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 Offload(OffloadHostDevice), CXX20HeaderType(HeaderMode_None), 196 ModulesModeCXX20(false), LTOMode(LTOK_None), 197 ClangExecutable(ClangExecutable), SysRoot(DEFAULT_SYSROOT), 198 DriverTitle(Title), CCCPrintBindings(false), CCPrintOptions(false), 199 CCPrintHeaders(false), CCLogDiagnostics(false), CCGenDiagnostics(false), 200 CCPrintProcessStats(false), TargetTriple(TargetTriple), Saver(Alloc), 201 CheckInputsExist(true), ProbePrecompiled(true), 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 = static_cast<std::string>(CfgDir); 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 = static_cast<std::string>(CfgDir); 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 1221 // Process -fproc-stat-report options. 1222 if (const Arg *A = Args.getLastArg(options::OPT_fproc_stat_report_EQ)) { 1223 CCPrintProcessStats = true; 1224 CCPrintStatReportFilename = A->getValue(); 1225 } 1226 if (Args.hasArg(options::OPT_fproc_stat_report)) 1227 CCPrintProcessStats = true; 1228 1229 // FIXME: TargetTriple is used by the target-prefixed calls to as/ld 1230 // and getToolChain is const. 1231 if (IsCLMode()) { 1232 // clang-cl targets MSVC-style Win32. 1233 llvm::Triple T(TargetTriple); 1234 T.setOS(llvm::Triple::Win32); 1235 T.setVendor(llvm::Triple::PC); 1236 T.setEnvironment(llvm::Triple::MSVC); 1237 T.setObjectFormat(llvm::Triple::COFF); 1238 TargetTriple = T.str(); 1239 } else if (IsDXCMode()) { 1240 // Build TargetTriple from target_profile option for clang-dxc. 1241 if (const Arg *A = Args.getLastArg(options::OPT_target_profile)) { 1242 StringRef TargetProfile = A->getValue(); 1243 if (auto Triple = 1244 toolchains::HLSLToolChain::parseTargetProfile(TargetProfile)) 1245 TargetTriple = Triple.getValue(); 1246 else 1247 Diag(diag::err_drv_invalid_directx_shader_module) << TargetProfile; 1248 1249 A->claim(); 1250 } else { 1251 Diag(diag::err_drv_dxc_missing_target_profile); 1252 } 1253 } 1254 1255 if (const Arg *A = Args.getLastArg(options::OPT_target)) 1256 TargetTriple = A->getValue(); 1257 if (const Arg *A = Args.getLastArg(options::OPT_ccc_install_dir)) 1258 Dir = InstalledDir = A->getValue(); 1259 for (const Arg *A : Args.filtered(options::OPT_B)) { 1260 A->claim(); 1261 PrefixDirs.push_back(A->getValue(0)); 1262 } 1263 if (Optional<std::string> CompilerPathValue = 1264 llvm::sys::Process::GetEnv("COMPILER_PATH")) { 1265 StringRef CompilerPath = *CompilerPathValue; 1266 while (!CompilerPath.empty()) { 1267 std::pair<StringRef, StringRef> Split = 1268 CompilerPath.split(llvm::sys::EnvPathSeparator); 1269 PrefixDirs.push_back(std::string(Split.first)); 1270 CompilerPath = Split.second; 1271 } 1272 } 1273 if (const Arg *A = Args.getLastArg(options::OPT__sysroot_EQ)) 1274 SysRoot = A->getValue(); 1275 if (const Arg *A = Args.getLastArg(options::OPT__dyld_prefix_EQ)) 1276 DyldPrefix = A->getValue(); 1277 1278 if (const Arg *A = Args.getLastArg(options::OPT_resource_dir)) 1279 ResourceDir = A->getValue(); 1280 1281 if (const Arg *A = Args.getLastArg(options::OPT_save_temps_EQ)) { 1282 SaveTemps = llvm::StringSwitch<SaveTempsMode>(A->getValue()) 1283 .Case("cwd", SaveTempsCwd) 1284 .Case("obj", SaveTempsObj) 1285 .Default(SaveTempsCwd); 1286 } 1287 1288 if (const Arg *A = Args.getLastArg(options::OPT_offload_host_only, 1289 options::OPT_offload_device_only, 1290 options::OPT_offload_host_device)) { 1291 if (A->getOption().matches(options::OPT_offload_host_only)) 1292 Offload = OffloadHost; 1293 else if (A->getOption().matches(options::OPT_offload_device_only)) 1294 Offload = OffloadDevice; 1295 else 1296 Offload = OffloadHostDevice; 1297 } 1298 1299 setLTOMode(Args); 1300 1301 // Process -fembed-bitcode= flags. 1302 if (Arg *A = Args.getLastArg(options::OPT_fembed_bitcode_EQ)) { 1303 StringRef Name = A->getValue(); 1304 unsigned Model = llvm::StringSwitch<unsigned>(Name) 1305 .Case("off", EmbedNone) 1306 .Case("all", EmbedBitcode) 1307 .Case("bitcode", EmbedBitcode) 1308 .Case("marker", EmbedMarker) 1309 .Default(~0U); 1310 if (Model == ~0U) { 1311 Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) 1312 << Name; 1313 } else 1314 BitcodeEmbed = static_cast<BitcodeEmbedMode>(Model); 1315 } 1316 1317 // Setting up the jobs for some precompile cases depends on whether we are 1318 // treating them as PCH, implicit modules or C++20 ones. 1319 // TODO: inferring the mode like this seems fragile (it meets the objective 1320 // of not requiring anything new for operation, however). 1321 const Arg *Std = Args.getLastArg(options::OPT_std_EQ); 1322 ModulesModeCXX20 = 1323 !Args.hasArg(options::OPT_fmodules) && Std && 1324 (Std->containsValue("c++20") || Std->containsValue("c++2b") || 1325 Std->containsValue("c++2a") || Std->containsValue("c++latest")); 1326 1327 // Process -fmodule-header{=} flags. 1328 if (Arg *A = Args.getLastArg(options::OPT_fmodule_header_EQ, 1329 options::OPT_fmodule_header)) { 1330 // These flags force C++20 handling of headers. 1331 ModulesModeCXX20 = true; 1332 if (A->getOption().matches(options::OPT_fmodule_header)) 1333 CXX20HeaderType = HeaderMode_Default; 1334 else { 1335 StringRef ArgName = A->getValue(); 1336 unsigned Kind = llvm::StringSwitch<unsigned>(ArgName) 1337 .Case("user", HeaderMode_User) 1338 .Case("system", HeaderMode_System) 1339 .Default(~0U); 1340 if (Kind == ~0U) { 1341 Diags.Report(diag::err_drv_invalid_value) 1342 << A->getAsString(Args) << ArgName; 1343 } else 1344 CXX20HeaderType = static_cast<ModuleHeaderMode>(Kind); 1345 } 1346 } 1347 1348 std::unique_ptr<llvm::opt::InputArgList> UArgs = 1349 std::make_unique<InputArgList>(std::move(Args)); 1350 1351 // Perform the default argument translations. 1352 DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs); 1353 1354 // Owned by the host. 1355 const ToolChain &TC = getToolChain( 1356 *UArgs, computeTargetTriple(*this, TargetTriple, *UArgs)); 1357 1358 // The compilation takes ownership of Args. 1359 Compilation *C = new Compilation(*this, TC, UArgs.release(), TranslatedArgs, 1360 ContainsError); 1361 1362 if (!HandleImmediateArgs(*C)) 1363 return C; 1364 1365 // Construct the list of inputs. 1366 InputList Inputs; 1367 BuildInputs(C->getDefaultToolChain(), *TranslatedArgs, Inputs); 1368 1369 // Populate the tool chains for the offloading devices, if any. 1370 CreateOffloadingDeviceToolChains(*C, Inputs); 1371 1372 // Construct the list of abstract actions to perform for this compilation. On 1373 // MachO targets this uses the driver-driver and universal actions. 1374 if (TC.getTriple().isOSBinFormatMachO()) 1375 BuildUniversalActions(*C, C->getDefaultToolChain(), Inputs); 1376 else 1377 BuildActions(*C, C->getArgs(), Inputs, C->getActions()); 1378 1379 if (CCCPrintPhases) { 1380 PrintActions(*C); 1381 return C; 1382 } 1383 1384 BuildJobs(*C); 1385 1386 return C; 1387 } 1388 1389 static void printArgList(raw_ostream &OS, const llvm::opt::ArgList &Args) { 1390 llvm::opt::ArgStringList ASL; 1391 for (const auto *A : Args) { 1392 // Use user's original spelling of flags. For example, use 1393 // `/source-charset:utf-8` instead of `-finput-charset=utf-8` if the user 1394 // wrote the former. 1395 while (A->getAlias()) 1396 A = A->getAlias(); 1397 A->render(Args, ASL); 1398 } 1399 1400 for (auto I = ASL.begin(), E = ASL.end(); I != E; ++I) { 1401 if (I != ASL.begin()) 1402 OS << ' '; 1403 llvm::sys::printArg(OS, *I, true); 1404 } 1405 OS << '\n'; 1406 } 1407 1408 bool Driver::getCrashDiagnosticFile(StringRef ReproCrashFilename, 1409 SmallString<128> &CrashDiagDir) { 1410 using namespace llvm::sys; 1411 assert(llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin() && 1412 "Only knows about .crash files on Darwin"); 1413 1414 // The .crash file can be found on at ~/Library/Logs/DiagnosticReports/ 1415 // (or /Library/Logs/DiagnosticReports for root) and has the filename pattern 1416 // clang-<VERSION>_<YYYY-MM-DD-HHMMSS>_<hostname>.crash. 1417 path::home_directory(CrashDiagDir); 1418 if (CrashDiagDir.startswith("/var/root")) 1419 CrashDiagDir = "/"; 1420 path::append(CrashDiagDir, "Library/Logs/DiagnosticReports"); 1421 int PID = 1422 #if LLVM_ON_UNIX 1423 getpid(); 1424 #else 1425 0; 1426 #endif 1427 std::error_code EC; 1428 fs::file_status FileStatus; 1429 TimePoint<> LastAccessTime; 1430 SmallString<128> CrashFilePath; 1431 // Lookup the .crash files and get the one generated by a subprocess spawned 1432 // by this driver invocation. 1433 for (fs::directory_iterator File(CrashDiagDir, EC), FileEnd; 1434 File != FileEnd && !EC; File.increment(EC)) { 1435 StringRef FileName = path::filename(File->path()); 1436 if (!FileName.startswith(Name)) 1437 continue; 1438 if (fs::status(File->path(), FileStatus)) 1439 continue; 1440 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CrashFile = 1441 llvm::MemoryBuffer::getFile(File->path()); 1442 if (!CrashFile) 1443 continue; 1444 // The first line should start with "Process:", otherwise this isn't a real 1445 // .crash file. 1446 StringRef Data = CrashFile.get()->getBuffer(); 1447 if (!Data.startswith("Process:")) 1448 continue; 1449 // Parse parent process pid line, e.g: "Parent Process: clang-4.0 [79141]" 1450 size_t ParentProcPos = Data.find("Parent Process:"); 1451 if (ParentProcPos == StringRef::npos) 1452 continue; 1453 size_t LineEnd = Data.find_first_of("\n", ParentProcPos); 1454 if (LineEnd == StringRef::npos) 1455 continue; 1456 StringRef ParentProcess = Data.slice(ParentProcPos+15, LineEnd).trim(); 1457 int OpenBracket = -1, CloseBracket = -1; 1458 for (size_t i = 0, e = ParentProcess.size(); i < e; ++i) { 1459 if (ParentProcess[i] == '[') 1460 OpenBracket = i; 1461 if (ParentProcess[i] == ']') 1462 CloseBracket = i; 1463 } 1464 // Extract the parent process PID from the .crash file and check whether 1465 // it matches this driver invocation pid. 1466 int CrashPID; 1467 if (OpenBracket < 0 || CloseBracket < 0 || 1468 ParentProcess.slice(OpenBracket + 1, CloseBracket) 1469 .getAsInteger(10, CrashPID) || CrashPID != PID) { 1470 continue; 1471 } 1472 1473 // Found a .crash file matching the driver pid. To avoid getting an older 1474 // and misleading crash file, continue looking for the most recent. 1475 // FIXME: the driver can dispatch multiple cc1 invocations, leading to 1476 // multiple crashes poiting to the same parent process. Since the driver 1477 // does not collect pid information for the dispatched invocation there's 1478 // currently no way to distinguish among them. 1479 const auto FileAccessTime = FileStatus.getLastModificationTime(); 1480 if (FileAccessTime > LastAccessTime) { 1481 CrashFilePath.assign(File->path()); 1482 LastAccessTime = FileAccessTime; 1483 } 1484 } 1485 1486 // If found, copy it over to the location of other reproducer files. 1487 if (!CrashFilePath.empty()) { 1488 EC = fs::copy_file(CrashFilePath, ReproCrashFilename); 1489 if (EC) 1490 return false; 1491 return true; 1492 } 1493 1494 return false; 1495 } 1496 1497 // When clang crashes, produce diagnostic information including the fully 1498 // preprocessed source file(s). Request that the developer attach the 1499 // diagnostic information to a bug report. 1500 void Driver::generateCompilationDiagnostics( 1501 Compilation &C, const Command &FailingCommand, 1502 StringRef AdditionalInformation, CompilationDiagnosticReport *Report) { 1503 if (C.getArgs().hasArg(options::OPT_fno_crash_diagnostics)) 1504 return; 1505 1506 // Don't try to generate diagnostics for link or dsymutil jobs. 1507 if (FailingCommand.getCreator().isLinkJob() || 1508 FailingCommand.getCreator().isDsymutilJob()) 1509 return; 1510 1511 // Print the version of the compiler. 1512 PrintVersion(C, llvm::errs()); 1513 1514 // Suppress driver output and emit preprocessor output to temp file. 1515 CCGenDiagnostics = true; 1516 1517 // Save the original job command(s). 1518 Command Cmd = FailingCommand; 1519 1520 // Keep track of whether we produce any errors while trying to produce 1521 // preprocessed sources. 1522 DiagnosticErrorTrap Trap(Diags); 1523 1524 // Suppress tool output. 1525 C.initCompilationForDiagnostics(); 1526 1527 // Construct the list of inputs. 1528 InputList Inputs; 1529 BuildInputs(C.getDefaultToolChain(), C.getArgs(), Inputs); 1530 1531 for (InputList::iterator it = Inputs.begin(), ie = Inputs.end(); it != ie;) { 1532 bool IgnoreInput = false; 1533 1534 // Ignore input from stdin or any inputs that cannot be preprocessed. 1535 // Check type first as not all linker inputs have a value. 1536 if (types::getPreprocessedType(it->first) == types::TY_INVALID) { 1537 IgnoreInput = true; 1538 } else if (!strcmp(it->second->getValue(), "-")) { 1539 Diag(clang::diag::note_drv_command_failed_diag_msg) 1540 << "Error generating preprocessed source(s) - " 1541 "ignoring input from stdin."; 1542 IgnoreInput = true; 1543 } 1544 1545 if (IgnoreInput) { 1546 it = Inputs.erase(it); 1547 ie = Inputs.end(); 1548 } else { 1549 ++it; 1550 } 1551 } 1552 1553 if (Inputs.empty()) { 1554 Diag(clang::diag::note_drv_command_failed_diag_msg) 1555 << "Error generating preprocessed source(s) - " 1556 "no preprocessable inputs."; 1557 return; 1558 } 1559 1560 // Don't attempt to generate preprocessed files if multiple -arch options are 1561 // used, unless they're all duplicates. 1562 llvm::StringSet<> ArchNames; 1563 for (const Arg *A : C.getArgs()) { 1564 if (A->getOption().matches(options::OPT_arch)) { 1565 StringRef ArchName = A->getValue(); 1566 ArchNames.insert(ArchName); 1567 } 1568 } 1569 if (ArchNames.size() > 1) { 1570 Diag(clang::diag::note_drv_command_failed_diag_msg) 1571 << "Error generating preprocessed source(s) - cannot generate " 1572 "preprocessed source with multiple -arch options."; 1573 return; 1574 } 1575 1576 // Construct the list of abstract actions to perform for this compilation. On 1577 // Darwin OSes this uses the driver-driver and builds universal actions. 1578 const ToolChain &TC = C.getDefaultToolChain(); 1579 if (TC.getTriple().isOSBinFormatMachO()) 1580 BuildUniversalActions(C, TC, Inputs); 1581 else 1582 BuildActions(C, C.getArgs(), Inputs, C.getActions()); 1583 1584 BuildJobs(C); 1585 1586 // If there were errors building the compilation, quit now. 1587 if (Trap.hasErrorOccurred()) { 1588 Diag(clang::diag::note_drv_command_failed_diag_msg) 1589 << "Error generating preprocessed source(s)."; 1590 return; 1591 } 1592 1593 // Generate preprocessed output. 1594 SmallVector<std::pair<int, const Command *>, 4> FailingCommands; 1595 C.ExecuteJobs(C.getJobs(), FailingCommands); 1596 1597 // If any of the preprocessing commands failed, clean up and exit. 1598 if (!FailingCommands.empty()) { 1599 Diag(clang::diag::note_drv_command_failed_diag_msg) 1600 << "Error generating preprocessed source(s)."; 1601 return; 1602 } 1603 1604 const ArgStringList &TempFiles = C.getTempFiles(); 1605 if (TempFiles.empty()) { 1606 Diag(clang::diag::note_drv_command_failed_diag_msg) 1607 << "Error generating preprocessed source(s)."; 1608 return; 1609 } 1610 1611 Diag(clang::diag::note_drv_command_failed_diag_msg) 1612 << "\n********************\n\n" 1613 "PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:\n" 1614 "Preprocessed source(s) and associated run script(s) are located at:"; 1615 1616 SmallString<128> VFS; 1617 SmallString<128> ReproCrashFilename; 1618 for (const char *TempFile : TempFiles) { 1619 Diag(clang::diag::note_drv_command_failed_diag_msg) << TempFile; 1620 if (Report) 1621 Report->TemporaryFiles.push_back(TempFile); 1622 if (ReproCrashFilename.empty()) { 1623 ReproCrashFilename = TempFile; 1624 llvm::sys::path::replace_extension(ReproCrashFilename, ".crash"); 1625 } 1626 if (StringRef(TempFile).endswith(".cache")) { 1627 // In some cases (modules) we'll dump extra data to help with reproducing 1628 // the crash into a directory next to the output. 1629 VFS = llvm::sys::path::filename(TempFile); 1630 llvm::sys::path::append(VFS, "vfs", "vfs.yaml"); 1631 } 1632 } 1633 1634 // Assume associated files are based off of the first temporary file. 1635 CrashReportInfo CrashInfo(TempFiles[0], VFS); 1636 1637 llvm::SmallString<128> Script(CrashInfo.Filename); 1638 llvm::sys::path::replace_extension(Script, "sh"); 1639 std::error_code EC; 1640 llvm::raw_fd_ostream ScriptOS(Script, EC, llvm::sys::fs::CD_CreateNew, 1641 llvm::sys::fs::FA_Write, 1642 llvm::sys::fs::OF_Text); 1643 if (EC) { 1644 Diag(clang::diag::note_drv_command_failed_diag_msg) 1645 << "Error generating run script: " << Script << " " << EC.message(); 1646 } else { 1647 ScriptOS << "# Crash reproducer for " << getClangFullVersion() << "\n" 1648 << "# Driver args: "; 1649 printArgList(ScriptOS, C.getInputArgs()); 1650 ScriptOS << "# Original command: "; 1651 Cmd.Print(ScriptOS, "\n", /*Quote=*/true); 1652 Cmd.Print(ScriptOS, "\n", /*Quote=*/true, &CrashInfo); 1653 if (!AdditionalInformation.empty()) 1654 ScriptOS << "\n# Additional information: " << AdditionalInformation 1655 << "\n"; 1656 if (Report) 1657 Report->TemporaryFiles.push_back(std::string(Script.str())); 1658 Diag(clang::diag::note_drv_command_failed_diag_msg) << Script; 1659 } 1660 1661 // On darwin, provide information about the .crash diagnostic report. 1662 if (llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin()) { 1663 SmallString<128> CrashDiagDir; 1664 if (getCrashDiagnosticFile(ReproCrashFilename, CrashDiagDir)) { 1665 Diag(clang::diag::note_drv_command_failed_diag_msg) 1666 << ReproCrashFilename.str(); 1667 } else { // Suggest a directory for the user to look for .crash files. 1668 llvm::sys::path::append(CrashDiagDir, Name); 1669 CrashDiagDir += "_<YYYY-MM-DD-HHMMSS>_<hostname>.crash"; 1670 Diag(clang::diag::note_drv_command_failed_diag_msg) 1671 << "Crash backtrace is located in"; 1672 Diag(clang::diag::note_drv_command_failed_diag_msg) 1673 << CrashDiagDir.str(); 1674 Diag(clang::diag::note_drv_command_failed_diag_msg) 1675 << "(choose the .crash file that corresponds to your crash)"; 1676 } 1677 } 1678 1679 for (const auto &A : C.getArgs().filtered(options::OPT_frewrite_map_file_EQ)) 1680 Diag(clang::diag::note_drv_command_failed_diag_msg) << A->getValue(); 1681 1682 Diag(clang::diag::note_drv_command_failed_diag_msg) 1683 << "\n\n********************"; 1684 } 1685 1686 void Driver::setUpResponseFiles(Compilation &C, Command &Cmd) { 1687 // Since commandLineFitsWithinSystemLimits() may underestimate system's 1688 // capacity if the tool does not support response files, there is a chance/ 1689 // that things will just work without a response file, so we silently just 1690 // skip it. 1691 if (Cmd.getResponseFileSupport().ResponseKind == 1692 ResponseFileSupport::RF_None || 1693 llvm::sys::commandLineFitsWithinSystemLimits(Cmd.getExecutable(), 1694 Cmd.getArguments())) 1695 return; 1696 1697 std::string TmpName = GetTemporaryPath("response", "txt"); 1698 Cmd.setResponseFile(C.addTempFile(C.getArgs().MakeArgString(TmpName))); 1699 } 1700 1701 int Driver::ExecuteCompilation( 1702 Compilation &C, 1703 SmallVectorImpl<std::pair<int, const Command *>> &FailingCommands) { 1704 if (C.getArgs().hasArg(options::OPT_fdriver_only)) { 1705 if (C.getArgs().hasArg(options::OPT_v)) 1706 C.getJobs().Print(llvm::errs(), "\n", true); 1707 1708 C.ExecuteJobs(C.getJobs(), FailingCommands, /*LogOnly=*/true); 1709 1710 // If there were errors building the compilation, quit now. 1711 if (!FailingCommands.empty() || Diags.hasErrorOccurred()) 1712 return 1; 1713 1714 return 0; 1715 } 1716 1717 // Just print if -### was present. 1718 if (C.getArgs().hasArg(options::OPT__HASH_HASH_HASH)) { 1719 C.getJobs().Print(llvm::errs(), "\n", true); 1720 return 0; 1721 } 1722 1723 // If there were errors building the compilation, quit now. 1724 if (Diags.hasErrorOccurred()) 1725 return 1; 1726 1727 // Set up response file names for each command, if necessary. 1728 for (auto &Job : C.getJobs()) 1729 setUpResponseFiles(C, Job); 1730 1731 C.ExecuteJobs(C.getJobs(), FailingCommands); 1732 1733 // If the command succeeded, we are done. 1734 if (FailingCommands.empty()) 1735 return 0; 1736 1737 // Otherwise, remove result files and print extra information about abnormal 1738 // failures. 1739 int Res = 0; 1740 for (const auto &CmdPair : FailingCommands) { 1741 int CommandRes = CmdPair.first; 1742 const Command *FailingCommand = CmdPair.second; 1743 1744 // Remove result files if we're not saving temps. 1745 if (!isSaveTempsEnabled()) { 1746 const JobAction *JA = cast<JobAction>(&FailingCommand->getSource()); 1747 C.CleanupFileMap(C.getResultFiles(), JA, true); 1748 1749 // Failure result files are valid unless we crashed. 1750 if (CommandRes < 0) 1751 C.CleanupFileMap(C.getFailureResultFiles(), JA, true); 1752 } 1753 1754 #if LLVM_ON_UNIX 1755 // llvm/lib/Support/Unix/Signals.inc will exit with a special return code 1756 // for SIGPIPE. Do not print diagnostics for this case. 1757 if (CommandRes == EX_IOERR) { 1758 Res = CommandRes; 1759 continue; 1760 } 1761 #endif 1762 1763 // Print extra information about abnormal failures, if possible. 1764 // 1765 // This is ad-hoc, but we don't want to be excessively noisy. If the result 1766 // status was 1, assume the command failed normally. In particular, if it 1767 // was the compiler then assume it gave a reasonable error code. Failures 1768 // in other tools are less common, and they generally have worse 1769 // diagnostics, so always print the diagnostic there. 1770 const Tool &FailingTool = FailingCommand->getCreator(); 1771 1772 if (!FailingCommand->getCreator().hasGoodDiagnostics() || CommandRes != 1) { 1773 // FIXME: See FIXME above regarding result code interpretation. 1774 if (CommandRes < 0) 1775 Diag(clang::diag::err_drv_command_signalled) 1776 << FailingTool.getShortName(); 1777 else 1778 Diag(clang::diag::err_drv_command_failed) 1779 << FailingTool.getShortName() << CommandRes; 1780 } 1781 } 1782 return Res; 1783 } 1784 1785 void Driver::PrintHelp(bool ShowHidden) const { 1786 unsigned IncludedFlagsBitmask; 1787 unsigned ExcludedFlagsBitmask; 1788 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 1789 getIncludeExcludeOptionFlagMasks(IsCLMode()); 1790 1791 ExcludedFlagsBitmask |= options::NoDriverOption; 1792 if (!ShowHidden) 1793 ExcludedFlagsBitmask |= HelpHidden; 1794 1795 if (IsFlangMode()) 1796 IncludedFlagsBitmask |= options::FlangOption; 1797 else 1798 ExcludedFlagsBitmask |= options::FlangOnlyOption; 1799 1800 std::string Usage = llvm::formatv("{0} [options] file...", Name).str(); 1801 getOpts().printHelp(llvm::outs(), Usage.c_str(), DriverTitle.c_str(), 1802 IncludedFlagsBitmask, ExcludedFlagsBitmask, 1803 /*ShowAllAliases=*/false); 1804 } 1805 1806 void Driver::PrintVersion(const Compilation &C, raw_ostream &OS) const { 1807 if (IsFlangMode()) { 1808 OS << getClangToolFullVersion("flang-new") << '\n'; 1809 } else { 1810 // FIXME: The following handlers should use a callback mechanism, we don't 1811 // know what the client would like to do. 1812 OS << getClangFullVersion() << '\n'; 1813 } 1814 const ToolChain &TC = C.getDefaultToolChain(); 1815 OS << "Target: " << TC.getTripleString() << '\n'; 1816 1817 // Print the threading model. 1818 if (Arg *A = C.getArgs().getLastArg(options::OPT_mthread_model)) { 1819 // Don't print if the ToolChain would have barfed on it already 1820 if (TC.isThreadModelSupported(A->getValue())) 1821 OS << "Thread model: " << A->getValue(); 1822 } else 1823 OS << "Thread model: " << TC.getThreadModel(); 1824 OS << '\n'; 1825 1826 // Print out the install directory. 1827 OS << "InstalledDir: " << InstalledDir << '\n'; 1828 1829 // If configuration file was used, print its path. 1830 if (!ConfigFile.empty()) 1831 OS << "Configuration file: " << ConfigFile << '\n'; 1832 } 1833 1834 /// PrintDiagnosticCategories - Implement the --print-diagnostic-categories 1835 /// option. 1836 static void PrintDiagnosticCategories(raw_ostream &OS) { 1837 // Skip the empty category. 1838 for (unsigned i = 1, max = DiagnosticIDs::getNumberOfCategories(); i != max; 1839 ++i) 1840 OS << i << ',' << DiagnosticIDs::getCategoryNameFromID(i) << '\n'; 1841 } 1842 1843 void Driver::HandleAutocompletions(StringRef PassedFlags) const { 1844 if (PassedFlags == "") 1845 return; 1846 // Print out all options that start with a given argument. This is used for 1847 // shell autocompletion. 