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; 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 // Remove existing compilation database so that each job can append to it. 1318 if (Arg *A = Args.getLastArg(options::OPT_MJ)) 1319 llvm::sys::fs::remove(A->getValue()); 1320 1321 // Setting up the jobs for some precompile cases depends on whether we are 1322 // treating them as PCH, implicit modules or C++20 ones. 1323 // TODO: inferring the mode like this seems fragile (it meets the objective 1324 // of not requiring anything new for operation, however). 1325 const Arg *Std = Args.getLastArg(options::OPT_std_EQ); 1326 ModulesModeCXX20 = 1327 !Args.hasArg(options::OPT_fmodules) && Std && 1328 (Std->containsValue("c++20") || Std->containsValue("c++2b") || 1329 Std->containsValue("c++2a") || Std->containsValue("c++latest")); 1330 1331 // Process -fmodule-header{=} flags. 1332 if (Arg *A = Args.getLastArg(options::OPT_fmodule_header_EQ, 1333 options::OPT_fmodule_header)) { 1334 // These flags force C++20 handling of headers. 1335 ModulesModeCXX20 = true; 1336 if (A->getOption().matches(options::OPT_fmodule_header)) 1337 CXX20HeaderType = HeaderMode_Default; 1338 else { 1339 StringRef ArgName = A->getValue(); 1340 unsigned Kind = llvm::StringSwitch<unsigned>(ArgName) 1341 .Case("user", HeaderMode_User) 1342 .Case("system", HeaderMode_System) 1343 .Default(~0U); 1344 if (Kind == ~0U) { 1345 Diags.Report(diag::err_drv_invalid_value) 1346 << A->getAsString(Args) << ArgName; 1347 } else 1348 CXX20HeaderType = static_cast<ModuleHeaderMode>(Kind); 1349 } 1350 } 1351 1352 std::unique_ptr<llvm::opt::InputArgList> UArgs = 1353 std::make_unique<InputArgList>(std::move(Args)); 1354 1355 // Perform the default argument translations. 1356 DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs); 1357 1358 // Owned by the host. 1359 const ToolChain &TC = getToolChain( 1360 *UArgs, computeTargetTriple(*this, TargetTriple, *UArgs)); 1361 1362 // The compilation takes ownership of Args. 1363 Compilation *C = new Compilation(*this, TC, UArgs.release(), TranslatedArgs, 1364 ContainsError); 1365 1366 if (!HandleImmediateArgs(*C)) 1367 return C; 1368 1369 // Construct the list of inputs. 1370 InputList Inputs; 1371 BuildInputs(C->getDefaultToolChain(), *TranslatedArgs, Inputs); 1372 1373 // Populate the tool chains for the offloading devices, if any. 1374 CreateOffloadingDeviceToolChains(*C, Inputs); 1375 1376 // Construct the list of abstract actions to perform for this compilation. On 1377 // MachO targets this uses the driver-driver and universal actions. 1378 if (TC.getTriple().isOSBinFormatMachO()) 1379 BuildUniversalActions(*C, C->getDefaultToolChain(), Inputs); 1380 else 1381 BuildActions(*C, C->getArgs(), Inputs, C->getActions()); 1382 1383 if (CCCPrintPhases) { 1384 PrintActions(*C); 1385 return C; 1386 } 1387 1388 BuildJobs(*C); 1389 1390 return C; 1391 } 1392 1393 static void printArgList(raw_ostream &OS, const llvm::opt::ArgList &Args) { 1394 llvm::opt::ArgStringList ASL; 1395 for (const auto *A : Args) { 1396 // Use user's original spelling of flags. For example, use 1397 // `/source-charset:utf-8` instead of `-finput-charset=utf-8` if the user 1398 // wrote the former. 1399 while (A->getAlias()) 1400 A = A->getAlias(); 1401 A->render(Args, ASL); 1402 } 1403 1404 for (auto I = ASL.begin(), E = ASL.end(); I != E; ++I) { 1405 if (I != ASL.begin()) 1406 OS << ' '; 1407 llvm::sys::printArg(OS, *I, true); 1408 } 1409 OS << '\n'; 1410 } 1411 1412 bool Driver::getCrashDiagnosticFile(StringRef ReproCrashFilename, 1413 SmallString<128> &CrashDiagDir) { 1414 using namespace llvm::sys; 1415 assert(llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin() && 1416 "Only knows about .crash files on Darwin"); 1417 1418 // The .crash file can be found on at ~/Library/Logs/DiagnosticReports/ 1419 // (or /Library/Logs/DiagnosticReports for root) and has the filename pattern 1420 // clang-<VERSION>_<YYYY-MM-DD-HHMMSS>_<hostname>.crash. 1421 path::home_directory(CrashDiagDir); 1422 if (CrashDiagDir.startswith("/var/root")) 1423 CrashDiagDir = "/"; 1424 path::append(CrashDiagDir, "Library/Logs/DiagnosticReports"); 1425 int PID = 1426 #if LLVM_ON_UNIX 1427 getpid(); 1428 #else 1429 0; 1430 #endif 1431 std::error_code EC; 1432 fs::file_status FileStatus; 1433 TimePoint<> LastAccessTime; 1434 SmallString<128> CrashFilePath; 1435 // Lookup the .crash files and get the one generated by a subprocess spawned 1436 // by this driver invocation. 1437 for (fs::directory_iterator File(CrashDiagDir, EC), FileEnd; 1438 File != FileEnd && !EC; File.increment(EC)) { 1439 StringRef FileName = path::filename(File->path()); 1440 if (!FileName.startswith(Name)) 1441 continue; 1442 if (fs::status(File->path(), FileStatus)) 1443 continue; 1444 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CrashFile = 1445 llvm::MemoryBuffer::getFile(File->path()); 1446 if (!CrashFile) 1447 continue; 1448 // The first line should start with "Process:", otherwise this isn't a real 1449 // .crash file. 1450 StringRef Data = CrashFile.get()->getBuffer(); 1451 if (!Data.startswith("Process:")) 1452 continue; 1453 // Parse parent process pid line, e.g: "Parent Process: clang-4.0 [79141]" 1454 size_t ParentProcPos = Data.find("Parent Process:"); 1455 if (ParentProcPos == StringRef::npos) 1456 continue; 1457 size_t LineEnd = Data.find_first_of("\n", ParentProcPos); 1458 if (LineEnd == StringRef::npos) 1459 continue; 1460 StringRef ParentProcess = Data.slice(ParentProcPos+15, LineEnd).trim(); 1461 int OpenBracket = -1, CloseBracket = -1; 1462 for (size_t i = 0, e = ParentProcess.size(); i < e; ++i) { 1463 if (ParentProcess[i] == '[') 1464 OpenBracket = i; 1465 if (ParentProcess[i] == ']') 1466 CloseBracket = i; 1467 } 1468 // Extract the parent process PID from the .crash file and check whether 1469 // it matches this driver invocation pid. 1470 int CrashPID; 1471 if (OpenBracket < 0 || CloseBracket < 0 || 1472 ParentProcess.slice(OpenBracket + 1, CloseBracket) 1473 .getAsInteger(10, CrashPID) || CrashPID != PID) { 1474 continue; 1475 } 1476 1477 // Found a .crash file matching the driver pid. To avoid getting an older 1478 // and misleading crash file, continue looking for the most recent. 1479 // FIXME: the driver can dispatch multiple cc1 invocations, leading to 1480 // multiple crashes poiting to the same parent process. Since the driver 1481 // does not collect pid information for the dispatched invocation there's 1482 // currently no way to distinguish among them. 1483 const auto FileAccessTime = FileStatus.getLastModificationTime(); 1484 if (FileAccessTime > LastAccessTime) { 1485 CrashFilePath.assign(File->path()); 1486 LastAccessTime = FileAccessTime; 1487 } 1488 } 1489 1490 // If found, copy it over to the location of other reproducer files. 1491 if (!CrashFilePath.empty()) { 1492 EC = fs::copy_file(CrashFilePath, ReproCrashFilename); 1493 if (EC) 1494 return false; 1495 return true; 1496 } 1497 1498 return false; 1499 } 1500 1501 // When clang crashes, produce diagnostic information including the fully 1502 // preprocessed source file(s). Request that the developer attach the 1503 // diagnostic information to a bug report. 1504 void Driver::generateCompilationDiagnostics( 1505 Compilation &C, const Command &FailingCommand, 1506 StringRef AdditionalInformation, CompilationDiagnosticReport *Report) { 1507 if (C.getArgs().hasArg(options::OPT_fno_crash_diagnostics)) 1508 return; 1509 1510 // Don't try to generate diagnostics for link or dsymutil jobs. 1511 if (FailingCommand.getCreator().isLinkJob() || 1512 FailingCommand.getCreator().isDsymutilJob()) 1513 return; 1514 1515 // Print the version of the compiler. 1516 PrintVersion(C, llvm::errs()); 1517 1518 // Suppress driver output and emit preprocessor output to temp file. 1519 CCGenDiagnostics = true; 1520 1521 // Save the original job command(s). 1522 Command Cmd = FailingCommand; 1523 1524 // Keep track of whether we produce any errors while trying to produce 1525 // preprocessed sources. 1526 DiagnosticErrorTrap Trap(Diags); 1527 1528 // Suppress tool output. 1529 C.initCompilationForDiagnostics(); 1530 1531 // Construct the list of inputs. 1532 InputList Inputs; 1533 BuildInputs(C.getDefaultToolChain(), C.getArgs(), Inputs); 1534 1535 for (InputList::iterator it = Inputs.begin(), ie = Inputs.end(); it != ie;) { 1536 bool IgnoreInput = false; 1537 1538 // Ignore input from stdin or any inputs that cannot be preprocessed. 1539 // Check type first as not all linker inputs have a value. 1540 if (types::getPreprocessedType(it->first) == types::TY_INVALID) { 1541 IgnoreInput = true; 1542 } else if (!strcmp(it->second->getValue(), "-")) { 1543 Diag(clang::diag::note_drv_command_failed_diag_msg) 1544 << "Error generating preprocessed source(s) - " 1545 "ignoring input from stdin."; 1546 IgnoreInput = true; 1547 } 1548 1549 if (IgnoreInput) { 1550 it = Inputs.erase(it); 1551 ie = Inputs.end(); 1552 } else { 1553 ++it; 1554 } 1555 } 1556 1557 if (Inputs.empty()) { 1558 Diag(clang::diag::note_drv_command_failed_diag_msg) 1559 << "Error generating preprocessed source(s) - " 1560 "no preprocessable inputs."; 1561 return; 1562 } 1563 1564 // Don't attempt to generate preprocessed files if multiple -arch options are 1565 // used, unless they're all duplicates. 1566 llvm::StringSet<> ArchNames; 1567 for (const Arg *A : C.getArgs()) { 1568 if (A->getOption().matches(options::OPT_arch)) { 1569 StringRef ArchName = A->getValue(); 1570 ArchNames.insert(ArchName); 1571 } 1572 } 1573 if (ArchNames.size() > 1) { 1574 Diag(clang::diag::note_drv_command_failed_diag_msg) 1575 << "Error generating preprocessed source(s) - cannot generate " 1576 "preprocessed source with multiple -arch options."; 1577 return; 1578 } 1579 1580 // Construct the list of abstract actions to perform for this compilation. On 1581 // Darwin OSes this uses the driver-driver and builds universal actions. 1582 const ToolChain &TC = C.getDefaultToolChain(); 1583 if (TC.getTriple().isOSBinFormatMachO()) 1584 BuildUniversalActions(C, TC, Inputs); 1585 else 1586 BuildActions(C, C.getArgs(), Inputs, C.getActions()); 1587 1588 BuildJobs(C); 1589 1590 // If there were errors building the compilation, quit now. 1591 if (Trap.hasErrorOccurred()) { 1592 Diag(clang::diag::note_drv_command_failed_diag_msg) 1593 << "Error generating preprocessed source(s)."; 1594 return; 1595 } 1596 1597 // Generate preprocessed output. 1598 SmallVector<std::pair<int, const Command *>, 4> FailingCommands; 1599 C.ExecuteJobs(C.getJobs(), FailingCommands); 1600 1601 // If any of the preprocessing commands failed, clean up and exit. 1602 if (!FailingCommands.empty()) { 1603 Diag(clang::diag::note_drv_command_failed_diag_msg) 1604 << "Error generating preprocessed source(s)."; 1605 return; 1606 } 1607 1608 const ArgStringList &TempFiles = C.getTempFiles(); 1609 if (TempFiles.empty()) { 1610 Diag(clang::diag::note_drv_command_failed_diag_msg) 1611 << "Error generating preprocessed source(s)."; 1612 return; 1613 } 1614 1615 Diag(clang::diag::note_drv_command_failed_diag_msg) 1616 << "\n********************\n\n" 1617 "PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:\n" 1618 "Preprocessed source(s) and associated run script(s) are located at:"; 1619 1620 SmallString<128> VFS; 1621 SmallString<128> ReproCrashFilename; 1622 for (const char *TempFile : TempFiles) { 1623 Diag(clang::diag::note_drv_command_failed_diag_msg) << TempFile; 1624 if (Report) 1625 Report->TemporaryFiles.push_back(TempFile); 1626 if (ReproCrashFilename.empty()) { 1627 ReproCrashFilename = TempFile; 1628 llvm::sys::path::replace_extension(ReproCrashFilename, ".crash"); 1629 } 1630 if (StringRef(TempFile).endswith(".cache")) { 1631 // In some cases (modules) we'll dump extra data to help with reproducing 1632 // the crash into a directory next to the output. 1633 VFS = llvm::sys::path::filename(TempFile); 1634 llvm::sys::path::append(VFS, "vfs", "vfs.yaml"); 1635 } 1636 } 1637 1638 // Assume associated files are based off of the first temporary file. 1639 CrashReportInfo CrashInfo(TempFiles[0], VFS); 1640 1641 llvm::SmallString<128> Script(CrashInfo.Filename); 1642 llvm::sys::path::replace_extension(Script, "sh"); 1643 std::error_code EC; 1644 llvm::raw_fd_ostream ScriptOS(Script, EC, llvm::sys::fs::CD_CreateNew, 1645 llvm::sys::fs::FA_Write, 1646 llvm::sys::fs::OF_Text); 1647 if (EC) { 1648 Diag(clang::diag::note_drv_command_failed_diag_msg) 1649 << "Error generating run script: " << Script << " " << EC.message(); 1650 } else { 1651 ScriptOS << "# Crash reproducer for " << getClangFullVersion() << "\n" 1652 << "# Driver args: "; 1653 printArgList(ScriptOS, C.getInputArgs()); 1654 ScriptOS << "# Original command: "; 1655 Cmd.Print(ScriptOS, "\n", /*Quote=*/true); 1656 Cmd.Print(ScriptOS, "\n", /*Quote=*/true, &CrashInfo); 1657 if (!AdditionalInformation.empty()) 1658 ScriptOS << "\n# Additional information: " << AdditionalInformation 1659 << "\n"; 1660 if (Report) 1661 Report->TemporaryFiles.push_back(std::string(Script.str())); 1662 Diag(clang::diag::note_drv_command_failed_diag_msg) << Script; 1663 } 1664 1665 // On darwin, provide information about the .crash diagnostic report. 1666 if (llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin()) { 1667 SmallString<128> CrashDiagDir; 1668 if (getCrashDiagnosticFile(ReproCrashFilename, CrashDiagDir)) { 1669 Diag(clang::diag::note_drv_command_failed_diag_msg) 1670 << ReproCrashFilename.str(); 1671 } else { // Suggest a directory for the user to look for .crash files. 1672 llvm::sys::path::append(CrashDiagDir, Name); 1673 CrashDiagDir += "_<YYYY-MM-DD-HHMMSS>_<hostname>.crash"; 1674 Diag(clang::diag::note_drv_command_failed_diag_msg) 1675 << "Crash backtrace is located in"; 1676 Diag(clang::diag::note_drv_command_failed_diag_msg) 1677 << CrashDiagDir.str(); 1678 Diag(clang::diag::note_drv_command_failed_diag_msg) 1679 << "(choose the .crash file that corresponds to your crash)"; 1680 } 1681 } 1682 1683 for (const auto &A : C.getArgs().filtered(options::OPT_frewrite_map_file_EQ)) 1684 Diag(clang::diag::note_drv_command_failed_diag_msg) << A->getValue(); 1685 1686 Diag(clang::diag::note_drv_command_failed_diag_msg) 1687 << "\n\n********************"; 1688 } 1689 1690 void Driver::setUpResponseFiles(Compilation &C, Command &Cmd) { 1691 // Since commandLineFitsWithinSystemLimits() may underestimate system's 1692 // capacity if the tool does not support response files, there is a chance/ 1693 // that things will just work without a response file, so we silently just 1694 // skip it. 1695 if (Cmd.getResponseFileSupport().ResponseKind == 1696 ResponseFileSupport::RF_None || 1697 llvm::sys::commandLineFitsWithinSystemLimits(Cmd.getExecutable(), 1698 Cmd.getArguments())) 1699 return; 1700 1701 std::string TmpName = GetTemporaryPath("response", "txt"); 1702 Cmd.setResponseFile(C.addTempFile(C.getArgs().MakeArgString(TmpName))); 1703 } 1704 1705 int Driver::ExecuteCompilation( 1706 Compilation &C, 1707 SmallVectorImpl<std::pair<int, const Command *>> &FailingCommands) { 1708 if (C.getArgs().hasArg(options::OPT_fdriver_only)) { 1709 if (C.getArgs().hasArg(options::OPT_v)) 1710 C.getJobs().Print(llvm::errs(), "\n", true); 1711 1712 C.ExecuteJobs(C.getJobs(), FailingCommands, /*LogOnly=*/true); 1713 1714 // If there were errors building the compilation, quit now. 1715 if (!FailingCommands.empty() || Diags.hasErrorOccurred()) 1716 return 1; 1717 1718 return 0; 1719 } 1720 1721 // Just print if -### was present. 1722 if (C.getArgs().hasArg(options::OPT__HASH_HASH_HASH)) { 1723 C.getJobs().Print(llvm::errs(), "\n", true); 1724 return 0; 1725 } 1726 1727 // If there were errors building the compilation, quit now. 1728 if (Diags.hasErrorOccurred()) 1729 return 1; 1730 1731 // Set up response file names for each command, if necessary. 1732 for (auto &Job : C.getJobs()) 1733 setUpResponseFiles(C, Job); 1734 1735 C.ExecuteJobs(C.getJobs(), FailingCommands); 1736 1737 // If the command succeeded, we are done. 1738 if (FailingCommands.empty()) 1739 return 0; 1740 1741 // Otherwise, remove result files and print extra information about abnormal 1742 // failures. 1743 int Res = 0; 1744 for (const auto &CmdPair : FailingCommands) { 1745 int CommandRes = CmdPair.first; 1746 const Command *FailingCommand = CmdPair.second; 1747 1748 // Remove result files if we're not saving temps. 1749 if (!isSaveTempsEnabled()) { 1750 const JobAction *JA = cast<JobAction>(&FailingCommand->getSource()); 1751 C.CleanupFileMap(C.getResultFiles(), JA, true); 1752 1753 // Failure result files are valid unless we crashed. 1754 if (CommandRes < 0) 1755 C.CleanupFileMap(C.getFailureResultFiles(), JA, true); 1756 } 1757 1758 #if LLVM_ON_UNIX 1759 // llvm/lib/Support/Unix/Signals.inc will exit with a special return code 1760 // for SIGPIPE. Do not print diagnostics for this case. 1761 if (CommandRes == EX_IOERR) { 1762 Res = CommandRes; 1763 continue; 1764 } 1765 #endif 1766 1767 // Print extra information about abnormal failures, if possible. 1768 // 1769 // This is ad-hoc, but we don't want to be excessively noisy. If the result 1770 // status was 1, assume the command failed normally. In particular, if it 1771 // was the compiler then assume it gave a reasonable error code. Failures 1772 // in other tools are less common, and they generally have worse 1773 // diagnostics, so always print the diagnostic there. 1774 const Tool &FailingTool = FailingCommand->getCreator(); 1775 1776 if (!FailingCommand->getCreator().hasGoodDiagnostics() || CommandRes != 1) { 1777 // FIXME: See FIXME above regarding result code interpretation. 1778 if (CommandRes < 0) 1779 Diag(clang::diag::err_drv_command_signalled) 1780 << FailingTool.getShortName(); 1781 else 1782 Diag(clang::diag::err_drv_command_failed) 1783 << FailingTool.getShortName() << CommandRes; 1784 } 1785 } 1786 return Res; 1787 } 1788 1789 void Driver::PrintHelp(bool ShowHidden) const { 1790 unsigned IncludedFlagsBitmask; 1791 unsigned ExcludedFlagsBitmask; 1792 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 1793 getIncludeExcludeOptionFlagMasks(IsCLMode()); 1794 1795 ExcludedFlagsBitmask |= options::NoDriverOption; 1796 if (!ShowHidden) 1797 ExcludedFlagsBitmask |= HelpHidden; 1798 1799 if (IsFlangMode()) 1800 IncludedFlagsBitmask |= options::FlangOption; 1801 else 1802 ExcludedFlagsBitmask |= options::FlangOnlyOption; 1803 1804 std::string Usage = llvm::formatv("{0} [options] file...", Name).str(); 1805 getOpts().printHelp(llvm::outs(), Usage.c_str(), DriverTitle.c_str(), 1806 IncludedFlagsBitmask, ExcludedFlagsBitmask, 1807 /*ShowAllAliases=*/false); 1808 } 1809 1810 void Driver::PrintVersion(const Compilation &C, raw_ostream &OS) const { 1811 if (IsFlangMode()) { 1812 OS << getClangToolFullVersion("flang-new") << '\n'; 1813 } else { 1814 // FIXME: The following handlers should use a callback mechanism, we don't 1815 // know what the client would like to do. 1816 OS << getClangFullVersion() << '\n'; 1817 } 1818 const ToolChain &TC = C.getDefaultToolChain(); 1819 OS << "Target: " << TC.getTripleString() << '\n'; 1820 1821 // Print the threading model. 1822 if (Arg *A = C.getArgs().getLastArg(options::OPT_mthread_model)) { 1823 // Don't print if the ToolChain would have barfed on it already 1824 if (TC.isThreadModelSupported(A->getValue())) 1825 OS << "Thread model: " << A->getValue(); 1826 } else 1827 OS << "Thread model: " << TC.getThreadModel(); 1828 OS << '\n'; 1829 1830 // Print out the install directory. 1831 OS << "InstalledDir: " << InstalledDir << '\n'; 1832 1833 // If configuration file was used, print its path. 1834 if (!ConfigFile.empty()) 1835 OS << "Configuration file: " << ConfigFile << '\n'; 1836 } 1837 1838 /// PrintDiagnosticCategories - Implement the --print-diagnostic-categories 1839 /// option. 1840 static void PrintDiagnosticCategories(raw_ostream &OS) { 1841 // Skip the empty category. 1842 for (unsigned i = 1, max = DiagnosticIDs::getNumberOfCategories(); i != max; 1843 ++i) 1844 OS << i << ',' << DiagnosticIDs::getCategoryNameFromID(i) << '\n'; 1845 } 1846 1847 void Driver::HandleAutocompletions(StringRef PassedFlags) const { 1848 if (PassedFlags == "") 1849 return; 1850 // Print out all options that start with a given argument. This is used for 1851 // shell autocompletion. 1852 std::vector<std::string> SuggestedCompletions; 1853 std::vector<std::string> Flags; 1854 1855 unsigned int DisableFlags = 1856 options::NoDriverOption | options::Unsupported | options::Ignored; 1857 1858 // Make sure that Flang-only options don't pollute the Clang output 1859 // TODO: Make sure that Clang-only options don't pollute Flang output 1860 if (!IsFlangMode()) 1861 DisableFlags |= options::FlangOnlyOption; 1862 1863 // Distinguish "--autocomplete=-someflag" and "--autocomplete=-someflag," 1864 // because the latter indicates that the user put space before pushing tab 1865 // which should end up in a file completion. 1866 const bool HasSpace = PassedFlags.endswith(","); 1867 1868 // Parse PassedFlags by "," as all the command-line flags are passed to this 1869 // function separated by "," 1870 StringRef TargetFlags = PassedFlags; 1871 while (TargetFlags != "") { 1872 StringRef CurFlag; 1873 std::tie(CurFlag, TargetFlags) = TargetFlags.split(","); 1874 Flags.push_back(std::string(CurFlag)); 1875 } 1876 1877 // We want to show cc1-only options only when clang is invoked with -cc1 or 1878 // -Xclang. 1879 if (llvm::is_contained(Flags, "-Xclang") || llvm::is_contained(Flags, "-cc1")) 1880 DisableFlags &= ~options::NoDriverOption; 1881 1882 const llvm::opt::OptTable &Opts = getOpts(); 1883 StringRef Cur; 1884 Cur = Flags.at(Flags.size() - 1); 1885 StringRef Prev; 1886 if (Flags.size() >= 2) { 1887 Prev = Flags.at(Flags.size() - 2); 1888 SuggestedCompletions = Opts.suggestValueCompletions(Prev, Cur); 1889 } 1890 1891 if (SuggestedCompletions.empty()) 1892 SuggestedCompletions = Opts.suggestValueCompletions(Cur, ""); 1893 1894 // If Flags were empty, it means the user typed `clang [tab]` where we should 1895 // list all possible flags. If there was no value completion and the user 1896 // pressed tab after a space, we should fall back to a file completion. 1897 // We're printing a newline to be consistent with what we print at the end of 1898 // this function. 1899 if (SuggestedCompletions.empty() && HasSpace && !Flags.empty()) { 1900 llvm::outs() << '\n'; 1901 return; 1902 } 1903 1904 // When flag ends with '=' and there was no value completion, return empty 1905 // string and fall back to the file autocompletion. 1906 if (SuggestedCompletions.empty() && !Cur.endswith("=")) { 1907 // If the flag is in the form of "--autocomplete=-foo", 1908 // we were requested to print out all option names that start with "-foo". 1909 // For example, "--autocomplete=-fsyn" is expanded to "-fsyntax-only". 1910 SuggestedCompletions = Opts.findByPrefix(Cur, DisableFlags); 1911 1912 // We have to query the -W flags manually as they're not in the OptTable. 1913 // TODO: Find a good way to add them to OptTable instead and them remove 1914 // this code. 1915 for (StringRef S : DiagnosticIDs::getDiagnosticFlags()) 1916 if (S.startswith(Cur)) 1917 SuggestedCompletions.push_back(std::string(S)); 1918 } 1919 1920 // Sort the autocomplete candidates so that shells print them out in a 1921 // deterministic order. We could sort in any way, but we chose 1922 // case-insensitive sorting for consistency with the -help option 1923 // which prints out options in the case-insensitive alphabetical order. 1924 llvm::sort(SuggestedCompletions, [](StringRef A, StringRef B) { 1925 if (int X = A.compare_insensitive(B)) 1926 return X < 0; 1927 return A.compare(B) > 0; 1928 }); 1929 1930 llvm::outs() << llvm::join(SuggestedCompletions, "\n") << '\n'; 1931 } 1932 1933 bool Driver::HandleImmediateArgs(const Compilation &C) { 1934 // The order these options are handled in gcc is all over the place, but we 1935 // don't expect inconsistencies w.r.t. that to matter in practice. 