1 //===--- Driver.cpp - Clang GCC Compatible Driver -------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "clang/Driver/Driver.h" 10 #include "ToolChains/AIX.h" 11 #include "ToolChains/AMDGPU.h" 12 #include "ToolChains/AMDGPUOpenMP.h" 13 #include "ToolChains/AVR.h" 14 #include "ToolChains/Ananas.h" 15 #include "ToolChains/BareMetal.h" 16 #include "ToolChains/CSKYToolChain.h" 17 #include "ToolChains/Clang.h" 18 #include "ToolChains/CloudABI.h" 19 #include "ToolChains/Contiki.h" 20 #include "ToolChains/CrossWindows.h" 21 #include "ToolChains/Cuda.h" 22 #include "ToolChains/Darwin.h" 23 #include "ToolChains/DragonFly.h" 24 #include "ToolChains/FreeBSD.h" 25 #include "ToolChains/Fuchsia.h" 26 #include "ToolChains/Gnu.h" 27 #include "ToolChains/HIPAMD.h" 28 #include "ToolChains/HIPSPV.h" 29 #include "ToolChains/HLSL.h" 30 #include "ToolChains/Haiku.h" 31 #include "ToolChains/Hexagon.h" 32 #include "ToolChains/Hurd.h" 33 #include "ToolChains/Lanai.h" 34 #include "ToolChains/Linux.h" 35 #include "ToolChains/MSP430.h" 36 #include "ToolChains/MSVC.h" 37 #include "ToolChains/MinGW.h" 38 #include "ToolChains/Minix.h" 39 #include "ToolChains/MipsLinux.h" 40 #include "ToolChains/Myriad.h" 41 #include "ToolChains/NaCl.h" 42 #include "ToolChains/NetBSD.h" 43 #include "ToolChains/OpenBSD.h" 44 #include "ToolChains/PPCFreeBSD.h" 45 #include "ToolChains/PPCLinux.h" 46 #include "ToolChains/PS4CPU.h" 47 #include "ToolChains/RISCVToolchain.h" 48 #include "ToolChains/SPIRV.h" 49 #include "ToolChains/Solaris.h" 50 #include "ToolChains/TCE.h" 51 #include "ToolChains/VEToolchain.h" 52 #include "ToolChains/WebAssembly.h" 53 #include "ToolChains/XCore.h" 54 #include "ToolChains/ZOS.h" 55 #include "clang/Basic/TargetID.h" 56 #include "clang/Basic/Version.h" 57 #include "clang/Config/config.h" 58 #include "clang/Driver/Action.h" 59 #include "clang/Driver/Compilation.h" 60 #include "clang/Driver/DriverDiagnostic.h" 61 #include "clang/Driver/InputInfo.h" 62 #include "clang/Driver/Job.h" 63 #include "clang/Driver/Options.h" 64 #include "clang/Driver/Phases.h" 65 #include "clang/Driver/SanitizerArgs.h" 66 #include "clang/Driver/Tool.h" 67 #include "clang/Driver/ToolChain.h" 68 #include "clang/Driver/Types.h" 69 #include "llvm/ADT/ArrayRef.h" 70 #include "llvm/ADT/STLExtras.h" 71 #include "llvm/ADT/SmallSet.h" 72 #include "llvm/ADT/StringExtras.h" 73 #include "llvm/ADT/StringRef.h" 74 #include "llvm/ADT/StringSet.h" 75 #include "llvm/ADT/StringSwitch.h" 76 #include "llvm/Config/llvm-config.h" 77 #include "llvm/MC/TargetRegistry.h" 78 #include "llvm/Option/Arg.h" 79 #include "llvm/Option/ArgList.h" 80 #include "llvm/Option/OptSpecifier.h" 81 #include "llvm/Option/OptTable.h" 82 #include "llvm/Option/Option.h" 83 #include "llvm/Support/CommandLine.h" 84 #include "llvm/Support/ErrorHandling.h" 85 #include "llvm/Support/ExitCodes.h" 86 #include "llvm/Support/FileSystem.h" 87 #include "llvm/Support/FormatVariadic.h" 88 #include "llvm/Support/Host.h" 89 #include "llvm/Support/MD5.h" 90 #include "llvm/Support/Path.h" 91 #include "llvm/Support/PrettyStackTrace.h" 92 #include "llvm/Support/Process.h" 93 #include "llvm/Support/Program.h" 94 #include "llvm/Support/StringSaver.h" 95 #include "llvm/Support/VirtualFileSystem.h" 96 #include "llvm/Support/raw_ostream.h" 97 #include <map> 98 #include <memory> 99 #include <utility> 100 #if LLVM_ON_UNIX 101 #include <unistd.h> // getpid 102 #endif 103 104 using namespace clang::driver; 105 using namespace clang; 106 using namespace llvm::opt; 107 108 static llvm::Optional<llvm::Triple> 109 getOffloadTargetTriple(const Driver &D, const ArgList &Args) { 110 auto OffloadTargets = Args.getAllArgValues(options::OPT_offload_EQ); 111 // Offload compilation flow does not support multiple targets for now. We 112 // need the HIPActionBuilder (and possibly the CudaActionBuilder{,Base}too) 113 // to support multiple tool chains first. 114 switch (OffloadTargets.size()) { 115 default: 116 D.Diag(diag::err_drv_only_one_offload_target_supported); 117 return llvm::None; 118 case 0: 119 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << ""; 120 return llvm::None; 121 case 1: 122 break; 123 } 124 return llvm::Triple(OffloadTargets[0]); 125 } 126 127 static llvm::Optional<llvm::Triple> 128 getNVIDIAOffloadTargetTriple(const Driver &D, const ArgList &Args, 129 const llvm::Triple &HostTriple) { 130 if (!Args.hasArg(options::OPT_offload_EQ)) { 131 return llvm::Triple(HostTriple.isArch64Bit() ? "nvptx64-nvidia-cuda" 132 : "nvptx-nvidia-cuda"); 133 } 134 auto TT = getOffloadTargetTriple(D, Args); 135 if (TT && (TT->getArch() == llvm::Triple::spirv32 || 136 TT->getArch() == llvm::Triple::spirv64)) { 137 if (Args.hasArg(options::OPT_emit_llvm)) 138 return TT; 139 D.Diag(diag::err_drv_cuda_offload_only_emit_bc); 140 return llvm::None; 141 } 142 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 143 return llvm::None; 144 } 145 static llvm::Optional<llvm::Triple> 146 getHIPOffloadTargetTriple(const Driver &D, const ArgList &Args) { 147 if (!Args.hasArg(options::OPT_offload_EQ)) { 148 return llvm::Triple("amdgcn-amd-amdhsa"); // Default HIP triple. 