false);
 774             }
 775             OutputType::Assembly => {
 776                 copy_if_one_unit("0.s", OutputType::Assembly, false);
 777             }
 778             OutputType::Object => {
 779                 user_wants_objects = true;
 780                 copy_if_one_unit("0.o", OutputType::Object, true);
 781             }
 782             OutputType::Exe |
 783             OutputType::DepInfo => {}
 784         }
 785     }
 786     let user_wants_bitcode = user_wants_bitcode;
 787
 788     // Clean up unwanted temporary files.
 789
 790     // We create the following files by default:
 791     //  - crate.0.bc
 792     //  - crate.0.o
 793     //  - crate.metadata.bc
 794     //  - crate.metadata.o
 795     //  - crate.o (linked from crate.##.o)
 796     //  - crate.bc (copied from crate.0.bc)
 797     // We may create additional files if requested by the user (through
 798     // `-C save-temps` or `--emit=` flags).
 799
 800     if !sess.opts.cg.save_temps {
 801         // Remove the temporary .0.o objects.  If the user didn't
 802         // explicitly request bitcode (with --emit=bc), and the bitcode is not
 803         // needed for building an rlib, then we must remove .0.bc as well.
 804
 805         // Specific rules for keeping .0.bc:
 806         //  - If we're building an rlib (`needs_crate_bitcode`), then keep
 807         //    it.
 808         //  - If the user requested bitcode (`user_wants_bitcode`), and
 809         //    codegen_units > 1, then keep it.
 810         //  - If the user requested bitcode but codegen_units == 1, then we
 811         //    can toss .0.bc because we copied it to .bc earlier.
 812         //  - If we're not building an rlib and the user didn't request
 813         //    bitcode, then delete .0.bc.
 814         // If you change how this works, also update back::link::link_rlib,
 815         // where .0.bc files are (maybe) deleted after making an rlib.
 816         let keep_numbered_bitcode = needs_crate_bitcode ||
 817                 (user_wants_bitcode && sess.opts.cg.codegen_units > 1);
 818
 819         let keep_numbered_objects = needs_crate_object ||
 820                 (user_wants_objects && sess.opts.cg.codegen_units > 1);
 821
 822         for i in 0..trans.modules.len() {
 823             if modules_config.emit_obj && !keep_numbered_objects {
 824                 let ext = format!("{}.o", i);
 825                 remove(sess, &crate_output.with_extension(&ext));
 826             }
 827
 828             if modules_config.emit_bc && !keep_numbered_bitcode {
 829                 let ext = format!("{}.bc", i);
 830                 remove(sess, &crate_output.with_extension(&ext));
 831             }
 832         }
 833
 834         if metadata_config.emit_bc && !user_wants_bitcode {
 835             remove(sess, &crate_output.with_extension("metadata.bc"));
 836         }
 837     }
 838
 839     // We leave the following files around by default:
 840     //  - crate.o
 841     //  - crate.metadata.o
 842     //  - crate.bc
 843     // These are used in linking steps and will be cleaned up afterward.
 844
 845     // FIXME: time_llvm_passes support - does this use a global context or
 846     // something?
 847     if sess.opts.cg.codegen_units == 1 && sess.time_llvm_passes() {
 848         unsafe { llvm::LLVMRustPrintPassTimings(); }
 849     }
 850 }
 851
 852 struct WorkItem {
 853     mtrans: ModuleTranslation,
 854     config: ModuleConfig,
 855     output_names: OutputFilenames,
 856     name_extra: String
 857 }
 858
 859 fn build_work_item(sess: &Session,
 860                    mtrans: ModuleTranslation,
 861                    config: ModuleConfig,
 862                    output_names: OutputFilenames,
 863                    name_extra: String)
 864                    -> WorkItem
 865 {
 866     let mut config = config;
 867     config.tm = create_target_machine(sess);
 868     WorkItem { mtrans: mtrans, config: config, output_names: output_names,
 869                name_extra: name_extra }
 870 }
 871
 872 fn execute_work_item(cgcx: &CodegenContext,
 873                      work_item: WorkItem) {
 874     unsafe {
 875         optimize_and_codegen(cgcx, work_item.mtrans, work_item.config,
 876                              work_item.name_extra, work_item.output_names);
 877     }
 878 }
 879
 880 fn run_work_singlethreaded(sess: &Session,
 881                            reachable: &[String],
 882                            work_items: Vec<WorkItem>) {
 883     let cgcx = CodegenContext::new_with_session(sess, reachable);
 884
 885     // Since we're running single-threaded, we can pass the session to
 886     // the proc, allowing `optimize_and_codegen` to perform LTO.
 887     for work in work_items.into_iter().rev() {
 888         execute_work_item(&cgcx, work);
 889     }
 890 }
 891
 892 fn run_work_multithreaded(sess: &Session,
 893                           work_items: Vec<WorkItem>,
 894                           num_workers: usize) {
 895     // Run some workers to process the work items.
 896     let work_items_arc = Arc::new(Mutex::new(work_items));
 897     let mut diag_emitter = SharedEmitter::new();
 898     let mut futures = Vec::with_capacity(num_workers);
 899
 900     for i in 0..num_workers {
 901         let work_items_arc = work_items_arc.clone();
 902         let diag_emitter = diag_emitter.clone();
 903         let plugin_passes = sess.plugin_llvm_passes.borrow().clone();
 904         let remark = sess.opts.cg.remark.clone();
 905
 906         let (tx, rx) = channel();
 907         let mut tx = Some(tx);
 908         futures.push(rx);
 909
 910         thread::Builder::new().name(format!("codegen-{}", i)).spawn(move || {
 911             let diag_handler = Handler::with_emitter(true, false, box diag_emitter);
 912
 913             // Must construct cgcx inside the proc because it has non-Send
 914             // fields.