1936 1937 if (C.getArgs().hasArg(options::OPT_dumpmachine)) { 1938 llvm::outs() << C.getDefaultToolChain().getTripleString() << '\n'; 1939 return false; 1940 } 1941 1942 if (C.getArgs().hasArg(options::OPT_dumpversion)) { 1943 // Since -dumpversion is only implemented for pedantic GCC compatibility, we 1944 // return an answer which matches our definition of __VERSION__. 1945 llvm::outs() << CLANG_VERSION_STRING << "\n"; 1946 return false; 1947 } 1948 1949 if (C.getArgs().hasArg(options::OPT__print_diagnostic_categories)) { 1950 PrintDiagnosticCategories(llvm::outs()); 1951 return false; 1952 } 1953 1954 if (C.getArgs().hasArg(options::OPT_help) || 1955 C.getArgs().hasArg(options::OPT__help_hidden)) { 1956 PrintHelp(C.getArgs().hasArg(options::OPT__help_hidden)); 1957 return false; 1958 } 1959 1960 if (C.getArgs().hasArg(options::OPT__version)) { 1961 // Follow gcc behavior and use stdout for --version and stderr for -v. 1962 PrintVersion(C, llvm::outs()); 1963 return false; 1964 } 1965 1966 if (C.getArgs().hasArg(options::OPT_v) || 1967 C.getArgs().hasArg(options::OPT__HASH_HASH_HASH) || 1968 C.getArgs().hasArg(options::OPT_print_supported_cpus)) { 1969 PrintVersion(C, llvm::errs()); 1970 SuppressMissingInputWarning = true; 1971 } 1972 1973 if (C.getArgs().hasArg(options::OPT_v)) { 1974 if (!SystemConfigDir.empty()) 1975 llvm::errs() << "System configuration file directory: " 1976 << SystemConfigDir << "\n"; 1977 if (!UserConfigDir.empty()) 1978 llvm::errs() << "User configuration file directory: " 1979 << UserConfigDir << "\n"; 1980 } 1981 1982 const ToolChain &TC = C.getDefaultToolChain(); 1983 1984 if (C.getArgs().hasArg(options::OPT_v)) 1985 TC.printVerboseInfo(llvm::errs()); 1986 1987 if (C.getArgs().hasArg(options::OPT_print_resource_dir)) { 1988 llvm::outs() << ResourceDir << '\n'; 1989 return false; 1990 } 1991 1992 if (C.getArgs().hasArg(options::OPT_print_search_dirs)) { 1993 llvm::outs() << "programs: ="; 1994 bool separator = false; 1995 // Print -B and COMPILER_PATH. 1996 for (const std::string &Path : PrefixDirs) { 1997 if (separator) 1998 llvm::outs() << llvm::sys::EnvPathSeparator; 1999 llvm::outs() << Path; 2000 separator = true; 2001 } 2002 for (const std::string &Path : TC.getProgramPaths()) { 2003 if (separator) 2004 llvm::outs() << llvm::sys::EnvPathSeparator; 2005 llvm::outs() << Path; 2006 separator = true; 2007 } 2008 llvm::outs() << "\n"; 2009 llvm::outs() << "libraries: =" << ResourceDir; 2010 2011 StringRef sysroot = C.getSysRoot(); 2012 2013 for (const std::string &Path : TC.getFilePaths()) { 2014 // Always print a separator. ResourceDir was the first item shown. 2015 llvm::outs() << llvm::sys::EnvPathSeparator; 2016 // Interpretation of leading '=' is needed only for NetBSD. 2017 if (Path[0] == '=') 2018 llvm::outs() << sysroot << Path.substr(1); 2019 else 2020 llvm::outs() << Path; 2021 } 2022 llvm::outs() << "\n"; 2023 return false; 2024 } 2025 2026 if (C.getArgs().hasArg(options::OPT_print_runtime_dir)) { 2027 std::string RuntimePath; 2028 // Get the first existing path, if any. 2029 for (auto Path : TC.getRuntimePaths()) { 2030 if (getVFS().exists(Path)) { 2031 RuntimePath = Path; 2032 break; 2033 } 2034 } 2035 if (!RuntimePath.empty()) 2036 llvm::outs() << RuntimePath << '\n'; 2037 else 2038 llvm::outs() << TC.getCompilerRTPath() << '\n'; 2039 return false; 2040 } 2041 2042 if (C.getArgs().hasArg(options::OPT_print_diagnostic_options)) { 2043 std::vector<std::string> Flags = DiagnosticIDs::getDiagnosticFlags(); 2044 for (std::size_t I = 0; I != Flags.size(); I += 2) 2045 llvm::outs() << " " << Flags[I] << "\n " << Flags[I + 1] << "\n\n"; 2046 return false; 2047 } 2048 2049 // FIXME: The following handlers should use a callback mechanism, we don't 2050 // know what the client would like to do. 2051 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_file_name_EQ)) { 2052 llvm::outs() << GetFilePath(A->getValue(), TC) << "\n"; 2053 return false; 2054 } 2055 2056 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_prog_name_EQ)) { 2057 StringRef ProgName = A->getValue(); 2058 2059 // Null program name cannot have a path. 2060 if (! ProgName.empty()) 2061 llvm::outs() << GetProgramPath(ProgName, TC); 2062 2063 llvm::outs() << "\n"; 2064 return false; 2065 } 2066 2067 if (Arg *A = C.getArgs().getLastArg(options::OPT_autocomplete)) { 2068 StringRef PassedFlags = A->getValue(); 2069 HandleAutocompletions(PassedFlags); 2070 return false; 2071 } 2072 2073 if (C.getArgs().hasArg(options::OPT_print_libgcc_file_name)) { 2074 ToolChain::RuntimeLibType RLT = TC.GetRuntimeLibType(C.getArgs()); 2075 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2076 RegisterEffectiveTriple TripleRAII(TC, Triple); 2077 switch (RLT) { 2078 case ToolChain::RLT_CompilerRT: 2079 llvm::outs() << TC.getCompilerRT(C.getArgs(), "builtins") << "\n"; 2080 break; 2081 case ToolChain::RLT_Libgcc: 2082 llvm::outs() << GetFilePath("libgcc.a", TC) << "\n"; 2083 break; 2084 } 2085 return false; 2086 } 2087 2088 if (C.getArgs().hasArg(options::OPT_print_multi_lib)) { 2089 for (const Multilib &Multilib : TC.getMultilibs()) 2090 llvm::outs() << Multilib << "\n"; 2091 return false; 2092 } 2093 2094 if (C.getArgs().hasArg(options::OPT_print_multi_directory)) { 2095 const Multilib &Multilib = TC.getMultilib(); 2096 if (Multilib.gccSuffix().empty()) 2097 llvm::outs() << ".\n"; 2098 else { 2099 StringRef Suffix(Multilib.gccSuffix()); 2100 assert(Suffix.front() == '/'); 2101 llvm::outs() << Suffix.substr(1) << "\n"; 2102 } 2103 return false; 2104 } 2105 2106 if (C.getArgs().hasArg(options::OPT_print_target_triple)) { 2107 llvm::outs() << TC.getTripleString() << "\n"; 2108 return false; 2109 } 2110 2111 if (C.getArgs().hasArg(options::OPT_print_effective_triple)) { 2112 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2113 llvm::outs() << Triple.getTriple() << "\n"; 2114 return false; 2115 } 2116 2117 if (C.getArgs().hasArg(options::OPT_print_multiarch)) { 2118 llvm::outs() << TC.getMultiarchTriple(*this, TC.getTriple(), SysRoot) 2119 << "\n"; 2120 return false; 2121 } 2122 2123 if (C.getArgs().hasArg(options::OPT_print_targets)) { 2124 llvm::TargetRegistry::printRegisteredTargetsForVersion(llvm::outs()); 2125 return false; 2126 } 2127 2128 return true; 2129 } 2130 2131 enum { 2132 TopLevelAction = 0, 2133 HeadSibAction = 1, 2134 OtherSibAction = 2, 2135 }; 2136 2137 // Display an action graph human-readably. Action A is the "sink" node 2138 // and latest-occuring action. Traversal is in pre-order, visiting the 2139 // inputs to each action before printing the action itself. 2140 static unsigned PrintActions1(const Compilation &C, Action *A, 2141 std::map<Action *, unsigned> &Ids, 2142 Twine Indent = {}, int Kind = TopLevelAction) { 2143 if (Ids.count(A)) // A was already visited. 2144 return Ids[A]; 2145 2146 std::string str; 2147 llvm::raw_string_ostream os(str); 2148 2149 auto getSibIndent = [](int K) -> Twine { 2150 return (K == HeadSibAction) ? " " : (K == OtherSibAction) ? "| " : ""; 2151 }; 2152 2153 Twine SibIndent = Indent + getSibIndent(Kind); 2154 int SibKind = HeadSibAction; 2155 os << Action::getClassName(A->getKind()) << ", "; 2156 if (InputAction *IA = dyn_cast<InputAction>(A)) { 2157 os << "\"" << IA->getInputArg().getValue() << "\""; 2158 } else if (BindArchAction *BIA = dyn_cast<BindArchAction>(A)) { 2159 os << '"' << BIA->getArchName() << '"' << ", {" 2160 << PrintActions1(C, *BIA->input_begin(), Ids, SibIndent, SibKind) << "}"; 2161 } else if (OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 2162 bool IsFirst = true; 2163 OA->doOnEachDependence( 2164 [&](Action *A, const ToolChain *TC, const char *BoundArch) { 2165 assert(TC && "Unknown host toolchain"); 2166 // E.g. for two CUDA device dependences whose bound arch is sm_20 and 2167 // sm_35 this will generate: 2168 // "cuda-device" (nvptx64-nvidia-cuda:sm_20) {#ID}, "cuda-device" 2169 // (nvptx64-nvidia-cuda:sm_35) {#ID} 2170 if (!IsFirst) 2171 os << ", "; 2172 os << '"'; 2173 os << A->getOffloadingKindPrefix(); 2174 os << " ("; 2175 os << TC->getTriple().normalize(); 2176 if (BoundArch) 2177 os << ":" << BoundArch; 2178 os << ")"; 2179 os << '"'; 2180 os << " {" << PrintActions1(C, A, Ids, SibIndent, SibKind) << "}"; 2181 IsFirst = false; 2182 SibKind = OtherSibAction; 2183 }); 2184 } else { 2185 const ActionList *AL = &A->getInputs(); 2186 2187 if (AL->size()) { 2188 const char *Prefix = "{"; 2189 for (Action *PreRequisite : *AL) { 2190 os << Prefix << PrintActions1(C, PreRequisite, Ids, SibIndent, SibKind); 2191 Prefix = ", "; 2192 SibKind = OtherSibAction; 2193 } 2194 os << "}"; 2195 } else 2196 os << "{}"; 2197 } 2198 2199 // Append offload info for all options other than the offloading action 2200 // itself (e.g. (cuda-device, sm_20) or (cuda-host)). 2201 std::string offload_str; 2202 llvm::raw_string_ostream offload_os(offload_str); 2203 if (!isa<OffloadAction>(A)) { 2204 auto S = A->getOffloadingKindPrefix(); 2205 if (!S.empty()) { 2206 offload_os << ", (" << S; 2207 if (A->getOffloadingArch()) 2208 offload_os << ", " << A->getOffloadingArch(); 2209 offload_os << ")"; 2210 } 2211 } 2212 2213 auto getSelfIndent = [](int K) -> Twine { 2214 return (K == HeadSibAction) ? "+- " : (K == OtherSibAction) ? "|- " : ""; 2215 }; 2216 2217 unsigned Id = Ids.size(); 2218 Ids[A] = Id; 2219 llvm::errs() << Indent + getSelfIndent(Kind) << Id << ": " << os.str() << ", " 2220 << types::getTypeName(A->getType()) << offload_os.str() << "\n"; 2221 2222 return Id; 2223 } 2224 2225 // Print the action graphs in a compilation C. 2226 // For example "clang -c file1.c file2.c" is composed of two subgraphs. 2227 void Driver::PrintActions(const Compilation &C) const { 2228 std::map<Action *, unsigned> Ids; 2229 for (Action *A : C.getActions()) 2230 PrintActions1(C, A, Ids); 2231 } 2232 2233 /// Check whether the given input tree contains any compilation or 2234 /// assembly actions. 2235 static bool ContainsCompileOrAssembleAction(const Action *A) { 2236 if (isa<CompileJobAction>(A) || isa<BackendJobAction>(A) || 2237 isa<AssembleJobAction>(A)) 2238 return true; 2239 2240 return llvm::any_of(A->inputs(), ContainsCompileOrAssembleAction); 2241 } 2242 2243 void Driver::BuildUniversalActions(Compilation &C, const ToolChain &TC, 2244 const InputList &BAInputs) const { 2245 DerivedArgList &Args = C.getArgs(); 2246 ActionList &Actions = C.getActions(); 2247 llvm::PrettyStackTraceString CrashInfo("Building universal build actions"); 2248 // Collect the list of architectures. Duplicates are allowed, but should only 2249 // be handled once (in the order seen). 2250 llvm::StringSet<> ArchNames; 2251 SmallVector<const char *, 4> Archs; 2252 for (Arg *A : Args) { 2253 if (A->getOption().matches(options::OPT_arch)) { 2254 // Validate the option here; we don't save the type here because its 2255 // particular spelling may participate in other driver choices. 2256 llvm::Triple::ArchType Arch = 2257 tools::darwin::getArchTypeForMachOArchName(A->getValue()); 2258 if (Arch == llvm::Triple::UnknownArch) { 2259 Diag(clang::diag::err_drv_invalid_arch_name) << A->getAsString(Args); 2260 continue; 2261 } 2262 2263 A->claim(); 2264 if (ArchNames.insert(A->getValue()).second) 2265 Archs.push_back(A->getValue()); 2266 } 2267 } 2268 2269 // When there is no explicit arch for this platform, make sure we still bind 2270 // the architecture (to the default) so that -Xarch_ is handled correctly. 2271 if (!Archs.size()) 2272 Archs.push_back(Args.MakeArgString(TC.getDefaultUniversalArchName())); 2273 2274 ActionList SingleActions; 2275 BuildActions(C, Args, BAInputs, SingleActions); 2276 2277 // Add in arch bindings for every top level action, as well as lipo and 2278 // dsymutil steps if needed. 2279 for (Action* Act : SingleActions) { 2280 // Make sure we can lipo this kind of output. If not (and it is an actual 2281 // output) then we disallow, since we can't create an output file with the 2282 // right name without overwriting it. We could remove this oddity by just 2283 // changing the output names to include the arch, which would also fix 2284 // -save-temps. Compatibility wins for now. 2285 2286 if (Archs.size() > 1 && !types::canLipoType(Act->getType())) 2287 Diag(clang::diag::err_drv_invalid_output_with_multiple_archs) 2288 << types::getTypeName(Act->getType()); 2289 2290 ActionList Inputs; 2291 for (unsigned i = 0, e = Archs.size(); i != e; ++i) 2292 Inputs.push_back(C.MakeAction<BindArchAction>(Act, Archs[i])); 2293 2294 // Lipo if necessary, we do it this way because we need to set the arch flag 2295 // so that -Xarch_ gets overwritten. 2296 if (Inputs.size() == 1 || Act->getType() == types::TY_Nothing) 2297 Actions.append(Inputs.begin(), Inputs.end()); 2298 else 2299 Actions.push_back(C.MakeAction<LipoJobAction>(Inputs, Act->getType())); 2300 2301 // Handle debug info queries. 2302 Arg *A = Args.getLastArg(options::OPT_g_Group); 2303 bool enablesDebugInfo = A && !A->getOption().matches(options::OPT_g0) && 2304 !A->getOption().matches(options::OPT_gstabs); 2305 if ((enablesDebugInfo || willEmitRemarks(Args)) && 2306 ContainsCompileOrAssembleAction(Actions.back())) { 2307 2308 // Add a 'dsymutil' step if necessary, when debug info is enabled and we 2309 // have a compile input. We need to run 'dsymutil' ourselves in such cases 2310 // because the debug info will refer to a temporary object file which 2311 // will be removed at the end of the compilation process. 2312 if (Act->getType() == types::TY_Image) { 2313 ActionList Inputs; 2314 Inputs.push_back(Actions.back()); 2315 Actions.pop_back(); 2316 Actions.push_back( 2317 C.MakeAction<DsymutilJobAction>(Inputs, types::TY_dSYM)); 2318 } 2319 2320 // Verify the debug info output. 2321 if (Args.hasArg(options::OPT_verify_debug_info)) { 2322 Action* LastAction = Actions.back(); 2323 Actions.pop_back(); 2324 Actions.push_back(C.MakeAction<VerifyDebugInfoJobAction>( 2325 LastAction, types::TY_Nothing)); 2326 } 2327 } 2328 } 2329 } 2330 2331 bool Driver::DiagnoseInputExistence(const DerivedArgList &Args, StringRef Value, 2332 types::ID Ty, bool TypoCorrect) const { 2333 if (!getCheckInputsExist()) 2334 return true; 2335 2336 // stdin always exists. 2337 if (Value == "-") 2338 return true; 2339 2340 // If it's a header to be found in the system or user search path, then defer 2341 // complaints about its absence until those searches can be done. When we 2342 // are definitely processing headers for C++20 header units, extend this to 2343 // allow the user to put "-fmodule-header -xc++-header vector" for example. 2344 if (Ty == types::TY_CXXSHeader || Ty == types::TY_CXXUHeader || 2345 (ModulesModeCXX20 && Ty == types::TY_CXXHeader)) 2346 return true; 2347 2348 if (getVFS().exists(Value)) 2349 return true; 2350 2351 if (TypoCorrect) { 2352 // Check if the filename is a typo for an option flag. OptTable thinks 2353 // that all args that are not known options and that start with / are 2354 // filenames, but e.g. `/diagnostic:caret` is more likely a typo for 2355 // the option `/diagnostics:caret` than a reference to a file in the root 2356 // directory. 2357 unsigned IncludedFlagsBitmask; 2358 unsigned ExcludedFlagsBitmask; 2359 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 2360 getIncludeExcludeOptionFlagMasks(IsCLMode()); 2361 std::string Nearest; 2362 if (getOpts().findNearest(Value, Nearest, IncludedFlagsBitmask, 2363 ExcludedFlagsBitmask) <= 1) { 2364 Diag(clang::diag::err_drv_no_such_file_with_suggestion) 2365 << Value << Nearest; 2366 return false; 2367 } 2368 } 2369 2370 // In CL mode, don't error on apparently non-existent linker inputs, because 2371 // they can be influenced by linker flags the clang driver might not 2372 // understand. 2373 // Examples: 2374 // - `clang-cl main.cc ole32.lib` in a a non-MSVC shell will make the driver 2375 // module look for an MSVC installation in the registry. (We could ask 2376 // the MSVCToolChain object if it can find `ole32.lib`, but the logic to 2377 // look in the registry might move into lld-link in the future so that 2378 // lld-link invocations in non-MSVC shells just work too.) 2379 // - `clang-cl ... /link ...` can pass arbitrary flags to the linker, 2380 // including /libpath:, which is used to find .lib and .obj files. 2381 // So do not diagnose this on the driver level. Rely on the linker diagnosing 2382 // it. (If we don't end up invoking the linker, this means we'll emit a 2383 // "'linker' input unused [-Wunused-command-line-argument]" warning instead 2384 // of an error.) 2385 // 2386 // Only do this skip after the typo correction step above. `/Brepo` is treated 2387 // as TY_Object, but it's clearly a typo for `/Brepro`. It seems fine to emit 2388 // an error if we have a flag that's within an edit distance of 1 from a 2389 // flag. (Users can use `-Wl,` or `/linker` to launder the flag past the 2390 // driver in the unlikely case they run into this.) 2391 // 2392 // Don't do this for inputs that start with a '/', else we'd pass options 2393 // like /libpath: through to the linker silently. 2394 // 2395 // Emitting an error for linker inputs can also cause incorrect diagnostics 2396 // with the gcc driver. The command 2397 // clang -fuse-ld=lld -Wl,--chroot,some/dir /file.o 2398 // will make lld look for some/dir/file.o, while we will diagnose here that 2399 // `/file.o` does not exist. However, configure scripts check if 2400 // `clang /GR-` compiles without error to see if the compiler is cl.exe, 2401 // so we can't downgrade diagnostics for `/GR-` from an error to a warning 2402 // in cc mode. (We can in cl mode because cl.exe itself only warns on 2403 // unknown flags.) 2404 if (IsCLMode() && Ty == types::TY_Object && !Value.startswith("/")) 2405 return true; 2406 2407 Diag(clang::diag::err_drv_no_such_file) << Value; 2408 return false; 2409 } 2410 2411 // Get the C++20 Header Unit type corresponding to the input type. 2412 static types::ID CXXHeaderUnitType(ModuleHeaderMode HM) { 2413 switch (HM) { 2414 case HeaderMode_User: 2415 return types::TY_CXXUHeader; 2416 case HeaderMode_System: 2417 return types::TY_CXXSHeader; 2418 case HeaderMode_Default: 2419 break; 2420 case HeaderMode_None: 2421 llvm_unreachable("should not be called in this case"); 2422 } 2423 return types::TY_CXXHUHeader; 2424 } 2425 2426 // Construct a the list of inputs and their types. 2427 void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args, 2428 InputList &Inputs) const { 2429 const llvm::opt::OptTable &Opts = getOpts(); 2430 // Track the current user specified (-x) input. We also explicitly track the 2431 // argument used to set the type; we only want to claim the type when we 2432 // actually use it, so we warn about unused -x arguments. 2433 types::ID InputType = types::TY_Nothing; 2434 Arg *InputTypeArg = nullptr; 2435 2436 // The last /TC or /TP option sets the input type to C or C++ globally. 2437 if (Arg *TCTP = Args.getLastArgNoClaim(options::OPT__SLASH_TC, 2438 options::OPT__SLASH_TP)) { 2439 InputTypeArg = TCTP; 2440 InputType = TCTP->getOption().matches(options::OPT__SLASH_TC) 2441 ? types::TY_C 2442 : types::TY_CXX; 2443 2444 Arg *Previous = nullptr; 2445 bool ShowNote = false; 2446 for (Arg *A : 2447 Args.filtered(options::OPT__SLASH_TC, options::OPT__SLASH_TP)) { 2448 if (Previous) { 2449 Diag(clang::diag::warn_drv_overriding_flag_option) 2450 << Previous->getSpelling() << A->getSpelling(); 2451 ShowNote = true; 2452 } 2453 Previous = A; 2454 } 2455 if (ShowNote) 2456 Diag(clang::diag::note_drv_t_option_is_global); 2457 2458 // No driver mode exposes -x and /TC or /TP; we don't support mixing them. 2459 assert(!Args.hasArg(options::OPT_x) && "-x and /TC or /TP is not allowed"); 2460 } 2461 2462 // Warn -x after last input file has no effect 2463 { 2464 Arg *LastXArg = Args.getLastArgNoClaim(options::OPT_x); 2465 Arg *LastInputArg = Args.getLastArgNoClaim(options::OPT_INPUT); 2466 if (LastXArg && LastInputArg && LastInputArg->getIndex() < LastXArg->getIndex()) 2467 Diag(clang::diag::warn_drv_unused_x) << LastXArg->getValue(); 2468 } 2469 2470 for (Arg *A : Args) { 2471 if (A->getOption().getKind() == Option::InputClass) { 2472 const char *Value = A->getValue(); 2473 types::ID Ty = types::TY_INVALID; 2474 2475 // Infer the input type if necessary. 2476 if (InputType == types::TY_Nothing) { 2477 // If there was an explicit arg for this, claim it. 2478 if (InputTypeArg) 2479 InputTypeArg->claim(); 2480 2481 // stdin must be handled specially. 2482 if (memcmp(Value, "-", 2) == 0) { 2483 if (IsFlangMode()) { 2484 Ty = types::TY_Fortran; 2485 } else { 2486 // If running with -E, treat as a C input (this changes the 2487 // builtin macros, for example). This may be overridden by -ObjC 2488 // below. 2489 // 2490 // Otherwise emit an error but still use a valid type to avoid 2491 // spurious errors (e.g., no inputs). 2492 assert(!CCGenDiagnostics && "stdin produces no crash reproducer"); 2493 if (!Args.hasArgNoClaim(options::OPT_E) && !CCCIsCPP()) 2494 Diag(IsCLMode() ? clang::diag::err_drv_unknown_stdin_type_clang_cl 2495 : clang::diag::err_drv_unknown_stdin_type); 2496 Ty = types::TY_C; 2497 } 2498 } else { 2499 // Otherwise lookup by extension. 2500 // Fallback is C if invoked as C preprocessor, C++ if invoked with 2501 // clang-cl /E, or Object otherwise. 2502 // We use a host hook here because Darwin at least has its own 2503 // idea of what .s is. 2504 if (const char *Ext = strrchr(Value, '.')) 2505 Ty = TC.LookupTypeForExtension(Ext + 1); 2506 2507 if (Ty == types::TY_INVALID) { 2508 if (IsCLMode() && (Args.hasArgNoClaim(options::OPT_E) || CCGenDiagnostics)) 2509 Ty = types::TY_CXX; 2510 else if (CCCIsCPP() || CCGenDiagnostics) 2511 Ty = types::TY_C; 2512 else 2513 Ty = types::TY_Object; 2514 } 2515 2516 // If the driver is invoked as C++ compiler (like clang++ or c++) it 2517 // should autodetect some input files as C++ for g++ compatibility. 2518 if (CCCIsCXX()) { 2519 types::ID OldTy = Ty; 2520 Ty = types::lookupCXXTypeForCType(Ty); 2521 2522 // Do not complain about foo.h, when we are known to be processing 2523 // it as a C++20 header unit. 2524 if (Ty != OldTy && !(OldTy == types::TY_CHeader && hasHeaderMode())) 2525 Diag(clang::diag::warn_drv_treating_input_as_cxx) 2526 << getTypeName(OldTy) << getTypeName(Ty); 2527 } 2528 2529 // If running with -fthinlto-index=, extensions that normally identify 2530 // native object files actually identify LLVM bitcode files. 2531 if (Args.hasArgNoClaim(options::OPT_fthinlto_index_EQ) && 2532 Ty == types::TY_Object) 2533 Ty = types::TY_LLVM_BC; 2534 } 2535 2536 // -ObjC and -ObjC++ override the default language, but only for "source 2537 // files". We just treat everything that isn't a linker input as a 2538 // source file. 2539 // 2540 // FIXME: Clean this up if we move the phase sequence into the type. 2541 if (Ty != types::TY_Object) { 2542 if (Args.hasArg(options::OPT_ObjC)) 2543 Ty = types::TY_ObjC; 2544 else if (Args.hasArg(options::OPT_ObjCXX)) 2545 Ty = types::TY_ObjCXX; 2546 } 2547 2548 // Disambiguate headers that are meant to be header units from those 2549 // intended to be PCH. Avoid missing '.h' cases that are counted as 2550 // C headers by default - we know we are in C++ mode and we do not 2551 // want to issue a complaint about compiling things in the wrong mode. 2552 if ((Ty == types::TY_CXXHeader || Ty == types::TY_CHeader) && 2553 hasHeaderMode()) 2554 Ty = CXXHeaderUnitType(CXX20HeaderType); 2555 } else { 2556 assert(InputTypeArg && "InputType set w/o InputTypeArg"); 2557 if (!InputTypeArg->getOption().matches(options::OPT_x)) { 2558 // If emulating cl.exe, make sure that /TC and /TP don't affect input 2559 // object files. 