149 } 150 auto TT = getOffloadTargetTriple(D, Args); 151 if (!TT) 152 return llvm::None; 153 if (TT->getArch() == llvm::Triple::amdgcn && 154 TT->getVendor() == llvm::Triple::AMD && 155 TT->getOS() == llvm::Triple::AMDHSA) 156 return TT; 157 if (TT->getArch() == llvm::Triple::spirv64) 158 return TT; 159 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 160 return llvm::None; 161 } 162 163 // static 164 std::string Driver::GetResourcesPath(StringRef BinaryPath, 165 StringRef CustomResourceDir) { 166 // Since the resource directory is embedded in the module hash, it's important 167 // that all places that need it call this function, so that they get the 168 // exact same string ("a/../b/" and "b/" get different hashes, for example). 169 170 // Dir is bin/ or lib/, depending on where BinaryPath is. 171 std::string Dir = std::string(llvm::sys::path::parent_path(BinaryPath)); 172 173 SmallString<128> P(Dir); 174 if (CustomResourceDir != "") { 175 llvm::sys::path::append(P, CustomResourceDir); 176 } else { 177 // On Windows, libclang.dll is in bin/. 178 // On non-Windows, libclang.so/.dylib is in lib/. 179 // With a static-library build of libclang, LibClangPath will contain the 180 // path of the embedding binary, which for LLVM binaries will be in bin/. 181 // ../lib gets us to lib/ in both cases. 182 P = llvm::sys::path::parent_path(Dir); 183 llvm::sys::path::append(P, Twine("lib") + CLANG_LIBDIR_SUFFIX, "clang", 184 CLANG_VERSION_STRING); 185 } 186 187 return std::string(P.str()); 188 } 189 190 Driver::Driver(StringRef ClangExecutable, StringRef TargetTriple, 191 DiagnosticsEngine &Diags, std::string Title, 192 IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS) 193 : Diags(Diags), VFS(std::move(VFS)), Mode(GCCMode), 194 SaveTemps(SaveTempsNone), BitcodeEmbed(EmbedNone), 195 CXX20HeaderType(HeaderMode_None), ModulesModeCXX20(false), 196 LTOMode(LTOK_None), ClangExecutable(ClangExecutable), 197 SysRoot(DEFAULT_SYSROOT), DriverTitle(Title), CCCPrintBindings(false), 198 CCPrintOptions(false), CCPrintHeaders(false), CCLogDiagnostics(false), 199 CCGenDiagnostics(false), CCPrintProcessStats(false), 200 TargetTriple(TargetTriple), Saver(Alloc), CheckInputsExist(true), 201 ProbePrecompiled(true), SuppressMissingInputWarning(false) { 202 // Provide a sane fallback if no VFS is specified. 203 if (!this->VFS) 204 this->VFS = llvm::vfs::getRealFileSystem(); 205 206 Name = std::string(llvm::sys::path::filename(ClangExecutable)); 207 Dir = std::string(llvm::sys::path::parent_path(ClangExecutable)); 208 InstalledDir = Dir; // Provide a sensible default installed dir. 209 210 if ((!SysRoot.empty()) && llvm::sys::path::is_relative(SysRoot)) { 211 // Prepend InstalledDir if SysRoot is relative 212 SmallString<128> P(InstalledDir); 213 llvm::sys::path::append(P, SysRoot); 214 SysRoot = std::string(P); 215 } 216 217 #if defined(CLANG_CONFIG_FILE_SYSTEM_DIR) 218 SystemConfigDir = CLANG_CONFIG_FILE_SYSTEM_DIR; 219 #endif 220 #if defined(CLANG_CONFIG_FILE_USER_DIR) 221 UserConfigDir = CLANG_CONFIG_FILE_USER_DIR; 222 #endif 223 224 // Compute the path to the resource directory. 225 ResourceDir = GetResourcesPath(ClangExecutable, CLANG_RESOURCE_DIR); 226 } 227 228 void Driver::setDriverMode(StringRef Value) { 229 static const std::string OptName = 230 getOpts().getOption(options::OPT_driver_mode).getPrefixedName(); 231 if (auto M = llvm::StringSwitch<llvm::Optional<DriverMode>>(Value) 232 .Case("gcc", GCCMode) 233 .Case("g++", GXXMode) 234 .Case("cpp", CPPMode) 235 .Case("cl", CLMode) 236 .Case("flang", FlangMode) 237 .Case("dxc", DXCMode) 238 .Default(None)) 239 Mode = *M; 240 else 241 Diag(diag::err_drv_unsupported_option_argument) << OptName << Value; 242 } 243 244 InputArgList Driver::ParseArgStrings(ArrayRef<const char *> ArgStrings, 245 bool IsClCompatMode, 246 bool &ContainsError) { 247 llvm::PrettyStackTraceString CrashInfo("Command line argument parsing"); 248 ContainsError = false; 249 250 unsigned IncludedFlagsBitmask; 251 unsigned ExcludedFlagsBitmask; 252 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 253 getIncludeExcludeOptionFlagMasks(IsClCompatMode); 254 255 // Make sure that Flang-only options don't pollute the Clang output 256 // TODO: Make sure that Clang-only options don't pollute Flang output 257 if (!IsFlangMode()) 258 ExcludedFlagsBitmask |= options::FlangOnlyOption; 259 260 unsigned MissingArgIndex, MissingArgCount; 261 InputArgList Args = 262 getOpts().ParseArgs(ArgStrings, MissingArgIndex, MissingArgCount, 263 IncludedFlagsBitmask, ExcludedFlagsBitmask); 264 265 // Check for missing argument error. 