1848 std::vector<std::string> SuggestedCompletions; 1849 std::vector<std::string> Flags; 1850 1851 unsigned int DisableFlags = 1852 options::NoDriverOption | options::Unsupported | options::Ignored; 1853 1854 // Make sure that Flang-only options don't pollute the Clang output 1855 // TODO: Make sure that Clang-only options don't pollute Flang output 1856 if (!IsFlangMode()) 1857 DisableFlags |= options::FlangOnlyOption; 1858 1859 // Distinguish "--autocomplete=-someflag" and "--autocomplete=-someflag," 1860 // because the latter indicates that the user put space before pushing tab 1861 // which should end up in a file completion. 1862 const bool HasSpace = PassedFlags.endswith(","); 1863 1864 // Parse PassedFlags by "," as all the command-line flags are passed to this 1865 // function separated by "," 1866 StringRef TargetFlags = PassedFlags; 1867 while (TargetFlags != "") { 1868 StringRef CurFlag; 1869 std::tie(CurFlag, TargetFlags) = TargetFlags.split(","); 1870 Flags.push_back(std::string(CurFlag)); 1871 } 1872 1873 // We want to show cc1-only options only when clang is invoked with -cc1 or 1874 // -Xclang. 1875 if (llvm::is_contained(Flags, "-Xclang") || llvm::is_contained(Flags, "-cc1")) 1876 DisableFlags &= ~options::NoDriverOption; 1877 1878 const llvm::opt::OptTable &Opts = getOpts(); 1879 StringRef Cur; 1880 Cur = Flags.at(Flags.size() - 1); 1881 StringRef Prev; 1882 if (Flags.size() >= 2) { 1883 Prev = Flags.at(Flags.size() - 2); 1884 SuggestedCompletions = Opts.suggestValueCompletions(Prev, Cur); 1885 } 1886 1887 if (SuggestedCompletions.empty()) 1888 SuggestedCompletions = Opts.suggestValueCompletions(Cur, ""); 1889 1890 // If Flags were empty, it means the user typed `clang [tab]` where we should 1891 // list all possible flags. If there was no value completion and the user 1892 // pressed tab after a space, we should fall back to a file completion. 1893 // We're printing a newline to be consistent with what we print at the end of 1894 // this function. 1895 if (SuggestedCompletions.empty() && HasSpace && !Flags.empty()) { 1896 llvm::outs() << '\n'; 1897 return; 1898 } 1899 1900 // When flag ends with '=' and there was no value completion, return empty 1901 // string and fall back to the file autocompletion. 1902 if (SuggestedCompletions.empty() && !Cur.endswith("=")) { 1903 // If the flag is in the form of "--autocomplete=-foo", 1904 // we were requested to print out all option names that start with "-foo". 1905 // For example, "--autocomplete=-fsyn" is expanded to "-fsyntax-only". 1906 SuggestedCompletions = Opts.findByPrefix(Cur, DisableFlags); 1907 1908 // We have to query the -W flags manually as they're not in the OptTable. 1909 // TODO: Find a good way to add them to OptTable instead and them remove 1910 // this code. 1911 for (StringRef S : DiagnosticIDs::getDiagnosticFlags()) 1912 if (S.startswith(Cur)) 1913 SuggestedCompletions.push_back(std::string(S)); 1914 } 1915 1916 // Sort the autocomplete candidates so that shells print them out in a 1917 // deterministic order. We could sort in any way, but we chose 1918 // case-insensitive sorting for consistency with the -help option 1919 // which prints out options in the case-insensitive alphabetical order. 1920 llvm::sort(SuggestedCompletions, [](StringRef A, StringRef B) { 1921 if (int X = A.compare_insensitive(B)) 1922 return X < 0; 1923 return A.compare(B) > 0; 1924 }); 1925 1926 llvm::outs() << llvm::join(SuggestedCompletions, "\n") << '\n'; 1927 } 1928 1929 bool Driver::HandleImmediateArgs(const Compilation &C) { 1930 // The order these options are handled in gcc is all over the place, but we 1931 // don't expect inconsistencies w.r.t. that to matter in practice. 1932 1933 if (C.getArgs().hasArg(options::OPT_dumpmachine)) { 1934 llvm::outs() << C.getDefaultToolChain().getTripleString() << '\n'; 1935 return false; 1936 } 1937 1938 if (C.getArgs().hasArg(options::OPT_dumpversion)) { 1939 // Since -dumpversion is only implemented for pedantic GCC compatibility, we 1940 // return an answer which matches our definition of __VERSION__. 1941 llvm::outs() << CLANG_VERSION_STRING << "\n"; 1942 return false; 1943 } 1944 1945 if (C.getArgs().hasArg(options::OPT__print_diagnostic_categories)) { 1946 PrintDiagnosticCategories(llvm::outs()); 1947 return false; 1948 } 1949 1950 if (C.getArgs().hasArg(options::OPT_help) || 1951 C.getArgs().hasArg(options::OPT__help_hidden)) { 1952 PrintHelp(C.getArgs().hasArg(options::OPT__help_hidden)); 1953 return false; 1954 } 1955 1956 if (C.getArgs().hasArg(options::OPT__version)) { 1957 // Follow gcc behavior and use stdout for --version and stderr for -v. 1958 PrintVersion(C, llvm::outs()); 1959 return false; 1960 } 1961 1962 if (C.getArgs().hasArg(options::OPT_v) || 1963 C.getArgs().hasArg(options::OPT__HASH_HASH_HASH) || 1964 C.getArgs().hasArg(options::OPT_print_supported_cpus)) { 1965 PrintVersion(C, llvm::errs()); 1966 SuppressMissingInputWarning = true; 1967 } 1968 1969 if (C.getArgs().hasArg(options::OPT_v)) { 1970 if (!SystemConfigDir.empty()) 1971 llvm::errs() << "System configuration file directory: " 1972 << SystemConfigDir << "\n"; 1973 if (!UserConfigDir.empty()) 1974 llvm::errs() << "User configuration file directory: " 1975 << UserConfigDir << "\n"; 1976 } 1977 1978 const ToolChain &TC = C.getDefaultToolChain(); 1979 1980 if (C.getArgs().hasArg(options::OPT_v)) 1981 TC.printVerboseInfo(llvm::errs()); 1982 1983 if (C.getArgs().hasArg(options::OPT_print_resource_dir)) { 1984 llvm::outs() << ResourceDir << '\n'; 1985 return false; 1986 } 1987 1988 if (C.getArgs().hasArg(options::OPT_print_search_dirs)) { 1989 llvm::outs() << "programs: ="; 1990 bool separator = false; 1991 // Print -B and COMPILER_PATH. 1992 for (const std::string &Path : PrefixDirs) { 1993 if (separator) 1994 llvm::outs() << llvm::sys::EnvPathSeparator; 1995 llvm::outs() << Path; 1996 separator = true; 1997 } 1998 for (const std::string &Path : TC.getProgramPaths()) { 1999 if (separator) 2000 llvm::outs() << llvm::sys::EnvPathSeparator; 2001 llvm::outs() << Path; 2002 separator = true; 2003 } 2004 llvm::outs() << "\n"; 2005 llvm::outs() << "libraries: =" << ResourceDir; 2006 2007 StringRef sysroot = C.getSysRoot(); 2008 2009 for (const std::string &Path : TC.getFilePaths()) { 2010 // Always print a separator. ResourceDir was the first item shown. 2011 llvm::outs() << llvm::sys::EnvPathSeparator; 2012 // Interpretation of leading '=' is needed only for NetBSD. 2013 if (Path[0] == '=') 2014 llvm::outs() << sysroot << Path.substr(1); 2015 else 2016 llvm::outs() << Path; 2017 } 2018 llvm::outs() << "\n"; 2019 return false; 2020 } 2021 2022 if (C.getArgs().hasArg(options::OPT_print_runtime_dir)) { 2023 std::string RuntimePath; 2024 // Get the first existing path, if any. 2025 for (auto Path : TC.getRuntimePaths()) { 2026 if (getVFS().exists(Path)) { 2027 RuntimePath = Path; 2028 break; 2029 } 2030 } 2031 if (!RuntimePath.empty()) 2032 llvm::outs() << RuntimePath << '\n'; 2033 else 2034 llvm::outs() << TC.getCompilerRTPath() << '\n'; 2035 return false; 2036 } 2037 2038 if (C.getArgs().hasArg(options::OPT_print_diagnostic_options)) { 2039 std::vector<std::string> Flags = DiagnosticIDs::getDiagnosticFlags(); 2040 for (std::size_t I = 0; I != Flags.size(); I += 2) 2041 llvm::outs() << " " << Flags[I] << "\n " << Flags[I + 1] << "\n\n"; 2042 return false; 2043 } 2044 2045 // FIXME: The following handlers should use a callback mechanism, we don't 2046 // know what the client would like to do. 2047 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_file_name_EQ)) { 2048 llvm::outs() << GetFilePath(A->getValue(), TC) << "\n"; 2049 return false; 2050 } 2051 2052 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_prog_name_EQ)) { 2053 StringRef ProgName = A->getValue(); 2054 2055 // Null program name cannot have a path. 2056 if (! ProgName.empty()) 2057 llvm::outs() << GetProgramPath(ProgName, TC); 2058 2059 llvm::outs() << "\n"; 2060 return false; 2061 } 2062 2063 if (Arg *A = C.getArgs().getLastArg(options::OPT_autocomplete)) { 2064 StringRef PassedFlags = A->getValue(); 2065 HandleAutocompletions(PassedFlags); 2066 return false; 2067 } 2068 2069 if (C.getArgs().hasArg(options::OPT_print_libgcc_file_name)) { 2070 ToolChain::RuntimeLibType RLT = TC.GetRuntimeLibType(C.getArgs()); 2071 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2072 RegisterEffectiveTriple TripleRAII(TC, Triple); 2073 switch (RLT) { 2074 case ToolChain::RLT_CompilerRT: 2075 llvm::outs() << TC.getCompilerRT(C.getArgs(), "builtins") << "\n"; 2076 break; 2077 case ToolChain::RLT_Libgcc: 2078 llvm::outs() << GetFilePath("libgcc.a", TC) << "\n"; 2079 break; 2080 } 2081 return false; 2082 } 2083 2084 if (C.getArgs().hasArg(options::OPT_print_multi_lib)) { 2085 for (const Multilib &Multilib : TC.getMultilibs()) 2086 llvm::outs() << Multilib << "\n"; 2087 return false; 2088 } 2089 2090 if (C.getArgs().hasArg(options::OPT_print_multi_directory)) { 2091 const Multilib &Multilib = TC.getMultilib(); 2092 if (Multilib.gccSuffix().empty()) 2093 llvm::outs() << ".\n"; 2094 else { 2095 StringRef Suffix(Multilib.gccSuffix()); 2096 assert(Suffix.front() == '/'); 2097 llvm::outs() << Suffix.substr(1) << "\n"; 2098 } 2099 return false; 2100 } 2101 2102 if (C.getArgs().hasArg(options::OPT_print_target_triple)) { 2103 llvm::outs() << TC.getTripleString() << "\n"; 2104 return false; 2105 } 2106 2107 if (C.getArgs().hasArg(options::OPT_print_effective_triple)) { 2108 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2109 llvm::outs() << Triple.getTriple() << "\n"; 2110 return false; 2111 } 2112 2113 if (C.getArgs().hasArg(options::OPT_print_multiarch)) { 2114 llvm::outs() << TC.getMultiarchTriple(*this, TC.getTriple(), SysRoot) 2115 << "\n"; 2116 return false; 2117 } 2118 2119 if (C.getArgs().hasArg(options::OPT_print_targets)) { 2120 llvm::TargetRegistry::printRegisteredTargetsForVersion(llvm::outs()); 2121 return false; 2122 } 2123 2124 return true; 2125 } 2126 2127 enum { 2128 TopLevelAction = 0, 2129 HeadSibAction = 1, 2130 OtherSibAction = 2, 2131 }; 2132 2133 // Display an action graph human-readably. Action A is the "sink" node 2134 // and latest-occuring action. Traversal is in pre-order, visiting the 2135 // inputs to each action before printing the action itself. 2136 static unsigned PrintActions1(const Compilation &C, Action *A, 2137 std::map<Action *, unsigned> &Ids, 2138 Twine Indent = {}, int Kind = TopLevelAction) { 2139 if (Ids.count(A)) // A was already visited. 2140 return Ids[A]; 2141 2142 std::string str; 2143 llvm::raw_string_ostream os(str); 2144 2145 auto getSibIndent = [](int K) -> Twine { 2146 return (K == HeadSibAction) ? " " : (K == OtherSibAction) ? "| " : ""; 2147 }; 2148 2149 Twine SibIndent = Indent + getSibIndent(Kind); 2150 int SibKind = HeadSibAction; 2151 os << Action::getClassName(A->getKind()) << ", "; 2152 if (InputAction *IA = dyn_cast<InputAction>(A)) { 2153 os << "\"" << IA->getInputArg().getValue() << "\""; 2154 } else if (BindArchAction *BIA = dyn_cast<BindArchAction>(A)) { 2155 os << '"' << BIA->getArchName() << '"' << ", {" 2156 << PrintActions1(C, *BIA->input_begin(), Ids, SibIndent, SibKind) << "}"; 2157 } else if (OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 2158 bool IsFirst = true; 2159 OA->doOnEachDependence( 2160 [&](Action *A, const ToolChain *TC, const char *BoundArch) { 2161 assert(TC && "Unknown host toolchain"); 2162 // E.g. for two CUDA device dependences whose bound arch is sm_20 and 2163 // sm_35 this will generate: 2164 // "cuda-device" (nvptx64-nvidia-cuda:sm_20) {#ID}, "cuda-device" 2165 // (nvptx64-nvidia-cuda:sm_35) {#ID} 2166 if (!IsFirst) 2167 os << ", "; 2168 os << '"'; 2169 os << A->getOffloadingKindPrefix(); 2170 os << " ("; 2171 os << TC->getTriple().normalize(); 2172 if (BoundArch) 2173 os << ":" << BoundArch; 2174 os << ")"; 2175 os << '"'; 2176 os << " {" << PrintActions1(C, A, Ids, SibIndent, SibKind) << "}"; 2177 IsFirst = false; 2178 SibKind = OtherSibAction; 2179 }); 2180 } else { 2181 const ActionList *AL = &A->getInputs(); 2182 2183 if (AL->size()) { 2184 const char *Prefix = "{"; 2185 for (Action *PreRequisite : *AL) { 2186 os << Prefix << PrintActions1(C, PreRequisite, Ids, SibIndent, SibKind); 2187 Prefix = ", "; 2188 SibKind = OtherSibAction; 2189 } 2190 os << "}"; 2191 } else 2192 os << "{}"; 2193 } 2194 2195 // Append offload info for all options other than the offloading action 2196 // itself (e.g. (cuda-device, sm_20) or (cuda-host)). 2197 std::string offload_str; 2198 llvm::raw_string_ostream offload_os(offload_str); 2199 if (!isa<OffloadAction>(A)) { 2200 auto S = A->getOffloadingKindPrefix(); 2201 if (!S.empty()) { 2202 offload_os << ", (" << S; 2203 if (A->getOffloadingArch()) 2204 offload_os << ", " << A->getOffloadingArch(); 2205 offload_os << ")"; 2206 } 2207 } 2208 2209 auto getSelfIndent = [](int K) -> Twine { 2210 return (K == HeadSibAction) ? "+- " : (K == OtherSibAction) ? "|- " : ""; 2211 }; 2212 2213 unsigned Id = Ids.size(); 2214 Ids[A] = Id; 2215 llvm::errs() << Indent + getSelfIndent(Kind) << Id << ": " << os.str() << ", " 2216 << types::getTypeName(A->getType()) << offload_os.str() << "\n"; 2217 2218 return Id; 2219 } 2220 2221 // Print the action graphs in a compilation C. 2222 // For example "clang -c file1.c file2.c" is composed of two subgraphs. 2223 void Driver::PrintActions(const Compilation &C) const { 2224 std::map<Action *, unsigned> Ids; 2225 for (Action *A : C.getActions()) 2226 PrintActions1(C, A, Ids); 2227 } 2228 2229 /// Check whether the given input tree contains any compilation or 2230 /// assembly actions. 2231 static bool ContainsCompileOrAssembleAction(const Action *A) { 2232 if (isa<CompileJobAction>(A) || isa<BackendJobAction>(A) || 2233 isa<AssembleJobAction>(A)) 2234 return true; 2235 2236 return llvm::any_of(A->inputs(), ContainsCompileOrAssembleAction); 2237 } 2238 2239 void Driver::BuildUniversalActions(Compilation &C, const ToolChain &TC, 2240 const InputList &BAInputs) const { 2241 DerivedArgList &Args = C.getArgs(); 2242 ActionList &Actions = C.getActions(); 2243 llvm::PrettyStackTraceString CrashInfo("Building universal build actions"); 2244 // Collect the list of architectures. Duplicates are allowed, but should only 2245 // be handled once (in the order seen). 2246 llvm::StringSet<> ArchNames; 2247 SmallVector<const char *, 4> Archs; 2248 for (Arg *A : Args) { 2249 if (A->getOption().matches(options::OPT_arch)) { 2250 // Validate the option here; we don't save the type here because its 2251 // particular spelling may participate in other driver choices. 2252 llvm::Triple::ArchType Arch = 2253 tools::darwin::getArchTypeForMachOArchName(A->getValue()); 2254 if (Arch == llvm::Triple::UnknownArch) { 2255 Diag(clang::diag::err_drv_invalid_arch_name) << A->getAsString(Args); 2256 continue; 2257 } 2258 2259 A->claim(); 2260 if (ArchNames.insert(A->getValue()).second) 2261 Archs.push_back(A->getValue()); 2262 } 2263 } 2264 2265 // When there is no explicit arch for this platform, make sure we still bind 2266 // the architecture (to the default) so that -Xarch_ is handled correctly. 2267 if (!Archs.size()) 2268 Archs.push_back(Args.MakeArgString(TC.getDefaultUniversalArchName())); 2269 2270 ActionList SingleActions; 2271 BuildActions(C, Args, BAInputs, SingleActions); 2272 2273 // Add in arch bindings for every top level action, as well as lipo and 2274 // dsymutil steps if needed. 2275 for (Action* Act : SingleActions) { 2276 // Make sure we can lipo this kind of output. If not (and it is an actual 2277 // output) then we disallow, since we can't create an output file with the 2278 // right name without overwriting it. We could remove this oddity by just 2279 // changing the output names to include the arch, which would also fix 2280 // -save-temps. Compatibility wins for now. 2281 2282 if (Archs.size() > 1 && !types::canLipoType(Act->getType())) 2283 Diag(clang::diag::err_drv_invalid_output_with_multiple_archs) 2284 << types::getTypeName(Act->getType()); 2285 2286 ActionList Inputs; 2287 for (unsigned i = 0, e = Archs.size(); i != e; ++i) 2288 Inputs.push_back(C.MakeAction<BindArchAction>(Act, Archs[i])); 2289 2290 // Lipo if necessary, we do it this way because we need to set the arch flag 2291 // so that -Xarch_ gets overwritten. 2292 if (Inputs.size() == 1 || Act->getType() == types::TY_Nothing) 2293 Actions.append(Inputs.begin(), Inputs.end()); 2294 else 2295 Actions.push_back(C.MakeAction<LipoJobAction>(Inputs, Act->getType())); 2296 2297 // Handle debug info queries. 2298 Arg *A = Args.getLastArg(options::OPT_g_Group); 2299 bool enablesDebugInfo = A && !A->getOption().matches(options::OPT_g0) && 2300 !A->getOption().matches(options::OPT_gstabs); 2301 if ((enablesDebugInfo || willEmitRemarks(Args)) && 2302 ContainsCompileOrAssembleAction(Actions.back())) { 2303 2304 // Add a 'dsymutil' step if necessary, when debug info is enabled and we 2305 // have a compile input. We need to run 'dsymutil' ourselves in such cases 2306 // because the debug info will refer to a temporary object file which 2307 // will be removed at the end of the compilation process. 2308 if (Act->getType() == types::TY_Image) { 2309 ActionList Inputs; 2310 Inputs.push_back(Actions.back()); 2311 Actions.pop_back(); 2312 Actions.push_back( 2313 C.MakeAction<DsymutilJobAction>(Inputs, types::TY_dSYM)); 2314 } 2315 2316 // Verify the debug info output. 2317 if (Args.hasArg(options::OPT_verify_debug_info)) { 2318 Action* LastAction = Actions.back(); 2319 Actions.pop_back(); 2320 Actions.push_back(C.MakeAction<VerifyDebugInfoJobAction>( 2321 LastAction, types::TY_Nothing)); 2322 } 2323 } 2324 } 2325 } 2326 2327 bool Driver::DiagnoseInputExistence(const DerivedArgList &Args, StringRef Value, 2328 types::ID Ty, bool TypoCorrect) const { 2329 if (!getCheckInputsExist()) 2330 return true; 2331 2332 // stdin always exists. 2333 if (Value == "-") 2334 return true; 2335 2336 // If it's a header to be found in the system or user search path, then defer 2337 // complaints about its absence until those searches can be done. When we 2338 // are definitely processing headers for C++20 header units, extend this to 2339 // allow the user to put "-fmodule-header -xc++-header vector" for example. 2340 if (Ty == types::TY_CXXSHeader || Ty == types::TY_CXXUHeader || 2341 (ModulesModeCXX20 && Ty == types::TY_CXXHeader)) 2342 return true; 2343 2344 if (getVFS().exists(Value)) 2345 return true; 2346 2347 if (TypoCorrect) { 2348 // Check if the filename is a typo for an option flag. OptTable thinks 2349 // that all args that are not known options and that start with / are 2350 // filenames, but e.g. `/diagnostic:caret` is more likely a typo for 2351 // the option `/diagnostics:caret` than a reference to a file in the root 2352 // directory. 2353 unsigned IncludedFlagsBitmask; 2354 unsigned ExcludedFlagsBitmask; 2355 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 2356 getIncludeExcludeOptionFlagMasks(IsCLMode()); 2357 std::string Nearest; 2358 if (getOpts().findNearest(Value, Nearest, IncludedFlagsBitmask, 2359 ExcludedFlagsBitmask) <= 1) { 2360 Diag(clang::diag::err_drv_no_such_file_with_suggestion) 2361 << Value << Nearest; 2362 return false; 2363 } 2364 } 2365 2366 // In CL mode, don't error on apparently non-existent linker inputs, because 2367 // they can be influenced by linker flags the clang driver might not 2368 // understand. 2369 // Examples: 2370 // - `clang-cl main.cc ole32.lib` in a a non-MSVC shell will make the driver 2371 // module look for an MSVC installation in the registry. (We could ask 2372 // the MSVCToolChain object if it can find `ole32.lib`, but the logic to 2373 // look in the registry might move into lld-link in the future so that 2374 // lld-link invocations in non-MSVC shells just work too.) 2375 // - `clang-cl ... /link ...` can pass arbitrary flags to the linker, 2376 // including /libpath:, which is used to find .lib and .obj files. 2377 // So do not diagnose this on the driver level. Rely on the linker diagnosing 2378 // it. (If we don't end up invoking the linker, this means we'll emit a 2379 // "'linker' input unused [-Wunused-command-line-argument]" warning instead 2380 // of an error.) 2381 // 2382 // Only do this skip after the typo correction step above. `/Brepo` is treated 2383 // as TY_Object, but it's clearly a typo for `/Brepro`. It seems fine to emit 2384 // an error if we have a flag that's within an edit distance of 1 from a 2385 // flag. (Users can use `-Wl,` or `/linker` to launder the flag past the 2386 // driver in the unlikely case they run into this.) 2387 // 2388 // Don't do this for inputs that start with a '/', else we'd pass options 2389 // like /libpath: through to the linker silently. 2390 // 2391 // Emitting an error for linker inputs can also cause incorrect diagnostics 2392 // with the gcc driver. The command 2393 // clang -fuse-ld=lld -Wl,--chroot,some/dir /file.o 2394 // will make lld look for some/dir/file.o, while we will diagnose here that 2395 // `/file.o` does not exist. However, configure scripts check if 2396 // `clang /GR-` compiles without error to see if the compiler is cl.exe, 2397 // so we can't downgrade diagnostics for `/GR-` from an error to a warning 2398 // in cc mode. (We can in cl mode because cl.exe itself only warns on 2399 // unknown flags.) 2400 if (IsCLMode() && Ty == types::TY_Object && !Value.startswith("/")) 2401 return true; 2402 2403 Diag(clang::diag::err_drv_no_such_file) << Value; 2404 return false; 2405 } 2406 2407 // Get the C++20 Header Unit type corresponding to the input type. 2408 static types::ID CXXHeaderUnitType(ModuleHeaderMode HM) { 2409 switch (HM) { 2410 case HeaderMode_User: 2411 return types::TY_CXXUHeader; 2412 case HeaderMode_System: 2413 return types::TY_CXXSHeader; 2414 case HeaderMode_Default: 2415 break; 2416 case HeaderMode_None: 2417 llvm_unreachable("should not be called in this case"); 2418 } 2419 return types::TY_CXXHUHeader; 2420 } 2421 2422 // Construct a the list of inputs and their types. 2423 void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args, 2424 InputList &Inputs) const { 2425 const llvm::opt::OptTable &Opts = getOpts(); 2426 // Track the current user specified (-x) input. We also explicitly track the 2427 // argument used to set the type; we only want to claim the type when we 2428 // actually use it, so we warn about unused -x arguments. 2429 types::ID InputType = types::TY_Nothing; 2430 Arg *InputTypeArg = nullptr; 2431 2432 // The last /TC or /TP option sets the input type to C or C++ globally. 2433 if (Arg *TCTP = Args.getLastArgNoClaim(options::OPT__SLASH_TC, 2434 options::OPT__SLASH_TP)) { 2435 InputTypeArg = TCTP; 2436 InputType = TCTP->getOption().matches(options::OPT__SLASH_TC) 2437 ? types::TY_C 2438 : types::TY_CXX; 2439 2440 Arg *Previous = nullptr; 2441 bool ShowNote = false; 2442 for (Arg *A : 2443 Args.filtered(options::OPT__SLASH_TC, options::OPT__SLASH_TP)) { 2444 if (Previous) { 2445 Diag(clang::diag::warn_drv_overriding_flag_option) 2446 << Previous->getSpelling() << A->getSpelling(); 2447 ShowNote = true; 2448 } 2449 Previous = A; 2450 } 2451 if (ShowNote) 2452 Diag(clang::diag::note_drv_t_option_is_global); 2453 2454 // No driver mode exposes -x and /TC or /TP; we don't support mixing them. 2455 assert(!Args.hasArg(options::OPT_x) && "-x and /TC or /TP is not allowed"); 2456 } 2457 2458 // Warn -x after last input file has no effect 2459 { 2460 Arg *LastXArg = Args.getLastArgNoClaim(options::OPT_x); 2461 Arg *LastInputArg = Args.getLastArgNoClaim(options::OPT_INPUT); 2462 if (LastXArg && LastInputArg && LastInputArg->getIndex() < LastXArg->getIndex()) 2463 Diag(clang::diag::warn_drv_unused_x) << LastXArg->getValue(); 2464 } 2465 2466 for (Arg *A : Args) { 2467 if (A->getOption().getKind() == Option::InputClass) { 2468 const char *Value = A->getValue(); 2469 types::ID Ty = types::TY_INVALID; 2470 2471 // Infer the input type if necessary. 2472 if (InputType == types::TY_Nothing) { 2473 // If there was an explicit arg for this, claim it. 2474 if (InputTypeArg) 2475 InputTypeArg->claim(); 2476 2477 // stdin must be handled specially. 2478 if (memcmp(Value, "-", 2) == 0) { 2479 if (IsFlangMode()) { 2480 Ty = types::TY_Fortran; 2481 } else { 2482 // If running with -E, treat as a C input (this changes the 2483 // builtin macros, for example). This may be overridden by -ObjC 2484 // below. 