2560 const char *Ext = strrchr(Value, '.'); 2561 if (Ext && TC.LookupTypeForExtension(Ext + 1) == types::TY_Object) 2562 Ty = types::TY_Object; 2563 } 2564 if (Ty == types::TY_INVALID) { 2565 Ty = InputType; 2566 InputTypeArg->claim(); 2567 } 2568 } 2569 2570 if (DiagnoseInputExistence(Args, Value, Ty, /*TypoCorrect=*/true)) 2571 Inputs.push_back(std::make_pair(Ty, A)); 2572 2573 } else if (A->getOption().matches(options::OPT__SLASH_Tc)) { 2574 StringRef Value = A->getValue(); 2575 if (DiagnoseInputExistence(Args, Value, types::TY_C, 2576 /*TypoCorrect=*/false)) { 2577 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2578 Inputs.push_back(std::make_pair(types::TY_C, InputArg)); 2579 } 2580 A->claim(); 2581 } else if (A->getOption().matches(options::OPT__SLASH_Tp)) { 2582 StringRef Value = A->getValue(); 2583 if (DiagnoseInputExistence(Args, Value, types::TY_CXX, 2584 /*TypoCorrect=*/false)) { 2585 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2586 Inputs.push_back(std::make_pair(types::TY_CXX, InputArg)); 2587 } 2588 A->claim(); 2589 } else if (A->getOption().hasFlag(options::LinkerInput)) { 2590 // Just treat as object type, we could make a special type for this if 2591 // necessary. 2592 Inputs.push_back(std::make_pair(types::TY_Object, A)); 2593 2594 } else if (A->getOption().matches(options::OPT_x)) { 2595 InputTypeArg = A; 2596 InputType = types::lookupTypeForTypeSpecifier(A->getValue()); 2597 A->claim(); 2598 2599 // Follow gcc behavior and treat as linker input for invalid -x 2600 // options. Its not clear why we shouldn't just revert to unknown; but 2601 // this isn't very important, we might as well be bug compatible. 2602 if (!InputType) { 2603 Diag(clang::diag::err_drv_unknown_language) << A->getValue(); 2604 InputType = types::TY_Object; 2605 } 2606 2607 // If the user has put -fmodule-header{,=} then we treat C++ headers as 2608 // header unit inputs. So we 'promote' -xc++-header appropriately. 2609 if (InputType == types::TY_CXXHeader && hasHeaderMode()) 2610 InputType = CXXHeaderUnitType(CXX20HeaderType); 2611 } else if (A->getOption().getID() == options::OPT_U) { 2612 assert(A->getNumValues() == 1 && "The /U option has one value."); 2613 StringRef Val = A->getValue(0); 2614 if (Val.find_first_of("/\\") != StringRef::npos) { 2615 // Warn about e.g. "/Users/me/myfile.c". 2616 Diag(diag::warn_slash_u_filename) << Val; 2617 Diag(diag::note_use_dashdash); 2618 } 2619 } 2620 } 2621 if (CCCIsCPP() && Inputs.empty()) { 2622 // If called as standalone preprocessor, stdin is processed 2623 // if no other input is present. 2624 Arg *A = MakeInputArg(Args, Opts, "-"); 2625 Inputs.push_back(std::make_pair(types::TY_C, A)); 2626 } 2627 } 2628 2629 namespace { 2630 /// Provides a convenient interface for different programming models to generate 2631 /// the required device actions. 2632 class OffloadingActionBuilder final { 2633 /// Flag used to trace errors in the builder. 2634 bool IsValid = false; 2635 2636 /// The compilation that is using this builder. 2637 Compilation &C; 2638 2639 /// Map between an input argument and the offload kinds used to process it. 2640 std::map<const Arg *, unsigned> InputArgToOffloadKindMap; 2641 2642 /// Map between a host action and its originating input argument. 2643 std::map<Action *, const Arg *> HostActionToInputArgMap; 2644 2645 /// Builder interface. It doesn't build anything or keep any state. 2646 class DeviceActionBuilder { 2647 public: 2648 typedef const llvm::SmallVectorImpl<phases::ID> PhasesTy; 2649 2650 enum ActionBuilderReturnCode { 2651 // The builder acted successfully on the current action. 2652 ABRT_Success, 2653 // The builder didn't have to act on the current action. 2654 ABRT_Inactive, 2655 // The builder was successful and requested the host action to not be 2656 // generated. 2657 ABRT_Ignore_Host, 2658 }; 2659 2660 protected: 2661 /// Compilation associated with this builder. 2662 Compilation &C; 2663 2664 /// Tool chains associated with this builder. The same programming 2665 /// model may have associated one or more tool chains. 2666 SmallVector<const ToolChain *, 2> ToolChains; 2667 2668 /// The derived arguments associated with this builder. 2669 DerivedArgList &Args; 2670 2671 /// The inputs associated with this builder. 2672 const Driver::InputList &Inputs; 2673 2674 /// The associated offload kind. 2675 Action::OffloadKind AssociatedOffloadKind = Action::OFK_None; 2676 2677 public: 2678 DeviceActionBuilder(Compilation &C, DerivedArgList &Args, 2679 const Driver::InputList &Inputs, 2680 Action::OffloadKind AssociatedOffloadKind) 2681 : C(C), Args(Args), Inputs(Inputs), 2682 AssociatedOffloadKind(AssociatedOffloadKind) {} 2683 virtual ~DeviceActionBuilder() {} 2684 2685 /// Fill up the array \a DA with all the device dependences that should be 2686 /// added to the provided host action \a HostAction. By default it is 2687 /// inactive. 2688 virtual ActionBuilderReturnCode 2689 getDeviceDependences(OffloadAction::DeviceDependences &DA, 2690 phases::ID CurPhase, phases::ID FinalPhase, 2691 PhasesTy &Phases) { 2692 return ABRT_Inactive; 2693 } 2694 2695 /// Update the state to include the provided host action \a HostAction as a 2696 /// dependency of the current device action. By default it is inactive. 2697 virtual ActionBuilderReturnCode addDeviceDepences(Action *HostAction) { 2698 return ABRT_Inactive; 2699 } 2700 2701 /// Append top level actions generated by the builder. 2702 virtual void appendTopLevelActions(ActionList &AL) {} 2703 2704 /// Append linker device actions generated by the builder. 2705 virtual void appendLinkDeviceActions(ActionList &AL) {} 2706 2707 /// Append linker host action generated by the builder. 2708 virtual Action* appendLinkHostActions(ActionList &AL) { return nullptr; } 2709 2710 /// Append linker actions generated by the builder. 2711 virtual void appendLinkDependences(OffloadAction::DeviceDependences &DA) {} 2712 2713 /// Initialize the builder. Return true if any initialization errors are 2714 /// found. 2715 virtual bool initialize() { return false; } 2716 2717 /// Return true if the builder can use bundling/unbundling. 2718 virtual bool canUseBundlerUnbundler() const { return false; } 2719 2720 /// Return true if this builder is valid. We have a valid builder if we have 2721 /// associated device tool chains. 2722 bool isValid() { return !ToolChains.empty(); } 2723 2724 /// Return the associated offload kind. 2725 Action::OffloadKind getAssociatedOffloadKind() { 2726 return AssociatedOffloadKind; 2727 } 2728 }; 2729 2730 /// Base class for CUDA/HIP action builder. It injects device code in 2731 /// the host backend action. 2732 class CudaActionBuilderBase : public DeviceActionBuilder { 2733 protected: 2734 /// Flags to signal if the user requested host-only or device-only 2735 /// compilation. 2736 bool CompileHostOnly = false; 2737 bool CompileDeviceOnly = false; 2738 bool EmitLLVM = false; 2739 bool EmitAsm = false; 2740 2741 /// ID to identify each device compilation. For CUDA it is simply the 2742 /// GPU arch string. For HIP it is either the GPU arch string or GPU 2743 /// arch string plus feature strings delimited by a plus sign, e.g. 2744 /// gfx906+xnack. 2745 struct TargetID { 2746 /// Target ID string which is persistent throughout the compilation. 2747 const char *ID; 2748 TargetID(CudaArch Arch) { ID = CudaArchToString(Arch); } 2749 TargetID(const char *ID) : ID(ID) {} 2750 operator const char *() { return ID; } 2751 operator StringRef() { return StringRef(ID); } 2752 }; 2753 /// List of GPU architectures to use in this compilation. 2754 SmallVector<TargetID, 4> GpuArchList; 2755 2756 /// The CUDA actions for the current input. 2757 ActionList CudaDeviceActions; 2758 2759 /// The CUDA fat binary if it was generated for the current input. 2760 Action *CudaFatBinary = nullptr; 2761 2762 /// Flag that is set to true if this builder acted on the current input. 2763 bool IsActive = false; 2764 2765 /// Flag for -fgpu-rdc. 2766 bool Relocatable = false; 2767 2768 /// Default GPU architecture if there's no one specified. 2769 CudaArch DefaultCudaArch = CudaArch::UNKNOWN; 2770 2771 /// Method to generate compilation unit ID specified by option 2772 /// '-fuse-cuid='. 2773 enum UseCUIDKind { CUID_Hash, CUID_Random, CUID_None, CUID_Invalid }; 2774 UseCUIDKind UseCUID = CUID_Hash; 2775 2776 /// Compilation unit ID specified by option '-cuid='. 2777 StringRef FixedCUID; 2778 2779 public: 2780 CudaActionBuilderBase(Compilation &C, DerivedArgList &Args, 2781 const Driver::InputList &Inputs, 2782 Action::OffloadKind OFKind) 2783 : DeviceActionBuilder(C, Args, Inputs, OFKind) {} 2784 2785 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 2786 // While generating code for CUDA, we only depend on the host input action 2787 // to trigger the creation of all the CUDA device actions. 2788 2789 // If we are dealing with an input action, replicate it for each GPU 2790 // architecture. If we are in host-only mode we return 'success' so that 2791 // the host uses the CUDA offload kind. 2792 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 2793 assert(!GpuArchList.empty() && 2794 "We should have at least one GPU architecture."); 2795 2796 // If the host input is not CUDA or HIP, we don't need to bother about 2797 // this input. 2798 if (!(IA->getType() == types::TY_CUDA || 2799 IA->getType() == types::TY_HIP || 2800 IA->getType() == types::TY_PP_HIP)) { 2801 // The builder will ignore this input. 2802 IsActive = false; 2803 return ABRT_Inactive; 2804 } 2805 2806 // Set the flag to true, so that the builder acts on the current input. 2807 IsActive = true; 2808 2809 if (CompileHostOnly) 2810 return ABRT_Success; 2811 2812 // Replicate inputs for each GPU architecture. 2813 auto Ty = IA->getType() == types::TY_HIP ? types::TY_HIP_DEVICE 2814 : types::TY_CUDA_DEVICE; 2815 std::string CUID = FixedCUID.str(); 2816 if (CUID.empty()) { 2817 if (UseCUID == CUID_Random) 2818 CUID = llvm::utohexstr(llvm::sys::Process::GetRandomNumber(), 2819 /*LowerCase=*/true); 2820 else if (UseCUID == CUID_Hash) { 2821 llvm::MD5 Hasher; 2822 llvm::MD5::MD5Result Hash; 2823 SmallString<256> RealPath; 2824 llvm::sys::fs::real_path(IA->getInputArg().getValue(), RealPath, 2825 /*expand_tilde=*/true); 2826 Hasher.update(RealPath); 2827 for (auto *A : Args) { 2828 if (A->getOption().matches(options::OPT_INPUT)) 2829 continue; 2830 Hasher.update(A->getAsString(Args)); 2831 } 2832 Hasher.final(Hash); 2833 CUID = llvm::utohexstr(Hash.low(), /*LowerCase=*/true); 2834 } 2835 } 2836 IA->setId(CUID); 2837 2838 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 2839 CudaDeviceActions.push_back( 2840 C.MakeAction<InputAction>(IA->getInputArg(), Ty, IA->getId())); 2841 } 2842 2843 return ABRT_Success; 2844 } 2845 2846 // If this is an unbundling action use it as is for each CUDA toolchain. 2847 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 2848 2849 // If -fgpu-rdc is disabled, should not unbundle since there is no 2850 // device code to link. 2851 if (UA->getType() == types::TY_Object && !Relocatable) 2852 return ABRT_Inactive; 2853 2854 CudaDeviceActions.clear(); 2855 auto *IA = cast<InputAction>(UA->getInputs().back()); 2856 std::string FileName = IA->getInputArg().getAsString(Args); 2857 // Check if the type of the file is the same as the action. Do not 2858 // unbundle it if it is not. Do not unbundle .so files, for example, 2859 // which are not object files. 2860 if (IA->getType() == types::TY_Object && 2861 (!llvm::sys::path::has_extension(FileName) || 2862 types::lookupTypeForExtension( 2863 llvm::sys::path::extension(FileName).drop_front()) != 2864 types::TY_Object)) 2865 return ABRT_Inactive; 2866 2867 for (auto Arch : GpuArchList) { 2868 CudaDeviceActions.push_back(UA); 2869 UA->registerDependentActionInfo(ToolChains[0], Arch, 2870 AssociatedOffloadKind); 2871 } 2872 IsActive = true; 2873 return ABRT_Success; 2874 } 2875 2876 return IsActive ? ABRT_Success : ABRT_Inactive; 2877 } 2878 2879 void appendTopLevelActions(ActionList &AL) override { 2880 // Utility to append actions to the top level list. 2881 auto AddTopLevel = [&](Action *A, TargetID TargetID) { 2882 OffloadAction::DeviceDependences Dep; 2883 Dep.add(*A, *ToolChains.front(), TargetID, AssociatedOffloadKind); 2884 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 2885 }; 2886 2887 // If we have a fat binary, add it to the list. 2888 if (CudaFatBinary) { 2889 AddTopLevel(CudaFatBinary, CudaArch::UNUSED); 2890 CudaDeviceActions.clear(); 2891 CudaFatBinary = nullptr; 2892 return; 2893 } 2894 2895 if (CudaDeviceActions.empty()) 2896 return; 2897 2898 // If we have CUDA actions at this point, that's because we have a have 2899 // partial compilation, so we should have an action for each GPU 2900 // architecture. 2901 assert(CudaDeviceActions.size() == GpuArchList.size() && 2902 "Expecting one action per GPU architecture."); 2903 assert(ToolChains.size() == 1 && 2904 "Expecting to have a single CUDA toolchain."); 2905 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) 2906 AddTopLevel(CudaDeviceActions[I], GpuArchList[I]); 2907 2908 CudaDeviceActions.clear(); 2909 } 2910 2911 /// Get canonicalized offload arch option. \returns empty StringRef if the 2912 /// option is invalid. 2913 virtual StringRef getCanonicalOffloadArch(StringRef Arch) = 0; 2914 2915 virtual llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 2916 getConflictOffloadArchCombination(const std::set<StringRef> &GpuArchs) = 0; 2917 2918 bool initialize() override { 2919 assert(AssociatedOffloadKind == Action::OFK_Cuda || 2920 AssociatedOffloadKind == Action::OFK_HIP); 2921 2922 // We don't need to support CUDA. 2923 if (AssociatedOffloadKind == Action::OFK_Cuda && 2924 !C.hasOffloadToolChain<Action::OFK_Cuda>()) 2925 return false; 2926 2927 // We don't need to support HIP. 2928 if (AssociatedOffloadKind == Action::OFK_HIP && 2929 !C.hasOffloadToolChain<Action::OFK_HIP>()) 2930 return false; 2931 2932 Relocatable = Args.hasFlag(options::OPT_fgpu_rdc, 2933 options::OPT_fno_gpu_rdc, /*Default=*/false); 2934 2935 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 2936 assert(HostTC && "No toolchain for host compilation."); 2937 if (HostTC->getTriple().isNVPTX() || 2938 HostTC->getTriple().getArch() == llvm::Triple::amdgcn) { 2939 // We do not support targeting NVPTX/AMDGCN for host compilation. Throw 2940 // an error and abort pipeline construction early so we don't trip 2941 // asserts that assume device-side compilation. 2942 C.getDriver().Diag(diag::err_drv_cuda_host_arch) 2943 << HostTC->getTriple().getArchName(); 2944 return true; 2945 } 2946 2947 ToolChains.push_back( 2948 AssociatedOffloadKind == Action::OFK_Cuda 2949 ? C.getSingleOffloadToolChain<Action::OFK_Cuda>() 2950 : C.getSingleOffloadToolChain<Action::OFK_HIP>()); 2951 2952 CompileHostOnly = C.getDriver().offloadHostOnly(); 2953 CompileDeviceOnly = C.getDriver().offloadDeviceOnly(); 2954 EmitLLVM = Args.getLastArg(options::OPT_emit_llvm); 2955 EmitAsm = Args.getLastArg(options::OPT_S); 2956 FixedCUID = Args.getLastArgValue(options::OPT_cuid_EQ); 2957 if (Arg *A = Args.getLastArg(options::OPT_fuse_cuid_EQ)) { 2958 StringRef UseCUIDStr = A->getValue(); 2959 UseCUID = llvm::StringSwitch<UseCUIDKind>(UseCUIDStr) 2960 .Case("hash", CUID_Hash) 2961 .Case("random", CUID_Random) 2962 .Case("none", CUID_None) 2963 .Default(CUID_Invalid); 2964 if (UseCUID == CUID_Invalid) { 2965 C.getDriver().Diag(diag::err_drv_invalid_value) 2966 << A->getAsString(Args) << UseCUIDStr; 2967 C.setContainsError(); 2968 return true; 2969 } 2970 } 2971 2972 // --offload and --offload-arch options are mutually exclusive. 2973 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 2974 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 2975 options::OPT_no_offload_arch_EQ)) { 2976 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) << "--offload-arch" 2977 << "--offload"; 2978 } 2979 2980 // Collect all offload arch parameters, removing duplicates. 2981 std::set<StringRef> GpuArchs; 2982 bool Error = false; 2983 for (Arg *A : Args) { 2984 if (!(A->getOption().matches(options::OPT_offload_arch_EQ) || 2985 A->getOption().matches(options::OPT_no_offload_arch_EQ))) 2986 continue; 2987 A->claim(); 2988 2989 for (StringRef ArchStr : llvm::split(A->getValue(), ",")) { 2990 if (A->getOption().matches(options::OPT_no_offload_arch_EQ) && 2991 ArchStr == "all") { 2992 GpuArchs.clear(); 2993 } else { 2994 ArchStr = getCanonicalOffloadArch(ArchStr); 2995 if (ArchStr.empty()) { 2996 Error = true; 2997 } else if (A->getOption().matches(options::OPT_offload_arch_EQ)) 2998 GpuArchs.insert(ArchStr); 2999 else if (A->getOption().matches(options::OPT_no_offload_arch_EQ)) 3000 GpuArchs.erase(ArchStr); 3001 else 3002 llvm_unreachable("Unexpected option."); 3003 } 3004 } 3005 } 3006 3007 auto &&ConflictingArchs = getConflictOffloadArchCombination(GpuArchs); 3008 if (ConflictingArchs) { 3009 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 3010 << ConflictingArchs->first << ConflictingArchs->second; 3011 C.setContainsError(); 3012 return true; 3013 } 3014 3015 // Collect list of GPUs remaining in the set. 3016 for (auto Arch : GpuArchs) 3017 GpuArchList.push_back(Arch.data()); 3018 3019 // Default to sm_20 which is the lowest common denominator for 3020 // supported GPUs. sm_20 code should work correctly, if 3021 // suboptimally, on all newer GPUs. 3022 if (GpuArchList.empty()) { 3023 if (ToolChains.front()->getTriple().isSPIRV()) 3024 GpuArchList.push_back(CudaArch::Generic); 3025 else 3026 GpuArchList.push_back(DefaultCudaArch); 3027 } 3028 3029 return Error; 3030 } 3031 }; 3032 3033 /// \brief CUDA action builder. It injects device code in the host backend 3034 /// action. 3035 class CudaActionBuilder final : public CudaActionBuilderBase { 3036 public: 3037 CudaActionBuilder(Compilation &C, DerivedArgList &Args, 3038 const Driver::InputList &Inputs) 3039 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) { 3040 DefaultCudaArch = CudaArch::SM_35; 3041 } 3042 3043 StringRef getCanonicalOffloadArch(StringRef ArchStr) override { 3044 CudaArch Arch = StringToCudaArch(ArchStr); 3045 if (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch)) { 3046 C.getDriver().Diag(clang::diag::err_drv_cuda_bad_gpu_arch) << ArchStr; 3047 return StringRef(); 3048 } 3049 return CudaArchToString(Arch); 3050 } 3051 3052 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3053 getConflictOffloadArchCombination( 3054 const std::set<StringRef> &GpuArchs) override { 3055 return llvm::None; 3056 } 3057 3058 ActionBuilderReturnCode 3059 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3060 phases::ID CurPhase, phases::ID FinalPhase, 3061 PhasesTy &Phases) override { 3062 if (!IsActive) 3063 return ABRT_Inactive; 3064 3065 // If we don't have more CUDA actions, we don't have any dependences to 3066 // create for the host. 3067 if (CudaDeviceActions.empty()) 3068 return ABRT_Success; 3069 3070 assert(CudaDeviceActions.size() == GpuArchList.size() && 3071 "Expecting one action per GPU architecture."); 3072 assert(!CompileHostOnly && 3073 "Not expecting CUDA actions in host-only compilation."); 3074 3075 // If we are generating code for the device or we are in a backend phase, 3076 // we attempt to generate the fat binary. We compile each arch to ptx and 3077 // assemble to cubin, then feed the cubin *and* the ptx into a device 3078 // "link" action, which uses fatbinary to combine these cubins into one 3079 // fatbin. The fatbin is then an input to the host action if not in 3080 // device-only mode. 3081 if (CompileDeviceOnly || CurPhase == phases::Backend) { 3082 ActionList DeviceActions; 3083 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3084 // Produce the device action from the current phase up to the assemble 3085 // phase. 3086 for (auto Ph : Phases) { 3087 // Skip the phases that were already dealt with. 3088 if (Ph < CurPhase) 3089 continue; 3090 // We have to be consistent with the host final phase. 3091 if (Ph > FinalPhase) 3092 break; 3093 3094 CudaDeviceActions[I] = C.getDriver().ConstructPhaseAction( 3095 C, Args, Ph, CudaDeviceActions[I], Action::OFK_Cuda); 3096 3097 if (Ph == phases::Assemble) 3098 break; 3099 } 3100 3101 // If we didn't reach the assemble phase, we can't generate the fat 3102 // binary. We don't need to generate the fat binary if we are not in 3103 // device-only mode. 3104 if (!isa<AssembleJobAction>(CudaDeviceActions[I]) || 3105 CompileDeviceOnly) 3106 continue; 3107 3108 Action *AssembleAction = CudaDeviceActions[I]; 3109 assert(AssembleAction->getType() == types::TY_Object); 3110 assert(AssembleAction->getInputs().size() == 1); 3111 3112 Action *BackendAction = AssembleAction->getInputs()[0]; 3113 assert(BackendAction->getType() == types::TY_PP_Asm); 3114 3115 for (auto &A : {AssembleAction, BackendAction}) { 3116 OffloadAction::DeviceDependences DDep; 3117 DDep.add(*A, *ToolChains.front(), GpuArchList[I], Action::OFK_Cuda); 3118 DeviceActions.push_back( 3119 C.MakeAction<OffloadAction>(DDep, A->getType())); 3120 } 3121 } 3122 3123 // We generate the fat binary if we have device input actions. 3124 if (!DeviceActions.empty()) { 3125 CudaFatBinary = 3126 C.MakeAction<LinkJobAction>(DeviceActions, types::TY_CUDA_FATBIN); 3127 3128 if (!CompileDeviceOnly) { 3129 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3130 Action::OFK_Cuda); 3131 // Clear the fat binary, it is already a dependence to an host 3132 // action. 3133 CudaFatBinary = nullptr; 3134 } 3135 3136 // Remove the CUDA actions as they are already connected to an host 3137 // action or fat binary. 3138 CudaDeviceActions.clear(); 3139 } 3140 3141 // We avoid creating host action in device-only mode. 3142 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3143 } else if (CurPhase > phases::Backend) { 3144 // If we are past the backend phase and still have a device action, we 3145 // don't have to do anything as this action is already a device 3146 // top-level action. 3147 return ABRT_Success; 3148 } 3149 3150 assert(CurPhase < phases::Backend && "Generating single CUDA " 3151 "instructions should only occur " 3152 "before the backend phase!"); 3153 3154 // By default, we produce an action for each device arch. 3155 for (Action *&A : CudaDeviceActions) 3156 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3157 3158 return ABRT_Success; 3159 } 3160 }; 3161 /// \brief HIP action builder. It injects device code in the host backend 3162 /// action. 3163 class HIPActionBuilder final : public CudaActionBuilderBase { 3164 /// The linker inputs obtained for each device arch. 3165 SmallVector<ActionList, 8> DeviceLinkerInputs; 3166 // The default bundling behavior depends on the type of output, therefore 3167 // BundleOutput needs to be tri-value: None, true, or false. 3168 // Bundle code objects except --no-gpu-output is specified for device 3169 // only compilation. Bundle other type of output files only if 3170 // --gpu-bundle-output is specified for device only compilation. 3171 Optional<bool> BundleOutput; 3172 3173 public: 3174 HIPActionBuilder(Compilation &C, DerivedArgList &Args, 3175 const Driver::InputList &Inputs) 3176 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) { 3177 DefaultCudaArch = CudaArch::GFX803; 3178 if (Args.hasArg(options::OPT_gpu_bundle_output, 3179 options::OPT_no_gpu_bundle_output)) 3180 BundleOutput = Args.hasFlag(options::OPT_gpu_bundle_output, 3181 options::OPT_no_gpu_bundle_output, true); 3182 } 3183 3184 bool canUseBundlerUnbundler() const override { return true; } 3185 3186 StringRef getCanonicalOffloadArch(StringRef IdStr) override { 3187 llvm::StringMap<bool> Features; 3188 // getHIPOffloadTargetTriple() is known to return valid value as it has 3189 // been called successfully in the CreateOffloadingDeviceToolChains(). 3190 auto ArchStr = parseTargetID( 3191 *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()), IdStr, 3192 &Features); 3193 if (!ArchStr) { 3194 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << IdStr; 3195 C.setContainsError(); 3196 return StringRef(); 3197 } 3198 auto CanId = getCanonicalTargetID(*ArchStr, Features); 3199 return Args.MakeArgStringRef(CanId); 3200 }; 3201 3202 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3203 getConflictOffloadArchCombination( 3204 const std::set<StringRef> &GpuArchs) override { 3205 return getConflictTargetIDCombination(GpuArchs); 3206 } 3207 3208 ActionBuilderReturnCode 3209 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3210 phases::ID CurPhase, phases::ID FinalPhase, 3211 PhasesTy &Phases) override { 3212 if (!IsActive) 3213 return ABRT_Inactive; 3214 3215 // amdgcn does not support linking of object files, therefore we skip 3216 // backend and assemble phases to output LLVM IR. Except for generating 3217 // non-relocatable device code, where we generate fat binary for device 3218 // code and pass to host in Backend phase. 