266 if (MissingArgCount) { 267 Diag(diag::err_drv_missing_argument) 268 << Args.getArgString(MissingArgIndex) << MissingArgCount; 269 ContainsError |= 270 Diags.getDiagnosticLevel(diag::err_drv_missing_argument, 271 SourceLocation()) > DiagnosticsEngine::Warning; 272 } 273 274 // Check for unsupported options. 275 for (const Arg *A : Args) { 276 if (A->getOption().hasFlag(options::Unsupported)) { 277 unsigned DiagID; 278 auto ArgString = A->getAsString(Args); 279 std::string Nearest; 280 if (getOpts().findNearest( 281 ArgString, Nearest, IncludedFlagsBitmask, 282 ExcludedFlagsBitmask | options::Unsupported) > 1) { 283 DiagID = diag::err_drv_unsupported_opt; 284 Diag(DiagID) << ArgString; 285 } else { 286 DiagID = diag::err_drv_unsupported_opt_with_suggestion; 287 Diag(DiagID) << ArgString << Nearest; 288 } 289 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 290 DiagnosticsEngine::Warning; 291 continue; 292 } 293 294 // Warn about -mcpu= without an argument. 295 if (A->getOption().matches(options::OPT_mcpu_EQ) && A->containsValue("")) { 296 Diag(diag::warn_drv_empty_joined_argument) << A->getAsString(Args); 297 ContainsError |= Diags.getDiagnosticLevel( 298 diag::warn_drv_empty_joined_argument, 299 SourceLocation()) > DiagnosticsEngine::Warning; 300 } 301 } 302 303 for (const Arg *A : Args.filtered(options::OPT_UNKNOWN)) { 304 unsigned DiagID; 305 auto ArgString = A->getAsString(Args); 306 std::string Nearest; 307 if (getOpts().findNearest( 308 ArgString, Nearest, IncludedFlagsBitmask, ExcludedFlagsBitmask) > 1) { 309 DiagID = IsCLMode() ? diag::warn_drv_unknown_argument_clang_cl 310 : diag::err_drv_unknown_argument; 311 Diags.Report(DiagID) << ArgString; 312 } else { 313 DiagID = IsCLMode() 314 ? diag::warn_drv_unknown_argument_clang_cl_with_suggestion 315 : diag::err_drv_unknown_argument_with_suggestion; 316 Diags.Report(DiagID) << ArgString << Nearest; 317 } 318 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 319 DiagnosticsEngine::Warning; 320 } 321 322 return Args; 323 } 324 325 // Determine which compilation mode we are in. We look for options which 326 // affect the phase, starting with the earliest phases, and record which 327 // option we used to determine the final phase. 328 phases::ID Driver::getFinalPhase(const DerivedArgList &DAL, 329 Arg **FinalPhaseArg) const { 330 Arg *PhaseArg = nullptr; 331 phases::ID FinalPhase; 332 333 // -{E,EP,P,M,MM} only run the preprocessor. 334 if (CCCIsCPP() || (PhaseArg = DAL.getLastArg(options::OPT_E)) || 335 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_EP)) || 336 (PhaseArg = DAL.getLastArg(options::OPT_M, options::OPT_MM)) || 337 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_P)) || 338 CCGenDiagnostics) { 339 FinalPhase = phases::Preprocess; 340 341 // --precompile only runs up to precompilation. 342 // Options that cause the output of C++20 compiled module interfaces or 343 // header units have the same effect. 344 } else if ((PhaseArg = DAL.getLastArg(options::OPT__precompile)) || 345 (PhaseArg = DAL.getLastArg(options::OPT_extract_api)) || 346 (PhaseArg = DAL.getLastArg(options::OPT_fmodule_header, 347 options::OPT_fmodule_header_EQ))) { 348 FinalPhase = phases::Precompile; 349 // -{fsyntax-only,-analyze,emit-ast} only run up to the compiler. 350 } else if ((PhaseArg = DAL.getLastArg(options::OPT_fsyntax_only)) || 351 (PhaseArg = DAL.getLastArg(options::OPT_print_supported_cpus)) || 352 (PhaseArg = DAL.getLastArg(options::OPT_module_file_info)) || 353 (PhaseArg = DAL.getLastArg(options::OPT_verify_pch)) || 354 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_objc)) || 355 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_legacy_objc)) || 356 (PhaseArg = DAL.getLastArg(options::OPT__migrate)) || 357 (PhaseArg = DAL.getLastArg(options::OPT__analyze)) || 358 (PhaseArg = DAL.getLastArg(options::OPT_emit_ast))) { 359 FinalPhase = phases::Compile; 360 361 // -S only runs up to the backend. 362 } else if ((PhaseArg = DAL.getLastArg(options::OPT_S))) { 363 FinalPhase = phases::Backend; 364 365 // -c compilation only runs up to the assembler. 366 } else if ((PhaseArg = DAL.getLastArg(options::OPT_c))) { 367 FinalPhase = phases::Assemble; 368 369 } else if ((PhaseArg = DAL.getLastArg(options::OPT_emit_interface_stubs))) { 370 FinalPhase = phases::IfsMerge; 371 372 // Otherwise do everything. 