2485 // 2486 // Otherwise emit an error but still use a valid type to avoid 2487 // spurious errors (e.g., no inputs). 2488 assert(!CCGenDiagnostics && "stdin produces no crash reproducer"); 2489 if (!Args.hasArgNoClaim(options::OPT_E) && !CCCIsCPP()) 2490 Diag(IsCLMode() ? clang::diag::err_drv_unknown_stdin_type_clang_cl 2491 : clang::diag::err_drv_unknown_stdin_type); 2492 Ty = types::TY_C; 2493 } 2494 } else { 2495 // Otherwise lookup by extension. 2496 // Fallback is C if invoked as C preprocessor, C++ if invoked with 2497 // clang-cl /E, or Object otherwise. 2498 // We use a host hook here because Darwin at least has its own 2499 // idea of what .s is. 2500 if (const char *Ext = strrchr(Value, '.')) 2501 Ty = TC.LookupTypeForExtension(Ext + 1); 2502 2503 if (Ty == types::TY_INVALID) { 2504 if (IsCLMode() && (Args.hasArgNoClaim(options::OPT_E) || CCGenDiagnostics)) 2505 Ty = types::TY_CXX; 2506 else if (CCCIsCPP() || CCGenDiagnostics) 2507 Ty = types::TY_C; 2508 else 2509 Ty = types::TY_Object; 2510 } 2511 2512 // If the driver is invoked as C++ compiler (like clang++ or c++) it 2513 // should autodetect some input files as C++ for g++ compatibility. 2514 if (CCCIsCXX()) { 2515 types::ID OldTy = Ty; 2516 Ty = types::lookupCXXTypeForCType(Ty); 2517 2518 // Do not complain about foo.h, when we are known to be processing 2519 // it as a C++20 header unit. 2520 if (Ty != OldTy && !(OldTy == types::TY_CHeader && hasHeaderMode())) 2521 Diag(clang::diag::warn_drv_treating_input_as_cxx) 2522 << getTypeName(OldTy) << getTypeName(Ty); 2523 } 2524 2525 // If running with -fthinlto-index=, extensions that normally identify 2526 // native object files actually identify LLVM bitcode files. 2527 if (Args.hasArgNoClaim(options::OPT_fthinlto_index_EQ) && 2528 Ty == types::TY_Object) 2529 Ty = types::TY_LLVM_BC; 2530 } 2531 2532 // -ObjC and -ObjC++ override the default language, but only for "source 2533 // files". We just treat everything that isn't a linker input as a 2534 // source file. 2535 // 2536 // FIXME: Clean this up if we move the phase sequence into the type. 2537 if (Ty != types::TY_Object) { 2538 if (Args.hasArg(options::OPT_ObjC)) 2539 Ty = types::TY_ObjC; 2540 else if (Args.hasArg(options::OPT_ObjCXX)) 2541 Ty = types::TY_ObjCXX; 2542 } 2543 2544 // Disambiguate headers that are meant to be header units from those 2545 // intended to be PCH. Avoid missing '.h' cases that are counted as 2546 // C headers by default - we know we are in C++ mode and we do not 2547 // want to issue a complaint about compiling things in the wrong mode. 2548 if ((Ty == types::TY_CXXHeader || Ty == types::TY_CHeader) && 2549 hasHeaderMode()) 2550 Ty = CXXHeaderUnitType(CXX20HeaderType); 2551 } else { 2552 assert(InputTypeArg && "InputType set w/o InputTypeArg"); 2553 if (!InputTypeArg->getOption().matches(options::OPT_x)) { 2554 // If emulating cl.exe, make sure that /TC and /TP don't affect input 2555 // object files. 2556 const char *Ext = strrchr(Value, '.'); 2557 if (Ext && TC.LookupTypeForExtension(Ext + 1) == types::TY_Object) 2558 Ty = types::TY_Object; 2559 } 2560 if (Ty == types::TY_INVALID) { 2561 Ty = InputType; 2562 InputTypeArg->claim(); 2563 } 2564 } 2565 2566 if (DiagnoseInputExistence(Args, Value, Ty, /*TypoCorrect=*/true)) 2567 Inputs.push_back(std::make_pair(Ty, A)); 2568 2569 } else if (A->getOption().matches(options::OPT__SLASH_Tc)) { 2570 StringRef Value = A->getValue(); 2571 if (DiagnoseInputExistence(Args, Value, types::TY_C, 2572 /*TypoCorrect=*/false)) { 2573 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2574 Inputs.push_back(std::make_pair(types::TY_C, InputArg)); 2575 } 2576 A->claim(); 2577 } else if (A->getOption().matches(options::OPT__SLASH_Tp)) { 2578 StringRef Value = A->getValue(); 2579 if (DiagnoseInputExistence(Args, Value, types::TY_CXX, 2580 /*TypoCorrect=*/false)) { 2581 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2582 Inputs.push_back(std::make_pair(types::TY_CXX, InputArg)); 2583 } 2584 A->claim(); 2585 } else if (A->getOption().hasFlag(options::LinkerInput)) { 2586 // Just treat as object type, we could make a special type for this if 2587 // necessary. 2588 Inputs.push_back(std::make_pair(types::TY_Object, A)); 2589 2590 } else if (A->getOption().matches(options::OPT_x)) { 2591 InputTypeArg = A; 2592 InputType = types::lookupTypeForTypeSpecifier(A->getValue()); 2593 A->claim(); 2594 2595 // Follow gcc behavior and treat as linker input for invalid -x 2596 // options. Its not clear why we shouldn't just revert to unknown; but 2597 // this isn't very important, we might as well be bug compatible. 2598 if (!InputType) { 2599 Diag(clang::diag::err_drv_unknown_language) << A->getValue(); 2600 InputType = types::TY_Object; 2601 } 2602 2603 // If the user has put -fmodule-header{,=} then we treat C++ headers as 2604 // header unit inputs. So we 'promote' -xc++-header appropriately. 2605 if (InputType == types::TY_CXXHeader && hasHeaderMode()) 2606 InputType = CXXHeaderUnitType(CXX20HeaderType); 2607 } else if (A->getOption().getID() == options::OPT_U) { 2608 assert(A->getNumValues() == 1 && "The /U option has one value."); 2609 StringRef Val = A->getValue(0); 2610 if (Val.find_first_of("/\\") != StringRef::npos) { 2611 // Warn about e.g. "/Users/me/myfile.c". 2612 Diag(diag::warn_slash_u_filename) << Val; 2613 Diag(diag::note_use_dashdash); 2614 } 2615 } 2616 } 2617 if (CCCIsCPP() && Inputs.empty()) { 2618 // If called as standalone preprocessor, stdin is processed 2619 // if no other input is present. 2620 Arg *A = MakeInputArg(Args, Opts, "-"); 2621 Inputs.push_back(std::make_pair(types::TY_C, A)); 2622 } 2623 } 2624 2625 namespace { 2626 /// Provides a convenient interface for different programming models to generate 2627 /// the required device actions. 2628 class OffloadingActionBuilder final { 2629 /// Flag used to trace errors in the builder. 2630 bool IsValid = false; 2631 2632 /// The compilation that is using this builder. 2633 Compilation &C; 2634 2635 /// Map between an input argument and the offload kinds used to process it. 2636 std::map<const Arg *, unsigned> InputArgToOffloadKindMap; 2637 2638 /// Map between a host action and its originating input argument. 2639 std::map<Action *, const Arg *> HostActionToInputArgMap; 2640 2641 /// Builder interface. It doesn't build anything or keep any state. 2642 class DeviceActionBuilder { 2643 public: 2644 typedef const llvm::SmallVectorImpl<phases::ID> PhasesTy; 2645 2646 enum ActionBuilderReturnCode { 2647 // The builder acted successfully on the current action. 2648 ABRT_Success, 2649 // The builder didn't have to act on the current action. 2650 ABRT_Inactive, 2651 // The builder was successful and requested the host action to not be 2652 // generated. 2653 ABRT_Ignore_Host, 2654 }; 2655 2656 protected: 2657 /// Compilation associated with this builder. 2658 Compilation &C; 2659 2660 /// Tool chains associated with this builder. The same programming 2661 /// model may have associated one or more tool chains. 2662 SmallVector<const ToolChain *, 2> ToolChains; 2663 2664 /// The derived arguments associated with this builder. 2665 DerivedArgList &Args; 2666 2667 /// The inputs associated with this builder. 2668 const Driver::InputList &Inputs; 2669 2670 /// The associated offload kind. 2671 Action::OffloadKind AssociatedOffloadKind = Action::OFK_None; 2672 2673 public: 2674 DeviceActionBuilder(Compilation &C, DerivedArgList &Args, 2675 const Driver::InputList &Inputs, 2676 Action::OffloadKind AssociatedOffloadKind) 2677 : C(C), Args(Args), Inputs(Inputs), 2678 AssociatedOffloadKind(AssociatedOffloadKind) {} 2679 virtual ~DeviceActionBuilder() {} 2680 2681 /// Fill up the array \a DA with all the device dependences that should be 2682 /// added to the provided host action \a HostAction. By default it is 2683 /// inactive. 2684 virtual ActionBuilderReturnCode 2685 getDeviceDependences(OffloadAction::DeviceDependences &DA, 2686 phases::ID CurPhase, phases::ID FinalPhase, 2687 PhasesTy &Phases) { 2688 return ABRT_Inactive; 2689 } 2690 2691 /// Update the state to include the provided host action \a HostAction as a 2692 /// dependency of the current device action. By default it is inactive. 2693 virtual ActionBuilderReturnCode addDeviceDepences(Action *HostAction) { 2694 return ABRT_Inactive; 2695 } 2696 2697 /// Append top level actions generated by the builder. 2698 virtual void appendTopLevelActions(ActionList &AL) {} 2699 2700 /// Append linker device actions generated by the builder. 2701 virtual void appendLinkDeviceActions(ActionList &AL) {} 2702 2703 /// Append linker host action generated by the builder. 2704 virtual Action* appendLinkHostActions(ActionList &AL) { return nullptr; } 2705 2706 /// Append linker actions generated by the builder. 2707 virtual void appendLinkDependences(OffloadAction::DeviceDependences &DA) {} 2708 2709 /// Initialize the builder. Return true if any initialization errors are 2710 /// found. 2711 virtual bool initialize() { return false; } 2712 2713 /// Return true if the builder can use bundling/unbundling. 2714 virtual bool canUseBundlerUnbundler() const { return false; } 2715 2716 /// Return true if this builder is valid. We have a valid builder if we have 2717 /// associated device tool chains. 2718 bool isValid() { return !ToolChains.empty(); } 2719 2720 /// Return the associated offload kind. 2721 Action::OffloadKind getAssociatedOffloadKind() { 2722 return AssociatedOffloadKind; 2723 } 2724 }; 2725 2726 /// Base class for CUDA/HIP action builder. It injects device code in 2727 /// the host backend action. 2728 class CudaActionBuilderBase : public DeviceActionBuilder { 2729 protected: 2730 /// Flags to signal if the user requested host-only or device-only 2731 /// compilation. 2732 bool CompileHostOnly = false; 2733 bool CompileDeviceOnly = false; 2734 bool EmitLLVM = false; 2735 bool EmitAsm = false; 2736 2737 /// ID to identify each device compilation. For CUDA it is simply the 2738 /// GPU arch string. For HIP it is either the GPU arch string or GPU 2739 /// arch string plus feature strings delimited by a plus sign, e.g. 2740 /// gfx906+xnack. 2741 struct TargetID { 2742 /// Target ID string which is persistent throughout the compilation. 2743 const char *ID; 2744 TargetID(CudaArch Arch) { ID = CudaArchToString(Arch); } 2745 TargetID(const char *ID) : ID(ID) {} 2746 operator const char *() { return ID; } 2747 operator StringRef() { return StringRef(ID); } 2748 }; 2749 /// List of GPU architectures to use in this compilation. 2750 SmallVector<TargetID, 4> GpuArchList; 2751 2752 /// The CUDA actions for the current input. 2753 ActionList CudaDeviceActions; 2754 2755 /// The CUDA fat binary if it was generated for the current input. 2756 Action *CudaFatBinary = nullptr; 2757 2758 /// Flag that is set to true if this builder acted on the current input. 2759 bool IsActive = false; 2760 2761 /// Flag for -fgpu-rdc. 2762 bool Relocatable = false; 2763 2764 /// Default GPU architecture if there's no one specified. 2765 CudaArch DefaultCudaArch = CudaArch::UNKNOWN; 2766 2767 /// Method to generate compilation unit ID specified by option 2768 /// '-fuse-cuid='. 2769 enum UseCUIDKind { CUID_Hash, CUID_Random, CUID_None, CUID_Invalid }; 2770 UseCUIDKind UseCUID = CUID_Hash; 2771 2772 /// Compilation unit ID specified by option '-cuid='. 2773 StringRef FixedCUID; 2774 2775 public: 2776 CudaActionBuilderBase(Compilation &C, DerivedArgList &Args, 2777 const Driver::InputList &Inputs, 2778 Action::OffloadKind OFKind) 2779 : DeviceActionBuilder(C, Args, Inputs, OFKind) {} 2780 2781 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 2782 // While generating code for CUDA, we only depend on the host input action 2783 // to trigger the creation of all the CUDA device actions. 2784 2785 // If we are dealing with an input action, replicate it for each GPU 2786 // architecture. If we are in host-only mode we return 'success' so that 2787 // the host uses the CUDA offload kind. 2788 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 2789 assert(!GpuArchList.empty() && 2790 "We should have at least one GPU architecture."); 2791 2792 // If the host input is not CUDA or HIP, we don't need to bother about 2793 // this input. 2794 if (!(IA->getType() == types::TY_CUDA || 2795 IA->getType() == types::TY_HIP || 2796 IA->getType() == types::TY_PP_HIP)) { 2797 // The builder will ignore this input. 2798 IsActive = false; 2799 return ABRT_Inactive; 2800 } 2801 2802 // Set the flag to true, so that the builder acts on the current input. 2803 IsActive = true; 2804 2805 if (CompileHostOnly) 2806 return ABRT_Success; 2807 2808 // Replicate inputs for each GPU architecture. 2809 auto Ty = IA->getType() == types::TY_HIP ? types::TY_HIP_DEVICE 2810 : types::TY_CUDA_DEVICE; 2811 std::string CUID = FixedCUID.str(); 2812 if (CUID.empty()) { 2813 if (UseCUID == CUID_Random) 2814 CUID = llvm::utohexstr(llvm::sys::Process::GetRandomNumber(), 2815 /*LowerCase=*/true); 2816 else if (UseCUID == CUID_Hash) { 2817 llvm::MD5 Hasher; 2818 llvm::MD5::MD5Result Hash; 2819 SmallString<256> RealPath; 2820 llvm::sys::fs::real_path(IA->getInputArg().getValue(), RealPath, 2821 /*expand_tilde=*/true); 2822 Hasher.update(RealPath); 2823 for (auto *A : Args) { 2824 if (A->getOption().matches(options::OPT_INPUT)) 2825 continue; 2826 Hasher.update(A->getAsString(Args)); 2827 } 2828 Hasher.final(Hash); 2829 CUID = llvm::utohexstr(Hash.low(), /*LowerCase=*/true); 2830 } 2831 } 2832 IA->setId(CUID); 2833 2834 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 2835 CudaDeviceActions.push_back( 2836 C.MakeAction<InputAction>(IA->getInputArg(), Ty, IA->getId())); 2837 } 2838 2839 return ABRT_Success; 2840 } 2841 2842 // If this is an unbundling action use it as is for each CUDA toolchain. 2843 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 2844 2845 // If -fgpu-rdc is disabled, should not unbundle since there is no 2846 // device code to link. 2847 if (UA->getType() == types::TY_Object && !Relocatable) 2848 return ABRT_Inactive; 2849 2850 CudaDeviceActions.clear(); 2851 auto *IA = cast<InputAction>(UA->getInputs().back()); 2852 std::string FileName = IA->getInputArg().getAsString(Args); 2853 // Check if the type of the file is the same as the action. Do not 2854 // unbundle it if it is not. Do not unbundle .so files, for example, 2855 // which are not object files. 2856 if (IA->getType() == types::TY_Object && 2857 (!llvm::sys::path::has_extension(FileName) || 2858 types::lookupTypeForExtension( 2859 llvm::sys::path::extension(FileName).drop_front()) != 2860 types::TY_Object)) 2861 return ABRT_Inactive; 2862 2863 for (auto Arch : GpuArchList) { 2864 CudaDeviceActions.push_back(UA); 2865 UA->registerDependentActionInfo(ToolChains[0], Arch, 2866 AssociatedOffloadKind); 2867 } 2868 IsActive = true; 2869 return ABRT_Success; 2870 } 2871 2872 return IsActive ? ABRT_Success : ABRT_Inactive; 2873 } 2874 2875 void appendTopLevelActions(ActionList &AL) override { 2876 // Utility to append actions to the top level list. 2877 auto AddTopLevel = [&](Action *A, TargetID TargetID) { 2878 OffloadAction::DeviceDependences Dep; 2879 Dep.add(*A, *ToolChains.front(), TargetID, AssociatedOffloadKind); 2880 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 2881 }; 2882 2883 // If we have a fat binary, add it to the list. 2884 if (CudaFatBinary) { 2885 AddTopLevel(CudaFatBinary, CudaArch::UNUSED); 2886 CudaDeviceActions.clear(); 2887 CudaFatBinary = nullptr; 2888 return; 2889 } 2890 2891 if (CudaDeviceActions.empty()) 2892 return; 2893 2894 // If we have CUDA actions at this point, that's because we have a have 2895 // partial compilation, so we should have an action for each GPU 2896 // architecture. 2897 assert(CudaDeviceActions.size() == GpuArchList.size() && 2898 "Expecting one action per GPU architecture."); 2899 assert(ToolChains.size() == 1 && 2900 "Expecting to have a single CUDA toolchain."); 2901 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) 2902 AddTopLevel(CudaDeviceActions[I], GpuArchList[I]); 2903 2904 CudaDeviceActions.clear(); 2905 } 2906 2907 /// Get canonicalized offload arch option. \returns empty StringRef if the 2908 /// option is invalid. 2909 virtual StringRef getCanonicalOffloadArch(StringRef Arch) = 0; 2910 2911 virtual llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 2912 getConflictOffloadArchCombination(const std::set<StringRef> &GpuArchs) = 0; 2913 2914 bool initialize() override { 2915 assert(AssociatedOffloadKind == Action::OFK_Cuda || 2916 AssociatedOffloadKind == Action::OFK_HIP); 2917 2918 // We don't need to support CUDA. 2919 if (AssociatedOffloadKind == Action::OFK_Cuda && 2920 !C.hasOffloadToolChain<Action::OFK_Cuda>()) 2921 return false; 2922 2923 // We don't need to support HIP. 2924 if (AssociatedOffloadKind == Action::OFK_HIP && 2925 !C.hasOffloadToolChain<Action::OFK_HIP>()) 2926 return false; 2927 2928 Relocatable = Args.hasFlag(options::OPT_fgpu_rdc, 2929 options::OPT_fno_gpu_rdc, /*Default=*/false); 2930 2931 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 2932 assert(HostTC && "No toolchain for host compilation."); 2933 if (HostTC->getTriple().isNVPTX() || 2934 HostTC->getTriple().getArch() == llvm::Triple::amdgcn) { 2935 // We do not support targeting NVPTX/AMDGCN for host compilation. Throw 2936 // an error and abort pipeline construction early so we don't trip 2937 // asserts that assume device-side compilation. 2938 C.getDriver().Diag(diag::err_drv_cuda_host_arch) 2939 << HostTC->getTriple().getArchName(); 2940 return true; 2941 } 2942 2943 ToolChains.push_back( 2944 AssociatedOffloadKind == Action::OFK_Cuda 2945 ? C.getSingleOffloadToolChain<Action::OFK_Cuda>() 2946 : C.getSingleOffloadToolChain<Action::OFK_HIP>()); 2947 2948 CompileHostOnly = C.getDriver().offloadHostOnly(); 2949 CompileDeviceOnly = C.getDriver().offloadDeviceOnly(); 2950 EmitLLVM = Args.getLastArg(options::OPT_emit_llvm); 2951 EmitAsm = Args.getLastArg(options::OPT_S); 2952 FixedCUID = Args.getLastArgValue(options::OPT_cuid_EQ); 2953 if (Arg *A = Args.getLastArg(options::OPT_fuse_cuid_EQ)) { 2954 StringRef UseCUIDStr = A->getValue(); 2955 UseCUID = llvm::StringSwitch<UseCUIDKind>(UseCUIDStr) 2956 .Case("hash", CUID_Hash) 2957 .Case("random", CUID_Random) 2958 .Case("none", CUID_None) 2959 .Default(CUID_Invalid); 2960 if (UseCUID == CUID_Invalid) { 2961 C.getDriver().Diag(diag::err_drv_invalid_value) 2962 << A->getAsString(Args) << UseCUIDStr; 2963 C.setContainsError(); 2964 return true; 2965 } 2966 } 2967 2968 // --offload and --offload-arch options are mutually exclusive. 2969 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 2970 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 2971 options::OPT_no_offload_arch_EQ)) { 2972 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) << "--offload-arch" 2973 << "--offload"; 2974 } 2975 2976 // Collect all cuda_gpu_arch parameters, removing duplicates. 2977 std::set<StringRef> GpuArchs; 2978 bool Error = false; 2979 for (Arg *A : Args) { 2980 if (!(A->getOption().matches(options::OPT_offload_arch_EQ) || 2981 A->getOption().matches(options::OPT_no_offload_arch_EQ))) 2982 continue; 2983 A->claim(); 2984 2985 StringRef ArchStr = A->getValue(); 2986 if (A->getOption().matches(options::OPT_no_offload_arch_EQ) && 2987 ArchStr == "all") { 2988 GpuArchs.clear(); 2989 continue; 2990 } 2991 ArchStr = getCanonicalOffloadArch(ArchStr); 2992 if (ArchStr.empty()) { 2993 Error = true; 2994 } else if (A->getOption().matches(options::OPT_offload_arch_EQ)) 2995 GpuArchs.insert(ArchStr); 2996 else if (A->getOption().matches(options::OPT_no_offload_arch_EQ)) 2997 GpuArchs.erase(ArchStr); 2998 else 2999 llvm_unreachable("Unexpected option."); 3000 } 3001 3002 auto &&ConflictingArchs = getConflictOffloadArchCombination(GpuArchs); 3003 if (ConflictingArchs) { 3004 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 3005 << ConflictingArchs.getValue().first 3006 << ConflictingArchs.getValue().second; 3007 C.setContainsError(); 3008 return true; 3009 } 3010 3011 // Collect list of GPUs remaining in the set. 3012 for (auto Arch : GpuArchs) 3013 GpuArchList.push_back(Arch.data()); 3014 3015 // Default to sm_20 which is the lowest common denominator for 3016 // supported GPUs. sm_20 code should work correctly, if 3017 // suboptimally, on all newer GPUs. 3018 if (GpuArchList.empty()) { 3019 if (ToolChains.front()->getTriple().isSPIRV()) 3020 GpuArchList.push_back(CudaArch::Generic); 3021 else 3022 GpuArchList.push_back(DefaultCudaArch); 3023 } 3024 3025 return Error; 3026 } 3027 }; 3028 3029 /// \brief CUDA action builder. It injects device code in the host backend 3030 /// action. 3031 class CudaActionBuilder final : public CudaActionBuilderBase { 3032 public: 3033 CudaActionBuilder(Compilation &C, DerivedArgList &Args, 3034 const Driver::InputList &Inputs) 3035 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) { 3036 DefaultCudaArch = CudaArch::SM_35; 3037 } 3038 3039 StringRef getCanonicalOffloadArch(StringRef ArchStr) override { 3040 CudaArch Arch = StringToCudaArch(ArchStr); 3041 if (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch)) { 3042 C.getDriver().Diag(clang::diag::err_drv_cuda_bad_gpu_arch) << ArchStr; 3043 return StringRef(); 3044 } 3045 return CudaArchToString(Arch); 3046 } 3047 3048 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3049 getConflictOffloadArchCombination( 3050 const std::set<StringRef> &GpuArchs) override { 3051 return llvm::None; 3052 } 3053 3054 ActionBuilderReturnCode 3055 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3056 phases::ID CurPhase, phases::ID FinalPhase, 3057 PhasesTy &Phases) override { 3058 if (!IsActive) 3059 return ABRT_Inactive; 3060 3061 // If we don't have more CUDA actions, we don't have any dependences to 3062 // create for the host. 3063 if (CudaDeviceActions.empty()) 3064 return ABRT_Success; 3065 3066 assert(CudaDeviceActions.size() == GpuArchList.size() && 3067 "Expecting one action per GPU architecture."); 3068 assert(!CompileHostOnly && 3069 "Not expecting CUDA actions in host-only compilation."); 3070 3071 // If we are generating code for the device or we are in a backend phase, 3072 // we attempt to generate the fat binary. We compile each arch to ptx and 3073 // assemble to cubin, then feed the cubin *and* the ptx into a device 3074 // "link" action, which uses fatbinary to combine these cubins into one 3075 // fatbin. The fatbin is then an input to the host action if not in 3076 // device-only mode. 3077 if (CompileDeviceOnly || CurPhase == phases::Backend) { 3078 ActionList DeviceActions; 3079 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3080 // Produce the device action from the current phase up to the assemble 3081 // phase. 3082 for (auto Ph : Phases) { 3083 // Skip the phases that were already dealt with. 3084 if (Ph < CurPhase) 3085 continue; 3086 // We have to be consistent with the host final phase. 3087 if (Ph > FinalPhase) 3088 break; 3089 3090 CudaDeviceActions[I] = C.getDriver().ConstructPhaseAction( 3091 C, Args, Ph, CudaDeviceActions[I], Action::OFK_Cuda); 3092 3093 if (Ph == phases::Assemble) 3094 break; 3095 } 3096 3097 // If we didn't reach the assemble phase, we can't generate the fat 3098 // binary. We don't need to generate the fat binary if we are not in 3099 // device-only mode. 3100 if (!isa<AssembleJobAction>(CudaDeviceActions[I]) || 3101 CompileDeviceOnly) 3102 continue; 3103 3104 Action *AssembleAction = CudaDeviceActions[I]; 3105 assert(AssembleAction->getType() == types::TY_Object); 3106 assert(AssembleAction->getInputs().size() == 1); 3107 3108 Action *BackendAction = AssembleAction->getInputs()[0]; 3109 assert(BackendAction->getType() == types::TY_PP_Asm); 3110 3111 for (auto &A : {AssembleAction, BackendAction}) { 3112 OffloadAction::DeviceDependences DDep; 3113 DDep.add(*A, *ToolChains.front(), GpuArchList[I], Action::OFK_Cuda); 3114 DeviceActions.push_back( 3115 C.MakeAction<OffloadAction>(DDep, A->getType())); 3116 } 3117 } 3118 3119 // We generate the fat binary if we have device input actions. 3120 if (!DeviceActions.empty()) { 3121 CudaFatBinary = 3122 C.MakeAction<LinkJobAction>(DeviceActions, types::TY_CUDA_FATBIN); 3123 3124 if (!CompileDeviceOnly) { 3125 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3126 Action::OFK_Cuda); 3127 // Clear the fat binary, it is already a dependence to an host 3128 // action. 