1 //===--- Driver.cpp - Clang GCC Compatible Driver -------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "clang/Driver/Driver.h" 10 #include "ToolChains/AIX.h" 11 #include "ToolChains/AMDGPU.h" 12 #include "ToolChains/AMDGPUOpenMP.h" 13 #include "ToolChains/AVR.h" 14 #include "ToolChains/Ananas.h" 15 #include "ToolChains/BareMetal.h" 16 #include "ToolChains/CSKYToolChain.h" 17 #include "ToolChains/Clang.h" 18 #include "ToolChains/CloudABI.h" 19 #include "ToolChains/Contiki.h" 20 #include "ToolChains/CrossWindows.h" 21 #include "ToolChains/Cuda.h" 22 #include "ToolChains/Darwin.h" 23 #include "ToolChains/DragonFly.h" 24 #include "ToolChains/FreeBSD.h" 25 #include "ToolChains/Fuchsia.h" 26 #include "ToolChains/Gnu.h" 27 #include "ToolChains/HIPAMD.h" 28 #include "ToolChains/HIPSPV.h" 29 #include "ToolChains/HLSL.h" 30 #include "ToolChains/Haiku.h" 31 #include "ToolChains/Hexagon.h" 32 #include "ToolChains/Hurd.h" 33 #include "ToolChains/Lanai.h" 34 #include "ToolChains/Linux.h" 35 #include "ToolChains/MSP430.h" 36 #include "ToolChains/MSVC.h" 37 #include "ToolChains/MinGW.h" 38 #include "ToolChains/Minix.h" 39 #include "ToolChains/MipsLinux.h" 40 #include "ToolChains/Myriad.h" 41 #include "ToolChains/NaCl.h" 42 #include "ToolChains/NetBSD.h" 43 #include "ToolChains/OpenBSD.h" 44 #include "ToolChains/PPCFreeBSD.h" 45 #include "ToolChains/PPCLinux.h" 46 #include "ToolChains/PS4CPU.h" 47 #include "ToolChains/RISCVToolchain.h" 48 #include "ToolChains/SPIRV.h" 49 #include "ToolChains/Solaris.h" 50 #include "ToolChains/TCE.h" 51 #include "ToolChains/VEToolchain.h" 52 #include "ToolChains/WebAssembly.h" 53 #include "ToolChains/XCore.h" 54 #include "ToolChains/ZOS.h" 55 #include "clang/Basic/TargetID.h" 56 #include "clang/Basic/Version.h" 57 #include "clang/Config/config.h" 58 #include "clang/Driver/Action.h" 59 #include "clang/Driver/Compilation.h" 60 #include "clang/Driver/DriverDiagnostic.h" 61 #include "clang/Driver/InputInfo.h" 62 #include "clang/Driver/Job.h" 63 #include "clang/Driver/Options.h" 64 #include "clang/Driver/Phases.h" 65 #include "clang/Driver/SanitizerArgs.h" 66 #include "clang/Driver/Tool.h" 67 #include "clang/Driver/ToolChain.h" 68 #include "clang/Driver/Types.h" 69 #include "llvm/ADT/ArrayRef.h" 70 #include "llvm/ADT/STLExtras.h" 71 #include "llvm/ADT/SmallSet.h" 72 #include "llvm/ADT/StringExtras.h" 73 #include "llvm/ADT/StringRef.h" 74 #include "llvm/ADT/StringSet.h" 75 #include "llvm/ADT/StringSwitch.h" 76 #include "llvm/Config/llvm-config.h" 77 #include "llvm/MC/TargetRegistry.h" 78 #include "llvm/Option/Arg.h" 79 #include "llvm/Option/ArgList.h" 80 #include "llvm/Option/OptSpecifier.h" 81 #include "llvm/Option/OptTable.h" 82 #include "llvm/Option/Option.h" 83 #include "llvm/Support/CommandLine.h" 84 #include "llvm/Support/ErrorHandling.h" 85 #include "llvm/Support/ExitCodes.h" 86 #include "llvm/Support/FileSystem.h" 87 #include "llvm/Support/FormatVariadic.h" 88 #include "llvm/Support/Host.h" 89 #include "llvm/Support/MD5.h" 90 #include "llvm/Support/Path.h" 91 #include "llvm/Support/PrettyStackTrace.h" 92 #include "llvm/Support/Process.h" 93 #include "llvm/Support/Program.h" 94 #include "llvm/Support/StringSaver.h" 95 #include "llvm/Support/VirtualFileSystem.h" 96 #include "llvm/Support/raw_ostream.h" 97 #include <map> 98 #include <memory> 99 #include <utility> 100 #if LLVM_ON_UNIX 101 #include <unistd.h> // getpid 102 #endif 103 104 using namespace clang::driver; 105 using namespace clang; 106 using namespace llvm::opt; 107 108 static llvm::Optional<llvm::Triple> 109 getOffloadTargetTriple(const Driver &D, const ArgList &Args) { 110 auto OffloadTargets = Args.getAllArgValues(options::OPT_offload_EQ); 111 // Offload compilation flow does not support multiple targets for now. We 112 // need the HIPActionBuilder (and possibly the CudaActionBuilder{,Base}too) 113 // to support multiple tool chains first. 114 switch (OffloadTargets.size()) { 115 default: 116 D.Diag(diag::err_drv_only_one_offload_target_supported); 117 return llvm::None; 118 case 0: 119 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << ""; 120 return llvm::None; 121 case 1: 122 break; 123 } 124 return llvm::Triple(OffloadTargets[0]); 125 } 126 127 static llvm::Optional<llvm::Triple> 128 getNVIDIAOffloadTargetTriple(const Driver &D, const ArgList &Args, 129 const llvm::Triple &HostTriple) { 130 if (!Args.hasArg(options::OPT_offload_EQ)) { 131 return llvm::Triple(HostTriple.isArch64Bit() ? "nvptx64-nvidia-cuda" 132 : "nvptx-nvidia-cuda"); 133 } 134 auto TT = getOffloadTargetTriple(D, Args); 135 if (TT && (TT->getArch() == llvm::Triple::spirv32 || 136 TT->getArch() == llvm::Triple::spirv64)) { 137 if (Args.hasArg(options::OPT_emit_llvm)) 138 return TT; 139 D.Diag(diag::err_drv_cuda_offload_only_emit_bc); 140 return llvm::None; 141 } 142 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 143 return llvm::None; 144 } 145 static llvm::Optional<llvm::Triple> 146 getHIPOffloadTargetTriple(const Driver &D, const ArgList &Args) { 147 if (!Args.hasArg(options::OPT_offload_EQ)) { 148 return llvm::Triple("amdgcn-amd-amdhsa"); // Default HIP triple. 149 } 150 auto TT = getOffloadTargetTriple(D, Args); 151 if (!TT) 152 return llvm::None; 153 if (TT->getArch() == llvm::Triple::amdgcn && 154 TT->getVendor() == llvm::Triple::AMD && 155 TT->getOS() == llvm::Triple::AMDHSA) 156 return TT; 157 if (TT->getArch() == llvm::Triple::spirv64) 158 return TT; 159 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 160 return llvm::None; 161 } 162 163 // static 164 std::string Driver::GetResourcesPath(StringRef BinaryPath, 165 StringRef CustomResourceDir) { 166 // Since the resource directory is embedded in the module hash, it's important 167 // that all places that need it call this function, so that they get the 168 // exact same string ("a/../b/" and "b/" get different hashes, for example). 169 170 // Dir is bin/ or lib/, depending on where BinaryPath is. 171 std::string Dir = std::string(llvm::sys::path::parent_path(BinaryPath)); 172 173 SmallString<128> P(Dir); 174 if (CustomResourceDir != "") { 175 llvm::sys::path::append(P, CustomResourceDir); 176 } else { 177 // On Windows, libclang.dll is in bin/. 178 // On non-Windows, libclang.so/.dylib is in lib/. 179 // With a static-library build of libclang, LibClangPath will contain the 180 // path of the embedding binary, which for LLVM binaries will be in bin/. 181 // ../lib gets us to lib/ in both cases. 182 P = llvm::sys::path::parent_path(Dir); 183 llvm::sys::path::append(P, Twine("lib") + CLANG_LIBDIR_SUFFIX, "clang", 184 CLANG_VERSION_STRING); 185 } 186 187 return std::string(P.str()); 188 } 189 190 Driver::Driver(StringRef ClangExecutable, StringRef TargetTriple, 191 DiagnosticsEngine &Diags, std::string Title, 192 IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS) 193 : Diags(Diags), VFS(std::move(VFS)), Mode(GCCMode), 194 SaveTemps(SaveTempsNone), BitcodeEmbed(EmbedNone), 195 CXX20HeaderType(HeaderMode_None), ModulesModeCXX20(false), 196 LTOMode(LTOK_None), ClangExecutable(ClangExecutable), 197 SysRoot(DEFAULT_SYSROOT), DriverTitle(Title), CCCPrintBindings(false), 198 CCPrintOptions(false), CCPrintHeaders(false), CCLogDiagnostics(false), 199 CCGenDiagnostics(false), CCPrintProcessStats(false), 200 TargetTriple(TargetTriple), Saver(Alloc), CheckInputsExist(true), 201 ProbePrecompiled(true), GenReproducer(false), 202 SuppressMissingInputWarning(false) { 203 // Provide a sane fallback if no VFS is specified. 204 if (!this->VFS) 205 this->VFS = llvm::vfs::getRealFileSystem(); 206 207 Name = std::string(llvm::sys::path::filename(ClangExecutable)); 208 Dir = std::string(llvm::sys::path::parent_path(ClangExecutable)); 209 InstalledDir = Dir; // Provide a sensible default installed dir. 210 211 if ((!SysRoot.empty()) && llvm::sys::path::is_relative(SysRoot)) { 212 // Prepend InstalledDir if SysRoot is relative 213 SmallString<128> P(InstalledDir); 214 llvm::sys::path::append(P, SysRoot); 215 SysRoot = std::string(P); 216 } 217 218 #if defined(CLANG_CONFIG_FILE_SYSTEM_DIR) 219 SystemConfigDir = CLANG_CONFIG_FILE_SYSTEM_DIR; 220 #endif 221 #if defined(CLANG_CONFIG_FILE_USER_DIR) 222 UserConfigDir = CLANG_CONFIG_FILE_USER_DIR; 223 #endif 224 225 // Compute the path to the resource directory. 226 ResourceDir = GetResourcesPath(ClangExecutable, CLANG_RESOURCE_DIR); 227 } 228 229 void Driver::setDriverMode(StringRef Value) { 230 static const std::string OptName = 231 getOpts().getOption(options::OPT_driver_mode).getPrefixedName(); 232 if (auto M = llvm::StringSwitch<llvm::Optional<DriverMode>>(Value) 233 .Case("gcc", GCCMode) 234 .Case("g++", GXXMode) 235 .Case("cpp", CPPMode) 236 .Case("cl", CLMode) 237 .Case("flang", FlangMode) 238 .Case("dxc", DXCMode) 239 .Default(None)) 240 Mode = *M; 241 else 242 Diag(diag::err_drv_unsupported_option_argument) << OptName << Value; 243 } 244 245 InputArgList Driver::ParseArgStrings(ArrayRef<const char *> ArgStrings, 246 bool IsClCompatMode, 247 bool &ContainsError) { 248 llvm::PrettyStackTraceString CrashInfo("Command line argument parsing"); 249 ContainsError = false; 250 251 unsigned IncludedFlagsBitmask; 252 unsigned ExcludedFlagsBitmask; 253 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 254 getIncludeExcludeOptionFlagMasks(IsClCompatMode); 255 256 // Make sure that Flang-only options don't pollute the Clang output 257 // TODO: Make sure that Clang-only options don't pollute Flang output 258 if (!IsFlangMode()) 259 ExcludedFlagsBitmask |= options::FlangOnlyOption; 260 261 unsigned MissingArgIndex, MissingArgCount; 262 InputArgList Args = 263 getOpts().ParseArgs(ArgStrings, MissingArgIndex, MissingArgCount, 264 IncludedFlagsBitmask, ExcludedFlagsBitmask); 265 266 // Check for missing argument error. 267 if (MissingArgCount) { 268 Diag(diag::err_drv_missing_argument) 269 << Args.getArgString(MissingArgIndex) << MissingArgCount; 270 ContainsError |= 271 Diags.getDiagnosticLevel(diag::err_drv_missing_argument, 272 SourceLocation()) > DiagnosticsEngine::Warning; 273 } 274 275 // Check for unsupported options. 276 for (const Arg *A : Args) { 277 if (A->getOption().hasFlag(options::Unsupported)) { 278 unsigned DiagID; 279 auto ArgString = A->getAsString(Args); 280 std::string Nearest; 281 if (getOpts().findNearest( 282 ArgString, Nearest, IncludedFlagsBitmask, 283 ExcludedFlagsBitmask | options::Unsupported) > 1) { 284 DiagID = diag::err_drv_unsupported_opt; 285 Diag(DiagID) << ArgString; 286 } else { 287 DiagID = diag::err_drv_unsupported_opt_with_suggestion; 288 Diag(DiagID) << ArgString << Nearest; 289 } 290 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 291 DiagnosticsEngine::Warning; 292 continue; 293 } 294 295 // Warn about -mcpu= without an argument. 296 if (A->getOption().matches(options::OPT_mcpu_EQ) && A->containsValue("")) { 297 Diag(diag::warn_drv_empty_joined_argument) << A->getAsString(Args); 298 ContainsError |= Diags.getDiagnosticLevel( 299 diag::warn_drv_empty_joined_argument, 300 SourceLocation()) > DiagnosticsEngine::Warning; 301 } 302 } 303 304 for (const Arg *A : Args.filtered(options::OPT_UNKNOWN)) { 305 unsigned DiagID; 306 auto ArgString = A->getAsString(Args); 307 std::string Nearest; 308 if (getOpts().findNearest( 309 ArgString, Nearest, IncludedFlagsBitmask, ExcludedFlagsBitmask) > 1) { 310 DiagID = IsCLMode() ? diag::warn_drv_unknown_argument_clang_cl 311 : diag::err_drv_unknown_argument; 312 Diags.Report(DiagID) << ArgString; 313 } else { 314 DiagID = IsCLMode() 315 ? diag::warn_drv_unknown_argument_clang_cl_with_suggestion 316 : diag::err_drv_unknown_argument_with_suggestion; 317 Diags.Report(DiagID) << ArgString << Nearest; 318 } 319 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 320 DiagnosticsEngine::Warning; 321 } 322 323 return Args; 324 } 325 326 // Determine which compilation mode we are in. We look for options which 327 // affect the phase, starting with the earliest phases, and record which 328 // option we used to determine the final phase. 329 phases::ID Driver::getFinalPhase(const DerivedArgList &DAL, 330 Arg **FinalPhaseArg) const { 331 Arg *PhaseArg = nullptr; 332 phases::ID FinalPhase; 333 334 // -{E,EP,P,M,MM} only run the preprocessor. 335 if (CCCIsCPP() || (PhaseArg = DAL.getLastArg(options::OPT_E)) || 336 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_EP)) || 337 (PhaseArg = DAL.getLastArg(options::OPT_M, options::OPT_MM)) || 338 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_P)) || 339 CCGenDiagnostics) { 340 FinalPhase = phases::Preprocess; 341 342 // --precompile only runs up to precompilation. 343 // Options that cause the output of C++20 compiled module interfaces or 344 // header units have the same effect. 345 } else if ((PhaseArg = DAL.getLastArg(options::OPT__precompile)) || 346 (PhaseArg = DAL.getLastArg(options::OPT_extract_api)) || 347 (PhaseArg = DAL.getLastArg(options::OPT_fmodule_header, 348 options::OPT_fmodule_header_EQ))) { 349 FinalPhase = phases::Precompile; 350 // -{fsyntax-only,-analyze,emit-ast} only run up to the compiler. 351 } else if ((PhaseArg = DAL.getLastArg(options::OPT_fsyntax_only)) || 352 (PhaseArg = DAL.getLastArg(options::OPT_print_supported_cpus)) || 353 (PhaseArg = DAL.getLastArg(options::OPT_module_file_info)) || 354 (PhaseArg = DAL.getLastArg(options::OPT_verify_pch)) || 355 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_objc)) || 356 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_legacy_objc)) || 357 (PhaseArg = DAL.getLastArg(options::OPT__migrate)) || 358 (PhaseArg = DAL.getLastArg(options::OPT__analyze)) || 359 (PhaseArg = DAL.getLastArg(options::OPT_emit_ast))) { 360 FinalPhase = phases::Compile; 361 362 // -S only runs up to the backend. 363 } else if ((PhaseArg = DAL.getLastArg(options::OPT_S))) { 364 FinalPhase = phases::Backend; 365 366 // -c compilation only runs up to the assembler. 367 } else if ((PhaseArg = DAL.getLastArg(options::OPT_c))) { 368 FinalPhase = phases::Assemble; 369 370 } else if ((PhaseArg = DAL.getLastArg(options::OPT_emit_interface_stubs))) { 371 FinalPhase = phases::IfsMerge; 372 373 // Otherwise do everything. 374 } else 375 FinalPhase = phases::Link; 376 377 if (FinalPhaseArg) 378 *FinalPhaseArg = PhaseArg; 379 380 return FinalPhase; 381 } 382 383 static Arg *MakeInputArg(DerivedArgList &Args, const OptTable &Opts, 384 StringRef Value, bool Claim = true) { 385 Arg *A = new Arg(Opts.getOption(options::OPT_INPUT), Value, 386 Args.getBaseArgs().MakeIndex(Value), Value.data()); 387 Args.AddSynthesizedArg(A); 388 if (Claim) 389 A->claim(); 390 return A; 391 } 392 393 DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args) const { 394 const llvm::opt::OptTable &Opts = getOpts(); 395 DerivedArgList *DAL = new DerivedArgList(Args); 396 397 bool HasNostdlib = Args.hasArg(options::OPT_nostdlib); 398 bool HasNostdlibxx = Args.hasArg(options::OPT_nostdlibxx); 399 bool HasNodefaultlib = Args.hasArg(options::OPT_nodefaultlibs); 400 bool IgnoreUnused = false; 401 for (Arg *A : Args) { 402 if (IgnoreUnused) 403 A->claim(); 404 405 if (A->getOption().matches(options::OPT_start_no_unused_arguments)) { 406 IgnoreUnused = true; 407 continue; 408 } 409 if (A->getOption().matches(options::OPT_end_no_unused_arguments)) { 410 IgnoreUnused = false; 411 continue; 412 } 413 414 // Unfortunately, we have to parse some forwarding options (-Xassembler, 415 // -Xlinker, -Xpreprocessor) because we either integrate their functionality 416 // (assembler and preprocessor), or bypass a previous driver ('collect2'). 417 418 // Rewrite linker options, to replace --no-demangle with a custom internal 419 // option. 420 if ((A->getOption().matches(options::OPT_Wl_COMMA) || 421 A->getOption().matches(options::OPT_Xlinker)) && 422 A->containsValue("--no-demangle")) { 423 // Add the rewritten no-demangle argument. 424 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_Xlinker__no_demangle)); 425 426 // Add the remaining values as Xlinker arguments. 427 for (StringRef Val : A->getValues()) 428 if (Val != "--no-demangle") 429 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_Xlinker), Val); 430 431 continue; 432 } 433 434 // Rewrite preprocessor options, to replace -Wp,-MD,FOO which is used by 435 // some build systems. We don't try to be complete here because we don't 436 // care to encourage this usage model. 437 if (A->getOption().matches(options::OPT_Wp_COMMA) && 438 (A->getValue(0) == StringRef("-MD") || 439 A->getValue(0) == StringRef("-MMD"))) { 440 // Rewrite to -MD/-MMD along with -MF. 441 if (A->getValue(0) == StringRef("-MD")) 442 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MD)); 443 else 444 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MMD)); 445 if (A->getNumValues() == 2) 446 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_MF), A->getValue(1)); 447 continue; 448 } 449 450 // Rewrite reserved library names. 451 if (A->getOption().matches(options::OPT_l)) { 452 StringRef Value = A->getValue(); 453 454 // Rewrite unless -nostdlib is present. 455 if (!HasNostdlib && !HasNodefaultlib && !HasNostdlibxx && 456 Value == "stdc++") { 457 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_stdcxx)); 458 continue; 459 } 460 461 // Rewrite unconditionally. 462 if (Value == "cc_kext") { 463 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_cckext)); 464 continue; 465 } 466 } 467 468 // Pick up inputs via the -- option. 469 if (A->getOption().matches(options::OPT__DASH_DASH)) { 470 A->claim(); 471 for (StringRef Val : A->getValues()) 472 DAL->append(MakeInputArg(*DAL, Opts, Val, false)); 473 continue; 474 } 475 476 DAL->append(A); 477 } 478 479 // Enforce -static if -miamcu is present. 480 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) 481 DAL->AddFlagArg(nullptr, Opts.getOption(options::OPT_static)); 482 483 // Add a default value of -mlinker-version=, if one was given and the user 484 // didn't specify one. 485 #if defined(HOST_LINK_VERSION) 486 if (!Args.hasArg(options::OPT_mlinker_version_EQ) && 487 strlen(HOST_LINK_VERSION) > 0) { 488 DAL->AddJoinedArg(0, Opts.getOption(options::OPT_mlinker_version_EQ), 489 HOST_LINK_VERSION); 490 DAL->getLastArg(options::OPT_mlinker_version_EQ)->claim(); 491 } 492 #endif 493 494 return DAL; 495 } 496 497 /// Compute target triple from args. 498 /// 499 /// This routine provides the logic to compute a target triple from various 500 /// args passed to the driver and the default triple string. 501 static llvm::Triple computeTargetTriple(const Driver &D, 502 StringRef TargetTriple, 503 const ArgList &Args, 504 StringRef DarwinArchName = "") { 505 // FIXME: Already done in Compilation *Driver::BuildCompilation 506 if (const Arg *A = Args.getLastArg(options::OPT_target)) 507 TargetTriple = A->getValue(); 508 509 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 510 511 // GNU/Hurd's triples should have been -hurd-gnu*, but were historically made 512 // -gnu* only, and we can not change this, so we have to detect that case as 513 // being the Hurd OS. 514 if (TargetTriple.contains("-unknown-gnu") || TargetTriple.contains("-pc-gnu")) 515 Target.setOSName("hurd"); 516 517 // Handle Apple-specific options available here. 518 if (Target.isOSBinFormatMachO()) { 519 // If an explicit Darwin arch name is given, that trumps all. 520 if (!DarwinArchName.empty()) { 521 tools::darwin::setTripleTypeForMachOArchName(Target, DarwinArchName); 522 return Target; 523 } 524 525 // Handle the Darwin '-arch' flag. 526 if (Arg *A = Args.getLastArg(options::OPT_arch)) { 527 StringRef ArchName = A->getValue(); 528 tools::darwin::setTripleTypeForMachOArchName(Target, ArchName); 529 } 530 } 531 532 // Handle pseudo-target flags '-mlittle-endian'/'-EL' and 533 // '-mbig-endian'/'-EB'. 534 if (Arg *A = Args.getLastArg(options::OPT_mlittle_endian, 535 options::OPT_mbig_endian)) { 536 if (A->getOption().matches(options::OPT_mlittle_endian)) { 537 llvm::Triple LE = Target.getLittleEndianArchVariant(); 538 if (LE.getArch() != llvm::Triple::UnknownArch) 539 Target = std::move(LE); 540 } else { 541 llvm::Triple BE = Target.getBigEndianArchVariant(); 542 if (BE.getArch() != llvm::Triple::UnknownArch) 543 Target = std::move(BE); 544 } 545 } 546 547 // Skip further flag support on OSes which don't support '-m32' or '-m64'. 548 if (Target.getArch() == llvm::Triple::tce || 549 Target.getOS() == llvm::Triple::Minix) 550 return Target; 551 552 // On AIX, the env OBJECT_MODE may affect the resulting arch variant. 553 if (Target.isOSAIX()) { 554 if (Optional<std::string> ObjectModeValue = 555 llvm::sys::Process::GetEnv("OBJECT_MODE")) { 556 StringRef ObjectMode = *ObjectModeValue; 557 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 558 559 if (ObjectMode.equals("64")) { 560 AT = Target.get64BitArchVariant().getArch(); 561 } else if (ObjectMode.equals("32")) { 562 AT = Target.get32BitArchVariant().getArch(); 563 } else { 564 D.Diag(diag::err_drv_invalid_object_mode) << ObjectMode; 565 } 566 567 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) 568 Target.setArch(AT); 569 } 570 } 571 572 // Handle pseudo-target flags '-m64', '-mx32', '-m32' and '-m16'. 573 Arg *A = Args.getLastArg(options::OPT_m64, options::OPT_mx32, 574 options::OPT_m32, options::OPT_m16); 575 if (A) { 576 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 577 578 if (A->getOption().matches(options::OPT_m64)) { 579 AT = Target.get64BitArchVariant().getArch(); 580 if (Target.getEnvironment() == llvm::Triple::GNUX32) 581 Target.setEnvironment(llvm::Triple::GNU); 582 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 583 Target.setEnvironment(llvm::Triple::Musl); 584 } else if (A->getOption().matches(options::OPT_mx32) && 585 Target.get64BitArchVariant().getArch() == llvm::Triple::x86_64) { 586 AT = llvm::Triple::x86_64; 587 if (Target.getEnvironment() == llvm::Triple::Musl) 588 Target.setEnvironment(llvm::Triple::MuslX32); 589 else 590 Target.setEnvironment(llvm::Triple::GNUX32); 591 } else if (A->getOption().matches(options::OPT_m32)) { 592 AT = Target.get32BitArchVariant().getArch(); 593 if (Target.getEnvironment() == llvm::Triple::GNUX32) 594 Target.setEnvironment(llvm::Triple::GNU); 595 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 596 Target.setEnvironment(llvm::Triple::Musl); 597 } else if (A->getOption().matches(options::OPT_m16) && 598 Target.get32BitArchVariant().getArch() == llvm::Triple::x86) { 599 AT = llvm::Triple::x86; 600 Target.setEnvironment(llvm::Triple::CODE16); 601 } 602 603 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) { 604 Target.setArch(AT); 605 if (Target.isWindowsGNUEnvironment()) 606 toolchains::MinGW::fixTripleArch(D, Target, Args); 607 } 608 } 609 610 // Handle -miamcu flag. 611 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) { 612 if (Target.get32BitArchVariant().getArch() != llvm::Triple::x86) 613 D.Diag(diag::err_drv_unsupported_opt_for_target) << "-miamcu" 614 << Target.str(); 615 616 if (A && !A->getOption().matches(options::OPT_m32)) 617 D.Diag(diag::err_drv_argument_not_allowed_with) 618 << "-miamcu" << A->getBaseArg().getAsString(Args); 619 620 Target.setArch(llvm::Triple::x86); 621 Target.setArchName("i586"); 622 Target.setEnvironment(llvm::Triple::UnknownEnvironment); 623 Target.setEnvironmentName(""); 624 Target.setOS(llvm::Triple::ELFIAMCU); 625 Target.setVendor(llvm::Triple::UnknownVendor); 626 Target.setVendorName("intel"); 627 } 628 629 // If target is MIPS adjust the target triple 630 // accordingly to provided ABI name. 631 A = Args.getLastArg(options::OPT_mabi_EQ); 632 if (A && Target.isMIPS()) { 633 StringRef ABIName = A->getValue(); 634 if (ABIName == "32") { 635 Target = Target.get32BitArchVariant(); 636 if (Target.getEnvironment() == llvm::Triple::GNUABI64 || 637 Target.getEnvironment() == llvm::Triple::GNUABIN32) 638 Target.setEnvironment(llvm::Triple::GNU); 639 } else if (ABIName == "n32") { 640 Target = Target.get64BitArchVariant(); 641 if (Target.getEnvironment() == llvm::Triple::GNU || 642 Target.getEnvironment() == llvm::Triple::GNUABI64) 643 Target.setEnvironment(llvm::Triple::GNUABIN32); 644 } else if (ABIName == "64") { 645 Target = Target.get64BitArchVariant(); 646 if (Target.getEnvironment() == llvm::Triple::GNU || 647 Target.getEnvironment() == llvm::Triple::GNUABIN32) 648 Target.setEnvironment(llvm::Triple::GNUABI64); 649 } 650 } 651 652 // If target is RISC-V adjust the target triple according to 653 // provided architecture name 654 A = Args.getLastArg(options::OPT_march_EQ); 655 if (A && Target.isRISCV()) { 656 StringRef ArchName = A->getValue(); 657 if (ArchName.startswith_insensitive("rv32")) 658 Target.setArch(llvm::Triple::riscv32); 659 else if (ArchName.startswith_insensitive("rv64")) 660 Target.setArch(llvm::Triple::riscv64); 661 } 662 663 return Target; 664 } 665 666 // Parse the LTO options and record the type of LTO compilation 667 // based on which -f(no-)?lto(=.*)? or -f(no-)?offload-lto(=.*)? 668 // option occurs last. 669 static driver::LTOKind parseLTOMode(Driver &D, const llvm::opt::ArgList &Args, 670 OptSpecifier OptEq, OptSpecifier OptNeg) { 671 if (!Args.hasFlag(OptEq, OptNeg, false)) 672 return LTOK_None; 673 674 const Arg *A = Args.getLastArg(OptEq); 675 StringRef LTOName = A->getValue(); 676 677 driver::LTOKind LTOMode = llvm::StringSwitch<LTOKind>(LTOName) 678 .Case("full", LTOK_Full) 679 .Case("thin", LTOK_Thin) 680 .Default(LTOK_Unknown); 681 682 if (LTOMode == LTOK_Unknown) { 683 D.Diag(diag::err_drv_unsupported_option_argument) 684 << A->getOption().getName() << A->getValue(); 685 return LTOK_None; 686 } 687 return LTOMode; 688 } 689 690 // Parse the LTO options. 691 void Driver::setLTOMode(const llvm::opt::ArgList &Args) { 692 LTOMode = 693 parseLTOMode(*this, Args, options::OPT_flto_EQ, options::OPT_fno_lto); 694 695 OffloadLTOMode = parseLTOMode(*this, Args, options::OPT_foffload_lto_EQ, 696 options::OPT_fno_offload_lto); 697 } 698 699 /// Compute the desired OpenMP runtime from the flags provided. 700 Driver::OpenMPRuntimeKind Driver::getOpenMPRuntime(const ArgList &Args) const { 701 StringRef RuntimeName(CLANG_DEFAULT_OPENMP_RUNTIME); 702 703 const Arg *A = Args.getLastArg(options::OPT_fopenmp_EQ); 704 if (A) 705 RuntimeName = A->getValue(); 706 707 auto RT = llvm::StringSwitch<OpenMPRuntimeKind>(RuntimeName) 708 .Case("libomp", OMPRT_OMP) 709 .Case("libgomp", OMPRT_GOMP) 710 .Case("libiomp5", OMPRT_IOMP5) 711 .Default(OMPRT_Unknown); 712 713 if (RT == OMPRT_Unknown) { 714 if (A) 715 Diag(diag::err_drv_unsupported_option_argument) 716 << A->getOption().getName() << A->getValue(); 717 else 718 // FIXME: We could use a nicer diagnostic here. 719 Diag(diag::err_drv_unsupported_opt) << "-fopenmp"; 720 } 721 722 return RT; 723 } 724 725 void Driver::CreateOffloadingDeviceToolChains(Compilation &C, 726 InputList &Inputs) { 727 728 // 729 // CUDA/HIP 730 // 731 // We need to generate a CUDA/HIP toolchain if any of the inputs has a CUDA 732 // or HIP type. However, mixed CUDA/HIP compilation is not supported. 733 bool IsCuda = 734 llvm::any_of(Inputs, [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 735 return types::isCuda(I.first); 736 }); 737 bool IsHIP = 738 llvm::any_of(Inputs, 739 [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 740 return types::isHIP(I.first); 741 }) || 742 C.getInputArgs().hasArg(options::OPT_hip_link); 743 if (IsCuda && IsHIP) { 744 Diag(clang::diag::err_drv_mix_cuda_hip); 745 return; 746 } 747 if (IsCuda) { 748 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 749 const llvm::Triple &HostTriple = HostTC->getTriple(); 750 auto OFK = Action::OFK_Cuda; 751 auto CudaTriple = 752 getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), HostTriple); 753 if (!CudaTriple) 754 return; 755 // Use the CUDA and host triples as the key into the ToolChains map, 756 // because the device toolchain we create depends on both. 757 auto &CudaTC = ToolChains[CudaTriple->str() + "/" + HostTriple.str()]; 758 if (!CudaTC) { 759 CudaTC = std::make_unique<toolchains::CudaToolChain>( 760 *this, *CudaTriple, *HostTC, C.getInputArgs(), OFK); 761 } 762 C.addOffloadDeviceToolChain(CudaTC.get(), OFK); 763 } else if (IsHIP) { 764 if (auto *OMPTargetArg = 765 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 766 Diag(clang::diag::err_drv_unsupported_opt_for_language_mode) 767 << OMPTargetArg->getSpelling() << "HIP"; 768 return; 769 } 770 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 771 auto OFK = Action::OFK_HIP; 772 auto HIPTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 773 if (!HIPTriple) 774 return; 775 auto *HIPTC = &getOffloadingDeviceToolChain(C.getInputArgs(), *HIPTriple, 776 *HostTC, OFK); 777 assert(HIPTC && "Could not create offloading device tool chain."); 778 C.addOffloadDeviceToolChain(HIPTC, OFK); 779 } 780 781 // 782 // OpenMP 783 // 784 // We need to generate an OpenMP toolchain if the user specified targets with 785 // the -fopenmp-targets option or used --offload-arch with OpenMP enabled. 786 bool IsOpenMPOffloading = 787 C.getInputArgs().hasFlag(options::OPT_fopenmp, options::OPT_fopenmp_EQ, 788 options::OPT_fno_openmp, false) && 789 (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ) || 790 C.getInputArgs().hasArg(options::OPT_offload_arch_EQ)); 791 if (IsOpenMPOffloading) { 792 // We expect that -fopenmp-targets is always used in conjunction with the 793 // option -fopenmp specifying a valid runtime with offloading support, i.e. 794 // libomp or libiomp. 795 OpenMPRuntimeKind RuntimeKind = getOpenMPRuntime(C.getInputArgs()); 796 if (RuntimeKind != OMPRT_OMP && RuntimeKind != OMPRT_IOMP5) { 797 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 798 return; 799 } 800 801 llvm::StringMap<llvm::DenseSet<StringRef>> DerivedArchs; 802 llvm::StringMap<StringRef> FoundNormalizedTriples; 803 llvm::SmallVector<StringRef, 4> OpenMPTriples; 804 805 // If the user specified -fopenmp-targets= we create a toolchain for each 806 // valid triple. Otherwise, if only --offload-arch= was specified we instead 807 // attempt to derive the appropriate toolchains from the arguments. 