 915             let cgcx = CodegenContext {
 916                 lto_ctxt: None,
 917                 handler: &diag_handler,
 918                 plugin_passes: plugin_passes,
 919                 remark: remark,
 920                 worker: i,
 921             };
 922
 923             loop {
 924                 // Avoid holding the lock for the entire duration of the match.
 925                 let maybe_work = work_items_arc.lock().unwrap().pop();
 926                 match maybe_work {
 927                     Some(work) => {
 928                         execute_work_item(&cgcx, work);
 929
 930                         // Make sure to fail the worker so the main thread can
 931                         // tell that there were errors.
 932                         cgcx.handler.abort_if_errors();
 933                     }
 934                     None => break,
 935                 }
 936             }
 937
 938             tx.take().unwrap().send(()).unwrap();
 939         }).unwrap();
 940     }
 941
 942     let mut panicked = false;
 943     for rx in futures {
 944         match rx.recv() {
 945             Ok(()) => {},
 946             Err(_) => {
 947                 panicked = true;
 948             },
 949         }
 950         // Display any new diagnostics.
 951         diag_emitter.dump(sess.diagnostic());
 952     }
 953     if panicked {
 954         sess.fatal("aborting due to worker thread panic");
 955     }
 956 }
 957
 958 pub fn run_assembler(sess: &Session, outputs: &OutputFilenames) {
 959     let (pname, mut cmd) = get_linker(sess);
 960
 961     cmd.arg("-c").arg("-o").arg(&outputs.path(OutputType::Object))
 962                            .arg(&outputs.temp_path(OutputType::Assembly));
 963     debug!("{:?}", cmd);
 964
 965     match cmd.output() {
 966         Ok(prog) => {
 967             if !prog.status.success() {
 968                 let mut note = prog.stderr.clone();
 969                 note.extend_from_slice(&prog.stdout);
 970
 971                 sess.struct_err(&format!("linking with `{}` failed: {}",
 972                                          pname,
 973                                          prog.status))
 974                     .note(&format!("{:?}", &cmd))
 975                     .note(str::from_utf8(&note[..]).unwrap())
 976                     .emit();
 977                 sess.abort_if_errors();
 978             }
 979         },
 980         Err(e) => {
 981             sess.err(&format!("could not exec the linker `{}`: {}", pname, e));
 982             sess.abort_if_errors();
 983         }
 984     }
 985 }
 986
 987 pub unsafe fn configure_llvm(sess: &Session) {
 988     let mut llvm_c_strs = Vec::new();
 989     let mut llvm_args = Vec::new();
 990
 991     {
 992         let mut add = |arg: &str| {
 993             let s = CString::new(arg).unwrap();
 994             llvm_args.push(s.as_ptr());
 995             llvm_c_strs.push(s);
 996         };
 997         add("rustc"); // fake program name
 998         if sess.time_llvm_passes() { add("-time-passes"); }
 999         if sess.print_llvm_passes() { add("-debug-pass=Structure"); }
1000
1001         // FIXME #21627 disable faulty FastISel on AArch64 (even for -O0)
1002         if sess.target.target.arch == "aarch64" { add("-fast-isel=0"); }
1003
1004         for arg in &sess.opts.cg.llvm_args {
1005             add(&(*arg));
1006         }
1007     }
1008
1009     llvm::LLVMInitializePasses();
1010
1011     llvm::initialize_available_targets();
1012
1013     llvm::LLVMRustSetLLVMOptions(llvm_args.len() as c_int,
1014                                  llvm_args.as_ptr());
1015 }
1016
1017 pub unsafe fn with_llvm_pmb(llmod: ModuleRef,
1018                             config: &ModuleConfig,
1019                             f: &mut FnMut(llvm::PassManagerBuilderRef)) {
1020     // Create the PassManagerBuilder for LLVM. We configure it with
1021     // reasonable defaults and prepare it to actually populate the pass
1022     // manager.
1023     let builder = llvm::LLVMPassManagerBuilderCreate();
1024     let opt = config.opt_level.unwrap_or(llvm::CodeGenLevelNone);
1025     let inline_threshold = config.inline_threshold;
1026
1027     llvm::LLVMRustConfigurePassManagerBuilder(builder, opt,
1028                                               config.merge_functions,
1029                                               config.vectorize_slp,
1030                                               config.vectorize_loop);
1031
1032     llvm::LLVMRustAddBuilderLibraryInfo(builder, llmod, config.no_builtins);
1033
1034     // Here we match what clang does (kinda). For O0 we only inline
1035     // always-inline functions (but don't add lifetime intrinsics), at O1 we
1036     // inline with lifetime intrinsics, and O2+ we add an inliner with a
1037     // thresholds copied from clang.
1038     match (opt, inline_threshold) {
1039         (_, Some(t)) => {
1040             llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, t as u32);
1041         }
1042         (llvm::CodeGenLevelNone, _) => {
1043             llvm::LLVMRustAddAlwaysInlinePass(builder, false);
1044         }
1045         (llvm::CodeGenLevelLess, _) => {
1046             llvm::LLVMRustAddAlwaysInlinePass(builder, true);
1047         }
1048         (llvm::CodeGenLevelDefault, _) => {
1049             llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 225);
1050         }
1051         (llvm::CodeGenLevelAggressive, _) => {
1052             llvm::LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 275);
1053         }
1054     }
1055
1056     f(builder);
1057     llvm::LLVMPassManagerBuilderDispose(builder);
1058 }