3129 CudaFatBinary = nullptr; 3130 } 3131 3132 // Remove the CUDA actions as they are already connected to an host 3133 // action or fat binary. 3134 CudaDeviceActions.clear(); 3135 } 3136 3137 // We avoid creating host action in device-only mode. 3138 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3139 } else if (CurPhase > phases::Backend) { 3140 // If we are past the backend phase and still have a device action, we 3141 // don't have to do anything as this action is already a device 3142 // top-level action. 3143 return ABRT_Success; 3144 } 3145 3146 assert(CurPhase < phases::Backend && "Generating single CUDA " 3147 "instructions should only occur " 3148 "before the backend phase!"); 3149 3150 // By default, we produce an action for each device arch. 3151 for (Action *&A : CudaDeviceActions) 3152 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3153 3154 return ABRT_Success; 3155 } 3156 }; 3157 /// \brief HIP action builder. It injects device code in the host backend 3158 /// action. 3159 class HIPActionBuilder final : public CudaActionBuilderBase { 3160 /// The linker inputs obtained for each device arch. 3161 SmallVector<ActionList, 8> DeviceLinkerInputs; 3162 // The default bundling behavior depends on the type of output, therefore 3163 // BundleOutput needs to be tri-value: None, true, or false. 3164 // Bundle code objects except --no-gpu-output is specified for device 3165 // only compilation. Bundle other type of output files only if 3166 // --gpu-bundle-output is specified for device only compilation. 3167 Optional<bool> BundleOutput; 3168 3169 public: 3170 HIPActionBuilder(Compilation &C, DerivedArgList &Args, 3171 const Driver::InputList &Inputs) 3172 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) { 3173 DefaultCudaArch = CudaArch::GFX803; 3174 if (Args.hasArg(options::OPT_gpu_bundle_output, 3175 options::OPT_no_gpu_bundle_output)) 3176 BundleOutput = Args.hasFlag(options::OPT_gpu_bundle_output, 3177 options::OPT_no_gpu_bundle_output, true); 3178 } 3179 3180 bool canUseBundlerUnbundler() const override { return true; } 3181 3182 StringRef getCanonicalOffloadArch(StringRef IdStr) override { 3183 llvm::StringMap<bool> Features; 3184 // getHIPOffloadTargetTriple() is known to return valid value as it has 3185 // been called successfully in the CreateOffloadingDeviceToolChains(). 3186 auto ArchStr = parseTargetID( 3187 *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()), IdStr, 3188 &Features); 3189 if (!ArchStr) { 3190 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << IdStr; 3191 C.setContainsError(); 3192 return StringRef(); 3193 } 3194 auto CanId = getCanonicalTargetID(ArchStr.getValue(), Features); 3195 return Args.MakeArgStringRef(CanId); 3196 }; 3197 3198 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3199 getConflictOffloadArchCombination( 3200 const std::set<StringRef> &GpuArchs) override { 3201 return getConflictTargetIDCombination(GpuArchs); 3202 } 3203 3204 ActionBuilderReturnCode 3205 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3206 phases::ID CurPhase, phases::ID FinalPhase, 3207 PhasesTy &Phases) override { 3208 if (!IsActive) 3209 return ABRT_Inactive; 3210 3211 // amdgcn does not support linking of object files, therefore we skip 3212 // backend and assemble phases to output LLVM IR. Except for generating 3213 // non-relocatable device code, where we generate fat binary for device 3214 // code and pass to host in Backend phase. 3215 if (CudaDeviceActions.empty()) 3216 return ABRT_Success; 3217 3218 assert(((CurPhase == phases::Link && Relocatable) || 3219 CudaDeviceActions.size() == GpuArchList.size()) && 3220 "Expecting one action per GPU architecture."); 3221 assert(!CompileHostOnly && 3222 "Not expecting HIP actions in host-only compilation."); 3223 3224 if (!Relocatable && CurPhase == phases::Backend && !EmitLLVM && 3225 !EmitAsm) { 3226 // If we are in backend phase, we attempt to generate the fat binary. 3227 // We compile each arch to IR and use a link action to generate code 3228 // object containing ISA. Then we use a special "link" action to create 3229 // a fat binary containing all the code objects for different GPU's. 3230 // The fat binary is then an input to the host action. 3231 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3232 if (C.getDriver().isUsingLTO(/*IsOffload=*/true)) { 3233 // When LTO is enabled, skip the backend and assemble phases and 3234 // use lld to link the bitcode. 3235 ActionList AL; 3236 AL.push_back(CudaDeviceActions[I]); 3237 // Create a link action to link device IR with device library 3238 // and generate ISA. 3239 CudaDeviceActions[I] = 3240 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3241 } else { 3242 // When LTO is not enabled, we follow the conventional 3243 // compiler phases, including backend and assemble phases. 3244 ActionList AL; 3245 Action *BackendAction = nullptr; 3246 if (ToolChains.front()->getTriple().isSPIRV()) { 3247 // Emit LLVM bitcode for SPIR-V targets. SPIR-V device tool chain 3248 // (HIPSPVToolChain) runs post-link LLVM IR passes. 3249 types::ID Output = Args.hasArg(options::OPT_S) 3250 ? types::TY_LLVM_IR 3251 : types::TY_LLVM_BC; 3252 BackendAction = 3253 C.MakeAction<BackendJobAction>(CudaDeviceActions[I], Output); 3254 } else 3255 BackendAction = C.getDriver().ConstructPhaseAction( 3256 C, Args, phases::Backend, CudaDeviceActions[I], 3257 AssociatedOffloadKind); 3258 auto AssembleAction = C.getDriver().ConstructPhaseAction( 3259 C, Args, phases::Assemble, BackendAction, 3260 AssociatedOffloadKind); 3261 AL.push_back(AssembleAction); 3262 // Create a link action to link device IR with device library 3263 // and generate ISA. 3264 CudaDeviceActions[I] = 3265 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3266 } 3267 3268 // OffloadingActionBuilder propagates device arch until an offload 3269 // action. Since the next action for creating fatbin does 3270 // not have device arch, whereas the above link action and its input 3271 // have device arch, an offload action is needed to stop the null 3272 // device arch of the next action being propagated to the above link 3273 // action. 3274 OffloadAction::DeviceDependences DDep; 3275 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3276 AssociatedOffloadKind); 3277 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3278 DDep, CudaDeviceActions[I]->getType()); 3279 } 3280 3281 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3282 BundleOutput.getValue()) { 3283 // Create HIP fat binary with a special "link" action. 3284 CudaFatBinary = C.MakeAction<LinkJobAction>(CudaDeviceActions, 3285 types::TY_HIP_FATBIN); 3286 3287 if (!CompileDeviceOnly) { 3288 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3289 AssociatedOffloadKind); 3290 // Clear the fat binary, it is already a dependence to an host 3291 // action. 3292 CudaFatBinary = nullptr; 3293 } 3294 3295 // Remove the CUDA actions as they are already connected to an host 3296 // action or fat binary. 3297 CudaDeviceActions.clear(); 3298 } 3299 3300 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3301 } else if (CurPhase == phases::Link) { 3302 // Save CudaDeviceActions to DeviceLinkerInputs for each GPU subarch. 3303 // This happens to each device action originated from each input file. 3304 // Later on, device actions in DeviceLinkerInputs are used to create 3305 // device link actions in appendLinkDependences and the created device 3306 // link actions are passed to the offload action as device dependence. 3307 DeviceLinkerInputs.resize(CudaDeviceActions.size()); 3308 auto LI = DeviceLinkerInputs.begin(); 3309 for (auto *A : CudaDeviceActions) { 3310 LI->push_back(A); 3311 ++LI; 3312 } 3313 3314 // We will pass the device action as a host dependence, so we don't 3315 // need to do anything else with them. 3316 CudaDeviceActions.clear(); 3317 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3318 } 3319 3320 // By default, we produce an action for each device arch. 3321 for (Action *&A : CudaDeviceActions) 3322 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A, 3323 AssociatedOffloadKind); 3324 3325 if (CompileDeviceOnly && CurPhase == FinalPhase && 3326 BundleOutput.hasValue() && BundleOutput.getValue()) { 3327 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3328 OffloadAction::DeviceDependences DDep; 3329 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3330 AssociatedOffloadKind); 3331 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3332 DDep, CudaDeviceActions[I]->getType()); 3333 } 3334 CudaFatBinary = 3335 C.MakeAction<OffloadBundlingJobAction>(CudaDeviceActions); 3336 CudaDeviceActions.clear(); 3337 } 3338 3339 return (CompileDeviceOnly && CurPhase == FinalPhase) ? ABRT_Ignore_Host 3340 : ABRT_Success; 3341 } 3342 3343 void appendLinkDeviceActions(ActionList &AL) override { 3344 if (DeviceLinkerInputs.size() == 0) 3345 return; 3346 3347 assert(DeviceLinkerInputs.size() == GpuArchList.size() && 3348 "Linker inputs and GPU arch list sizes do not match."); 3349 3350 ActionList Actions; 3351 unsigned I = 0; 3352 // Append a new link action for each device. 3353 // Each entry in DeviceLinkerInputs corresponds to a GPU arch. 3354 for (auto &LI : DeviceLinkerInputs) { 3355 3356 types::ID Output = Args.hasArg(options::OPT_emit_llvm) 3357 ? types::TY_LLVM_BC 3358 : types::TY_Image; 3359 3360 auto *DeviceLinkAction = C.MakeAction<LinkJobAction>(LI, Output); 3361 // Linking all inputs for the current GPU arch. 3362 // LI contains all the inputs for the linker. 3363 OffloadAction::DeviceDependences DeviceLinkDeps; 3364 DeviceLinkDeps.add(*DeviceLinkAction, *ToolChains[0], 3365 GpuArchList[I], AssociatedOffloadKind); 3366 Actions.push_back(C.MakeAction<OffloadAction>( 3367 DeviceLinkDeps, DeviceLinkAction->getType())); 3368 ++I; 3369 } 3370 DeviceLinkerInputs.clear(); 3371 3372 // If emitting LLVM, do not generate final host/device compilation action 3373 if (Args.hasArg(options::OPT_emit_llvm)) { 3374 AL.append(Actions); 3375 return; 3376 } 3377 3378 // Create a host object from all the device images by embedding them 3379 // in a fat binary for mixed host-device compilation. For device-only 3380 // compilation, creates a fat binary. 3381 OffloadAction::DeviceDependences DDeps; 3382 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3383 BundleOutput.getValue()) { 3384 auto *TopDeviceLinkAction = C.MakeAction<LinkJobAction>( 3385 Actions, 3386 CompileDeviceOnly ? types::TY_HIP_FATBIN : types::TY_Object); 3387 DDeps.add(*TopDeviceLinkAction, *ToolChains[0], nullptr, 3388 AssociatedOffloadKind); 3389 // Offload the host object to the host linker. 3390 AL.push_back( 3391 C.MakeAction<OffloadAction>(DDeps, TopDeviceLinkAction->getType())); 3392 } else { 3393 AL.append(Actions); 3394 } 3395 } 3396 3397 Action* appendLinkHostActions(ActionList &AL) override { return AL.back(); } 3398 3399 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3400 }; 3401 3402 /// OpenMP action builder. The host bitcode is passed to the device frontend 3403 /// and all the device linked images are passed to the host link phase. 3404 class OpenMPActionBuilder final : public DeviceActionBuilder { 3405 /// The OpenMP actions for the current input. 3406 ActionList OpenMPDeviceActions; 3407 3408 /// The linker inputs obtained for each toolchain. 3409 SmallVector<ActionList, 8> DeviceLinkerInputs; 3410 3411 public: 3412 OpenMPActionBuilder(Compilation &C, DerivedArgList &Args, 3413 const Driver::InputList &Inputs) 3414 : DeviceActionBuilder(C, Args, Inputs, Action::OFK_OpenMP) {} 3415 3416 ActionBuilderReturnCode 3417 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3418 phases::ID CurPhase, phases::ID FinalPhase, 3419 PhasesTy &Phases) override { 3420 if (OpenMPDeviceActions.empty()) 3421 return ABRT_Inactive; 3422 3423 // We should always have an action for each input. 3424 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3425 "Number of OpenMP actions and toolchains do not match."); 3426 3427 // The host only depends on device action in the linking phase, when all 3428 // the device images have to be embedded in the host image. 3429 if (CurPhase == phases::Link) { 3430 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3431 "Toolchains and linker inputs sizes do not match."); 3432 auto LI = DeviceLinkerInputs.begin(); 3433 for (auto *A : OpenMPDeviceActions) { 3434 LI->push_back(A); 3435 ++LI; 3436 } 3437 3438 // We passed the device action as a host dependence, so we don't need to 3439 // do anything else with them. 3440 OpenMPDeviceActions.clear(); 3441 return ABRT_Success; 3442 } 3443 3444 // By default, we produce an action for each device arch. 3445 for (Action *&A : OpenMPDeviceActions) 3446 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3447 3448 return ABRT_Success; 3449 } 3450 3451 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 3452 3453 // If this is an input action replicate it for each OpenMP toolchain. 3454 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 3455 OpenMPDeviceActions.clear(); 3456 for (unsigned I = 0; I < ToolChains.size(); ++I) 3457 OpenMPDeviceActions.push_back( 3458 C.MakeAction<InputAction>(IA->getInputArg(), IA->getType())); 3459 return ABRT_Success; 3460 } 3461 3462 // If this is an unbundling action use it as is for each OpenMP toolchain. 3463 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 3464 OpenMPDeviceActions.clear(); 3465 auto *IA = cast<InputAction>(UA->getInputs().back()); 3466 std::string FileName = IA->getInputArg().getAsString(Args); 3467 // Check if the type of the file is the same as the action. Do not 3468 // unbundle it if it is not. Do not unbundle .so files, for example, 3469 // which are not object files. 3470 if (IA->getType() == types::TY_Object && 3471 (!llvm::sys::path::has_extension(FileName) || 3472 types::lookupTypeForExtension( 3473 llvm::sys::path::extension(FileName).drop_front()) != 3474 types::TY_Object)) 3475 return ABRT_Inactive; 3476 for (unsigned I = 0; I < ToolChains.size(); ++I) { 3477 OpenMPDeviceActions.push_back(UA); 3478 UA->registerDependentActionInfo( 3479 ToolChains[I], /*BoundArch=*/StringRef(), Action::OFK_OpenMP); 3480 } 3481 return ABRT_Success; 3482 } 3483 3484 // When generating code for OpenMP we use the host compile phase result as 3485 // a dependence to the device compile phase so that it can learn what 3486 // declarations should be emitted. However, this is not the only use for 3487 // the host action, so we prevent it from being collapsed. 3488 if (isa<CompileJobAction>(HostAction)) { 3489 HostAction->setCannotBeCollapsedWithNextDependentAction(); 3490 assert(ToolChains.size() == OpenMPDeviceActions.size() && 3491 "Toolchains and device action sizes do not match."); 3492 OffloadAction::HostDependence HDep( 3493 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3494 /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3495 auto TC = ToolChains.begin(); 3496 for (Action *&A : OpenMPDeviceActions) { 3497 assert(isa<CompileJobAction>(A)); 3498 OffloadAction::DeviceDependences DDep; 3499 DDep.add(*A, **TC, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3500 A = C.MakeAction<OffloadAction>(HDep, DDep); 3501 ++TC; 3502 } 3503 } 3504 return ABRT_Success; 3505 } 3506 3507 void appendTopLevelActions(ActionList &AL) override { 3508 if (OpenMPDeviceActions.empty()) 3509 return; 3510 3511 // We should always have an action for each input. 3512 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3513 "Number of OpenMP actions and toolchains do not match."); 3514 3515 // Append all device actions followed by the proper offload action. 3516 auto TI = ToolChains.begin(); 3517 for (auto *A : OpenMPDeviceActions) { 3518 OffloadAction::DeviceDependences Dep; 3519 Dep.add(*A, **TI, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3520 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 3521 ++TI; 3522 } 3523 // We no longer need the action stored in this builder. 3524 OpenMPDeviceActions.clear(); 3525 } 3526 3527 void appendLinkDeviceActions(ActionList &AL) override { 3528 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3529 "Toolchains and linker inputs sizes do not match."); 3530 3531 // Append a new link action for each device. 3532 auto TC = ToolChains.begin(); 3533 for (auto &LI : DeviceLinkerInputs) { 3534 auto *DeviceLinkAction = 3535 C.MakeAction<LinkJobAction>(LI, types::TY_Image); 3536 OffloadAction::DeviceDependences DeviceLinkDeps; 3537 DeviceLinkDeps.add(*DeviceLinkAction, **TC, /*BoundArch=*/nullptr, 3538 Action::OFK_OpenMP); 3539 AL.push_back(C.MakeAction<OffloadAction>(DeviceLinkDeps, 3540 DeviceLinkAction->getType())); 3541 ++TC; 3542 } 3543 DeviceLinkerInputs.clear(); 3544 } 3545 3546 Action* appendLinkHostActions(ActionList &AL) override { 3547 // Create wrapper bitcode from the result of device link actions and compile 3548 // it to an object which will be added to the host link command. 3549 auto *BC = C.MakeAction<OffloadWrapperJobAction>(AL, types::TY_LLVM_BC); 3550 auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm); 3551 return C.MakeAction<AssembleJobAction>(ASM, types::TY_Object); 3552 } 3553 3554 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3555 3556 bool initialize() override { 3557 // Get the OpenMP toolchains. If we don't get any, the action builder will 3558 // know there is nothing to do related to OpenMP offloading. 3559 auto OpenMPTCRange = C.getOffloadToolChains<Action::OFK_OpenMP>(); 3560 for (auto TI = OpenMPTCRange.first, TE = OpenMPTCRange.second; TI != TE; 3561 ++TI) 3562 ToolChains.push_back(TI->second); 3563 3564 DeviceLinkerInputs.resize(ToolChains.size()); 3565 return false; 3566 } 3567 3568 bool canUseBundlerUnbundler() const override { 3569 // OpenMP should use bundled files whenever possible. 3570 return true; 3571 } 3572 }; 3573 3574 /// 3575 /// TODO: Add the implementation for other specialized builders here. 3576 /// 3577 3578 /// Specialized builders being used by this offloading action builder. 3579 SmallVector<DeviceActionBuilder *, 4> SpecializedBuilders; 3580 3581 /// Flag set to true if all valid builders allow file bundling/unbundling. 3582 bool CanUseBundler; 3583 3584 public: 3585 OffloadingActionBuilder(Compilation &C, DerivedArgList &Args, 3586 const Driver::InputList &Inputs) 3587 : C(C) { 3588 // Create a specialized builder for each device toolchain. 3589 3590 IsValid = true; 3591 3592 // Create a specialized builder for CUDA. 3593 SpecializedBuilders.push_back(new CudaActionBuilder(C, Args, Inputs)); 3594 3595 // Create a specialized builder for HIP. 3596 SpecializedBuilders.push_back(new HIPActionBuilder(C, Args, Inputs)); 3597 3598 // Create a specialized builder for OpenMP. 3599 SpecializedBuilders.push_back(new OpenMPActionBuilder(C, Args, Inputs)); 3600 3601 // 3602 // TODO: Build other specialized builders here. 3603 // 3604 3605 // Initialize all the builders, keeping track of errors. If all valid 3606 // builders agree that we can use bundling, set the flag to true. 3607 unsigned ValidBuilders = 0u; 3608 unsigned ValidBuildersSupportingBundling = 0u; 3609 for (auto *SB : SpecializedBuilders) { 3610 IsValid = IsValid && !SB->initialize(); 3611 3612 // Update the counters if the builder is valid. 3613 if (SB->isValid()) { 3614 ++ValidBuilders; 3615 if (SB->canUseBundlerUnbundler()) 3616 ++ValidBuildersSupportingBundling; 3617 } 3618 } 3619 CanUseBundler = 3620 ValidBuilders && ValidBuilders == ValidBuildersSupportingBundling; 3621 } 3622 3623 ~OffloadingActionBuilder() { 3624 for (auto *SB : SpecializedBuilders) 3625 delete SB; 3626 } 3627 3628 /// Record a host action and its originating input argument. 3629 void recordHostAction(Action *HostAction, const Arg *InputArg) { 3630 assert(HostAction && "Invalid host action"); 3631 assert(InputArg && "Invalid input argument"); 3632 auto Loc = HostActionToInputArgMap.find(HostAction); 3633 if (Loc == HostActionToInputArgMap.end()) 3634 HostActionToInputArgMap[HostAction] = InputArg; 3635 assert(HostActionToInputArgMap[HostAction] == InputArg && 3636 "host action mapped to multiple input arguments"); 3637 } 3638 3639 /// Generate an action that adds device dependences (if any) to a host action. 3640 /// If no device dependence actions exist, just return the host action \a 3641 /// HostAction. If an error is found or if no builder requires the host action 3642 /// to be generated, return nullptr. 3643 Action * 3644 addDeviceDependencesToHostAction(Action *HostAction, const Arg *InputArg, 3645 phases::ID CurPhase, phases::ID FinalPhase, 3646 DeviceActionBuilder::PhasesTy &Phases) { 3647 if (!IsValid) 3648 return nullptr; 3649 3650 if (SpecializedBuilders.empty()) 3651 return HostAction; 3652 3653 assert(HostAction && "Invalid host action!"); 3654 recordHostAction(HostAction, InputArg); 3655 3656 OffloadAction::DeviceDependences DDeps; 3657 // Check if all the programming models agree we should not emit the host 3658 // action. Also, keep track of the offloading kinds employed. 3659 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3660 unsigned InactiveBuilders = 0u; 3661 unsigned IgnoringBuilders = 0u; 3662 for (auto *SB : SpecializedBuilders) { 3663 if (!SB->isValid()) { 3664 ++InactiveBuilders; 3665 continue; 3666 } 3667 3668 auto RetCode = 3669 SB->getDeviceDependences(DDeps, CurPhase, FinalPhase, Phases); 3670 3671 // If the builder explicitly says the host action should be ignored, 3672 // we need to increment the variable that tracks the builders that request 3673 // the host object to be ignored. 3674 if (RetCode == DeviceActionBuilder::ABRT_Ignore_Host) 3675 ++IgnoringBuilders; 3676 3677 // Unless the builder was inactive for this action, we have to record the 3678 // offload kind because the host will have to use it. 3679 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3680 OffloadKind |= SB->getAssociatedOffloadKind(); 3681 } 3682 3683 // If all builders agree that the host object should be ignored, just return 3684 // nullptr. 3685 if (IgnoringBuilders && 3686 SpecializedBuilders.size() == (InactiveBuilders + IgnoringBuilders)) 3687 return nullptr; 3688 3689 if (DDeps.getActions().empty()) 3690 return HostAction; 3691 3692 // We have dependences we need to bundle together. We use an offload action 3693 // for that. 3694 OffloadAction::HostDependence HDep( 3695 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3696 /*BoundArch=*/nullptr, DDeps); 3697 return C.MakeAction<OffloadAction>(HDep, DDeps); 3698 } 3699 3700 /// Generate an action that adds a host dependence to a device action. The 3701 /// results will be kept in this action builder. Return true if an error was 3702 /// found. 3703 bool addHostDependenceToDeviceActions(Action *&HostAction, 3704 const Arg *InputArg) { 3705 if (!IsValid) 3706 return true; 3707 3708 recordHostAction(HostAction, InputArg); 3709 3710 // If we are supporting bundling/unbundling and the current action is an 3711 // input action of non-source file, we replace the host action by the 3712 // unbundling action. The bundler tool has the logic to detect if an input 3713 // is a bundle or not and if the input is not a bundle it assumes it is a 3714 // host file. Therefore it is safe to create an unbundling action even if 3715 // the input is not a bundle. 3716 if (CanUseBundler && isa<InputAction>(HostAction) && 3717 InputArg->getOption().getKind() == llvm::opt::Option::InputClass && 3718 (!types::isSrcFile(HostAction->getType()) || 3719 HostAction->getType() == types::TY_PP_HIP)) { 3720 auto UnbundlingHostAction = 3721 C.MakeAction<OffloadUnbundlingJobAction>(HostAction); 3722 UnbundlingHostAction->registerDependentActionInfo( 3723 C.getSingleOffloadToolChain<Action::OFK_Host>(), 3724 /*BoundArch=*/StringRef(), Action::OFK_Host); 3725 HostAction = UnbundlingHostAction; 3726 recordHostAction(HostAction, InputArg); 3727 } 3728 3729 assert(HostAction && "Invalid host action!"); 3730 3731 // Register the offload kinds that are used. 3732 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3733 for (auto *SB : SpecializedBuilders) { 3734 if (!SB->isValid()) 3735 continue; 3736 3737 auto RetCode = SB->addDeviceDepences(HostAction); 3738 3739 // Host dependences for device actions are not compatible with that same 3740 // action being ignored. 