3219 if (CudaDeviceActions.empty()) 3220 return ABRT_Success; 3221 3222 assert(((CurPhase == phases::Link && Relocatable) || 3223 CudaDeviceActions.size() == GpuArchList.size()) && 3224 "Expecting one action per GPU architecture."); 3225 assert(!CompileHostOnly && 3226 "Not expecting HIP actions in host-only compilation."); 3227 3228 if (!Relocatable && CurPhase == phases::Backend && !EmitLLVM && 3229 !EmitAsm) { 3230 // If we are in backend phase, we attempt to generate the fat binary. 3231 // We compile each arch to IR and use a link action to generate code 3232 // object containing ISA. Then we use a special "link" action to create 3233 // a fat binary containing all the code objects for different GPU's. 3234 // The fat binary is then an input to the host action. 3235 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3236 if (C.getDriver().isUsingLTO(/*IsOffload=*/true)) { 3237 // When LTO is enabled, skip the backend and assemble phases and 3238 // use lld to link the bitcode. 3239 ActionList AL; 3240 AL.push_back(CudaDeviceActions[I]); 3241 // Create a link action to link device IR with device library 3242 // and generate ISA. 3243 CudaDeviceActions[I] = 3244 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3245 } else { 3246 // When LTO is not enabled, we follow the conventional 3247 // compiler phases, including backend and assemble phases. 3248 ActionList AL; 3249 Action *BackendAction = nullptr; 3250 if (ToolChains.front()->getTriple().isSPIRV()) { 3251 // Emit LLVM bitcode for SPIR-V targets. SPIR-V device tool chain 3252 // (HIPSPVToolChain) runs post-link LLVM IR passes. 3253 types::ID Output = Args.hasArg(options::OPT_S) 3254 ? types::TY_LLVM_IR 3255 : types::TY_LLVM_BC; 3256 BackendAction = 3257 C.MakeAction<BackendJobAction>(CudaDeviceActions[I], Output); 3258 } else 3259 BackendAction = C.getDriver().ConstructPhaseAction( 3260 C, Args, phases::Backend, CudaDeviceActions[I], 3261 AssociatedOffloadKind); 3262 auto AssembleAction = C.getDriver().ConstructPhaseAction( 3263 C, Args, phases::Assemble, BackendAction, 3264 AssociatedOffloadKind); 3265 AL.push_back(AssembleAction); 3266 // Create a link action to link device IR with device library 3267 // and generate ISA. 3268 CudaDeviceActions[I] = 3269 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3270 } 3271 3272 // OffloadingActionBuilder propagates device arch until an offload 3273 // action. Since the next action for creating fatbin does 3274 // not have device arch, whereas the above link action and its input 3275 // have device arch, an offload action is needed to stop the null 3276 // device arch of the next action being propagated to the above link 3277 // action. 3278 OffloadAction::DeviceDependences DDep; 3279 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3280 AssociatedOffloadKind); 3281 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3282 DDep, CudaDeviceActions[I]->getType()); 3283 } 3284 3285 if (!CompileDeviceOnly || !BundleOutput || *BundleOutput) { 3286 // Create HIP fat binary with a special "link" action. 3287 CudaFatBinary = C.MakeAction<LinkJobAction>(CudaDeviceActions, 3288 types::TY_HIP_FATBIN); 3289 3290 if (!CompileDeviceOnly) { 3291 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3292 AssociatedOffloadKind); 3293 // Clear the fat binary, it is already a dependence to an host 3294 // action. 3295 CudaFatBinary = nullptr; 3296 } 3297 3298 // Remove the CUDA actions as they are already connected to an host 3299 // action or fat binary. 3300 CudaDeviceActions.clear(); 3301 } 3302 3303 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3304 } else if (CurPhase == phases::Link) { 3305 // Save CudaDeviceActions to DeviceLinkerInputs for each GPU subarch. 3306 // This happens to each device action originated from each input file. 3307 // Later on, device actions in DeviceLinkerInputs are used to create 3308 // device link actions in appendLinkDependences and the created device 3309 // link actions are passed to the offload action as device dependence. 3310 DeviceLinkerInputs.resize(CudaDeviceActions.size()); 3311 auto LI = DeviceLinkerInputs.begin(); 3312 for (auto *A : CudaDeviceActions) { 3313 LI->push_back(A); 3314 ++LI; 3315 } 3316 3317 // We will pass the device action as a host dependence, so we don't 3318 // need to do anything else with them. 3319 CudaDeviceActions.clear(); 3320 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3321 } 3322 3323 // By default, we produce an action for each device arch. 3324 for (Action *&A : CudaDeviceActions) 3325 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A, 3326 AssociatedOffloadKind); 3327 3328 if (CompileDeviceOnly && CurPhase == FinalPhase && BundleOutput && 3329 BundleOutput.value()) { 3330 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3331 OffloadAction::DeviceDependences DDep; 3332 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3333 AssociatedOffloadKind); 3334 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3335 DDep, CudaDeviceActions[I]->getType()); 3336 } 3337 CudaFatBinary = 3338 C.MakeAction<OffloadBundlingJobAction>(CudaDeviceActions); 3339 CudaDeviceActions.clear(); 3340 } 3341 3342 return (CompileDeviceOnly && CurPhase == FinalPhase) ? ABRT_Ignore_Host 3343 : ABRT_Success; 3344 } 3345 3346 void appendLinkDeviceActions(ActionList &AL) override { 3347 if (DeviceLinkerInputs.size() == 0) 3348 return; 3349 3350 assert(DeviceLinkerInputs.size() == GpuArchList.size() && 3351 "Linker inputs and GPU arch list sizes do not match."); 3352 3353 ActionList Actions; 3354 unsigned I = 0; 3355 // Append a new link action for each device. 3356 // Each entry in DeviceLinkerInputs corresponds to a GPU arch. 3357 for (auto &LI : DeviceLinkerInputs) { 3358 3359 types::ID Output = Args.hasArg(options::OPT_emit_llvm) 3360 ? types::TY_LLVM_BC 3361 : types::TY_Image; 3362 3363 auto *DeviceLinkAction = C.MakeAction<LinkJobAction>(LI, Output); 3364 // Linking all inputs for the current GPU arch. 3365 // LI contains all the inputs for the linker. 3366 OffloadAction::DeviceDependences DeviceLinkDeps; 3367 DeviceLinkDeps.add(*DeviceLinkAction, *ToolChains[0], 3368 GpuArchList[I], AssociatedOffloadKind); 3369 Actions.push_back(C.MakeAction<OffloadAction>( 3370 DeviceLinkDeps, DeviceLinkAction->getType())); 3371 ++I; 3372 } 3373 DeviceLinkerInputs.clear(); 3374 3375 // If emitting LLVM, do not generate final host/device compilation action 3376 if (Args.hasArg(options::OPT_emit_llvm)) { 3377 AL.append(Actions); 3378 return; 3379 } 3380 3381 // Create a host object from all the device images by embedding them 3382 // in a fat binary for mixed host-device compilation. For device-only 3383 // compilation, creates a fat binary. 3384 OffloadAction::DeviceDependences DDeps; 3385 if (!CompileDeviceOnly || !BundleOutput || *BundleOutput) { 3386 auto *TopDeviceLinkAction = C.MakeAction<LinkJobAction>( 3387 Actions, 3388 CompileDeviceOnly ? types::TY_HIP_FATBIN : types::TY_Object); 3389 DDeps.add(*TopDeviceLinkAction, *ToolChains[0], nullptr, 3390 AssociatedOffloadKind); 3391 // Offload the host object to the host linker. 3392 AL.push_back( 3393 C.MakeAction<OffloadAction>(DDeps, TopDeviceLinkAction->getType())); 3394 } else { 3395 AL.append(Actions); 3396 } 3397 } 3398 3399 Action* appendLinkHostActions(ActionList &AL) override { return AL.back(); } 3400 3401 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3402 }; 3403 3404 /// OpenMP action builder. The host bitcode is passed to the device frontend 3405 /// and all the device linked images are passed to the host link phase. 3406 class OpenMPActionBuilder final : public DeviceActionBuilder { 3407 /// The OpenMP actions for the current input. 3408 ActionList OpenMPDeviceActions; 3409 3410 /// The linker inputs obtained for each toolchain. 3411 SmallVector<ActionList, 8> DeviceLinkerInputs; 3412 3413 public: 3414 OpenMPActionBuilder(Compilation &C, DerivedArgList &Args, 3415 const Driver::InputList &Inputs) 3416 : DeviceActionBuilder(C, Args, Inputs, Action::OFK_OpenMP) {} 3417 3418 ActionBuilderReturnCode 3419 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3420 phases::ID CurPhase, phases::ID FinalPhase, 3421 PhasesTy &Phases) override { 3422 if (OpenMPDeviceActions.empty()) 3423 return ABRT_Inactive; 3424 3425 // We should always have an action for each input. 3426 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3427 "Number of OpenMP actions and toolchains do not match."); 3428 3429 // The host only depends on device action in the linking phase, when all 3430 // the device images have to be embedded in the host image. 3431 if (CurPhase == phases::Link) { 3432 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3433 "Toolchains and linker inputs sizes do not match."); 3434 auto LI = DeviceLinkerInputs.begin(); 3435 for (auto *A : OpenMPDeviceActions) { 3436 LI->push_back(A); 3437 ++LI; 3438 } 3439 3440 // We passed the device action as a host dependence, so we don't need to 3441 // do anything else with them. 3442 OpenMPDeviceActions.clear(); 3443 return ABRT_Success; 3444 } 3445 3446 // By default, we produce an action for each device arch. 3447 for (Action *&A : OpenMPDeviceActions) 3448 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3449 3450 return ABRT_Success; 3451 } 3452 3453 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 3454 3455 // If this is an input action replicate it for each OpenMP toolchain. 3456 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 3457 OpenMPDeviceActions.clear(); 3458 for (unsigned I = 0; I < ToolChains.size(); ++I) 3459 OpenMPDeviceActions.push_back( 3460 C.MakeAction<InputAction>(IA->getInputArg(), IA->getType())); 3461 return ABRT_Success; 3462 } 3463 3464 // If this is an unbundling action use it as is for each OpenMP toolchain. 3465 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 3466 OpenMPDeviceActions.clear(); 3467 auto *IA = cast<InputAction>(UA->getInputs().back()); 3468 std::string FileName = IA->getInputArg().getAsString(Args); 3469 // Check if the type of the file is the same as the action. Do not 3470 // unbundle it if it is not. Do not unbundle .so files, for example, 3471 // which are not object files. 3472 if (IA->getType() == types::TY_Object && 3473 (!llvm::sys::path::has_extension(FileName) || 3474 types::lookupTypeForExtension( 3475 llvm::sys::path::extension(FileName).drop_front()) != 3476 types::TY_Object)) 3477 return ABRT_Inactive; 3478 for (unsigned I = 0; I < ToolChains.size(); ++I) { 3479 OpenMPDeviceActions.push_back(UA); 3480 UA->registerDependentActionInfo( 3481 ToolChains[I], /*BoundArch=*/StringRef(), Action::OFK_OpenMP); 3482 } 3483 return ABRT_Success; 3484 } 3485 3486 // When generating code for OpenMP we use the host compile phase result as 3487 // a dependence to the device compile phase so that it can learn what 3488 // declarations should be emitted. However, this is not the only use for 3489 // the host action, so we prevent it from being collapsed. 3490 if (isa<CompileJobAction>(HostAction)) { 3491 HostAction->setCannotBeCollapsedWithNextDependentAction(); 3492 assert(ToolChains.size() == OpenMPDeviceActions.size() && 3493 "Toolchains and device action sizes do not match."); 3494 OffloadAction::HostDependence HDep( 3495 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3496 /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3497 auto TC = ToolChains.begin(); 3498 for (Action *&A : OpenMPDeviceActions) { 3499 assert(isa<CompileJobAction>(A)); 3500 OffloadAction::DeviceDependences DDep; 3501 DDep.add(*A, **TC, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3502 A = C.MakeAction<OffloadAction>(HDep, DDep); 3503 ++TC; 3504 } 3505 } 3506 return ABRT_Success; 3507 } 3508 3509 void appendTopLevelActions(ActionList &AL) override { 3510 if (OpenMPDeviceActions.empty()) 3511 return; 3512 3513 // We should always have an action for each input. 3514 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3515 "Number of OpenMP actions and toolchains do not match."); 3516 3517 // Append all device actions followed by the proper offload action. 3518 auto TI = ToolChains.begin(); 3519 for (auto *A : OpenMPDeviceActions) { 3520 OffloadAction::DeviceDependences Dep; 3521 Dep.add(*A, **TI, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3522 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 3523 ++TI; 3524 } 3525 // We no longer need the action stored in this builder. 3526 OpenMPDeviceActions.clear(); 3527 } 3528 3529 void appendLinkDeviceActions(ActionList &AL) override { 3530 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3531 "Toolchains and linker inputs sizes do not match."); 3532 3533 // Append a new link action for each device. 3534 auto TC = ToolChains.begin(); 3535 for (auto &LI : DeviceLinkerInputs) { 3536 auto *DeviceLinkAction = 3537 C.MakeAction<LinkJobAction>(LI, types::TY_Image); 3538 OffloadAction::DeviceDependences DeviceLinkDeps; 3539 DeviceLinkDeps.add(*DeviceLinkAction, **TC, /*BoundArch=*/nullptr, 3540 Action::OFK_OpenMP); 3541 AL.push_back(C.MakeAction<OffloadAction>(DeviceLinkDeps, 3542 DeviceLinkAction->getType())); 3543 ++TC; 3544 } 3545 DeviceLinkerInputs.clear(); 3546 } 3547 3548 Action* appendLinkHostActions(ActionList &AL) override { 3549 // Create wrapper bitcode from the result of device link actions and compile 3550 // it to an object which will be added to the host link command. 3551 auto *BC = C.MakeAction<OffloadWrapperJobAction>(AL, types::TY_LLVM_BC); 3552 auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm); 3553 return C.MakeAction<AssembleJobAction>(ASM, types::TY_Object); 3554 } 3555 3556 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3557 3558 bool initialize() override { 3559 // Get the OpenMP toolchains. If we don't get any, the action builder will 3560 // know there is nothing to do related to OpenMP offloading. 3561 auto OpenMPTCRange = C.getOffloadToolChains<Action::OFK_OpenMP>(); 3562 for (auto TI = OpenMPTCRange.first, TE = OpenMPTCRange.second; TI != TE; 3563 ++TI) 3564 ToolChains.push_back(TI->second); 3565 3566 DeviceLinkerInputs.resize(ToolChains.size()); 3567 return false; 3568 } 3569 3570 bool canUseBundlerUnbundler() const override { 3571 // OpenMP should use bundled files whenever possible. 3572 return true; 3573 } 3574 }; 3575 3576 /// 3577 /// TODO: Add the implementation for other specialized builders here. 3578 /// 3579 3580 /// Specialized builders being used by this offloading action builder. 3581 SmallVector<DeviceActionBuilder *, 4> SpecializedBuilders; 3582 3583 /// Flag set to true if all valid builders allow file bundling/unbundling. 3584 bool CanUseBundler; 3585 3586 public: 3587 OffloadingActionBuilder(Compilation &C, DerivedArgList &Args, 3588 const Driver::InputList &Inputs) 3589 : C(C) { 3590 // Create a specialized builder for each device toolchain. 3591 3592 IsValid = true; 3593 3594 // Create a specialized builder for CUDA. 3595 SpecializedBuilders.push_back(new CudaActionBuilder(C, Args, Inputs)); 3596 3597 // Create a specialized builder for HIP. 3598 SpecializedBuilders.push_back(new HIPActionBuilder(C, Args, Inputs)); 3599 3600 // Create a specialized builder for OpenMP. 3601 SpecializedBuilders.push_back(new OpenMPActionBuilder(C, Args, Inputs)); 3602 3603 // 3604 // TODO: Build other specialized builders here. 3605 // 3606 3607 // Initialize all the builders, keeping track of errors. If all valid 3608 // builders agree that we can use bundling, set the flag to true. 3609 unsigned ValidBuilders = 0u; 3610 unsigned ValidBuildersSupportingBundling = 0u; 3611 for (auto *SB : SpecializedBuilders) { 3612 IsValid = IsValid && !SB->initialize(); 3613 3614 // Update the counters if the builder is valid. 3615 if (SB->isValid()) { 3616 ++ValidBuilders; 3617 if (SB->canUseBundlerUnbundler()) 3618 ++ValidBuildersSupportingBundling; 3619 } 3620 } 3621 CanUseBundler = 3622 ValidBuilders && ValidBuilders == ValidBuildersSupportingBundling; 3623 } 3624 3625 ~OffloadingActionBuilder() { 3626 for (auto *SB : SpecializedBuilders) 3627 delete SB; 3628 } 3629 3630 /// Record a host action and its originating input argument. 3631 void recordHostAction(Action *HostAction, const Arg *InputArg) { 3632 assert(HostAction && "Invalid host action"); 3633 assert(InputArg && "Invalid input argument"); 3634 auto Loc = HostActionToInputArgMap.find(HostAction); 3635 if (Loc == HostActionToInputArgMap.end()) 3636 HostActionToInputArgMap[HostAction] = InputArg; 3637 assert(HostActionToInputArgMap[HostAction] == InputArg && 3638 "host action mapped to multiple input arguments"); 3639 } 3640 3641 /// Generate an action that adds device dependences (if any) to a host action. 3642 /// If no device dependence actions exist, just return the host action \a 3643 /// HostAction. If an error is found or if no builder requires the host action 3644 /// to be generated, return nullptr. 3645 Action * 3646 addDeviceDependencesToHostAction(Action *HostAction, const Arg *InputArg, 3647 phases::ID CurPhase, phases::ID FinalPhase, 3648 DeviceActionBuilder::PhasesTy &Phases) { 3649 if (!IsValid) 3650 return nullptr; 3651 3652 if (SpecializedBuilders.empty()) 3653 return HostAction; 3654 3655 assert(HostAction && "Invalid host action!"); 3656 recordHostAction(HostAction, InputArg); 3657 3658 OffloadAction::DeviceDependences DDeps; 3659 // Check if all the programming models agree we should not emit the host 3660 // action. Also, keep track of the offloading kinds employed. 3661 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3662 unsigned InactiveBuilders = 0u; 3663 unsigned IgnoringBuilders = 0u; 3664 for (auto *SB : SpecializedBuilders) { 3665 if (!SB->isValid()) { 3666 ++InactiveBuilders; 3667 continue; 3668 } 3669 3670 auto RetCode = 3671 SB->getDeviceDependences(DDeps, CurPhase, FinalPhase, Phases); 3672 3673 // If the builder explicitly says the host action should be ignored, 3674 // we need to increment the variable that tracks the builders that request 3675 // the host object to be ignored. 3676 if (RetCode == DeviceActionBuilder::ABRT_Ignore_Host) 3677 ++IgnoringBuilders; 3678 3679 // Unless the builder was inactive for this action, we have to record the 3680 // offload kind because the host will have to use it. 3681 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3682 OffloadKind |= SB->getAssociatedOffloadKind(); 3683 } 3684 3685 // If all builders agree that the host object should be ignored, just return 3686 // nullptr. 3687 if (IgnoringBuilders && 3688 SpecializedBuilders.size() == (InactiveBuilders + IgnoringBuilders)) 3689 return nullptr; 3690 3691 if (DDeps.getActions().empty()) 3692 return HostAction; 3693 3694 // We have dependences we need to bundle together. We use an offload action 3695 // for that. 3696 OffloadAction::HostDependence HDep( 3697 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3698 /*BoundArch=*/nullptr, DDeps); 3699 return C.MakeAction<OffloadAction>(HDep, DDeps); 3700 } 3701 3702 /// Generate an action that adds a host dependence to a device action. The 3703 /// results will be kept in this action builder. Return true if an error was 3704 /// found. 3705 bool addHostDependenceToDeviceActions(Action *&HostAction, 3706 const Arg *InputArg) { 3707 if (!IsValid) 3708 return true; 3709 3710 recordHostAction(HostAction, InputArg); 3711 3712 // If we are supporting bundling/unbundling and the current action is an 3713 // input action of non-source file, we replace the host action by the 3714 // unbundling action. The bundler tool has the logic to detect if an input 3715 // is a bundle or not and if the input is not a bundle it assumes it is a 3716 // host file. Therefore it is safe to create an unbundling action even if 3717 // the input is not a bundle. 3718 if (CanUseBundler && isa<InputAction>(HostAction) && 3719 InputArg->getOption().getKind() == llvm::opt::Option::InputClass && 3720 (!types::isSrcFile(HostAction->getType()) || 3721 HostAction->getType() == types::TY_PP_HIP)) { 3722 auto UnbundlingHostAction = 3723 C.MakeAction<OffloadUnbundlingJobAction>(HostAction); 3724 UnbundlingHostAction->registerDependentActionInfo( 3725 C.getSingleOffloadToolChain<Action::OFK_Host>(), 3726 /*BoundArch=*/StringRef(), Action::OFK_Host); 3727 HostAction = UnbundlingHostAction; 3728 recordHostAction(HostAction, InputArg); 3729 } 3730 3731 assert(HostAction && "Invalid host action!"); 3732 3733 // Register the offload kinds that are used. 3734 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3735 for (auto *SB : SpecializedBuilders) { 3736 if (!SB->isValid()) 3737 continue; 3738 3739 auto RetCode = SB->addDeviceDepences(HostAction); 3740 3741 // Host dependences for device actions are not compatible with that same 3742 // action being ignored. 3743 assert(RetCode != DeviceActionBuilder::ABRT_Ignore_Host && 3744 "Host dependence not expected to be ignored.!"); 3745 3746 // Unless the builder was inactive for this action, we have to record the 3747 // offload kind because the host will have to use it. 3748 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3749 OffloadKind |= SB->getAssociatedOffloadKind(); 3750 } 3751 3752 // Do not use unbundler if the Host does not depend on device action. 3753 if (OffloadKind == Action::OFK_None && CanUseBundler) 3754 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) 3755 HostAction = UA->getInputs().back(); 3756 3757 return false; 3758 } 3759 3760 /// Add the offloading top level actions to the provided action list. This 3761 /// function can replace the host action by a bundling action if the 3762 /// programming models allow it. 3763 bool appendTopLevelActions(ActionList &AL, Action *HostAction, 3764 const Arg *InputArg) { 3765 if (HostAction) 3766 recordHostAction(HostAction, InputArg); 3767 3768 // Get the device actions to be appended. 3769 ActionList OffloadAL; 3770 for (auto *SB : SpecializedBuilders) { 3771 if (!SB->isValid()) 3772 continue; 3773 SB->appendTopLevelActions(OffloadAL); 3774 } 3775 3776 // If we can use the bundler, replace the host action by the bundling one in 3777 // the resulting list. Otherwise, just append the device actions. For 3778 // device only compilation, HostAction is a null pointer, therefore only do 3779 // this when HostAction is not a null pointer. 3780 if (CanUseBundler && HostAction && 3781 HostAction->getType() != types::TY_Nothing && !OffloadAL.empty()) { 3782 // Add the host action to the list in order to create the bundling action. 3783 OffloadAL.push_back(HostAction); 3784 3785 // We expect that the host action was just appended to the action list 3786 // before this method was called. 3787 assert(HostAction == AL.back() && "Host action not in the list??"); 3788 HostAction = C.MakeAction<OffloadBundlingJobAction>(OffloadAL); 3789 recordHostAction(HostAction, InputArg); 3790 AL.back() = HostAction; 3791 } else 3792 AL.append(OffloadAL.begin(), OffloadAL.end()); 3793 3794 // Propagate to the current host action (if any) the offload information 3795 // associated with the current input. 3796 if (HostAction) 3797 HostAction->propagateHostOffloadInfo(InputArgToOffloadKindMap[InputArg], 3798 /*BoundArch=*/nullptr); 3799 return false; 3800 } 3801 3802 void appendDeviceLinkActions(ActionList &AL) { 3803 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3804 if (!SB->isValid()) 3805 continue; 3806 SB->appendLinkDeviceActions(AL); 3807 } 3808 } 3809 3810 Action *makeHostLinkAction() { 3811 // Build a list of device linking actions. 3812 ActionList DeviceAL; 3813 appendDeviceLinkActions(DeviceAL); 3814 if (DeviceAL.empty()) 3815 return nullptr; 3816 3817 // Let builders add host linking actions. 3818 Action* HA = nullptr; 3819 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3820 if (!SB->isValid()) 3821 continue; 3822 HA = SB->appendLinkHostActions(DeviceAL); 3823 // This created host action has no originating input argument, therefore 3824 // needs to set its offloading kind directly. 