808 if (Arg *OpenMPTargets = 809 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 810 if (OpenMPTargets && !OpenMPTargets->getNumValues()) { 811 Diag(clang::diag::warn_drv_empty_joined_argument) 812 << OpenMPTargets->getAsString(C.getInputArgs()); 813 return; 814 } 815 llvm::copy(OpenMPTargets->getValues(), std::back_inserter(OpenMPTriples)); 816 } else if (C.getInputArgs().hasArg(options::OPT_offload_arch_EQ) && 817 !IsHIP && !IsCuda) { 818 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 819 auto AMDTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 820 auto NVPTXTriple = getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), 821 HostTC->getTriple()); 822 823 // Attempt to deduce the offloading triple from the set of architectures. 824 // We can only correctly deduce NVPTX / AMDGPU triples currently. 825 llvm::DenseSet<StringRef> Archs = 826 getOffloadArchs(C, C.getArgs(), Action::OFK_OpenMP, nullptr); 827 for (StringRef Arch : Archs) { 828 if (NVPTXTriple && IsNVIDIAGpuArch(StringToCudaArch( 829 getProcessorFromTargetID(*NVPTXTriple, Arch)))) { 830 DerivedArchs[NVPTXTriple->getTriple()].insert(Arch); 831 } else if (AMDTriple && 832 IsAMDGpuArch(StringToCudaArch( 833 getProcessorFromTargetID(*AMDTriple, Arch)))) { 834 DerivedArchs[AMDTriple->getTriple()].insert(Arch); 835 } else { 836 Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch) << Arch; 837 return; 838 } 839 } 840 841 for (const auto &TripleAndArchs : DerivedArchs) 842 OpenMPTriples.push_back(TripleAndArchs.first()); 843 } 844 845 for (StringRef Val : OpenMPTriples) { 846 llvm::Triple TT(ToolChain::getOpenMPTriple(Val)); 847 std::string NormalizedName = TT.normalize(); 848 849 // Make sure we don't have a duplicate triple. 850 auto Duplicate = FoundNormalizedTriples.find(NormalizedName); 851 if (Duplicate != FoundNormalizedTriples.end()) { 852 Diag(clang::diag::warn_drv_omp_offload_target_duplicate) 853 << Val << Duplicate->second; 854 continue; 855 } 856 857 // Store the current triple so that we can check for duplicates in the 858 // following iterations. 859 FoundNormalizedTriples[NormalizedName] = Val; 860 861 // If the specified target is invalid, emit a diagnostic. 862 if (TT.getArch() == llvm::Triple::UnknownArch) 863 Diag(clang::diag::err_drv_invalid_omp_target) << Val; 864 else { 865 const ToolChain *TC; 866 // Device toolchains have to be selected differently. They pair host 867 // and device in their implementation. 868 if (TT.isNVPTX() || TT.isAMDGCN()) { 869 const ToolChain *HostTC = 870 C.getSingleOffloadToolChain<Action::OFK_Host>(); 871 assert(HostTC && "Host toolchain should be always defined."); 872 auto &DeviceTC = 873 ToolChains[TT.str() + "/" + HostTC->getTriple().normalize()]; 874 if (!DeviceTC) { 875 if (TT.isNVPTX()) 876 DeviceTC = std::make_unique<toolchains::CudaToolChain>( 877 *this, TT, *HostTC, C.getInputArgs(), Action::OFK_OpenMP); 878 else if (TT.isAMDGCN()) 879 DeviceTC = std::make_unique<toolchains::AMDGPUOpenMPToolChain>( 880 *this, TT, *HostTC, C.getInputArgs()); 881 else 882 assert(DeviceTC && "Device toolchain not defined."); 883 } 884 885 TC = DeviceTC.get(); 886 } else 887 TC = &getToolChain(C.getInputArgs(), TT); 888 C.addOffloadDeviceToolChain(TC, Action::OFK_OpenMP); 889 if (DerivedArchs.find(TT.getTriple()) != DerivedArchs.end()) 890 KnownArchs[TC] = DerivedArchs[TT.getTriple()]; 891 } 892 } 893 } else if (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ)) { 894 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 895 return; 896 } 897 898 // 899 // TODO: Add support for other offloading programming models here. 900 // 901 } 902 903 /// Looks the given directories for the specified file. 904 /// 905 /// \param[out] FilePath File path, if the file was found. 906 /// \param[in] Dirs Directories used for the search. 907 /// \param[in] FileName Name of the file to search for. 908 /// \return True if file was found. 909 /// 910 /// Looks for file specified by FileName sequentially in directories specified 911 /// by Dirs. 912 /// 913 static bool searchForFile(SmallVectorImpl<char> &FilePath, 914 ArrayRef<StringRef> Dirs, StringRef FileName) { 915 SmallString<128> WPath; 916 for (const StringRef &Dir : Dirs) { 917 if (Dir.empty()) 918 continue; 919 WPath.clear(); 920 llvm::sys::path::append(WPath, Dir, FileName); 921 llvm::sys::path::native(WPath); 922 if (llvm::sys::fs::is_regular_file(WPath)) { 923 FilePath = std::move(WPath); 924 return true; 925 } 926 } 927 return false; 928 } 929 930 bool Driver::readConfigFile(StringRef FileName) { 931 // Try reading the given file. 932 SmallVector<const char *, 32> NewCfgArgs; 933 if (!llvm::cl::readConfigFile(FileName, Saver, NewCfgArgs)) { 934 Diag(diag::err_drv_cannot_read_config_file) << FileName; 935 return true; 936 } 937 938 // Read options from config file. 939 llvm::SmallString<128> CfgFileName(FileName); 940 llvm::sys::path::native(CfgFileName); 941 ConfigFile = std::string(CfgFileName); 942 bool ContainErrors; 943 CfgOptions = std::make_unique<InputArgList>( 944 ParseArgStrings(NewCfgArgs, IsCLMode(), ContainErrors)); 945 if (ContainErrors) { 946 CfgOptions.reset(); 947 return true; 948 } 949 950 if (CfgOptions->hasArg(options::OPT_config)) { 951 CfgOptions.reset(); 952 Diag(diag::err_drv_nested_config_file); 953 return true; 954 } 955 956 // Claim all arguments that come from a configuration file so that the driver 957 // does not warn on any that is unused. 958 for (Arg *A : *CfgOptions) 959 A->claim(); 960 return false; 961 } 962 963 bool Driver::loadConfigFile() { 964 std::string CfgFileName; 965 bool FileSpecifiedExplicitly = false; 966 967 // Process options that change search path for config files. 968 if (CLOptions) { 969 if (CLOptions->hasArg(options::OPT_config_system_dir_EQ)) { 970 SmallString<128> CfgDir; 971 CfgDir.append( 972 CLOptions->getLastArgValue(options::OPT_config_system_dir_EQ)); 973 if (!CfgDir.empty()) { 974 if (llvm::sys::fs::make_absolute(CfgDir).value() != 0) 975 SystemConfigDir.clear(); 976 else 977 SystemConfigDir = std::string(CfgDir.begin(), CfgDir.end()); 978 } 979 } 980 if (CLOptions->hasArg(options::OPT_config_user_dir_EQ)) { 981 SmallString<128> CfgDir; 982 CfgDir.append( 983 CLOptions->getLastArgValue(options::OPT_config_user_dir_EQ)); 984 if (!CfgDir.empty()) { 985 if (llvm::sys::fs::make_absolute(CfgDir).value() != 0) 986 UserConfigDir.clear(); 987 else 988 UserConfigDir = std::string(CfgDir.begin(), CfgDir.end()); 989 } 990 } 991 } 992 993 // First try to find config file specified in command line. 994 if (CLOptions) { 995 std::vector<std::string> ConfigFiles = 996 CLOptions->getAllArgValues(options::OPT_config); 997 if (ConfigFiles.size() > 1) { 998 if (!llvm::all_of(ConfigFiles, [ConfigFiles](const std::string &s) { 999 return s == ConfigFiles[0]; 1000 })) { 1001 Diag(diag::err_drv_duplicate_config); 1002 return true; 1003 } 1004 } 1005 1006 if (!ConfigFiles.empty()) { 1007 CfgFileName = ConfigFiles.front(); 1008 assert(!CfgFileName.empty()); 1009 1010 // If argument contains directory separator, treat it as a path to 1011 // configuration file. 1012 if (llvm::sys::path::has_parent_path(CfgFileName)) { 1013 SmallString<128> CfgFilePath; 1014 if (llvm::sys::path::is_relative(CfgFileName)) 1015 llvm::sys::fs::current_path(CfgFilePath); 1016 llvm::sys::path::append(CfgFilePath, CfgFileName); 1017 if (!llvm::sys::fs::is_regular_file(CfgFilePath)) { 1018 Diag(diag::err_drv_config_file_not_exist) << CfgFilePath; 1019 return true; 1020 } 1021 return readConfigFile(CfgFilePath); 1022 } 1023 1024 FileSpecifiedExplicitly = true; 1025 } 1026 } 1027 1028 // If config file is not specified explicitly, try to deduce configuration 1029 // from executable name. For instance, an executable 'armv7l-clang' will 1030 // search for config file 'armv7l-clang.cfg'. 1031 if (CfgFileName.empty() && !ClangNameParts.TargetPrefix.empty()) 1032 CfgFileName = ClangNameParts.TargetPrefix + '-' + ClangNameParts.ModeSuffix; 1033 1034 if (CfgFileName.empty()) 1035 return false; 1036 1037 // Determine architecture part of the file name, if it is present. 1038 StringRef CfgFileArch = CfgFileName; 1039 size_t ArchPrefixLen = CfgFileArch.find('-'); 1040 if (ArchPrefixLen == StringRef::npos) 1041 ArchPrefixLen = CfgFileArch.size(); 1042 llvm::Triple CfgTriple; 1043 CfgFileArch = CfgFileArch.take_front(ArchPrefixLen); 1044 CfgTriple = llvm::Triple(llvm::Triple::normalize(CfgFileArch)); 1045 if (CfgTriple.getArch() == llvm::Triple::ArchType::UnknownArch) 1046 ArchPrefixLen = 0; 1047 1048 if (!StringRef(CfgFileName).endswith(".cfg")) 1049 CfgFileName += ".cfg"; 1050 1051 // If config file starts with architecture name and command line options 1052 // redefine architecture (with options like -m32 -LE etc), try finding new 1053 // config file with that architecture. 1054 SmallString<128> FixedConfigFile; 1055 size_t FixedArchPrefixLen = 0; 1056 if (ArchPrefixLen) { 1057 // Get architecture name from config file name like 'i386.cfg' or 1058 // 'armv7l-clang.cfg'. 1059 // Check if command line options changes effective triple. 1060 llvm::Triple EffectiveTriple = computeTargetTriple(*this, 1061 CfgTriple.getTriple(), *CLOptions); 1062 if (CfgTriple.getArch() != EffectiveTriple.getArch()) { 1063 FixedConfigFile = EffectiveTriple.getArchName(); 1064 FixedArchPrefixLen = FixedConfigFile.size(); 1065 // Append the rest of original file name so that file name transforms 1066 // like: i386-clang.cfg -> x86_64-clang.cfg. 1067 if (ArchPrefixLen < CfgFileName.size()) 1068 FixedConfigFile += CfgFileName.substr(ArchPrefixLen); 1069 } 1070 } 1071 1072 // Prepare list of directories where config file is searched for. 1073 StringRef CfgFileSearchDirs[] = {UserConfigDir, SystemConfigDir, Dir}; 1074 1075 // Try to find config file. First try file with corrected architecture. 1076 llvm::SmallString<128> CfgFilePath; 1077 if (!FixedConfigFile.empty()) { 1078 if (searchForFile(CfgFilePath, CfgFileSearchDirs, FixedConfigFile)) 1079 return readConfigFile(CfgFilePath); 1080 // If 'x86_64-clang.cfg' was not found, try 'x86_64.cfg'. 1081 FixedConfigFile.resize(FixedArchPrefixLen); 1082 FixedConfigFile.append(".cfg"); 1083 if (searchForFile(CfgFilePath, CfgFileSearchDirs, FixedConfigFile)) 1084 return readConfigFile(CfgFilePath); 1085 } 1086 1087 // Then try original file name. 1088 if (searchForFile(CfgFilePath, CfgFileSearchDirs, CfgFileName)) 1089 return readConfigFile(CfgFilePath); 1090 1091 // Finally try removing driver mode part: 'x86_64-clang.cfg' -> 'x86_64.cfg'. 1092 if (!ClangNameParts.ModeSuffix.empty() && 1093 !ClangNameParts.TargetPrefix.empty()) { 1094 CfgFileName.assign(ClangNameParts.TargetPrefix); 1095 CfgFileName.append(".cfg"); 1096 if (searchForFile(CfgFilePath, CfgFileSearchDirs, CfgFileName)) 1097 return readConfigFile(CfgFilePath); 1098 } 1099 1100 // Report error but only if config file was specified explicitly, by option 1101 // --config. If it was deduced from executable name, it is not an error. 1102 if (FileSpecifiedExplicitly) { 1103 Diag(diag::err_drv_config_file_not_found) << CfgFileName; 1104 for (const StringRef &SearchDir : CfgFileSearchDirs) 1105 if (!SearchDir.empty()) 1106 Diag(diag::note_drv_config_file_searched_in) << SearchDir; 1107 return true; 1108 } 1109 1110 return false; 1111 } 1112 1113 Compilation *Driver::BuildCompilation(ArrayRef<const char *> ArgList) { 1114 llvm::PrettyStackTraceString CrashInfo("Compilation construction"); 1115 1116 // FIXME: Handle environment options which affect driver behavior, somewhere 1117 // (client?). GCC_EXEC_PREFIX, LPATH, CC_PRINT_OPTIONS. 1118 1119 // We look for the driver mode option early, because the mode can affect 1120 // how other options are parsed. 1121 1122 auto DriverMode = getDriverMode(ClangExecutable, ArgList.slice(1)); 1123 if (!DriverMode.empty()) 1124 setDriverMode(DriverMode); 1125 1126 // FIXME: What are we going to do with -V and -b? 1127 1128 // Arguments specified in command line. 1129 bool ContainsError; 1130 CLOptions = std::make_unique<InputArgList>( 1131 ParseArgStrings(ArgList.slice(1), IsCLMode(), ContainsError)); 1132 1133 // Try parsing configuration file. 1134 if (!ContainsError) 1135 ContainsError = loadConfigFile(); 1136 bool HasConfigFile = !ContainsError && (CfgOptions.get() != nullptr); 1137 1138 // All arguments, from both config file and command line. 1139 InputArgList Args = std::move(HasConfigFile ? std::move(*CfgOptions) 1140 : std::move(*CLOptions)); 1141 1142 // The args for config files or /clang: flags belong to different InputArgList 1143 // objects than Args. This copies an Arg from one of those other InputArgLists 1144 // to the ownership of Args. 1145 auto appendOneArg = [&Args](const Arg *Opt, const Arg *BaseArg) { 1146 unsigned Index = Args.MakeIndex(Opt->getSpelling()); 1147 Arg *Copy = new llvm::opt::Arg(Opt->getOption(), Args.getArgString(Index), 1148 Index, BaseArg); 1149 Copy->getValues() = Opt->getValues(); 1150 if (Opt->isClaimed()) 1151 Copy->claim(); 1152 Copy->setOwnsValues(Opt->getOwnsValues()); 1153 Opt->setOwnsValues(false); 1154 Args.append(Copy); 1155 }; 1156 1157 if (HasConfigFile) 1158 for (auto *Opt : *CLOptions) { 1159 if (Opt->getOption().matches(options::OPT_config)) 1160 continue; 1161 const Arg *BaseArg = &Opt->getBaseArg(); 1162 if (BaseArg == Opt) 1163 BaseArg = nullptr; 1164 appendOneArg(Opt, BaseArg); 1165 } 1166 1167 // In CL mode, look for any pass-through arguments 1168 if (IsCLMode() && !ContainsError) { 1169 SmallVector<const char *, 16> CLModePassThroughArgList; 1170 for (const auto *A : Args.filtered(options::OPT__SLASH_clang)) { 1171 A->claim(); 1172 CLModePassThroughArgList.push_back(A->getValue()); 1173 } 1174 1175 if (!CLModePassThroughArgList.empty()) { 1176 // Parse any pass through args using default clang processing rather 1177 // than clang-cl processing. 1178 auto CLModePassThroughOptions = std::make_unique<InputArgList>( 1179 ParseArgStrings(CLModePassThroughArgList, false, ContainsError)); 1180 1181 if (!ContainsError) 1182 for (auto *Opt : *CLModePassThroughOptions) { 1183 appendOneArg(Opt, nullptr); 1184 } 1185 } 1186 } 1187 1188 // Check for working directory option before accessing any files 1189 if (Arg *WD = Args.getLastArg(options::OPT_working_directory)) 1190 if (VFS->setCurrentWorkingDirectory(WD->getValue())) 1191 Diag(diag::err_drv_unable_to_set_working_directory) << WD->getValue(); 1192 1193 // FIXME: This stuff needs to go into the Compilation, not the driver. 1194 bool CCCPrintPhases; 1195 1196 // Silence driver warnings if requested 1197 Diags.setIgnoreAllWarnings(Args.hasArg(options::OPT_w)); 1198 1199 // -canonical-prefixes, -no-canonical-prefixes are used very early in main. 1200 Args.ClaimAllArgs(options::OPT_canonical_prefixes); 1201 Args.ClaimAllArgs(options::OPT_no_canonical_prefixes); 1202 1203 // f(no-)integated-cc1 is also used very early in main. 1204 Args.ClaimAllArgs(options::OPT_fintegrated_cc1); 1205 Args.ClaimAllArgs(options::OPT_fno_integrated_cc1); 1206 1207 // Ignore -pipe. 1208 Args.ClaimAllArgs(options::OPT_pipe); 1209 1210 // Extract -ccc args. 1211 // 1212 // FIXME: We need to figure out where this behavior should live. Most of it 1213 // should be outside in the client; the parts that aren't should have proper 1214 // options, either by introducing new ones or by overloading gcc ones like -V 1215 // or -b. 1216 CCCPrintPhases = Args.hasArg(options::OPT_ccc_print_phases); 1217 CCCPrintBindings = Args.hasArg(options::OPT_ccc_print_bindings); 1218 if (const Arg *A = Args.getLastArg(options::OPT_ccc_gcc_name)) 1219 CCCGenericGCCName = A->getValue(); 1220 GenReproducer = Args.hasFlag(options::OPT_gen_reproducer, 1221 options::OPT_fno_crash_diagnostics, 1222 !!::getenv("FORCE_CLANG_DIAGNOSTICS_CRASH")); 1223 1224 // Process -fproc-stat-report options. 1225 if (const Arg *A = Args.getLastArg(options::OPT_fproc_stat_report_EQ)) { 1226 CCPrintProcessStats = true; 1227 CCPrintStatReportFilename = A->getValue(); 1228 } 1229 if (Args.hasArg(options::OPT_fproc_stat_report)) 1230 CCPrintProcessStats = true; 1231 1232 // FIXME: TargetTriple is used by the target-prefixed calls to as/ld 1233 // and getToolChain is const. 1234 if (IsCLMode()) { 1235 // clang-cl targets MSVC-style Win32. 1236 llvm::Triple T(TargetTriple); 1237 T.setOS(llvm::Triple::Win32); 1238 T.setVendor(llvm::Triple::PC); 1239 T.setEnvironment(llvm::Triple::MSVC); 1240 T.setObjectFormat(llvm::Triple::COFF); 1241 TargetTriple = T.str(); 1242 } else if (IsDXCMode()) { 1243 // clang-dxc target is build from target_profile option. 1244 // Just set OS to shader model to select HLSLToolChain. 1245 llvm::Triple T(TargetTriple); 1246 T.setOS(llvm::Triple::ShaderModel); 1247 TargetTriple = T.str(); 1248 } 1249 1250 if (const Arg *A = Args.getLastArg(options::OPT_target)) 1251 TargetTriple = A->getValue(); 1252 if (const Arg *A = Args.getLastArg(options::OPT_ccc_install_dir)) 1253 Dir = InstalledDir = A->getValue(); 1254 for (const Arg *A : Args.filtered(options::OPT_B)) { 1255 A->claim(); 1256 PrefixDirs.push_back(A->getValue(0)); 1257 } 1258 if (Optional<std::string> CompilerPathValue = 1259 llvm::sys::Process::GetEnv("COMPILER_PATH")) { 1260 StringRef CompilerPath = *CompilerPathValue; 1261 while (!CompilerPath.empty()) { 1262 std::pair<StringRef, StringRef> Split = 1263 CompilerPath.split(llvm::sys::EnvPathSeparator); 1264 PrefixDirs.push_back(std::string(Split.first)); 1265 CompilerPath = Split.second; 1266 } 1267 } 1268 if (const Arg *A = Args.getLastArg(options::OPT__sysroot_EQ)) 1269 SysRoot = A->getValue(); 1270 if (const Arg *A = Args.getLastArg(options::OPT__dyld_prefix_EQ)) 1271 DyldPrefix = A->getValue(); 1272 1273 if (const Arg *A = Args.getLastArg(options::OPT_resource_dir)) 1274 ResourceDir = A->getValue(); 1275 1276 if (const Arg *A = Args.getLastArg(options::OPT_save_temps_EQ)) { 1277 SaveTemps = llvm::StringSwitch<SaveTempsMode>(A->getValue()) 1278 .Case("cwd", SaveTempsCwd) 1279 .Case("obj", SaveTempsObj) 1280 .Default(SaveTempsCwd); 1281 } 1282 1283 setLTOMode(Args); 1284 1285 // Process -fembed-bitcode= flags. 1286 if (Arg *A = Args.getLastArg(options::OPT_fembed_bitcode_EQ)) { 1287 StringRef Name = A->getValue(); 1288 unsigned Model = llvm::StringSwitch<unsigned>(Name) 1289 .Case("off", EmbedNone) 1290 .Case("all", EmbedBitcode) 1291 .Case("bitcode", EmbedBitcode) 1292 .Case("marker", EmbedMarker) 1293 .Default(~0U); 1294 if (Model == ~0U) { 1295 Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) 1296 << Name; 1297 } else 1298 BitcodeEmbed = static_cast<BitcodeEmbedMode>(Model); 1299 } 1300 1301 // Setting up the jobs for some precompile cases depends on whether we are 1302 // treating them as PCH, implicit modules or C++20 ones. 1303 // TODO: inferring the mode like this seems fragile (it meets the objective 1304 // of not requiring anything new for operation, however). 1305 const Arg *Std = Args.getLastArg(options::OPT_std_EQ); 1306 ModulesModeCXX20 = 1307 !Args.hasArg(options::OPT_fmodules) && Std && 1308 (Std->containsValue("c++20") || Std->containsValue("c++2b") || 1309 Std->containsValue("c++2a") || Std->containsValue("c++latest")); 1310 1311 // Process -fmodule-header{=} flags. 1312 if (Arg *A = Args.getLastArg(options::OPT_fmodule_header_EQ, 1313 options::OPT_fmodule_header)) { 1314 // These flags force C++20 handling of headers. 1315 ModulesModeCXX20 = true; 1316 if (A->getOption().matches(options::OPT_fmodule_header)) 1317 CXX20HeaderType = HeaderMode_Default; 1318 else { 1319 StringRef ArgName = A->getValue(); 1320 unsigned Kind = llvm::StringSwitch<unsigned>(ArgName) 1321 .Case("user", HeaderMode_User) 1322 .Case("system", HeaderMode_System) 1323 .Default(~0U); 1324 if (Kind == ~0U) { 1325 Diags.Report(diag::err_drv_invalid_value) 1326 << A->getAsString(Args) << ArgName; 1327 } else 1328 CXX20HeaderType = static_cast<ModuleHeaderMode>(Kind); 1329 } 1330 } 1331 1332 std::unique_ptr<llvm::opt::InputArgList> UArgs = 1333 std::make_unique<InputArgList>(std::move(Args)); 1334 1335 // Perform the default argument translations. 1336 DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs); 1337 1338 // Owned by the host. 1339 const ToolChain &TC = getToolChain( 1340 *UArgs, computeTargetTriple(*this, TargetTriple, *UArgs)); 1341 1342 // The compilation takes ownership of Args. 1343 Compilation *C = new Compilation(*this, TC, UArgs.release(), TranslatedArgs, 1344 ContainsError); 1345 1346 if (!HandleImmediateArgs(*C)) 1347 return C; 1348 1349 // Construct the list of inputs. 1350 InputList Inputs; 1351 BuildInputs(C->getDefaultToolChain(), *TranslatedArgs, Inputs); 1352 1353 // Populate the tool chains for the offloading devices, if any. 1354 CreateOffloadingDeviceToolChains(*C, Inputs); 1355 1356 // Construct the list of abstract actions to perform for this compilation. On 1357 // MachO targets this uses the driver-driver and universal actions. 1358 if (TC.getTriple().isOSBinFormatMachO()) 1359 BuildUniversalActions(*C, C->getDefaultToolChain(), Inputs); 1360 else 1361 BuildActions(*C, C->getArgs(), Inputs, C->getActions()); 1362 1363 if (CCCPrintPhases) { 1364 PrintActions(*C); 1365 return C; 1366 } 1367 1368 BuildJobs(*C); 1369 1370 return C; 1371 } 1372 1373 static void printArgList(raw_ostream &OS, const llvm::opt::ArgList &Args) { 1374 llvm::opt::ArgStringList ASL; 1375 for (const auto *A : Args) { 1376 // Use user's original spelling of flags. For example, use 1377 // `/source-charset:utf-8` instead of `-finput-charset=utf-8` if the user 1378 // wrote the former. 1379 while (A->getAlias()) 1380 A = A->getAlias(); 1381 A->render(Args, ASL); 1382 } 1383 1384 for (auto I = ASL.begin(), E = ASL.end(); I != E; ++I) { 1385 if (I != ASL.begin()) 1386 OS << ' '; 1387 llvm::sys::printArg(OS, *I, true); 1388 } 1389 OS << '\n'; 1390 } 1391 1392 bool Driver::getCrashDiagnosticFile(StringRef ReproCrashFilename, 1393 SmallString<128> &CrashDiagDir) { 1394 using namespace llvm::sys; 1395 assert(llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin() && 1396 "Only knows about .crash files on Darwin"); 1397 1398 // The .crash file can be found on at ~/Library/Logs/DiagnosticReports/ 1399 // (or /Library/Logs/DiagnosticReports for root) and has the filename pattern 1400 // clang-<VERSION>_<YYYY-MM-DD-HHMMSS>_<hostname>.crash. 1401 path::home_directory(CrashDiagDir); 1402 if (CrashDiagDir.startswith("/var/root")) 1403 CrashDiagDir = "/"; 1404 path::append(CrashDiagDir, "Library/Logs/DiagnosticReports"); 1405 int PID = 1406 #if LLVM_ON_UNIX 1407 getpid(); 1408 #else 1409 0; 1410 #endif 1411 std::error_code EC; 1412 fs::file_status FileStatus; 1413 TimePoint<> LastAccessTime; 1414 SmallString<128> CrashFilePath; 1415 // Lookup the .crash files and get the one generated by a subprocess spawned 1416 // by this driver invocation. 1417 for (fs::directory_iterator File(CrashDiagDir, EC), FileEnd; 1418 File != FileEnd && !EC; File.increment(EC)) { 1419 StringRef FileName = path::filename(File->path()); 1420 if (!FileName.startswith(Name)) 1421 continue; 1422 if (fs::status(File->path(), FileStatus)) 1423 continue; 1424 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CrashFile = 1425 llvm::MemoryBuffer::getFile(File->path()); 1426 if (!CrashFile) 1427 continue; 1428 // The first line should start with "Process:", otherwise this isn't a real 1429 // .crash file. 1430 StringRef Data = CrashFile.get()->getBuffer(); 1431 if (!Data.startswith("Process:")) 1432 continue; 1433 // Parse parent process pid line, e.g: "Parent Process: clang-4.0 [79141]" 1434 size_t ParentProcPos = Data.find("Parent Process:"); 1435 if (ParentProcPos == StringRef::npos) 1436 continue; 1437 size_t LineEnd = Data.find_first_of("\n", ParentProcPos); 1438 if (LineEnd == StringRef::npos) 1439 continue; 1440 StringRef ParentProcess = Data.slice(ParentProcPos+15, LineEnd).trim(); 1441 int OpenBracket = -1, CloseBracket = -1; 1442 for (size_t i = 0, e = ParentProcess.size(); i < e; ++i) { 1443 if (ParentProcess[i] == '[') 1444 OpenBracket = i; 1445 if (ParentProcess[i] == ']') 1446 CloseBracket = i; 1447 } 1448 // Extract the parent process PID from the .crash file and check whether 1449 // it matches this driver invocation pid. 1450 int CrashPID; 1451 if (OpenBracket < 0 || CloseBracket < 0 || 1452 ParentProcess.slice(OpenBracket + 1, CloseBracket) 1453 .getAsInteger(10, CrashPID) || CrashPID != PID) { 1454 continue; 1455 } 1456 1457 // Found a .crash file matching the driver pid. To avoid getting an older 1458 // and misleading crash file, continue looking for the most recent. 1459 // FIXME: the driver can dispatch multiple cc1 invocations, leading to 1460 // multiple crashes poiting to the same parent process. Since the driver 1461 // does not collect pid information for the dispatched invocation there's 1462 // currently no way to distinguish among them. 1463 const auto FileAccessTime = FileStatus.getLastModificationTime(); 1464 if (FileAccessTime > LastAccessTime) { 1465 CrashFilePath.assign(File->path()); 1466 LastAccessTime = FileAccessTime; 1467 } 1468 } 1469 1470 // If found, copy it over to the location of other reproducer files. 1471 if (!CrashFilePath.empty()) { 1472 EC = fs::copy_file(CrashFilePath, ReproCrashFilename); 1473 if (EC) 1474 return false; 1475 return true; 1476 } 1477 1478 return false; 1479 } 1480 1481 // When clang crashes, produce diagnostic information including the fully 1482 // preprocessed source file(s). Request that the developer attach the 1483 // diagnostic information to a bug report. 1484 void Driver::generateCompilationDiagnostics( 1485 Compilation &C, const Command &FailingCommand, 1486 StringRef AdditionalInformation, CompilationDiagnosticReport *Report) { 1487 if (C.getArgs().hasArg(options::OPT_fno_crash_diagnostics)) 1488 return; 1489 1490 // Don't try to generate diagnostics for link or dsymutil jobs. 1491 if (FailingCommand.getCreator().isLinkJob() || 1492 FailingCommand.getCreator().isDsymutilJob()) 1493 return; 1494 1495 // Print the version of the compiler. 1496 PrintVersion(C, llvm::errs()); 1497 1498 // Suppress driver output and emit preprocessor output to temp file. 1499 CCGenDiagnostics = true; 1500 1501 // Save the original job command(s). 1502 Command Cmd = FailingCommand; 1503 1504 // Keep track of whether we produce any errors while trying to produce 1505 // preprocessed sources. 1506 DiagnosticErrorTrap Trap(Diags); 1507 1508 // Suppress tool output. 1509 C.initCompilationForDiagnostics(); 1510 1511 // Construct the list of inputs. 1512 InputList Inputs; 1513 BuildInputs(C.getDefaultToolChain(), C.getArgs(), Inputs); 1514 1515 for (InputList::iterator it = Inputs.begin(), ie = Inputs.end(); it != ie;) { 1516 bool IgnoreInput = false; 1517 1518 // Ignore input from stdin or any inputs that cannot be preprocessed. 1519 // Check type first as not all linker inputs have a value. 1520 if (types::getPreprocessedType(it->first) == types::TY_INVALID) { 1521 IgnoreInput = true; 1522 } else if (!strcmp(it->second->getValue(), "-")) { 1523 Diag(clang::diag::note_drv_command_failed_diag_msg) 1524 << "Error generating preprocessed source(s) - " 1525 "ignoring input from stdin."; 1526 IgnoreInput = true; 1527 } 1528 1529 if (IgnoreInput) { 1530 it = Inputs.erase(it); 1531 ie = Inputs.end(); 1532 } else { 1533 ++it; 1534 } 1535 } 1536 1537 if (Inputs.empty()) { 1538 Diag(clang::diag::note_drv_command_failed_diag_msg) 1539 << "Error generating preprocessed source(s) - " 1540 "no preprocessable inputs."; 1541 return; 1542 } 1543 1544 // Don't attempt to generate preprocessed files if multiple -arch options are 1545 // used, unless they're all duplicates. 1546 llvm::StringSet<> ArchNames; 1547 for (const Arg *A : C.getArgs()) { 1548 if (A->getOption().matches(options::OPT_arch)) { 1549 StringRef ArchName = A->getValue(); 1550 ArchNames.insert(ArchName); 1551 } 1552 } 1553 if (ArchNames.size() > 1) { 1554 Diag(clang::diag::note_drv_command_failed_diag_msg) 1555 << "Error generating preprocessed source(s) - cannot generate " 1556 "preprocessed source with multiple -arch options."; 1557 return; 1558 } 1559 1560 // Construct the list of abstract actions to perform for this compilation. On 1561 // Darwin OSes this uses the driver-driver and builds universal actions. 1562 const ToolChain &TC = C.getDefaultToolChain(); 1563 if (TC.getTriple().isOSBinFormatMachO()) 1564 BuildUniversalActions(C, TC, Inputs); 1565 else 1566 BuildActions(C, C.getArgs(), Inputs, C.getActions()); 1567 1568 BuildJobs(C); 1569 1570 // If there were errors building the compilation, quit now. 1571 if (Trap.hasErrorOccurred()) { 1572 Diag(clang::diag::note_drv_command_failed_diag_msg) 1573 << "Error generating preprocessed source(s)."; 1574 return; 1575 } 1576 1577 // Generate preprocessed output. 1578 SmallVector<std::pair<int, const Command *>, 4> FailingCommands; 1579 C.ExecuteJobs(C.getJobs(), FailingCommands); 1580 1581 // If any of the preprocessing commands failed, clean up and exit. 1582 if (!FailingCommands.empty()) { 1583 Diag(clang::diag::note_drv_command_failed_diag_msg) 1584 << "Error generating preprocessed source(s)."; 1585 return; 1586 } 1587 1588 const ArgStringList &TempFiles = C.getTempFiles(); 1589 if (TempFiles.empty()) { 1590 Diag(clang::diag::note_drv_command_failed_diag_msg) 1591 << "Error generating preprocessed source(s)."; 1592 return; 1593 } 1594 1595 Diag(clang::diag::note_drv_command_failed_diag_msg) 1596 << "\n********************\n\n" 1597 "PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:\n" 1598 "Preprocessed source(s) and associated run script(s) are located at:"; 1599 1600 SmallString<128> VFS; 1601 SmallString<128> ReproCrashFilename; 1602 for (const char *TempFile : TempFiles) { 1603 Diag(clang::diag::note_drv_command_failed_diag_msg) << TempFile; 1604 if (Report) 1605 Report->TemporaryFiles.push_back(TempFile); 1606 if (ReproCrashFilename.empty()) { 1607 ReproCrashFilename = TempFile; 1608 llvm::sys::path::replace_extension(ReproCrashFilename, ".crash"); 1609 } 1610 if (StringRef(TempFile).endswith(".cache")) { 1611 // In some cases (modules) we'll dump extra data to help with reproducing 1612 // the crash into a directory next to the output. 