3741 assert(RetCode != DeviceActionBuilder::ABRT_Ignore_Host && 3742 "Host dependence not expected to be ignored.!"); 3743 3744 // Unless the builder was inactive for this action, we have to record the 3745 // offload kind because the host will have to use it. 3746 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3747 OffloadKind |= SB->getAssociatedOffloadKind(); 3748 } 3749 3750 // Do not use unbundler if the Host does not depend on device action. 3751 if (OffloadKind == Action::OFK_None && CanUseBundler) 3752 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) 3753 HostAction = UA->getInputs().back(); 3754 3755 return false; 3756 } 3757 3758 /// Add the offloading top level actions to the provided action list. This 3759 /// function can replace the host action by a bundling action if the 3760 /// programming models allow it. 3761 bool appendTopLevelActions(ActionList &AL, Action *HostAction, 3762 const Arg *InputArg) { 3763 if (HostAction) 3764 recordHostAction(HostAction, InputArg); 3765 3766 // Get the device actions to be appended. 3767 ActionList OffloadAL; 3768 for (auto *SB : SpecializedBuilders) { 3769 if (!SB->isValid()) 3770 continue; 3771 SB->appendTopLevelActions(OffloadAL); 3772 } 3773 3774 // If we can use the bundler, replace the host action by the bundling one in 3775 // the resulting list. Otherwise, just append the device actions. For 3776 // device only compilation, HostAction is a null pointer, therefore only do 3777 // this when HostAction is not a null pointer. 3778 if (CanUseBundler && HostAction && 3779 HostAction->getType() != types::TY_Nothing && !OffloadAL.empty()) { 3780 // Add the host action to the list in order to create the bundling action. 3781 OffloadAL.push_back(HostAction); 3782 3783 // We expect that the host action was just appended to the action list 3784 // before this method was called. 3785 assert(HostAction == AL.back() && "Host action not in the list??"); 3786 HostAction = C.MakeAction<OffloadBundlingJobAction>(OffloadAL); 3787 recordHostAction(HostAction, InputArg); 3788 AL.back() = HostAction; 3789 } else 3790 AL.append(OffloadAL.begin(), OffloadAL.end()); 3791 3792 // Propagate to the current host action (if any) the offload information 3793 // associated with the current input. 3794 if (HostAction) 3795 HostAction->propagateHostOffloadInfo(InputArgToOffloadKindMap[InputArg], 3796 /*BoundArch=*/nullptr); 3797 return false; 3798 } 3799 3800 void appendDeviceLinkActions(ActionList &AL) { 3801 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3802 if (!SB->isValid()) 3803 continue; 3804 SB->appendLinkDeviceActions(AL); 3805 } 3806 } 3807 3808 Action *makeHostLinkAction() { 3809 // Build a list of device linking actions. 3810 ActionList DeviceAL; 3811 appendDeviceLinkActions(DeviceAL); 3812 if (DeviceAL.empty()) 3813 return nullptr; 3814 3815 // Let builders add host linking actions. 3816 Action* HA = nullptr; 3817 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3818 if (!SB->isValid()) 3819 continue; 3820 HA = SB->appendLinkHostActions(DeviceAL); 3821 // This created host action has no originating input argument, therefore 3822 // needs to set its offloading kind directly. 3823 if (HA) 3824 HA->propagateHostOffloadInfo(SB->getAssociatedOffloadKind(), 3825 /*BoundArch=*/nullptr); 3826 } 3827 return HA; 3828 } 3829 3830 /// Processes the host linker action. This currently consists of replacing it 3831 /// with an offload action if there are device link objects and propagate to 3832 /// the host action all the offload kinds used in the current compilation. The 3833 /// resulting action is returned. 3834 Action *processHostLinkAction(Action *HostAction) { 3835 // Add all the dependences from the device linking actions. 3836 OffloadAction::DeviceDependences DDeps; 3837 for (auto *SB : SpecializedBuilders) { 3838 if (!SB->isValid()) 3839 continue; 3840 3841 SB->appendLinkDependences(DDeps); 3842 } 3843 3844 // Calculate all the offload kinds used in the current compilation. 3845 unsigned ActiveOffloadKinds = 0u; 3846 for (auto &I : InputArgToOffloadKindMap) 3847 ActiveOffloadKinds |= I.second; 3848 3849 // If we don't have device dependencies, we don't have to create an offload 3850 // action. 3851 if (DDeps.getActions().empty()) { 3852 // Set all the active offloading kinds to the link action. Given that it 3853 // is a link action it is assumed to depend on all actions generated so 3854 // far. 3855 HostAction->setHostOffloadInfo(ActiveOffloadKinds, 3856 /*BoundArch=*/nullptr); 3857 // Propagate active offloading kinds for each input to the link action. 3858 // Each input may have different active offloading kind. 3859 for (auto A : HostAction->inputs()) { 3860 auto ArgLoc = HostActionToInputArgMap.find(A); 3861 if (ArgLoc == HostActionToInputArgMap.end()) 3862 continue; 3863 auto OFKLoc = InputArgToOffloadKindMap.find(ArgLoc->second); 3864 if (OFKLoc == InputArgToOffloadKindMap.end()) 3865 continue; 3866 A->propagateHostOffloadInfo(OFKLoc->second, /*BoundArch=*/nullptr); 3867 } 3868 return HostAction; 3869 } 3870 3871 // Create the offload action with all dependences. When an offload action 3872 // is created the kinds are propagated to the host action, so we don't have 3873 // to do that explicitly here. 3874 OffloadAction::HostDependence HDep( 3875 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3876 /*BoundArch*/ nullptr, ActiveOffloadKinds); 3877 return C.MakeAction<OffloadAction>(HDep, DDeps); 3878 } 3879 }; 3880 } // anonymous namespace. 3881 3882 void Driver::handleArguments(Compilation &C, DerivedArgList &Args, 3883 const InputList &Inputs, 3884 ActionList &Actions) const { 3885 3886 // Ignore /Yc/Yu if both /Yc and /Yu passed but with different filenames. 3887 Arg *YcArg = Args.getLastArg(options::OPT__SLASH_Yc); 3888 Arg *YuArg = Args.getLastArg(options::OPT__SLASH_Yu); 3889 if (YcArg && YuArg && strcmp(YcArg->getValue(), YuArg->getValue()) != 0) { 3890 Diag(clang::diag::warn_drv_ycyu_different_arg_clang_cl); 3891 Args.eraseArg(options::OPT__SLASH_Yc); 3892 Args.eraseArg(options::OPT__SLASH_Yu); 3893 YcArg = YuArg = nullptr; 3894 } 3895 if (YcArg && Inputs.size() > 1) { 3896 Diag(clang::diag::warn_drv_yc_multiple_inputs_clang_cl); 3897 Args.eraseArg(options::OPT__SLASH_Yc); 3898 YcArg = nullptr; 3899 } 3900 3901 Arg *FinalPhaseArg; 3902 phases::ID FinalPhase = getFinalPhase(Args, &FinalPhaseArg); 3903 3904 if (FinalPhase == phases::Link) { 3905 // Emitting LLVM while linking disabled except in HIPAMD Toolchain 3906 if (Args.hasArg(options::OPT_emit_llvm) && !Args.hasArg(options::OPT_hip_link)) 3907 Diag(clang::diag::err_drv_emit_llvm_link); 3908 if (IsCLMode() && LTOMode != LTOK_None && 3909 !Args.getLastArgValue(options::OPT_fuse_ld_EQ) 3910 .equals_insensitive("lld")) 3911 Diag(clang::diag::err_drv_lto_without_lld); 3912 } 3913 3914 if (FinalPhase == phases::Preprocess || Args.hasArg(options::OPT__SLASH_Y_)) { 3915 // If only preprocessing or /Y- is used, all pch handling is disabled. 3916 // Rather than check for it everywhere, just remove clang-cl pch-related 3917 // flags here. 3918 Args.eraseArg(options::OPT__SLASH_Fp); 3919 Args.eraseArg(options::OPT__SLASH_Yc); 3920 Args.eraseArg(options::OPT__SLASH_Yu); 3921 YcArg = YuArg = nullptr; 3922 } 3923 3924 unsigned LastPLSize = 0; 3925 for (auto &I : Inputs) { 3926 types::ID InputType = I.first; 3927 const Arg *InputArg = I.second; 3928 3929 auto PL = types::getCompilationPhases(InputType); 3930 LastPLSize = PL.size(); 3931 3932 // If the first step comes after the final phase we are doing as part of 3933 // this compilation, warn the user about it. 3934 phases::ID InitialPhase = PL[0]; 3935 if (InitialPhase > FinalPhase) { 3936 if (InputArg->isClaimed()) 3937 continue; 3938 3939 // Claim here to avoid the more general unused warning. 3940 InputArg->claim(); 3941 3942 // Suppress all unused style warnings with -Qunused-arguments 3943 if (Args.hasArg(options::OPT_Qunused_arguments)) 3944 continue; 3945 3946 // Special case when final phase determined by binary name, rather than 3947 // by a command-line argument with a corresponding Arg. 3948 if (CCCIsCPP()) 3949 Diag(clang::diag::warn_drv_input_file_unused_by_cpp) 3950 << InputArg->getAsString(Args) << getPhaseName(InitialPhase); 3951 // Special case '-E' warning on a previously preprocessed file to make 3952 // more sense. 3953 else if (InitialPhase == phases::Compile && 3954 (Args.getLastArg(options::OPT__SLASH_EP, 3955 options::OPT__SLASH_P) || 3956 Args.getLastArg(options::OPT_E) || 3957 Args.getLastArg(options::OPT_M, options::OPT_MM)) && 3958 getPreprocessedType(InputType) == types::TY_INVALID) 3959 Diag(clang::diag::warn_drv_preprocessed_input_file_unused) 3960 << InputArg->getAsString(Args) << !!FinalPhaseArg 3961 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3962 else 3963 Diag(clang::diag::warn_drv_input_file_unused) 3964 << InputArg->getAsString(Args) << getPhaseName(InitialPhase) 3965 << !!FinalPhaseArg 3966 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3967 continue; 3968 } 3969 3970 if (YcArg) { 3971 // Add a separate precompile phase for the compile phase. 3972 if (FinalPhase >= phases::Compile) { 3973 const types::ID HeaderType = lookupHeaderTypeForSourceType(InputType); 3974 // Build the pipeline for the pch file. 3975 Action *ClangClPch = C.MakeAction<InputAction>(*InputArg, HeaderType); 3976 for (phases::ID Phase : types::getCompilationPhases(HeaderType)) 3977 ClangClPch = ConstructPhaseAction(C, Args, Phase, ClangClPch); 3978 assert(ClangClPch); 3979 Actions.push_back(ClangClPch); 3980 // The driver currently exits after the first failed command. This 3981 // relies on that behavior, to make sure if the pch generation fails, 3982 // the main compilation won't run. 3983 // FIXME: If the main compilation fails, the PCH generation should 3984 // probably not be considered successful either. 3985 } 3986 } 3987 } 3988 3989 // If we are linking, claim any options which are obviously only used for 3990 // compilation. 3991 // FIXME: Understand why the last Phase List length is used here. 3992 if (FinalPhase == phases::Link && LastPLSize == 1) { 3993 Args.ClaimAllArgs(options::OPT_CompileOnly_Group); 3994 Args.ClaimAllArgs(options::OPT_cl_compile_Group); 3995 } 3996 } 3997 3998 void Driver::BuildActions(Compilation &C, DerivedArgList &Args, 3999 const InputList &Inputs, ActionList &Actions) const { 4000 llvm::PrettyStackTraceString CrashInfo("Building compilation actions"); 4001 4002 if (!SuppressMissingInputWarning && Inputs.empty()) { 4003 Diag(clang::diag::err_drv_no_input_files); 4004 return; 4005 } 4006 4007 // Reject -Z* at the top level, these options should never have been exposed 4008 // by gcc. 4009 if (Arg *A = Args.getLastArg(options::OPT_Z_Joined)) 4010 Diag(clang::diag::err_drv_use_of_Z_option) << A->getAsString(Args); 4011 4012 // Diagnose misuse of /Fo. 4013 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fo)) { 4014 StringRef V = A->getValue(); 4015 if (Inputs.size() > 1 && !V.empty() && 4016 !llvm::sys::path::is_separator(V.back())) { 4017 // Check whether /Fo tries to name an output file for multiple inputs. 4018 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4019 << A->getSpelling() << V; 4020 Args.eraseArg(options::OPT__SLASH_Fo); 4021 } 4022 } 4023 4024 // Diagnose misuse of /Fa. 4025 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fa)) { 4026 StringRef V = A->getValue(); 4027 if (Inputs.size() > 1 && !V.empty() && 4028 !llvm::sys::path::is_separator(V.back())) { 4029 // Check whether /Fa tries to name an asm file for multiple inputs. 4030 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4031 << A->getSpelling() << V; 4032 Args.eraseArg(options::OPT__SLASH_Fa); 4033 } 4034 } 4035 4036 // Diagnose misuse of /o. 4037 if (Arg *A = Args.getLastArg(options::OPT__SLASH_o)) { 4038 if (A->getValue()[0] == '\0') { 4039 // It has to have a value. 4040 Diag(clang::diag::err_drv_missing_argument) << A->getSpelling() << 1; 4041 Args.eraseArg(options::OPT__SLASH_o); 4042 } 4043 } 4044 4045 handleArguments(C, Args, Inputs, Actions); 4046 4047 // Builder to be used to build offloading actions. 4048 OffloadingActionBuilder OffloadBuilder(C, Args, Inputs); 4049 4050 bool UseNewOffloadingDriver = 4051 (C.isOffloadingHostKind(Action::OFK_OpenMP) && 4052 Args.hasFlag(options::OPT_fopenmp_new_driver, 4053 options::OPT_no_offload_new_driver, true)) || 4054 Args.hasFlag(options::OPT_offload_new_driver, 4055 options::OPT_no_offload_new_driver, false); 4056 4057 // Construct the actions to perform. 4058 HeaderModulePrecompileJobAction *HeaderModuleAction = nullptr; 4059 ExtractAPIJobAction *ExtractAPIAction = nullptr; 4060 ActionList LinkerInputs; 4061 ActionList MergerInputs; 4062 4063 for (auto &I : Inputs) { 4064 types::ID InputType = I.first; 4065 const Arg *InputArg = I.second; 4066 4067 auto PL = types::getCompilationPhases(*this, Args, InputType); 4068 if (PL.empty()) 4069 continue; 4070 4071 auto FullPL = types::getCompilationPhases(InputType); 4072 4073 // Build the pipeline for this file. 4074 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4075 4076 // Use the current host action in any of the offloading actions, if 4077 // required. 4078 if (!UseNewOffloadingDriver) 4079 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4080 break; 4081 4082 for (phases::ID Phase : PL) { 4083 4084 // Add any offload action the host action depends on. 4085 if (!UseNewOffloadingDriver) 4086 Current = OffloadBuilder.addDeviceDependencesToHostAction( 4087 Current, InputArg, Phase, PL.back(), FullPL); 4088 if (!Current) 4089 break; 4090 4091 // Queue linker inputs. 4092 if (Phase == phases::Link) { 4093 assert(Phase == PL.back() && "linking must be final compilation step."); 4094 // We don't need to generate additional link commands if emitting AMD bitcode 4095 if (!(C.getInputArgs().hasArg(options::OPT_hip_link) && 4096 (C.getInputArgs().hasArg(options::OPT_emit_llvm)))) 4097 LinkerInputs.push_back(Current); 4098 Current = nullptr; 4099 break; 4100 } 4101 4102 // TODO: Consider removing this because the merged may not end up being 4103 // the final Phase in the pipeline. Perhaps the merged could just merge 4104 // and then pass an artifact of some sort to the Link Phase. 4105 // Queue merger inputs. 4106 if (Phase == phases::IfsMerge) { 4107 assert(Phase == PL.back() && "merging must be final compilation step."); 4108 MergerInputs.push_back(Current); 4109 Current = nullptr; 4110 break; 4111 } 4112 4113 // Each precompiled header file after a module file action is a module 4114 // header of that same module file, rather than being compiled to a 4115 // separate PCH. 4116 if (Phase == phases::Precompile && HeaderModuleAction && 4117 getPrecompiledType(InputType) == types::TY_PCH) { 4118 HeaderModuleAction->addModuleHeaderInput(Current); 4119 Current = nullptr; 4120 break; 4121 } 4122 4123 if (Phase == phases::Precompile && ExtractAPIAction) { 4124 ExtractAPIAction->addHeaderInput(Current); 4125 Current = nullptr; 4126 break; 4127 } 4128 4129 // FIXME: Should we include any prior module file outputs as inputs of 4130 // later actions in the same command line? 4131 4132 // Otherwise construct the appropriate action. 4133 Action *NewCurrent = ConstructPhaseAction(C, Args, Phase, Current); 4134 4135 // We didn't create a new action, so we will just move to the next phase. 4136 if (NewCurrent == Current) 4137 continue; 4138 4139 if (auto *HMA = dyn_cast<HeaderModulePrecompileJobAction>(NewCurrent)) 4140 HeaderModuleAction = HMA; 4141 else if (auto *EAA = dyn_cast<ExtractAPIJobAction>(NewCurrent)) 4142 ExtractAPIAction = EAA; 4143 4144 Current = NewCurrent; 4145 4146 // Use the current host action in any of the offloading actions, if 4147 // required. 4148 if (!UseNewOffloadingDriver) 4149 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4150 break; 4151 4152 // Try to build the offloading actions and add the result as a dependency 4153 // to the host. 4154 if (UseNewOffloadingDriver) 4155 Current = BuildOffloadingActions(C, Args, I, Current); 4156 4157 if (Current->getType() == types::TY_Nothing) 4158 break; 4159 } 4160 4161 // If we ended with something, add to the output list. 4162 if (Current) 4163 Actions.push_back(Current); 4164 4165 // Add any top level actions generated for offloading. 4166 if (!UseNewOffloadingDriver) 4167 OffloadBuilder.appendTopLevelActions(Actions, Current, InputArg); 4168 else if (Current) 4169 Current->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4170 /*BoundArch=*/nullptr); 4171 } 4172 4173 // Add a link action if necessary. 4174 4175 if (LinkerInputs.empty()) { 4176 Arg *FinalPhaseArg; 4177 if (getFinalPhase(Args, &FinalPhaseArg) == phases::Link) 4178 if (!UseNewOffloadingDriver) 4179 OffloadBuilder.appendDeviceLinkActions(Actions); 4180 } 4181 4182 if (!LinkerInputs.empty()) { 4183 if (!UseNewOffloadingDriver) 4184 if (Action *Wrapper = OffloadBuilder.makeHostLinkAction()) 4185 LinkerInputs.push_back(Wrapper); 4186 Action *LA; 4187 // Check if this Linker Job should emit a static library. 4188 if (ShouldEmitStaticLibrary(Args)) { 4189 LA = C.MakeAction<StaticLibJobAction>(LinkerInputs, types::TY_Image); 4190 } else if (UseNewOffloadingDriver || 4191 Args.hasArg(options::OPT_offload_link)) { 4192 LA = C.MakeAction<LinkerWrapperJobAction>(LinkerInputs, types::TY_Image); 4193 LA->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4194 /*BoundArch=*/nullptr); 4195 } else { 4196 LA = C.MakeAction<LinkJobAction>(LinkerInputs, types::TY_Image); 4197 } 4198 if (!UseNewOffloadingDriver) 4199 LA = OffloadBuilder.processHostLinkAction(LA); 4200 Actions.push_back(LA); 4201 } 4202 4203 // Add an interface stubs merge action if necessary. 4204 if (!MergerInputs.empty()) 4205 Actions.push_back( 4206 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4207 4208 if (Args.hasArg(options::OPT_emit_interface_stubs)) { 4209 auto PhaseList = types::getCompilationPhases( 4210 types::TY_IFS_CPP, 4211 Args.hasArg(options::OPT_c) ? phases::Compile : phases::IfsMerge); 4212 4213 ActionList MergerInputs; 4214 4215 for (auto &I : Inputs) { 4216 types::ID InputType = I.first; 4217 const Arg *InputArg = I.second; 4218 4219 // Currently clang and the llvm assembler do not support generating symbol 4220 // stubs from assembly, so we skip the input on asm files. For ifs files 4221 // we rely on the normal pipeline setup in the pipeline setup code above. 4222 if (InputType == types::TY_IFS || InputType == types::TY_PP_Asm || 4223 InputType == types::TY_Asm) 4224 continue; 4225 4226 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4227 4228 for (auto Phase : PhaseList) { 4229 switch (Phase) { 4230 default: 4231 llvm_unreachable( 4232 "IFS Pipeline can only consist of Compile followed by IfsMerge."); 4233 case phases::Compile: { 4234 // Only IfsMerge (llvm-ifs) can handle .o files by looking for ifs 4235 // files where the .o file is located. The compile action can not 4236 // handle this. 4237 if (InputType == types::TY_Object) 4238 break; 4239 4240 Current = C.MakeAction<CompileJobAction>(Current, types::TY_IFS_CPP); 4241 break; 4242 } 4243 case phases::IfsMerge: { 4244 assert(Phase == PhaseList.back() && 4245 "merging must be final compilation step."); 4246 MergerInputs.push_back(Current); 4247 Current = nullptr; 4248 break; 4249 } 4250 } 4251 } 4252 4253 // If we ended with something, add to the output list. 4254 if (Current) 4255 Actions.push_back(Current); 4256 } 4257 4258 // Add an interface stubs merge action if necessary. 4259 if (!MergerInputs.empty()) 4260 Actions.push_back( 4261 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4262 } 4263 4264 // If --print-supported-cpus, -mcpu=? or -mtune=? is specified, build a custom 4265 // Compile phase that prints out supported cpu models and quits. 4266 if (Arg *A = Args.getLastArg(options::OPT_print_supported_cpus)) { 4267 // Use the -mcpu=? flag as the dummy input to cc1. 4268 Actions.clear(); 4269 Action *InputAc = C.MakeAction<InputAction>(*A, types::TY_C); 4270 Actions.push_back( 4271 C.MakeAction<PrecompileJobAction>(InputAc, types::TY_Nothing)); 4272 for (auto &I : Inputs) 4273 I.second->claim(); 4274 } 4275 4276 // Claim ignored clang-cl options. 4277 Args.ClaimAllArgs(options::OPT_cl_ignored_Group); 4278 } 4279 4280 /// Returns the canonical name for the offloading architecture when using a HIP 4281 /// or CUDA architecture. 4282 static StringRef getCanonicalArchString(Compilation &C, 4283 const llvm::opt::DerivedArgList &Args, 4284 StringRef ArchStr, 4285 const llvm::Triple &Triple) { 4286 // Lookup the CUDA / HIP architecture string. Only report an error if we were 4287 // expecting the triple to be only NVPTX / AMDGPU. 4288 CudaArch Arch = StringToCudaArch(getProcessorFromTargetID(Triple, ArchStr)); 4289 if (Triple.isNVPTX() && 4290 (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch))) { 4291 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4292 << "CUDA" << ArchStr; 4293 return StringRef(); 4294 } else if (Triple.isAMDGPU() && 4295 (Arch == CudaArch::UNKNOWN || !IsAMDGpuArch(Arch))) { 4296 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4297 << "HIP" << ArchStr; 4298 return StringRef(); 4299 } 4300 4301 if (IsNVIDIAGpuArch(Arch)) 4302 return Args.MakeArgStringRef(CudaArchToString(Arch)); 4303 4304 if (IsAMDGpuArch(Arch)) { 4305 llvm::StringMap<bool> Features; 4306 auto HIPTriple = getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()); 4307 if (!HIPTriple) 4308 return StringRef(); 4309 auto Arch = parseTargetID(*HIPTriple, ArchStr, &Features); 4310 if (!Arch) { 4311 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << ArchStr; 4312 C.setContainsError(); 4313 return StringRef(); 4314 } 4315 return Args.MakeArgStringRef( 4316 getCanonicalTargetID(Arch.getValue(), Features)); 4317 } 4318 4319 // If the input isn't CUDA or HIP just return the architecture. 4320 return ArchStr; 4321 } 4322 4323 /// Checks if the set offloading architectures does not conflict. Returns the 4324 /// incompatible pair if a conflict occurs. 4325 static llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 4326 getConflictOffloadArchCombination(const llvm::DenseSet<StringRef> &Archs, 4327 Action::OffloadKind Kind) { 4328 if (Kind != Action::OFK_HIP) 4329 return None; 4330 4331 std::set<StringRef> ArchSet; 4332 llvm::copy(Archs, std::inserter(ArchSet, ArchSet.begin())); 4333 return getConflictTargetIDCombination(ArchSet); 4334 } 4335 4336 llvm::DenseSet<StringRef> 4337 Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args, 4338 Action::OffloadKind Kind, const ToolChain *TC) const { 4339 if (!TC) 4340 TC = &C.getDefaultToolChain(); 4341 4342 // --offload and --offload-arch options are mutually exclusive. 4343 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 4344 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 4345 options::OPT_no_offload_arch_EQ)) { 4346 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) 4347 << "--offload" 4348 << (Args.hasArgNoClaim(options::OPT_offload_arch_EQ) 4349 ? "--offload-arch" 4350 : "--no-offload-arch"); 4351 } 4352 4353 if (KnownArchs.find(TC) != KnownArchs.end()) 4354 return KnownArchs.lookup(TC); 4355 4356 llvm::DenseSet<StringRef> Archs; 4357 for (auto &Arg : Args) { 4358 if (Arg->getOption().matches(options::OPT_offload_arch_EQ)) { 4359 Archs.insert( 4360 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4361 } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) { 4362 if (Arg->getValue() == StringRef("all")) 4363 Archs.clear(); 4364 else 4365 Archs.erase( 4366 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4367 } 4368 } 4369 4370 if (auto ConflictingArchs = getConflictOffloadArchCombination(Archs, Kind)) { 4371 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 4372 << ConflictingArchs.getValue().first 4373 << ConflictingArchs.getValue().second; 4374 C.setContainsError(); 4375 } 4376 4377 if (Archs.empty()) { 4378 if (Kind == Action::OFK_Cuda) 4379 Archs.insert(CudaArchToString(CudaArch::CudaDefault)); 4380 else if (Kind == Action::OFK_HIP) 4381 Archs.insert(CudaArchToString(CudaArch::HIPDefault)); 4382 else if (Kind == Action::OFK_OpenMP) 4383 Archs.insert(StringRef()); 4384 } else { 4385 Args.ClaimAllArgs(options::OPT_offload_arch_EQ); 4386 Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ); 4387 } 4388 4389 return Archs; 4390 } 4391 4392 Action *Driver::BuildOffloadingActions(Compilation &C, 4393 llvm::opt::DerivedArgList &Args, 4394 const InputTy &Input, 4395 Action *HostAction) const { 4396 // Don't build offloading actions if explicitly disabled or we do not have a 4397 // valid source input and compile action to embed it in. If preprocessing only 4398 // ignore embedding. 