3825 if (HA) 3826 HA->propagateHostOffloadInfo(SB->getAssociatedOffloadKind(), 3827 /*BoundArch=*/nullptr); 3828 } 3829 return HA; 3830 } 3831 3832 /// Processes the host linker action. This currently consists of replacing it 3833 /// with an offload action if there are device link objects and propagate to 3834 /// the host action all the offload kinds used in the current compilation. The 3835 /// resulting action is returned. 3836 Action *processHostLinkAction(Action *HostAction) { 3837 // Add all the dependences from the device linking actions. 3838 OffloadAction::DeviceDependences DDeps; 3839 for (auto *SB : SpecializedBuilders) { 3840 if (!SB->isValid()) 3841 continue; 3842 3843 SB->appendLinkDependences(DDeps); 3844 } 3845 3846 // Calculate all the offload kinds used in the current compilation. 3847 unsigned ActiveOffloadKinds = 0u; 3848 for (auto &I : InputArgToOffloadKindMap) 3849 ActiveOffloadKinds |= I.second; 3850 3851 // If we don't have device dependencies, we don't have to create an offload 3852 // action. 3853 if (DDeps.getActions().empty()) { 3854 // Set all the active offloading kinds to the link action. Given that it 3855 // is a link action it is assumed to depend on all actions generated so 3856 // far. 3857 HostAction->setHostOffloadInfo(ActiveOffloadKinds, 3858 /*BoundArch=*/nullptr); 3859 // Propagate active offloading kinds for each input to the link action. 3860 // Each input may have different active offloading kind. 3861 for (auto A : HostAction->inputs()) { 3862 auto ArgLoc = HostActionToInputArgMap.find(A); 3863 if (ArgLoc == HostActionToInputArgMap.end()) 3864 continue; 3865 auto OFKLoc = InputArgToOffloadKindMap.find(ArgLoc->second); 3866 if (OFKLoc == InputArgToOffloadKindMap.end()) 3867 continue; 3868 A->propagateHostOffloadInfo(OFKLoc->second, /*BoundArch=*/nullptr); 3869 } 3870 return HostAction; 3871 } 3872 3873 // Create the offload action with all dependences. When an offload action 3874 // is created the kinds are propagated to the host action, so we don't have 3875 // to do that explicitly here. 3876 OffloadAction::HostDependence HDep( 3877 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3878 /*BoundArch*/ nullptr, ActiveOffloadKinds); 3879 return C.MakeAction<OffloadAction>(HDep, DDeps); 3880 } 3881 }; 3882 } // anonymous namespace. 3883 3884 void Driver::handleArguments(Compilation &C, DerivedArgList &Args, 3885 const InputList &Inputs, 3886 ActionList &Actions) const { 3887 3888 // Ignore /Yc/Yu if both /Yc and /Yu passed but with different filenames. 3889 Arg *YcArg = Args.getLastArg(options::OPT__SLASH_Yc); 3890 Arg *YuArg = Args.getLastArg(options::OPT__SLASH_Yu); 3891 if (YcArg && YuArg && strcmp(YcArg->getValue(), YuArg->getValue()) != 0) { 3892 Diag(clang::diag::warn_drv_ycyu_different_arg_clang_cl); 3893 Args.eraseArg(options::OPT__SLASH_Yc); 3894 Args.eraseArg(options::OPT__SLASH_Yu); 3895 YcArg = YuArg = nullptr; 3896 } 3897 if (YcArg && Inputs.size() > 1) { 3898 Diag(clang::diag::warn_drv_yc_multiple_inputs_clang_cl); 3899 Args.eraseArg(options::OPT__SLASH_Yc); 3900 YcArg = nullptr; 3901 } 3902 3903 Arg *FinalPhaseArg; 3904 phases::ID FinalPhase = getFinalPhase(Args, &FinalPhaseArg); 3905 3906 if (FinalPhase == phases::Link) { 3907 // Emitting LLVM while linking disabled except in HIPAMD Toolchain 3908 if (Args.hasArg(options::OPT_emit_llvm) && !Args.hasArg(options::OPT_hip_link)) 3909 Diag(clang::diag::err_drv_emit_llvm_link); 3910 if (IsCLMode() && LTOMode != LTOK_None && 3911 !Args.getLastArgValue(options::OPT_fuse_ld_EQ) 3912 .equals_insensitive("lld")) 3913 Diag(clang::diag::err_drv_lto_without_lld); 3914 } 3915 3916 if (FinalPhase == phases::Preprocess || Args.hasArg(options::OPT__SLASH_Y_)) { 3917 // If only preprocessing or /Y- is used, all pch handling is disabled. 3918 // Rather than check for it everywhere, just remove clang-cl pch-related 3919 // flags here. 3920 Args.eraseArg(options::OPT__SLASH_Fp); 3921 Args.eraseArg(options::OPT__SLASH_Yc); 3922 Args.eraseArg(options::OPT__SLASH_Yu); 3923 YcArg = YuArg = nullptr; 3924 } 3925 3926 unsigned LastPLSize = 0; 3927 for (auto &I : Inputs) { 3928 types::ID InputType = I.first; 3929 const Arg *InputArg = I.second; 3930 3931 auto PL = types::getCompilationPhases(InputType); 3932 LastPLSize = PL.size(); 3933 3934 // If the first step comes after the final phase we are doing as part of 3935 // this compilation, warn the user about it. 3936 phases::ID InitialPhase = PL[0]; 3937 if (InitialPhase > FinalPhase) { 3938 if (InputArg->isClaimed()) 3939 continue; 3940 3941 // Claim here to avoid the more general unused warning. 3942 InputArg->claim(); 3943 3944 // Suppress all unused style warnings with -Qunused-arguments 3945 if (Args.hasArg(options::OPT_Qunused_arguments)) 3946 continue; 3947 3948 // Special case when final phase determined by binary name, rather than 3949 // by a command-line argument with a corresponding Arg. 3950 if (CCCIsCPP()) 3951 Diag(clang::diag::warn_drv_input_file_unused_by_cpp) 3952 << InputArg->getAsString(Args) << getPhaseName(InitialPhase); 3953 // Special case '-E' warning on a previously preprocessed file to make 3954 // more sense. 3955 else if (InitialPhase == phases::Compile && 3956 (Args.getLastArg(options::OPT__SLASH_EP, 3957 options::OPT__SLASH_P) || 3958 Args.getLastArg(options::OPT_E) || 3959 Args.getLastArg(options::OPT_M, options::OPT_MM)) && 3960 getPreprocessedType(InputType) == types::TY_INVALID) 3961 Diag(clang::diag::warn_drv_preprocessed_input_file_unused) 3962 << InputArg->getAsString(Args) << !!FinalPhaseArg 3963 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3964 else 3965 Diag(clang::diag::warn_drv_input_file_unused) 3966 << InputArg->getAsString(Args) << getPhaseName(InitialPhase) 3967 << !!FinalPhaseArg 3968 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3969 continue; 3970 } 3971 3972 if (YcArg) { 3973 // Add a separate precompile phase for the compile phase. 3974 if (FinalPhase >= phases::Compile) { 3975 const types::ID HeaderType = lookupHeaderTypeForSourceType(InputType); 3976 // Build the pipeline for the pch file. 3977 Action *ClangClPch = C.MakeAction<InputAction>(*InputArg, HeaderType); 3978 for (phases::ID Phase : types::getCompilationPhases(HeaderType)) 3979 ClangClPch = ConstructPhaseAction(C, Args, Phase, ClangClPch); 3980 assert(ClangClPch); 3981 Actions.push_back(ClangClPch); 3982 // The driver currently exits after the first failed command. This 3983 // relies on that behavior, to make sure if the pch generation fails, 3984 // the main compilation won't run. 3985 // FIXME: If the main compilation fails, the PCH generation should 3986 // probably not be considered successful either. 3987 } 3988 } 3989 } 3990 3991 // If we are linking, claim any options which are obviously only used for 3992 // compilation. 3993 // FIXME: Understand why the last Phase List length is used here. 3994 if (FinalPhase == phases::Link && LastPLSize == 1) { 3995 Args.ClaimAllArgs(options::OPT_CompileOnly_Group); 3996 Args.ClaimAllArgs(options::OPT_cl_compile_Group); 3997 } 3998 } 3999 4000 void Driver::BuildActions(Compilation &C, DerivedArgList &Args, 4001 const InputList &Inputs, ActionList &Actions) const { 4002 llvm::PrettyStackTraceString CrashInfo("Building compilation actions"); 4003 4004 if (!SuppressMissingInputWarning && Inputs.empty()) { 4005 Diag(clang::diag::err_drv_no_input_files); 4006 return; 4007 } 4008 4009 // Reject -Z* at the top level, these options should never have been exposed 4010 // by gcc. 4011 if (Arg *A = Args.getLastArg(options::OPT_Z_Joined)) 4012 Diag(clang::diag::err_drv_use_of_Z_option) << A->getAsString(Args); 4013 4014 // Diagnose misuse of /Fo. 4015 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fo)) { 4016 StringRef V = A->getValue(); 4017 if (Inputs.size() > 1 && !V.empty() && 4018 !llvm::sys::path::is_separator(V.back())) { 4019 // Check whether /Fo tries to name an output file for multiple inputs. 4020 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4021 << A->getSpelling() << V; 4022 Args.eraseArg(options::OPT__SLASH_Fo); 4023 } 4024 } 4025 4026 // Diagnose misuse of /Fa. 4027 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fa)) { 4028 StringRef V = A->getValue(); 4029 if (Inputs.size() > 1 && !V.empty() && 4030 !llvm::sys::path::is_separator(V.back())) { 4031 // Check whether /Fa tries to name an asm file for multiple inputs. 4032 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4033 << A->getSpelling() << V; 4034 Args.eraseArg(options::OPT__SLASH_Fa); 4035 } 4036 } 4037 4038 // Diagnose misuse of /o. 4039 if (Arg *A = Args.getLastArg(options::OPT__SLASH_o)) { 4040 if (A->getValue()[0] == '\0') { 4041 // It has to have a value. 4042 Diag(clang::diag::err_drv_missing_argument) << A->getSpelling() << 1; 4043 Args.eraseArg(options::OPT__SLASH_o); 4044 } 4045 } 4046 4047 handleArguments(C, Args, Inputs, Actions); 4048 4049 // Builder to be used to build offloading actions. 4050 OffloadingActionBuilder OffloadBuilder(C, Args, Inputs); 4051 4052 bool UseNewOffloadingDriver = 4053 (C.isOffloadingHostKind(Action::OFK_OpenMP) && 4054 Args.hasFlag(options::OPT_fopenmp_new_driver, 4055 options::OPT_no_offload_new_driver, true)) || 4056 Args.hasFlag(options::OPT_offload_new_driver, 4057 options::OPT_no_offload_new_driver, false); 4058 4059 // Construct the actions to perform. 4060 HeaderModulePrecompileJobAction *HeaderModuleAction = nullptr; 4061 ExtractAPIJobAction *ExtractAPIAction = nullptr; 4062 ActionList LinkerInputs; 4063 ActionList MergerInputs; 4064 4065 for (auto &I : Inputs) { 4066 types::ID InputType = I.first; 4067 const Arg *InputArg = I.second; 4068 4069 auto PL = types::getCompilationPhases(*this, Args, InputType); 4070 if (PL.empty()) 4071 continue; 4072 4073 auto FullPL = types::getCompilationPhases(InputType); 4074 4075 // Build the pipeline for this file. 4076 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4077 4078 // Use the current host action in any of the offloading actions, if 4079 // required. 4080 if (!UseNewOffloadingDriver) 4081 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4082 break; 4083 4084 for (phases::ID Phase : PL) { 4085 4086 // Add any offload action the host action depends on. 4087 if (!UseNewOffloadingDriver) 4088 Current = OffloadBuilder.addDeviceDependencesToHostAction( 4089 Current, InputArg, Phase, PL.back(), FullPL); 4090 if (!Current) 4091 break; 4092 4093 // Queue linker inputs. 4094 if (Phase == phases::Link) { 4095 assert(Phase == PL.back() && "linking must be final compilation step."); 4096 // We don't need to generate additional link commands if emitting AMD bitcode 4097 if (!(C.getInputArgs().hasArg(options::OPT_hip_link) && 4098 (C.getInputArgs().hasArg(options::OPT_emit_llvm)))) 4099 LinkerInputs.push_back(Current); 4100 Current = nullptr; 4101 break; 4102 } 4103 4104 // TODO: Consider removing this because the merged may not end up being 4105 // the final Phase in the pipeline. Perhaps the merged could just merge 4106 // and then pass an artifact of some sort to the Link Phase. 4107 // Queue merger inputs. 4108 if (Phase == phases::IfsMerge) { 4109 assert(Phase == PL.back() && "merging must be final compilation step."); 4110 MergerInputs.push_back(Current); 4111 Current = nullptr; 4112 break; 4113 } 4114 4115 // Each precompiled header file after a module file action is a module 4116 // header of that same module file, rather than being compiled to a 4117 // separate PCH. 4118 if (Phase == phases::Precompile && HeaderModuleAction && 4119 getPrecompiledType(InputType) == types::TY_PCH) { 4120 HeaderModuleAction->addModuleHeaderInput(Current); 4121 Current = nullptr; 4122 break; 4123 } 4124 4125 if (Phase == phases::Precompile && ExtractAPIAction) { 4126 ExtractAPIAction->addHeaderInput(Current); 4127 Current = nullptr; 4128 break; 4129 } 4130 4131 // FIXME: Should we include any prior module file outputs as inputs of 4132 // later actions in the same command line? 4133 4134 // Otherwise construct the appropriate action. 4135 Action *NewCurrent = ConstructPhaseAction(C, Args, Phase, Current); 4136 4137 // We didn't create a new action, so we will just move to the next phase. 4138 if (NewCurrent == Current) 4139 continue; 4140 4141 if (auto *HMA = dyn_cast<HeaderModulePrecompileJobAction>(NewCurrent)) 4142 HeaderModuleAction = HMA; 4143 else if (auto *EAA = dyn_cast<ExtractAPIJobAction>(NewCurrent)) 4144 ExtractAPIAction = EAA; 4145 4146 Current = NewCurrent; 4147 4148 // Use the current host action in any of the offloading actions, if 4149 // required. 4150 if (!UseNewOffloadingDriver) 4151 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4152 break; 4153 4154 // Try to build the offloading actions and add the result as a dependency 4155 // to the host. 4156 if (UseNewOffloadingDriver) 4157 Current = BuildOffloadingActions(C, Args, I, Current); 4158 4159 if (Current->getType() == types::TY_Nothing) 4160 break; 4161 } 4162 4163 // If we ended with something, add to the output list. 4164 if (Current) 4165 Actions.push_back(Current); 4166 4167 // Add any top level actions generated for offloading. 4168 if (!UseNewOffloadingDriver) 4169 OffloadBuilder.appendTopLevelActions(Actions, Current, InputArg); 4170 else if (Current) 4171 Current->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4172 /*BoundArch=*/nullptr); 4173 } 4174 4175 // Add a link action if necessary. 4176 4177 if (LinkerInputs.empty()) { 4178 Arg *FinalPhaseArg; 4179 if (getFinalPhase(Args, &FinalPhaseArg) == phases::Link) 4180 if (!UseNewOffloadingDriver) 4181 OffloadBuilder.appendDeviceLinkActions(Actions); 4182 } 4183 4184 if (!LinkerInputs.empty()) { 4185 if (!UseNewOffloadingDriver) 4186 if (Action *Wrapper = OffloadBuilder.makeHostLinkAction()) 4187 LinkerInputs.push_back(Wrapper); 4188 Action *LA; 4189 // Check if this Linker Job should emit a static library. 4190 if (ShouldEmitStaticLibrary(Args)) { 4191 LA = C.MakeAction<StaticLibJobAction>(LinkerInputs, types::TY_Image); 4192 } else if (UseNewOffloadingDriver || 4193 Args.hasArg(options::OPT_offload_link)) { 4194 LA = C.MakeAction<LinkerWrapperJobAction>(LinkerInputs, types::TY_Image); 4195 LA->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4196 /*BoundArch=*/nullptr); 4197 } else { 4198 LA = C.MakeAction<LinkJobAction>(LinkerInputs, types::TY_Image); 4199 } 4200 if (!UseNewOffloadingDriver) 4201 LA = OffloadBuilder.processHostLinkAction(LA); 4202 Actions.push_back(LA); 4203 } 4204 4205 // Add an interface stubs merge action if necessary. 4206 if (!MergerInputs.empty()) 4207 Actions.push_back( 4208 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4209 4210 if (Args.hasArg(options::OPT_emit_interface_stubs)) { 4211 auto PhaseList = types::getCompilationPhases( 4212 types::TY_IFS_CPP, 4213 Args.hasArg(options::OPT_c) ? phases::Compile : phases::IfsMerge); 4214 4215 ActionList MergerInputs; 4216 4217 for (auto &I : Inputs) { 4218 types::ID InputType = I.first; 4219 const Arg *InputArg = I.second; 4220 4221 // Currently clang and the llvm assembler do not support generating symbol 4222 // stubs from assembly, so we skip the input on asm files. For ifs files 4223 // we rely on the normal pipeline setup in the pipeline setup code above. 4224 if (InputType == types::TY_IFS || InputType == types::TY_PP_Asm || 4225 InputType == types::TY_Asm) 4226 continue; 4227 4228 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4229 4230 for (auto Phase : PhaseList) { 4231 switch (Phase) { 4232 default: 4233 llvm_unreachable( 4234 "IFS Pipeline can only consist of Compile followed by IfsMerge."); 4235 case phases::Compile: { 4236 // Only IfsMerge (llvm-ifs) can handle .o files by looking for ifs 4237 // files where the .o file is located. The compile action can not 4238 // handle this. 4239 if (InputType == types::TY_Object) 4240 break; 4241 4242 Current = C.MakeAction<CompileJobAction>(Current, types::TY_IFS_CPP); 4243 break; 4244 } 4245 case phases::IfsMerge: { 4246 assert(Phase == PhaseList.back() && 4247 "merging must be final compilation step."); 4248 MergerInputs.push_back(Current); 4249 Current = nullptr; 4250 break; 4251 } 4252 } 4253 } 4254 4255 // If we ended with something, add to the output list. 4256 if (Current) 4257 Actions.push_back(Current); 4258 } 4259 4260 // Add an interface stubs merge action if necessary. 4261 if (!MergerInputs.empty()) 4262 Actions.push_back( 4263 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4264 } 4265 4266 // If --print-supported-cpus, -mcpu=? or -mtune=? is specified, build a custom 4267 // Compile phase that prints out supported cpu models and quits. 4268 if (Arg *A = Args.getLastArg(options::OPT_print_supported_cpus)) { 4269 // Use the -mcpu=? flag as the dummy input to cc1. 4270 Actions.clear(); 4271 Action *InputAc = C.MakeAction<InputAction>(*A, types::TY_C); 4272 Actions.push_back( 4273 C.MakeAction<PrecompileJobAction>(InputAc, types::TY_Nothing)); 4274 for (auto &I : Inputs) 4275 I.second->claim(); 4276 } 4277 4278 // Claim ignored clang-cl options. 4279 Args.ClaimAllArgs(options::OPT_cl_ignored_Group); 4280 } 4281 4282 /// Returns the canonical name for the offloading architecture when using a HIP 4283 /// or CUDA architecture. 4284 static StringRef getCanonicalArchString(Compilation &C, 4285 const llvm::opt::DerivedArgList &Args, 4286 StringRef ArchStr, 4287 const llvm::Triple &Triple) { 4288 // Lookup the CUDA / HIP architecture string. Only report an error if we were 4289 // expecting the triple to be only NVPTX / AMDGPU. 4290 CudaArch Arch = StringToCudaArch(getProcessorFromTargetID(Triple, ArchStr)); 4291 if (Triple.isNVPTX() && 4292 (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch))) { 4293 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4294 << "CUDA" << ArchStr; 4295 return StringRef(); 4296 } else if (Triple.isAMDGPU() && 4297 (Arch == CudaArch::UNKNOWN || !IsAMDGpuArch(Arch))) { 4298 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4299 << "HIP" << ArchStr; 4300 return StringRef(); 4301 } 4302 4303 if (IsNVIDIAGpuArch(Arch)) 4304 return Args.MakeArgStringRef(CudaArchToString(Arch)); 4305 4306 if (IsAMDGpuArch(Arch)) { 4307 llvm::StringMap<bool> Features; 4308 auto HIPTriple = getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()); 4309 if (!HIPTriple) 4310 return StringRef(); 4311 auto Arch = parseTargetID(*HIPTriple, ArchStr, &Features); 4312 if (!Arch) { 4313 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << ArchStr; 4314 C.setContainsError(); 4315 return StringRef(); 4316 } 4317 return Args.MakeArgStringRef(getCanonicalTargetID(*Arch, Features)); 4318 } 4319 4320 // If the input isn't CUDA or HIP just return the architecture. 4321 return ArchStr; 4322 } 4323 4324 /// Checks if the set offloading architectures does not conflict. Returns the 4325 /// incompatible pair if a conflict occurs. 4326 static llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 4327 getConflictOffloadArchCombination(const llvm::DenseSet<StringRef> &Archs, 4328 Action::OffloadKind Kind) { 4329 if (Kind != Action::OFK_HIP) 4330 return None; 4331 4332 std::set<StringRef> ArchSet; 4333 llvm::copy(Archs, std::inserter(ArchSet, ArchSet.begin())); 4334 return getConflictTargetIDCombination(ArchSet); 4335 } 4336 4337 llvm::DenseSet<StringRef> 4338 Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args, 4339 Action::OffloadKind Kind, const ToolChain *TC) const { 4340 if (!TC) 4341 TC = &C.getDefaultToolChain(); 4342 4343 // --offload and --offload-arch options are mutually exclusive. 4344 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 4345 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 4346 options::OPT_no_offload_arch_EQ)) { 4347 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) 4348 << "--offload" 4349 << (Args.hasArgNoClaim(options::OPT_offload_arch_EQ) 4350 ? "--offload-arch" 4351 : "--no-offload-arch"); 4352 } 4353 4354 if (KnownArchs.find(TC) != KnownArchs.end()) 4355 return KnownArchs.lookup(TC); 4356 4357 llvm::DenseSet<StringRef> Archs; 4358 for (auto *Arg : Args) { 4359 // Extract any '--[no-]offload-arch' arguments intended for this toolchain. 4360 std::unique_ptr<llvm::opt::Arg> ExtractedArg = nullptr; 4361 if (Arg->getOption().matches(options::OPT_Xopenmp_target_EQ) && 4362 ToolChain::getOpenMPTriple(Arg->getValue(0)) == TC->getTriple()) { 4363 Arg->claim(); 4364 unsigned Index = Args.getBaseArgs().MakeIndex(Arg->getValue(1)); 4365 ExtractedArg = getOpts().ParseOneArg(Args, Index); 4366 Arg = ExtractedArg.get(); 4367 } 4368 4369 if (Arg->getOption().matches(options::OPT_offload_arch_EQ)) { 4370 for (StringRef Arch : llvm::split(Arg->getValue(), ",")) 4371 Archs.insert(getCanonicalArchString(C, Args, Arch, TC->getTriple())); 4372 } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) { 4373 for (StringRef Arch : llvm::split(Arg->getValue(), ",")) { 4374 if (Arch == StringRef("all")) 4375 Archs.clear(); 4376 else 4377 Archs.erase(getCanonicalArchString(C, Args, Arch, TC->getTriple())); 4378 } 4379 } 4380 } 4381 4382 if (auto ConflictingArchs = getConflictOffloadArchCombination(Archs, Kind)) { 4383 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 4384 << ConflictingArchs->first << ConflictingArchs->second; 4385 C.setContainsError(); 4386 } 4387 4388 if (Archs.empty()) { 4389 if (Kind == Action::OFK_Cuda) 4390 Archs.insert(CudaArchToString(CudaArch::CudaDefault)); 4391 else if (Kind == Action::OFK_HIP) 4392 Archs.insert(CudaArchToString(CudaArch::HIPDefault)); 4393 else if (Kind == Action::OFK_OpenMP) 4394 Archs.insert(StringRef()); 4395 } else { 4396 Args.ClaimAllArgs(options::OPT_offload_arch_EQ); 4397 Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ); 4398 } 4399 4400 return Archs; 4401 } 4402 4403 Action *Driver::BuildOffloadingActions(Compilation &C, 4404 llvm::opt::DerivedArgList &Args, 4405 const InputTy &Input, 4406 Action *HostAction) const { 4407 // Don't build offloading actions if explicitly disabled or we do not have a 4408 // valid source input and compile action to embed it in. If preprocessing only 4409 // ignore embedding. 4410 if (offloadHostOnly() || !types::isSrcFile(Input.first) || 4411 !(isa<CompileJobAction>(HostAction) || 4412 getFinalPhase(Args) == phases::Preprocess)) 4413 return HostAction; 4414 4415 ActionList OffloadActions; 4416 OffloadAction::DeviceDependences DDeps; 4417 4418 const Action::OffloadKind OffloadKinds[] = { 4419 Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP}; 4420 4421 for (Action::OffloadKind Kind : OffloadKinds) { 4422 SmallVector<const ToolChain *, 2> ToolChains; 4423 ActionList DeviceActions; 4424 4425 auto TCRange = C.getOffloadToolChains(Kind); 4426 for (auto TI = TCRange.first, TE = TCRange.second; TI != TE; ++TI) 4427 ToolChains.push_back(TI->second); 4428 4429 if (ToolChains.empty()) 4430 continue; 4431 4432 types::ID InputType = Input.first; 4433 const Arg *InputArg = Input.second; 4434 4435 // The toolchain can be active for unsupported file types. 4436 if ((Kind == Action::OFK_Cuda && !types::isCuda(InputType)) || 4437 (Kind == Action::OFK_HIP && !types::isHIP(InputType))) 4438 continue; 4439 4440 // Get the product of all bound architectures and toolchains. 4441 SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs; 4442 for (const ToolChain *TC : ToolChains) 4443 for (StringRef Arch : getOffloadArchs(C, Args, Kind, TC)) 4444 TCAndArchs.push_back(std::make_pair(TC, Arch)); 4445 4446 for (unsigned I = 0, E = TCAndArchs.size(); I != E; ++I) 4447 DeviceActions.push_back(C.MakeAction<InputAction>(*InputArg, InputType)); 4448 4449 if (DeviceActions.empty()) 4450 return HostAction; 4451 4452 auto PL = types::getCompilationPhases(*this, Args, InputType); 4453 4454 for (phases::ID Phase : PL) { 4455 if (Phase == phases::Link) { 4456 assert(Phase == PL.back() && "linking must be final compilation step."); 4457 break; 4458 } 4459 4460 auto TCAndArch = TCAndArchs.begin(); 4461 for (Action *&A : DeviceActions) { 4462 A = ConstructPhaseAction(C, Args, Phase, A, Kind); 4463 4464 if (isa<CompileJobAction>(A) && isa<CompileJobAction>(HostAction) && 4465 Kind == Action::OFK_OpenMP) { 4466 // OpenMP offloading has a dependency on the host compile action to 4467 // identify which declarations need to be emitted. This shouldn't be 4468 // collapsed with any other actions so we can use it in the device. 