1613 VFS = llvm::sys::path::filename(TempFile); 1614 llvm::sys::path::append(VFS, "vfs", "vfs.yaml"); 1615 } 1616 } 1617 1618 // Assume associated files are based off of the first temporary file. 1619 CrashReportInfo CrashInfo(TempFiles[0], VFS); 1620 1621 llvm::SmallString<128> Script(CrashInfo.Filename); 1622 llvm::sys::path::replace_extension(Script, "sh"); 1623 std::error_code EC; 1624 llvm::raw_fd_ostream ScriptOS(Script, EC, llvm::sys::fs::CD_CreateNew, 1625 llvm::sys::fs::FA_Write, 1626 llvm::sys::fs::OF_Text); 1627 if (EC) { 1628 Diag(clang::diag::note_drv_command_failed_diag_msg) 1629 << "Error generating run script: " << Script << " " << EC.message(); 1630 } else { 1631 ScriptOS << "# Crash reproducer for " << getClangFullVersion() << "\n" 1632 << "# Driver args: "; 1633 printArgList(ScriptOS, C.getInputArgs()); 1634 ScriptOS << "# Original command: "; 1635 Cmd.Print(ScriptOS, "\n", /*Quote=*/true); 1636 Cmd.Print(ScriptOS, "\n", /*Quote=*/true, &CrashInfo); 1637 if (!AdditionalInformation.empty()) 1638 ScriptOS << "\n# Additional information: " << AdditionalInformation 1639 << "\n"; 1640 if (Report) 1641 Report->TemporaryFiles.push_back(std::string(Script.str())); 1642 Diag(clang::diag::note_drv_command_failed_diag_msg) << Script; 1643 } 1644 1645 // On darwin, provide information about the .crash diagnostic report. 1646 if (llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin()) { 1647 SmallString<128> CrashDiagDir; 1648 if (getCrashDiagnosticFile(ReproCrashFilename, CrashDiagDir)) { 1649 Diag(clang::diag::note_drv_command_failed_diag_msg) 1650 << ReproCrashFilename.str(); 1651 } else { // Suggest a directory for the user to look for .crash files. 1652 llvm::sys::path::append(CrashDiagDir, Name); 1653 CrashDiagDir += "_<YYYY-MM-DD-HHMMSS>_<hostname>.crash"; 1654 Diag(clang::diag::note_drv_command_failed_diag_msg) 1655 << "Crash backtrace is located in"; 1656 Diag(clang::diag::note_drv_command_failed_diag_msg) 1657 << CrashDiagDir.str(); 1658 Diag(clang::diag::note_drv_command_failed_diag_msg) 1659 << "(choose the .crash file that corresponds to your crash)"; 1660 } 1661 } 1662 1663 for (const auto &A : C.getArgs().filtered(options::OPT_frewrite_map_file_EQ)) 1664 Diag(clang::diag::note_drv_command_failed_diag_msg) << A->getValue(); 1665 1666 Diag(clang::diag::note_drv_command_failed_diag_msg) 1667 << "\n\n********************"; 1668 } 1669 1670 void Driver::setUpResponseFiles(Compilation &C, Command &Cmd) { 1671 // Since commandLineFitsWithinSystemLimits() may underestimate system's 1672 // capacity if the tool does not support response files, there is a chance/ 1673 // that things will just work without a response file, so we silently just 1674 // skip it. 1675 if (Cmd.getResponseFileSupport().ResponseKind == 1676 ResponseFileSupport::RF_None || 1677 llvm::sys::commandLineFitsWithinSystemLimits(Cmd.getExecutable(), 1678 Cmd.getArguments())) 1679 return; 1680 1681 std::string TmpName = GetTemporaryPath("response", "txt"); 1682 Cmd.setResponseFile(C.addTempFile(C.getArgs().MakeArgString(TmpName))); 1683 } 1684 1685 int Driver::ExecuteCompilation( 1686 Compilation &C, 1687 SmallVectorImpl<std::pair<int, const Command *>> &FailingCommands) { 1688 // Just print if -### was present. 1689 if (C.getArgs().hasArg(options::OPT__HASH_HASH_HASH)) { 1690 C.getJobs().Print(llvm::errs(), "\n", true); 1691 return 0; 1692 } 1693 1694 // If there were errors building the compilation, quit now. 1695 if (Diags.hasErrorOccurred()) 1696 return 1; 1697 1698 // Set up response file names for each command, if necessary. 1699 for (auto &Job : C.getJobs()) 1700 setUpResponseFiles(C, Job); 1701 1702 C.ExecuteJobs(C.getJobs(), FailingCommands); 1703 1704 // If the command succeeded, we are done. 1705 if (FailingCommands.empty()) 1706 return 0; 1707 1708 // Otherwise, remove result files and print extra information about abnormal 1709 // failures. 1710 int Res = 0; 1711 for (const auto &CmdPair : FailingCommands) { 1712 int CommandRes = CmdPair.first; 1713 const Command *FailingCommand = CmdPair.second; 1714 1715 // Remove result files if we're not saving temps. 1716 if (!isSaveTempsEnabled()) { 1717 const JobAction *JA = cast<JobAction>(&FailingCommand->getSource()); 1718 C.CleanupFileMap(C.getResultFiles(), JA, true); 1719 1720 // Failure result files are valid unless we crashed. 1721 if (CommandRes < 0) 1722 C.CleanupFileMap(C.getFailureResultFiles(), JA, true); 1723 } 1724 1725 #if LLVM_ON_UNIX 1726 // llvm/lib/Support/Unix/Signals.inc will exit with a special return code 1727 // for SIGPIPE. Do not print diagnostics for this case. 1728 if (CommandRes == EX_IOERR) { 1729 Res = CommandRes; 1730 continue; 1731 } 1732 #endif 1733 1734 // Print extra information about abnormal failures, if possible. 1735 // 1736 // This is ad-hoc, but we don't want to be excessively noisy. If the result 1737 // status was 1, assume the command failed normally. In particular, if it 1738 // was the compiler then assume it gave a reasonable error code. Failures 1739 // in other tools are less common, and they generally have worse 1740 // diagnostics, so always print the diagnostic there. 1741 const Tool &FailingTool = FailingCommand->getCreator(); 1742 1743 if (!FailingCommand->getCreator().hasGoodDiagnostics() || CommandRes != 1) { 1744 // FIXME: See FIXME above regarding result code interpretation. 1745 if (CommandRes < 0) 1746 Diag(clang::diag::err_drv_command_signalled) 1747 << FailingTool.getShortName(); 1748 else 1749 Diag(clang::diag::err_drv_command_failed) 1750 << FailingTool.getShortName() << CommandRes; 1751 } 1752 } 1753 return Res; 1754 } 1755 1756 void Driver::PrintHelp(bool ShowHidden) const { 1757 unsigned IncludedFlagsBitmask; 1758 unsigned ExcludedFlagsBitmask; 1759 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 1760 getIncludeExcludeOptionFlagMasks(IsCLMode()); 1761 1762 ExcludedFlagsBitmask |= options::NoDriverOption; 1763 if (!ShowHidden) 1764 ExcludedFlagsBitmask |= HelpHidden; 1765 1766 if (IsFlangMode()) 1767 IncludedFlagsBitmask |= options::FlangOption; 1768 else 1769 ExcludedFlagsBitmask |= options::FlangOnlyOption; 1770 1771 std::string Usage = llvm::formatv("{0} [options] file...", Name).str(); 1772 getOpts().printHelp(llvm::outs(), Usage.c_str(), DriverTitle.c_str(), 1773 IncludedFlagsBitmask, ExcludedFlagsBitmask, 1774 /*ShowAllAliases=*/false); 1775 } 1776 1777 void Driver::PrintVersion(const Compilation &C, raw_ostream &OS) const { 1778 if (IsFlangMode()) { 1779 OS << getClangToolFullVersion("flang-new") << '\n'; 1780 } else { 1781 // FIXME: The following handlers should use a callback mechanism, we don't 1782 // know what the client would like to do. 1783 OS << getClangFullVersion() << '\n'; 1784 } 1785 const ToolChain &TC = C.getDefaultToolChain(); 1786 OS << "Target: " << TC.getTripleString() << '\n'; 1787 1788 // Print the threading model. 1789 if (Arg *A = C.getArgs().getLastArg(options::OPT_mthread_model)) { 1790 // Don't print if the ToolChain would have barfed on it already 1791 if (TC.isThreadModelSupported(A->getValue())) 1792 OS << "Thread model: " << A->getValue(); 1793 } else 1794 OS << "Thread model: " << TC.getThreadModel(); 1795 OS << '\n'; 1796 1797 // Print out the install directory. 1798 OS << "InstalledDir: " << InstalledDir << '\n'; 1799 1800 // If configuration file was used, print its path. 1801 if (!ConfigFile.empty()) 1802 OS << "Configuration file: " << ConfigFile << '\n'; 1803 } 1804 1805 /// PrintDiagnosticCategories - Implement the --print-diagnostic-categories 1806 /// option. 1807 static void PrintDiagnosticCategories(raw_ostream &OS) { 1808 // Skip the empty category. 1809 for (unsigned i = 1, max = DiagnosticIDs::getNumberOfCategories(); i != max; 1810 ++i) 1811 OS << i << ',' << DiagnosticIDs::getCategoryNameFromID(i) << '\n'; 1812 } 1813 1814 void Driver::HandleAutocompletions(StringRef PassedFlags) const { 1815 if (PassedFlags == "") 1816 return; 1817 // Print out all options that start with a given argument. This is used for 1818 // shell autocompletion. 1819 std::vector<std::string> SuggestedCompletions; 1820 std::vector<std::string> Flags; 1821 1822 unsigned int DisableFlags = 1823 options::NoDriverOption | options::Unsupported | options::Ignored; 1824 1825 // Make sure that Flang-only options don't pollute the Clang output 1826 // TODO: Make sure that Clang-only options don't pollute Flang output 1827 if (!IsFlangMode()) 1828 DisableFlags |= options::FlangOnlyOption; 1829 1830 // Distinguish "--autocomplete=-someflag" and "--autocomplete=-someflag," 1831 // because the latter indicates that the user put space before pushing tab 1832 // which should end up in a file completion. 1833 const bool HasSpace = PassedFlags.endswith(","); 1834 1835 // Parse PassedFlags by "," as all the command-line flags are passed to this 1836 // function separated by "," 1837 StringRef TargetFlags = PassedFlags; 1838 while (TargetFlags != "") { 1839 StringRef CurFlag; 1840 std::tie(CurFlag, TargetFlags) = TargetFlags.split(","); 1841 Flags.push_back(std::string(CurFlag)); 1842 } 1843 1844 // We want to show cc1-only options only when clang is invoked with -cc1 or 1845 // -Xclang. 1846 if (llvm::is_contained(Flags, "-Xclang") || llvm::is_contained(Flags, "-cc1")) 1847 DisableFlags &= ~options::NoDriverOption; 1848 1849 const llvm::opt::OptTable &Opts = getOpts(); 1850 StringRef Cur; 1851 Cur = Flags.at(Flags.size() - 1); 1852 StringRef Prev; 1853 if (Flags.size() >= 2) { 1854 Prev = Flags.at(Flags.size() - 2); 1855 SuggestedCompletions = Opts.suggestValueCompletions(Prev, Cur); 1856 } 1857 1858 if (SuggestedCompletions.empty()) 1859 SuggestedCompletions = Opts.suggestValueCompletions(Cur, ""); 1860 1861 // If Flags were empty, it means the user typed `clang [tab]` where we should 1862 // list all possible flags. If there was no value completion and the user 1863 // pressed tab after a space, we should fall back to a file completion. 1864 // We're printing a newline to be consistent with what we print at the end of 1865 // this function. 1866 if (SuggestedCompletions.empty() && HasSpace && !Flags.empty()) { 1867 llvm::outs() << '\n'; 1868 return; 1869 } 1870 1871 // When flag ends with '=' and there was no value completion, return empty 1872 // string and fall back to the file autocompletion. 1873 if (SuggestedCompletions.empty() && !Cur.endswith("=")) { 1874 // If the flag is in the form of "--autocomplete=-foo", 1875 // we were requested to print out all option names that start with "-foo". 1876 // For example, "--autocomplete=-fsyn" is expanded to "-fsyntax-only". 1877 SuggestedCompletions = Opts.findByPrefix(Cur, DisableFlags); 1878 1879 // We have to query the -W flags manually as they're not in the OptTable. 1880 // TODO: Find a good way to add them to OptTable instead and them remove 1881 // this code. 1882 for (StringRef S : DiagnosticIDs::getDiagnosticFlags()) 1883 if (S.startswith(Cur)) 1884 SuggestedCompletions.push_back(std::string(S)); 1885 } 1886 1887 // Sort the autocomplete candidates so that shells print them out in a 1888 // deterministic order. We could sort in any way, but we chose 1889 // case-insensitive sorting for consistency with the -help option 1890 // which prints out options in the case-insensitive alphabetical order. 1891 llvm::sort(SuggestedCompletions, [](StringRef A, StringRef B) { 1892 if (int X = A.compare_insensitive(B)) 1893 return X < 0; 1894 return A.compare(B) > 0; 1895 }); 1896 1897 llvm::outs() << llvm::join(SuggestedCompletions, "\n") << '\n'; 1898 } 1899 1900 bool Driver::HandleImmediateArgs(const Compilation &C) { 1901 // The order these options are handled in gcc is all over the place, but we 1902 // don't expect inconsistencies w.r.t. that to matter in practice. 1903 1904 if (C.getArgs().hasArg(options::OPT_dumpmachine)) { 1905 llvm::outs() << C.getDefaultToolChain().getTripleString() << '\n'; 1906 return false; 1907 } 1908 1909 if (C.getArgs().hasArg(options::OPT_dumpversion)) { 1910 // Since -dumpversion is only implemented for pedantic GCC compatibility, we 1911 // return an answer which matches our definition of __VERSION__. 1912 llvm::outs() << CLANG_VERSION_STRING << "\n"; 1913 return false; 1914 } 1915 1916 if (C.getArgs().hasArg(options::OPT__print_diagnostic_categories)) { 1917 PrintDiagnosticCategories(llvm::outs()); 1918 return false; 1919 } 1920 1921 if (C.getArgs().hasArg(options::OPT_help) || 1922 C.getArgs().hasArg(options::OPT__help_hidden)) { 1923 PrintHelp(C.getArgs().hasArg(options::OPT__help_hidden)); 1924 return false; 1925 } 1926 1927 if (C.getArgs().hasArg(options::OPT__version)) { 1928 // Follow gcc behavior and use stdout for --version and stderr for -v. 1929 PrintVersion(C, llvm::outs()); 1930 return false; 1931 } 1932 1933 if (C.getArgs().hasArg(options::OPT_v) || 1934 C.getArgs().hasArg(options::OPT__HASH_HASH_HASH) || 1935 C.getArgs().hasArg(options::OPT_print_supported_cpus)) { 1936 PrintVersion(C, llvm::errs()); 1937 SuppressMissingInputWarning = true; 1938 } 1939 1940 if (C.getArgs().hasArg(options::OPT_v)) { 1941 if (!SystemConfigDir.empty()) 1942 llvm::errs() << "System configuration file directory: " 1943 << SystemConfigDir << "\n"; 1944 if (!UserConfigDir.empty()) 1945 llvm::errs() << "User configuration file directory: " 1946 << UserConfigDir << "\n"; 1947 } 1948 1949 const ToolChain &TC = C.getDefaultToolChain(); 1950 1951 if (C.getArgs().hasArg(options::OPT_v)) 1952 TC.printVerboseInfo(llvm::errs()); 1953 1954 if (C.getArgs().hasArg(options::OPT_print_resource_dir)) { 1955 llvm::outs() << ResourceDir << '\n'; 1956 return false; 1957 } 1958 1959 if (C.getArgs().hasArg(options::OPT_print_search_dirs)) { 1960 llvm::outs() << "programs: ="; 1961 bool separator = false; 1962 // Print -B and COMPILER_PATH. 1963 for (const std::string &Path : PrefixDirs) { 1964 if (separator) 1965 llvm::outs() << llvm::sys::EnvPathSeparator; 1966 llvm::outs() << Path; 1967 separator = true; 1968 } 1969 for (const std::string &Path : TC.getProgramPaths()) { 1970 if (separator) 1971 llvm::outs() << llvm::sys::EnvPathSeparator; 1972 llvm::outs() << Path; 1973 separator = true; 1974 } 1975 llvm::outs() << "\n"; 1976 llvm::outs() << "libraries: =" << ResourceDir; 1977 1978 StringRef sysroot = C.getSysRoot(); 1979 1980 for (const std::string &Path : TC.getFilePaths()) { 1981 // Always print a separator. ResourceDir was the first item shown. 1982 llvm::outs() << llvm::sys::EnvPathSeparator; 1983 // Interpretation of leading '=' is needed only for NetBSD. 1984 if (Path[0] == '=') 1985 llvm::outs() << sysroot << Path.substr(1); 1986 else 1987 llvm::outs() << Path; 1988 } 1989 llvm::outs() << "\n"; 1990 return false; 1991 } 1992 1993 if (C.getArgs().hasArg(options::OPT_print_runtime_dir)) { 1994 std::string RuntimePath; 1995 // Get the first existing path, if any. 1996 for (auto Path : TC.getRuntimePaths()) { 1997 if (getVFS().exists(Path)) { 1998 RuntimePath = Path; 1999 break; 2000 } 2001 } 2002 if (!RuntimePath.empty()) 2003 llvm::outs() << RuntimePath << '\n'; 2004 else 2005 llvm::outs() << TC.getCompilerRTPath() << '\n'; 2006 return false; 2007 } 2008 2009 // FIXME: The following handlers should use a callback mechanism, we don't 2010 // know what the client would like to do. 2011 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_file_name_EQ)) { 2012 llvm::outs() << GetFilePath(A->getValue(), TC) << "\n"; 2013 return false; 2014 } 2015 2016 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_prog_name_EQ)) { 2017 StringRef ProgName = A->getValue(); 2018 2019 // Null program name cannot have a path. 2020 if (! ProgName.empty()) 2021 llvm::outs() << GetProgramPath(ProgName, TC); 2022 2023 llvm::outs() << "\n"; 2024 return false; 2025 } 2026 2027 if (Arg *A = C.getArgs().getLastArg(options::OPT_autocomplete)) { 2028 StringRef PassedFlags = A->getValue(); 2029 HandleAutocompletions(PassedFlags); 2030 return false; 2031 } 2032 2033 if (C.getArgs().hasArg(options::OPT_print_libgcc_file_name)) { 2034 ToolChain::RuntimeLibType RLT = TC.GetRuntimeLibType(C.getArgs()); 2035 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2036 RegisterEffectiveTriple TripleRAII(TC, Triple); 2037 switch (RLT) { 2038 case ToolChain::RLT_CompilerRT: 2039 llvm::outs() << TC.getCompilerRT(C.getArgs(), "builtins") << "\n"; 2040 break; 2041 case ToolChain::RLT_Libgcc: 2042 llvm::outs() << GetFilePath("libgcc.a", TC) << "\n"; 2043 break; 2044 } 2045 return false; 2046 } 2047 2048 if (C.getArgs().hasArg(options::OPT_print_multi_lib)) { 2049 for (const Multilib &Multilib : TC.getMultilibs()) 2050 llvm::outs() << Multilib << "\n"; 2051 return false; 2052 } 2053 2054 if (C.getArgs().hasArg(options::OPT_print_multi_directory)) { 2055 const Multilib &Multilib = TC.getMultilib(); 2056 if (Multilib.gccSuffix().empty()) 2057 llvm::outs() << ".\n"; 2058 else { 2059 StringRef Suffix(Multilib.gccSuffix()); 2060 assert(Suffix.front() == '/'); 2061 llvm::outs() << Suffix.substr(1) << "\n"; 2062 } 2063 return false; 2064 } 2065 2066 if (C.getArgs().hasArg(options::OPT_print_target_triple)) { 2067 llvm::outs() << TC.getTripleString() << "\n"; 2068 return false; 2069 } 2070 2071 if (C.getArgs().hasArg(options::OPT_print_effective_triple)) { 2072 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2073 llvm::outs() << Triple.getTriple() << "\n"; 2074 return false; 2075 } 2076 2077 if (C.getArgs().hasArg(options::OPT_print_multiarch)) { 2078 llvm::outs() << TC.getMultiarchTriple(*this, TC.getTriple(), SysRoot) 2079 << "\n"; 2080 return false; 2081 } 2082 2083 if (C.getArgs().hasArg(options::OPT_print_targets)) { 2084 llvm::TargetRegistry::printRegisteredTargetsForVersion(llvm::outs()); 2085 return false; 2086 } 2087 2088 return true; 2089 } 2090 2091 enum { 2092 TopLevelAction = 0, 2093 HeadSibAction = 1, 2094 OtherSibAction = 2, 2095 }; 2096 2097 // Display an action graph human-readably. Action A is the "sink" node 2098 // and latest-occuring action. Traversal is in pre-order, visiting the 2099 // inputs to each action before printing the action itself. 2100 static unsigned PrintActions1(const Compilation &C, Action *A, 2101 std::map<Action *, unsigned> &Ids, 2102 Twine Indent = {}, int Kind = TopLevelAction) { 2103 if (Ids.count(A)) // A was already visited. 2104 return Ids[A]; 2105 2106 std::string str; 2107 llvm::raw_string_ostream os(str); 2108 2109 auto getSibIndent = [](int K) -> Twine { 2110 return (K == HeadSibAction) ? " " : (K == OtherSibAction) ? "| " : ""; 2111 }; 2112 2113 Twine SibIndent = Indent + getSibIndent(Kind); 2114 int SibKind = HeadSibAction; 2115 os << Action::getClassName(A->getKind()) << ", "; 2116 if (InputAction *IA = dyn_cast<InputAction>(A)) { 2117 os << "\"" << IA->getInputArg().getValue() << "\""; 2118 } else if (BindArchAction *BIA = dyn_cast<BindArchAction>(A)) { 2119 os << '"' << BIA->getArchName() << '"' << ", {" 2120 << PrintActions1(C, *BIA->input_begin(), Ids, SibIndent, SibKind) << "}"; 2121 } else if (OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 2122 bool IsFirst = true; 2123 OA->doOnEachDependence( 2124 [&](Action *A, const ToolChain *TC, const char *BoundArch) { 2125 assert(TC && "Unknown host toolchain"); 2126 // E.g. for two CUDA device dependences whose bound arch is sm_20 and 2127 // sm_35 this will generate: 2128 // "cuda-device" (nvptx64-nvidia-cuda:sm_20) {#ID}, "cuda-device" 2129 // (nvptx64-nvidia-cuda:sm_35) {#ID} 2130 if (!IsFirst) 2131 os << ", "; 2132 os << '"'; 2133 os << A->getOffloadingKindPrefix(); 2134 os << " ("; 2135 os << TC->getTriple().normalize(); 2136 if (BoundArch) 2137 os << ":" << BoundArch; 2138 os << ")"; 2139 os << '"'; 2140 os << " {" << PrintActions1(C, A, Ids, SibIndent, SibKind) << "}"; 2141 IsFirst = false; 2142 SibKind = OtherSibAction; 2143 }); 2144 } else { 2145 const ActionList *AL = &A->getInputs(); 2146 2147 if (AL->size()) { 2148 const char *Prefix = "{"; 2149 for (Action *PreRequisite : *AL) { 2150 os << Prefix << PrintActions1(C, PreRequisite, Ids, SibIndent, SibKind); 2151 Prefix = ", "; 2152 SibKind = OtherSibAction; 2153 } 2154 os << "}"; 2155 } else 2156 os << "{}"; 2157 } 2158 2159 // Append offload info for all options other than the offloading action 2160 // itself (e.g. (cuda-device, sm_20) or (cuda-host)). 2161 std::string offload_str; 2162 llvm::raw_string_ostream offload_os(offload_str); 2163 if (!isa<OffloadAction>(A)) { 2164 auto S = A->getOffloadingKindPrefix(); 2165 if (!S.empty()) { 2166 offload_os << ", (" << S; 2167 if (A->getOffloadingArch()) 2168 offload_os << ", " << A->getOffloadingArch(); 2169 offload_os << ")"; 2170 } 2171 } 2172 2173 auto getSelfIndent = [](int K) -> Twine { 2174 return (K == HeadSibAction) ? "+- " : (K == OtherSibAction) ? "|- " : ""; 2175 }; 2176 2177 unsigned Id = Ids.size(); 2178 Ids[A] = Id; 2179 llvm::errs() << Indent + getSelfIndent(Kind) << Id << ": " << os.str() << ", " 2180 << types::getTypeName(A->getType()) << offload_os.str() << "\n"; 2181 2182 return Id; 2183 } 2184 2185 // Print the action graphs in a compilation C. 2186 // For example "clang -c file1.c file2.c" is composed of two subgraphs. 2187 void Driver::PrintActions(const Compilation &C) const { 2188 std::map<Action *, unsigned> Ids; 2189 for (Action *A : C.getActions()) 2190 PrintActions1(C, A, Ids); 2191 } 2192 2193 /// Check whether the given input tree contains any compilation or 2194 /// assembly actions. 2195 static bool ContainsCompileOrAssembleAction(const Action *A) { 2196 if (isa<CompileJobAction>(A) || isa<BackendJobAction>(A) || 2197 isa<AssembleJobAction>(A)) 2198 return true; 2199 2200 return llvm::any_of(A->inputs(), ContainsCompileOrAssembleAction); 2201 } 2202 2203 void Driver::BuildUniversalActions(Compilation &C, const ToolChain &TC, 2204 const InputList &BAInputs) const { 2205 DerivedArgList &Args = C.getArgs(); 2206 ActionList &Actions = C.getActions(); 2207 llvm::PrettyStackTraceString CrashInfo("Building universal build actions"); 2208 // Collect the list of architectures. Duplicates are allowed, but should only 2209 // be handled once (in the order seen). 2210 llvm::StringSet<> ArchNames; 2211 SmallVector<const char *, 4> Archs; 2212 for (Arg *A : Args) { 2213 if (A->getOption().matches(options::OPT_arch)) { 2214 // Validate the option here; we don't save the type here because its 2215 // particular spelling may participate in other driver choices. 2216 llvm::Triple::ArchType Arch = 2217 tools::darwin::getArchTypeForMachOArchName(A->getValue()); 2218 if (Arch == llvm::Triple::UnknownArch) { 2219 Diag(clang::diag::err_drv_invalid_arch_name) << A->getAsString(Args); 2220 continue; 2221 } 2222 2223 A->claim(); 2224 if (ArchNames.insert(A->getValue()).second) 2225 Archs.push_back(A->getValue()); 2226 } 2227 } 2228 2229 // When there is no explicit arch for this platform, make sure we still bind 2230 // the architecture (to the default) so that -Xarch_ is handled correctly. 2231 if (!Archs.size()) 2232 Archs.push_back(Args.MakeArgString(TC.getDefaultUniversalArchName())); 2233 2234 ActionList SingleActions; 2235 BuildActions(C, Args, BAInputs, SingleActions); 2236 2237 // Add in arch bindings for every top level action, as well as lipo and 2238 // dsymutil steps if needed. 2239 for (Action* Act : SingleActions) { 2240 // Make sure we can lipo this kind of output. If not (and it is an actual 2241 // output) then we disallow, since we can't create an output file with the 2242 // right name without overwriting it. We could remove this oddity by just 2243 // changing the output names to include the arch, which would also fix 2244 // -save-temps. Compatibility wins for now. 2245 2246 if (Archs.size() > 1 && !types::canLipoType(Act->getType())) 2247 Diag(clang::diag::err_drv_invalid_output_with_multiple_archs) 2248 << types::getTypeName(Act->getType()); 2249 2250 ActionList Inputs; 2251 for (unsigned i = 0, e = Archs.size(); i != e; ++i) 2252 Inputs.push_back(C.MakeAction<BindArchAction>(Act, Archs[i])); 2253 2254 // Lipo if necessary, we do it this way because we need to set the arch flag 2255 // so that -Xarch_ gets overwritten. 2256 if (Inputs.size() == 1 || Act->getType() == types::TY_Nothing) 2257 Actions.append(Inputs.begin(), Inputs.end()); 2258 else 2259 Actions.push_back(C.MakeAction<LipoJobAction>(Inputs, Act->getType())); 2260 2261 // Handle debug info queries. 2262 Arg *A = Args.getLastArg(options::OPT_g_Group); 2263 bool enablesDebugInfo = A && !A->getOption().matches(options::OPT_g0) && 2264 !A->getOption().matches(options::OPT_gstabs); 2265 if ((enablesDebugInfo || willEmitRemarks(Args)) && 2266 ContainsCompileOrAssembleAction(Actions.back())) { 2267 2268 // Add a 'dsymutil' step if necessary, when debug info is enabled and we 2269 // have a compile input. We need to run 'dsymutil' ourselves in such cases 2270 // because the debug info will refer to a temporary object file which 2271 // will be removed at the end of the compilation process. 2272 if (Act->getType() == types::TY_Image) { 2273 ActionList Inputs; 2274 Inputs.push_back(Actions.back()); 2275 Actions.pop_back(); 2276 Actions.push_back( 2277 C.MakeAction<DsymutilJobAction>(Inputs, types::TY_dSYM)); 2278 } 2279 2280 // Verify the debug info output. 2281 if (Args.hasArg(options::OPT_verify_debug_info)) { 2282 Action* LastAction = Actions.back(); 2283 Actions.pop_back(); 2284 Actions.push_back(C.MakeAction<VerifyDebugInfoJobAction>( 2285 LastAction, types::TY_Nothing)); 2286 } 2287 } 2288 } 2289 } 2290 2291 bool Driver::DiagnoseInputExistence(const DerivedArgList &Args, StringRef Value, 2292 types::ID Ty, bool TypoCorrect) const { 2293 if (!getCheckInputsExist()) 2294 return true; 2295 2296 // stdin always exists. 2297 if (Value == "-") 2298 return true; 2299 2300 // If it's a header to be found in the system or user search path, then defer 2301 // complaints about its absence until those searches can be done. When we 2302 // are definitely processing headers for C++20 header units, extend this to 2303 // allow the user to put "-fmodule-header -xc++-header vector" for example. 2304 if (Ty == types::TY_CXXSHeader || Ty == types::TY_CXXUHeader || 2305 (ModulesModeCXX20 && Ty == types::TY_CXXHeader)) 2306 return true; 2307 2308 if (getVFS().exists(Value)) 2309 return true; 2310 2311 if (TypoCorrect) { 2312 // Check if the filename is a typo for an option flag. OptTable thinks 2313 // that all args that are not known options and that start with / are 2314 // filenames, but e.g. `/diagnostic:caret` is more likely a typo for 2315 // the option `/diagnostics:caret` than a reference to a file in the root 2316 // directory. 2317 unsigned IncludedFlagsBitmask; 2318 unsigned ExcludedFlagsBitmask; 2319 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 2320 getIncludeExcludeOptionFlagMasks(IsCLMode()); 2321 std::string Nearest; 2322 if (getOpts().findNearest(Value, Nearest, IncludedFlagsBitmask, 2323 ExcludedFlagsBitmask) <= 1) { 2324 Diag(clang::diag::err_drv_no_such_file_with_suggestion) 2325 << Value << Nearest; 2326 return false; 2327 } 2328 } 2329 2330 // In CL mode, don't error on apparently non-existent linker inputs, because 2331 // they can be influenced by linker flags the clang driver might not 2332 // understand. 2333 // Examples: 2334 // - `clang-cl main.cc ole32.lib` in a a non-MSVC shell will make the driver 2335 // module look for an MSVC installation in the registry. (We could ask 2336 // the MSVCToolChain object if it can find `ole32.lib`, but the logic to 2337 // look in the registry might move into lld-link in the future so that 2338 // lld-link invocations in non-MSVC shells just work too.) 2339 // - `clang-cl ... /link ...` can pass arbitrary flags to the linker, 2340 // including /libpath:, which is used to find .lib and .obj files. 2341 // So do not diagnose this on the driver level. Rely on the linker diagnosing 2342 // it. (If we don't end up invoking the linker, this means we'll emit a 2343 // "'linker' input unused [-Wunused-command-line-argument]" warning instead 2344 // of an error.) 2345 // 2346 // Only do this skip after the typo correction step above. `/Brepo` is treated 2347 // as TY_Object, but it's clearly a typo for `/Brepro`. It seems fine to emit 2348 // an error if we have a flag that's within an edit distance of 1 from a 2349 // flag. (Users can use `-Wl,` or `/linker` to launder the flag past the 2350 // driver in the unlikely case they run into this.) 2351 // 2352 // Don't do this for inputs that start with a '/', else we'd pass options 2353 // like /libpath: through to the linker silently. 2354 // 2355 // Emitting an error for linker inputs can also cause incorrect diagnostics 2356 // with the gcc driver. The command 2357 // clang -fuse-ld=lld -Wl,--chroot,some/dir /file.o 2358 // will make lld look for some/dir/file.o, while we will diagnose here that 2359 // `/file.o` does not exist. However, configure scripts check if 2360 // `clang /GR-` compiles without error to see if the compiler is cl.exe, 2361 // so we can't downgrade diagnostics for `/GR-` from an error to a warning 2362 // in cc mode. (We can in cl mode because cl.exe itself only warns on 2363 // unknown flags.) 2364 if (IsCLMode() && Ty == types::TY_Object && !Value.startswith("/")) 2365 return true; 2366 2367 Diag(clang::diag::err_drv_no_such_file) << Value; 2368 return false; 2369 } 2370 2371 // Get the C++20 Header Unit type corresponding to the input type. 2372 static types::ID CXXHeaderUnitType(ModuleHeaderMode HM) { 2373 switch (HM) { 2374 case HeaderMode_User: 2375 return types::TY_CXXUHeader; 2376 case HeaderMode_System: 2377 return types::TY_CXXSHeader; 2378 case HeaderMode_Default: 2379 break; 2380 case HeaderMode_None: 2381 llvm_unreachable("should not be called in this case"); 2382 } 2383 return types::TY_CXXHUHeader; 2384 } 2385 2386 // Construct a the list of inputs and their types. 