4399 if (offloadHostOnly() || !types::isSrcFile(Input.first) || 4400 !(isa<CompileJobAction>(HostAction) || 4401 getFinalPhase(Args) == phases::Preprocess)) 4402 return HostAction; 4403 4404 ActionList OffloadActions; 4405 OffloadAction::DeviceDependences DDeps; 4406 4407 const Action::OffloadKind OffloadKinds[] = { 4408 Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP}; 4409 4410 for (Action::OffloadKind Kind : OffloadKinds) { 4411 SmallVector<const ToolChain *, 2> ToolChains; 4412 ActionList DeviceActions; 4413 4414 auto TCRange = C.getOffloadToolChains(Kind); 4415 for (auto TI = TCRange.first, TE = TCRange.second; TI != TE; ++TI) 4416 ToolChains.push_back(TI->second); 4417 4418 if (ToolChains.empty()) 4419 continue; 4420 4421 types::ID InputType = Input.first; 4422 const Arg *InputArg = Input.second; 4423 4424 // Get the product of all bound architectures and toolchains. 4425 SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs; 4426 for (const ToolChain *TC : ToolChains) 4427 for (StringRef Arch : getOffloadArchs( 4428 C, C.getArgsForToolChain(TC, "generic", Kind), Kind, TC)) 4429 TCAndArchs.push_back(std::make_pair(TC, Arch)); 4430 4431 for (unsigned I = 0, E = TCAndArchs.size(); I != E; ++I) 4432 DeviceActions.push_back(C.MakeAction<InputAction>(*InputArg, InputType)); 4433 4434 if (DeviceActions.empty()) 4435 return HostAction; 4436 4437 auto PL = types::getCompilationPhases(*this, Args, InputType); 4438 4439 for (phases::ID Phase : PL) { 4440 if (Phase == phases::Link) { 4441 assert(Phase == PL.back() && "linking must be final compilation step."); 4442 break; 4443 } 4444 4445 auto TCAndArch = TCAndArchs.begin(); 4446 for (Action *&A : DeviceActions) { 4447 A = ConstructPhaseAction(C, Args, Phase, A, Kind); 4448 4449 if (isa<CompileJobAction>(A) && isa<CompileJobAction>(HostAction) && 4450 Kind == Action::OFK_OpenMP) { 4451 // OpenMP offloading has a dependency on the host compile action to 4452 // identify which declarations need to be emitted. This shouldn't be 4453 // collapsed with any other actions so we can use it in the device. 4454 HostAction->setCannotBeCollapsedWithNextDependentAction(); 4455 OffloadAction::HostDependence HDep( 4456 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4457 TCAndArch->second.data(), Kind); 4458 OffloadAction::DeviceDependences DDep; 4459 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4460 A = C.MakeAction<OffloadAction>(HDep, DDep); 4461 } else if (isa<AssembleJobAction>(A) && Kind == Action::OFK_Cuda) { 4462 // The Cuda toolchain uses fatbinary as the linker phase to bundle the 4463 // PTX and Cubin output. 4464 ActionList FatbinActions; 4465 for (Action *A : {A, A->getInputs()[0]}) { 4466 OffloadAction::DeviceDependences DDep; 4467 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4468 FatbinActions.emplace_back( 4469 C.MakeAction<OffloadAction>(DDep, A->getType())); 4470 } 4471 A = C.MakeAction<LinkJobAction>(FatbinActions, types::TY_CUDA_FATBIN); 4472 } 4473 ++TCAndArch; 4474 } 4475 } 4476 4477 auto TCAndArch = TCAndArchs.begin(); 4478 for (Action *A : DeviceActions) { 4479 DDeps.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4480 OffloadAction::DeviceDependences DDep; 4481 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4482 OffloadActions.push_back(C.MakeAction<OffloadAction>(DDep, A->getType())); 4483 ++TCAndArch; 4484 } 4485 } 4486 4487 if (offloadDeviceOnly()) 4488 return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing); 4489 4490 Action *OffloadPackager = 4491 C.MakeAction<OffloadPackagerJobAction>(OffloadActions, types::TY_Image); 4492 OffloadAction::DeviceDependences DDep; 4493 DDep.add(*OffloadPackager, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4494 nullptr, Action::OFK_None); 4495 OffloadAction::HostDependence HDep( 4496 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4497 /*BoundArch=*/nullptr, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4498 return C.MakeAction<OffloadAction>( 4499 HDep, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4500 } 4501 4502 Action *Driver::ConstructPhaseAction( 4503 Compilation &C, const ArgList &Args, phases::ID Phase, Action *Input, 4504 Action::OffloadKind TargetDeviceOffloadKind) const { 4505 llvm::PrettyStackTraceString CrashInfo("Constructing phase actions"); 4506 4507 // Some types skip the assembler phase (e.g., llvm-bc), but we can't 4508 // encode this in the steps because the intermediate type depends on 4509 // arguments. Just special case here. 4510 if (Phase == phases::Assemble && Input->getType() != types::TY_PP_Asm) 4511 return Input; 4512 4513 // Build the appropriate action. 4514 switch (Phase) { 4515 case phases::Link: 4516 llvm_unreachable("link action invalid here."); 4517 case phases::IfsMerge: 4518 llvm_unreachable("ifsmerge action invalid here."); 4519 case phases::Preprocess: { 4520 types::ID OutputTy; 4521 // -M and -MM specify the dependency file name by altering the output type, 4522 // -if -MD and -MMD are not specified. 4523 if (Args.hasArg(options::OPT_M, options::OPT_MM) && 4524 !Args.hasArg(options::OPT_MD, options::OPT_MMD)) { 4525 OutputTy = types::TY_Dependencies; 4526 } else { 4527 OutputTy = Input->getType(); 4528 // For these cases, the preprocessor is only translating forms, the Output 4529 // still needs preprocessing. 4530 if (!Args.hasFlag(options::OPT_frewrite_includes, 4531 options::OPT_fno_rewrite_includes, false) && 4532 !Args.hasFlag(options::OPT_frewrite_imports, 4533 options::OPT_fno_rewrite_imports, false) && 4534 !Args.hasFlag(options::OPT_fdirectives_only, 4535 options::OPT_fno_directives_only, false) && 4536 !CCGenDiagnostics) 4537 OutputTy = types::getPreprocessedType(OutputTy); 4538 assert(OutputTy != types::TY_INVALID && 4539 "Cannot preprocess this input type!"); 4540 } 4541 return C.MakeAction<PreprocessJobAction>(Input, OutputTy); 4542 } 4543 case phases::Precompile: { 4544 // API extraction should not generate an actual precompilation action. 4545 if (Args.hasArg(options::OPT_extract_api)) 4546 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4547 4548 types::ID OutputTy = getPrecompiledType(Input->getType()); 4549 assert(OutputTy != types::TY_INVALID && 4550 "Cannot precompile this input type!"); 4551 4552 // If we're given a module name, precompile header file inputs as a 4553 // module, not as a precompiled header. 4554 const char *ModName = nullptr; 4555 if (OutputTy == types::TY_PCH) { 4556 if (Arg *A = Args.getLastArg(options::OPT_fmodule_name_EQ)) 4557 ModName = A->getValue(); 4558 if (ModName) 4559 OutputTy = types::TY_ModuleFile; 4560 } 4561 4562 if (Args.hasArg(options::OPT_fsyntax_only)) { 4563 // Syntax checks should not emit a PCH file 4564 OutputTy = types::TY_Nothing; 4565 } 4566 4567 if (ModName) 4568 return C.MakeAction<HeaderModulePrecompileJobAction>(Input, OutputTy, 4569 ModName); 4570 return C.MakeAction<PrecompileJobAction>(Input, OutputTy); 4571 } 4572 case phases::Compile: { 4573 if (Args.hasArg(options::OPT_fsyntax_only)) 4574 return C.MakeAction<CompileJobAction>(Input, types::TY_Nothing); 4575 if (Args.hasArg(options::OPT_rewrite_objc)) 4576 return C.MakeAction<CompileJobAction>(Input, types::TY_RewrittenObjC); 4577 if (Args.hasArg(options::OPT_rewrite_legacy_objc)) 4578 return C.MakeAction<CompileJobAction>(Input, 4579 types::TY_RewrittenLegacyObjC); 4580 if (Args.hasArg(options::OPT__analyze)) 4581 return C.MakeAction<AnalyzeJobAction>(Input, types::TY_Plist); 4582 if (Args.hasArg(options::OPT__migrate)) 4583 return C.MakeAction<MigrateJobAction>(Input, types::TY_Remap); 4584 if (Args.hasArg(options::OPT_emit_ast)) 4585 return C.MakeAction<CompileJobAction>(Input, types::TY_AST); 4586 if (Args.hasArg(options::OPT_module_file_info)) 4587 return C.MakeAction<CompileJobAction>(Input, types::TY_ModuleFile); 4588 if (Args.hasArg(options::OPT_verify_pch)) 4589 return C.MakeAction<VerifyPCHJobAction>(Input, types::TY_Nothing); 4590 if (Args.hasArg(options::OPT_extract_api)) 4591 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4592 return C.MakeAction<CompileJobAction>(Input, types::TY_LLVM_BC); 4593 } 4594 case phases::Backend: { 4595 if (isUsingLTO() && TargetDeviceOffloadKind == Action::OFK_None) { 4596 types::ID Output = 4597 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4598 return C.MakeAction<BackendJobAction>(Input, Output); 4599 } 4600 if (isUsingLTO(/* IsOffload */ true) && 4601 TargetDeviceOffloadKind != Action::OFK_None) { 4602 types::ID Output = 4603 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4604 return C.MakeAction<BackendJobAction>(Input, Output); 4605 } 4606 if (Args.hasArg(options::OPT_emit_llvm) || 4607 (TargetDeviceOffloadKind == Action::OFK_HIP && 4608 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4609 false))) { 4610 types::ID Output = 4611 Args.hasArg(options::OPT_S) ? types::TY_LLVM_IR : types::TY_LLVM_BC; 4612 return C.MakeAction<BackendJobAction>(Input, Output); 4613 } 4614 return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm); 4615 } 4616 case phases::Assemble: 4617 return C.MakeAction<AssembleJobAction>(std::move(Input), types::TY_Object); 4618 } 4619 4620 llvm_unreachable("invalid phase in ConstructPhaseAction"); 4621 } 4622 4623 void Driver::BuildJobs(Compilation &C) const { 4624 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 4625 4626 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 4627 4628 // It is an error to provide a -o option if we are making multiple output 4629 // files. There are exceptions: 4630 // 4631 // IfsMergeJob: when generating interface stubs enabled we want to be able to 4632 // generate the stub file at the same time that we generate the real 4633 // library/a.out. So when a .o, .so, etc are the output, with clang interface 4634 // stubs there will also be a .ifs and .ifso at the same location. 4635 // 4636 // CompileJob of type TY_IFS_CPP: when generating interface stubs is enabled 4637 // and -c is passed, we still want to be able to generate a .ifs file while 4638 // we are also generating .o files. So we allow more than one output file in 4639 // this case as well. 4640 // 4641 if (FinalOutput) { 4642 unsigned NumOutputs = 0; 4643 unsigned NumIfsOutputs = 0; 4644 for (const Action *A : C.getActions()) 4645 if (A->getType() != types::TY_Nothing && 4646 !(A->getKind() == Action::IfsMergeJobClass || 4647 (A->getType() == clang::driver::types::TY_IFS_CPP && 4648 A->getKind() == clang::driver::Action::CompileJobClass && 4649 0 == NumIfsOutputs++) || 4650 (A->getKind() == Action::BindArchClass && A->getInputs().size() && 4651 A->getInputs().front()->getKind() == Action::IfsMergeJobClass))) 4652 ++NumOutputs; 4653 4654 if (NumOutputs > 1) { 4655 Diag(clang::diag::err_drv_output_argument_with_multiple_files); 4656 FinalOutput = nullptr; 4657 } 4658 } 4659 4660 const llvm::Triple &RawTriple = C.getDefaultToolChain().getTriple(); 4661 if (RawTriple.isOSAIX()) { 4662 if (Arg *A = C.getArgs().getLastArg(options::OPT_G)) 4663 Diag(diag::err_drv_unsupported_opt_for_target) 4664 << A->getSpelling() << RawTriple.str(); 4665 if (LTOMode == LTOK_Thin) 4666 Diag(diag::err_drv_clang_unsupported) << "thinLTO on AIX"; 4667 } 4668 4669 // Collect the list of architectures. 4670 llvm::StringSet<> ArchNames; 4671 if (RawTriple.isOSBinFormatMachO()) 4672 for (const Arg *A : C.getArgs()) 4673 if (A->getOption().matches(options::OPT_arch)) 4674 ArchNames.insert(A->getValue()); 4675 4676 // Set of (Action, canonical ToolChain triple) pairs we've built jobs for. 4677 std::map<std::pair<const Action *, std::string>, InputInfoList> CachedResults; 4678 for (Action *A : C.getActions()) { 4679 // If we are linking an image for multiple archs then the linker wants 4680 // -arch_multiple and -final_output <final image name>. Unfortunately, this 4681 // doesn't fit in cleanly because we have to pass this information down. 4682 // 4683 // FIXME: This is a hack; find a cleaner way to integrate this into the 4684 // process. 4685 const char *LinkingOutput = nullptr; 4686 if (isa<LipoJobAction>(A)) { 4687 if (FinalOutput) 4688 LinkingOutput = FinalOutput->getValue(); 4689 else 4690 LinkingOutput = getDefaultImageName(); 4691 } 4692 4693 BuildJobsForAction(C, A, &C.getDefaultToolChain(), 4694 /*BoundArch*/ StringRef(), 4695 /*AtTopLevel*/ true, 4696 /*MultipleArchs*/ ArchNames.size() > 1, 4697 /*LinkingOutput*/ LinkingOutput, CachedResults, 4698 /*TargetDeviceOffloadKind*/ Action::OFK_None); 4699 } 4700 4701 // If we have more than one job, then disable integrated-cc1 for now. Do this 4702 // also when we need to report process execution statistics. 4703 if (C.getJobs().size() > 1 || CCPrintProcessStats) 4704 for (auto &J : C.getJobs()) 4705 J.InProcess = false; 4706 4707 if (CCPrintProcessStats) { 4708 C.setPostCallback([=](const Command &Cmd, int Res) { 4709 Optional<llvm::sys::ProcessStatistics> ProcStat = 4710 Cmd.getProcessStatistics(); 4711 if (!ProcStat) 4712 return; 4713 4714 const char *LinkingOutput = nullptr; 4715 if (FinalOutput) 4716 LinkingOutput = FinalOutput->getValue(); 4717 else if (!Cmd.getOutputFilenames().empty()) 4718 LinkingOutput = Cmd.getOutputFilenames().front().c_str(); 4719 else 4720 LinkingOutput = getDefaultImageName(); 4721 4722 if (CCPrintStatReportFilename.empty()) { 4723 using namespace llvm; 4724 // Human readable output. 4725 outs() << sys::path::filename(Cmd.getExecutable()) << ": " 4726 << "output=" << LinkingOutput; 4727 outs() << ", total=" 4728 << format("%.3f", ProcStat->TotalTime.count() / 1000.) << " ms" 4729 << ", user=" 4730 << format("%.3f", ProcStat->UserTime.count() / 1000.) << " ms" 4731 << ", mem=" << ProcStat->PeakMemory << " Kb\n"; 4732 } else { 4733 // CSV format. 4734 std::string Buffer; 4735 llvm::raw_string_ostream Out(Buffer); 4736 llvm::sys::printArg(Out, llvm::sys::path::filename(Cmd.getExecutable()), 4737 /*Quote*/ true); 4738 Out << ','; 4739 llvm::sys::printArg(Out, LinkingOutput, true); 4740 Out << ',' << ProcStat->TotalTime.count() << ',' 4741 << ProcStat->UserTime.count() << ',' << ProcStat->PeakMemory 4742 << '\n'; 4743 Out.flush(); 4744 std::error_code EC; 4745 llvm::raw_fd_ostream OS(CCPrintStatReportFilename, EC, 4746 llvm::sys::fs::OF_Append | 4747 llvm::sys::fs::OF_Text); 4748 if (EC) 4749 return; 4750 auto L = OS.lock(); 4751 if (!L) { 4752 llvm::errs() << "ERROR: Cannot lock file " 4753 << CCPrintStatReportFilename << ": " 4754 << toString(L.takeError()) << "\n"; 4755 return; 4756 } 4757 OS << Buffer; 4758 OS.flush(); 4759 } 4760 }); 4761 } 4762 4763 // If the user passed -Qunused-arguments or there were errors, don't warn 4764 // about any unused arguments. 4765 if (Diags.hasErrorOccurred() || 4766 C.getArgs().hasArg(options::OPT_Qunused_arguments)) 4767 return; 4768 4769 // Claim -fdriver-only here. 4770 (void)C.getArgs().hasArg(options::OPT_fdriver_only); 4771 // Claim -### here. 4772 (void)C.getArgs().hasArg(options::OPT__HASH_HASH_HASH); 4773 4774 // Claim --driver-mode, --rsp-quoting, it was handled earlier. 4775 (void)C.getArgs().hasArg(options::OPT_driver_mode); 4776 (void)C.getArgs().hasArg(options::OPT_rsp_quoting); 4777 4778 for (Arg *A : C.getArgs()) { 4779 // FIXME: It would be nice to be able to send the argument to the 4780 // DiagnosticsEngine, so that extra values, position, and so on could be 4781 // printed. 4782 if (!A->isClaimed()) { 4783 if (A->getOption().hasFlag(options::NoArgumentUnused)) 4784 continue; 4785 4786 // Suppress the warning automatically if this is just a flag, and it is an 4787 // instance of an argument we already claimed. 4788 const Option &Opt = A->getOption(); 4789 if (Opt.getKind() == Option::FlagClass) { 4790 bool DuplicateClaimed = false; 4791 4792 for (const Arg *AA : C.getArgs().filtered(&Opt)) { 4793 if (AA->isClaimed()) { 4794 DuplicateClaimed = true; 4795 break; 4796 } 4797 } 4798 4799 if (DuplicateClaimed) 4800 continue; 4801 } 4802 4803 // In clang-cl, don't mention unknown arguments here since they have 4804 // already been warned about. 4805 if (!IsCLMode() || !A->getOption().matches(options::OPT_UNKNOWN)) 4806 Diag(clang::diag::warn_drv_unused_argument) 4807 << A->getAsString(C.getArgs()); 4808 } 4809 } 4810 } 4811 4812 namespace { 4813 /// Utility class to control the collapse of dependent actions and select the 4814 /// tools accordingly. 4815 class ToolSelector final { 4816 /// The tool chain this selector refers to. 4817 const ToolChain &TC; 4818 4819 /// The compilation this selector refers to. 4820 const Compilation &C; 4821 4822 /// The base action this selector refers to. 4823 const JobAction *BaseAction; 4824 4825 /// Set to true if the current toolchain refers to host actions. 4826 bool IsHostSelector; 4827 4828 /// Set to true if save-temps and embed-bitcode functionalities are active. 4829 bool SaveTemps; 4830 bool EmbedBitcode; 4831 4832 /// Get previous dependent action or null if that does not exist. If 4833 /// \a CanBeCollapsed is false, that action must be legal to collapse or 4834 /// null will be returned. 4835 const JobAction *getPrevDependentAction(const ActionList &Inputs, 4836 ActionList &SavedOffloadAction, 4837 bool CanBeCollapsed = true) { 4838 // An option can be collapsed only if it has a single input. 4839 if (Inputs.size() != 1) 4840 return nullptr; 4841 4842 Action *CurAction = *Inputs.begin(); 4843 if (CanBeCollapsed && 4844 !CurAction->isCollapsingWithNextDependentActionLegal()) 4845 return nullptr; 4846 4847 // If the input action is an offload action. Look through it and save any 4848 // offload action that can be dropped in the event of a collapse. 4849 if (auto *OA = dyn_cast<OffloadAction>(CurAction)) { 4850 // If the dependent action is a device action, we will attempt to collapse 4851 // only with other device actions. Otherwise, we would do the same but 4852 // with host actions only. 4853 if (!IsHostSelector) { 4854 if (OA->hasSingleDeviceDependence(/*DoNotConsiderHostActions=*/true)) { 4855 CurAction = 4856 OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true); 4857 if (CanBeCollapsed && 4858 !CurAction->isCollapsingWithNextDependentActionLegal()) 4859 return nullptr; 4860 SavedOffloadAction.push_back(OA); 4861 return dyn_cast<JobAction>(CurAction); 4862 } 4863 } else if (OA->hasHostDependence()) { 4864 CurAction = OA->getHostDependence(); 4865 if (CanBeCollapsed && 4866 !CurAction->isCollapsingWithNextDependentActionLegal()) 4867 return nullptr; 4868 SavedOffloadAction.push_back(OA); 4869 return dyn_cast<JobAction>(CurAction); 4870 } 4871 return nullptr; 4872 } 4873 4874 return dyn_cast<JobAction>(CurAction); 4875 } 4876 4877 /// Return true if an assemble action can be collapsed. 4878 bool canCollapseAssembleAction() const { 4879 return TC.useIntegratedAs() && !SaveTemps && 4880 !C.getArgs().hasArg(options::OPT_via_file_asm) && 4881 !C.getArgs().hasArg(options::OPT__SLASH_FA) && 4882 !C.getArgs().hasArg(options::OPT__SLASH_Fa); 4883 } 4884 4885 /// Return true if a preprocessor action can be collapsed. 4886 bool canCollapsePreprocessorAction() const { 4887 return !C.getArgs().hasArg(options::OPT_no_integrated_cpp) && 4888 !C.getArgs().hasArg(options::OPT_traditional_cpp) && !SaveTemps && 4889 !C.getArgs().hasArg(options::OPT_rewrite_objc); 4890 } 4891 4892 /// Struct that relates an action with the offload actions that would be 4893 /// collapsed with it. 4894 struct JobActionInfo final { 4895 /// The action this info refers to. 4896 const JobAction *JA = nullptr; 4897 /// The offload actions we need to take care off if this action is 4898 /// collapsed. 4899 ActionList SavedOffloadAction; 4900 }; 4901 4902 /// Append collapsed offload actions from the give nnumber of elements in the 4903 /// action info array. 4904 static void AppendCollapsedOffloadAction(ActionList &CollapsedOffloadAction, 4905 ArrayRef<JobActionInfo> &ActionInfo, 4906 unsigned ElementNum) { 4907 assert(ElementNum <= ActionInfo.size() && "Invalid number of elements."); 4908 for (unsigned I = 0; I < ElementNum; ++I) 4909 CollapsedOffloadAction.append(ActionInfo[I].SavedOffloadAction.begin(), 4910 ActionInfo[I].SavedOffloadAction.end()); 4911 } 4912 4913 /// Functions that attempt to perform the combining. They detect if that is 4914 /// legal, and if so they update the inputs \a Inputs and the offload action 4915 /// that were collapsed in \a CollapsedOffloadAction. A tool that deals with 4916 /// the combined action is returned. If the combining is not legal or if the 4917 /// tool does not exist, null is returned. 4918 /// Currently three kinds of collapsing are supported: 4919 /// - Assemble + Backend + Compile; 4920 /// - Assemble + Backend ; 4921 /// - Backend + Compile. 4922 const Tool * 4923 combineAssembleBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4924 ActionList &Inputs, 4925 ActionList &CollapsedOffloadAction) { 4926 if (ActionInfo.size() < 3 || !canCollapseAssembleAction()) 4927 return nullptr; 4928 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4929 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4930 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[2].JA); 4931 if (!AJ || !BJ || !CJ) 4932 return nullptr; 4933 4934 // Get compiler tool. 4935 const Tool *T = TC.SelectTool(*CJ); 4936 if (!T) 4937 return nullptr; 4938 4939 // Can't collapse if we don't have codegen support unless we are 4940 // emitting LLVM IR. 4941 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 4942 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 4943 return nullptr; 4944 4945 // When using -fembed-bitcode, it is required to have the same tool (clang) 4946 // for both CompilerJA and BackendJA. Otherwise, combine two stages. 4947 if (EmbedBitcode) { 4948 const Tool *BT = TC.SelectTool(*BJ); 4949 if (BT == T) 4950 return nullptr; 4951 } 4952 4953 if (!T->hasIntegratedAssembler()) 4954 return nullptr; 4955 4956 Inputs = CJ->getInputs(); 4957 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4958 /*NumElements=*/3); 4959 return T; 4960 } 4961 const Tool *combineAssembleBackend(ArrayRef<JobActionInfo> ActionInfo, 4962 ActionList &Inputs, 4963 ActionList &CollapsedOffloadAction) { 4964 if (ActionInfo.size() < 2 || !canCollapseAssembleAction()) 4965 return nullptr; 4966 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4967 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4968 if (!AJ || !BJ) 4969 return nullptr; 4970 4971 // Get backend tool. 4972 const Tool *T = TC.SelectTool(*BJ); 4973 if (!T) 4974 return nullptr; 4975 4976 if (!T->hasIntegratedAssembler()) 4977 return nullptr; 4978 4979 Inputs = BJ->getInputs(); 4980 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4981 /*NumElements=*/2); 4982 return T; 4983 } 4984 const Tool *combineBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4985 ActionList &Inputs, 4986 ActionList &CollapsedOffloadAction) { 4987 if (ActionInfo.size() < 2) 4988 return nullptr; 4989 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[0].JA); 4990 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[1].JA); 4991 if (!BJ || !CJ) 4992 return nullptr; 4993 4994 // Check if the initial input (to the compile job or its predessor if one 4995 // exists) is LLVM bitcode. In that case, no preprocessor step is required 4996 // and we can still collapse the compile and backend jobs when we have 4997 // -save-temps. I.e. there is no need for a separate compile job just to 4998 // emit unoptimized bitcode. 4999 bool InputIsBitcode = true; 5000 for (size_t i = 1; i < ActionInfo.size(); i++) 5001 if (ActionInfo[i].JA->getType() != types::TY_LLVM_BC && 5002 ActionInfo[i].JA->getType() != types::TY_LTO_BC) { 5003 InputIsBitcode = false; 5004 break; 5005 } 5006 if (!InputIsBitcode && !canCollapsePreprocessorAction()) 5007 return nullptr; 5008 5009 // Get compiler tool. 5010 const Tool *T = TC.SelectTool(*CJ); 5011 if (!T) 5012 return nullptr; 5013 5014 // Can't collapse if we don't have codegen support unless we are 5015 // emitting LLVM IR. 5016 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 5017 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 5018 return nullptr; 5019 5020 if (T->canEmitIR() && ((SaveTemps && !InputIsBitcode) || EmbedBitcode)) 5021 return nullptr; 5022 5023 Inputs = CJ->getInputs(); 5024 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5025 /*NumElements=*/2); 5026 return T; 5027 } 5028 5029 /// Updates the inputs if the obtained tool supports combining with 5030 /// preprocessor action, and the current input is indeed a preprocessor 5031 /// action. If combining results in the collapse of offloading actions, those 5032 /// are appended to \a CollapsedOffloadAction. 