4469 HostAction->setCannotBeCollapsedWithNextDependentAction(); 4470 OffloadAction::HostDependence HDep( 4471 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4472 TCAndArch->second.data(), Kind); 4473 OffloadAction::DeviceDependences DDep; 4474 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4475 A = C.MakeAction<OffloadAction>(HDep, DDep); 4476 } 4477 ++TCAndArch; 4478 } 4479 } 4480 4481 // Compiling HIP in non-RDC mode requires linking each action individually. 4482 for (Action *&A : DeviceActions) { 4483 if (A->getType() != types::TY_Object || Kind != Action::OFK_HIP || 4484 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false)) 4485 continue; 4486 ActionList LinkerInput = {A}; 4487 A = C.MakeAction<LinkJobAction>(LinkerInput, types::TY_Image); 4488 } 4489 4490 auto TCAndArch = TCAndArchs.begin(); 4491 for (Action *A : DeviceActions) { 4492 DDeps.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4493 OffloadAction::DeviceDependences DDep; 4494 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4495 OffloadActions.push_back(C.MakeAction<OffloadAction>(DDep, A->getType())); 4496 ++TCAndArch; 4497 } 4498 } 4499 4500 if (offloadDeviceOnly()) 4501 return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing); 4502 4503 if (OffloadActions.empty()) 4504 return HostAction; 4505 4506 OffloadAction::DeviceDependences DDep; 4507 if (C.isOffloadingHostKind(Action::OFK_Cuda) && 4508 !Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false)) { 4509 // If we are not in RDC-mode we just emit the final CUDA fatbinary for 4510 // each translation unit without requiring any linking. 4511 Action *FatbinAction = 4512 C.MakeAction<LinkJobAction>(OffloadActions, types::TY_CUDA_FATBIN); 4513 DDep.add(*FatbinAction, *C.getSingleOffloadToolChain<Action::OFK_Cuda>(), 4514 nullptr, Action::OFK_Cuda); 4515 } else if (C.isOffloadingHostKind(Action::OFK_HIP) && 4516 !Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4517 false)) { 4518 // If we are not in RDC-mode we just emit the final HIP fatbinary for each 4519 // translation unit, linking each input individually. 4520 Action *FatbinAction = 4521 C.MakeAction<LinkJobAction>(OffloadActions, types::TY_HIP_FATBIN); 4522 DDep.add(*FatbinAction, *C.getSingleOffloadToolChain<Action::OFK_HIP>(), 4523 nullptr, Action::OFK_HIP); 4524 } else { 4525 // Package all the offloading actions into a single output that can be 4526 // embedded in the host and linked. 4527 Action *PackagerAction = 4528 C.MakeAction<OffloadPackagerJobAction>(OffloadActions, types::TY_Image); 4529 DDep.add(*PackagerAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4530 nullptr, Action::OFK_None); 4531 } 4532 4533 OffloadAction::HostDependence HDep( 4534 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4535 /*BoundArch=*/nullptr, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4536 return C.MakeAction<OffloadAction>( 4537 HDep, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4538 } 4539 4540 Action *Driver::ConstructPhaseAction( 4541 Compilation &C, const ArgList &Args, phases::ID Phase, Action *Input, 4542 Action::OffloadKind TargetDeviceOffloadKind) const { 4543 llvm::PrettyStackTraceString CrashInfo("Constructing phase actions"); 4544 4545 // Some types skip the assembler phase (e.g., llvm-bc), but we can't 4546 // encode this in the steps because the intermediate type depends on 4547 // arguments. Just special case here. 4548 if (Phase == phases::Assemble && Input->getType() != types::TY_PP_Asm) 4549 return Input; 4550 4551 // Build the appropriate action. 4552 switch (Phase) { 4553 case phases::Link: 4554 llvm_unreachable("link action invalid here."); 4555 case phases::IfsMerge: 4556 llvm_unreachable("ifsmerge action invalid here."); 4557 case phases::Preprocess: { 4558 types::ID OutputTy; 4559 // -M and -MM specify the dependency file name by altering the output type, 4560 // -if -MD and -MMD are not specified. 4561 if (Args.hasArg(options::OPT_M, options::OPT_MM) && 4562 !Args.hasArg(options::OPT_MD, options::OPT_MMD)) { 4563 OutputTy = types::TY_Dependencies; 4564 } else { 4565 OutputTy = Input->getType(); 4566 // For these cases, the preprocessor is only translating forms, the Output 4567 // still needs preprocessing. 4568 if (!Args.hasFlag(options::OPT_frewrite_includes, 4569 options::OPT_fno_rewrite_includes, false) && 4570 !Args.hasFlag(options::OPT_frewrite_imports, 4571 options::OPT_fno_rewrite_imports, false) && 4572 !Args.hasFlag(options::OPT_fdirectives_only, 4573 options::OPT_fno_directives_only, false) && 4574 !CCGenDiagnostics) 4575 OutputTy = types::getPreprocessedType(OutputTy); 4576 assert(OutputTy != types::TY_INVALID && 4577 "Cannot preprocess this input type!"); 4578 } 4579 return C.MakeAction<PreprocessJobAction>(Input, OutputTy); 4580 } 4581 case phases::Precompile: { 4582 // API extraction should not generate an actual precompilation action. 4583 if (Args.hasArg(options::OPT_extract_api)) 4584 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4585 4586 types::ID OutputTy = getPrecompiledType(Input->getType()); 4587 assert(OutputTy != types::TY_INVALID && 4588 "Cannot precompile this input type!"); 4589 4590 // If we're given a module name, precompile header file inputs as a 4591 // module, not as a precompiled header. 4592 const char *ModName = nullptr; 4593 if (OutputTy == types::TY_PCH) { 4594 if (Arg *A = Args.getLastArg(options::OPT_fmodule_name_EQ)) 4595 ModName = A->getValue(); 4596 if (ModName) 4597 OutputTy = types::TY_ModuleFile; 4598 } 4599 4600 if (Args.hasArg(options::OPT_fsyntax_only)) { 4601 // Syntax checks should not emit a PCH file 4602 OutputTy = types::TY_Nothing; 4603 } 4604 4605 if (ModName) 4606 return C.MakeAction<HeaderModulePrecompileJobAction>(Input, OutputTy, 4607 ModName); 4608 return C.MakeAction<PrecompileJobAction>(Input, OutputTy); 4609 } 4610 case phases::Compile: { 4611 if (Args.hasArg(options::OPT_fsyntax_only)) 4612 return C.MakeAction<CompileJobAction>(Input, types::TY_Nothing); 4613 if (Args.hasArg(options::OPT_rewrite_objc)) 4614 return C.MakeAction<CompileJobAction>(Input, types::TY_RewrittenObjC); 4615 if (Args.hasArg(options::OPT_rewrite_legacy_objc)) 4616 return C.MakeAction<CompileJobAction>(Input, 4617 types::TY_RewrittenLegacyObjC); 4618 if (Args.hasArg(options::OPT__analyze)) 4619 return C.MakeAction<AnalyzeJobAction>(Input, types::TY_Plist); 4620 if (Args.hasArg(options::OPT__migrate)) 4621 return C.MakeAction<MigrateJobAction>(Input, types::TY_Remap); 4622 if (Args.hasArg(options::OPT_emit_ast)) 4623 return C.MakeAction<CompileJobAction>(Input, types::TY_AST); 4624 if (Args.hasArg(options::OPT_module_file_info)) 4625 return C.MakeAction<CompileJobAction>(Input, types::TY_ModuleFile); 4626 if (Args.hasArg(options::OPT_verify_pch)) 4627 return C.MakeAction<VerifyPCHJobAction>(Input, types::TY_Nothing); 4628 if (Args.hasArg(options::OPT_extract_api)) 4629 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4630 return C.MakeAction<CompileJobAction>(Input, types::TY_LLVM_BC); 4631 } 4632 case phases::Backend: { 4633 if (isUsingLTO() && TargetDeviceOffloadKind == Action::OFK_None) { 4634 types::ID Output = 4635 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4636 return C.MakeAction<BackendJobAction>(Input, Output); 4637 } 4638 if (isUsingLTO(/* IsOffload */ true) && 4639 TargetDeviceOffloadKind != Action::OFK_None) { 4640 types::ID Output = 4641 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4642 return C.MakeAction<BackendJobAction>(Input, Output); 4643 } 4644 if (Args.hasArg(options::OPT_emit_llvm) || 4645 (TargetDeviceOffloadKind == Action::OFK_HIP && 4646 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4647 false))) { 4648 types::ID Output = 4649 Args.hasArg(options::OPT_S) ? types::TY_LLVM_IR : types::TY_LLVM_BC; 4650 return C.MakeAction<BackendJobAction>(Input, Output); 4651 } 4652 return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm); 4653 } 4654 case phases::Assemble: 4655 return C.MakeAction<AssembleJobAction>(std::move(Input), types::TY_Object); 4656 } 4657 4658 llvm_unreachable("invalid phase in ConstructPhaseAction"); 4659 } 4660 4661 void Driver::BuildJobs(Compilation &C) const { 4662 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 4663 4664 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 4665 4666 // It is an error to provide a -o option if we are making multiple output 4667 // files. There are exceptions: 4668 // 4669 // IfsMergeJob: when generating interface stubs enabled we want to be able to 4670 // generate the stub file at the same time that we generate the real 4671 // library/a.out. So when a .o, .so, etc are the output, with clang interface 4672 // stubs there will also be a .ifs and .ifso at the same location. 4673 // 4674 // CompileJob of type TY_IFS_CPP: when generating interface stubs is enabled 4675 // and -c is passed, we still want to be able to generate a .ifs file while 4676 // we are also generating .o files. So we allow more than one output file in 4677 // this case as well. 4678 // 4679 if (FinalOutput) { 4680 unsigned NumOutputs = 0; 4681 unsigned NumIfsOutputs = 0; 4682 for (const Action *A : C.getActions()) 4683 if (A->getType() != types::TY_Nothing && 4684 !(A->getKind() == Action::IfsMergeJobClass || 4685 (A->getType() == clang::driver::types::TY_IFS_CPP && 4686 A->getKind() == clang::driver::Action::CompileJobClass && 4687 0 == NumIfsOutputs++) || 4688 (A->getKind() == Action::BindArchClass && A->getInputs().size() && 4689 A->getInputs().front()->getKind() == Action::IfsMergeJobClass))) 4690 ++NumOutputs; 4691 4692 if (NumOutputs > 1) { 4693 Diag(clang::diag::err_drv_output_argument_with_multiple_files); 4694 FinalOutput = nullptr; 4695 } 4696 } 4697 4698 const llvm::Triple &RawTriple = C.getDefaultToolChain().getTriple(); 4699 if (RawTriple.isOSAIX()) { 4700 if (Arg *A = C.getArgs().getLastArg(options::OPT_G)) 4701 Diag(diag::err_drv_unsupported_opt_for_target) 4702 << A->getSpelling() << RawTriple.str(); 4703 if (LTOMode == LTOK_Thin) 4704 Diag(diag::err_drv_clang_unsupported) << "thinLTO on AIX"; 4705 } 4706 4707 // Collect the list of architectures. 4708 llvm::StringSet<> ArchNames; 4709 if (RawTriple.isOSBinFormatMachO()) 4710 for (const Arg *A : C.getArgs()) 4711 if (A->getOption().matches(options::OPT_arch)) 4712 ArchNames.insert(A->getValue()); 4713 4714 // Set of (Action, canonical ToolChain triple) pairs we've built jobs for. 4715 std::map<std::pair<const Action *, std::string>, InputInfoList> CachedResults; 4716 for (Action *A : C.getActions()) { 4717 // If we are linking an image for multiple archs then the linker wants 4718 // -arch_multiple and -final_output <final image name>. Unfortunately, this 4719 // doesn't fit in cleanly because we have to pass this information down. 4720 // 4721 // FIXME: This is a hack; find a cleaner way to integrate this into the 4722 // process. 4723 const char *LinkingOutput = nullptr; 4724 if (isa<LipoJobAction>(A)) { 4725 if (FinalOutput) 4726 LinkingOutput = FinalOutput->getValue(); 4727 else 4728 LinkingOutput = getDefaultImageName(); 4729 } 4730 4731 BuildJobsForAction(C, A, &C.getDefaultToolChain(), 4732 /*BoundArch*/ StringRef(), 4733 /*AtTopLevel*/ true, 4734 /*MultipleArchs*/ ArchNames.size() > 1, 4735 /*LinkingOutput*/ LinkingOutput, CachedResults, 4736 /*TargetDeviceOffloadKind*/ Action::OFK_None); 4737 } 4738 4739 // If we have more than one job, then disable integrated-cc1 for now. Do this 4740 // also when we need to report process execution statistics. 4741 if (C.getJobs().size() > 1 || CCPrintProcessStats) 4742 for (auto &J : C.getJobs()) 4743 J.InProcess = false; 4744 4745 if (CCPrintProcessStats) { 4746 C.setPostCallback([=](const Command &Cmd, int Res) { 4747 Optional<llvm::sys::ProcessStatistics> ProcStat = 4748 Cmd.getProcessStatistics(); 4749 if (!ProcStat) 4750 return; 4751 4752 const char *LinkingOutput = nullptr; 4753 if (FinalOutput) 4754 LinkingOutput = FinalOutput->getValue(); 4755 else if (!Cmd.getOutputFilenames().empty()) 4756 LinkingOutput = Cmd.getOutputFilenames().front().c_str(); 4757 else 4758 LinkingOutput = getDefaultImageName(); 4759 4760 if (CCPrintStatReportFilename.empty()) { 4761 using namespace llvm; 4762 // Human readable output. 4763 outs() << sys::path::filename(Cmd.getExecutable()) << ": " 4764 << "output=" << LinkingOutput; 4765 outs() << ", total=" 4766 << format("%.3f", ProcStat->TotalTime.count() / 1000.) << " ms" 4767 << ", user=" 4768 << format("%.3f", ProcStat->UserTime.count() / 1000.) << " ms" 4769 << ", mem=" << ProcStat->PeakMemory << " Kb\n"; 4770 } else { 4771 // CSV format. 4772 std::string Buffer; 4773 llvm::raw_string_ostream Out(Buffer); 4774 llvm::sys::printArg(Out, llvm::sys::path::filename(Cmd.getExecutable()), 4775 /*Quote*/ true); 4776 Out << ','; 4777 llvm::sys::printArg(Out, LinkingOutput, true); 4778 Out << ',' << ProcStat->TotalTime.count() << ',' 4779 << ProcStat->UserTime.count() << ',' << ProcStat->PeakMemory 4780 << '\n'; 4781 Out.flush(); 4782 std::error_code EC; 4783 llvm::raw_fd_ostream OS(CCPrintStatReportFilename, EC, 4784 llvm::sys::fs::OF_Append | 4785 llvm::sys::fs::OF_Text); 4786 if (EC) 4787 return; 4788 auto L = OS.lock(); 4789 if (!L) { 4790 llvm::errs() << "ERROR: Cannot lock file " 4791 << CCPrintStatReportFilename << ": " 4792 << toString(L.takeError()) << "\n"; 4793 return; 4794 } 4795 OS << Buffer; 4796 OS.flush(); 4797 } 4798 }); 4799 } 4800 4801 // If the user passed -Qunused-arguments or there were errors, don't warn 4802 // about any unused arguments. 4803 if (Diags.hasErrorOccurred() || 4804 C.getArgs().hasArg(options::OPT_Qunused_arguments)) 4805 return; 4806 4807 // Claim -fdriver-only here. 4808 (void)C.getArgs().hasArg(options::OPT_fdriver_only); 4809 // Claim -### here. 4810 (void)C.getArgs().hasArg(options::OPT__HASH_HASH_HASH); 4811 4812 // Claim --driver-mode, --rsp-quoting, it was handled earlier. 4813 (void)C.getArgs().hasArg(options::OPT_driver_mode); 4814 (void)C.getArgs().hasArg(options::OPT_rsp_quoting); 4815 4816 for (Arg *A : C.getArgs()) { 4817 // FIXME: It would be nice to be able to send the argument to the 4818 // DiagnosticsEngine, so that extra values, position, and so on could be 4819 // printed. 4820 if (!A->isClaimed()) { 4821 if (A->getOption().hasFlag(options::NoArgumentUnused)) 4822 continue; 4823 4824 // Suppress the warning automatically if this is just a flag, and it is an 4825 // instance of an argument we already claimed. 4826 const Option &Opt = A->getOption(); 4827 if (Opt.getKind() == Option::FlagClass) { 4828 bool DuplicateClaimed = false; 4829 4830 for (const Arg *AA : C.getArgs().filtered(&Opt)) { 4831 if (AA->isClaimed()) { 4832 DuplicateClaimed = true; 4833 break; 4834 } 4835 } 4836 4837 if (DuplicateClaimed) 4838 continue; 4839 } 4840 4841 // In clang-cl, don't mention unknown arguments here since they have 4842 // already been warned about. 4843 if (!IsCLMode() || !A->getOption().matches(options::OPT_UNKNOWN)) 4844 Diag(clang::diag::warn_drv_unused_argument) 4845 << A->getAsString(C.getArgs()); 4846 } 4847 } 4848 } 4849 4850 namespace { 4851 /// Utility class to control the collapse of dependent actions and select the 4852 /// tools accordingly. 4853 class ToolSelector final { 4854 /// The tool chain this selector refers to. 4855 const ToolChain &TC; 4856 4857 /// The compilation this selector refers to. 4858 const Compilation &C; 4859 4860 /// The base action this selector refers to. 4861 const JobAction *BaseAction; 4862 4863 /// Set to true if the current toolchain refers to host actions. 4864 bool IsHostSelector; 4865 4866 /// Set to true if save-temps and embed-bitcode functionalities are active. 4867 bool SaveTemps; 4868 bool EmbedBitcode; 4869 4870 /// Get previous dependent action or null if that does not exist. If 4871 /// \a CanBeCollapsed is false, that action must be legal to collapse or 4872 /// null will be returned. 4873 const JobAction *getPrevDependentAction(const ActionList &Inputs, 4874 ActionList &SavedOffloadAction, 4875 bool CanBeCollapsed = true) { 4876 // An option can be collapsed only if it has a single input. 4877 if (Inputs.size() != 1) 4878 return nullptr; 4879 4880 Action *CurAction = *Inputs.begin(); 4881 if (CanBeCollapsed && 4882 !CurAction->isCollapsingWithNextDependentActionLegal()) 4883 return nullptr; 4884 4885 // If the input action is an offload action. Look through it and save any 4886 // offload action that can be dropped in the event of a collapse. 4887 if (auto *OA = dyn_cast<OffloadAction>(CurAction)) { 4888 // If the dependent action is a device action, we will attempt to collapse 4889 // only with other device actions. Otherwise, we would do the same but 4890 // with host actions only. 4891 if (!IsHostSelector) { 4892 if (OA->hasSingleDeviceDependence(/*DoNotConsiderHostActions=*/true)) { 4893 CurAction = 4894 OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true); 4895 if (CanBeCollapsed && 4896 !CurAction->isCollapsingWithNextDependentActionLegal()) 4897 return nullptr; 4898 SavedOffloadAction.push_back(OA); 4899 return dyn_cast<JobAction>(CurAction); 4900 } 4901 } else if (OA->hasHostDependence()) { 4902 CurAction = OA->getHostDependence(); 4903 if (CanBeCollapsed && 4904 !CurAction->isCollapsingWithNextDependentActionLegal()) 4905 return nullptr; 4906 SavedOffloadAction.push_back(OA); 4907 return dyn_cast<JobAction>(CurAction); 4908 } 4909 return nullptr; 4910 } 4911 4912 return dyn_cast<JobAction>(CurAction); 4913 } 4914 4915 /// Return true if an assemble action can be collapsed. 4916 bool canCollapseAssembleAction() const { 4917 return TC.useIntegratedAs() && !SaveTemps && 4918 !C.getArgs().hasArg(options::OPT_via_file_asm) && 4919 !C.getArgs().hasArg(options::OPT__SLASH_FA) && 4920 !C.getArgs().hasArg(options::OPT__SLASH_Fa); 4921 } 4922 4923 /// Return true if a preprocessor action can be collapsed. 4924 bool canCollapsePreprocessorAction() const { 4925 return !C.getArgs().hasArg(options::OPT_no_integrated_cpp) && 4926 !C.getArgs().hasArg(options::OPT_traditional_cpp) && !SaveTemps && 4927 !C.getArgs().hasArg(options::OPT_rewrite_objc); 4928 } 4929 4930 /// Struct that relates an action with the offload actions that would be 4931 /// collapsed with it. 4932 struct JobActionInfo final { 4933 /// The action this info refers to. 4934 const JobAction *JA = nullptr; 4935 /// The offload actions we need to take care off if this action is 4936 /// collapsed. 4937 ActionList SavedOffloadAction; 4938 }; 4939 4940 /// Append collapsed offload actions from the give nnumber of elements in the 4941 /// action info array. 4942 static void AppendCollapsedOffloadAction(ActionList &CollapsedOffloadAction, 4943 ArrayRef<JobActionInfo> &ActionInfo, 4944 unsigned ElementNum) { 4945 assert(ElementNum <= ActionInfo.size() && "Invalid number of elements."); 4946 for (unsigned I = 0; I < ElementNum; ++I) 4947 CollapsedOffloadAction.append(ActionInfo[I].SavedOffloadAction.begin(), 4948 ActionInfo[I].SavedOffloadAction.end()); 4949 } 4950 4951 /// Functions that attempt to perform the combining. They detect if that is 4952 /// legal, and if so they update the inputs \a Inputs and the offload action 4953 /// that were collapsed in \a CollapsedOffloadAction. A tool that deals with 4954 /// the combined action is returned. If the combining is not legal or if the 4955 /// tool does not exist, null is returned. 4956 /// Currently three kinds of collapsing are supported: 4957 /// - Assemble + Backend + Compile; 4958 /// - Assemble + Backend ; 4959 /// - Backend + Compile. 4960 const Tool * 4961 combineAssembleBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4962 ActionList &Inputs, 4963 ActionList &CollapsedOffloadAction) { 4964 if (ActionInfo.size() < 3 || !canCollapseAssembleAction()) 4965 return nullptr; 4966 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4967 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4968 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[2].JA); 4969 if (!AJ || !BJ || !CJ) 4970 return nullptr; 4971 4972 // Get compiler tool. 4973 const Tool *T = TC.SelectTool(*CJ); 4974 if (!T) 4975 return nullptr; 4976 4977 // Can't collapse if we don't have codegen support unless we are 4978 // emitting LLVM IR. 4979 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 4980 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 4981 return nullptr; 4982 4983 // When using -fembed-bitcode, it is required to have the same tool (clang) 4984 // for both CompilerJA and BackendJA. Otherwise, combine two stages. 4985 if (EmbedBitcode) { 4986 const Tool *BT = TC.SelectTool(*BJ); 4987 if (BT == T) 4988 return nullptr; 4989 } 4990 4991 if (!T->hasIntegratedAssembler()) 4992 return nullptr; 4993 4994 Inputs = CJ->getInputs(); 4995 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4996 /*NumElements=*/3); 4997 return T; 4998 } 4999 const Tool *combineAssembleBackend(ArrayRef<JobActionInfo> ActionInfo, 5000 ActionList &Inputs, 5001 ActionList &CollapsedOffloadAction) { 5002 if (ActionInfo.size() < 2 || !canCollapseAssembleAction()) 5003 return nullptr; 5004 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 5005 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 5006 if (!AJ || !BJ) 5007 return nullptr; 5008 5009 // Get backend tool. 5010 const Tool *T = TC.SelectTool(*BJ); 5011 if (!T) 5012 return nullptr; 5013 5014 if (!T->hasIntegratedAssembler()) 5015 return nullptr; 5016 5017 Inputs = BJ->getInputs(); 5018 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5019 /*NumElements=*/2); 5020 return T; 5021 } 5022 const Tool *combineBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 5023 ActionList &Inputs, 5024 ActionList &CollapsedOffloadAction) { 5025 if (ActionInfo.size() < 2) 5026 return nullptr; 5027 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[0].JA); 5028 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[1].JA); 5029 if (!BJ || !CJ) 5030 return nullptr; 5031 5032 // Check if the initial input (to the compile job or its predessor if one 5033 // exists) is LLVM bitcode. In that case, no preprocessor step is required 5034 // and we can still collapse the compile and backend jobs when we have 5035 // -save-temps. I.e. there is no need for a separate compile job just to 5036 // emit unoptimized bitcode. 5037 bool InputIsBitcode = true; 5038 for (size_t i = 1; i < ActionInfo.size(); i++) 5039 if (ActionInfo[i].JA->getType() != types::TY_LLVM_BC && 5040 ActionInfo[i].JA->getType() != types::TY_LTO_BC) { 5041 InputIsBitcode = false; 5042 break; 5043 } 5044 if (!InputIsBitcode && !canCollapsePreprocessorAction()) 5045 return nullptr; 5046 5047 // Get compiler tool. 5048 const Tool *T = TC.SelectTool(*CJ); 5049 if (!T) 5050 return nullptr; 5051 5052 // Can't collapse if we don't have codegen support unless we are 5053 // emitting LLVM IR. 5054 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 5055 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 5056 return nullptr; 5057 5058 if (T->canEmitIR() && ((SaveTemps && !InputIsBitcode) || EmbedBitcode)) 5059 return nullptr; 5060 5061 Inputs = CJ->getInputs(); 5062 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5063 /*NumElements=*/2); 5064 return T; 5065 } 5066 5067 /// Updates the inputs if the obtained tool supports combining with 5068 /// preprocessor action, and the current input is indeed a preprocessor 5069 /// action. If combining results in the collapse of offloading actions, those 5070 /// are appended to \a CollapsedOffloadAction. 5071 void combineWithPreprocessor(const Tool *T, ActionList &Inputs, 5072 ActionList &CollapsedOffloadAction) { 5073 if (!T || !canCollapsePreprocessorAction() || !T->hasIntegratedCPP()) 5074 return; 5075 5076 // Attempt to get a preprocessor action dependence. 5077 ActionList PreprocessJobOffloadActions; 5078 ActionList NewInputs; 5079 for (Action *A : Inputs) { 5080 auto *PJ = getPrevDependentAction({A}, PreprocessJobOffloadActions); 5081 if (!PJ || !isa<PreprocessJobAction>(PJ)) { 5082 NewInputs.push_back(A); 5083 continue; 5084 } 5085 5086 // This is legal to combine. Append any offload action we found and add the 5087 // current input to preprocessor inputs. 