2387 void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args, 2388 InputList &Inputs) const { 2389 const llvm::opt::OptTable &Opts = getOpts(); 2390 // Track the current user specified (-x) input. We also explicitly track the 2391 // argument used to set the type; we only want to claim the type when we 2392 // actually use it, so we warn about unused -x arguments. 2393 types::ID InputType = types::TY_Nothing; 2394 Arg *InputTypeArg = nullptr; 2395 2396 // The last /TC or /TP option sets the input type to C or C++ globally. 2397 if (Arg *TCTP = Args.getLastArgNoClaim(options::OPT__SLASH_TC, 2398 options::OPT__SLASH_TP)) { 2399 InputTypeArg = TCTP; 2400 InputType = TCTP->getOption().matches(options::OPT__SLASH_TC) 2401 ? types::TY_C 2402 : types::TY_CXX; 2403 2404 Arg *Previous = nullptr; 2405 bool ShowNote = false; 2406 for (Arg *A : 2407 Args.filtered(options::OPT__SLASH_TC, options::OPT__SLASH_TP)) { 2408 if (Previous) { 2409 Diag(clang::diag::warn_drv_overriding_flag_option) 2410 << Previous->getSpelling() << A->getSpelling(); 2411 ShowNote = true; 2412 } 2413 Previous = A; 2414 } 2415 if (ShowNote) 2416 Diag(clang::diag::note_drv_t_option_is_global); 2417 2418 // No driver mode exposes -x and /TC or /TP; we don't support mixing them. 2419 assert(!Args.hasArg(options::OPT_x) && "-x and /TC or /TP is not allowed"); 2420 } 2421 2422 // Warn -x after last input file has no effect 2423 { 2424 Arg *LastXArg = Args.getLastArgNoClaim(options::OPT_x); 2425 Arg *LastInputArg = Args.getLastArgNoClaim(options::OPT_INPUT); 2426 if (LastXArg && LastInputArg && LastInputArg->getIndex() < LastXArg->getIndex()) 2427 Diag(clang::diag::warn_drv_unused_x) << LastXArg->getValue(); 2428 } 2429 2430 for (Arg *A : Args) { 2431 if (A->getOption().getKind() == Option::InputClass) { 2432 const char *Value = A->getValue(); 2433 types::ID Ty = types::TY_INVALID; 2434 2435 // Infer the input type if necessary. 2436 if (InputType == types::TY_Nothing) { 2437 // If there was an explicit arg for this, claim it. 2438 if (InputTypeArg) 2439 InputTypeArg->claim(); 2440 2441 // stdin must be handled specially. 2442 if (memcmp(Value, "-", 2) == 0) { 2443 if (IsFlangMode()) { 2444 Ty = types::TY_Fortran; 2445 } else { 2446 // If running with -E, treat as a C input (this changes the 2447 // builtin macros, for example). This may be overridden by -ObjC 2448 // below. 2449 // 2450 // Otherwise emit an error but still use a valid type to avoid 2451 // spurious errors (e.g., no inputs). 2452 assert(!CCGenDiagnostics && "stdin produces no crash reproducer"); 2453 if (!Args.hasArgNoClaim(options::OPT_E) && !CCCIsCPP()) 2454 Diag(IsCLMode() ? clang::diag::err_drv_unknown_stdin_type_clang_cl 2455 : clang::diag::err_drv_unknown_stdin_type); 2456 Ty = types::TY_C; 2457 } 2458 } else { 2459 // Otherwise lookup by extension. 2460 // Fallback is C if invoked as C preprocessor, C++ if invoked with 2461 // clang-cl /E, or Object otherwise. 2462 // We use a host hook here because Darwin at least has its own 2463 // idea of what .s is. 2464 if (const char *Ext = strrchr(Value, '.')) 2465 Ty = TC.LookupTypeForExtension(Ext + 1); 2466 2467 if (Ty == types::TY_INVALID) { 2468 if (IsCLMode() && (Args.hasArgNoClaim(options::OPT_E) || CCGenDiagnostics)) 2469 Ty = types::TY_CXX; 2470 else if (CCCIsCPP() || CCGenDiagnostics) 2471 Ty = types::TY_C; 2472 else 2473 Ty = types::TY_Object; 2474 } 2475 2476 // If the driver is invoked as C++ compiler (like clang++ or c++) it 2477 // should autodetect some input files as C++ for g++ compatibility. 2478 if (CCCIsCXX()) { 2479 types::ID OldTy = Ty; 2480 Ty = types::lookupCXXTypeForCType(Ty); 2481 2482 // Do not complain about foo.h, when we are known to be processing 2483 // it as a C++20 header unit. 2484 if (Ty != OldTy && !(OldTy == types::TY_CHeader && hasHeaderMode())) 2485 Diag(clang::diag::warn_drv_treating_input_as_cxx) 2486 << getTypeName(OldTy) << getTypeName(Ty); 2487 } 2488 2489 // If running with -fthinlto-index=, extensions that normally identify 2490 // native object files actually identify LLVM bitcode files. 2491 if (Args.hasArgNoClaim(options::OPT_fthinlto_index_EQ) && 2492 Ty == types::TY_Object) 2493 Ty = types::TY_LLVM_BC; 2494 } 2495 2496 // -ObjC and -ObjC++ override the default language, but only for "source 2497 // files". We just treat everything that isn't a linker input as a 2498 // source file. 2499 // 2500 // FIXME: Clean this up if we move the phase sequence into the type. 2501 if (Ty != types::TY_Object) { 2502 if (Args.hasArg(options::OPT_ObjC)) 2503 Ty = types::TY_ObjC; 2504 else if (Args.hasArg(options::OPT_ObjCXX)) 2505 Ty = types::TY_ObjCXX; 2506 } 2507 2508 // Disambiguate headers that are meant to be header units from those 2509 // intended to be PCH. Avoid missing '.h' cases that are counted as 2510 // C headers by default - we know we are in C++ mode and we do not 2511 // want to issue a complaint about compiling things in the wrong mode. 2512 if ((Ty == types::TY_CXXHeader || Ty == types::TY_CHeader) && 2513 hasHeaderMode()) 2514 Ty = CXXHeaderUnitType(CXX20HeaderType); 2515 } else { 2516 assert(InputTypeArg && "InputType set w/o InputTypeArg"); 2517 if (!InputTypeArg->getOption().matches(options::OPT_x)) { 2518 // If emulating cl.exe, make sure that /TC and /TP don't affect input 2519 // object files. 2520 const char *Ext = strrchr(Value, '.'); 2521 if (Ext && TC.LookupTypeForExtension(Ext + 1) == types::TY_Object) 2522 Ty = types::TY_Object; 2523 } 2524 if (Ty == types::TY_INVALID) { 2525 Ty = InputType; 2526 InputTypeArg->claim(); 2527 } 2528 } 2529 2530 if (DiagnoseInputExistence(Args, Value, Ty, /*TypoCorrect=*/true)) 2531 Inputs.push_back(std::make_pair(Ty, A)); 2532 2533 } else if (A->getOption().matches(options::OPT__SLASH_Tc)) { 2534 StringRef Value = A->getValue(); 2535 if (DiagnoseInputExistence(Args, Value, types::TY_C, 2536 /*TypoCorrect=*/false)) { 2537 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2538 Inputs.push_back(std::make_pair(types::TY_C, InputArg)); 2539 } 2540 A->claim(); 2541 } else if (A->getOption().matches(options::OPT__SLASH_Tp)) { 2542 StringRef Value = A->getValue(); 2543 if (DiagnoseInputExistence(Args, Value, types::TY_CXX, 2544 /*TypoCorrect=*/false)) { 2545 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2546 Inputs.push_back(std::make_pair(types::TY_CXX, InputArg)); 2547 } 2548 A->claim(); 2549 } else if (A->getOption().hasFlag(options::LinkerInput)) { 2550 // Just treat as object type, we could make a special type for this if 2551 // necessary. 2552 Inputs.push_back(std::make_pair(types::TY_Object, A)); 2553 2554 } else if (A->getOption().matches(options::OPT_x)) { 2555 InputTypeArg = A; 2556 InputType = types::lookupTypeForTypeSpecifier(A->getValue()); 2557 A->claim(); 2558 2559 // Follow gcc behavior and treat as linker input for invalid -x 2560 // options. Its not clear why we shouldn't just revert to unknown; but 2561 // this isn't very important, we might as well be bug compatible. 2562 if (!InputType) { 2563 Diag(clang::diag::err_drv_unknown_language) << A->getValue(); 2564 InputType = types::TY_Object; 2565 } 2566 2567 // If the user has put -fmodule-header{,=} then we treat C++ headers as 2568 // header unit inputs. So we 'promote' -xc++-header appropriately. 2569 if (InputType == types::TY_CXXHeader && hasHeaderMode()) 2570 InputType = CXXHeaderUnitType(CXX20HeaderType); 2571 } else if (A->getOption().getID() == options::OPT_U) { 2572 assert(A->getNumValues() == 1 && "The /U option has one value."); 2573 StringRef Val = A->getValue(0); 2574 if (Val.find_first_of("/\\") != StringRef::npos) { 2575 // Warn about e.g. "/Users/me/myfile.c". 2576 Diag(diag::warn_slash_u_filename) << Val; 2577 Diag(diag::note_use_dashdash); 2578 } 2579 } 2580 } 2581 if (CCCIsCPP() && Inputs.empty()) { 2582 // If called as standalone preprocessor, stdin is processed 2583 // if no other input is present. 2584 Arg *A = MakeInputArg(Args, Opts, "-"); 2585 Inputs.push_back(std::make_pair(types::TY_C, A)); 2586 } 2587 } 2588 2589 namespace { 2590 /// Provides a convenient interface for different programming models to generate 2591 /// the required device actions. 2592 class OffloadingActionBuilder final { 2593 /// Flag used to trace errors in the builder. 2594 bool IsValid = false; 2595 2596 /// The compilation that is using this builder. 2597 Compilation &C; 2598 2599 /// Map between an input argument and the offload kinds used to process it. 2600 std::map<const Arg *, unsigned> InputArgToOffloadKindMap; 2601 2602 /// Map between a host action and its originating input argument. 2603 std::map<Action *, const Arg *> HostActionToInputArgMap; 2604 2605 /// Builder interface. It doesn't build anything or keep any state. 2606 class DeviceActionBuilder { 2607 public: 2608 typedef const llvm::SmallVectorImpl<phases::ID> PhasesTy; 2609 2610 enum ActionBuilderReturnCode { 2611 // The builder acted successfully on the current action. 2612 ABRT_Success, 2613 // The builder didn't have to act on the current action. 2614 ABRT_Inactive, 2615 // The builder was successful and requested the host action to not be 2616 // generated. 2617 ABRT_Ignore_Host, 2618 }; 2619 2620 protected: 2621 /// Compilation associated with this builder. 2622 Compilation &C; 2623 2624 /// Tool chains associated with this builder. The same programming 2625 /// model may have associated one or more tool chains. 2626 SmallVector<const ToolChain *, 2> ToolChains; 2627 2628 /// The derived arguments associated with this builder. 2629 DerivedArgList &Args; 2630 2631 /// The inputs associated with this builder. 2632 const Driver::InputList &Inputs; 2633 2634 /// The associated offload kind. 2635 Action::OffloadKind AssociatedOffloadKind = Action::OFK_None; 2636 2637 public: 2638 DeviceActionBuilder(Compilation &C, DerivedArgList &Args, 2639 const Driver::InputList &Inputs, 2640 Action::OffloadKind AssociatedOffloadKind) 2641 : C(C), Args(Args), Inputs(Inputs), 2642 AssociatedOffloadKind(AssociatedOffloadKind) {} 2643 virtual ~DeviceActionBuilder() {} 2644 2645 /// Fill up the array \a DA with all the device dependences that should be 2646 /// added to the provided host action \a HostAction. By default it is 2647 /// inactive. 2648 virtual ActionBuilderReturnCode 2649 getDeviceDependences(OffloadAction::DeviceDependences &DA, 2650 phases::ID CurPhase, phases::ID FinalPhase, 2651 PhasesTy &Phases) { 2652 return ABRT_Inactive; 2653 } 2654 2655 /// Update the state to include the provided host action \a HostAction as a 2656 /// dependency of the current device action. By default it is inactive. 2657 virtual ActionBuilderReturnCode addDeviceDepences(Action *HostAction) { 2658 return ABRT_Inactive; 2659 } 2660 2661 /// Append top level actions generated by the builder. 2662 virtual void appendTopLevelActions(ActionList &AL) {} 2663 2664 /// Append linker device actions generated by the builder. 2665 virtual void appendLinkDeviceActions(ActionList &AL) {} 2666 2667 /// Append linker host action generated by the builder. 2668 virtual Action* appendLinkHostActions(ActionList &AL) { return nullptr; } 2669 2670 /// Append linker actions generated by the builder. 2671 virtual void appendLinkDependences(OffloadAction::DeviceDependences &DA) {} 2672 2673 /// Initialize the builder. Return true if any initialization errors are 2674 /// found. 2675 virtual bool initialize() { return false; } 2676 2677 /// Return true if the builder can use bundling/unbundling. 2678 virtual bool canUseBundlerUnbundler() const { return false; } 2679 2680 /// Return true if this builder is valid. We have a valid builder if we have 2681 /// associated device tool chains. 2682 bool isValid() { return !ToolChains.empty(); } 2683 2684 /// Return the associated offload kind. 2685 Action::OffloadKind getAssociatedOffloadKind() { 2686 return AssociatedOffloadKind; 2687 } 2688 }; 2689 2690 /// Base class for CUDA/HIP action builder. It injects device code in 2691 /// the host backend action. 2692 class CudaActionBuilderBase : public DeviceActionBuilder { 2693 protected: 2694 /// Flags to signal if the user requested host-only or device-only 2695 /// compilation. 2696 bool CompileHostOnly = false; 2697 bool CompileDeviceOnly = false; 2698 bool EmitLLVM = false; 2699 bool EmitAsm = false; 2700 2701 /// ID to identify each device compilation. For CUDA it is simply the 2702 /// GPU arch string. For HIP it is either the GPU arch string or GPU 2703 /// arch string plus feature strings delimited by a plus sign, e.g. 2704 /// gfx906+xnack. 2705 struct TargetID { 2706 /// Target ID string which is persistent throughout the compilation. 2707 const char *ID; 2708 TargetID(CudaArch Arch) { ID = CudaArchToString(Arch); } 2709 TargetID(const char *ID) : ID(ID) {} 2710 operator const char *() { return ID; } 2711 operator StringRef() { return StringRef(ID); } 2712 }; 2713 /// List of GPU architectures to use in this compilation. 2714 SmallVector<TargetID, 4> GpuArchList; 2715 2716 /// The CUDA actions for the current input. 2717 ActionList CudaDeviceActions; 2718 2719 /// The CUDA fat binary if it was generated for the current input. 2720 Action *CudaFatBinary = nullptr; 2721 2722 /// Flag that is set to true if this builder acted on the current input. 2723 bool IsActive = false; 2724 2725 /// Flag for -fgpu-rdc. 2726 bool Relocatable = false; 2727 2728 /// Default GPU architecture if there's no one specified. 2729 CudaArch DefaultCudaArch = CudaArch::UNKNOWN; 2730 2731 /// Method to generate compilation unit ID specified by option 2732 /// '-fuse-cuid='. 2733 enum UseCUIDKind { CUID_Hash, CUID_Random, CUID_None, CUID_Invalid }; 2734 UseCUIDKind UseCUID = CUID_Hash; 2735 2736 /// Compilation unit ID specified by option '-cuid='. 2737 StringRef FixedCUID; 2738 2739 public: 2740 CudaActionBuilderBase(Compilation &C, DerivedArgList &Args, 2741 const Driver::InputList &Inputs, 2742 Action::OffloadKind OFKind) 2743 : DeviceActionBuilder(C, Args, Inputs, OFKind) {} 2744 2745 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 2746 // While generating code for CUDA, we only depend on the host input action 2747 // to trigger the creation of all the CUDA device actions. 2748 2749 // If we are dealing with an input action, replicate it for each GPU 2750 // architecture. If we are in host-only mode we return 'success' so that 2751 // the host uses the CUDA offload kind. 2752 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 2753 assert(!GpuArchList.empty() && 2754 "We should have at least one GPU architecture."); 2755 2756 // If the host input is not CUDA or HIP, we don't need to bother about 2757 // this input. 2758 if (!(IA->getType() == types::TY_CUDA || 2759 IA->getType() == types::TY_HIP || 2760 IA->getType() == types::TY_PP_HIP)) { 2761 // The builder will ignore this input. 2762 IsActive = false; 2763 return ABRT_Inactive; 2764 } 2765 2766 // Set the flag to true, so that the builder acts on the current input. 2767 IsActive = true; 2768 2769 if (CompileHostOnly) 2770 return ABRT_Success; 2771 2772 // Replicate inputs for each GPU architecture. 2773 auto Ty = IA->getType() == types::TY_HIP ? types::TY_HIP_DEVICE 2774 : types::TY_CUDA_DEVICE; 2775 std::string CUID = FixedCUID.str(); 2776 if (CUID.empty()) { 2777 if (UseCUID == CUID_Random) 2778 CUID = llvm::utohexstr(llvm::sys::Process::GetRandomNumber(), 2779 /*LowerCase=*/true); 2780 else if (UseCUID == CUID_Hash) { 2781 llvm::MD5 Hasher; 2782 llvm::MD5::MD5Result Hash; 2783 SmallString<256> RealPath; 2784 llvm::sys::fs::real_path(IA->getInputArg().getValue(), RealPath, 2785 /*expand_tilde=*/true); 2786 Hasher.update(RealPath); 2787 for (auto *A : Args) { 2788 if (A->getOption().matches(options::OPT_INPUT)) 2789 continue; 2790 Hasher.update(A->getAsString(Args)); 2791 } 2792 Hasher.final(Hash); 2793 CUID = llvm::utohexstr(Hash.low(), /*LowerCase=*/true); 2794 } 2795 } 2796 IA->setId(CUID); 2797 2798 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 2799 CudaDeviceActions.push_back( 2800 C.MakeAction<InputAction>(IA->getInputArg(), Ty, IA->getId())); 2801 } 2802 2803 return ABRT_Success; 2804 } 2805 2806 // If this is an unbundling action use it as is for each CUDA toolchain. 2807 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 2808 2809 // If -fgpu-rdc is disabled, should not unbundle since there is no 2810 // device code to link. 2811 if (UA->getType() == types::TY_Object && !Relocatable) 2812 return ABRT_Inactive; 2813 2814 CudaDeviceActions.clear(); 2815 auto *IA = cast<InputAction>(UA->getInputs().back()); 2816 std::string FileName = IA->getInputArg().getAsString(Args); 2817 // Check if the type of the file is the same as the action. Do not 2818 // unbundle it if it is not. Do not unbundle .so files, for example, 2819 // which are not object files. 2820 if (IA->getType() == types::TY_Object && 2821 (!llvm::sys::path::has_extension(FileName) || 2822 types::lookupTypeForExtension( 2823 llvm::sys::path::extension(FileName).drop_front()) != 2824 types::TY_Object)) 2825 return ABRT_Inactive; 2826 2827 for (auto Arch : GpuArchList) { 2828 CudaDeviceActions.push_back(UA); 2829 UA->registerDependentActionInfo(ToolChains[0], Arch, 2830 AssociatedOffloadKind); 2831 } 2832 IsActive = true; 2833 return ABRT_Success; 2834 } 2835 2836 return IsActive ? ABRT_Success : ABRT_Inactive; 2837 } 2838 2839 void appendTopLevelActions(ActionList &AL) override { 2840 // Utility to append actions to the top level list. 2841 auto AddTopLevel = [&](Action *A, TargetID TargetID) { 2842 OffloadAction::DeviceDependences Dep; 2843 Dep.add(*A, *ToolChains.front(), TargetID, AssociatedOffloadKind); 2844 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 2845 }; 2846 2847 // If we have a fat binary, add it to the list. 2848 if (CudaFatBinary) { 2849 AddTopLevel(CudaFatBinary, CudaArch::UNUSED); 2850 CudaDeviceActions.clear(); 2851 CudaFatBinary = nullptr; 2852 return; 2853 } 2854 2855 if (CudaDeviceActions.empty()) 2856 return; 2857 2858 // If we have CUDA actions at this point, that's because we have a have 2859 // partial compilation, so we should have an action for each GPU 2860 // architecture. 2861 assert(CudaDeviceActions.size() == GpuArchList.size() && 2862 "Expecting one action per GPU architecture."); 2863 assert(ToolChains.size() == 1 && 2864 "Expecting to have a single CUDA toolchain."); 2865 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) 2866 AddTopLevel(CudaDeviceActions[I], GpuArchList[I]); 2867 2868 CudaDeviceActions.clear(); 2869 } 2870 2871 /// Get canonicalized offload arch option. \returns empty StringRef if the 2872 /// option is invalid. 2873 virtual StringRef getCanonicalOffloadArch(StringRef Arch) = 0; 2874 2875 virtual llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 2876 getConflictOffloadArchCombination(const std::set<StringRef> &GpuArchs) = 0; 2877 2878 bool initialize() override { 2879 assert(AssociatedOffloadKind == Action::OFK_Cuda || 2880 AssociatedOffloadKind == Action::OFK_HIP); 2881 2882 // We don't need to support CUDA. 2883 if (AssociatedOffloadKind == Action::OFK_Cuda && 2884 !C.hasOffloadToolChain<Action::OFK_Cuda>()) 2885 return false; 2886 2887 // We don't need to support HIP. 2888 if (AssociatedOffloadKind == Action::OFK_HIP && 2889 !C.hasOffloadToolChain<Action::OFK_HIP>()) 2890 return false; 2891 2892 Relocatable = Args.hasFlag(options::OPT_fgpu_rdc, 2893 options::OPT_fno_gpu_rdc, /*Default=*/false); 2894 2895 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 2896 assert(HostTC && "No toolchain for host compilation."); 2897 if (HostTC->getTriple().isNVPTX() || 2898 HostTC->getTriple().getArch() == llvm::Triple::amdgcn) { 2899 // We do not support targeting NVPTX/AMDGCN for host compilation. Throw 2900 // an error and abort pipeline construction early so we don't trip 2901 // asserts that assume device-side compilation. 2902 C.getDriver().Diag(diag::err_drv_cuda_host_arch) 2903 << HostTC->getTriple().getArchName(); 2904 return true; 2905 } 2906 2907 ToolChains.push_back( 2908 AssociatedOffloadKind == Action::OFK_Cuda 2909 ? C.getSingleOffloadToolChain<Action::OFK_Cuda>() 2910 : C.getSingleOffloadToolChain<Action::OFK_HIP>()); 2911 2912 Arg *PartialCompilationArg = Args.getLastArg( 2913 options::OPT_offload_host_only, options::OPT_offload_device_only, 2914 options::OPT_offload_host_device); 2915 CompileHostOnly = 2916 PartialCompilationArg && PartialCompilationArg->getOption().matches( 2917 options::OPT_offload_host_only); 2918 CompileDeviceOnly = 2919 PartialCompilationArg && PartialCompilationArg->getOption().matches( 2920 options::OPT_offload_device_only); 2921 EmitLLVM = Args.getLastArg(options::OPT_emit_llvm); 2922 EmitAsm = Args.getLastArg(options::OPT_S); 2923 FixedCUID = Args.getLastArgValue(options::OPT_cuid_EQ); 2924 if (Arg *A = Args.getLastArg(options::OPT_fuse_cuid_EQ)) { 2925 StringRef UseCUIDStr = A->getValue(); 2926 UseCUID = llvm::StringSwitch<UseCUIDKind>(UseCUIDStr) 2927 .Case("hash", CUID_Hash) 2928 .Case("random", CUID_Random) 2929 .Case("none", CUID_None) 2930 .Default(CUID_Invalid); 2931 if (UseCUID == CUID_Invalid) { 2932 C.getDriver().Diag(diag::err_drv_invalid_value) 2933 << A->getAsString(Args) << UseCUIDStr; 2934 C.setContainsError(); 2935 return true; 2936 } 2937 } 2938 2939 // --offload and --offload-arch options are mutually exclusive. 2940 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 2941 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 2942 options::OPT_no_offload_arch_EQ)) { 2943 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) << "--offload-arch" 2944 << "--offload"; 2945 } 2946 2947 // Collect all cuda_gpu_arch parameters, removing duplicates. 2948 std::set<StringRef> GpuArchs; 2949 bool Error = false; 2950 for (Arg *A : Args) { 2951 if (!(A->getOption().matches(options::OPT_offload_arch_EQ) || 2952 A->getOption().matches(options::OPT_no_offload_arch_EQ))) 2953 continue; 2954 A->claim(); 2955 2956 StringRef ArchStr = A->getValue(); 2957 if (A->getOption().matches(options::OPT_no_offload_arch_EQ) && 2958 ArchStr == "all") { 2959 GpuArchs.clear(); 2960 continue; 2961 } 2962 ArchStr = getCanonicalOffloadArch(ArchStr); 2963 if (ArchStr.empty()) { 2964 Error = true; 2965 } else if (A->getOption().matches(options::OPT_offload_arch_EQ)) 2966 GpuArchs.insert(ArchStr); 2967 else if (A->getOption().matches(options::OPT_no_offload_arch_EQ)) 2968 GpuArchs.erase(ArchStr); 2969 else 2970 llvm_unreachable("Unexpected option."); 2971 } 2972 2973 auto &&ConflictingArchs = getConflictOffloadArchCombination(GpuArchs); 2974 if (ConflictingArchs) { 2975 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 2976 << ConflictingArchs.getValue().first 2977 << ConflictingArchs.getValue().second; 2978 C.setContainsError(); 2979 return true; 2980 } 2981 2982 // Collect list of GPUs remaining in the set. 2983 for (auto Arch : GpuArchs) 2984 GpuArchList.push_back(Arch.data()); 2985 2986 // Default to sm_20 which is the lowest common denominator for 2987 // supported GPUs. sm_20 code should work correctly, if 2988 // suboptimally, on all newer GPUs. 2989 if (GpuArchList.empty()) { 2990 if (ToolChains.front()->getTriple().isSPIRV()) 2991 GpuArchList.push_back(CudaArch::Generic); 2992 else 2993 GpuArchList.push_back(DefaultCudaArch); 2994 } 2995 2996 return Error; 2997 } 2998 }; 2999 3000 /// \brief CUDA action builder. It injects device code in the host backend 3001 /// action. 3002 class CudaActionBuilder final : public CudaActionBuilderBase { 3003 public: 3004 CudaActionBuilder(Compilation &C, DerivedArgList &Args, 3005 const Driver::InputList &Inputs) 3006 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) { 3007 DefaultCudaArch = CudaArch::SM_35; 3008 } 3009 3010 StringRef getCanonicalOffloadArch(StringRef ArchStr) override { 3011 CudaArch Arch = StringToCudaArch(ArchStr); 3012 if (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch)) { 3013 C.getDriver().Diag(clang::diag::err_drv_cuda_bad_gpu_arch) << ArchStr; 3014 return StringRef(); 3015 } 3016 return CudaArchToString(Arch); 3017 } 3018 3019 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3020 getConflictOffloadArchCombination( 3021 const std::set<StringRef> &GpuArchs) override { 3022 return llvm::None; 3023 } 3024 3025 ActionBuilderReturnCode 3026 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3027 phases::ID CurPhase, phases::ID FinalPhase, 3028 PhasesTy &Phases) override { 3029 if (!IsActive) 3030 return ABRT_Inactive; 3031 3032 // If we don't have more CUDA actions, we don't have any dependences to 3033 // create for the host. 3034 if (CudaDeviceActions.empty()) 3035 return ABRT_Success; 3036 3037 assert(CudaDeviceActions.size() == GpuArchList.size() && 3038 "Expecting one action per GPU architecture."); 3039 assert(!CompileHostOnly && 3040 "Not expecting CUDA actions in host-only compilation."); 3041 3042 // If we are generating code for the device or we are in a backend phase, 3043 // we attempt to generate the fat binary. We compile each arch to ptx and 3044 // assemble to cubin, then feed the cubin *and* the ptx into a device 3045 // "link" action, which uses fatbinary to combine these cubins into one 3046 // fatbin. The fatbin is then an input to the host action if not in 3047 // device-only mode. 3048 if (CompileDeviceOnly || CurPhase == phases::Backend) { 3049 ActionList DeviceActions; 3050 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3051 // Produce the device action from the current phase up to the assemble 3052 // phase. 3053 for (auto Ph : Phases) { 3054 // Skip the phases that were already dealt with. 3055 if (Ph < CurPhase) 3056 continue; 3057 // We have to be consistent with the host final phase. 3058 if (Ph > FinalPhase) 3059 break; 3060 3061 CudaDeviceActions[I] = C.getDriver().ConstructPhaseAction( 3062 C, Args, Ph, CudaDeviceActions[I], Action::OFK_Cuda); 3063 3064 if (Ph == phases::Assemble) 3065 break; 3066 } 3067 3068 // If we didn't reach the assemble phase, we can't generate the fat 3069 // binary. We don't need to generate the fat binary if we are not in 3070 // device-only mode. 3071 if (!isa<AssembleJobAction>(CudaDeviceActions[I]) || 3072 CompileDeviceOnly) 3073 continue; 3074 3075 Action *AssembleAction = CudaDeviceActions[I]; 3076 assert(AssembleAction->getType() == types::TY_Object); 3077 assert(AssembleAction->getInputs().size() == 1); 3078 3079 Action *BackendAction = AssembleAction->getInputs()[0]; 3080 assert(BackendAction->getType() == types::TY_PP_Asm); 3081 3082 for (auto &A : {AssembleAction, BackendAction}) { 3083 OffloadAction::DeviceDependences DDep; 3084 DDep.add(*A, *ToolChains.front(), GpuArchList[I], Action::OFK_Cuda); 3085 DeviceActions.push_back( 3086 C.MakeAction<OffloadAction>(DDep, A->getType())); 3087 } 3088 } 3089 3090 // We generate the fat binary if we have device input actions. 3091 if (!DeviceActions.empty()) { 3092 CudaFatBinary = 3093 C.MakeAction<LinkJobAction>(DeviceActions, types::TY_CUDA_FATBIN); 3094 3095 if (!CompileDeviceOnly) { 3096 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3097 Action::OFK_Cuda); 3098 // Clear the fat binary, it is already a dependence to an host 3099 // action. 3100 CudaFatBinary = nullptr; 3101 } 3102 3103 // Remove the CUDA actions as they are already connected to an host 3104 // action or fat binary. 3105 CudaDeviceActions.clear(); 3106 } 3107 3108 // We avoid creating host action in device-only mode. 3109 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3110 } else if (CurPhase > phases::Backend) { 3111 // If we are past the backend phase and still have a device action, we 3112 // don't have to do anything as this action is already a device 3113 // top-level action. 3114 return ABRT_Success; 3115 } 3116 3117 assert(CurPhase < phases::Backend && "Generating single CUDA " 3118 "instructions should only occur " 3119 "before the backend phase!"); 3120 3121 // By default, we produce an action for each device arch. 3122 for (Action *&A : CudaDeviceActions) 3123 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3124 3125 return ABRT_Success; 3126 } 3127 }; 3128 /// \brief HIP action builder. It injects device code in the host backend 3129 /// action. 3130 class HIPActionBuilder final : public CudaActionBuilderBase { 3131 /// The linker inputs obtained for each device arch. 3132 SmallVector<ActionList, 8> DeviceLinkerInputs; 3133 // The default bundling behavior depends on the type of output, therefore 3134 // BundleOutput needs to be tri-value: None, true, or false. 3135 // Bundle code objects except --no-gpu-output is specified for device 3136 // only compilation. Bundle other type of output files only if 3137 // --gpu-bundle-output is specified for device only compilation. 3138 Optional<bool> BundleOutput; 3139 3140 public: 3141 HIPActionBuilder(Compilation &C, DerivedArgList &Args, 3142 const Driver::InputList &Inputs) 3143 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) { 3144 DefaultCudaArch = CudaArch::GFX803; 3145 if (Args.hasArg(options::OPT_gpu_bundle_output, 3146 options::OPT_no_gpu_bundle_output)) 3147 BundleOutput = Args.hasFlag(options::OPT_gpu_bundle_output, 3148 options::OPT_no_gpu_bundle_output, true); 3149 } 3150 3151 bool canUseBundlerUnbundler() const override { return true; } 3152 3153 StringRef getCanonicalOffloadArch(StringRef IdStr) override { 3154 llvm::StringMap<bool> Features; 3155 // getHIPOffloadTargetTriple() is known to return valid value as it has 3156 // been called successfully in the CreateOffloadingDeviceToolChains(). 