373 } else 374 FinalPhase = phases::Link; 375 376 if (FinalPhaseArg) 377 *FinalPhaseArg = PhaseArg; 378 379 return FinalPhase; 380 } 381 382 static Arg *MakeInputArg(DerivedArgList &Args, const OptTable &Opts, 383 StringRef Value, bool Claim = true) { 384 Arg *A = new Arg(Opts.getOption(options::OPT_INPUT), Value, 385 Args.getBaseArgs().MakeIndex(Value), Value.data()); 386 Args.AddSynthesizedArg(A); 387 if (Claim) 388 A->claim(); 389 return A; 390 } 391 392 DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args) const { 393 const llvm::opt::OptTable &Opts = getOpts(); 394 DerivedArgList *DAL = new DerivedArgList(Args); 395 396 bool HasNostdlib = Args.hasArg(options::OPT_nostdlib); 397 bool HasNostdlibxx = Args.hasArg(options::OPT_nostdlibxx); 398 bool HasNodefaultlib = Args.hasArg(options::OPT_nodefaultlibs); 399 bool IgnoreUnused = false; 400 for (Arg *A : Args) { 401 if (IgnoreUnused) 402 A->claim(); 403 404 if (A->getOption().matches(options::OPT_start_no_unused_arguments)) { 405 IgnoreUnused = true; 406 continue; 407 } 408 if (A->getOption().matches(options::OPT_end_no_unused_arguments)) { 409 IgnoreUnused = false; 410 continue; 411 } 412 413 // Unfortunately, we have to parse some forwarding options (-Xassembler, 414 // -Xlinker, -Xpreprocessor) because we either integrate their functionality 415 // (assembler and preprocessor), or bypass a previous driver ('collect2'). 416 417 // Rewrite linker options, to replace --no-demangle with a custom internal 418 // option. 419 if ((A->getOption().matches(options::OPT_Wl_COMMA) || 420 A->getOption().matches(options::OPT_Xlinker)) && 421 A->containsValue("--no-demangle")) { 422 // Add the rewritten no-demangle argument. 423 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_Xlinker__no_demangle)); 424 425 // Add the remaining values as Xlinker arguments. 426 for (StringRef Val : A->getValues()) 427 if (Val != "--no-demangle") 428 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_Xlinker), Val); 429 430 continue; 431 } 432 433 // Rewrite preprocessor options, to replace -Wp,-MD,FOO which is used by 434 // some build systems. We don't try to be complete here because we don't 435 // care to encourage this usage model. 436 if (A->getOption().matches(options::OPT_Wp_COMMA) && 437 (A->getValue(0) == StringRef("-MD") || 438 A->getValue(0) == StringRef("-MMD"))) { 439 // Rewrite to -MD/-MMD along with -MF. 440 if (A->getValue(0) == StringRef("-MD")) 441 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MD)); 442 else 443 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MMD)); 444 if (A->getNumValues() == 2) 445 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_MF), A->getValue(1)); 446 continue; 447 } 448 449 // Rewrite reserved library names. 450 if (A->getOption().matches(options::OPT_l)) { 451 StringRef Value = A->getValue(); 452 453 // Rewrite unless -nostdlib is present. 454 if (!HasNostdlib && !HasNodefaultlib && !HasNostdlibxx && 455 Value == "stdc++") { 456 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_stdcxx)); 457 continue; 458 } 459 460 // Rewrite unconditionally. 461 if (Value == "cc_kext") { 462 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_cckext)); 463 continue; 464 } 465 } 466 467 // Pick up inputs via the -- option. 468 if (A->getOption().matches(options::OPT__DASH_DASH)) { 469 A->claim(); 470 for (StringRef Val : A->getValues()) 471 DAL->append(MakeInputArg(*DAL, Opts, Val, false)); 472 continue; 473 } 474 475 DAL->append(A); 476 } 477 478 // Enforce -static if -miamcu is present. 479 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) 480 DAL->AddFlagArg(nullptr, Opts.getOption(options::OPT_static)); 481 482 // Add a default value of -mlinker-version=, if one was given and the user 483 // didn't specify one. 484 #if defined(HOST_LINK_VERSION) 485 if (!Args.hasArg(options::OPT_mlinker_version_EQ) && 486 strlen(HOST_LINK_VERSION) > 0) { 487 DAL->AddJoinedArg(0, Opts.getOption(options::OPT_mlinker_version_EQ), 488 HOST_LINK_VERSION); 489 DAL->getLastArg(options::OPT_mlinker_version_EQ)->claim(); 490 } 491 #endif 492 493 return DAL; 494 } 495 496 /// Compute target triple from args. 497 /// 498 /// This routine provides the logic to compute a target triple from various 499 /// args passed to the driver and the default triple string. 500 static llvm::Triple computeTargetTriple(const Driver &D, 501 StringRef TargetTriple, 502 const ArgList &Args, 503 StringRef DarwinArchName = "") { 504 // FIXME: Already done in Compilation *Driver::BuildCompilation 505 if (const Arg *A = Args.getLastArg(options::OPT_target)) 506 TargetTriple = A->getValue(); 507 508 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 509 510 // GNU/Hurd's triples should have been -hurd-gnu*, but were historically made 511 // -gnu* only, and we can not change this, so we have to detect that case as 512 // being the Hurd OS. 513 if (TargetTriple.contains("-unknown-gnu") || TargetTriple.contains("-pc-gnu")) 514 Target.setOSName("hurd"); 515 516 // Handle Apple-specific options available here. 517 if (Target.isOSBinFormatMachO()) { 518 // If an explicit Darwin arch name is given, that trumps all. 519 if (!DarwinArchName.empty()) { 520 tools::darwin::setTripleTypeForMachOArchName(Target, DarwinArchName); 521 return Target; 522 } 523 524 // Handle the Darwin '-arch' flag. 525 if (Arg *A = Args.getLastArg(options::OPT_arch)) { 526 StringRef ArchName = A->getValue(); 527 tools::darwin::setTripleTypeForMachOArchName(Target, ArchName); 528 } 529 } 530 531 // Handle pseudo-target flags '-mlittle-endian'/'-EL' and 532 // '-mbig-endian'/'-EB'. 