5088 CollapsedOffloadAction.append(PreprocessJobOffloadActions.begin(), 5089 PreprocessJobOffloadActions.end()); 5090 NewInputs.append(PJ->input_begin(), PJ->input_end()); 5091 } 5092 Inputs = NewInputs; 5093 } 5094 5095 public: 5096 ToolSelector(const JobAction *BaseAction, const ToolChain &TC, 5097 const Compilation &C, bool SaveTemps, bool EmbedBitcode) 5098 : TC(TC), C(C), BaseAction(BaseAction), SaveTemps(SaveTemps), 5099 EmbedBitcode(EmbedBitcode) { 5100 assert(BaseAction && "Invalid base action."); 5101 IsHostSelector = BaseAction->getOffloadingDeviceKind() == Action::OFK_None; 5102 } 5103 5104 /// Check if a chain of actions can be combined and return the tool that can 5105 /// handle the combination of actions. The pointer to the current inputs \a 5106 /// Inputs and the list of offload actions \a CollapsedOffloadActions 5107 /// connected to collapsed actions are updated accordingly. The latter enables 5108 /// the caller of the selector to process them afterwards instead of just 5109 /// dropping them. If no suitable tool is found, null will be returned. 5110 const Tool *getTool(ActionList &Inputs, 5111 ActionList &CollapsedOffloadAction) { 5112 // 5113 // Get the largest chain of actions that we could combine. 5114 // 5115 5116 SmallVector<JobActionInfo, 5> ActionChain(1); 5117 ActionChain.back().JA = BaseAction; 5118 while (ActionChain.back().JA) { 5119 const Action *CurAction = ActionChain.back().JA; 5120 5121 // Grow the chain by one element. 5122 ActionChain.resize(ActionChain.size() + 1); 5123 JobActionInfo &AI = ActionChain.back(); 5124 5125 // Attempt to fill it with the 5126 AI.JA = 5127 getPrevDependentAction(CurAction->getInputs(), AI.SavedOffloadAction); 5128 } 5129 5130 // Pop the last action info as it could not be filled. 5131 ActionChain.pop_back(); 5132 5133 // 5134 // Attempt to combine actions. If all combining attempts failed, just return 5135 // the tool of the provided action. At the end we attempt to combine the 5136 // action with any preprocessor action it may depend on. 5137 // 5138 5139 const Tool *T = combineAssembleBackendCompile(ActionChain, Inputs, 5140 CollapsedOffloadAction); 5141 if (!T) 5142 T = combineAssembleBackend(ActionChain, Inputs, CollapsedOffloadAction); 5143 if (!T) 5144 T = combineBackendCompile(ActionChain, Inputs, CollapsedOffloadAction); 5145 if (!T) { 5146 Inputs = BaseAction->getInputs(); 5147 T = TC.SelectTool(*BaseAction); 5148 } 5149 5150 combineWithPreprocessor(T, Inputs, CollapsedOffloadAction); 5151 return T; 5152 } 5153 }; 5154 } 5155 5156 /// Return a string that uniquely identifies the result of a job. The bound arch 5157 /// is not necessarily represented in the toolchain's triple -- for example, 5158 /// armv7 and armv7s both map to the same triple -- so we need both in our map. 5159 /// Also, we need to add the offloading device kind, as the same tool chain can 5160 /// be used for host and device for some programming models, e.g. OpenMP. 5161 static std::string GetTriplePlusArchString(const ToolChain *TC, 5162 StringRef BoundArch, 5163 Action::OffloadKind OffloadKind) { 5164 std::string TriplePlusArch = TC->getTriple().normalize(); 5165 if (!BoundArch.empty()) { 5166 TriplePlusArch += "-"; 5167 TriplePlusArch += BoundArch; 5168 } 5169 TriplePlusArch += "-"; 5170 TriplePlusArch += Action::GetOffloadKindName(OffloadKind); 5171 return TriplePlusArch; 5172 } 5173 5174 InputInfoList Driver::BuildJobsForAction( 5175 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5176 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5177 std::map<std::pair<const Action *, std::string>, InputInfoList> 5178 &CachedResults, 5179 Action::OffloadKind TargetDeviceOffloadKind) const { 5180 std::pair<const Action *, std::string> ActionTC = { 5181 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5182 auto CachedResult = CachedResults.find(ActionTC); 5183 if (CachedResult != CachedResults.end()) { 5184 return CachedResult->second; 5185 } 5186 InputInfoList Result = BuildJobsForActionNoCache( 5187 C, A, TC, BoundArch, AtTopLevel, MultipleArchs, LinkingOutput, 5188 CachedResults, TargetDeviceOffloadKind); 5189 CachedResults[ActionTC] = Result; 5190 return Result; 5191 } 5192 5193 InputInfoList Driver::BuildJobsForActionNoCache( 5194 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5195 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5196 std::map<std::pair<const Action *, std::string>, InputInfoList> 5197 &CachedResults, 5198 Action::OffloadKind TargetDeviceOffloadKind) const { 5199 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 5200 5201 InputInfoList OffloadDependencesInputInfo; 5202 bool BuildingForOffloadDevice = TargetDeviceOffloadKind != Action::OFK_None; 5203 if (const OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 5204 // The 'Darwin' toolchain is initialized only when its arguments are 5205 // computed. Get the default arguments for OFK_None to ensure that 5206 // initialization is performed before processing the offload action. 5207 // FIXME: Remove when darwin's toolchain is initialized during construction. 5208 C.getArgsForToolChain(TC, BoundArch, Action::OFK_None); 5209 5210 // The offload action is expected to be used in four different situations. 5211 // 5212 // a) Set a toolchain/architecture/kind for a host action: 5213 // Host Action 1 -> OffloadAction -> Host Action 2 5214 // 5215 // b) Set a toolchain/architecture/kind for a device action; 5216 // Device Action 1 -> OffloadAction -> Device Action 2 5217 // 5218 // c) Specify a device dependence to a host action; 5219 // Device Action 1 _ 5220 // \ 5221 // Host Action 1 ---> OffloadAction -> Host Action 2 5222 // 5223 // d) Specify a host dependence to a device action. 5224 // Host Action 1 _ 5225 // \ 5226 // Device Action 1 ---> OffloadAction -> Device Action 2 5227 // 5228 // For a) and b), we just return the job generated for the dependence. For 5229 // c) and d) we override the current action with the host/device dependence 5230 // if the current toolchain is host/device and set the offload dependences 5231 // info with the jobs obtained from the device/host dependence(s). 5232 5233 // If there is a single device option, just generate the job for it. 5234 if (OA->hasSingleDeviceDependence()) { 5235 InputInfoList DevA; 5236 OA->doOnEachDeviceDependence([&](Action *DepA, const ToolChain *DepTC, 5237 const char *DepBoundArch) { 5238 DevA = 5239 BuildJobsForAction(C, DepA, DepTC, DepBoundArch, AtTopLevel, 5240 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, 5241 CachedResults, DepA->getOffloadingDeviceKind()); 5242 }); 5243 return DevA; 5244 } 5245 5246 // If 'Action 2' is host, we generate jobs for the device dependences and 5247 // override the current action with the host dependence. Otherwise, we 5248 // generate the host dependences and override the action with the device 5249 // dependence. The dependences can't therefore be a top-level action. 5250 OA->doOnEachDependence( 5251 /*IsHostDependence=*/BuildingForOffloadDevice, 5252 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5253 OffloadDependencesInputInfo.append(BuildJobsForAction( 5254 C, DepA, DepTC, DepBoundArch, /*AtTopLevel=*/false, 5255 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, CachedResults, 5256 DepA->getOffloadingDeviceKind())); 5257 }); 5258 5259 A = BuildingForOffloadDevice 5260 ? OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true) 5261 : OA->getHostDependence(); 5262 5263 // We may have already built this action as a part of the offloading 5264 // toolchain, return the cached input if so. 5265 std::pair<const Action *, std::string> ActionTC = { 5266 OA->getHostDependence(), 5267 GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5268 if (CachedResults.find(ActionTC) != CachedResults.end()) { 5269 InputInfoList Inputs = CachedResults[ActionTC]; 5270 Inputs.append(OffloadDependencesInputInfo); 5271 return Inputs; 5272 } 5273 } 5274 5275 if (const InputAction *IA = dyn_cast<InputAction>(A)) { 5276 // FIXME: It would be nice to not claim this here; maybe the old scheme of 5277 // just using Args was better? 5278 const Arg &Input = IA->getInputArg(); 5279 Input.claim(); 5280 if (Input.getOption().matches(options::OPT_INPUT)) { 5281 const char *Name = Input.getValue(); 5282 return {InputInfo(A, Name, /* _BaseInput = */ Name)}; 5283 } 5284 return {InputInfo(A, &Input, /* _BaseInput = */ "")}; 5285 } 5286 5287 if (const BindArchAction *BAA = dyn_cast<BindArchAction>(A)) { 5288 const ToolChain *TC; 5289 StringRef ArchName = BAA->getArchName(); 5290 5291 if (!ArchName.empty()) 5292 TC = &getToolChain(C.getArgs(), 5293 computeTargetTriple(*this, TargetTriple, 5294 C.getArgs(), ArchName)); 5295 else 5296 TC = &C.getDefaultToolChain(); 5297 5298 return BuildJobsForAction(C, *BAA->input_begin(), TC, ArchName, AtTopLevel, 5299 MultipleArchs, LinkingOutput, CachedResults, 5300 TargetDeviceOffloadKind); 5301 } 5302 5303 5304 ActionList Inputs = A->getInputs(); 5305 5306 const JobAction *JA = cast<JobAction>(A); 5307 ActionList CollapsedOffloadActions; 5308 5309 ToolSelector TS(JA, *TC, C, isSaveTempsEnabled(), 5310 embedBitcodeInObject() && !isUsingLTO()); 5311 const Tool *T = TS.getTool(Inputs, CollapsedOffloadActions); 5312 5313 if (!T) 5314 return {InputInfo()}; 5315 5316 if (BuildingForOffloadDevice && 5317 A->getOffloadingDeviceKind() == Action::OFK_OpenMP) { 5318 if (TC->getTriple().isAMDGCN()) { 5319 // AMDGCN treats backend and assemble actions as no-op because 5320 // linker does not support object files. 5321 if (const BackendJobAction *BA = dyn_cast<BackendJobAction>(A)) { 5322 return BuildJobsForAction(C, *BA->input_begin(), TC, BoundArch, 5323 AtTopLevel, MultipleArchs, LinkingOutput, 5324 CachedResults, TargetDeviceOffloadKind); 5325 } 5326 5327 if (const AssembleJobAction *AA = dyn_cast<AssembleJobAction>(A)) { 5328 return BuildJobsForAction(C, *AA->input_begin(), TC, BoundArch, 5329 AtTopLevel, MultipleArchs, LinkingOutput, 5330 CachedResults, TargetDeviceOffloadKind); 5331 } 5332 } 5333 } 5334 5335 // If we've collapsed action list that contained OffloadAction we 5336 // need to build jobs for host/device-side inputs it may have held. 5337 for (const auto *OA : CollapsedOffloadActions) 5338 cast<OffloadAction>(OA)->doOnEachDependence( 5339 /*IsHostDependence=*/BuildingForOffloadDevice, 5340 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5341 OffloadDependencesInputInfo.append(BuildJobsForAction( 5342 C, DepA, DepTC, DepBoundArch, /* AtTopLevel */ false, 5343 /*MultipleArchs=*/!!DepBoundArch, LinkingOutput, CachedResults, 5344 DepA->getOffloadingDeviceKind())); 5345 }); 5346 5347 // Only use pipes when there is exactly one input. 5348 InputInfoList InputInfos; 5349 for (const Action *Input : Inputs) { 5350 // Treat dsymutil and verify sub-jobs as being at the top-level too, they 5351 // shouldn't get temporary output names. 5352 // FIXME: Clean this up. 5353 bool SubJobAtTopLevel = 5354 AtTopLevel && (isa<DsymutilJobAction>(A) || isa<VerifyJobAction>(A)); 5355 InputInfos.append(BuildJobsForAction( 5356 C, Input, TC, BoundArch, SubJobAtTopLevel, MultipleArchs, LinkingOutput, 5357 CachedResults, A->getOffloadingDeviceKind())); 5358 } 5359 5360 // Always use the first file input as the base input. 5361 const char *BaseInput = InputInfos[0].getBaseInput(); 5362 for (auto &Info : InputInfos) { 5363 if (Info.isFilename()) { 5364 BaseInput = Info.getBaseInput(); 5365 break; 5366 } 5367 } 5368 5369 // ... except dsymutil actions, which use their actual input as the base 5370 // input. 5371 if (JA->getType() == types::TY_dSYM) 5372 BaseInput = InputInfos[0].getFilename(); 5373 5374 // ... and in header module compilations, which use the module name. 5375 if (auto *ModuleJA = dyn_cast<HeaderModulePrecompileJobAction>(JA)) 5376 BaseInput = ModuleJA->getModuleName(); 5377 5378 // Append outputs of offload device jobs to the input list 5379 if (!OffloadDependencesInputInfo.empty()) 5380 InputInfos.append(OffloadDependencesInputInfo.begin(), 5381 OffloadDependencesInputInfo.end()); 5382 5383 // Set the effective triple of the toolchain for the duration of this job. 5384 llvm::Triple EffectiveTriple; 5385 const ToolChain &ToolTC = T->getToolChain(); 5386 const ArgList &Args = 5387 C.getArgsForToolChain(TC, BoundArch, A->getOffloadingDeviceKind()); 5388 if (InputInfos.size() != 1) { 5389 EffectiveTriple = llvm::Triple(ToolTC.ComputeEffectiveClangTriple(Args)); 5390 } else { 5391 // Pass along the input type if it can be unambiguously determined. 5392 EffectiveTriple = llvm::Triple( 5393 ToolTC.ComputeEffectiveClangTriple(Args, InputInfos[0].getType())); 5394 } 5395 RegisterEffectiveTriple TripleRAII(ToolTC, EffectiveTriple); 5396 5397 // Determine the place to write output to, if any. 5398 InputInfo Result; 5399 InputInfoList UnbundlingResults; 5400 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(JA)) { 5401 // If we have an unbundling job, we need to create results for all the 5402 // outputs. We also update the results cache so that other actions using 5403 // this unbundling action can get the right results. 5404 for (auto &UI : UA->getDependentActionsInfo()) { 5405 assert(UI.DependentOffloadKind != Action::OFK_None && 5406 "Unbundling with no offloading??"); 5407 5408 // Unbundling actions are never at the top level. When we generate the 5409 // offloading prefix, we also do that for the host file because the 5410 // unbundling action does not change the type of the output which can 5411 // cause a overwrite. 5412 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5413 UI.DependentOffloadKind, 5414 UI.DependentToolChain->getTriple().normalize(), 5415 /*CreatePrefixForHost=*/true); 5416 auto CurI = InputInfo( 5417 UA, 5418 GetNamedOutputPath(C, *UA, BaseInput, UI.DependentBoundArch, 5419 /*AtTopLevel=*/false, 5420 MultipleArchs || 5421 UI.DependentOffloadKind == Action::OFK_HIP, 5422 OffloadingPrefix), 5423 BaseInput); 5424 // Save the unbundling result. 5425 UnbundlingResults.push_back(CurI); 5426 5427 // Get the unique string identifier for this dependence and cache the 5428 // result. 5429 StringRef Arch; 5430 if (TargetDeviceOffloadKind == Action::OFK_HIP) { 5431 if (UI.DependentOffloadKind == Action::OFK_Host) 5432 Arch = StringRef(); 5433 else 5434 Arch = UI.DependentBoundArch; 5435 } else 5436 Arch = BoundArch; 5437 5438 CachedResults[{A, GetTriplePlusArchString(UI.DependentToolChain, Arch, 5439 UI.DependentOffloadKind)}] = { 5440 CurI}; 5441 } 5442 5443 // Now that we have all the results generated, select the one that should be 5444 // returned for the current depending action. 5445 std::pair<const Action *, std::string> ActionTC = { 5446 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5447 assert(CachedResults.find(ActionTC) != CachedResults.end() && 5448 "Result does not exist??"); 5449 Result = CachedResults[ActionTC].front(); 5450 } else if (JA->getType() == types::TY_Nothing) 5451 Result = {InputInfo(A, BaseInput)}; 5452 else { 5453 // We only have to generate a prefix for the host if this is not a top-level 5454 // action. 5455 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5456 A->getOffloadingDeviceKind(), TC->getTriple().normalize(), 5457 /*CreatePrefixForHost=*/isa<OffloadPackagerJobAction>(A) || 5458 !(A->getOffloadingHostActiveKinds() == Action::OFK_None || 5459 AtTopLevel)); 5460 if (isa<OffloadWrapperJobAction>(JA)) { 5461 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5462 BaseInput = FinalOutput->getValue(); 5463 else 5464 BaseInput = getDefaultImageName(); 5465 BaseInput = 5466 C.getArgs().MakeArgString(std::string(BaseInput) + "-wrapper"); 5467 } 5468 Result = InputInfo(A, GetNamedOutputPath(C, *JA, BaseInput, BoundArch, 5469 AtTopLevel, MultipleArchs, 5470 OffloadingPrefix), 5471 BaseInput); 5472 } 5473 5474 if (CCCPrintBindings && !CCGenDiagnostics) { 5475 llvm::errs() << "# \"" << T->getToolChain().getTripleString() << '"' 5476 << " - \"" << T->getName() << "\", inputs: ["; 5477 for (unsigned i = 0, e = InputInfos.size(); i != e; ++i) { 5478 llvm::errs() << InputInfos[i].getAsString(); 5479 if (i + 1 != e) 5480 llvm::errs() << ", "; 5481 } 5482 if (UnbundlingResults.empty()) 5483 llvm::errs() << "], output: " << Result.getAsString() << "\n"; 5484 else { 5485 llvm::errs() << "], outputs: ["; 5486 for (unsigned i = 0, e = UnbundlingResults.size(); i != e; ++i) { 5487 llvm::errs() << UnbundlingResults[i].getAsString(); 5488 if (i + 1 != e) 5489 llvm::errs() << ", "; 5490 } 5491 llvm::errs() << "] \n"; 5492 } 5493 } else { 5494 if (UnbundlingResults.empty()) 5495 T->ConstructJob( 5496 C, *JA, Result, InputInfos, 5497 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5498 LinkingOutput); 5499 else 5500 T->ConstructJobMultipleOutputs( 5501 C, *JA, UnbundlingResults, InputInfos, 5502 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5503 LinkingOutput); 5504 } 5505 return {Result}; 5506 } 5507 5508 const char *Driver::getDefaultImageName() const { 5509 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 5510 return Target.isOSWindows() ? "a.exe" : "a.out"; 5511 } 5512 5513 /// Create output filename based on ArgValue, which could either be a 5514 /// full filename, filename without extension, or a directory. If ArgValue 5515 /// does not provide a filename, then use BaseName, and use the extension 5516 /// suitable for FileType. 5517 static const char *MakeCLOutputFilename(const ArgList &Args, StringRef ArgValue, 5518 StringRef BaseName, 5519 types::ID FileType) { 5520 SmallString<128> Filename = ArgValue; 5521 5522 if (ArgValue.empty()) { 5523 // If the argument is empty, output to BaseName in the current dir. 5524 Filename = BaseName; 5525 } else if (llvm::sys::path::is_separator(Filename.back())) { 5526 // If the argument is a directory, output to BaseName in that dir. 5527 llvm::sys::path::append(Filename, BaseName); 5528 } 5529 5530 if (!llvm::sys::path::has_extension(ArgValue)) { 5531 // If the argument didn't provide an extension, then set it. 5532 const char *Extension = types::getTypeTempSuffix(FileType, true); 5533 5534 if (FileType == types::TY_Image && 5535 Args.hasArg(options::OPT__SLASH_LD, options::OPT__SLASH_LDd)) { 5536 // The output file is a dll. 5537 Extension = "dll"; 5538 } 5539 5540 llvm::sys::path::replace_extension(Filename, Extension); 5541 } 5542 5543 return Args.MakeArgString(Filename.c_str()); 5544 } 5545 5546 static bool HasPreprocessOutput(const Action &JA) { 5547 if (isa<PreprocessJobAction>(JA)) 5548 return true; 5549 if (isa<OffloadAction>(JA) && isa<PreprocessJobAction>(JA.getInputs()[0])) 5550 return true; 5551 if (isa<OffloadBundlingJobAction>(JA) && 5552 HasPreprocessOutput(*(JA.getInputs()[0]))) 5553 return true; 5554 return false; 5555 } 5556 5557 const char *Driver::GetNamedOutputPath(Compilation &C, const JobAction &JA, 5558 const char *BaseInput, 5559 StringRef OrigBoundArch, bool AtTopLevel, 5560 bool MultipleArchs, 5561 StringRef OffloadingPrefix) const { 5562 std::string BoundArch = OrigBoundArch.str(); 5563 if (is_style_windows(llvm::sys::path::Style::native)) { 5564 // BoundArch may contains ':', which is invalid in file names on Windows, 5565 // therefore replace it with '%'. 5566 std::replace(BoundArch.begin(), BoundArch.end(), ':', '@'); 5567 } 5568 5569 llvm::PrettyStackTraceString CrashInfo("Computing output path"); 5570 // Output to a user requested destination? 5571 if (AtTopLevel && !isa<DsymutilJobAction>(JA) && !isa<VerifyJobAction>(JA)) { 5572 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5573 return C.addResultFile(FinalOutput->getValue(), &JA); 5574 } 5575 5576 // For /P, preprocess to file named after BaseInput. 5577 if (C.getArgs().hasArg(options::OPT__SLASH_P)) { 5578 assert(AtTopLevel && isa<PreprocessJobAction>(JA)); 5579 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5580 StringRef NameArg; 5581 if (Arg *A = C.getArgs().getLastArg(options::OPT__SLASH_Fi)) 5582 NameArg = A->getValue(); 5583 return C.addResultFile( 5584 MakeCLOutputFilename(C.getArgs(), NameArg, BaseName, types::TY_PP_C), 5585 &JA); 5586 } 5587 5588 // Default to writing to stdout? 5589 if (AtTopLevel && !CCGenDiagnostics && HasPreprocessOutput(JA)) { 5590 return "-"; 5591 } 5592 5593 if (JA.getType() == types::TY_ModuleFile && 5594 C.getArgs().getLastArg(options::OPT_module_file_info)) { 5595 return "-"; 5596 } 5597 5598 // Is this the assembly listing for /FA? 5599 if (JA.getType() == types::TY_PP_Asm && 5600 (C.getArgs().hasArg(options::OPT__SLASH_FA) || 5601 C.getArgs().hasArg(options::OPT__SLASH_Fa))) { 5602 // Use /Fa and the input filename to determine the asm file name. 5603 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5604 StringRef FaValue = C.getArgs().getLastArgValue(options::OPT__SLASH_Fa); 5605 return C.addResultFile( 5606 MakeCLOutputFilename(C.getArgs(), FaValue, BaseName, JA.getType()), 5607 &JA); 5608 } 5609 5610 // Output to a temporary file? 5611 if ((!AtTopLevel && !isSaveTempsEnabled() && 5612 !C.getArgs().hasArg(options::OPT__SLASH_Fo)) || 5613 CCGenDiagnostics) { 5614 StringRef Name = llvm::sys::path::filename(BaseInput); 5615 std::pair<StringRef, StringRef> Split = Name.split('.'); 5616 SmallString<128> TmpName; 5617 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5618 Arg *A = C.getArgs().getLastArg(options::OPT_fcrash_diagnostics_dir); 5619 if (CCGenDiagnostics && A) { 5620 SmallString<128> CrashDirectory(A->getValue()); 5621 if (!getVFS().exists(CrashDirectory)) 5622 llvm::sys::fs::create_directories(CrashDirectory); 5623 llvm::sys::path::append(CrashDirectory, Split.first); 5624 const char *Middle = Suffix ? "-%%%%%%." : "-%%%%%%"; 5625 std::error_code EC = llvm::sys::fs::createUniqueFile( 5626 CrashDirectory + Middle + Suffix, TmpName); 5627 if (EC) { 5628 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5629 return ""; 5630 } 5631 } else { 5632 if (MultipleArchs && !BoundArch.empty()) { 5633 TmpName = GetTemporaryDirectory(Split.first); 5634 llvm::sys::path::append(TmpName, 5635 Split.first + "-" + BoundArch + "." + Suffix); 5636 } else { 5637 TmpName = GetTemporaryPath(Split.first, Suffix); 5638 } 5639 } 5640 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5641 } 5642 5643 SmallString<128> BasePath(BaseInput); 5644 SmallString<128> ExternalPath(""); 5645 StringRef BaseName; 5646 5647 // Dsymutil actions should use the full path. 5648 if (isa<DsymutilJobAction>(JA) && C.getArgs().hasArg(options::OPT_dsym_dir)) { 5649 ExternalPath += C.getArgs().getLastArg(options::OPT_dsym_dir)->getValue(); 5650 // We use posix style here because the tests (specifically 5651 // darwin-dsymutil.c) demonstrate that posix style paths are acceptable 5652 // even on Windows and if we don't then the similar test covering this 5653 // fails. 5654 llvm::sys::path::append(ExternalPath, llvm::sys::path::Style::posix, 5655 llvm::sys::path::filename(BasePath)); 5656 BaseName = ExternalPath; 5657 } else if (isa<DsymutilJobAction>(JA) || isa<VerifyJobAction>(JA)) 5658 BaseName = BasePath; 5659 else 5660 BaseName = llvm::sys::path::filename(BasePath); 5661 5662 // Determine what the derived output name should be. 5663 const char *NamedOutput; 5664 5665 if ((JA.getType() == types::TY_Object || JA.getType() == types::TY_LTO_BC) && 5666 C.getArgs().hasArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o)) { 5667 // The /Fo or /o flag decides the object filename. 