3157 auto ArchStr = parseTargetID( 3158 *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()), IdStr, 3159 &Features); 3160 if (!ArchStr) { 3161 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << IdStr; 3162 C.setContainsError(); 3163 return StringRef(); 3164 } 3165 auto CanId = getCanonicalTargetID(ArchStr.getValue(), Features); 3166 return Args.MakeArgStringRef(CanId); 3167 }; 3168 3169 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3170 getConflictOffloadArchCombination( 3171 const std::set<StringRef> &GpuArchs) override { 3172 return getConflictTargetIDCombination(GpuArchs); 3173 } 3174 3175 ActionBuilderReturnCode 3176 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3177 phases::ID CurPhase, phases::ID FinalPhase, 3178 PhasesTy &Phases) override { 3179 if (!IsActive) 3180 return ABRT_Inactive; 3181 3182 // amdgcn does not support linking of object files, therefore we skip 3183 // backend and assemble phases to output LLVM IR. Except for generating 3184 // non-relocatable device code, where we generate fat binary for device 3185 // code and pass to host in Backend phase. 3186 if (CudaDeviceActions.empty()) 3187 return ABRT_Success; 3188 3189 assert(((CurPhase == phases::Link && Relocatable) || 3190 CudaDeviceActions.size() == GpuArchList.size()) && 3191 "Expecting one action per GPU architecture."); 3192 assert(!CompileHostOnly && 3193 "Not expecting HIP actions in host-only compilation."); 3194 3195 if (!Relocatable && CurPhase == phases::Backend && !EmitLLVM && 3196 !EmitAsm) { 3197 // If we are in backend phase, we attempt to generate the fat binary. 3198 // We compile each arch to IR and use a link action to generate code 3199 // object containing ISA. Then we use a special "link" action to create 3200 // a fat binary containing all the code objects for different GPU's. 3201 // The fat binary is then an input to the host action. 3202 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3203 if (C.getDriver().isUsingLTO(/*IsOffload=*/true)) { 3204 // When LTO is enabled, skip the backend and assemble phases and 3205 // use lld to link the bitcode. 3206 ActionList AL; 3207 AL.push_back(CudaDeviceActions[I]); 3208 // Create a link action to link device IR with device library 3209 // and generate ISA. 3210 CudaDeviceActions[I] = 3211 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3212 } else { 3213 // When LTO is not enabled, we follow the conventional 3214 // compiler phases, including backend and assemble phases. 3215 ActionList AL; 3216 Action *BackendAction = nullptr; 3217 if (ToolChains.front()->getTriple().isSPIRV()) { 3218 // Emit LLVM bitcode for SPIR-V targets. SPIR-V device tool chain 3219 // (HIPSPVToolChain) runs post-link LLVM IR passes. 3220 types::ID Output = Args.hasArg(options::OPT_S) 3221 ? types::TY_LLVM_IR 3222 : types::TY_LLVM_BC; 3223 BackendAction = 3224 C.MakeAction<BackendJobAction>(CudaDeviceActions[I], Output); 3225 } else 3226 BackendAction = C.getDriver().ConstructPhaseAction( 3227 C, Args, phases::Backend, CudaDeviceActions[I], 3228 AssociatedOffloadKind); 3229 auto AssembleAction = C.getDriver().ConstructPhaseAction( 3230 C, Args, phases::Assemble, BackendAction, 3231 AssociatedOffloadKind); 3232 AL.push_back(AssembleAction); 3233 // Create a link action to link device IR with device library 3234 // and generate ISA. 3235 CudaDeviceActions[I] = 3236 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3237 } 3238 3239 // OffloadingActionBuilder propagates device arch until an offload 3240 // action. Since the next action for creating fatbin does 3241 // not have device arch, whereas the above link action and its input 3242 // have device arch, an offload action is needed to stop the null 3243 // device arch of the next action being propagated to the above link 3244 // action. 3245 OffloadAction::DeviceDependences DDep; 3246 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3247 AssociatedOffloadKind); 3248 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3249 DDep, CudaDeviceActions[I]->getType()); 3250 } 3251 3252 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3253 BundleOutput.getValue()) { 3254 // Create HIP fat binary with a special "link" action. 3255 CudaFatBinary = C.MakeAction<LinkJobAction>(CudaDeviceActions, 3256 types::TY_HIP_FATBIN); 3257 3258 if (!CompileDeviceOnly) { 3259 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3260 AssociatedOffloadKind); 3261 // Clear the fat binary, it is already a dependence to an host 3262 // action. 3263 CudaFatBinary = nullptr; 3264 } 3265 3266 // Remove the CUDA actions as they are already connected to an host 3267 // action or fat binary. 3268 CudaDeviceActions.clear(); 3269 } 3270 3271 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3272 } else if (CurPhase == phases::Link) { 3273 // Save CudaDeviceActions to DeviceLinkerInputs for each GPU subarch. 3274 // This happens to each device action originated from each input file. 3275 // Later on, device actions in DeviceLinkerInputs are used to create 3276 // device link actions in appendLinkDependences and the created device 3277 // link actions are passed to the offload action as device dependence. 3278 DeviceLinkerInputs.resize(CudaDeviceActions.size()); 3279 auto LI = DeviceLinkerInputs.begin(); 3280 for (auto *A : CudaDeviceActions) { 3281 LI->push_back(A); 3282 ++LI; 3283 } 3284 3285 // We will pass the device action as a host dependence, so we don't 3286 // need to do anything else with them. 3287 CudaDeviceActions.clear(); 3288 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3289 } 3290 3291 // By default, we produce an action for each device arch. 3292 for (Action *&A : CudaDeviceActions) 3293 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A, 3294 AssociatedOffloadKind); 3295 3296 if (CompileDeviceOnly && CurPhase == FinalPhase && 3297 BundleOutput.hasValue() && BundleOutput.getValue()) { 3298 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3299 OffloadAction::DeviceDependences DDep; 3300 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3301 AssociatedOffloadKind); 3302 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3303 DDep, CudaDeviceActions[I]->getType()); 3304 } 3305 CudaFatBinary = 3306 C.MakeAction<OffloadBundlingJobAction>(CudaDeviceActions); 3307 CudaDeviceActions.clear(); 3308 } 3309 3310 return (CompileDeviceOnly && CurPhase == FinalPhase) ? ABRT_Ignore_Host 3311 : ABRT_Success; 3312 } 3313 3314 void appendLinkDeviceActions(ActionList &AL) override { 3315 if (DeviceLinkerInputs.size() == 0) 3316 return; 3317 3318 assert(DeviceLinkerInputs.size() == GpuArchList.size() && 3319 "Linker inputs and GPU arch list sizes do not match."); 3320 3321 ActionList Actions; 3322 unsigned I = 0; 3323 // Append a new link action for each device. 3324 // Each entry in DeviceLinkerInputs corresponds to a GPU arch. 3325 for (auto &LI : DeviceLinkerInputs) { 3326 3327 types::ID Output = Args.hasArg(options::OPT_emit_llvm) 3328 ? types::TY_LLVM_BC 3329 : types::TY_Image; 3330 3331 auto *DeviceLinkAction = C.MakeAction<LinkJobAction>(LI, Output); 3332 // Linking all inputs for the current GPU arch. 3333 // LI contains all the inputs for the linker. 3334 OffloadAction::DeviceDependences DeviceLinkDeps; 3335 DeviceLinkDeps.add(*DeviceLinkAction, *ToolChains[0], 3336 GpuArchList[I], AssociatedOffloadKind); 3337 Actions.push_back(C.MakeAction<OffloadAction>( 3338 DeviceLinkDeps, DeviceLinkAction->getType())); 3339 ++I; 3340 } 3341 DeviceLinkerInputs.clear(); 3342 3343 // If emitting LLVM, do not generate final host/device compilation action 3344 if (Args.hasArg(options::OPT_emit_llvm)) { 3345 AL.append(Actions); 3346 return; 3347 } 3348 3349 // Create a host object from all the device images by embedding them 3350 // in a fat binary for mixed host-device compilation. For device-only 3351 // compilation, creates a fat binary. 3352 OffloadAction::DeviceDependences DDeps; 3353 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3354 BundleOutput.getValue()) { 3355 auto *TopDeviceLinkAction = C.MakeAction<LinkJobAction>( 3356 Actions, 3357 CompileDeviceOnly ? types::TY_HIP_FATBIN : types::TY_Object); 3358 DDeps.add(*TopDeviceLinkAction, *ToolChains[0], nullptr, 3359 AssociatedOffloadKind); 3360 // Offload the host object to the host linker. 3361 AL.push_back( 3362 C.MakeAction<OffloadAction>(DDeps, TopDeviceLinkAction->getType())); 3363 } else { 3364 AL.append(Actions); 3365 } 3366 } 3367 3368 Action* appendLinkHostActions(ActionList &AL) override { return AL.back(); } 3369 3370 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3371 }; 3372 3373 /// OpenMP action builder. The host bitcode is passed to the device frontend 3374 /// and all the device linked images are passed to the host link phase. 3375 class OpenMPActionBuilder final : public DeviceActionBuilder { 3376 /// The OpenMP actions for the current input. 3377 ActionList OpenMPDeviceActions; 3378 3379 /// The linker inputs obtained for each toolchain. 3380 SmallVector<ActionList, 8> DeviceLinkerInputs; 3381 3382 public: 3383 OpenMPActionBuilder(Compilation &C, DerivedArgList &Args, 3384 const Driver::InputList &Inputs) 3385 : DeviceActionBuilder(C, Args, Inputs, Action::OFK_OpenMP) {} 3386 3387 ActionBuilderReturnCode 3388 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3389 phases::ID CurPhase, phases::ID FinalPhase, 3390 PhasesTy &Phases) override { 3391 if (OpenMPDeviceActions.empty()) 3392 return ABRT_Inactive; 3393 3394 // We should always have an action for each input. 3395 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3396 "Number of OpenMP actions and toolchains do not match."); 3397 3398 // The host only depends on device action in the linking phase, when all 3399 // the device images have to be embedded in the host image. 3400 if (CurPhase == phases::Link) { 3401 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3402 "Toolchains and linker inputs sizes do not match."); 3403 auto LI = DeviceLinkerInputs.begin(); 3404 for (auto *A : OpenMPDeviceActions) { 3405 LI->push_back(A); 3406 ++LI; 3407 } 3408 3409 // We passed the device action as a host dependence, so we don't need to 3410 // do anything else with them. 3411 OpenMPDeviceActions.clear(); 3412 return ABRT_Success; 3413 } 3414 3415 // By default, we produce an action for each device arch. 3416 for (Action *&A : OpenMPDeviceActions) 3417 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3418 3419 return ABRT_Success; 3420 } 3421 3422 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 3423 3424 // If this is an input action replicate it for each OpenMP toolchain. 3425 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 3426 OpenMPDeviceActions.clear(); 3427 for (unsigned I = 0; I < ToolChains.size(); ++I) 3428 OpenMPDeviceActions.push_back( 3429 C.MakeAction<InputAction>(IA->getInputArg(), IA->getType())); 3430 return ABRT_Success; 3431 } 3432 3433 // If this is an unbundling action use it as is for each OpenMP toolchain. 3434 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 3435 OpenMPDeviceActions.clear(); 3436 auto *IA = cast<InputAction>(UA->getInputs().back()); 3437 std::string FileName = IA->getInputArg().getAsString(Args); 3438 // Check if the type of the file is the same as the action. Do not 3439 // unbundle it if it is not. Do not unbundle .so files, for example, 3440 // which are not object files. 3441 if (IA->getType() == types::TY_Object && 3442 (!llvm::sys::path::has_extension(FileName) || 3443 types::lookupTypeForExtension( 3444 llvm::sys::path::extension(FileName).drop_front()) != 3445 types::TY_Object)) 3446 return ABRT_Inactive; 3447 for (unsigned I = 0; I < ToolChains.size(); ++I) { 3448 OpenMPDeviceActions.push_back(UA); 3449 UA->registerDependentActionInfo( 3450 ToolChains[I], /*BoundArch=*/StringRef(), Action::OFK_OpenMP); 3451 } 3452 return ABRT_Success; 3453 } 3454 3455 // When generating code for OpenMP we use the host compile phase result as 3456 // a dependence to the device compile phase so that it can learn what 3457 // declarations should be emitted. However, this is not the only use for 3458 // the host action, so we prevent it from being collapsed. 3459 if (isa<CompileJobAction>(HostAction)) { 3460 HostAction->setCannotBeCollapsedWithNextDependentAction(); 3461 assert(ToolChains.size() == OpenMPDeviceActions.size() && 3462 "Toolchains and device action sizes do not match."); 3463 OffloadAction::HostDependence HDep( 3464 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3465 /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3466 auto TC = ToolChains.begin(); 3467 for (Action *&A : OpenMPDeviceActions) { 3468 assert(isa<CompileJobAction>(A)); 3469 OffloadAction::DeviceDependences DDep; 3470 DDep.add(*A, **TC, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3471 A = C.MakeAction<OffloadAction>(HDep, DDep); 3472 ++TC; 3473 } 3474 } 3475 return ABRT_Success; 3476 } 3477 3478 void appendTopLevelActions(ActionList &AL) override { 3479 if (OpenMPDeviceActions.empty()) 3480 return; 3481 3482 // We should always have an action for each input. 3483 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3484 "Number of OpenMP actions and toolchains do not match."); 3485 3486 // Append all device actions followed by the proper offload action. 3487 auto TI = ToolChains.begin(); 3488 for (auto *A : OpenMPDeviceActions) { 3489 OffloadAction::DeviceDependences Dep; 3490 Dep.add(*A, **TI, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3491 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 3492 ++TI; 3493 } 3494 // We no longer need the action stored in this builder. 3495 OpenMPDeviceActions.clear(); 3496 } 3497 3498 void appendLinkDeviceActions(ActionList &AL) override { 3499 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3500 "Toolchains and linker inputs sizes do not match."); 3501 3502 // Append a new link action for each device. 3503 auto TC = ToolChains.begin(); 3504 for (auto &LI : DeviceLinkerInputs) { 3505 auto *DeviceLinkAction = 3506 C.MakeAction<LinkJobAction>(LI, types::TY_Image); 3507 OffloadAction::DeviceDependences DeviceLinkDeps; 3508 DeviceLinkDeps.add(*DeviceLinkAction, **TC, /*BoundArch=*/nullptr, 3509 Action::OFK_OpenMP); 3510 AL.push_back(C.MakeAction<OffloadAction>(DeviceLinkDeps, 3511 DeviceLinkAction->getType())); 3512 ++TC; 3513 } 3514 DeviceLinkerInputs.clear(); 3515 } 3516 3517 Action* appendLinkHostActions(ActionList &AL) override { 3518 // Create wrapper bitcode from the result of device link actions and compile 3519 // it to an object which will be added to the host link command. 3520 auto *BC = C.MakeAction<OffloadWrapperJobAction>(AL, types::TY_LLVM_BC); 3521 auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm); 3522 return C.MakeAction<AssembleJobAction>(ASM, types::TY_Object); 3523 } 3524 3525 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3526 3527 bool initialize() override { 3528 // Get the OpenMP toolchains. If we don't get any, the action builder will 3529 // know there is nothing to do related to OpenMP offloading. 3530 auto OpenMPTCRange = C.getOffloadToolChains<Action::OFK_OpenMP>(); 3531 for (auto TI = OpenMPTCRange.first, TE = OpenMPTCRange.second; TI != TE; 3532 ++TI) 3533 ToolChains.push_back(TI->second); 3534 3535 DeviceLinkerInputs.resize(ToolChains.size()); 3536 return false; 3537 } 3538 3539 bool canUseBundlerUnbundler() const override { 3540 // OpenMP should use bundled files whenever possible. 3541 return true; 3542 } 3543 }; 3544 3545 /// 3546 /// TODO: Add the implementation for other specialized builders here. 3547 /// 3548 3549 /// Specialized builders being used by this offloading action builder. 3550 SmallVector<DeviceActionBuilder *, 4> SpecializedBuilders; 3551 3552 /// Flag set to true if all valid builders allow file bundling/unbundling. 3553 bool CanUseBundler; 3554 3555 public: 3556 OffloadingActionBuilder(Compilation &C, DerivedArgList &Args, 3557 const Driver::InputList &Inputs) 3558 : C(C) { 3559 // Create a specialized builder for each device toolchain. 3560 3561 IsValid = true; 3562 3563 // Create a specialized builder for CUDA. 3564 SpecializedBuilders.push_back(new CudaActionBuilder(C, Args, Inputs)); 3565 3566 // Create a specialized builder for HIP. 3567 SpecializedBuilders.push_back(new HIPActionBuilder(C, Args, Inputs)); 3568 3569 // Create a specialized builder for OpenMP. 3570 SpecializedBuilders.push_back(new OpenMPActionBuilder(C, Args, Inputs)); 3571 3572 // 3573 // TODO: Build other specialized builders here. 3574 // 3575 3576 // Initialize all the builders, keeping track of errors. If all valid 3577 // builders agree that we can use bundling, set the flag to true. 3578 unsigned ValidBuilders = 0u; 3579 unsigned ValidBuildersSupportingBundling = 0u; 3580 for (auto *SB : SpecializedBuilders) { 3581 IsValid = IsValid && !SB->initialize(); 3582 3583 // Update the counters if the builder is valid. 3584 if (SB->isValid()) { 3585 ++ValidBuilders; 3586 if (SB->canUseBundlerUnbundler()) 3587 ++ValidBuildersSupportingBundling; 3588 } 3589 } 3590 CanUseBundler = 3591 ValidBuilders && ValidBuilders == ValidBuildersSupportingBundling; 3592 } 3593 3594 ~OffloadingActionBuilder() { 3595 for (auto *SB : SpecializedBuilders) 3596 delete SB; 3597 } 3598 3599 /// Record a host action and its originating input argument. 3600 void recordHostAction(Action *HostAction, const Arg *InputArg) { 3601 assert(HostAction && "Invalid host action"); 3602 assert(InputArg && "Invalid input argument"); 3603 auto Loc = HostActionToInputArgMap.find(HostAction); 3604 if (Loc == HostActionToInputArgMap.end()) 3605 HostActionToInputArgMap[HostAction] = InputArg; 3606 assert(HostActionToInputArgMap[HostAction] == InputArg && 3607 "host action mapped to multiple input arguments"); 3608 } 3609 3610 /// Generate an action that adds device dependences (if any) to a host action. 3611 /// If no device dependence actions exist, just return the host action \a 3612 /// HostAction. If an error is found or if no builder requires the host action 3613 /// to be generated, return nullptr. 3614 Action * 3615 addDeviceDependencesToHostAction(Action *HostAction, const Arg *InputArg, 3616 phases::ID CurPhase, phases::ID FinalPhase, 3617 DeviceActionBuilder::PhasesTy &Phases) { 3618 if (!IsValid) 3619 return nullptr; 3620 3621 if (SpecializedBuilders.empty()) 3622 return HostAction; 3623 3624 assert(HostAction && "Invalid host action!"); 3625 recordHostAction(HostAction, InputArg); 3626 3627 OffloadAction::DeviceDependences DDeps; 3628 // Check if all the programming models agree we should not emit the host 3629 // action. Also, keep track of the offloading kinds employed. 3630 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3631 unsigned InactiveBuilders = 0u; 3632 unsigned IgnoringBuilders = 0u; 3633 for (auto *SB : SpecializedBuilders) { 3634 if (!SB->isValid()) { 3635 ++InactiveBuilders; 3636 continue; 3637 } 3638 3639 auto RetCode = 3640 SB->getDeviceDependences(DDeps, CurPhase, FinalPhase, Phases); 3641 3642 // If the builder explicitly says the host action should be ignored, 3643 // we need to increment the variable that tracks the builders that request 3644 // the host object to be ignored. 3645 if (RetCode == DeviceActionBuilder::ABRT_Ignore_Host) 3646 ++IgnoringBuilders; 3647 3648 // Unless the builder was inactive for this action, we have to record the 3649 // offload kind because the host will have to use it. 3650 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3651 OffloadKind |= SB->getAssociatedOffloadKind(); 3652 } 3653 3654 // If all builders agree that the host object should be ignored, just return 3655 // nullptr. 3656 if (IgnoringBuilders && 3657 SpecializedBuilders.size() == (InactiveBuilders + IgnoringBuilders)) 3658 return nullptr; 3659 3660 if (DDeps.getActions().empty()) 3661 return HostAction; 3662 3663 // We have dependences we need to bundle together. We use an offload action 3664 // for that. 3665 OffloadAction::HostDependence HDep( 3666 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3667 /*BoundArch=*/nullptr, DDeps); 3668 return C.MakeAction<OffloadAction>(HDep, DDeps); 3669 } 3670 3671 /// Generate an action that adds a host dependence to a device action. The 3672 /// results will be kept in this action builder. Return true if an error was 3673 /// found. 3674 bool addHostDependenceToDeviceActions(Action *&HostAction, 3675 const Arg *InputArg) { 3676 if (!IsValid) 3677 return true; 3678 3679 recordHostAction(HostAction, InputArg); 3680 3681 // If we are supporting bundling/unbundling and the current action is an 3682 // input action of non-source file, we replace the host action by the 3683 // unbundling action. The bundler tool has the logic to detect if an input 3684 // is a bundle or not and if the input is not a bundle it assumes it is a 3685 // host file. Therefore it is safe to create an unbundling action even if 3686 // the input is not a bundle. 3687 if (CanUseBundler && isa<InputAction>(HostAction) && 3688 InputArg->getOption().getKind() == llvm::opt::Option::InputClass && 3689 (!types::isSrcFile(HostAction->getType()) || 3690 HostAction->getType() == types::TY_PP_HIP)) { 3691 auto UnbundlingHostAction = 3692 C.MakeAction<OffloadUnbundlingJobAction>(HostAction); 3693 UnbundlingHostAction->registerDependentActionInfo( 3694 C.getSingleOffloadToolChain<Action::OFK_Host>(), 3695 /*BoundArch=*/StringRef(), Action::OFK_Host); 3696 HostAction = UnbundlingHostAction; 3697 recordHostAction(HostAction, InputArg); 3698 } 3699 3700 assert(HostAction && "Invalid host action!"); 3701 3702 // Register the offload kinds that are used. 3703 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3704 for (auto *SB : SpecializedBuilders) { 3705 if (!SB->isValid()) 3706 continue; 3707 3708 auto RetCode = SB->addDeviceDepences(HostAction); 3709 3710 // Host dependences for device actions are not compatible with that same 3711 // action being ignored. 3712 assert(RetCode != DeviceActionBuilder::ABRT_Ignore_Host && 3713 "Host dependence not expected to be ignored.!"); 3714 3715 // Unless the builder was inactive for this action, we have to record the 3716 // offload kind because the host will have to use it. 3717 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3718 OffloadKind |= SB->getAssociatedOffloadKind(); 3719 } 3720 3721 // Do not use unbundler if the Host does not depend on device action. 3722 if (OffloadKind == Action::OFK_None && CanUseBundler) 3723 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) 3724 HostAction = UA->getInputs().back(); 3725 3726 return false; 3727 } 3728 3729 /// Add the offloading top level actions to the provided action list. This 3730 /// function can replace the host action by a bundling action if the 3731 /// programming models allow it. 3732 bool appendTopLevelActions(ActionList &AL, Action *HostAction, 3733 const Arg *InputArg) { 3734 if (HostAction) 3735 recordHostAction(HostAction, InputArg); 3736 3737 // Get the device actions to be appended. 3738 ActionList OffloadAL; 3739 for (auto *SB : SpecializedBuilders) { 3740 if (!SB->isValid()) 3741 continue; 3742 SB->appendTopLevelActions(OffloadAL); 3743 } 3744 3745 // If we can use the bundler, replace the host action by the bundling one in 3746 // the resulting list. Otherwise, just append the device actions. For 3747 // device only compilation, HostAction is a null pointer, therefore only do 3748 // this when HostAction is not a null pointer. 3749 if (CanUseBundler && HostAction && 3750 HostAction->getType() != types::TY_Nothing && !OffloadAL.empty()) { 3751 // Add the host action to the list in order to create the bundling action. 3752 OffloadAL.push_back(HostAction); 3753 3754 // We expect that the host action was just appended to the action list 3755 // before this method was called. 3756 assert(HostAction == AL.back() && "Host action not in the list??"); 3757 HostAction = C.MakeAction<OffloadBundlingJobAction>(OffloadAL); 3758 recordHostAction(HostAction, InputArg); 3759 AL.back() = HostAction; 3760 } else 3761 AL.append(OffloadAL.begin(), OffloadAL.end()); 3762 3763 // Propagate to the current host action (if any) the offload information 3764 // associated with the current input. 3765 if (HostAction) 3766 HostAction->propagateHostOffloadInfo(InputArgToOffloadKindMap[InputArg], 3767 /*BoundArch=*/nullptr); 3768 return false; 3769 } 3770 3771 void appendDeviceLinkActions(ActionList &AL) { 3772 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3773 if (!SB->isValid()) 3774 continue; 3775 SB->appendLinkDeviceActions(AL); 3776 } 3777 } 3778 3779 Action *makeHostLinkAction() { 3780 // Build a list of device linking actions. 3781 ActionList DeviceAL; 3782 appendDeviceLinkActions(DeviceAL); 3783 if (DeviceAL.empty()) 3784 return nullptr; 3785 3786 // Let builders add host linking actions. 3787 Action* HA = nullptr; 3788 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3789 if (!SB->isValid()) 3790 continue; 3791 HA = SB->appendLinkHostActions(DeviceAL); 3792 // This created host action has no originating input argument, therefore 3793 // needs to set its offloading kind directly. 3794 if (HA) 3795 HA->propagateHostOffloadInfo(SB->getAssociatedOffloadKind(), 3796 /*BoundArch=*/nullptr); 3797 } 3798 return HA; 3799 } 3800 3801 /// Processes the host linker action. This currently consists of replacing it 3802 /// with an offload action if there are device link objects and propagate to 3803 /// the host action all the offload kinds used in the current compilation. The 3804 /// resulting action is returned. 3805 Action *processHostLinkAction(Action *HostAction) { 3806 // Add all the dependences from the device linking actions. 3807 OffloadAction::DeviceDependences DDeps; 3808 for (auto *SB : SpecializedBuilders) { 3809 if (!SB->isValid()) 3810 continue; 3811 3812 SB->appendLinkDependences(DDeps); 3813 } 3814 3815 // Calculate all the offload kinds used in the current compilation. 3816 unsigned ActiveOffloadKinds = 0u; 3817 for (auto &I : InputArgToOffloadKindMap) 3818 ActiveOffloadKinds |= I.second; 3819 3820 // If we don't have device dependencies, we don't have to create an offload 3821 // action. 3822 if (DDeps.getActions().empty()) { 3823 // Set all the active offloading kinds to the link action. Given that it 3824 // is a link action it is assumed to depend on all actions generated so 3825 // far. 3826 HostAction->setHostOffloadInfo(ActiveOffloadKinds, 3827 /*BoundArch=*/nullptr); 3828 // Propagate active offloading kinds for each input to the link action. 3829 // Each input may have different active offloading kind. 3830 for (auto A : HostAction->inputs()) { 3831 auto ArgLoc = HostActionToInputArgMap.find(A); 3832 if (ArgLoc == HostActionToInputArgMap.end()) 3833 continue; 3834 auto OFKLoc = InputArgToOffloadKindMap.find(ArgLoc->second); 3835 if (OFKLoc == InputArgToOffloadKindMap.end()) 3836 continue; 3837 A->propagateHostOffloadInfo(OFKLoc->second, /*BoundArch=*/nullptr); 3838 } 3839 return HostAction; 3840 } 3841 3842 // Create the offload action with all dependences. When an offload action 3843 // is created the kinds are propagated to the host action, so we don't have 3844 // to do that explicitly here. 3845 OffloadAction::HostDependence HDep( 3846 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3847 /*BoundArch*/ nullptr, ActiveOffloadKinds); 3848 return C.MakeAction<OffloadAction>(HDep, DDeps); 3849 } 3850 }; 3851 } // anonymous namespace. 3852 3853 void Driver::handleArguments(Compilation &C, DerivedArgList &Args, 3854 const InputList &Inputs, 3855 ActionList &Actions) const { 3856 3857 // Ignore /Yc/Yu if both /Yc and /Yu passed but with different filenames. 3858 Arg *YcArg = Args.getLastArg(options::OPT__SLASH_Yc); 3859 Arg *YuArg = Args.getLastArg(options::OPT__SLASH_Yu); 3860 if (YcArg && YuArg && strcmp(YcArg->getValue(), YuArg->getValue()) != 0) { 3861 Diag(clang::diag::warn_drv_ycyu_different_arg_clang_cl); 3862 Args.eraseArg(options::OPT__SLASH_Yc); 3863 Args.eraseArg(options::OPT__SLASH_Yu); 3864 YcArg = YuArg = nullptr; 3865 } 3866 if (YcArg && Inputs.size() > 1) { 3867 Diag(clang::diag::warn_drv_yc_multiple_inputs_clang_cl); 3868 Args.eraseArg(options::OPT__SLASH_Yc); 3869 YcArg = nullptr; 3870 } 3871 3872 Arg *FinalPhaseArg; 3873 phases::ID FinalPhase = getFinalPhase(Args, &FinalPhaseArg); 3874 3875 if (FinalPhase == phases::Link) { 3876 // Emitting LLVM while linking disabled except in HIPAMD Toolchain 3877 if (Args.hasArg(options::OPT_emit_llvm) && !Args.hasArg(options::OPT_hip_link)) 3878 Diag(clang::diag::err_drv_emit_llvm_link); 3879 if (IsCLMode() && LTOMode != LTOK_None && 3880 !Args.getLastArgValue(options::OPT_fuse_ld_EQ) 3881 .equals_insensitive("lld")) 3882 Diag(clang::diag::err_drv_lto_without_lld); 3883 } 3884 3885 if (FinalPhase == phases::Preprocess || Args.hasArg(options::OPT__SLASH_Y_)) { 3886 // If only preprocessing or /Y- is used, all pch handling is disabled. 3887 // Rather than check for it everywhere, just remove clang-cl pch-related 3888 // flags here. 3889 Args.eraseArg(options::OPT__SLASH_Fp); 3890 Args.eraseArg(options::OPT__SLASH_Yc); 3891 Args.eraseArg(options::OPT__SLASH_Yu); 3892 YcArg = YuArg = nullptr; 3893 } 3894 3895 unsigned LastPLSize = 0; 3896 for (auto &I : Inputs) { 3897 types::ID InputType = I.first; 3898 const Arg *InputArg = I.second; 3899 3900 auto PL = types::getCompilationPhases(InputType); 3901 LastPLSize = PL.size(); 3902 3903 // If the first step comes after the final phase we are doing as part of 3904 // this compilation, warn the user about it. 3905 phases::ID InitialPhase = PL[0]; 3906 if (InitialPhase > FinalPhase) { 3907 if (InputArg->isClaimed()) 3908 continue; 3909 3910 // Claim here to avoid the more general unused warning. 3911 InputArg->claim(); 3912 3913 // Suppress all unused style warnings with -Qunused-arguments 3914 if (Args.hasArg(options::OPT_Qunused_arguments)) 3915 continue; 3916 3917 // Special case when final phase determined by binary name, rather than 3918 // by a command-line argument with a corresponding Arg. 3919 if (CCCIsCPP()) 3920 Diag(clang::diag::warn_drv_input_file_unused_by_cpp) 3921 << InputArg->getAsString(Args) << getPhaseName(InitialPhase); 3922 // Special case '-E' warning on a previously preprocessed file to make 3923 // more sense. 