5033 void combineWithPreprocessor(const Tool *T, ActionList &Inputs, 5034 ActionList &CollapsedOffloadAction) { 5035 if (!T || !canCollapsePreprocessorAction() || !T->hasIntegratedCPP()) 5036 return; 5037 5038 // Attempt to get a preprocessor action dependence. 5039 ActionList PreprocessJobOffloadActions; 5040 ActionList NewInputs; 5041 for (Action *A : Inputs) { 5042 auto *PJ = getPrevDependentAction({A}, PreprocessJobOffloadActions); 5043 if (!PJ || !isa<PreprocessJobAction>(PJ)) { 5044 NewInputs.push_back(A); 5045 continue; 5046 } 5047 5048 // This is legal to combine. Append any offload action we found and add the 5049 // current input to preprocessor inputs. 5050 CollapsedOffloadAction.append(PreprocessJobOffloadActions.begin(), 5051 PreprocessJobOffloadActions.end()); 5052 NewInputs.append(PJ->input_begin(), PJ->input_end()); 5053 } 5054 Inputs = NewInputs; 5055 } 5056 5057 public: 5058 ToolSelector(const JobAction *BaseAction, const ToolChain &TC, 5059 const Compilation &C, bool SaveTemps, bool EmbedBitcode) 5060 : TC(TC), C(C), BaseAction(BaseAction), SaveTemps(SaveTemps), 5061 EmbedBitcode(EmbedBitcode) { 5062 assert(BaseAction && "Invalid base action."); 5063 IsHostSelector = BaseAction->getOffloadingDeviceKind() == Action::OFK_None; 5064 } 5065 5066 /// Check if a chain of actions can be combined and return the tool that can 5067 /// handle the combination of actions. The pointer to the current inputs \a 5068 /// Inputs and the list of offload actions \a CollapsedOffloadActions 5069 /// connected to collapsed actions are updated accordingly. The latter enables 5070 /// the caller of the selector to process them afterwards instead of just 5071 /// dropping them. If no suitable tool is found, null will be returned. 5072 const Tool *getTool(ActionList &Inputs, 5073 ActionList &CollapsedOffloadAction) { 5074 // 5075 // Get the largest chain of actions that we could combine. 5076 // 5077 5078 SmallVector<JobActionInfo, 5> ActionChain(1); 5079 ActionChain.back().JA = BaseAction; 5080 while (ActionChain.back().JA) { 5081 const Action *CurAction = ActionChain.back().JA; 5082 5083 // Grow the chain by one element. 5084 ActionChain.resize(ActionChain.size() + 1); 5085 JobActionInfo &AI = ActionChain.back(); 5086 5087 // Attempt to fill it with the 5088 AI.JA = 5089 getPrevDependentAction(CurAction->getInputs(), AI.SavedOffloadAction); 5090 } 5091 5092 // Pop the last action info as it could not be filled. 5093 ActionChain.pop_back(); 5094 5095 // 5096 // Attempt to combine actions. If all combining attempts failed, just return 5097 // the tool of the provided action. At the end we attempt to combine the 5098 // action with any preprocessor action it may depend on. 5099 // 5100 5101 const Tool *T = combineAssembleBackendCompile(ActionChain, Inputs, 5102 CollapsedOffloadAction); 5103 if (!T) 5104 T = combineAssembleBackend(ActionChain, Inputs, CollapsedOffloadAction); 5105 if (!T) 5106 T = combineBackendCompile(ActionChain, Inputs, CollapsedOffloadAction); 5107 if (!T) { 5108 Inputs = BaseAction->getInputs(); 5109 T = TC.SelectTool(*BaseAction); 5110 } 5111 5112 combineWithPreprocessor(T, Inputs, CollapsedOffloadAction); 5113 return T; 5114 } 5115 }; 5116 } 5117 5118 /// Return a string that uniquely identifies the result of a job. The bound arch 5119 /// is not necessarily represented in the toolchain's triple -- for example, 5120 /// armv7 and armv7s both map to the same triple -- so we need both in our map. 5121 /// Also, we need to add the offloading device kind, as the same tool chain can 5122 /// be used for host and device for some programming models, e.g. OpenMP. 5123 static std::string GetTriplePlusArchString(const ToolChain *TC, 5124 StringRef BoundArch, 5125 Action::OffloadKind OffloadKind) { 5126 std::string TriplePlusArch = TC->getTriple().normalize(); 5127 if (!BoundArch.empty()) { 5128 TriplePlusArch += "-"; 5129 TriplePlusArch += BoundArch; 5130 } 5131 TriplePlusArch += "-"; 5132 TriplePlusArch += Action::GetOffloadKindName(OffloadKind); 5133 return TriplePlusArch; 5134 } 5135 5136 InputInfoList Driver::BuildJobsForAction( 5137 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5138 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5139 std::map<std::pair<const Action *, std::string>, InputInfoList> 5140 &CachedResults, 5141 Action::OffloadKind TargetDeviceOffloadKind) const { 5142 std::pair<const Action *, std::string> ActionTC = { 5143 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5144 auto CachedResult = CachedResults.find(ActionTC); 5145 if (CachedResult != CachedResults.end()) { 5146 return CachedResult->second; 5147 } 5148 InputInfoList Result = BuildJobsForActionNoCache( 5149 C, A, TC, BoundArch, AtTopLevel, MultipleArchs, LinkingOutput, 5150 CachedResults, TargetDeviceOffloadKind); 5151 CachedResults[ActionTC] = Result; 5152 return Result; 5153 } 5154 5155 InputInfoList Driver::BuildJobsForActionNoCache( 5156 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5157 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5158 std::map<std::pair<const Action *, std::string>, InputInfoList> 5159 &CachedResults, 5160 Action::OffloadKind TargetDeviceOffloadKind) const { 5161 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 5162 5163 InputInfoList OffloadDependencesInputInfo; 5164 bool BuildingForOffloadDevice = TargetDeviceOffloadKind != Action::OFK_None; 5165 if (const OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 5166 // The 'Darwin' toolchain is initialized only when its arguments are 5167 // computed. Get the default arguments for OFK_None to ensure that 5168 // initialization is performed before processing the offload action. 5169 // FIXME: Remove when darwin's toolchain is initialized during construction. 5170 C.getArgsForToolChain(TC, BoundArch, Action::OFK_None); 5171 5172 // The offload action is expected to be used in four different situations. 5173 // 5174 // a) Set a toolchain/architecture/kind for a host action: 5175 // Host Action 1 -> OffloadAction -> Host Action 2 5176 // 5177 // b) Set a toolchain/architecture/kind for a device action; 5178 // Device Action 1 -> OffloadAction -> Device Action 2 5179 // 5180 // c) Specify a device dependence to a host action; 5181 // Device Action 1 _ 5182 // \ 5183 // Host Action 1 ---> OffloadAction -> Host Action 2 5184 // 5185 // d) Specify a host dependence to a device action. 5186 // Host Action 1 _ 5187 // \ 5188 // Device Action 1 ---> OffloadAction -> Device Action 2 5189 // 5190 // For a) and b), we just return the job generated for the dependence. For 5191 // c) and d) we override the current action with the host/device dependence 5192 // if the current toolchain is host/device and set the offload dependences 5193 // info with the jobs obtained from the device/host dependence(s). 5194 5195 // If there is a single device option, just generate the job for it. 5196 if (OA->hasSingleDeviceDependence()) { 5197 InputInfoList DevA; 5198 OA->doOnEachDeviceDependence([&](Action *DepA, const ToolChain *DepTC, 5199 const char *DepBoundArch) { 5200 DevA = 5201 BuildJobsForAction(C, DepA, DepTC, DepBoundArch, AtTopLevel, 5202 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, 5203 CachedResults, DepA->getOffloadingDeviceKind()); 5204 }); 5205 return DevA; 5206 } 5207 5208 // If 'Action 2' is host, we generate jobs for the device dependences and 5209 // override the current action with the host dependence. Otherwise, we 5210 // generate the host dependences and override the action with the device 5211 // dependence. The dependences can't therefore be a top-level action. 5212 OA->doOnEachDependence( 5213 /*IsHostDependence=*/BuildingForOffloadDevice, 5214 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5215 OffloadDependencesInputInfo.append(BuildJobsForAction( 5216 C, DepA, DepTC, DepBoundArch, /*AtTopLevel=*/false, 5217 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, CachedResults, 5218 DepA->getOffloadingDeviceKind())); 5219 }); 5220 5221 A = BuildingForOffloadDevice 5222 ? OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true) 5223 : OA->getHostDependence(); 5224 5225 // We may have already built this action as a part of the offloading 5226 // toolchain, return the cached input if so. 5227 std::pair<const Action *, std::string> ActionTC = { 5228 OA->getHostDependence(), 5229 GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5230 if (CachedResults.find(ActionTC) != CachedResults.end()) { 5231 InputInfoList Inputs = CachedResults[ActionTC]; 5232 Inputs.append(OffloadDependencesInputInfo); 5233 return Inputs; 5234 } 5235 } 5236 5237 if (const InputAction *IA = dyn_cast<InputAction>(A)) { 5238 // FIXME: It would be nice to not claim this here; maybe the old scheme of 5239 // just using Args was better? 5240 const Arg &Input = IA->getInputArg(); 5241 Input.claim(); 5242 if (Input.getOption().matches(options::OPT_INPUT)) { 5243 const char *Name = Input.getValue(); 5244 return {InputInfo(A, Name, /* _BaseInput = */ Name)}; 5245 } 5246 return {InputInfo(A, &Input, /* _BaseInput = */ "")}; 5247 } 5248 5249 if (const BindArchAction *BAA = dyn_cast<BindArchAction>(A)) { 5250 const ToolChain *TC; 5251 StringRef ArchName = BAA->getArchName(); 5252 5253 if (!ArchName.empty()) 5254 TC = &getToolChain(C.getArgs(), 5255 computeTargetTriple(*this, TargetTriple, 5256 C.getArgs(), ArchName)); 5257 else 5258 TC = &C.getDefaultToolChain(); 5259 5260 return BuildJobsForAction(C, *BAA->input_begin(), TC, ArchName, AtTopLevel, 5261 MultipleArchs, LinkingOutput, CachedResults, 5262 TargetDeviceOffloadKind); 5263 } 5264 5265 5266 ActionList Inputs = A->getInputs(); 5267 5268 const JobAction *JA = cast<JobAction>(A); 5269 ActionList CollapsedOffloadActions; 5270 5271 ToolSelector TS(JA, *TC, C, isSaveTempsEnabled(), 5272 embedBitcodeInObject() && !isUsingLTO()); 5273 const Tool *T = TS.getTool(Inputs, CollapsedOffloadActions); 5274 5275 if (!T) 5276 return {InputInfo()}; 5277 5278 if (BuildingForOffloadDevice && 5279 A->getOffloadingDeviceKind() == Action::OFK_OpenMP) { 5280 if (TC->getTriple().isAMDGCN()) { 5281 // AMDGCN treats backend and assemble actions as no-op because 5282 // linker does not support object files. 5283 if (const BackendJobAction *BA = dyn_cast<BackendJobAction>(A)) { 5284 return BuildJobsForAction(C, *BA->input_begin(), TC, BoundArch, 5285 AtTopLevel, MultipleArchs, LinkingOutput, 5286 CachedResults, TargetDeviceOffloadKind); 5287 } 5288 5289 if (const AssembleJobAction *AA = dyn_cast<AssembleJobAction>(A)) { 5290 return BuildJobsForAction(C, *AA->input_begin(), TC, BoundArch, 5291 AtTopLevel, MultipleArchs, LinkingOutput, 5292 CachedResults, TargetDeviceOffloadKind); 5293 } 5294 } 5295 } 5296 5297 // If we've collapsed action list that contained OffloadAction we 5298 // need to build jobs for host/device-side inputs it may have held. 5299 for (const auto *OA : CollapsedOffloadActions) 5300 cast<OffloadAction>(OA)->doOnEachDependence( 5301 /*IsHostDependence=*/BuildingForOffloadDevice, 5302 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5303 OffloadDependencesInputInfo.append(BuildJobsForAction( 5304 C, DepA, DepTC, DepBoundArch, /* AtTopLevel */ false, 5305 /*MultipleArchs=*/!!DepBoundArch, LinkingOutput, CachedResults, 5306 DepA->getOffloadingDeviceKind())); 5307 }); 5308 5309 // Only use pipes when there is exactly one input. 5310 InputInfoList InputInfos; 5311 for (const Action *Input : Inputs) { 5312 // Treat dsymutil and verify sub-jobs as being at the top-level too, they 5313 // shouldn't get temporary output names. 5314 // FIXME: Clean this up. 5315 bool SubJobAtTopLevel = 5316 AtTopLevel && (isa<DsymutilJobAction>(A) || isa<VerifyJobAction>(A)); 5317 InputInfos.append(BuildJobsForAction( 5318 C, Input, TC, BoundArch, SubJobAtTopLevel, MultipleArchs, LinkingOutput, 5319 CachedResults, A->getOffloadingDeviceKind())); 5320 } 5321 5322 // Always use the first file input as the base input. 5323 const char *BaseInput = InputInfos[0].getBaseInput(); 5324 for (auto &Info : InputInfos) { 5325 if (Info.isFilename()) { 5326 BaseInput = Info.getBaseInput(); 5327 break; 5328 } 5329 } 5330 5331 // ... except dsymutil actions, which use their actual input as the base 5332 // input. 5333 if (JA->getType() == types::TY_dSYM) 5334 BaseInput = InputInfos[0].getFilename(); 5335 5336 // ... and in header module compilations, which use the module name. 5337 if (auto *ModuleJA = dyn_cast<HeaderModulePrecompileJobAction>(JA)) 5338 BaseInput = ModuleJA->getModuleName(); 5339 5340 // Append outputs of offload device jobs to the input list 5341 if (!OffloadDependencesInputInfo.empty()) 5342 InputInfos.append(OffloadDependencesInputInfo.begin(), 5343 OffloadDependencesInputInfo.end()); 5344 5345 // Set the effective triple of the toolchain for the duration of this job. 5346 llvm::Triple EffectiveTriple; 5347 const ToolChain &ToolTC = T->getToolChain(); 5348 const ArgList &Args = 5349 C.getArgsForToolChain(TC, BoundArch, A->getOffloadingDeviceKind()); 5350 if (InputInfos.size() != 1) { 5351 EffectiveTriple = llvm::Triple(ToolTC.ComputeEffectiveClangTriple(Args)); 5352 } else { 5353 // Pass along the input type if it can be unambiguously determined. 5354 EffectiveTriple = llvm::Triple( 5355 ToolTC.ComputeEffectiveClangTriple(Args, InputInfos[0].getType())); 5356 } 5357 RegisterEffectiveTriple TripleRAII(ToolTC, EffectiveTriple); 5358 5359 // Determine the place to write output to, if any. 5360 InputInfo Result; 5361 InputInfoList UnbundlingResults; 5362 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(JA)) { 5363 // If we have an unbundling job, we need to create results for all the 5364 // outputs. We also update the results cache so that other actions using 5365 // this unbundling action can get the right results. 5366 for (auto &UI : UA->getDependentActionsInfo()) { 5367 assert(UI.DependentOffloadKind != Action::OFK_None && 5368 "Unbundling with no offloading??"); 5369 5370 // Unbundling actions are never at the top level. When we generate the 5371 // offloading prefix, we also do that for the host file because the 5372 // unbundling action does not change the type of the output which can 5373 // cause a overwrite. 5374 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5375 UI.DependentOffloadKind, 5376 UI.DependentToolChain->getTriple().normalize(), 5377 /*CreatePrefixForHost=*/true); 5378 auto CurI = InputInfo( 5379 UA, 5380 GetNamedOutputPath(C, *UA, BaseInput, UI.DependentBoundArch, 5381 /*AtTopLevel=*/false, 5382 MultipleArchs || 5383 UI.DependentOffloadKind == Action::OFK_HIP, 5384 OffloadingPrefix), 5385 BaseInput); 5386 // Save the unbundling result. 5387 UnbundlingResults.push_back(CurI); 5388 5389 // Get the unique string identifier for this dependence and cache the 5390 // result. 5391 StringRef Arch; 5392 if (TargetDeviceOffloadKind == Action::OFK_HIP) { 5393 if (UI.DependentOffloadKind == Action::OFK_Host) 5394 Arch = StringRef(); 5395 else 5396 Arch = UI.DependentBoundArch; 5397 } else 5398 Arch = BoundArch; 5399 5400 CachedResults[{A, GetTriplePlusArchString(UI.DependentToolChain, Arch, 5401 UI.DependentOffloadKind)}] = { 5402 CurI}; 5403 } 5404 5405 // Now that we have all the results generated, select the one that should be 5406 // returned for the current depending action. 5407 std::pair<const Action *, std::string> ActionTC = { 5408 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5409 assert(CachedResults.find(ActionTC) != CachedResults.end() && 5410 "Result does not exist??"); 5411 Result = CachedResults[ActionTC].front(); 5412 } else if (JA->getType() == types::TY_Nothing) 5413 Result = {InputInfo(A, BaseInput)}; 5414 else { 5415 // We only have to generate a prefix for the host if this is not a top-level 5416 // action. 5417 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5418 A->getOffloadingDeviceKind(), TC->getTriple().normalize(), 5419 /*CreatePrefixForHost=*/!!A->getOffloadingHostActiveKinds() && 5420 !AtTopLevel); 5421 if (isa<OffloadWrapperJobAction>(JA)) { 5422 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5423 BaseInput = FinalOutput->getValue(); 5424 else 5425 BaseInput = getDefaultImageName(); 5426 BaseInput = 5427 C.getArgs().MakeArgString(std::string(BaseInput) + "-wrapper"); 5428 } 5429 Result = InputInfo(A, GetNamedOutputPath(C, *JA, BaseInput, BoundArch, 5430 AtTopLevel, MultipleArchs, 5431 OffloadingPrefix), 5432 BaseInput); 5433 } 5434 5435 if (CCCPrintBindings && !CCGenDiagnostics) { 5436 llvm::errs() << "# \"" << T->getToolChain().getTripleString() << '"' 5437 << " - \"" << T->getName() << "\", inputs: ["; 5438 for (unsigned i = 0, e = InputInfos.size(); i != e; ++i) { 5439 llvm::errs() << InputInfos[i].getAsString(); 5440 if (i + 1 != e) 5441 llvm::errs() << ", "; 5442 } 5443 if (UnbundlingResults.empty()) 5444 llvm::errs() << "], output: " << Result.getAsString() << "\n"; 5445 else { 5446 llvm::errs() << "], outputs: ["; 5447 for (unsigned i = 0, e = UnbundlingResults.size(); i != e; ++i) { 5448 llvm::errs() << UnbundlingResults[i].getAsString(); 5449 if (i + 1 != e) 5450 llvm::errs() << ", "; 5451 } 5452 llvm::errs() << "] \n"; 5453 } 5454 } else { 5455 if (UnbundlingResults.empty()) 5456 T->ConstructJob( 5457 C, *JA, Result, InputInfos, 5458 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5459 LinkingOutput); 5460 else 5461 T->ConstructJobMultipleOutputs( 5462 C, *JA, UnbundlingResults, InputInfos, 5463 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5464 LinkingOutput); 5465 } 5466 return {Result}; 5467 } 5468 5469 const char *Driver::getDefaultImageName() const { 5470 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 5471 return Target.isOSWindows() ? "a.exe" : "a.out"; 5472 } 5473 5474 /// Create output filename based on ArgValue, which could either be a 5475 /// full filename, filename without extension, or a directory. If ArgValue 5476 /// does not provide a filename, then use BaseName, and use the extension 5477 /// suitable for FileType. 5478 static const char *MakeCLOutputFilename(const ArgList &Args, StringRef ArgValue, 5479 StringRef BaseName, 5480 types::ID FileType) { 5481 SmallString<128> Filename = ArgValue; 5482 5483 if (ArgValue.empty()) { 5484 // If the argument is empty, output to BaseName in the current dir. 5485 Filename = BaseName; 5486 } else if (llvm::sys::path::is_separator(Filename.back())) { 5487 // If the argument is a directory, output to BaseName in that dir. 5488 llvm::sys::path::append(Filename, BaseName); 5489 } 5490 5491 if (!llvm::sys::path::has_extension(ArgValue)) { 5492 // If the argument didn't provide an extension, then set it. 5493 const char *Extension = types::getTypeTempSuffix(FileType, true); 5494 5495 if (FileType == types::TY_Image && 5496 Args.hasArg(options::OPT__SLASH_LD, options::OPT__SLASH_LDd)) { 5497 // The output file is a dll. 5498 Extension = "dll"; 5499 } 5500 5501 llvm::sys::path::replace_extension(Filename, Extension); 5502 } 5503 5504 return Args.MakeArgString(Filename.c_str()); 5505 } 5506 5507 static bool HasPreprocessOutput(const Action &JA) { 5508 if (isa<PreprocessJobAction>(JA)) 5509 return true; 5510 if (isa<OffloadAction>(JA) && isa<PreprocessJobAction>(JA.getInputs()[0])) 5511 return true; 5512 if (isa<OffloadBundlingJobAction>(JA) && 5513 HasPreprocessOutput(*(JA.getInputs()[0]))) 5514 return true; 5515 return false; 5516 } 5517 5518 const char *Driver::GetNamedOutputPath(Compilation &C, const JobAction &JA, 5519 const char *BaseInput, 5520 StringRef OrigBoundArch, bool AtTopLevel, 5521 bool MultipleArchs, 5522 StringRef OffloadingPrefix) const { 5523 std::string BoundArch = OrigBoundArch.str(); 5524 if (is_style_windows(llvm::sys::path::Style::native)) { 5525 // BoundArch may contains ':', which is invalid in file names on Windows, 5526 // therefore replace it with '%'. 5527 std::replace(BoundArch.begin(), BoundArch.end(), ':', '@'); 5528 } 5529 5530 llvm::PrettyStackTraceString CrashInfo("Computing output path"); 5531 // Output to a user requested destination? 5532 if (AtTopLevel && !isa<DsymutilJobAction>(JA) && !isa<VerifyJobAction>(JA)) { 5533 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5534 return C.addResultFile(FinalOutput->getValue(), &JA); 5535 } 5536 5537 // For /P, preprocess to file named after BaseInput. 5538 if (C.getArgs().hasArg(options::OPT__SLASH_P)) { 5539 assert(AtTopLevel && isa<PreprocessJobAction>(JA)); 5540 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5541 StringRef NameArg; 5542 if (Arg *A = C.getArgs().getLastArg(options::OPT__SLASH_Fi)) 5543 NameArg = A->getValue(); 5544 return C.addResultFile( 5545 MakeCLOutputFilename(C.getArgs(), NameArg, BaseName, types::TY_PP_C), 5546 &JA); 5547 } 5548 5549 // Default to writing to stdout? 5550 if (AtTopLevel && !CCGenDiagnostics && HasPreprocessOutput(JA)) { 5551 return "-"; 5552 } 5553 5554 if (JA.getType() == types::TY_ModuleFile && 5555 C.getArgs().getLastArg(options::OPT_module_file_info)) { 5556 return "-"; 5557 } 5558 5559 // Is this the assembly listing for /FA? 5560 if (JA.getType() == types::TY_PP_Asm && 5561 (C.getArgs().hasArg(options::OPT__SLASH_FA) || 5562 C.getArgs().hasArg(options::OPT__SLASH_Fa))) { 5563 // Use /Fa and the input filename to determine the asm file name. 5564 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5565 StringRef FaValue = C.getArgs().getLastArgValue(options::OPT__SLASH_Fa); 5566 return C.addResultFile( 5567 MakeCLOutputFilename(C.getArgs(), FaValue, BaseName, JA.getType()), 5568 &JA); 5569 } 5570 5571 // Output to a temporary file? 5572 if ((!AtTopLevel && !isSaveTempsEnabled() && 5573 !C.getArgs().hasArg(options::OPT__SLASH_Fo)) || 5574 CCGenDiagnostics) { 5575 StringRef Name = llvm::sys::path::filename(BaseInput); 5576 std::pair<StringRef, StringRef> Split = Name.split('.'); 5577 SmallString<128> TmpName; 5578 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5579 Arg *A = C.getArgs().getLastArg(options::OPT_fcrash_diagnostics_dir); 5580 if (CCGenDiagnostics && A) { 5581 SmallString<128> CrashDirectory(A->getValue()); 5582 if (!getVFS().exists(CrashDirectory)) 5583 llvm::sys::fs::create_directories(CrashDirectory); 5584 llvm::sys::path::append(CrashDirectory, Split.first); 5585 const char *Middle = Suffix ? "-%%%%%%." : "-%%%%%%"; 5586 std::error_code EC = llvm::sys::fs::createUniqueFile( 5587 CrashDirectory + Middle + Suffix, TmpName); 5588 if (EC) { 5589 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5590 return ""; 5591 } 5592 } else { 5593 if (MultipleArchs && !BoundArch.empty()) { 5594 TmpName = GetTemporaryDirectory(Split.first); 5595 llvm::sys::path::append(TmpName, 5596 Split.first + "-" + BoundArch + "." + Suffix); 5597 } else { 5598 TmpName = GetTemporaryPath(Split.first, Suffix); 5599 } 5600 } 5601 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5602 } 5603 5604 SmallString<128> BasePath(BaseInput); 5605 SmallString<128> ExternalPath(""); 5606 StringRef BaseName; 5607 5608 // Dsymutil actions should use the full path. 5609 if (isa<DsymutilJobAction>(JA) && C.getArgs().hasArg(options::OPT_dsym_dir)) { 5610 ExternalPath += C.getArgs().getLastArg(options::OPT_dsym_dir)->getValue(); 5611 // We use posix style here because the tests (specifically 5612 // darwin-dsymutil.c) demonstrate that posix style paths are acceptable 5613 // even on Windows and if we don't then the similar test covering this 5614 // fails. 5615 llvm::sys::path::append(ExternalPath, llvm::sys::path::Style::posix, 5616 llvm::sys::path::filename(BasePath)); 5617 BaseName = ExternalPath; 5618 } else if (isa<DsymutilJobAction>(JA) || isa<VerifyJobAction>(JA)) 5619 BaseName = BasePath; 5620 else 5621 BaseName = llvm::sys::path::filename(BasePath); 5622 5623 // Determine what the derived output name should be. 5624 const char *NamedOutput; 5625 5626 if ((JA.getType() == types::TY_Object || JA.getType() == types::TY_LTO_BC) && 5627 C.getArgs().hasArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o)) { 5628 // The /Fo or /o flag decides the object filename. 