533 if (Arg *A = Args.getLastArg(options::OPT_mlittle_endian, 534 options::OPT_mbig_endian)) { 535 if (A->getOption().matches(options::OPT_mlittle_endian)) { 536 llvm::Triple LE = Target.getLittleEndianArchVariant(); 537 if (LE.getArch() != llvm::Triple::UnknownArch) 538 Target = std::move(LE); 539 } else { 540 llvm::Triple BE = Target.getBigEndianArchVariant(); 541 if (BE.getArch() != llvm::Triple::UnknownArch) 542 Target = std::move(BE); 543 } 544 } 545 546 // Skip further flag support on OSes which don't support '-m32' or '-m64'. 547 if (Target.getArch() == llvm::Triple::tce || 548 Target.getOS() == llvm::Triple::Minix) 549 return Target; 550 551 // On AIX, the env OBJECT_MODE may affect the resulting arch variant. 552 if (Target.isOSAIX()) { 553 if (Optional<std::string> ObjectModeValue = 554 llvm::sys::Process::GetEnv("OBJECT_MODE")) { 555 StringRef ObjectMode = *ObjectModeValue; 556 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 557 558 if (ObjectMode.equals("64")) { 559 AT = Target.get64BitArchVariant().getArch(); 560 } else if (ObjectMode.equals("32")) { 561 AT = Target.get32BitArchVariant().getArch(); 562 } else { 563 D.Diag(diag::err_drv_invalid_object_mode) << ObjectMode; 564 } 565 566 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) 567 Target.setArch(AT); 568 } 569 } 570 571 // Handle pseudo-target flags '-m64', '-mx32', '-m32' and '-m16'. 572 Arg *A = Args.getLastArg(options::OPT_m64, options::OPT_mx32, 573 options::OPT_m32, options::OPT_m16); 574 if (A) { 575 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 576 577 if (A->getOption().matches(options::OPT_m64)) { 578 AT = Target.get64BitArchVariant().getArch(); 579 if (Target.getEnvironment() == llvm::Triple::GNUX32) 580 Target.setEnvironment(llvm::Triple::GNU); 581 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 582 Target.setEnvironment(llvm::Triple::Musl); 583 } else if (A->getOption().matches(options::OPT_mx32) && 584 Target.get64BitArchVariant().getArch() == llvm::Triple::x86_64) { 585 AT = llvm::Triple::x86_64; 586 if (Target.getEnvironment() == llvm::Triple::Musl) 587 Target.setEnvironment(llvm::Triple::MuslX32); 588 else 589 Target.setEnvironment(llvm::Triple::GNUX32); 590 } else if (A->getOption().matches(options::OPT_m32)) { 591 AT = Target.get32BitArchVariant().getArch(); 592 if (Target.getEnvironment() == llvm::Triple::GNUX32) 593 Target.setEnvironment(llvm::Triple::GNU); 594 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 595 Target.setEnvironment(llvm::Triple::Musl); 596 } else if (A->getOption().matches(options::OPT_m16) && 597 Target.get32BitArchVariant().getArch() == llvm::Triple::x86) { 598 AT = llvm::Triple::x86; 599 Target.setEnvironment(llvm::Triple::CODE16); 600 } 601 602 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) { 603 Target.setArch(AT); 604 if (Target.isWindowsGNUEnvironment()) 605 toolchains::MinGW::fixTripleArch(D, Target, Args); 606 } 607 } 608 609 // Handle -miamcu flag. 610 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) { 611 if (Target.get32BitArchVariant().getArch() != llvm::Triple::x86) 612 D.Diag(diag::err_drv_unsupported_opt_for_target) << "-miamcu" 613 << Target.str(); 614 615 if (A && !A->getOption().matches(options::OPT_m32)) 616 D.Diag(diag::err_drv_argument_not_allowed_with) 617 << "-miamcu" << A->getBaseArg().getAsString(Args); 618 619 Target.setArch(llvm::Triple::x86); 620 Target.setArchName("i586"); 621 Target.setEnvironment(llvm::Triple::UnknownEnvironment); 622 Target.setEnvironmentName(""); 623 Target.setOS(llvm::Triple::ELFIAMCU); 624 Target.setVendor(llvm::Triple::UnknownVendor); 625 Target.setVendorName("intel"); 626 } 627 628 // If target is MIPS adjust the target triple 629 // accordingly to provided ABI name. 630 A = Args.getLastArg(options::OPT_mabi_EQ); 631 if (A && Target.isMIPS()) { 632 StringRef ABIName = A->getValue(); 633 if (ABIName == "32") { 634 Target = Target.get32BitArchVariant(); 635 if (Target.getEnvironment() == llvm::Triple::GNUABI64 || 636 Target.getEnvironment() == llvm::Triple::GNUABIN32) 637 Target.setEnvironment(llvm::Triple::GNU); 638 } else if (ABIName == "n32") { 639 Target = Target.get64BitArchVariant(); 640 if (Target.getEnvironment() == llvm::Triple::GNU || 641 Target.getEnvironment() == llvm::Triple::GNUABI64) 642 Target.setEnvironment(llvm::Triple::GNUABIN32); 643 } else if (ABIName == "64") { 644 Target = Target.get64BitArchVariant(); 645 if (Target.getEnvironment() == llvm::Triple::GNU || 646 Target.getEnvironment() == llvm::Triple::GNUABIN32) 647 Target.setEnvironment(llvm::Triple::GNUABI64); 648 } 649 } 650 651 // If target is RISC-V adjust the target triple according to 652 // provided architecture name 653 A = Args.getLastArg(options::OPT_march_EQ); 654 if (A && Target.isRISCV()) { 655 StringRef ArchName = A->getValue(); 656 if (ArchName.startswith_insensitive("rv32")) 657 Target.setArch(llvm::Triple::riscv32); 658 else if (ArchName.startswith_insensitive("rv64")) 659 Target.setArch(llvm::Triple::riscv64); 660 } 661 662 return Target; 663 } 664 665 // Parse the LTO options and record the type of LTO compilation 666 // based on which -f(no-)?lto(=.