3924 else if (InitialPhase == phases::Compile && 3925 (Args.getLastArg(options::OPT__SLASH_EP, 3926 options::OPT__SLASH_P) || 3927 Args.getLastArg(options::OPT_E) || 3928 Args.getLastArg(options::OPT_M, options::OPT_MM)) && 3929 getPreprocessedType(InputType) == types::TY_INVALID) 3930 Diag(clang::diag::warn_drv_preprocessed_input_file_unused) 3931 << InputArg->getAsString(Args) << !!FinalPhaseArg 3932 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3933 else 3934 Diag(clang::diag::warn_drv_input_file_unused) 3935 << InputArg->getAsString(Args) << getPhaseName(InitialPhase) 3936 << !!FinalPhaseArg 3937 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3938 continue; 3939 } 3940 3941 if (YcArg) { 3942 // Add a separate precompile phase for the compile phase. 3943 if (FinalPhase >= phases::Compile) { 3944 const types::ID HeaderType = lookupHeaderTypeForSourceType(InputType); 3945 // Build the pipeline for the pch file. 3946 Action *ClangClPch = C.MakeAction<InputAction>(*InputArg, HeaderType); 3947 for (phases::ID Phase : types::getCompilationPhases(HeaderType)) 3948 ClangClPch = ConstructPhaseAction(C, Args, Phase, ClangClPch); 3949 assert(ClangClPch); 3950 Actions.push_back(ClangClPch); 3951 // The driver currently exits after the first failed command. This 3952 // relies on that behavior, to make sure if the pch generation fails, 3953 // the main compilation won't run. 3954 // FIXME: If the main compilation fails, the PCH generation should 3955 // probably not be considered successful either. 3956 } 3957 } 3958 } 3959 3960 // If we are linking, claim any options which are obviously only used for 3961 // compilation. 3962 // FIXME: Understand why the last Phase List length is used here. 3963 if (FinalPhase == phases::Link && LastPLSize == 1) { 3964 Args.ClaimAllArgs(options::OPT_CompileOnly_Group); 3965 Args.ClaimAllArgs(options::OPT_cl_compile_Group); 3966 } 3967 } 3968 3969 void Driver::BuildActions(Compilation &C, DerivedArgList &Args, 3970 const InputList &Inputs, ActionList &Actions) const { 3971 llvm::PrettyStackTraceString CrashInfo("Building compilation actions"); 3972 3973 if (!SuppressMissingInputWarning && Inputs.empty()) { 3974 Diag(clang::diag::err_drv_no_input_files); 3975 return; 3976 } 3977 3978 // Reject -Z* at the top level, these options should never have been exposed 3979 // by gcc. 3980 if (Arg *A = Args.getLastArg(options::OPT_Z_Joined)) 3981 Diag(clang::diag::err_drv_use_of_Z_option) << A->getAsString(Args); 3982 3983 // Diagnose misuse of /Fo. 3984 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fo)) { 3985 StringRef V = A->getValue(); 3986 if (Inputs.size() > 1 && !V.empty() && 3987 !llvm::sys::path::is_separator(V.back())) { 3988 // Check whether /Fo tries to name an output file for multiple inputs. 3989 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 3990 << A->getSpelling() << V; 3991 Args.eraseArg(options::OPT__SLASH_Fo); 3992 } 3993 } 3994 3995 // Diagnose misuse of /Fa. 3996 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fa)) { 3997 StringRef V = A->getValue(); 3998 if (Inputs.size() > 1 && !V.empty() && 3999 !llvm::sys::path::is_separator(V.back())) { 4000 // Check whether /Fa tries to name an asm file for multiple inputs. 4001 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4002 << A->getSpelling() << V; 4003 Args.eraseArg(options::OPT__SLASH_Fa); 4004 } 4005 } 4006 4007 // Diagnose misuse of /o. 4008 if (Arg *A = Args.getLastArg(options::OPT__SLASH_o)) { 4009 if (A->getValue()[0] == '\0') { 4010 // It has to have a value. 4011 Diag(clang::diag::err_drv_missing_argument) << A->getSpelling() << 1; 4012 Args.eraseArg(options::OPT__SLASH_o); 4013 } 4014 } 4015 4016 handleArguments(C, Args, Inputs, Actions); 4017 4018 // Builder to be used to build offloading actions. 4019 OffloadingActionBuilder OffloadBuilder(C, Args, Inputs); 4020 4021 bool UseNewOffloadingDriver = 4022 (C.isOffloadingHostKind(Action::OFK_OpenMP) && 4023 Args.hasFlag(options::OPT_fopenmp_new_driver, 4024 options::OPT_no_offload_new_driver, true)) || 4025 Args.hasFlag(options::OPT_offload_new_driver, 4026 options::OPT_no_offload_new_driver, false); 4027 4028 // Construct the actions to perform. 4029 HeaderModulePrecompileJobAction *HeaderModuleAction = nullptr; 4030 ExtractAPIJobAction *ExtractAPIAction = nullptr; 4031 ActionList LinkerInputs; 4032 ActionList MergerInputs; 4033 4034 for (auto &I : Inputs) { 4035 types::ID InputType = I.first; 4036 const Arg *InputArg = I.second; 4037 4038 auto PL = types::getCompilationPhases(*this, Args, InputType); 4039 if (PL.empty()) 4040 continue; 4041 4042 auto FullPL = types::getCompilationPhases(InputType); 4043 4044 // Build the pipeline for this file. 4045 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4046 4047 // Use the current host action in any of the offloading actions, if 4048 // required. 4049 if (!UseNewOffloadingDriver) 4050 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4051 break; 4052 4053 for (phases::ID Phase : PL) { 4054 4055 // Add any offload action the host action depends on. 4056 if (!UseNewOffloadingDriver) 4057 Current = OffloadBuilder.addDeviceDependencesToHostAction( 4058 Current, InputArg, Phase, PL.back(), FullPL); 4059 if (!Current) 4060 break; 4061 4062 // Queue linker inputs. 4063 if (Phase == phases::Link) { 4064 assert(Phase == PL.back() && "linking must be final compilation step."); 4065 // We don't need to generate additional link commands if emitting AMD bitcode 4066 if (!(C.getInputArgs().hasArg(options::OPT_hip_link) && 4067 (C.getInputArgs().hasArg(options::OPT_emit_llvm)))) 4068 LinkerInputs.push_back(Current); 4069 Current = nullptr; 4070 break; 4071 } 4072 4073 // TODO: Consider removing this because the merged may not end up being 4074 // the final Phase in the pipeline. Perhaps the merged could just merge 4075 // and then pass an artifact of some sort to the Link Phase. 4076 // Queue merger inputs. 4077 if (Phase == phases::IfsMerge) { 4078 assert(Phase == PL.back() && "merging must be final compilation step."); 4079 MergerInputs.push_back(Current); 4080 Current = nullptr; 4081 break; 4082 } 4083 4084 // Each precompiled header file after a module file action is a module 4085 // header of that same module file, rather than being compiled to a 4086 // separate PCH. 4087 if (Phase == phases::Precompile && HeaderModuleAction && 4088 getPrecompiledType(InputType) == types::TY_PCH) { 4089 HeaderModuleAction->addModuleHeaderInput(Current); 4090 Current = nullptr; 4091 break; 4092 } 4093 4094 if (Phase == phases::Precompile && ExtractAPIAction) { 4095 ExtractAPIAction->addHeaderInput(Current); 4096 Current = nullptr; 4097 break; 4098 } 4099 4100 // FIXME: Should we include any prior module file outputs as inputs of 4101 // later actions in the same command line? 4102 4103 // Otherwise construct the appropriate action. 4104 Action *NewCurrent = ConstructPhaseAction(C, Args, Phase, Current); 4105 4106 // We didn't create a new action, so we will just move to the next phase. 4107 if (NewCurrent == Current) 4108 continue; 4109 4110 if (auto *HMA = dyn_cast<HeaderModulePrecompileJobAction>(NewCurrent)) 4111 HeaderModuleAction = HMA; 4112 else if (auto *EAA = dyn_cast<ExtractAPIJobAction>(NewCurrent)) 4113 ExtractAPIAction = EAA; 4114 4115 Current = NewCurrent; 4116 4117 // Use the current host action in any of the offloading actions, if 4118 // required. 4119 if (!UseNewOffloadingDriver) 4120 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4121 break; 4122 4123 // Try to build the offloading actions and add the result as a dependency 4124 // to the host. 4125 if (UseNewOffloadingDriver) 4126 Current = BuildOffloadingActions(C, Args, I, Current); 4127 4128 if (Current->getType() == types::TY_Nothing) 4129 break; 4130 } 4131 4132 // If we ended with something, add to the output list. 4133 if (Current) 4134 Actions.push_back(Current); 4135 4136 // Add any top level actions generated for offloading. 4137 if (!UseNewOffloadingDriver) 4138 OffloadBuilder.appendTopLevelActions(Actions, Current, InputArg); 4139 else if (Current) 4140 Current->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4141 /*BoundArch=*/nullptr); 4142 } 4143 4144 // Add a link action if necessary. 4145 4146 if (LinkerInputs.empty()) { 4147 Arg *FinalPhaseArg; 4148 if (getFinalPhase(Args, &FinalPhaseArg) == phases::Link) 4149 if (!UseNewOffloadingDriver) 4150 OffloadBuilder.appendDeviceLinkActions(Actions); 4151 } 4152 4153 if (!LinkerInputs.empty()) { 4154 if (!UseNewOffloadingDriver) 4155 if (Action *Wrapper = OffloadBuilder.makeHostLinkAction()) 4156 LinkerInputs.push_back(Wrapper); 4157 Action *LA; 4158 // Check if this Linker Job should emit a static library. 4159 if (ShouldEmitStaticLibrary(Args)) { 4160 LA = C.MakeAction<StaticLibJobAction>(LinkerInputs, types::TY_Image); 4161 } else if (UseNewOffloadingDriver) { 4162 LA = C.MakeAction<LinkerWrapperJobAction>(LinkerInputs, types::TY_Image); 4163 LA->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4164 /*BoundArch=*/nullptr); 4165 } else { 4166 LA = C.MakeAction<LinkJobAction>(LinkerInputs, types::TY_Image); 4167 } 4168 if (!UseNewOffloadingDriver) 4169 LA = OffloadBuilder.processHostLinkAction(LA); 4170 Actions.push_back(LA); 4171 } 4172 4173 // Add an interface stubs merge action if necessary. 4174 if (!MergerInputs.empty()) 4175 Actions.push_back( 4176 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4177 4178 if (Args.hasArg(options::OPT_emit_interface_stubs)) { 4179 auto PhaseList = types::getCompilationPhases( 4180 types::TY_IFS_CPP, 4181 Args.hasArg(options::OPT_c) ? phases::Compile : phases::IfsMerge); 4182 4183 ActionList MergerInputs; 4184 4185 for (auto &I : Inputs) { 4186 types::ID InputType = I.first; 4187 const Arg *InputArg = I.second; 4188 4189 // Currently clang and the llvm assembler do not support generating symbol 4190 // stubs from assembly, so we skip the input on asm files. For ifs files 4191 // we rely on the normal pipeline setup in the pipeline setup code above. 4192 if (InputType == types::TY_IFS || InputType == types::TY_PP_Asm || 4193 InputType == types::TY_Asm) 4194 continue; 4195 4196 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4197 4198 for (auto Phase : PhaseList) { 4199 switch (Phase) { 4200 default: 4201 llvm_unreachable( 4202 "IFS Pipeline can only consist of Compile followed by IfsMerge."); 4203 case phases::Compile: { 4204 // Only IfsMerge (llvm-ifs) can handle .o files by looking for ifs 4205 // files where the .o file is located. The compile action can not 4206 // handle this. 4207 if (InputType == types::TY_Object) 4208 break; 4209 4210 Current = C.MakeAction<CompileJobAction>(Current, types::TY_IFS_CPP); 4211 break; 4212 } 4213 case phases::IfsMerge: { 4214 assert(Phase == PhaseList.back() && 4215 "merging must be final compilation step."); 4216 MergerInputs.push_back(Current); 4217 Current = nullptr; 4218 break; 4219 } 4220 } 4221 } 4222 4223 // If we ended with something, add to the output list. 4224 if (Current) 4225 Actions.push_back(Current); 4226 } 4227 4228 // Add an interface stubs merge action if necessary. 4229 if (!MergerInputs.empty()) 4230 Actions.push_back( 4231 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4232 } 4233 4234 // If --print-supported-cpus, -mcpu=? or -mtune=? is specified, build a custom 4235 // Compile phase that prints out supported cpu models and quits. 4236 if (Arg *A = Args.getLastArg(options::OPT_print_supported_cpus)) { 4237 // Use the -mcpu=? flag as the dummy input to cc1. 4238 Actions.clear(); 4239 Action *InputAc = C.MakeAction<InputAction>(*A, types::TY_C); 4240 Actions.push_back( 4241 C.MakeAction<PrecompileJobAction>(InputAc, types::TY_Nothing)); 4242 for (auto &I : Inputs) 4243 I.second->claim(); 4244 } 4245 4246 // Claim ignored clang-cl options. 4247 Args.ClaimAllArgs(options::OPT_cl_ignored_Group); 4248 4249 // Claim --offload-host-only and --offload-compile-host-device, which may be 4250 // passed to non-CUDA compilations and should not trigger warnings there. 4251 Args.ClaimAllArgs(options::OPT_offload_host_only); 4252 Args.ClaimAllArgs(options::OPT_offload_host_device); 4253 } 4254 4255 /// Returns the canonical name for the offloading architecture when using a HIP 4256 /// or CUDA architecture. 4257 static StringRef getCanonicalArchString(Compilation &C, 4258 const llvm::opt::DerivedArgList &Args, 4259 StringRef ArchStr, 4260 const llvm::Triple &Triple) { 4261 // Lookup the CUDA / HIP architecture string. Only report an error if we were 4262 // expecting the triple to be only NVPTX / AMDGPU. 4263 CudaArch Arch = StringToCudaArch(getProcessorFromTargetID(Triple, ArchStr)); 4264 if (Triple.isNVPTX() && 4265 (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch))) { 4266 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4267 << "CUDA" << ArchStr; 4268 return StringRef(); 4269 } else if (Triple.isAMDGPU() && 4270 (Arch == CudaArch::UNKNOWN || !IsAMDGpuArch(Arch))) { 4271 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4272 << "HIP" << ArchStr; 4273 return StringRef(); 4274 } 4275 4276 if (IsNVIDIAGpuArch(Arch)) 4277 return Args.MakeArgStringRef(CudaArchToString(Arch)); 4278 4279 if (IsAMDGpuArch(Arch)) { 4280 llvm::StringMap<bool> Features; 4281 auto HIPTriple = getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()); 4282 if (!HIPTriple) 4283 return StringRef(); 4284 auto Arch = parseTargetID(*HIPTriple, ArchStr, &Features); 4285 if (!Arch) { 4286 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << ArchStr; 4287 C.setContainsError(); 4288 return StringRef(); 4289 } 4290 return Args.MakeArgStringRef( 4291 getCanonicalTargetID(Arch.getValue(), Features)); 4292 } 4293 4294 // If the input isn't CUDA or HIP just return the architecture. 4295 return ArchStr; 4296 } 4297 4298 /// Checks if the set offloading architectures does not conflict. Returns the 4299 /// incompatible pair if a conflict occurs. 4300 static llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 4301 getConflictOffloadArchCombination(const llvm::DenseSet<StringRef> &Archs, 4302 Action::OffloadKind Kind) { 4303 if (Kind != Action::OFK_HIP) 4304 return None; 4305 4306 std::set<StringRef> ArchSet; 4307 llvm::copy(Archs, std::inserter(ArchSet, ArchSet.begin())); 4308 return getConflictTargetIDCombination(ArchSet); 4309 } 4310 4311 llvm::DenseSet<StringRef> 4312 Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args, 4313 Action::OffloadKind Kind, const ToolChain *TC) const { 4314 if (!TC) 4315 TC = &C.getDefaultToolChain(); 4316 4317 // --offload and --offload-arch options are mutually exclusive. 4318 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 4319 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 4320 options::OPT_no_offload_arch_EQ)) { 4321 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) 4322 << "--offload" 4323 << (Args.hasArgNoClaim(options::OPT_offload_arch_EQ) 4324 ? "--offload-arch" 4325 : "--no-offload-arch"); 4326 } 4327 4328 if (KnownArchs.find(TC) != KnownArchs.end()) 4329 return KnownArchs.lookup(TC); 4330 4331 llvm::DenseSet<StringRef> Archs; 4332 for (auto &Arg : Args) { 4333 if (Arg->getOption().matches(options::OPT_offload_arch_EQ)) { 4334 Archs.insert( 4335 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4336 } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) { 4337 if (Arg->getValue() == StringRef("all")) 4338 Archs.clear(); 4339 else 4340 Archs.erase( 4341 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4342 } 4343 } 4344 4345 if (auto ConflictingArchs = getConflictOffloadArchCombination(Archs, Kind)) { 4346 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 4347 << ConflictingArchs.getValue().first 4348 << ConflictingArchs.getValue().second; 4349 C.setContainsError(); 4350 } 4351 4352 if (Archs.empty()) { 4353 if (Kind == Action::OFK_Cuda) 4354 Archs.insert(CudaArchToString(CudaArch::CudaDefault)); 4355 else if (Kind == Action::OFK_HIP) 4356 Archs.insert(CudaArchToString(CudaArch::HIPDefault)); 4357 else if (Kind == Action::OFK_OpenMP) 4358 Archs.insert(StringRef()); 4359 } else { 4360 Args.ClaimAllArgs(options::OPT_offload_arch_EQ); 4361 Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ); 4362 } 4363 4364 return Archs; 4365 } 4366 4367 Action *Driver::BuildOffloadingActions(Compilation &C, 4368 llvm::opt::DerivedArgList &Args, 4369 const InputTy &Input, 4370 Action *HostAction) const { 4371 const Arg *Mode = Args.getLastArg(options::OPT_offload_host_only, 4372 options::OPT_offload_device_only, 4373 options::OPT_offload_host_device); 4374 const bool HostOnly = 4375 Mode && Mode->getOption().matches(options::OPT_offload_host_only); 4376 const bool DeviceOnly = 4377 Mode && Mode->getOption().matches(options::OPT_offload_device_only); 4378 4379 // Don't build offloading actions if explicitly disabled or we do not have a 4380 // compile action to embed it in. If preprocessing only ignore embedding. 4381 if (HostOnly || !(isa<CompileJobAction>(HostAction) || 4382 getFinalPhase(Args) == phases::Preprocess)) 4383 return HostAction; 4384 4385 ActionList OffloadActions; 4386 OffloadAction::DeviceDependences DDeps; 4387 4388 const Action::OffloadKind OffloadKinds[] = { 4389 Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP}; 4390 4391 for (Action::OffloadKind Kind : OffloadKinds) { 4392 SmallVector<const ToolChain *, 2> ToolChains; 4393 ActionList DeviceActions; 4394 4395 auto TCRange = C.getOffloadToolChains(Kind); 4396 for (auto TI = TCRange.first, TE = TCRange.second; TI != TE; ++TI) 4397 ToolChains.push_back(TI->second); 4398 4399 if (ToolChains.empty()) 4400 continue; 4401 4402 types::ID InputType = Input.first; 4403 const Arg *InputArg = Input.second; 4404 4405 // Get the product of all bound architectures and toolchains. 4406 SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs; 4407 for (const ToolChain *TC : ToolChains) 4408 for (StringRef Arch : getOffloadArchs( 4409 C, C.getArgsForToolChain(TC, "generic", Kind), Kind, TC)) 4410 TCAndArchs.push_back(std::make_pair(TC, Arch)); 4411 4412 for (unsigned I = 0, E = TCAndArchs.size(); I != E; ++I) 4413 DeviceActions.push_back(C.MakeAction<InputAction>(*InputArg, InputType)); 4414 4415 if (DeviceActions.empty()) 4416 return HostAction; 4417 4418 auto PL = types::getCompilationPhases(*this, Args, InputType); 4419 4420 for (phases::ID Phase : PL) { 4421 if (Phase == phases::Link) { 4422 assert(Phase == PL.back() && "linking must be final compilation step."); 4423 break; 4424 } 4425 4426 auto TCAndArch = TCAndArchs.begin(); 4427 for (Action *&A : DeviceActions) { 4428 A = ConstructPhaseAction(C, Args, Phase, A, Kind); 4429 4430 if (isa<CompileJobAction>(A) && isa<CompileJobAction>(HostAction) && 4431 Kind == Action::OFK_OpenMP) { 4432 // OpenMP offloading has a dependency on the host compile action to 4433 // identify which declarations need to be emitted. This shouldn't be 4434 // collapsed with any other actions so we can use it in the device. 4435 HostAction->setCannotBeCollapsedWithNextDependentAction(); 4436 OffloadAction::HostDependence HDep( 4437 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4438 TCAndArch->second.data(), Kind); 4439 OffloadAction::DeviceDependences DDep; 4440 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4441 A = C.MakeAction<OffloadAction>(HDep, DDep); 4442 } else if (isa<AssembleJobAction>(A) && Kind == Action::OFK_Cuda) { 4443 // The Cuda toolchain uses fatbinary as the linker phase to bundle the 4444 // PTX and Cubin output. 4445 ActionList FatbinActions; 4446 for (Action *A : {A, A->getInputs()[0]}) { 4447 OffloadAction::DeviceDependences DDep; 4448 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4449 FatbinActions.emplace_back( 4450 C.MakeAction<OffloadAction>(DDep, A->getType())); 4451 } 4452 A = C.MakeAction<LinkJobAction>(FatbinActions, types::TY_CUDA_FATBIN); 4453 } 4454 ++TCAndArch; 4455 } 4456 } 4457 4458 auto TCAndArch = TCAndArchs.begin(); 4459 for (Action *A : DeviceActions) { 4460 DDeps.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4461 OffloadAction::DeviceDependences DDep; 4462 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4463 OffloadActions.push_back(C.MakeAction<OffloadAction>(DDep, A->getType())); 4464 ++TCAndArch; 4465 } 4466 } 4467 4468 if (DeviceOnly) 4469 return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing); 4470 4471 Action *OffloadPackager = 4472 C.MakeAction<OffloadPackagerJobAction>(OffloadActions, types::TY_Image); 4473 OffloadAction::DeviceDependences DDep; 4474 DDep.add(*OffloadPackager, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4475 nullptr, Action::OFK_None); 4476 OffloadAction::HostDependence HDep( 4477 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4478 /*BoundArch=*/nullptr, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4479 return C.MakeAction<OffloadAction>( 4480 HDep, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4481 } 4482 4483 Action *Driver::ConstructPhaseAction( 4484 Compilation &C, const ArgList &Args, phases::ID Phase, Action *Input, 4485 Action::OffloadKind TargetDeviceOffloadKind) const { 4486 llvm::PrettyStackTraceString CrashInfo("Constructing phase actions"); 4487 4488 // Some types skip the assembler phase (e.g., llvm-bc), but we can't 4489 // encode this in the steps because the intermediate type depends on 4490 // arguments. Just special case here. 4491 if (Phase == phases::Assemble && Input->getType() != types::TY_PP_Asm) 4492 return Input; 4493 4494 // Build the appropriate action. 4495 switch (Phase) { 4496 case phases::Link: 4497 llvm_unreachable("link action invalid here."); 4498 case phases::IfsMerge: 4499 llvm_unreachable("ifsmerge action invalid here."); 4500 case phases::Preprocess: { 4501 types::ID OutputTy; 4502 // -M and -MM specify the dependency file name by altering the output type, 4503 // -if -MD and -MMD are not specified. 4504 if (Args.hasArg(options::OPT_M, options::OPT_MM) && 4505 !Args.hasArg(options::OPT_MD, options::OPT_MMD)) { 4506 OutputTy = types::TY_Dependencies; 4507 } else { 4508 OutputTy = Input->getType(); 4509 // For these cases, the preprocessor is only translating forms, the Output 4510 // still needs preprocessing. 4511 if (!Args.hasFlag(options::OPT_frewrite_includes, 4512 options::OPT_fno_rewrite_includes, false) && 4513 !Args.hasFlag(options::OPT_frewrite_imports, 4514 options::OPT_fno_rewrite_imports, false) && 4515 !Args.hasFlag(options::OPT_fdirectives_only, 4516 options::OPT_fno_directives_only, false) && 4517 !CCGenDiagnostics) 4518 OutputTy = types::getPreprocessedType(OutputTy); 4519 assert(OutputTy != types::TY_INVALID && 4520 "Cannot preprocess this input type!"); 4521 } 4522 return C.MakeAction<PreprocessJobAction>(Input, OutputTy); 4523 } 4524 case phases::Precompile: { 4525 // API extraction should not generate an actual precompilation action. 4526 if (Args.hasArg(options::OPT_extract_api)) 4527 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4528 4529 types::ID OutputTy = getPrecompiledType(Input->getType()); 4530 assert(OutputTy != types::TY_INVALID && 4531 "Cannot precompile this input type!"); 4532 4533 // If we're given a module name, precompile header file inputs as a 4534 // module, not as a precompiled header. 4535 const char *ModName = nullptr; 4536 if (OutputTy == types::TY_PCH) { 4537 if (Arg *A = Args.getLastArg(options::OPT_fmodule_name_EQ)) 4538 ModName = A->getValue(); 4539 if (ModName) 4540 OutputTy = types::TY_ModuleFile; 4541 } 4542 4543 if (Args.hasArg(options::OPT_fsyntax_only)) { 4544 // Syntax checks should not emit a PCH file 4545 OutputTy = types::TY_Nothing; 4546 } 4547 4548 if (ModName) 4549 return C.MakeAction<HeaderModulePrecompileJobAction>(Input, OutputTy, 4550 ModName); 4551 return C.MakeAction<PrecompileJobAction>(Input, OutputTy); 4552 } 4553 case phases::Compile: { 4554 if (Args.hasArg(options::OPT_fsyntax_only)) 4555 return C.MakeAction<CompileJobAction>(Input, types::TY_Nothing); 4556 if (Args.hasArg(options::OPT_rewrite_objc)) 4557 return C.MakeAction<CompileJobAction>(Input, types::TY_RewrittenObjC); 4558 if (Args.hasArg(options::OPT_rewrite_legacy_objc)) 4559 return C.MakeAction<CompileJobAction>(Input, 4560 types::TY_RewrittenLegacyObjC); 4561 if (Args.hasArg(options::OPT__analyze)) 4562 return C.MakeAction<AnalyzeJobAction>(Input, types::TY_Plist); 4563 if (Args.hasArg(options::OPT__migrate)) 4564 return C.MakeAction<MigrateJobAction>(Input, types::TY_Remap); 4565 if (Args.hasArg(options::OPT_emit_ast)) 4566 return C.MakeAction<CompileJobAction>(Input, types::TY_AST); 4567 if (Args.hasArg(options::OPT_module_file_info)) 4568 return C.MakeAction<CompileJobAction>(Input, types::TY_ModuleFile); 4569 if (Args.hasArg(options::OPT_verify_pch)) 4570 return C.MakeAction<VerifyPCHJobAction>(Input, types::TY_Nothing); 4571 if (Args.hasArg(options::OPT_extract_api)) 4572 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4573 return C.MakeAction<CompileJobAction>(Input, types::TY_LLVM_BC); 4574 } 4575 case phases::Backend: { 4576 if (isUsingLTO() && TargetDeviceOffloadKind == Action::OFK_None) { 4577 types::ID Output = 4578 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4579 return C.MakeAction<BackendJobAction>(Input, Output); 4580 } 4581 if (isUsingLTO(/* IsOffload */ true) && 4582 TargetDeviceOffloadKind != Action::OFK_None) { 4583 types::ID Output = 4584 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4585 return C.MakeAction<BackendJobAction>(Input, Output); 4586 } 4587 if (Args.hasArg(options::OPT_emit_llvm) || 4588 (TargetDeviceOffloadKind == Action::OFK_HIP && 4589 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4590 false))) { 4591 types::ID Output = 4592 Args.hasArg(options::OPT_S) ? types::TY_LLVM_IR : types::TY_LLVM_BC; 4593 return C.MakeAction<BackendJobAction>(Input, Output); 4594 } 4595 return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm); 4596 } 4597 case phases::Assemble: 4598 return C.MakeAction<AssembleJobAction>(std::move(Input), types::TY_Object); 4599 } 4600 4601 llvm_unreachable("invalid phase in ConstructPhaseAction"); 4602 } 4603 4604 void Driver::BuildJobs(Compilation &C) const { 4605 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 4606 4607 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 4608 4609 // It is an error to provide a -o option if we are making multiple output 4610 // files. There are exceptions: 4611 // 4612 // IfsMergeJob: when generating interface stubs enabled we want to be able to 4613 // generate the stub file at the same time that we generate the real 4614 // library/a.out. So when a .o, .so, etc are the output, with clang interface 4615 // stubs there will also be a .ifs and .ifso at the same location. 4616 // 4617 // CompileJob of type TY_IFS_CPP: when generating interface stubs is enabled 4618 // and -c is passed, we still want to be able to generate a .ifs file while 4619 // we are also generating .o files. So we allow more than one output file in 4620 // this case as well. 4621 // 4622 if (FinalOutput) { 4623 unsigned NumOutputs = 0; 4624 unsigned NumIfsOutputs = 0; 4625 for (const Action *A : C.getActions()) 4626 if (A->getType() != types::TY_Nothing && 4627 !(A->getKind() == Action::IfsMergeJobClass || 4628 (A->getType() == clang::driver::types::TY_IFS_CPP && 4629 A->getKind() == clang::driver::Action::CompileJobClass && 4630 0 == NumIfsOutputs++) || 4631 (A->getKind() == Action::BindArchClass && A->getInputs().size() && 4632 A->getInputs().front()->getKind() == Action::IfsMergeJobClass))) 4633 ++NumOutputs; 4634 4635 if (NumOutputs > 1) { 4636 Diag(clang::diag::err_drv_output_argument_with_multiple_files); 4637 FinalOutput = nullptr; 4638 } 4639 } 4640 4641 const llvm::Triple &RawTriple = C.getDefaultToolChain().getTriple(); 4642 if (RawTriple.isOSAIX()) { 4643 if (Arg *A = C.getArgs().getLastArg(options::OPT_G)) 4644 Diag(diag::err_drv_unsupported_opt_for_target) 4645 << A->getSpelling() << RawTriple.str(); 4646 if (LTOMode == LTOK_Thin) 4647 Diag(diag::err_drv_clang_unsupported) << "thinLTO on AIX"; 4648 } 4649 4650 // Collect the list of architectures. 4651 llvm::StringSet<> ArchNames; 4652 if (RawTriple.isOSBinFormatMachO()) 4653 for (const Arg *A : C.getArgs()) 4654 if (A->getOption().matches(options::OPT_arch)) 4655 ArchNames.insert(A->getValue()); 4656 4657 // Set of (Action, canonical ToolChain triple) pairs we've built jobs for. 4658 std::map<std::pair<const Action *, std::string>, InputInfoList> CachedResults; 4659 for (Action *A : C.getActions()) { 4660 // If we are linking an image for multiple archs then the linker wants 4661 // -arch_multiple and -final_output <final image name>. Unfortunately, this 4662 // doesn't fit in cleanly because we have to pass this information down. 4663 // 4664 // FIXME: This is a hack; find a cleaner way to integrate this into the 4665 // process. 4666 const char *LinkingOutput = nullptr; 4667 if (isa<LipoJobAction>(A)) { 4668 if (FinalOutput) 4669 LinkingOutput = FinalOutput->getValue(); 4670 else 4671 LinkingOutput = getDefaultImageName(); 4672 } 4673 4674 BuildJobsForAction(C, A, &C.getDefaultToolChain(), 4675 /*BoundArch*/ StringRef(), 4676 /*AtTopLevel*/ true, 4677 /*MultipleArchs*/ ArchNames.size() > 1, 4678 /*LinkingOutput*/ LinkingOutput, CachedResults, 4679 /*TargetDeviceOffloadKind*/ Action::OFK_None); 4680 } 4681 4682 // If we have more than one job, then disable integrated-cc1 for now. Do this 4683 // also when we need to report process execution statistics. 