5629 StringRef Val = 5630 C.getArgs() 5631 .getLastArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o) 5632 ->getValue(); 5633 NamedOutput = 5634 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5635 } else if (JA.getType() == types::TY_Image && 5636 C.getArgs().hasArg(options::OPT__SLASH_Fe, 5637 options::OPT__SLASH_o)) { 5638 // The /Fe or /o flag names the linked file. 5639 StringRef Val = 5640 C.getArgs() 5641 .getLastArg(options::OPT__SLASH_Fe, options::OPT__SLASH_o) 5642 ->getValue(); 5643 NamedOutput = 5644 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Image); 5645 } else if (JA.getType() == types::TY_Image) { 5646 if (IsCLMode()) { 5647 // clang-cl uses BaseName for the executable name. 5648 NamedOutput = 5649 MakeCLOutputFilename(C.getArgs(), "", BaseName, types::TY_Image); 5650 } else { 5651 SmallString<128> Output(getDefaultImageName()); 5652 // HIP image for device compilation with -fno-gpu-rdc is per compilation 5653 // unit. 5654 bool IsHIPNoRDC = JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5655 !C.getArgs().hasFlag(options::OPT_fgpu_rdc, 5656 options::OPT_fno_gpu_rdc, false); 5657 if (IsHIPNoRDC) { 5658 Output = BaseName; 5659 llvm::sys::path::replace_extension(Output, ""); 5660 } 5661 Output += OffloadingPrefix; 5662 if (MultipleArchs && !BoundArch.empty()) { 5663 Output += "-"; 5664 Output.append(BoundArch); 5665 } 5666 if (IsHIPNoRDC) 5667 Output += ".out"; 5668 NamedOutput = C.getArgs().MakeArgString(Output.c_str()); 5669 } 5670 } else if (JA.getType() == types::TY_PCH && IsCLMode()) { 5671 NamedOutput = C.getArgs().MakeArgString(GetClPchPath(C, BaseName)); 5672 } else { 5673 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5674 assert(Suffix && "All types used for output should have a suffix."); 5675 5676 std::string::size_type End = std::string::npos; 5677 if (!types::appendSuffixForType(JA.getType())) 5678 End = BaseName.rfind('.'); 5679 SmallString<128> Suffixed(BaseName.substr(0, End)); 5680 Suffixed += OffloadingPrefix; 5681 if (MultipleArchs && !BoundArch.empty()) { 5682 Suffixed += "-"; 5683 Suffixed.append(BoundArch); 5684 } 5685 // When using both -save-temps and -emit-llvm, use a ".tmp.bc" suffix for 5686 // the unoptimized bitcode so that it does not get overwritten by the ".bc" 5687 // optimized bitcode output. 5688 auto IsHIPRDCInCompilePhase = [](const JobAction &JA, 5689 const llvm::opt::DerivedArgList &Args) { 5690 // The relocatable compilation in HIP implies -emit-llvm. Similarly, use a 5691 // ".tmp.bc" suffix for the unoptimized bitcode (generated in the compile 5692 // phase.) 5693 return isa<CompileJobAction>(JA) && 5694 JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5695 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 5696 false); 5697 }; 5698 if (!AtTopLevel && JA.getType() == types::TY_LLVM_BC && 5699 (C.getArgs().hasArg(options::OPT_emit_llvm) || 5700 IsHIPRDCInCompilePhase(JA, C.getArgs()))) 5701 Suffixed += ".tmp"; 5702 Suffixed += '.'; 5703 Suffixed += Suffix; 5704 NamedOutput = C.getArgs().MakeArgString(Suffixed.c_str()); 5705 } 5706 5707 // Prepend object file path if -save-temps=obj 5708 if (!AtTopLevel && isSaveTempsObj() && C.getArgs().hasArg(options::OPT_o) && 5709 JA.getType() != types::TY_PCH) { 5710 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 5711 SmallString<128> TempPath(FinalOutput->getValue()); 5712 llvm::sys::path::remove_filename(TempPath); 5713 StringRef OutputFileName = llvm::sys::path::filename(NamedOutput); 5714 llvm::sys::path::append(TempPath, OutputFileName); 5715 NamedOutput = C.getArgs().MakeArgString(TempPath.c_str()); 5716 } 5717 5718 // If we're saving temps and the temp file conflicts with the input file, 5719 // then avoid overwriting input file. 5720 if (!AtTopLevel && isSaveTempsEnabled() && NamedOutput == BaseName) { 5721 bool SameFile = false; 5722 SmallString<256> Result; 5723 llvm::sys::fs::current_path(Result); 5724 llvm::sys::path::append(Result, BaseName); 5725 llvm::sys::fs::equivalent(BaseInput, Result.c_str(), SameFile); 5726 // Must share the same path to conflict. 5727 if (SameFile) { 5728 StringRef Name = llvm::sys::path::filename(BaseInput); 5729 std::pair<StringRef, StringRef> Split = Name.split('.'); 5730 std::string TmpName = GetTemporaryPath( 5731 Split.first, types::getTypeTempSuffix(JA.getType(), IsCLMode())); 5732 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5733 } 5734 } 5735 5736 // As an annoying special case, PCH generation doesn't strip the pathname. 5737 if (JA.getType() == types::TY_PCH && !IsCLMode()) { 5738 llvm::sys::path::remove_filename(BasePath); 5739 if (BasePath.empty()) 5740 BasePath = NamedOutput; 5741 else 5742 llvm::sys::path::append(BasePath, NamedOutput); 5743 return C.addResultFile(C.getArgs().MakeArgString(BasePath.c_str()), &JA); 5744 } else { 5745 return C.addResultFile(NamedOutput, &JA); 5746 } 5747 } 5748 5749 std::string Driver::GetFilePath(StringRef Name, const ToolChain &TC) const { 5750 // Search for Name in a list of paths. 5751 auto SearchPaths = [&](const llvm::SmallVectorImpl<std::string> &P) 5752 -> llvm::Optional<std::string> { 5753 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5754 // attempting to use this prefix when looking for file paths. 5755 for (const auto &Dir : P) { 5756 if (Dir.empty()) 5757 continue; 5758 SmallString<128> P(Dir[0] == '=' ? SysRoot + Dir.substr(1) : Dir); 5759 llvm::sys::path::append(P, Name); 5760 if (llvm::sys::fs::exists(Twine(P))) 5761 return std::string(P); 5762 } 5763 return None; 5764 }; 5765 5766 if (auto P = SearchPaths(PrefixDirs)) 5767 return *P; 5768 5769 SmallString<128> R(ResourceDir); 5770 llvm::sys::path::append(R, Name); 5771 if (llvm::sys::fs::exists(Twine(R))) 5772 return std::string(R.str()); 5773 5774 SmallString<128> P(TC.getCompilerRTPath()); 5775 llvm::sys::path::append(P, Name); 5776 if (llvm::sys::fs::exists(Twine(P))) 5777 return std::string(P.str()); 5778 5779 SmallString<128> D(Dir); 5780 llvm::sys::path::append(D, "..", Name); 5781 if (llvm::sys::fs::exists(Twine(D))) 5782 return std::string(D.str()); 5783 5784 if (auto P = SearchPaths(TC.getLibraryPaths())) 5785 return *P; 5786 5787 if (auto P = SearchPaths(TC.getFilePaths())) 5788 return *P; 5789 5790 return std::string(Name); 5791 } 5792 5793 void Driver::generatePrefixedToolNames( 5794 StringRef Tool, const ToolChain &TC, 5795 SmallVectorImpl<std::string> &Names) const { 5796 // FIXME: Needs a better variable than TargetTriple 5797 Names.emplace_back((TargetTriple + "-" + Tool).str()); 5798 Names.emplace_back(Tool); 5799 } 5800 5801 static bool ScanDirForExecutable(SmallString<128> &Dir, StringRef Name) { 5802 llvm::sys::path::append(Dir, Name); 5803 if (llvm::sys::fs::can_execute(Twine(Dir))) 5804 return true; 5805 llvm::sys::path::remove_filename(Dir); 5806 return false; 5807 } 5808 5809 std::string Driver::GetProgramPath(StringRef Name, const ToolChain &TC) const { 5810 SmallVector<std::string, 2> TargetSpecificExecutables; 5811 generatePrefixedToolNames(Name, TC, TargetSpecificExecutables); 5812 5813 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5814 // attempting to use this prefix when looking for program paths. 5815 for (const auto &PrefixDir : PrefixDirs) { 5816 if (llvm::sys::fs::is_directory(PrefixDir)) { 5817 SmallString<128> P(PrefixDir); 5818 if (ScanDirForExecutable(P, Name)) 5819 return std::string(P.str()); 5820 } else { 5821 SmallString<128> P((PrefixDir + Name).str()); 5822 if (llvm::sys::fs::can_execute(Twine(P))) 5823 return std::string(P.str()); 5824 } 5825 } 5826 5827 const ToolChain::path_list &List = TC.getProgramPaths(); 5828 for (const auto &TargetSpecificExecutable : TargetSpecificExecutables) { 5829 // For each possible name of the tool look for it in 5830 // program paths first, then the path. 5831 // Higher priority names will be first, meaning that 5832 // a higher priority name in the path will be found 5833 // instead of a lower priority name in the program path. 5834 // E.g. <triple>-gcc on the path will be found instead 5835 // of gcc in the program path 5836 for (const auto &Path : List) { 5837 SmallString<128> P(Path); 5838 if (ScanDirForExecutable(P, TargetSpecificExecutable)) 5839 return std::string(P.str()); 5840 } 5841 5842 // Fall back to the path 5843 if (llvm::ErrorOr<std::string> P = 5844 llvm::sys::findProgramByName(TargetSpecificExecutable)) 5845 return *P; 5846 } 5847 5848 return std::string(Name); 5849 } 5850 5851 std::string Driver::GetTemporaryPath(StringRef Prefix, StringRef Suffix) const { 5852 SmallString<128> Path; 5853 std::error_code EC = llvm::sys::fs::createTemporaryFile(Prefix, Suffix, Path); 5854 if (EC) { 5855 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5856 return ""; 5857 } 5858 5859 return std::string(Path.str()); 5860 } 5861 5862 std::string Driver::GetTemporaryDirectory(StringRef Prefix) const { 5863 SmallString<128> Path; 5864 std::error_code EC = llvm::sys::fs::createUniqueDirectory(Prefix, Path); 5865 if (EC) { 5866 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5867 return ""; 5868 } 5869 5870 return std::string(Path.str()); 5871 } 5872 5873 std::string Driver::GetClPchPath(Compilation &C, StringRef BaseName) const { 5874 SmallString<128> Output; 5875 if (Arg *FpArg = C.getArgs().getLastArg(options::OPT__SLASH_Fp)) { 5876 // FIXME: If anybody needs it, implement this obscure rule: 5877 // "If you specify a directory without a file name, the default file name 5878 // is VCx0.pch., where x is the major version of Visual C++ in use." 5879 Output = FpArg->getValue(); 5880 5881 // "If you do not specify an extension as part of the path name, an 5882 // extension of .pch is assumed. " 5883 if (!llvm::sys::path::has_extension(Output)) 5884 Output += ".pch"; 5885 } else { 5886 if (Arg *YcArg = C.getArgs().getLastArg(options::OPT__SLASH_Yc)) 5887 Output = YcArg->getValue(); 5888 if (Output.empty()) 5889 Output = BaseName; 5890 llvm::sys::path::replace_extension(Output, ".pch"); 5891 } 5892 return std::string(Output.str()); 5893 } 5894 5895 const ToolChain &Driver::getToolChain(const ArgList &Args, 5896 const llvm::Triple &Target) const { 5897 5898 auto &TC = ToolChains[Target.str()]; 5899 if (!TC) { 5900 switch (Target.getOS()) { 5901 case llvm::Triple::AIX: 5902 TC = std::make_unique<toolchains::AIX>(*this, Target, Args); 5903 break; 5904 case llvm::Triple::Haiku: 5905 TC = std::make_unique<toolchains::Haiku>(*this, Target, Args); 5906 break; 5907 case llvm::Triple::Ananas: 5908 TC = std::make_unique<toolchains::Ananas>(*this, Target, Args); 5909 break; 5910 case llvm::Triple::CloudABI: 5911 TC = std::make_unique<toolchains::CloudABI>(*this, Target, Args); 5912 break; 5913 case llvm::Triple::Darwin: 5914 case llvm::Triple::MacOSX: 5915 case llvm::Triple::IOS: 5916 case llvm::Triple::TvOS: 5917 case llvm::Triple::WatchOS: 5918 case llvm::Triple::DriverKit: 5919 TC = std::make_unique<toolchains::DarwinClang>(*this, Target, Args); 5920 break; 5921 case llvm::Triple::DragonFly: 5922 TC = std::make_unique<toolchains::DragonFly>(*this, Target, Args); 5923 break; 5924 case llvm::Triple::OpenBSD: 5925 TC = std::make_unique<toolchains::OpenBSD>(*this, Target, Args); 5926 break; 5927 case llvm::Triple::NetBSD: 5928 TC = std::make_unique<toolchains::NetBSD>(*this, Target, Args); 5929 break; 5930 case llvm::Triple::FreeBSD: 5931 if (Target.isPPC()) 5932 TC = std::make_unique<toolchains::PPCFreeBSDToolChain>(*this, Target, 5933 Args); 5934 else 5935 TC = std::make_unique<toolchains::FreeBSD>(*this, Target, Args); 5936 break; 5937 case llvm::Triple::Minix: 5938 TC = std::make_unique<toolchains::Minix>(*this, Target, Args); 5939 break; 5940 case llvm::Triple::Linux: 5941 case llvm::Triple::ELFIAMCU: 5942 if (Target.getArch() == llvm::Triple::hexagon) 5943 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 5944 Args); 5945 else if ((Target.getVendor() == llvm::Triple::MipsTechnologies) && 5946 !Target.hasEnvironment()) 5947 TC = std::make_unique<toolchains::MipsLLVMToolChain>(*this, Target, 5948 Args); 5949 else if (Target.isPPC()) 5950 TC = std::make_unique<toolchains::PPCLinuxToolChain>(*this, Target, 5951 Args); 5952 else if (Target.getArch() == llvm::Triple::ve) 5953 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 5954 5955 else 5956 TC = std::make_unique<toolchains::Linux>(*this, Target, Args); 5957 break; 5958 case llvm::Triple::NaCl: 5959 TC = std::make_unique<toolchains::NaClToolChain>(*this, Target, Args); 5960 break; 5961 case llvm::Triple::Fuchsia: 5962 TC = std::make_unique<toolchains::Fuchsia>(*this, Target, Args); 5963 break; 5964 case llvm::Triple::Solaris: 5965 TC = std::make_unique<toolchains::Solaris>(*this, Target, Args); 5966 break; 5967 case llvm::Triple::AMDHSA: 5968 TC = std::make_unique<toolchains::ROCMToolChain>(*this, Target, Args); 5969 break; 5970 case llvm::Triple::AMDPAL: 5971 case llvm::Triple::Mesa3D: 5972 TC = std::make_unique<toolchains::AMDGPUToolChain>(*this, Target, Args); 5973 break; 5974 case llvm::Triple::Win32: 5975 switch (Target.getEnvironment()) { 5976 default: 5977 if (Target.isOSBinFormatELF()) 5978 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 5979 else if (Target.isOSBinFormatMachO()) 5980 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 5981 else 5982 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 5983 break; 5984 case llvm::Triple::GNU: 5985 TC = std::make_unique<toolchains::MinGW>(*this, Target, Args); 5986 break; 5987 case llvm::Triple::Itanium: 5988 TC = std::make_unique<toolchains::CrossWindowsToolChain>(*this, Target, 5989 Args); 5990 break; 5991 case llvm::Triple::MSVC: 5992 case llvm::Triple::UnknownEnvironment: 5993 if (Args.getLastArgValue(options::OPT_fuse_ld_EQ) 5994 .startswith_insensitive("bfd")) 5995 TC = std::make_unique<toolchains::CrossWindowsToolChain>( 5996 *this, Target, Args); 5997 else 5998 TC = 5999 std::make_unique<toolchains::MSVCToolChain>(*this, Target, Args); 6000 break; 6001 } 6002 break; 6003 case llvm::Triple::PS4: 6004 TC = std::make_unique<toolchains::PS4CPU>(*this, Target, Args); 6005 break; 6006 case llvm::Triple::PS5: 6007 TC = std::make_unique<toolchains::PS5CPU>(*this, Target, Args); 6008 break; 6009 case llvm::Triple::Contiki: 6010 TC = std::make_unique<toolchains::Contiki>(*this, Target, Args); 6011 break; 6012 case llvm::Triple::Hurd: 6013 TC = std::make_unique<toolchains::Hurd>(*this, Target, Args); 6014 break; 6015 case llvm::Triple::ZOS: 6016 TC = std::make_unique<toolchains::ZOS>(*this, Target, Args); 6017 break; 6018 case llvm::Triple::ShaderModel: 6019 TC = std::make_unique<toolchains::HLSLToolChain>(*this, Target, Args); 6020 break; 6021 default: 6022 // Of these targets, Hexagon is the only one that might have 6023 // an OS of Linux, in which case it got handled above already. 6024 switch (Target.getArch()) { 6025 case llvm::Triple::tce: 6026 TC = std::make_unique<toolchains::TCEToolChain>(*this, Target, Args); 6027 break; 6028 case llvm::Triple::tcele: 6029 TC = std::make_unique<toolchains::TCELEToolChain>(*this, Target, Args); 6030 break; 6031 case llvm::Triple::hexagon: 6032 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 6033 Args); 6034 break; 6035 case llvm::Triple::lanai: 6036 TC = std::make_unique<toolchains::LanaiToolChain>(*this, Target, Args); 6037 break; 6038 case llvm::Triple::xcore: 6039 TC = std::make_unique<toolchains::XCoreToolChain>(*this, Target, Args); 6040 break; 6041 case llvm::Triple::wasm32: 6042 case llvm::Triple::wasm64: 6043 TC = std::make_unique<toolchains::WebAssembly>(*this, Target, Args); 6044 break; 6045 case llvm::Triple::avr: 6046 TC = std::make_unique<toolchains::AVRToolChain>(*this, Target, Args); 6047 break; 6048 case llvm::Triple::msp430: 6049 TC = 6050 std::make_unique<toolchains::MSP430ToolChain>(*this, Target, Args); 6051 break; 6052 case llvm::Triple::riscv32: 6053 case llvm::Triple::riscv64: 6054 if (toolchains::RISCVToolChain::hasGCCToolchain(*this, Args)) 6055 TC = 6056 std::make_unique<toolchains::RISCVToolChain>(*this, Target, Args); 6057 else 6058 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6059 break; 6060 case llvm::Triple::ve: 6061 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6062 break; 6063 case llvm::Triple::spirv32: 6064 case llvm::Triple::spirv64: 6065 TC = std::make_unique<toolchains::SPIRVToolChain>(*this, Target, Args); 6066 break; 6067 case llvm::Triple::csky: 6068 TC = std::make_unique<toolchains::CSKYToolChain>(*this, Target, Args); 6069 break; 6070 default: 6071 if (Target.getVendor() == llvm::Triple::Myriad) 6072 TC = std::make_unique<toolchains::MyriadToolChain>(*this, Target, 6073 Args); 6074 else if (toolchains::BareMetal::handlesTarget(Target)) 6075 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6076 else if (Target.isOSBinFormatELF()) 6077 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6078 else if (Target.isOSBinFormatMachO()) 6079 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6080 else 6081 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6082 } 6083 } 6084 } 6085 6086 // Intentionally omitted from the switch above: llvm::Triple::CUDA. CUDA 6087 // compiles always need two toolchains, the CUDA toolchain and the host 6088 // toolchain. So the only valid way to create a CUDA toolchain is via 6089 // CreateOffloadingDeviceToolChains. 6090 6091 return *TC; 6092 } 6093 6094 const ToolChain &Driver::getOffloadingDeviceToolChain( 6095 const ArgList &Args, const llvm::Triple &Target, const ToolChain &HostTC, 6096 const Action::OffloadKind &TargetDeviceOffloadKind) const { 6097 // Use device / host triples as the key into the ToolChains map because the 6098 // device ToolChain we create depends on both. 6099 auto &TC = ToolChains[Target.str() + "/" + HostTC.getTriple().str()]; 6100 if (!TC) { 6101 // Categorized by offload kind > arch rather than OS > arch like 6102 // the normal getToolChain call, as it seems a reasonable way to categorize 6103 // things. 6104 switch (TargetDeviceOffloadKind) { 6105 case Action::OFK_HIP: { 6106 if (Target.getArch() == llvm::Triple::amdgcn && 6107 Target.getVendor() == llvm::Triple::AMD && 6108 Target.getOS() == llvm::Triple::AMDHSA) 6109 TC = std::make_unique<toolchains::HIPAMDToolChain>(*this, Target, 6110 HostTC, Args); 6111 else if (Target.getArch() == llvm::Triple::spirv64 && 6112 Target.getVendor() == llvm::Triple::UnknownVendor && 6113 Target.getOS() == llvm::Triple::UnknownOS) 6114 TC = std::make_unique<toolchains::HIPSPVToolChain>(*this, Target, 6115 HostTC, Args); 6116 break; 6117 } 6118 default: 6119 break; 6120 } 6121 } 6122 6123 return *TC; 6124 } 6125 6126 bool Driver::ShouldUseClangCompiler(const JobAction &JA) const { 6127 // Say "no" if there is not exactly one input of a type clang understands. 6128 if (JA.size() != 1 || 6129 !types::isAcceptedByClang((*JA.input_begin())->getType())) 6130 return false; 6131 6132 // And say "no" if this is not a kind of action clang understands. 6133 if (!isa<PreprocessJobAction>(JA) && !isa<PrecompileJobAction>(JA) && 6134 !isa<CompileJobAction>(JA) && !isa<BackendJobAction>(JA) && 6135 !isa<ExtractAPIJobAction>(JA)) 6136 return false; 6137 6138 return true; 6139 } 6140 6141 bool Driver::ShouldUseFlangCompiler(const JobAction &JA) const { 6142 // Say "no" if there is not exactly one input of a type flang understands. 6143 if (JA.size() != 1 || 6144 !types::isAcceptedByFlang((*JA.input_begin())->getType())) 6145 return false; 6146 6147 // And say "no" if this is not a kind of action flang understands. 6148 if (!isa<PreprocessJobAction>(JA) && !isa<CompileJobAction>(JA) && 6149 !isa<BackendJobAction>(JA)) 6150 return false; 6151 6152 return true; 6153 } 6154 6155 bool Driver::ShouldEmitStaticLibrary(const ArgList &Args) const { 6156 // Only emit static library if the flag is set explicitly. 6157 if (Args.hasArg(options::OPT_emit_static_lib)) 6158 return true; 6159 return false; 6160 } 6161 6162 /// GetReleaseVersion - Parse (([0-9]+)(.([0-9]+)(.([0-9]+)?))?)? and return the 6163 /// grouped values as integers. Numbers which are not provided are set to 0. 6164 /// 6165 /// \return True if the entire string was parsed (9.2), or all groups were 6166 /// parsed (10.3.5extrastuff). 6167 bool Driver::GetReleaseVersion(StringRef Str, unsigned &Major, unsigned &Minor, 6168 unsigned &Micro, bool &HadExtra) { 6169 HadExtra = false; 6170 6171 Major = Minor = Micro = 0; 6172 if (Str.empty()) 6173 return false; 6174 6175 if (Str.consumeInteger(10, Major)) 6176 return false; 6177 if (Str.empty()) 6178 return true; 6179 if (Str[0] != '.') 6180 return false; 6181 6182 Str = Str.drop_front(1); 6183 6184 if (Str.consumeInteger(10, Minor)) 6185 return false; 6186 if (Str.empty()) 6187 return true; 6188 if (Str[0] != '.') 6189 return false; 6190 Str = Str.drop_front(1); 6191 6192 if (Str.consumeInteger(10, Micro)) 6193 return false; 6194 if (!Str.empty()) 6195 HadExtra = true; 6196 return true; 6197 } 6198 6199 /// Parse digits from a string \p Str and fulfill \p Digits with 6200 /// the parsed numbers. This method assumes that the max number of 6201 /// digits to look for is equal to Digits.size(). 6202 /// 6203 /// \return True if the entire string was parsed and there are 6204 /// no extra characters remaining at the end. 6205 bool Driver::GetReleaseVersion(StringRef Str, 6206 MutableArrayRef<unsigned> Digits) { 6207 if (Str.empty()) 6208 return false; 6209 6210 unsigned CurDigit = 0; 6211 while (CurDigit < Digits.size()) { 6212 unsigned Digit; 6213 if (Str.consumeInteger(10, Digit)) 6214 return false; 6215 Digits[CurDigit] = Digit; 6216 if (Str.empty()) 6217 return true; 6218 if (Str[0] != '.') 6219 return false; 6220 Str = Str.drop_front(1); 6221 CurDigit++; 6222 } 6223 6224 // More digits than requested, bail out... 6225 return false; 6226 } 6227 6228 std::pair<unsigned, unsigned> 6229 Driver::getIncludeExcludeOptionFlagMasks(bool IsClCompatMode) const { 6230 unsigned IncludedFlagsBitmask = 0; 6231 unsigned ExcludedFlagsBitmask = options::NoDriverOption; 6232 6233 if (IsClCompatMode) { 6234 // Include CL and Core options. 6235 IncludedFlagsBitmask |= options::CLOption; 6236 IncludedFlagsBitmask |= options::CoreOption; 6237 } else { 6238 ExcludedFlagsBitmask |= options::CLOption; 6239 } 6240 if (IsDXCMode()) { 6241 // Include DXC and Core options. 6242 IncludedFlagsBitmask |= options::DXCOption; 6243 IncludedFlagsBitmask |= options::CoreOption; 6244 } else { 6245 ExcludedFlagsBitmask |= options::DXCOption; 6246 } 6247 return std::make_pair(IncludedFlagsBitmask, ExcludedFlagsBitmask); 6248 } 6249 6250 bool clang::driver::isOptimizationLevelFast(const ArgList &Args) { 6251 return Args.hasFlag(options::OPT_Ofast, options::OPT_O_Group, false); 6252 } 6253 6254 bool clang::driver::willEmitRemarks(const ArgList &Args) { 6255 // -fsave-optimization-record enables it. 6256 if (Args.hasFlag(options::OPT_fsave_optimization_record, 6257 options::OPT_fno_save_optimization_record, false)) 6258 return true; 6259 6260 // -fsave-optimization-record=<format> enables it as well. 6261 if (Args.hasFlag(options::OPT_fsave_optimization_record_EQ, 6262 options::OPT_fno_save_optimization_record, false)) 6263 return true; 6264 6265 // -foptimization-record-file alone enables it too. 6266 if (Args.hasFlag(options::OPT_foptimization_record_file_EQ, 6267 options::OPT_fno_save_optimization_record, false)) 6268 return true; 6269 6270 // -foptimization-record-passes alone enables it too. 6271 if (Args.hasFlag(options::OPT_foptimization_record_passes_EQ, 6272 options::OPT_fno_save_optimization_record, false)) 6273 return true; 6274 return false; 6275 } 6276 6277 llvm::StringRef clang::driver::getDriverMode(StringRef ProgName, 6278 ArrayRef<const char *> Args) { 6279 static const std::string OptName = 6280 getDriverOptTable().getOption(options::OPT_driver_mode).getPrefixedName(); 6281 llvm::StringRef Opt; 6282 for (StringRef Arg : Args) { 6283 if (!Arg.startswith(OptName)) 6284 continue; 6285 Opt = Arg; 6286 } 6287 if (Opt.empty()) 6288 Opt = ToolChain::getTargetAndModeFromProgramName(ProgName).DriverMode; 6289 return Opt.consume_front(OptName) ? Opt : ""; 6290 } 6291 6292 bool driver::IsClangCL(StringRef DriverMode) { return DriverMode.equals("cl"); } 6293