*)? or -f(no-)?offload-lto(=.*)? 667 // option occurs last. 668 static driver::LTOKind parseLTOMode(Driver &D, const llvm::opt::ArgList &Args, 669 OptSpecifier OptEq, OptSpecifier OptNeg) { 670 if (!Args.hasFlag(OptEq, OptNeg, false)) 671 return LTOK_None; 672 673 const Arg *A = Args.getLastArg(OptEq); 674 StringRef LTOName = A->getValue(); 675 676 driver::LTOKind LTOMode = llvm::StringSwitch<LTOKind>(LTOName) 677 .Case("full", LTOK_Full) 678 .Case("thin", LTOK_Thin) 679 .Default(LTOK_Unknown); 680 681 if (LTOMode == LTOK_Unknown) { 682 D.Diag(diag::err_drv_unsupported_option_argument) 683 << A->getOption().getName() << A->getValue(); 684 return LTOK_None; 685 } 686 return LTOMode; 687 } 688 689 // Parse the LTO options. 690 void Driver::setLTOMode(const llvm::opt::ArgList &Args) { 691 LTOMode = 692 parseLTOMode(*this, Args, options::OPT_flto_EQ, options::OPT_fno_lto); 693 694 OffloadLTOMode = parseLTOMode(*this, Args, options::OPT_foffload_lto_EQ, 695 options::OPT_fno_offload_lto); 696 } 697 698 /// Compute the desired OpenMP runtime from the flags provided. 699 Driver::OpenMPRuntimeKind Driver::getOpenMPRuntime(const ArgList &Args) const { 700 StringRef RuntimeName(CLANG_DEFAULT_OPENMP_RUNTIME); 701 702 const Arg *A = Args.getLastArg(options::OPT_fopenmp_EQ); 703 if (A) 704 RuntimeName = A->getValue(); 705 706 auto RT = llvm::StringSwitch<OpenMPRuntimeKind>(RuntimeName) 707 .Case("libomp", OMPRT_OMP) 708 .Case("libgomp", OMPRT_GOMP) 709 .Case("libiomp5", OMPRT_IOMP5) 710 .Default(OMPRT_Unknown); 711 712 if (RT == OMPRT_Unknown) { 713 if (A) 714 Diag(diag::err_drv_unsupported_option_argument) 715 << A->getOption().getName() << A->getValue(); 716 else 717 // FIXME: We could use a nicer diagnostic here. 718 Diag(diag::err_drv_unsupported_opt) << "-fopenmp"; 719 } 720 721 return RT; 722 } 723 724 void Driver::CreateOffloadingDeviceToolChains(Compilation &C, 725 InputList &Inputs) { 726 727 // 728 // CUDA/HIP 729 // 730 // We need to generate a CUDA/HIP toolchain if any of the inputs has a CUDA 731 // or HIP type. However, mixed CUDA/HIP compilation is not supported. 732 bool IsCuda = 733 llvm::any_of(Inputs, [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 734 return types::isCuda(I.first); 735 }); 736 bool IsHIP = 737 llvm::any_of(Inputs, 738 [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 739 return types::isHIP(I.first); 740 }) || 741 C.getInputArgs().hasArg(options::OPT_hip_link); 742 if (IsCuda && IsHIP) { 743 Diag(clang::diag::err_drv_mix_cuda_hip); 744 return; 745 } 746 if (IsCuda) { 747 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 748 const llvm::Triple &HostTriple = HostTC->getTriple(); 749 auto OFK = Action::OFK_Cuda; 750 auto CudaTriple = 751 getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), HostTriple); 752 if (!CudaTriple) 753 return; 754 // Use the CUDA and host triples as the key into the ToolChains map, 755 // because the device toolchain we create depends on both. 756 auto &CudaTC = ToolChains[CudaTriple->str() + "/" + HostTriple.str()]; 757 if (!CudaTC) { 758 CudaTC = std::make_unique<toolchains::CudaToolChain>( 759 *this, *CudaTriple, *HostTC, C.getInputArgs(), OFK); 760 } 761 C.addOffloadDeviceToolChain(CudaTC.get(), OFK); 762 } else if (IsHIP) { 763 if (auto *OMPTargetArg = 764 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 765 Diag(clang::diag::err_drv_unsupported_opt_for_language_mode) 766 << OMPTargetArg->getSpelling() << "HIP"; 767 return; 768 } 769 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 770 auto OFK = Action::OFK_HIP; 771 auto HIPTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 772 if (!HIPTriple) 773 return; 774 auto *HIPTC = &getOffloadingDeviceToolChain(C.getInputArgs(), *HIPTriple, 775 *HostTC, OFK); 776 assert(HIPTC && "Could not create offloading device tool chain."); 777 C.addOffloadDeviceToolChain(HIPTC, OFK); 778 } 779 780 // 781 // OpenMP 782 // 783 // We need to generate an OpenMP toolchain if the user specified targets with 784 // the -fopenmp-targets option or used --offload-arch with OpenMP enabled. 