4684 if (C.getJobs().size() > 1 || CCPrintProcessStats) 4685 for (auto &J : C.getJobs()) 4686 J.InProcess = false; 4687 4688 if (CCPrintProcessStats) { 4689 C.setPostCallback([=](const Command &Cmd, int Res) { 4690 Optional<llvm::sys::ProcessStatistics> ProcStat = 4691 Cmd.getProcessStatistics(); 4692 if (!ProcStat) 4693 return; 4694 4695 const char *LinkingOutput = nullptr; 4696 if (FinalOutput) 4697 LinkingOutput = FinalOutput->getValue(); 4698 else if (!Cmd.getOutputFilenames().empty()) 4699 LinkingOutput = Cmd.getOutputFilenames().front().c_str(); 4700 else 4701 LinkingOutput = getDefaultImageName(); 4702 4703 if (CCPrintStatReportFilename.empty()) { 4704 using namespace llvm; 4705 // Human readable output. 4706 outs() << sys::path::filename(Cmd.getExecutable()) << ": " 4707 << "output=" << LinkingOutput; 4708 outs() << ", total=" 4709 << format("%.3f", ProcStat->TotalTime.count() / 1000.) << " ms" 4710 << ", user=" 4711 << format("%.3f", ProcStat->UserTime.count() / 1000.) << " ms" 4712 << ", mem=" << ProcStat->PeakMemory << " Kb\n"; 4713 } else { 4714 // CSV format. 4715 std::string Buffer; 4716 llvm::raw_string_ostream Out(Buffer); 4717 llvm::sys::printArg(Out, llvm::sys::path::filename(Cmd.getExecutable()), 4718 /*Quote*/ true); 4719 Out << ','; 4720 llvm::sys::printArg(Out, LinkingOutput, true); 4721 Out << ',' << ProcStat->TotalTime.count() << ',' 4722 << ProcStat->UserTime.count() << ',' << ProcStat->PeakMemory 4723 << '\n'; 4724 Out.flush(); 4725 std::error_code EC; 4726 llvm::raw_fd_ostream OS(CCPrintStatReportFilename, EC, 4727 llvm::sys::fs::OF_Append | 4728 llvm::sys::fs::OF_Text); 4729 if (EC) 4730 return; 4731 auto L = OS.lock(); 4732 if (!L) { 4733 llvm::errs() << "ERROR: Cannot lock file " 4734 << CCPrintStatReportFilename << ": " 4735 << toString(L.takeError()) << "\n"; 4736 return; 4737 } 4738 OS << Buffer; 4739 OS.flush(); 4740 } 4741 }); 4742 } 4743 4744 // If the user passed -Qunused-arguments or there were errors, don't warn 4745 // about any unused arguments. 4746 if (Diags.hasErrorOccurred() || 4747 C.getArgs().hasArg(options::OPT_Qunused_arguments)) 4748 return; 4749 4750 // Claim -### here. 4751 (void)C.getArgs().hasArg(options::OPT__HASH_HASH_HASH); 4752 4753 // Claim --driver-mode, --rsp-quoting, it was handled earlier. 4754 (void)C.getArgs().hasArg(options::OPT_driver_mode); 4755 (void)C.getArgs().hasArg(options::OPT_rsp_quoting); 4756 4757 for (Arg *A : C.getArgs()) { 4758 // FIXME: It would be nice to be able to send the argument to the 4759 // DiagnosticsEngine, so that extra values, position, and so on could be 4760 // printed. 4761 if (!A->isClaimed()) { 4762 if (A->getOption().hasFlag(options::NoArgumentUnused)) 4763 continue; 4764 4765 // Suppress the warning automatically if this is just a flag, and it is an 4766 // instance of an argument we already claimed. 4767 const Option &Opt = A->getOption(); 4768 if (Opt.getKind() == Option::FlagClass) { 4769 bool DuplicateClaimed = false; 4770 4771 for (const Arg *AA : C.getArgs().filtered(&Opt)) { 4772 if (AA->isClaimed()) { 4773 DuplicateClaimed = true; 4774 break; 4775 } 4776 } 4777 4778 if (DuplicateClaimed) 4779 continue; 4780 } 4781 4782 // In clang-cl, don't mention unknown arguments here since they have 4783 // already been warned about. 4784 if (!IsCLMode() || !A->getOption().matches(options::OPT_UNKNOWN)) 4785 Diag(clang::diag::warn_drv_unused_argument) 4786 << A->getAsString(C.getArgs()); 4787 } 4788 } 4789 } 4790 4791 namespace { 4792 /// Utility class to control the collapse of dependent actions and select the 4793 /// tools accordingly. 4794 class ToolSelector final { 4795 /// The tool chain this selector refers to. 4796 const ToolChain &TC; 4797 4798 /// The compilation this selector refers to. 4799 const Compilation &C; 4800 4801 /// The base action this selector refers to. 4802 const JobAction *BaseAction; 4803 4804 /// Set to true if the current toolchain refers to host actions. 4805 bool IsHostSelector; 4806 4807 /// Set to true if save-temps and embed-bitcode functionalities are active. 4808 bool SaveTemps; 4809 bool EmbedBitcode; 4810 4811 /// Get previous dependent action or null if that does not exist. If 4812 /// \a CanBeCollapsed is false, that action must be legal to collapse or 4813 /// null will be returned. 4814 const JobAction *getPrevDependentAction(const ActionList &Inputs, 4815 ActionList &SavedOffloadAction, 4816 bool CanBeCollapsed = true) { 4817 // An option can be collapsed only if it has a single input. 4818 if (Inputs.size() != 1) 4819 return nullptr; 4820 4821 Action *CurAction = *Inputs.begin(); 4822 if (CanBeCollapsed && 4823 !CurAction->isCollapsingWithNextDependentActionLegal()) 4824 return nullptr; 4825 4826 // If the input action is an offload action. Look through it and save any 4827 // offload action that can be dropped in the event of a collapse. 4828 if (auto *OA = dyn_cast<OffloadAction>(CurAction)) { 4829 // If the dependent action is a device action, we will attempt to collapse 4830 // only with other device actions. Otherwise, we would do the same but 4831 // with host actions only. 4832 if (!IsHostSelector) { 4833 if (OA->hasSingleDeviceDependence(/*DoNotConsiderHostActions=*/true)) { 4834 CurAction = 4835 OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true); 4836 if (CanBeCollapsed && 4837 !CurAction->isCollapsingWithNextDependentActionLegal()) 4838 return nullptr; 4839 SavedOffloadAction.push_back(OA); 4840 return dyn_cast<JobAction>(CurAction); 4841 } 4842 } else if (OA->hasHostDependence()) { 4843 CurAction = OA->getHostDependence(); 4844 if (CanBeCollapsed && 4845 !CurAction->isCollapsingWithNextDependentActionLegal()) 4846 return nullptr; 4847 SavedOffloadAction.push_back(OA); 4848 return dyn_cast<JobAction>(CurAction); 4849 } 4850 return nullptr; 4851 } 4852 4853 return dyn_cast<JobAction>(CurAction); 4854 } 4855 4856 /// Return true if an assemble action can be collapsed. 4857 bool canCollapseAssembleAction() const { 4858 return TC.useIntegratedAs() && !SaveTemps && 4859 !C.getArgs().hasArg(options::OPT_via_file_asm) && 4860 !C.getArgs().hasArg(options::OPT__SLASH_FA) && 4861 !C.getArgs().hasArg(options::OPT__SLASH_Fa); 4862 } 4863 4864 /// Return true if a preprocessor action can be collapsed. 4865 bool canCollapsePreprocessorAction() const { 4866 return !C.getArgs().hasArg(options::OPT_no_integrated_cpp) && 4867 !C.getArgs().hasArg(options::OPT_traditional_cpp) && !SaveTemps && 4868 !C.getArgs().hasArg(options::OPT_rewrite_objc); 4869 } 4870 4871 /// Struct that relates an action with the offload actions that would be 4872 /// collapsed with it. 4873 struct JobActionInfo final { 4874 /// The action this info refers to. 4875 const JobAction *JA = nullptr; 4876 /// The offload actions we need to take care off if this action is 4877 /// collapsed. 4878 ActionList SavedOffloadAction; 4879 }; 4880 4881 /// Append collapsed offload actions from the give nnumber of elements in the 4882 /// action info array. 4883 static void AppendCollapsedOffloadAction(ActionList &CollapsedOffloadAction, 4884 ArrayRef<JobActionInfo> &ActionInfo, 4885 unsigned ElementNum) { 4886 assert(ElementNum <= ActionInfo.size() && "Invalid number of elements."); 4887 for (unsigned I = 0; I < ElementNum; ++I) 4888 CollapsedOffloadAction.append(ActionInfo[I].SavedOffloadAction.begin(), 4889 ActionInfo[I].SavedOffloadAction.end()); 4890 } 4891 4892 /// Functions that attempt to perform the combining. They detect if that is 4893 /// legal, and if so they update the inputs \a Inputs and the offload action 4894 /// that were collapsed in \a CollapsedOffloadAction. A tool that deals with 4895 /// the combined action is returned. If the combining is not legal or if the 4896 /// tool does not exist, null is returned. 4897 /// Currently three kinds of collapsing are supported: 4898 /// - Assemble + Backend + Compile; 4899 /// - Assemble + Backend ; 4900 /// - Backend + Compile. 4901 const Tool * 4902 combineAssembleBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4903 ActionList &Inputs, 4904 ActionList &CollapsedOffloadAction) { 4905 if (ActionInfo.size() < 3 || !canCollapseAssembleAction()) 4906 return nullptr; 4907 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4908 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4909 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[2].JA); 4910 if (!AJ || !BJ || !CJ) 4911 return nullptr; 4912 4913 // Get compiler tool. 4914 const Tool *T = TC.SelectTool(*CJ); 4915 if (!T) 4916 return nullptr; 4917 4918 // Can't collapse if we don't have codegen support unless we are 4919 // emitting LLVM IR. 4920 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 4921 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 4922 return nullptr; 4923 4924 // When using -fembed-bitcode, it is required to have the same tool (clang) 4925 // for both CompilerJA and BackendJA. Otherwise, combine two stages. 4926 if (EmbedBitcode) { 4927 const Tool *BT = TC.SelectTool(*BJ); 4928 if (BT == T) 4929 return nullptr; 4930 } 4931 4932 if (!T->hasIntegratedAssembler()) 4933 return nullptr; 4934 4935 Inputs = CJ->getInputs(); 4936 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4937 /*NumElements=*/3); 4938 return T; 4939 } 4940 const Tool *combineAssembleBackend(ArrayRef<JobActionInfo> ActionInfo, 4941 ActionList &Inputs, 4942 ActionList &CollapsedOffloadAction) { 4943 if (ActionInfo.size() < 2 || !canCollapseAssembleAction()) 4944 return nullptr; 4945 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4946 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4947 if (!AJ || !BJ) 4948 return nullptr; 4949 4950 // Get backend tool. 4951 const Tool *T = TC.SelectTool(*BJ); 4952 if (!T) 4953 return nullptr; 4954 4955 if (!T->hasIntegratedAssembler()) 4956 return nullptr; 4957 4958 Inputs = BJ->getInputs(); 4959 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4960 /*NumElements=*/2); 4961 return T; 4962 } 4963 const Tool *combineBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4964 ActionList &Inputs, 4965 ActionList &CollapsedOffloadAction) { 4966 if (ActionInfo.size() < 2) 4967 return nullptr; 4968 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[0].JA); 4969 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[1].JA); 4970 if (!BJ || !CJ) 4971 return nullptr; 4972 4973 // Check if the initial input (to the compile job or its predessor if one 4974 // exists) is LLVM bitcode. In that case, no preprocessor step is required 4975 // and we can still collapse the compile and backend jobs when we have 4976 // -save-temps. I.e. there is no need for a separate compile job just to 4977 // emit unoptimized bitcode. 4978 bool InputIsBitcode = true; 4979 for (size_t i = 1; i < ActionInfo.size(); i++) 4980 if (ActionInfo[i].JA->getType() != types::TY_LLVM_BC && 4981 ActionInfo[i].JA->getType() != types::TY_LTO_BC) { 4982 InputIsBitcode = false; 4983 break; 4984 } 4985 if (!InputIsBitcode && !canCollapsePreprocessorAction()) 4986 return nullptr; 4987 4988 // Get compiler tool. 4989 const Tool *T = TC.SelectTool(*CJ); 4990 if (!T) 4991 return nullptr; 4992 4993 // Can't collapse if we don't have codegen support unless we are 4994 // emitting LLVM IR. 4995 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 4996 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 4997 return nullptr; 4998 4999 if (T->canEmitIR() && ((SaveTemps && !InputIsBitcode) || EmbedBitcode)) 5000 return nullptr; 5001 5002 Inputs = CJ->getInputs(); 5003 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5004 /*NumElements=*/2); 5005 return T; 5006 } 5007 5008 /// Updates the inputs if the obtained tool supports combining with 5009 /// preprocessor action, and the current input is indeed a preprocessor 5010 /// action. If combining results in the collapse of offloading actions, those 5011 /// are appended to \a CollapsedOffloadAction. 5012 void combineWithPreprocessor(const Tool *T, ActionList &Inputs, 5013 ActionList &CollapsedOffloadAction) { 5014 if (!T || !canCollapsePreprocessorAction() || !T->hasIntegratedCPP()) 5015 return; 5016 5017 // Attempt to get a preprocessor action dependence. 5018 ActionList PreprocessJobOffloadActions; 5019 ActionList NewInputs; 5020 for (Action *A : Inputs) { 5021 auto *PJ = getPrevDependentAction({A}, PreprocessJobOffloadActions); 5022 if (!PJ || !isa<PreprocessJobAction>(PJ)) { 5023 NewInputs.push_back(A); 5024 continue; 5025 } 5026 5027 // This is legal to combine. Append any offload action we found and add the 5028 // current input to preprocessor inputs. 5029 CollapsedOffloadAction.append(PreprocessJobOffloadActions.begin(), 5030 PreprocessJobOffloadActions.end()); 5031 NewInputs.append(PJ->input_begin(), PJ->input_end()); 5032 } 5033 Inputs = NewInputs; 5034 } 5035 5036 public: 5037 ToolSelector(const JobAction *BaseAction, const ToolChain &TC, 5038 const Compilation &C, bool SaveTemps, bool EmbedBitcode) 5039 : TC(TC), C(C), BaseAction(BaseAction), SaveTemps(SaveTemps), 5040 EmbedBitcode(EmbedBitcode) { 5041 assert(BaseAction && "Invalid base action."); 5042 IsHostSelector = BaseAction->getOffloadingDeviceKind() == Action::OFK_None; 5043 } 5044 5045 /// Check if a chain of actions can be combined and return the tool that can 5046 /// handle the combination of actions. The pointer to the current inputs \a 5047 /// Inputs and the list of offload actions \a CollapsedOffloadActions 5048 /// connected to collapsed actions are updated accordingly. The latter enables 5049 /// the caller of the selector to process them afterwards instead of just 5050 /// dropping them. If no suitable tool is found, null will be returned. 5051 const Tool *getTool(ActionList &Inputs, 5052 ActionList &CollapsedOffloadAction) { 5053 // 5054 // Get the largest chain of actions that we could combine. 5055 // 5056 5057 SmallVector<JobActionInfo, 5> ActionChain(1); 5058 ActionChain.back().JA = BaseAction; 5059 while (ActionChain.back().JA) { 5060 const Action *CurAction = ActionChain.back().JA; 5061 5062 // Grow the chain by one element. 5063 ActionChain.resize(ActionChain.size() + 1); 5064 JobActionInfo &AI = ActionChain.back(); 5065 5066 // Attempt to fill it with the 5067 AI.JA = 5068 getPrevDependentAction(CurAction->getInputs(), AI.SavedOffloadAction); 5069 } 5070 5071 // Pop the last action info as it could not be filled. 5072 ActionChain.pop_back(); 5073 5074 // 5075 // Attempt to combine actions. If all combining attempts failed, just return 5076 // the tool of the provided action. At the end we attempt to combine the 5077 // action with any preprocessor action it may depend on. 5078 // 5079 5080 const Tool *T = combineAssembleBackendCompile(ActionChain, Inputs, 5081 CollapsedOffloadAction); 5082 if (!T) 5083 T = combineAssembleBackend(ActionChain, Inputs, CollapsedOffloadAction); 5084 if (!T) 5085 T = combineBackendCompile(ActionChain, Inputs, CollapsedOffloadAction); 5086 if (!T) { 5087 Inputs = BaseAction->getInputs(); 5088 T = TC.SelectTool(*BaseAction); 5089 } 5090 5091 combineWithPreprocessor(T, Inputs, CollapsedOffloadAction); 5092 return T; 5093 } 5094 }; 5095 } 5096 5097 /// Return a string that uniquely identifies the result of a job. The bound arch 5098 /// is not necessarily represented in the toolchain's triple -- for example, 5099 /// armv7 and armv7s both map to the same triple -- so we need both in our map. 5100 /// Also, we need to add the offloading device kind, as the same tool chain can 5101 /// be used for host and device for some programming models, e.g. OpenMP. 5102 static std::string GetTriplePlusArchString(const ToolChain *TC, 5103 StringRef BoundArch, 5104 Action::OffloadKind OffloadKind) { 5105 std::string TriplePlusArch = TC->getTriple().normalize(); 5106 if (!BoundArch.empty()) { 5107 TriplePlusArch += "-"; 5108 TriplePlusArch += BoundArch; 5109 } 5110 TriplePlusArch += "-"; 5111 TriplePlusArch += Action::GetOffloadKindName(OffloadKind); 5112 return TriplePlusArch; 5113 } 5114 5115 InputInfoList Driver::BuildJobsForAction( 5116 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5117 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5118 std::map<std::pair<const Action *, std::string>, InputInfoList> 5119 &CachedResults, 5120 Action::OffloadKind TargetDeviceOffloadKind) const { 5121 std::pair<const Action *, std::string> ActionTC = { 5122 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5123 auto CachedResult = CachedResults.find(ActionTC); 5124 if (CachedResult != CachedResults.end()) { 5125 return CachedResult->second; 5126 } 5127 InputInfoList Result = BuildJobsForActionNoCache( 5128 C, A, TC, BoundArch, AtTopLevel, MultipleArchs, LinkingOutput, 5129 CachedResults, TargetDeviceOffloadKind); 5130 CachedResults[ActionTC] = Result; 5131 return Result; 5132 } 5133 5134 InputInfoList Driver::BuildJobsForActionNoCache( 5135 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5136 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5137 std::map<std::pair<const Action *, std::string>, InputInfoList> 5138 &CachedResults, 5139 Action::OffloadKind TargetDeviceOffloadKind) const { 5140 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 5141 5142 InputInfoList OffloadDependencesInputInfo; 5143 bool BuildingForOffloadDevice = TargetDeviceOffloadKind != Action::OFK_None; 5144 if (const OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 5145 // The 'Darwin' toolchain is initialized only when its arguments are 5146 // computed. Get the default arguments for OFK_None to ensure that 5147 // initialization is performed before processing the offload action. 5148 // FIXME: Remove when darwin's toolchain is initialized during construction. 5149 C.getArgsForToolChain(TC, BoundArch, Action::OFK_None); 5150 5151 // The offload action is expected to be used in four different situations. 5152 // 5153 // a) Set a toolchain/architecture/kind for a host action: 5154 // Host Action 1 -> OffloadAction -> Host Action 2 5155 // 5156 // b) Set a toolchain/architecture/kind for a device action; 5157 // Device Action 1 -> OffloadAction -> Device Action 2 5158 // 5159 // c) Specify a device dependence to a host action; 5160 // Device Action 1 _ 5161 // \ 5162 // Host Action 1 ---> OffloadAction -> Host Action 2 5163 // 5164 // d) Specify a host dependence to a device action. 5165 // Host Action 1 _ 5166 // \ 5167 // Device Action 1 ---> OffloadAction -> Device Action 2 5168 // 5169 // For a) and b), we just return the job generated for the dependence. For 5170 // c) and d) we override the current action with the host/device dependence 5171 // if the current toolchain is host/device and set the offload dependences 5172 // info with the jobs obtained from the device/host dependence(s). 5173 5174 // If there is a single device option, just generate the job for it. 5175 if (OA->hasSingleDeviceDependence()) { 5176 InputInfoList DevA; 5177 OA->doOnEachDeviceDependence([&](Action *DepA, const ToolChain *DepTC, 5178 const char *DepBoundArch) { 5179 DevA = 5180 BuildJobsForAction(C, DepA, DepTC, DepBoundArch, AtTopLevel, 5181 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, 5182 CachedResults, DepA->getOffloadingDeviceKind()); 5183 }); 5184 return DevA; 5185 } 5186 5187 // If 'Action 2' is host, we generate jobs for the device dependences and 5188 // override the current action with the host dependence. Otherwise, we 5189 // generate the host dependences and override the action with the device 5190 // dependence. The dependences can't therefore be a top-level action. 5191 OA->doOnEachDependence( 5192 /*IsHostDependence=*/BuildingForOffloadDevice, 5193 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5194 OffloadDependencesInputInfo.append(BuildJobsForAction( 5195 C, DepA, DepTC, DepBoundArch, /*AtTopLevel=*/false, 5196 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, CachedResults, 5197 DepA->getOffloadingDeviceKind())); 5198 }); 5199 5200 A = BuildingForOffloadDevice 5201 ? OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true) 5202 : OA->getHostDependence(); 5203 5204 // We may have already built this action as a part of the offloading 5205 // toolchain, return the cached input if so. 5206 std::pair<const Action *, std::string> ActionTC = { 5207 OA->getHostDependence(), 5208 GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5209 if (CachedResults.find(ActionTC) != CachedResults.end()) { 5210 InputInfoList Inputs = CachedResults[ActionTC]; 5211 Inputs.append(OffloadDependencesInputInfo); 5212 return Inputs; 5213 } 5214 } 5215 5216 if (const InputAction *IA = dyn_cast<InputAction>(A)) { 5217 // FIXME: It would be nice to not claim this here; maybe the old scheme of 5218 // just using Args was better? 5219 const Arg &Input = IA->getInputArg(); 5220 Input.claim(); 5221 if (Input.getOption().matches(options::OPT_INPUT)) { 5222 const char *Name = Input.getValue(); 5223 return {InputInfo(A, Name, /* _BaseInput = */ Name)}; 5224 } 5225 return {InputInfo(A, &Input, /* _BaseInput = */ "")}; 5226 } 5227 5228 if (const BindArchAction *BAA = dyn_cast<BindArchAction>(A)) { 5229 const ToolChain *TC; 5230 StringRef ArchName = BAA->getArchName(); 5231 5232 if (!ArchName.empty()) 5233 TC = &getToolChain(C.getArgs(), 5234 computeTargetTriple(*this, TargetTriple, 5235 C.getArgs(), ArchName)); 5236 else 5237 TC = &C.getDefaultToolChain(); 5238 5239 return BuildJobsForAction(C, *BAA->input_begin(), TC, ArchName, AtTopLevel, 5240 MultipleArchs, LinkingOutput, CachedResults, 5241 TargetDeviceOffloadKind); 5242 } 5243 5244 5245 ActionList Inputs = A->getInputs(); 5246 5247 const JobAction *JA = cast<JobAction>(A); 5248 ActionList CollapsedOffloadActions; 5249 5250 ToolSelector TS(JA, *TC, C, isSaveTempsEnabled(), 5251 embedBitcodeInObject() && !isUsingLTO()); 5252 const Tool *T = TS.getTool(Inputs, CollapsedOffloadActions); 5253 5254 if (!T) 5255 return {InputInfo()}; 5256 5257 if (BuildingForOffloadDevice && 5258 A->getOffloadingDeviceKind() == Action::OFK_OpenMP) { 5259 if (TC->getTriple().isAMDGCN()) { 5260 // AMDGCN treats backend and assemble actions as no-op because 5261 // linker does not support object files. 5262 if (const BackendJobAction *BA = dyn_cast<BackendJobAction>(A)) { 5263 return BuildJobsForAction(C, *BA->input_begin(), TC, BoundArch, 5264 AtTopLevel, MultipleArchs, LinkingOutput, 5265 CachedResults, TargetDeviceOffloadKind); 5266 } 5267 5268 if (const AssembleJobAction *AA = dyn_cast<AssembleJobAction>(A)) { 5269 return BuildJobsForAction(C, *AA->input_begin(), TC, BoundArch, 5270 AtTopLevel, MultipleArchs, LinkingOutput, 5271 CachedResults, TargetDeviceOffloadKind); 5272 } 5273 } 5274 } 5275 5276 // If we've collapsed action list that contained OffloadAction we 5277 // need to build jobs for host/device-side inputs it may have held. 5278 for (const auto *OA : CollapsedOffloadActions) 5279 cast<OffloadAction>(OA)->doOnEachDependence( 5280 /*IsHostDependence=*/BuildingForOffloadDevice, 5281 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5282 OffloadDependencesInputInfo.append(BuildJobsForAction( 5283 C, DepA, DepTC, DepBoundArch, /* AtTopLevel */ false, 5284 /*MultipleArchs=*/!!DepBoundArch, LinkingOutput, CachedResults, 5285 DepA->getOffloadingDeviceKind())); 5286 }); 5287 5288 // Only use pipes when there is exactly one input. 5289 InputInfoList InputInfos; 5290 for (const Action *Input : Inputs) { 5291 // Treat dsymutil and verify sub-jobs as being at the top-level too, they 5292 // shouldn't get temporary output names. 5293 // FIXME: Clean this up. 5294 bool SubJobAtTopLevel = 5295 AtTopLevel && (isa<DsymutilJobAction>(A) || isa<VerifyJobAction>(A)); 5296 InputInfos.append(BuildJobsForAction( 5297 C, Input, TC, BoundArch, SubJobAtTopLevel, MultipleArchs, LinkingOutput, 5298 CachedResults, A->getOffloadingDeviceKind())); 5299 } 5300 5301 // Always use the first file input as the base input. 5302 const char *BaseInput = InputInfos[0].getBaseInput(); 5303 for (auto &Info : InputInfos) { 5304 if (Info.isFilename()) { 5305 BaseInput = Info.getBaseInput(); 5306 break; 5307 } 5308 } 5309 5310 // ... except dsymutil actions, which use their actual input as the base 5311 // input. 5312 if (JA->getType() == types::TY_dSYM) 5313 BaseInput = InputInfos[0].getFilename(); 5314 5315 // ... and in header module compilations, which use the module name. 5316 if (auto *ModuleJA = dyn_cast<HeaderModulePrecompileJobAction>(JA)) 5317 BaseInput = ModuleJA->getModuleName(); 5318 5319 // Append outputs of offload device jobs to the input list 5320 if (!OffloadDependencesInputInfo.empty()) 5321 InputInfos.append(OffloadDependencesInputInfo.begin(), 5322 OffloadDependencesInputInfo.end()); 5323 5324 // Set the effective triple of the toolchain for the duration of this job. 5325 llvm::Triple EffectiveTriple; 5326 const ToolChain &ToolTC = T->getToolChain(); 5327 const ArgList &Args = 5328 C.getArgsForToolChain(TC, BoundArch, A->getOffloadingDeviceKind()); 5329 if (InputInfos.size() != 1) { 5330 EffectiveTriple = llvm::Triple(ToolTC.ComputeEffectiveClangTriple(Args)); 5331 } else { 5332 // Pass along the input type if it can be unambiguously determined. 5333 EffectiveTriple = llvm::Triple( 5334 ToolTC.ComputeEffectiveClangTriple(Args, InputInfos[0].getType())); 5335 } 5336 RegisterEffectiveTriple TripleRAII(ToolTC, EffectiveTriple); 5337 5338 // Determine the place to write output to, if any. 5339 InputInfo Result; 5340 InputInfoList UnbundlingResults; 5341 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(JA)) { 5342 // If we have an unbundling job, we need to create results for all the 5343 // outputs. We also update the results cache so that other actions using 5344 // this unbundling action can get the right results. 5345 for (auto &UI : UA->getDependentActionsInfo()) { 5346 assert(UI.DependentOffloadKind != Action::OFK_None && 5347 "Unbundling with no offloading??"); 5348 5349 // Unbundling actions are never at the top level. When we generate the 5350 // offloading prefix, we also do that for the host file because the 5351 // unbundling action does not change the type of the output which can 5352 // cause a overwrite. 5353 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5354 UI.DependentOffloadKind, 5355 UI.DependentToolChain->getTriple().normalize(), 5356 /*CreatePrefixForHost=*/true); 5357 auto CurI = InputInfo( 5358 UA, 5359 GetNamedOutputPath(C, *UA, BaseInput, UI.DependentBoundArch, 5360 /*AtTopLevel=*/false, 5361 MultipleArchs || 5362 UI.DependentOffloadKind == Action::OFK_HIP, 5363 OffloadingPrefix), 5364 BaseInput); 5365 // Save the unbundling result. 5366 UnbundlingResults.push_back(CurI); 5367 5368 // Get the unique string identifier for this dependence and cache the 5369 // result. 5370 StringRef Arch; 5371 if (TargetDeviceOffloadKind == Action::OFK_HIP) { 5372 if (UI.DependentOffloadKind == Action::OFK_Host) 5373 Arch = StringRef(); 5374 else 5375 Arch = UI.DependentBoundArch; 5376 } else 5377 Arch = BoundArch; 5378 5379 CachedResults[{A, GetTriplePlusArchString(UI.DependentToolChain, Arch, 5380 UI.DependentOffloadKind)}] = { 5381 CurI}; 5382 } 5383 5384 // Now that we have all the results generated, select the one that should be 5385 // returned for the current depending action. 5386 std::pair<const Action *, std::string> ActionTC = { 5387 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5388 assert(CachedResults.find(ActionTC) != CachedResults.end() && 5389 "Result does not exist??"); 5390 Result = CachedResults[ActionTC].front(); 5391 } else if (JA->getType() == types::TY_Nothing) 5392 Result = {InputInfo(A, BaseInput)}; 5393 else { 5394 // We only have to generate a prefix for the host if this is not a top-level 5395 // action. 5396 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5397 A->getOffloadingDeviceKind(), TC->getTriple().normalize(), 5398 /*CreatePrefixForHost=*/!!A->getOffloadingHostActiveKinds() && 5399 !AtTopLevel); 5400 if (isa<OffloadWrapperJobAction>(JA)) { 5401 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5402 BaseInput = FinalOutput->getValue(); 5403 else 5404 BaseInput = getDefaultImageName(); 5405 BaseInput = 5406 C.getArgs().MakeArgString(std::string(BaseInput) + "-wrapper"); 5407 } 5408 Result = InputInfo(A, GetNamedOutputPath(C, *JA, BaseInput, BoundArch, 5409 AtTopLevel, MultipleArchs, 5410 OffloadingPrefix), 5411 BaseInput); 5412 } 5413 5414 if (CCCPrintBindings && !CCGenDiagnostics) { 5415 llvm::errs() << "# \"" << T->getToolChain().getTripleString() << '"' 5416 << " - \"" << T->getName() << "\", inputs: ["; 5417 for (unsigned i = 0, e = InputInfos.size(); i != e; ++i) { 5418 llvm::errs() << InputInfos[i].getAsString(); 5419 if (i + 1 != e) 5420 llvm::errs() << ", "; 5421 } 5422 if (UnbundlingResults.empty()) 5423 llvm::errs() << "], output: " << Result.getAsString() << "\n"; 5424 else { 5425 llvm::errs() << "], outputs: ["; 5426 for (unsigned i = 0, e = UnbundlingResults.size(); i != e; ++i) { 5427 llvm::errs() << UnbundlingResults[i].getAsString(); 5428 if (i + 1 != e) 5429 llvm::errs() << ", "; 5430 } 5431 llvm::errs() << "] \n"; 5432 } 5433 } else { 5434 if (UnbundlingResults.empty()) 5435 T->ConstructJob( 5436 C, *JA, Result, InputInfos, 5437 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5438 LinkingOutput); 5439 else 5440 T->ConstructJobMultipleOutputs( 5441 C, *JA, UnbundlingResults, InputInfos, 5442 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5443 LinkingOutput); 5444 } 5445 return {Result}; 5446 } 5447 5448 const char *Driver::getDefaultImageName() const { 5449 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 5450 return Target.isOSWindows() ? "a.exe" : "a.out"; 5451 } 5452 5453 /// Create output filename based on ArgValue, which could either be a 5454 /// full filename, filename without extension, or a directory. If ArgValue 5455 /// does not provide a filename, then use BaseName, and use the extension 5456 /// suitable for FileType. 