5668 StringRef Val = 5669 C.getArgs() 5670 .getLastArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o) 5671 ->getValue(); 5672 NamedOutput = 5673 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5674 } else if (JA.getType() == types::TY_Image && 5675 C.getArgs().hasArg(options::OPT__SLASH_Fe, 5676 options::OPT__SLASH_o)) { 5677 // The /Fe or /o flag names the linked file. 5678 StringRef Val = 5679 C.getArgs() 5680 .getLastArg(options::OPT__SLASH_Fe, options::OPT__SLASH_o) 5681 ->getValue(); 5682 NamedOutput = 5683 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Image); 5684 } else if (JA.getType() == types::TY_Image) { 5685 if (IsCLMode()) { 5686 // clang-cl uses BaseName for the executable name. 5687 NamedOutput = 5688 MakeCLOutputFilename(C.getArgs(), "", BaseName, types::TY_Image); 5689 } else { 5690 SmallString<128> Output(getDefaultImageName()); 5691 // HIP image for device compilation with -fno-gpu-rdc is per compilation 5692 // unit. 5693 bool IsHIPNoRDC = JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5694 !C.getArgs().hasFlag(options::OPT_fgpu_rdc, 5695 options::OPT_fno_gpu_rdc, false); 5696 bool UseOutExtension = IsHIPNoRDC || isa<OffloadPackagerJobAction>(JA); 5697 if (UseOutExtension) { 5698 Output = BaseName; 5699 llvm::sys::path::replace_extension(Output, ""); 5700 } 5701 Output += OffloadingPrefix; 5702 if (MultipleArchs && !BoundArch.empty()) { 5703 Output += "-"; 5704 Output.append(BoundArch); 5705 } 5706 if (UseOutExtension) 5707 Output += ".out"; 5708 NamedOutput = C.getArgs().MakeArgString(Output.c_str()); 5709 } 5710 } else if (JA.getType() == types::TY_PCH && IsCLMode()) { 5711 NamedOutput = C.getArgs().MakeArgString(GetClPchPath(C, BaseName)); 5712 } else if ((JA.getType() == types::TY_Plist || JA.getType() == types::TY_AST) && 5713 C.getArgs().hasArg(options::OPT__SLASH_o)) { 5714 StringRef Val = 5715 C.getArgs() 5716 .getLastArg(options::OPT__SLASH_o) 5717 ->getValue(); 5718 NamedOutput = 5719 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5720 } else { 5721 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5722 assert(Suffix && "All types used for output should have a suffix."); 5723 5724 std::string::size_type End = std::string::npos; 5725 if (!types::appendSuffixForType(JA.getType())) 5726 End = BaseName.rfind('.'); 5727 SmallString<128> Suffixed(BaseName.substr(0, End)); 5728 Suffixed += OffloadingPrefix; 5729 if (MultipleArchs && !BoundArch.empty()) { 5730 Suffixed += "-"; 5731 Suffixed.append(BoundArch); 5732 } 5733 // When using both -save-temps and -emit-llvm, use a ".tmp.bc" suffix for 5734 // the unoptimized bitcode so that it does not get overwritten by the ".bc" 5735 // optimized bitcode output. 5736 auto IsHIPRDCInCompilePhase = [](const JobAction &JA, 5737 const llvm::opt::DerivedArgList &Args) { 5738 // The relocatable compilation in HIP implies -emit-llvm. Similarly, use a 5739 // ".tmp.bc" suffix for the unoptimized bitcode (generated in the compile 5740 // phase.) 5741 return isa<CompileJobAction>(JA) && 5742 JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5743 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 5744 false); 5745 }; 5746 if (!AtTopLevel && JA.getType() == types::TY_LLVM_BC && 5747 (C.getArgs().hasArg(options::OPT_emit_llvm) || 5748 IsHIPRDCInCompilePhase(JA, C.getArgs()))) 5749 Suffixed += ".tmp"; 5750 Suffixed += '.'; 5751 Suffixed += Suffix; 5752 NamedOutput = C.getArgs().MakeArgString(Suffixed.c_str()); 5753 } 5754 5755 // Prepend object file path if -save-temps=obj 5756 if (!AtTopLevel && isSaveTempsObj() && C.getArgs().hasArg(options::OPT_o) && 5757 JA.getType() != types::TY_PCH) { 5758 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 5759 SmallString<128> TempPath(FinalOutput->getValue()); 5760 llvm::sys::path::remove_filename(TempPath); 5761 StringRef OutputFileName = llvm::sys::path::filename(NamedOutput); 5762 llvm::sys::path::append(TempPath, OutputFileName); 5763 NamedOutput = C.getArgs().MakeArgString(TempPath.c_str()); 5764 } 5765 5766 // If we're saving temps and the temp file conflicts with the input file, 5767 // then avoid overwriting input file. 5768 if (!AtTopLevel && isSaveTempsEnabled() && NamedOutput == BaseName) { 5769 bool SameFile = false; 5770 SmallString<256> Result; 5771 llvm::sys::fs::current_path(Result); 5772 llvm::sys::path::append(Result, BaseName); 5773 llvm::sys::fs::equivalent(BaseInput, Result.c_str(), SameFile); 5774 // Must share the same path to conflict. 5775 if (SameFile) { 5776 StringRef Name = llvm::sys::path::filename(BaseInput); 5777 std::pair<StringRef, StringRef> Split = Name.split('.'); 5778 std::string TmpName = GetTemporaryPath( 5779 Split.first, types::getTypeTempSuffix(JA.getType(), IsCLMode())); 5780 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5781 } 5782 } 5783 5784 // As an annoying special case, PCH generation doesn't strip the pathname. 5785 if (JA.getType() == types::TY_PCH && !IsCLMode()) { 5786 llvm::sys::path::remove_filename(BasePath); 5787 if (BasePath.empty()) 5788 BasePath = NamedOutput; 5789 else 5790 llvm::sys::path::append(BasePath, NamedOutput); 5791 return C.addResultFile(C.getArgs().MakeArgString(BasePath.c_str()), &JA); 5792 } else { 5793 return C.addResultFile(NamedOutput, &JA); 5794 } 5795 } 5796 5797 std::string Driver::GetFilePath(StringRef Name, const ToolChain &TC) const { 5798 // Search for Name in a list of paths. 5799 auto SearchPaths = [&](const llvm::SmallVectorImpl<std::string> &P) 5800 -> llvm::Optional<std::string> { 5801 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5802 // attempting to use this prefix when looking for file paths. 5803 for (const auto &Dir : P) { 5804 if (Dir.empty()) 5805 continue; 5806 SmallString<128> P(Dir[0] == '=' ? SysRoot + Dir.substr(1) : Dir); 5807 llvm::sys::path::append(P, Name); 5808 if (llvm::sys::fs::exists(Twine(P))) 5809 return std::string(P); 5810 } 5811 return None; 5812 }; 5813 5814 if (auto P = SearchPaths(PrefixDirs)) 5815 return *P; 5816 5817 SmallString<128> R(ResourceDir); 5818 llvm::sys::path::append(R, Name); 5819 if (llvm::sys::fs::exists(Twine(R))) 5820 return std::string(R.str()); 5821 5822 SmallString<128> P(TC.getCompilerRTPath()); 5823 llvm::sys::path::append(P, Name); 5824 if (llvm::sys::fs::exists(Twine(P))) 5825 return std::string(P.str()); 5826 5827 SmallString<128> D(Dir); 5828 llvm::sys::path::append(D, "..", Name); 5829 if (llvm::sys::fs::exists(Twine(D))) 5830 return std::string(D.str()); 5831 5832 if (auto P = SearchPaths(TC.getLibraryPaths())) 5833 return *P; 5834 5835 if (auto P = SearchPaths(TC.getFilePaths())) 5836 return *P; 5837 5838 return std::string(Name); 5839 } 5840 5841 void Driver::generatePrefixedToolNames( 5842 StringRef Tool, const ToolChain &TC, 5843 SmallVectorImpl<std::string> &Names) const { 5844 // FIXME: Needs a better variable than TargetTriple 5845 Names.emplace_back((TargetTriple + "-" + Tool).str()); 5846 Names.emplace_back(Tool); 5847 } 5848 5849 static bool ScanDirForExecutable(SmallString<128> &Dir, StringRef Name) { 5850 llvm::sys::path::append(Dir, Name); 5851 if (llvm::sys::fs::can_execute(Twine(Dir))) 5852 return true; 5853 llvm::sys::path::remove_filename(Dir); 5854 return false; 5855 } 5856 5857 std::string Driver::GetProgramPath(StringRef Name, const ToolChain &TC) const { 5858 SmallVector<std::string, 2> TargetSpecificExecutables; 5859 generatePrefixedToolNames(Name, TC, TargetSpecificExecutables); 5860 5861 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5862 // attempting to use this prefix when looking for program paths. 5863 for (const auto &PrefixDir : PrefixDirs) { 5864 if (llvm::sys::fs::is_directory(PrefixDir)) { 5865 SmallString<128> P(PrefixDir); 5866 if (ScanDirForExecutable(P, Name)) 5867 return std::string(P.str()); 5868 } else { 5869 SmallString<128> P((PrefixDir + Name).str()); 5870 if (llvm::sys::fs::can_execute(Twine(P))) 5871 return std::string(P.str()); 5872 } 5873 } 5874 5875 const ToolChain::path_list &List = TC.getProgramPaths(); 5876 for (const auto &TargetSpecificExecutable : TargetSpecificExecutables) { 5877 // For each possible name of the tool look for it in 5878 // program paths first, then the path. 5879 // Higher priority names will be first, meaning that 5880 // a higher priority name in the path will be found 5881 // instead of a lower priority name in the program path. 5882 // E.g. <triple>-gcc on the path will be found instead 5883 // of gcc in the program path 5884 for (const auto &Path : List) { 5885 SmallString<128> P(Path); 5886 if (ScanDirForExecutable(P, TargetSpecificExecutable)) 5887 return std::string(P.str()); 5888 } 5889 5890 // Fall back to the path 5891 if (llvm::ErrorOr<std::string> P = 5892 llvm::sys::findProgramByName(TargetSpecificExecutable)) 5893 return *P; 5894 } 5895 5896 return std::string(Name); 5897 } 5898 5899 std::string Driver::GetTemporaryPath(StringRef Prefix, StringRef Suffix) const { 5900 SmallString<128> Path; 5901 std::error_code EC = llvm::sys::fs::createTemporaryFile(Prefix, Suffix, Path); 5902 if (EC) { 5903 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5904 return ""; 5905 } 5906 5907 return std::string(Path.str()); 5908 } 5909 5910 std::string Driver::GetTemporaryDirectory(StringRef Prefix) const { 5911 SmallString<128> Path; 5912 std::error_code EC = llvm::sys::fs::createUniqueDirectory(Prefix, Path); 5913 if (EC) { 5914 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5915 return ""; 5916 } 5917 5918 return std::string(Path.str()); 5919 } 5920 5921 std::string Driver::GetClPchPath(Compilation &C, StringRef BaseName) const { 5922 SmallString<128> Output; 5923 if (Arg *FpArg = C.getArgs().getLastArg(options::OPT__SLASH_Fp)) { 5924 // FIXME: If anybody needs it, implement this obscure rule: 5925 // "If you specify a directory without a file name, the default file name 5926 // is VCx0.pch., where x is the major version of Visual C++ in use." 5927 Output = FpArg->getValue(); 5928 5929 // "If you do not specify an extension as part of the path name, an 5930 // extension of .pch is assumed. " 5931 if (!llvm::sys::path::has_extension(Output)) 5932 Output += ".pch"; 5933 } else { 5934 if (Arg *YcArg = C.getArgs().getLastArg(options::OPT__SLASH_Yc)) 5935 Output = YcArg->getValue(); 5936 if (Output.empty()) 5937 Output = BaseName; 5938 llvm::sys::path::replace_extension(Output, ".pch"); 5939 } 5940 return std::string(Output.str()); 5941 } 5942 5943 const ToolChain &Driver::getToolChain(const ArgList &Args, 5944 const llvm::Triple &Target) const { 5945 5946 auto &TC = ToolChains[Target.str()]; 5947 if (!TC) { 5948 switch (Target.getOS()) { 5949 case llvm::Triple::AIX: 5950 TC = std::make_unique<toolchains::AIX>(*this, Target, Args); 5951 break; 5952 case llvm::Triple::Haiku: 5953 TC = std::make_unique<toolchains::Haiku>(*this, Target, Args); 5954 break; 5955 case llvm::Triple::Ananas: 5956 TC = std::make_unique<toolchains::Ananas>(*this, Target, Args); 5957 break; 5958 case llvm::Triple::CloudABI: 5959 TC = std::make_unique<toolchains::CloudABI>(*this, Target, Args); 5960 break; 5961 case llvm::Triple::Darwin: 5962 case llvm::Triple::MacOSX: 5963 case llvm::Triple::IOS: 5964 case llvm::Triple::TvOS: 5965 case llvm::Triple::WatchOS: 5966 case llvm::Triple::DriverKit: 5967 TC = std::make_unique<toolchains::DarwinClang>(*this, Target, Args); 5968 break; 5969 case llvm::Triple::DragonFly: 5970 TC = std::make_unique<toolchains::DragonFly>(*this, Target, Args); 5971 break; 5972 case llvm::Triple::OpenBSD: 5973 TC = std::make_unique<toolchains::OpenBSD>(*this, Target, Args); 5974 break; 5975 case llvm::Triple::NetBSD: 5976 TC = std::make_unique<toolchains::NetBSD>(*this, Target, Args); 5977 break; 5978 case llvm::Triple::FreeBSD: 5979 if (Target.isPPC()) 5980 TC = std::make_unique<toolchains::PPCFreeBSDToolChain>(*this, Target, 5981 Args); 5982 else 5983 TC = std::make_unique<toolchains::FreeBSD>(*this, Target, Args); 5984 break; 5985 case llvm::Triple::Minix: 5986 TC = std::make_unique<toolchains::Minix>(*this, Target, Args); 5987 break; 5988 case llvm::Triple::Linux: 5989 case llvm::Triple::ELFIAMCU: 5990 if (Target.getArch() == llvm::Triple::hexagon) 5991 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 5992 Args); 5993 else if ((Target.getVendor() == llvm::Triple::MipsTechnologies) && 5994 !Target.hasEnvironment()) 5995 TC = std::make_unique<toolchains::MipsLLVMToolChain>(*this, Target, 5996 Args); 5997 else if (Target.isPPC()) 5998 TC = std::make_unique<toolchains::PPCLinuxToolChain>(*this, Target, 5999 Args); 6000 else if (Target.getArch() == llvm::Triple::ve) 6001 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6002 6003 else 6004 TC = std::make_unique<toolchains::Linux>(*this, Target, Args); 6005 break; 6006 case llvm::Triple::NaCl: 6007 TC = std::make_unique<toolchains::NaClToolChain>(*this, Target, Args); 6008 break; 6009 case llvm::Triple::Fuchsia: 6010 TC = std::make_unique<toolchains::Fuchsia>(*this, Target, Args); 6011 break; 6012 case llvm::Triple::Solaris: 6013 TC = std::make_unique<toolchains::Solaris>(*this, Target, Args); 6014 break; 6015 case llvm::Triple::AMDHSA: 6016 TC = std::make_unique<toolchains::ROCMToolChain>(*this, Target, Args); 6017 break; 6018 case llvm::Triple::AMDPAL: 6019 case llvm::Triple::Mesa3D: 6020 TC = std::make_unique<toolchains::AMDGPUToolChain>(*this, Target, Args); 6021 break; 6022 case llvm::Triple::Win32: 6023 switch (Target.getEnvironment()) { 6024 default: 6025 if (Target.isOSBinFormatELF()) 6026 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6027 else if (Target.isOSBinFormatMachO()) 6028 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6029 else 6030 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6031 break; 6032 case llvm::Triple::GNU: 6033 TC = std::make_unique<toolchains::MinGW>(*this, Target, Args); 6034 break; 6035 case llvm::Triple::Itanium: 6036 TC = std::make_unique<toolchains::CrossWindowsToolChain>(*this, Target, 6037 Args); 6038 break; 6039 case llvm::Triple::MSVC: 6040 case llvm::Triple::UnknownEnvironment: 6041 if (Args.getLastArgValue(options::OPT_fuse_ld_EQ) 6042 .startswith_insensitive("bfd")) 6043 TC = std::make_unique<toolchains::CrossWindowsToolChain>( 6044 *this, Target, Args); 6045 else 6046 TC = 6047 std::make_unique<toolchains::MSVCToolChain>(*this, Target, Args); 6048 break; 6049 } 6050 break; 6051 case llvm::Triple::PS4: 6052 TC = std::make_unique<toolchains::PS4CPU>(*this, Target, Args); 6053 break; 6054 case llvm::Triple::PS5: 6055 TC = std::make_unique<toolchains::PS5CPU>(*this, Target, Args); 6056 break; 6057 case llvm::Triple::Contiki: 6058 TC = std::make_unique<toolchains::Contiki>(*this, Target, Args); 6059 break; 6060 case llvm::Triple::Hurd: 6061 TC = std::make_unique<toolchains::Hurd>(*this, Target, Args); 6062 break; 6063 case llvm::Triple::ZOS: 6064 TC = std::make_unique<toolchains::ZOS>(*this, Target, Args); 6065 break; 6066 case llvm::Triple::ShaderModel: 6067 TC = std::make_unique<toolchains::HLSLToolChain>(*this, Target, Args); 6068 break; 6069 default: 6070 // Of these targets, Hexagon is the only one that might have 6071 // an OS of Linux, in which case it got handled above already. 6072 switch (Target.getArch()) { 6073 case llvm::Triple::tce: 6074 TC = std::make_unique<toolchains::TCEToolChain>(*this, Target, Args); 6075 break; 6076 case llvm::Triple::tcele: 6077 TC = std::make_unique<toolchains::TCELEToolChain>(*this, Target, Args); 6078 break; 6079 case llvm::Triple::hexagon: 6080 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 6081 Args); 6082 break; 6083 case llvm::Triple::lanai: 6084 TC = std::make_unique<toolchains::LanaiToolChain>(*this, Target, Args); 6085 break; 6086 case llvm::Triple::xcore: 6087 TC = std::make_unique<toolchains::XCoreToolChain>(*this, Target, Args); 6088 break; 6089 case llvm::Triple::wasm32: 6090 case llvm::Triple::wasm64: 6091 TC = std::make_unique<toolchains::WebAssembly>(*this, Target, Args); 6092 break; 6093 case llvm::Triple::avr: 6094 TC = std::make_unique<toolchains::AVRToolChain>(*this, Target, Args); 6095 break; 6096 case llvm::Triple::msp430: 6097 TC = 6098 std::make_unique<toolchains::MSP430ToolChain>(*this, Target, Args); 6099 break; 6100 case llvm::Triple::riscv32: 6101 case llvm::Triple::riscv64: 6102 if (toolchains::RISCVToolChain::hasGCCToolchain(*this, Args)) 6103 TC = 6104 std::make_unique<toolchains::RISCVToolChain>(*this, Target, Args); 6105 else 6106 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6107 break; 6108 case llvm::Triple::ve: 6109 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6110 break; 6111 case llvm::Triple::spirv32: 6112 case llvm::Triple::spirv64: 6113 TC = std::make_unique<toolchains::SPIRVToolChain>(*this, Target, Args); 6114 break; 6115 case llvm::Triple::csky: 6116 TC = std::make_unique<toolchains::CSKYToolChain>(*this, Target, Args); 6117 break; 6118 default: 6119 if (Target.getVendor() == llvm::Triple::Myriad) 6120 TC = std::make_unique<toolchains::MyriadToolChain>(*this, Target, 6121 Args); 6122 else if (toolchains::BareMetal::handlesTarget(Target)) 6123 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6124 else if (Target.isOSBinFormatELF()) 6125 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6126 else if (Target.isOSBinFormatMachO()) 6127 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6128 else 6129 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6130 } 6131 } 6132 } 6133 6134 // Intentionally omitted from the switch above: llvm::Triple::CUDA. CUDA 6135 // compiles always need two toolchains, the CUDA toolchain and the host 6136 // toolchain. So the only valid way to create a CUDA toolchain is via 6137 // CreateOffloadingDeviceToolChains. 6138 6139 return *TC; 6140 } 6141 6142 const ToolChain &Driver::getOffloadingDeviceToolChain( 6143 const ArgList &Args, const llvm::Triple &Target, const ToolChain &HostTC, 6144 const Action::OffloadKind &TargetDeviceOffloadKind) const { 6145 // Use device / host triples as the key into the ToolChains map because the 6146 // device ToolChain we create depends on both. 6147 auto &TC = ToolChains[Target.str() + "/" + HostTC.getTriple().str()]; 6148 if (!TC) { 6149 // Categorized by offload kind > arch rather than OS > arch like 6150 // the normal getToolChain call, as it seems a reasonable way to categorize 6151 // things. 6152 switch (TargetDeviceOffloadKind) { 6153 case Action::OFK_HIP: { 6154 if (Target.getArch() == llvm::Triple::amdgcn && 6155 Target.getVendor() == llvm::Triple::AMD && 6156 Target.getOS() == llvm::Triple::AMDHSA) 6157 TC = std::make_unique<toolchains::HIPAMDToolChain>(*this, Target, 6158 HostTC, Args); 6159 else if (Target.getArch() == llvm::Triple::spirv64 && 6160 Target.getVendor() == llvm::Triple::UnknownVendor && 6161 Target.getOS() == llvm::Triple::UnknownOS) 6162 TC = std::make_unique<toolchains::HIPSPVToolChain>(*this, Target, 6163 HostTC, Args); 6164 break; 6165 } 6166 default: 6167 break; 6168 } 6169 } 6170 6171 return *TC; 6172 } 6173 6174 bool Driver::ShouldUseClangCompiler(const JobAction &JA) const { 6175 // Say "no" if there is not exactly one input of a type clang understands. 6176 if (JA.size() != 1 || 6177 !types::isAcceptedByClang((*JA.input_begin())->getType())) 6178 return false; 6179 6180 // And say "no" if this is not a kind of action clang understands. 6181 if (!isa<PreprocessJobAction>(JA) && !isa<PrecompileJobAction>(JA) && 6182 !isa<CompileJobAction>(JA) && !isa<BackendJobAction>(JA) && 6183 !isa<ExtractAPIJobAction>(JA)) 6184 return false; 6185 6186 return true; 6187 } 6188 6189 bool Driver::ShouldUseFlangCompiler(const JobAction &JA) const { 6190 // Say "no" if there is not exactly one input of a type flang understands. 6191 if (JA.size() != 1 || 6192 !types::isAcceptedByFlang((*JA.input_begin())->getType())) 6193 return false; 6194 6195 // And say "no" if this is not a kind of action flang understands. 6196 if (!isa<PreprocessJobAction>(JA) && !isa<CompileJobAction>(JA) && 6197 !isa<BackendJobAction>(JA)) 6198 return false; 6199 6200 return true; 6201 } 6202 6203 bool Driver::ShouldEmitStaticLibrary(const ArgList &Args) const { 6204 // Only emit static library if the flag is set explicitly. 6205 if (Args.hasArg(options::OPT_emit_static_lib)) 6206 return true; 6207 return false; 6208 } 6209 6210 /// GetReleaseVersion - Parse (([0-9]+)(.([0-9]+)(.([0-9]+)?))?)? and return the 6211 /// grouped values as integers. Numbers which are not provided are set to 0. 6212 /// 6213 /// \return True if the entire string was parsed (9.2), or all groups were 6214 /// parsed (10.3.5extrastuff). 6215 bool Driver::GetReleaseVersion(StringRef Str, unsigned &Major, unsigned &Minor, 6216 unsigned &Micro, bool &HadExtra) { 6217 HadExtra = false; 6218 6219 Major = Minor = Micro = 0; 6220 if (Str.empty()) 6221 return false; 6222 6223 if (Str.consumeInteger(10, Major)) 6224 return false; 6225 if (Str.empty()) 6226 return true; 6227 if (Str[0] != '.') 6228 return false; 6229 6230 Str = Str.drop_front(1); 6231 6232 if (Str.consumeInteger(10, Minor)) 6233 return false; 6234 if (Str.empty()) 6235 return true; 6236 if (Str[0] != '.') 6237 return false; 6238 Str = Str.drop_front(1); 6239 6240 if (Str.consumeInteger(10, Micro)) 6241 return false; 6242 if (!Str.empty()) 6243 HadExtra = true; 6244 return true; 6245 } 6246 6247 /// Parse digits from a string \p Str and fulfill \p Digits with 6248 /// the parsed numbers. This method assumes that the max number of 6249 /// digits to look for is equal to Digits.size(). 6250 /// 6251 /// \return True if the entire string was parsed and there are 6252 /// no extra characters remaining at the end. 6253 bool Driver::GetReleaseVersion(StringRef Str, 6254 MutableArrayRef<unsigned> Digits) { 6255 if (Str.empty()) 6256 return false; 6257 6258 unsigned CurDigit = 0; 6259 while (CurDigit < Digits.size()) { 6260 unsigned Digit; 6261 if (Str.consumeInteger(10, Digit)) 6262 return false; 6263 Digits[CurDigit] = Digit; 6264 if (Str.empty()) 6265 return true; 6266 if (Str[0] != '.') 6267 return false; 6268 Str = Str.drop_front(1); 6269 CurDigit++; 6270 } 6271 6272 // More digits than requested, bail out... 6273 return false; 6274 } 6275 6276 std::pair<unsigned, unsigned> 6277 Driver::getIncludeExcludeOptionFlagMasks(bool IsClCompatMode) const { 6278 unsigned IncludedFlagsBitmask = 0; 6279 unsigned ExcludedFlagsBitmask = options::NoDriverOption; 6280 6281 if (IsClCompatMode) { 6282 // Include CL and Core options. 6283 IncludedFlagsBitmask |= options::CLOption; 6284 IncludedFlagsBitmask |= options::CLDXCOption; 6285 IncludedFlagsBitmask |= options::CoreOption; 6286 } else { 6287 ExcludedFlagsBitmask |= options::CLOption; 6288 } 6289 if (IsDXCMode()) { 6290 // Include DXC and Core options. 6291 IncludedFlagsBitmask |= options::DXCOption; 6292 IncludedFlagsBitmask |= options::CLDXCOption; 6293 IncludedFlagsBitmask |= options::CoreOption; 6294 } else { 6295 ExcludedFlagsBitmask |= options::DXCOption; 6296 } 6297 if (!IsClCompatMode && !IsDXCMode()) 6298 ExcludedFlagsBitmask |= options::CLDXCOption; 6299 6300 return std::make_pair(IncludedFlagsBitmask, ExcludedFlagsBitmask); 6301 } 6302 6303 bool clang::driver::isOptimizationLevelFast(const ArgList &Args) { 6304 return Args.hasFlag(options::OPT_Ofast, options::OPT_O_Group, false); 6305 } 6306 6307 bool clang::driver::willEmitRemarks(const ArgList &Args) { 6308 // -fsave-optimization-record enables it. 6309 if (Args.hasFlag(options::OPT_fsave_optimization_record, 6310 options::OPT_fno_save_optimization_record, false)) 6311 return true; 6312 6313 // -fsave-optimization-record=<format> enables it as well. 6314 if (Args.hasFlag(options::OPT_fsave_optimization_record_EQ, 6315 options::OPT_fno_save_optimization_record, false)) 6316 return true; 6317 6318 // -foptimization-record-file alone enables it too. 6319 if (Args.hasFlag(options::OPT_foptimization_record_file_EQ, 6320 options::OPT_fno_save_optimization_record, false)) 6321 return true; 6322 6323 // -foptimization-record-passes alone enables it too. 6324 if (Args.hasFlag(options::OPT_foptimization_record_passes_EQ, 6325 options::OPT_fno_save_optimization_record, false)) 6326 return true; 6327 return false; 6328 } 6329 6330 llvm::StringRef clang::driver::getDriverMode(StringRef ProgName, 6331 ArrayRef<const char *> Args) { 6332 static const std::string OptName = 6333 getDriverOptTable().getOption(options::OPT_driver_mode).getPrefixedName(); 6334 llvm::StringRef Opt; 6335 for (StringRef Arg : Args) { 6336 if (!Arg.startswith(OptName)) 6337 continue; 6338 Opt = Arg; 6339 } 6340 if (Opt.empty()) 6341 Opt = ToolChain::getTargetAndModeFromProgramName(ProgName).DriverMode; 6342 return Opt.consume_front(OptName) ? Opt : ""; 6343 } 6344 6345 bool driver::IsClangCL(StringRef DriverMode) { return DriverMode.equals("cl"); } 6346