785 bool IsOpenMPOffloading = 786 C.getInputArgs().hasFlag(options::OPT_fopenmp, options::OPT_fopenmp_EQ, 787 options::OPT_fno_openmp, false) && 788 (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ) || 789 C.getInputArgs().hasArg(options::OPT_offload_arch_EQ)); 790 if (IsOpenMPOffloading) { 791 // We expect that -fopenmp-targets is always used in conjunction with the 792 // option -fopenmp specifying a valid runtime with offloading support, i.e. 793 // libomp or libiomp. 794 OpenMPRuntimeKind RuntimeKind = getOpenMPRuntime(C.getInputArgs()); 795 if (RuntimeKind != OMPRT_OMP && RuntimeKind != OMPRT_IOMP5) { 796 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 797 return; 798 } 799 800 llvm::StringMap<llvm::DenseSet<StringRef>> DerivedArchs; 801 llvm::StringMap<StringRef> FoundNormalizedTriples; 802 llvm::SmallVector<StringRef, 4> OpenMPTriples; 803 804 // If the user specified -fopenmp-targets= we create a toolchain for each 805 // valid triple. Otherwise, if only --offload-arch= was specified we instead 806 // attempt to derive the appropriate toolchains from the arguments. 807 if (Arg *OpenMPTargets = 808 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 809 if (OpenMPTargets && !OpenMPTargets->getNumValues()) { 810 Diag(clang::diag::warn_drv_empty_joined_argument) 811 << OpenMPTargets->getAsString(C.getInputArgs()); 812 return; 813 } 814 llvm::copy(OpenMPTargets->getValues(), std::back_inserter(OpenMPTriples)); 815 } else if (C.getInputArgs().hasArg(options::OPT_offload_arch_EQ) && 816 !IsHIP && !IsCuda) { 817 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 818 auto AMDTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 819 auto NVPTXTriple = getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), 820 HostTC->getTriple()); 821 822 // Attempt to deduce the offloading triple from the set of architectures. 823 // We can only correctly deduce NVPTX / AMDGPU triples currently. 824 llvm::DenseSet<StringRef> Archs = 825 getOffloadArchs(C, C.getArgs(), Action::OFK_OpenMP, nullptr); 826 for (StringRef Arch : Archs) { 827 if (NVPTXTriple && IsNVIDIAGpuArch(StringToCudaArch( 828 getProcessorFromTargetID(*NVPTXTriple, Arch)))) { 829 DerivedArchs[NVPTXTriple->getTriple()].insert(Arch); 830 } else if (AMDTriple && 831 IsAMDGpuArch(StringToCudaArch( 832 getProcessorFromTargetID(*AMDTriple, Arch)))) { 833 DerivedArchs[AMDTriple->getTriple()].insert(Arch); 834 } else { 835 Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch) << Arch; 836 return; 837 } 838 } 839 840 for (const auto &TripleAndArchs : DerivedArchs) 841 OpenMPTriples.push_back(TripleAndArchs.first()); 842 } 843 844 for (StringRef Val : OpenMPTriples) { 845 llvm::Triple TT(ToolChain::getOpenMPTriple(Val)); 846 std::string NormalizedName = TT.normalize(); 847 848 // Make sure we don't have a duplicate triple. 849 auto Duplicate = FoundNormalizedTriples.find(NormalizedName); 850 if (Duplicate != FoundNormalizedTriples.end()) { 851 Diag(clang::diag::warn_drv_omp_offload_target_duplicate) 852 << Val << Duplicate->second; 853 continue; 854 } 855 856 // Store the current triple so that we can check for duplicates in the 857 // following iterations. 858 FoundNormalizedTriples[NormalizedName] = Val; 859 860 // If the specified target is invalid, emit a diagnostic. 861 if (TT.getArch() == llvm::Triple::UnknownArch) 862 Diag(clang::diag::err_drv_invalid_omp_target) << Val; 863 else { 864 const ToolChain *TC; 865 // Device toolchains have to be selected differently. They pair host 866 // and device in their implementation. 867 if (TT.isNVPTX() || TT.isAMDGCN()) { 868 const ToolChain *HostTC = 869 C.getSingleOffloadToolChain<Action::OFK_Host>(); 870 assert(HostTC && "Host toolchain should be always defined."); 871 auto &DeviceTC = 872 ToolChains[TT.str() + "/" + HostTC->getTriple().normalize()]; 873 if (!DeviceTC) { 874 if (TT.isNVPTX()) 875 DeviceTC = std::make_unique<toolchains::CudaToolChain>( 876 *this, TT, *HostTC, C.getInputArgs(), Action::OFK_OpenMP); 877 else if (TT.isAMDGCN()) 878 DeviceTC = std::make_unique<toolchains::AMDGPUOpenMPToolChain>( 879 *this, TT, *HostTC, C.getInputArgs()); 880 else 881 assert(DeviceTC && "Device toolchain not defined."); 882 } 883 884 TC = DeviceTC.get(); 885 } else 886 TC = &getToolChain(C.getInputArgs(), TT); 887 C.addOffloadDeviceToolChain(TC, Action::OFK_OpenMP); 888 if (DerivedArchs.find(TT.getTriple()) != DerivedArchs.end()) 889 KnownArchs[TC] = DerivedArchs[TT.getTriple()]; 890 } 891 } 892 } else if (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ)) { 893 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 894 return; 895 } 896 897 // 898 // TODO: Add support for other offloading programming models here. 899 // 900 } 901 902 /// Looks the given directories for the specified file. 903 /// 904 /// \param[out] FilePath File path, if the file was found. 