5457 static const char *MakeCLOutputFilename(const ArgList &Args, StringRef ArgValue, 5458 StringRef BaseName, 5459 types::ID FileType) { 5460 SmallString<128> Filename = ArgValue; 5461 5462 if (ArgValue.empty()) { 5463 // If the argument is empty, output to BaseName in the current dir. 5464 Filename = BaseName; 5465 } else if (llvm::sys::path::is_separator(Filename.back())) { 5466 // If the argument is a directory, output to BaseName in that dir. 5467 llvm::sys::path::append(Filename, BaseName); 5468 } 5469 5470 if (!llvm::sys::path::has_extension(ArgValue)) { 5471 // If the argument didn't provide an extension, then set it. 5472 const char *Extension = types::getTypeTempSuffix(FileType, true); 5473 5474 if (FileType == types::TY_Image && 5475 Args.hasArg(options::OPT__SLASH_LD, options::OPT__SLASH_LDd)) { 5476 // The output file is a dll. 5477 Extension = "dll"; 5478 } 5479 5480 llvm::sys::path::replace_extension(Filename, Extension); 5481 } 5482 5483 return Args.MakeArgString(Filename.c_str()); 5484 } 5485 5486 static bool HasPreprocessOutput(const Action &JA) { 5487 if (isa<PreprocessJobAction>(JA)) 5488 return true; 5489 if (isa<OffloadAction>(JA) && isa<PreprocessJobAction>(JA.getInputs()[0])) 5490 return true; 5491 if (isa<OffloadBundlingJobAction>(JA) && 5492 HasPreprocessOutput(*(JA.getInputs()[0]))) 5493 return true; 5494 return false; 5495 } 5496 5497 const char *Driver::GetNamedOutputPath(Compilation &C, const JobAction &JA, 5498 const char *BaseInput, 5499 StringRef OrigBoundArch, bool AtTopLevel, 5500 bool MultipleArchs, 5501 StringRef OffloadingPrefix) const { 5502 std::string BoundArch = OrigBoundArch.str(); 5503 if (is_style_windows(llvm::sys::path::Style::native)) { 5504 // BoundArch may contains ':', which is invalid in file names on Windows, 5505 // therefore replace it with '%'. 5506 std::replace(BoundArch.begin(), BoundArch.end(), ':', '@'); 5507 } 5508 5509 llvm::PrettyStackTraceString CrashInfo("Computing output path"); 5510 // Output to a user requested destination? 5511 if (AtTopLevel && !isa<DsymutilJobAction>(JA) && !isa<VerifyJobAction>(JA)) { 5512 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5513 return C.addResultFile(FinalOutput->getValue(), &JA); 5514 } 5515 5516 // For /P, preprocess to file named after BaseInput. 5517 if (C.getArgs().hasArg(options::OPT__SLASH_P)) { 5518 assert(AtTopLevel && isa<PreprocessJobAction>(JA)); 5519 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5520 StringRef NameArg; 5521 if (Arg *A = C.getArgs().getLastArg(options::OPT__SLASH_Fi)) 5522 NameArg = A->getValue(); 5523 return C.addResultFile( 5524 MakeCLOutputFilename(C.getArgs(), NameArg, BaseName, types::TY_PP_C), 5525 &JA); 5526 } 5527 5528 // Default to writing to stdout? 5529 if (AtTopLevel && !CCGenDiagnostics && HasPreprocessOutput(JA)) { 5530 return "-"; 5531 } 5532 5533 if (JA.getType() == types::TY_ModuleFile && 5534 C.getArgs().getLastArg(options::OPT_module_file_info)) { 5535 return "-"; 5536 } 5537 5538 // Is this the assembly listing for /FA? 5539 if (JA.getType() == types::TY_PP_Asm && 5540 (C.getArgs().hasArg(options::OPT__SLASH_FA) || 5541 C.getArgs().hasArg(options::OPT__SLASH_Fa))) { 5542 // Use /Fa and the input filename to determine the asm file name. 5543 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5544 StringRef FaValue = C.getArgs().getLastArgValue(options::OPT__SLASH_Fa); 5545 return C.addResultFile( 5546 MakeCLOutputFilename(C.getArgs(), FaValue, BaseName, JA.getType()), 5547 &JA); 5548 } 5549 5550 // Output to a temporary file? 5551 if ((!AtTopLevel && !isSaveTempsEnabled() && 5552 !C.getArgs().hasArg(options::OPT__SLASH_Fo)) || 5553 CCGenDiagnostics) { 5554 StringRef Name = llvm::sys::path::filename(BaseInput); 5555 std::pair<StringRef, StringRef> Split = Name.split('.'); 5556 SmallString<128> TmpName; 5557 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5558 Arg *A = C.getArgs().getLastArg(options::OPT_fcrash_diagnostics_dir); 5559 if (CCGenDiagnostics && A) { 5560 SmallString<128> CrashDirectory(A->getValue()); 5561 if (!getVFS().exists(CrashDirectory)) 5562 llvm::sys::fs::create_directories(CrashDirectory); 5563 llvm::sys::path::append(CrashDirectory, Split.first); 5564 const char *Middle = Suffix ? "-%%%%%%." : "-%%%%%%"; 5565 std::error_code EC = llvm::sys::fs::createUniqueFile( 5566 CrashDirectory + Middle + Suffix, TmpName); 5567 if (EC) { 5568 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5569 return ""; 5570 } 5571 } else { 5572 if (MultipleArchs && !BoundArch.empty()) { 5573 TmpName = GetTemporaryDirectory(Split.first); 5574 llvm::sys::path::append(TmpName, 5575 Split.first + "-" + BoundArch + "." + Suffix); 5576 } else { 5577 TmpName = GetTemporaryPath(Split.first, Suffix); 5578 } 5579 } 5580 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5581 } 5582 5583 SmallString<128> BasePath(BaseInput); 5584 SmallString<128> ExternalPath(""); 5585 StringRef BaseName; 5586 5587 // Dsymutil actions should use the full path. 5588 if (isa<DsymutilJobAction>(JA) && C.getArgs().hasArg(options::OPT_dsym_dir)) { 5589 ExternalPath += C.getArgs().getLastArg(options::OPT_dsym_dir)->getValue(); 5590 // We use posix style here because the tests (specifically 5591 // darwin-dsymutil.c) demonstrate that posix style paths are acceptable 5592 // even on Windows and if we don't then the similar test covering this 5593 // fails. 5594 llvm::sys::path::append(ExternalPath, llvm::sys::path::Style::posix, 5595 llvm::sys::path::filename(BasePath)); 5596 BaseName = ExternalPath; 5597 } else if (isa<DsymutilJobAction>(JA) || isa<VerifyJobAction>(JA)) 5598 BaseName = BasePath; 5599 else 5600 BaseName = llvm::sys::path::filename(BasePath); 5601 5602 // Determine what the derived output name should be. 5603 const char *NamedOutput; 5604 5605 if ((JA.getType() == types::TY_Object || JA.getType() == types::TY_LTO_BC) && 5606 C.getArgs().hasArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o)) { 5607 // The /Fo or /o flag decides the object filename. 5608 StringRef Val = 5609 C.getArgs() 5610 .getLastArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o) 5611 ->getValue(); 5612 NamedOutput = 5613 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5614 } else if (JA.getType() == types::TY_Image && 5615 C.getArgs().hasArg(options::OPT__SLASH_Fe, 5616 options::OPT__SLASH_o)) { 5617 // The /Fe or /o flag names the linked file. 5618 StringRef Val = 5619 C.getArgs() 5620 .getLastArg(options::OPT__SLASH_Fe, options::OPT__SLASH_o) 5621 ->getValue(); 5622 NamedOutput = 5623 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Image); 5624 } else if (JA.getType() == types::TY_Image) { 5625 if (IsCLMode()) { 5626 // clang-cl uses BaseName for the executable name. 5627 NamedOutput = 5628 MakeCLOutputFilename(C.getArgs(), "", BaseName, types::TY_Image); 5629 } else { 5630 SmallString<128> Output(getDefaultImageName()); 5631 // HIP image for device compilation with -fno-gpu-rdc is per compilation 5632 // unit. 5633 bool IsHIPNoRDC = JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5634 !C.getArgs().hasFlag(options::OPT_fgpu_rdc, 5635 options::OPT_fno_gpu_rdc, false); 5636 if (IsHIPNoRDC) { 5637 Output = BaseName; 5638 llvm::sys::path::replace_extension(Output, ""); 5639 } 5640 Output += OffloadingPrefix; 5641 if (MultipleArchs && !BoundArch.empty()) { 5642 Output += "-"; 5643 Output.append(BoundArch); 5644 } 5645 if (IsHIPNoRDC) 5646 Output += ".out"; 5647 NamedOutput = C.getArgs().MakeArgString(Output.c_str()); 5648 } 5649 } else if (JA.getType() == types::TY_PCH && IsCLMode()) { 5650 NamedOutput = C.getArgs().MakeArgString(GetClPchPath(C, BaseName)); 5651 } else { 5652 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5653 assert(Suffix && "All types used for output should have a suffix."); 5654 5655 std::string::size_type End = std::string::npos; 5656 if (!types::appendSuffixForType(JA.getType())) 5657 End = BaseName.rfind('.'); 5658 SmallString<128> Suffixed(BaseName.substr(0, End)); 5659 Suffixed += OffloadingPrefix; 5660 if (MultipleArchs && !BoundArch.empty()) { 5661 Suffixed += "-"; 5662 Suffixed.append(BoundArch); 5663 } 5664 // When using both -save-temps and -emit-llvm, use a ".tmp.bc" suffix for 5665 // the unoptimized bitcode so that it does not get overwritten by the ".bc" 5666 // optimized bitcode output. 5667 auto IsHIPRDCInCompilePhase = [](const JobAction &JA, 5668 const llvm::opt::DerivedArgList &Args) { 5669 // The relocatable compilation in HIP implies -emit-llvm. Similarly, use a 5670 // ".tmp.bc" suffix for the unoptimized bitcode (generated in the compile 5671 // phase.) 5672 return isa<CompileJobAction>(JA) && 5673 JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5674 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 5675 false); 5676 }; 5677 if (!AtTopLevel && JA.getType() == types::TY_LLVM_BC && 5678 (C.getArgs().hasArg(options::OPT_emit_llvm) || 5679 IsHIPRDCInCompilePhase(JA, C.getArgs()))) 5680 Suffixed += ".tmp"; 5681 Suffixed += '.'; 5682 Suffixed += Suffix; 5683 NamedOutput = C.getArgs().MakeArgString(Suffixed.c_str()); 5684 } 5685 5686 // Prepend object file path if -save-temps=obj 5687 if (!AtTopLevel && isSaveTempsObj() && C.getArgs().hasArg(options::OPT_o) && 5688 JA.getType() != types::TY_PCH) { 5689 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 5690 SmallString<128> TempPath(FinalOutput->getValue()); 5691 llvm::sys::path::remove_filename(TempPath); 5692 StringRef OutputFileName = llvm::sys::path::filename(NamedOutput); 5693 llvm::sys::path::append(TempPath, OutputFileName); 5694 NamedOutput = C.getArgs().MakeArgString(TempPath.c_str()); 5695 } 5696 5697 // If we're saving temps and the temp file conflicts with the input file, 5698 // then avoid overwriting input file. 5699 if (!AtTopLevel && isSaveTempsEnabled() && NamedOutput == BaseName) { 5700 bool SameFile = false; 5701 SmallString<256> Result; 5702 llvm::sys::fs::current_path(Result); 5703 llvm::sys::path::append(Result, BaseName); 5704 llvm::sys::fs::equivalent(BaseInput, Result.c_str(), SameFile); 5705 // Must share the same path to conflict. 5706 if (SameFile) { 5707 StringRef Name = llvm::sys::path::filename(BaseInput); 5708 std::pair<StringRef, StringRef> Split = Name.split('.'); 5709 std::string TmpName = GetTemporaryPath( 5710 Split.first, types::getTypeTempSuffix(JA.getType(), IsCLMode())); 5711 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5712 } 5713 } 5714 5715 // As an annoying special case, PCH generation doesn't strip the pathname. 5716 if (JA.getType() == types::TY_PCH && !IsCLMode()) { 5717 llvm::sys::path::remove_filename(BasePath); 5718 if (BasePath.empty()) 5719 BasePath = NamedOutput; 5720 else 5721 llvm::sys::path::append(BasePath, NamedOutput); 5722 return C.addResultFile(C.getArgs().MakeArgString(BasePath.c_str()), &JA); 5723 } else { 5724 return C.addResultFile(NamedOutput, &JA); 5725 } 5726 } 5727 5728 std::string Driver::GetFilePath(StringRef Name, const ToolChain &TC) const { 5729 // Search for Name in a list of paths. 5730 auto SearchPaths = [&](const llvm::SmallVectorImpl<std::string> &P) 5731 -> llvm::Optional<std::string> { 5732 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5733 // attempting to use this prefix when looking for file paths. 5734 for (const auto &Dir : P) { 5735 if (Dir.empty()) 5736 continue; 5737 SmallString<128> P(Dir[0] == '=' ? SysRoot + Dir.substr(1) : Dir); 5738 llvm::sys::path::append(P, Name); 5739 if (llvm::sys::fs::exists(Twine(P))) 5740 return std::string(P); 5741 } 5742 return None; 5743 }; 5744 5745 if (auto P = SearchPaths(PrefixDirs)) 5746 return *P; 5747 5748 SmallString<128> R(ResourceDir); 5749 llvm::sys::path::append(R, Name); 5750 if (llvm::sys::fs::exists(Twine(R))) 5751 return std::string(R.str()); 5752 5753 SmallString<128> P(TC.getCompilerRTPath()); 5754 llvm::sys::path::append(P, Name); 5755 if (llvm::sys::fs::exists(Twine(P))) 5756 return std::string(P.str()); 5757 5758 SmallString<128> D(Dir); 5759 llvm::sys::path::append(D, "..", Name); 5760 if (llvm::sys::fs::exists(Twine(D))) 5761 return std::string(D.str()); 5762 5763 if (auto P = SearchPaths(TC.getLibraryPaths())) 5764 return *P; 5765 5766 if (auto P = SearchPaths(TC.getFilePaths())) 5767 return *P; 5768 5769 return std::string(Name); 5770 } 5771 5772 void Driver::generatePrefixedToolNames( 5773 StringRef Tool, const ToolChain &TC, 5774 SmallVectorImpl<std::string> &Names) const { 5775 // FIXME: Needs a better variable than TargetTriple 5776 Names.emplace_back((TargetTriple + "-" + Tool).str()); 5777 Names.emplace_back(Tool); 5778 } 5779 5780 static bool ScanDirForExecutable(SmallString<128> &Dir, StringRef Name) { 5781 llvm::sys::path::append(Dir, Name); 5782 if (llvm::sys::fs::can_execute(Twine(Dir))) 5783 return true; 5784 llvm::sys::path::remove_filename(Dir); 5785 return false; 5786 } 5787 5788 std::string Driver::GetProgramPath(StringRef Name, const ToolChain &TC) const { 5789 SmallVector<std::string, 2> TargetSpecificExecutables; 5790 generatePrefixedToolNames(Name, TC, TargetSpecificExecutables); 5791 5792 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5793 // attempting to use this prefix when looking for program paths. 5794 for (const auto &PrefixDir : PrefixDirs) { 5795 if (llvm::sys::fs::is_directory(PrefixDir)) { 5796 SmallString<128> P(PrefixDir); 5797 if (ScanDirForExecutable(P, Name)) 5798 return std::string(P.str()); 5799 } else { 5800 SmallString<128> P((PrefixDir + Name).str()); 5801 if (llvm::sys::fs::can_execute(Twine(P))) 5802 return std::string(P.str()); 5803 } 5804 } 5805 5806 const ToolChain::path_list &List = TC.getProgramPaths(); 5807 for (const auto &TargetSpecificExecutable : TargetSpecificExecutables) { 5808 // For each possible name of the tool look for it in 5809 // program paths first, then the path. 5810 // Higher priority names will be first, meaning that 5811 // a higher priority name in the path will be found 5812 // instead of a lower priority name in the program path. 5813 // E.g. <triple>-gcc on the path will be found instead 5814 // of gcc in the program path 5815 for (const auto &Path : List) { 5816 SmallString<128> P(Path); 5817 if (ScanDirForExecutable(P, TargetSpecificExecutable)) 5818 return std::string(P.str()); 5819 } 5820 5821 // Fall back to the path 5822 if (llvm::ErrorOr<std::string> P = 5823 llvm::sys::findProgramByName(TargetSpecificExecutable)) 5824 return *P; 5825 } 5826 5827 return std::string(Name); 5828 } 5829 5830 std::string Driver::GetTemporaryPath(StringRef Prefix, StringRef Suffix) const { 5831 SmallString<128> Path; 5832 std::error_code EC = llvm::sys::fs::createTemporaryFile(Prefix, Suffix, Path); 5833 if (EC) { 5834 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5835 return ""; 5836 } 5837 5838 return std::string(Path.str()); 5839 } 5840 5841 std::string Driver::GetTemporaryDirectory(StringRef Prefix) const { 5842 SmallString<128> Path; 5843 std::error_code EC = llvm::sys::fs::createUniqueDirectory(Prefix, Path); 5844 if (EC) { 5845 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5846 return ""; 5847 } 5848 5849 return std::string(Path.str()); 5850 } 5851 5852 std::string Driver::GetClPchPath(Compilation &C, StringRef BaseName) const { 5853 SmallString<128> Output; 5854 if (Arg *FpArg = C.getArgs().getLastArg(options::OPT__SLASH_Fp)) { 5855 // FIXME: If anybody needs it, implement this obscure rule: 5856 // "If you specify a directory without a file name, the default file name 5857 // is VCx0.pch., where x is the major version of Visual C++ in use." 5858 Output = FpArg->getValue(); 5859 5860 // "If you do not specify an extension as part of the path name, an 5861 // extension of .pch is assumed. " 5862 if (!llvm::sys::path::has_extension(Output)) 5863 Output += ".pch"; 5864 } else { 5865 if (Arg *YcArg = C.getArgs().getLastArg(options::OPT__SLASH_Yc)) 5866 Output = YcArg->getValue(); 5867 if (Output.empty()) 5868 Output = BaseName; 5869 llvm::sys::path::replace_extension(Output, ".pch"); 5870 } 5871 return std::string(Output.str()); 5872 } 5873 5874 const ToolChain &Driver::getToolChain(const ArgList &Args, 5875 const llvm::Triple &Target) const { 5876 5877 auto &TC = ToolChains[Target.str()]; 5878 if (!TC) { 5879 switch (Target.getOS()) { 5880 case llvm::Triple::AIX: 5881 TC = std::make_unique<toolchains::AIX>(*this, Target, Args); 5882 break; 5883 case llvm::Triple::Haiku: 5884 TC = std::make_unique<toolchains::Haiku>(*this, Target, Args); 5885 break; 5886 case llvm::Triple::Ananas: 5887 TC = std::make_unique<toolchains::Ananas>(*this, Target, Args); 5888 break; 5889 case llvm::Triple::CloudABI: 5890 TC = std::make_unique<toolchains::CloudABI>(*this, Target, Args); 5891 break; 5892 case llvm::Triple::Darwin: 5893 case llvm::Triple::MacOSX: 5894 case llvm::Triple::IOS: 5895 case llvm::Triple::TvOS: 5896 case llvm::Triple::WatchOS: 5897 case llvm::Triple::DriverKit: 5898 TC = std::make_unique<toolchains::DarwinClang>(*this, Target, Args); 5899 break; 5900 case llvm::Triple::DragonFly: 5901 TC = std::make_unique<toolchains::DragonFly>(*this, Target, Args); 5902 break; 5903 case llvm::Triple::OpenBSD: 5904 TC = std::make_unique<toolchains::OpenBSD>(*this, Target, Args); 5905 break; 5906 case llvm::Triple::NetBSD: 5907 TC = std::make_unique<toolchains::NetBSD>(*this, Target, Args); 5908 break; 5909 case llvm::Triple::FreeBSD: 5910 if (Target.isPPC()) 5911 TC = std::make_unique<toolchains::PPCFreeBSDToolChain>(*this, Target, 5912 Args); 5913 else 5914 TC = std::make_unique<toolchains::FreeBSD>(*this, Target, Args); 5915 break; 5916 case llvm::Triple::Minix: 5917 TC = std::make_unique<toolchains::Minix>(*this, Target, Args); 5918 break; 5919 case llvm::Triple::Linux: 5920 case llvm::Triple::ELFIAMCU: 5921 if (Target.getArch() == llvm::Triple::hexagon) 5922 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 5923 Args); 5924 else if ((Target.getVendor() == llvm::Triple::MipsTechnologies) && 5925 !Target.hasEnvironment()) 5926 TC = std::make_unique<toolchains::MipsLLVMToolChain>(*this, Target, 5927 Args); 5928 else if (Target.isPPC()) 5929 TC = std::make_unique<toolchains::PPCLinuxToolChain>(*this, Target, 5930 Args); 5931 else if (Target.getArch() == llvm::Triple::ve) 5932 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 5933 5934 else 5935 TC = std::make_unique<toolchains::Linux>(*this, Target, Args); 5936 break; 5937 case llvm::Triple::NaCl: 5938 TC = std::make_unique<toolchains::NaClToolChain>(*this, Target, Args); 5939 break; 5940 case llvm::Triple::Fuchsia: 5941 TC = std::make_unique<toolchains::Fuchsia>(*this, Target, Args); 5942 break; 5943 case llvm::Triple::Solaris: 5944 TC = std::make_unique<toolchains::Solaris>(*this, Target, Args); 5945 break; 5946 case llvm::Triple::AMDHSA: 5947 TC = std::make_unique<toolchains::ROCMToolChain>(*this, Target, Args); 5948 break; 5949 case llvm::Triple::AMDPAL: 5950 case llvm::Triple::Mesa3D: 5951 TC = std::make_unique<toolchains::AMDGPUToolChain>(*this, Target, Args); 5952 break; 5953 case llvm::Triple::Win32: 5954 switch (Target.getEnvironment()) { 5955 default: 5956 if (Target.isOSBinFormatELF()) 5957 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 5958 else if (Target.isOSBinFormatMachO()) 5959 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 5960 else 5961 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 5962 break; 5963 case llvm::Triple::GNU: 5964 TC = std::make_unique<toolchains::MinGW>(*this, Target, Args); 5965 break; 5966 case llvm::Triple::Itanium: 5967 TC = std::make_unique<toolchains::CrossWindowsToolChain>(*this, Target, 5968 Args); 5969 break; 5970 case llvm::Triple::MSVC: 5971 case llvm::Triple::UnknownEnvironment: 5972 if (Args.getLastArgValue(options::OPT_fuse_ld_EQ) 5973 .startswith_insensitive("bfd")) 5974 TC = std::make_unique<toolchains::CrossWindowsToolChain>( 5975 *this, Target, Args); 5976 else 5977 TC = 5978 std::make_unique<toolchains::MSVCToolChain>(*this, Target, Args); 5979 break; 5980 } 5981 break; 5982 case llvm::Triple::PS4: 5983 TC = std::make_unique<toolchains::PS4CPU>(*this, Target, Args); 5984 break; 5985 case llvm::Triple::PS5: 5986 TC = std::make_unique<toolchains::PS5CPU>(*this, Target, Args); 5987 break; 5988 case llvm::Triple::Contiki: 5989 TC = std::make_unique<toolchains::Contiki>(*this, Target, Args); 5990 break; 5991 case llvm::Triple::Hurd: 5992 TC = std::make_unique<toolchains::Hurd>(*this, Target, Args); 5993 break; 5994 case llvm::Triple::ZOS: 5995 TC = std::make_unique<toolchains::ZOS>(*this, Target, Args); 5996 break; 5997 case llvm::Triple::ShaderModel: 5998 TC = std::make_unique<toolchains::HLSLToolChain>(*this, Target, Args); 5999 break; 6000 default: 6001 // Of these targets, Hexagon is the only one that might have 6002 // an OS of Linux, in which case it got handled above already. 6003 switch (Target.getArch()) { 6004 case llvm::Triple::tce: 6005 TC = std::make_unique<toolchains::TCEToolChain>(*this, Target, Args); 6006 break; 6007 case llvm::Triple::tcele: 6008 TC = std::make_unique<toolchains::TCELEToolChain>(*this, Target, Args); 6009 break; 6010 case llvm::Triple::hexagon: 6011 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 6012 Args); 6013 break; 6014 case llvm::Triple::lanai: 6015 TC = std::make_unique<toolchains::LanaiToolChain>(*this, Target, Args); 6016 break; 6017 case llvm::Triple::xcore: 6018 TC = std::make_unique<toolchains::XCoreToolChain>(*this, Target, Args); 6019 break; 6020 case llvm::Triple::wasm32: 6021 case llvm::Triple::wasm64: 6022 TC = std::make_unique<toolchains::WebAssembly>(*this, Target, Args); 6023 break; 6024 case llvm::Triple::avr: 6025 TC = std::make_unique<toolchains::AVRToolChain>(*this, Target, Args); 6026 break; 6027 case llvm::Triple::msp430: 6028 TC = 6029 std::make_unique<toolchains::MSP430ToolChain>(*this, Target, Args); 6030 break; 6031 case llvm::Triple::riscv32: 6032 case llvm::Triple::riscv64: 6033 if (toolchains::RISCVToolChain::hasGCCToolchain(*this, Args)) 6034 TC = 6035 std::make_unique<toolchains::RISCVToolChain>(*this, Target, Args); 6036 else 6037 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6038 break; 6039 case llvm::Triple::ve: 6040 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6041 break; 6042 case llvm::Triple::spirv32: 6043 case llvm::Triple::spirv64: 6044 TC = std::make_unique<toolchains::SPIRVToolChain>(*this, Target, Args); 6045 break; 6046 case llvm::Triple::csky: 6047 TC = std::make_unique<toolchains::CSKYToolChain>(*this, Target, Args); 6048 break; 6049 default: 6050 if (Target.getVendor() == llvm::Triple::Myriad) 6051 TC = std::make_unique<toolchains::MyriadToolChain>(*this, Target, 6052 Args); 6053 else if (toolchains::BareMetal::handlesTarget(Target)) 6054 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6055 else if (Target.isOSBinFormatELF()) 6056 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6057 else if (Target.isOSBinFormatMachO()) 6058 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6059 else 6060 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6061 } 6062 } 6063 } 6064 6065 // Intentionally omitted from the switch above: llvm::Triple::CUDA. CUDA 6066 // compiles always need two toolchains, the CUDA toolchain and the host 6067 // toolchain. So the only valid way to create a CUDA toolchain is via 6068 // CreateOffloadingDeviceToolChains. 6069 6070 return *TC; 6071 } 6072 6073 const ToolChain &Driver::getOffloadingDeviceToolChain( 6074 const ArgList &Args, const llvm::Triple &Target, const ToolChain &HostTC, 6075 const Action::OffloadKind &TargetDeviceOffloadKind) const { 6076 // Use device / host triples as the key into the ToolChains map because the 6077 // device ToolChain we create depends on both. 6078 auto &TC = ToolChains[Target.str() + "/" + HostTC.getTriple().str()]; 6079 if (!TC) { 6080 // Categorized by offload kind > arch rather than OS > arch like 6081 // the normal getToolChain call, as it seems a reasonable way to categorize 6082 // things. 6083 switch (TargetDeviceOffloadKind) { 6084 case Action::OFK_HIP: { 6085 if (Target.getArch() == llvm::Triple::amdgcn && 6086 Target.getVendor() == llvm::Triple::AMD && 6087 Target.getOS() == llvm::Triple::AMDHSA) 6088 TC = std::make_unique<toolchains::HIPAMDToolChain>(*this, Target, 6089 HostTC, Args); 6090 else if (Target.getArch() == llvm::Triple::spirv64 && 6091 Target.getVendor() == llvm::Triple::UnknownVendor && 6092 Target.getOS() == llvm::Triple::UnknownOS) 6093 TC = std::make_unique<toolchains::HIPSPVToolChain>(*this, Target, 6094 HostTC, Args); 6095 break; 6096 } 6097 default: 6098 break; 6099 } 6100 } 6101 6102 return *TC; 6103 } 6104 6105 bool Driver::ShouldUseClangCompiler(const JobAction &JA) const { 6106 // Say "no" if there is not exactly one input of a type clang understands. 6107 if (JA.size() != 1 || 6108 !types::isAcceptedByClang((*JA.input_begin())->getType())) 6109 return false; 6110 6111 // And say "no" if this is not a kind of action clang understands. 6112 if (!isa<PreprocessJobAction>(JA) && !isa<PrecompileJobAction>(JA) && 6113 !isa<CompileJobAction>(JA) && !isa<BackendJobAction>(JA) && 6114 !isa<ExtractAPIJobAction>(JA)) 6115 return false; 6116 6117 return true; 6118 } 6119 6120 bool Driver::ShouldUseFlangCompiler(const JobAction &JA) const { 6121 // Say "no" if there is not exactly one input of a type flang understands. 6122 if (JA.size() != 1 || 6123 !types::isAcceptedByFlang((*JA.input_begin())->getType())) 6124 return false; 6125 6126 // And say "no" if this is not a kind of action flang understands. 6127 if (!isa<PreprocessJobAction>(JA) && !isa<CompileJobAction>(JA) && 6128 !isa<BackendJobAction>(JA)) 6129 return false; 6130 6131 return true; 6132 } 6133 6134 bool Driver::ShouldEmitStaticLibrary(const ArgList &Args) const { 6135 // Only emit static library if the flag is set explicitly. 6136 if (Args.hasArg(options::OPT_emit_static_lib)) 6137 return true; 6138 return false; 6139 } 6140 6141 /// GetReleaseVersion - Parse (([0-9]+)(.([0-9]+)(.([0-9]+)?))?)? and return the 6142 /// grouped values as integers. Numbers which are not provided are set to 0. 6143 /// 6144 /// \return True if the entire string was parsed (9.2), or all groups were 6145 /// parsed (10.3.5extrastuff). 6146 bool Driver::GetReleaseVersion(StringRef Str, unsigned &Major, unsigned &Minor, 6147 unsigned &Micro, bool &HadExtra) { 6148 HadExtra = false; 6149 6150 Major = Minor = Micro = 0; 6151 if (Str.empty()) 6152 return false; 6153 6154 if (Str.consumeInteger(10, Major)) 6155 return false; 6156 if (Str.empty()) 6157 return true; 6158 if (Str[0] != '.') 6159 return false; 6160 6161 Str = Str.drop_front(1); 6162 6163 if (Str.consumeInteger(10, Minor)) 6164 return false; 6165 if (Str.empty()) 6166 return true; 6167 if (Str[0] != '.') 6168 return false; 6169 Str = Str.drop_front(1); 6170 6171 if (Str.consumeInteger(10, Micro)) 6172 return false; 6173 if (!Str.empty()) 6174 HadExtra = true; 6175 return true; 6176 } 6177 6178 /// Parse digits from a string \p Str and fulfill \p Digits with 6179 /// the parsed numbers. This method assumes that the max number of 6180 /// digits to look for is equal to Digits.size(). 6181 /// 6182 /// \return True if the entire string was parsed and there are 6183 /// no extra characters remaining at the end. 6184 bool Driver::GetReleaseVersion(StringRef Str, 6185 MutableArrayRef<unsigned> Digits) { 6186 if (Str.empty()) 6187 return false; 6188 6189 unsigned CurDigit = 0; 6190 while (CurDigit < Digits.size()) { 6191 unsigned Digit; 6192 if (Str.consumeInteger(10, Digit)) 6193 return false; 6194 Digits[CurDigit] = Digit; 6195 if (Str.empty()) 6196 return true; 6197 if (Str[0] != '.') 6198 return false; 6199 Str = Str.drop_front(1); 6200 CurDigit++; 6201 } 6202 6203 // More digits than requested, bail out... 6204 return false; 6205 } 6206 6207 std::pair<unsigned, unsigned> 6208 Driver::getIncludeExcludeOptionFlagMasks(bool IsClCompatMode) const { 6209 unsigned IncludedFlagsBitmask = 0; 6210 unsigned ExcludedFlagsBitmask = options::NoDriverOption; 6211 6212 if (IsClCompatMode) { 6213 // Include CL and Core options. 6214 IncludedFlagsBitmask |= options::CLOption; 6215 IncludedFlagsBitmask |= options::CoreOption; 6216 } else { 6217 ExcludedFlagsBitmask |= options::CLOption; 6218 } 6219 if (IsDXCMode()) { 6220 // Include DXC and Core options. 6221 IncludedFlagsBitmask |= options::DXCOption; 6222 IncludedFlagsBitmask |= options::CoreOption; 6223 } else { 6224 ExcludedFlagsBitmask |= options::DXCOption; 6225 } 6226 return std::make_pair(IncludedFlagsBitmask, ExcludedFlagsBitmask); 6227 } 6228 6229 bool clang::driver::isOptimizationLevelFast(const ArgList &Args) { 6230 return Args.hasFlag(options::OPT_Ofast, options::OPT_O_Group, false); 6231 } 6232 6233 bool clang::driver::willEmitRemarks(const ArgList &Args) { 6234 // -fsave-optimization-record enables it. 6235 if (Args.hasFlag(options::OPT_fsave_optimization_record, 6236 options::OPT_fno_save_optimization_record, false)) 6237 return true; 6238 6239 // -fsave-optimization-record=<format> enables it as well. 6240 if (Args.hasFlag(options::OPT_fsave_optimization_record_EQ, 6241 options::OPT_fno_save_optimization_record, false)) 6242 return true; 6243 6244 // -foptimization-record-file alone enables it too. 6245 if (Args.hasFlag(options::OPT_foptimization_record_file_EQ, 6246 options::OPT_fno_save_optimization_record, false)) 6247 return true; 6248 6249 // -foptimization-record-passes alone enables it too. 6250 if (Args.hasFlag(options::OPT_foptimization_record_passes_EQ, 6251 options::OPT_fno_save_optimization_record, false)) 6252 return true; 6253 return false; 6254 } 6255 6256 llvm::StringRef clang::driver::getDriverMode(StringRef ProgName, 6257 ArrayRef<const char *> Args) { 6258 static const std::string OptName = 6259 getDriverOptTable().getOption(options::OPT_driver_mode).getPrefixedName(); 6260 llvm::StringRef Opt; 6261 for (StringRef Arg : Args) { 6262 if (!Arg.startswith(OptName)) 6263 continue; 6264 Opt = Arg; 6265 } 6266 if (Opt.empty()) 6267 Opt = ToolChain::getTargetAndModeFromProgramName(ProgName).DriverMode; 6268 return Opt.consume_front(OptName) ? Opt : ""; 6269 } 6270 6271 bool driver::IsClangCL(StringRef DriverMode) { return DriverMode.equals("cl"); } 6272