905 /// \param[in] Dirs Directories used for the search. 906 /// \param[in] FileName Name of the file to search for. 907 /// \return True if file was found. 908 /// 909 /// Looks for file specified by FileName sequentially in directories specified 910 /// by Dirs. 911 /// 912 static bool searchForFile(SmallVectorImpl<char> &FilePath, 913 ArrayRef<StringRef> Dirs, StringRef FileName) { 914 SmallString<128> WPath; 915 for (const StringRef &Dir : Dirs) { 916 if (Dir.empty()) 917 continue; 918 WPath.clear(); 919 llvm::sys::path::append(WPath, Dir, FileName); 920 llvm::sys::path::native(WPath); 921 if (llvm::sys::fs::is_regular_file(WPath)) { 922 FilePath = std::move(WPath); 923 return true; 924 } 925 } 926 return false; 927 } 928 929 bool Driver::readConfigFile(StringRef FileName) { 930 // Try reading the given file. 931 SmallVector<const char *, 32> NewCfgArgs; 932 if (!llvm::cl::readConfigFile(FileName, Saver, NewCfgArgs)) { 933 Diag(diag::err_drv_cannot_read_config_file) << FileName; 934 return true; 935 } 936 937 // Read options from config file. 938 llvm::SmallString<128> CfgFileName(FileName); 939 llvm::sys::path::native(CfgFileName); 940 ConfigFile = std::string(CfgFileName); 941 bool ContainErrors; 942 CfgOptions = std::make_unique<InputArgList>( 943 ParseArgStrings(NewCfgArgs, IsCLMode(), ContainErrors)); 944 if (ContainErrors) { 945 CfgOptions.reset(); 946 return true; 947 } 948 949 if (CfgOptions->hasArg(options::OPT_config)) { 950 CfgOptions.reset(); 951 Diag(diag::err_drv_nested_config_file); 952 return true; 953 } 954 955 // Claim all arguments that come from a configuration file so that the driver 956 // does not warn on any that is unused. 957 for (Arg *A : *CfgOptions) 958 A->claim(); 959 return false; 960 } 961 962 bool Driver::loadConfigFile() { 963 std::string CfgFileName; 964 bool FileSpecifiedExplicitly = false; 965 966 // Process options that change search path for config files. 967 if (CLOptions) { 968 if (CLOptions->hasArg(options::OPT_config_system_dir_EQ)) { 969 SmallString<128> CfgDir; 970 CfgDir.append( 971 CLOptions->getLastArgValue(options::OPT_config_system_dir_EQ)); 972 if (!CfgDir.empty()) { 973 if (llvm::sys::fs::make_absolute(CfgDir).value() != 0) 974 SystemConfigDir.clear(); 975 else 976 SystemConfigDir = static_cast<std::string>(CfgDir); 977 } 978 } 979 if (CLOptions->hasArg(options::OPT_config_user_dir_EQ)) { 980 SmallString<128> CfgDir; 981 CfgDir.append( 982 CLOptions->getLastArgValue(options::OPT_config_user_dir_EQ)); 983 if (!CfgDir.empty()) { 984 if (llvm::sys::fs::make_absolute(CfgDir).value() != 0) 985 UserConfigDir.clear(); 986 else 987 UserConfigDir = static_cast<std::string>(CfgDir); 988 } 989 } 990 } 991 992 // First try to find config file specified in command line. 993 if (CLOptions) { 994 std::vector<std::string> ConfigFiles = 995 CLOptions->getAllArgValues(options::OPT_config); 996 if (ConfigFiles.size() > 1) { 997 if (!llvm::all_of(ConfigFiles, [ConfigFiles](const std::string &s) { 998 return s == ConfigFiles[0]; 999 })) { 1000 Diag(diag::err_drv_duplicate_config); 1001 return true; 1002 } 1003 } 1004 1005 if (!ConfigFiles.empty()) { 1006 CfgFileName = ConfigFiles.front(); 1007 assert(!CfgFileName.empty()); 1008 1009 // If argument contains directory separator, treat it as a path to 1010 // configuration file. 1011 if (llvm::sys::path::has_parent_path(CfgFileName)) { 1012 SmallString<128> CfgFilePath; 1013 if (llvm::sys::path::is_relative(CfgFileName)) 1014 llvm::sys::fs::current_path(CfgFilePath); 1015 llvm::sys::path::append(CfgFilePath, CfgFileName); 1016 if (!llvm::sys::fs::is_regular_file(CfgFilePath)) { 1017 Diag(diag::err_drv_config_file_not_exist) << CfgFilePath; 1018 return true; 1019 } 1020 return readConfigFile(CfgFilePath); 1021 } 1022 1023 FileSpecifiedExplicitly = true; 1024 } 1025 } 1026 1027 // If config file is not specified explicitly, try to deduce configuration 1028 // from executable name. For instance, an executable 'armv7l-clang' will 1029 // search for config file 'armv7l-clang.cfg'. 1030 if (CfgFileName.empty() && !ClangNameParts.TargetPrefix.empty()) 1031 CfgFileName = ClangNameParts.TargetPrefix + '-' + ClangNameParts.ModeSuffix; 1032 1033 if (CfgFileName.empty()) 1034 return false; 1035 1036 // Determine architecture part of the file name, if it is present. 1037 StringRef CfgFileArch = CfgFileName; 1038 size_t ArchPrefixLen = CfgFileArch.find('-'); 1039 if (ArchPrefixLen == StringRef::npos) 1040 ArchPrefixLen = CfgFileArch.size(); 1041 llvm::Triple CfgTriple; 1042 CfgFileArch = CfgFileArch.take_front(ArchPrefixLen); 1043 CfgTriple = llvm::Triple(llvm::Triple::normalize(CfgFileArch)); 1044 if (CfgTriple.getArch() == llvm::Triple::ArchType::UnknownArch) 1045 ArchPrefixLen = 0; 1046 1047 if (!StringRef(CfgFileName).endswith(".cfg")) 1048 CfgFileName += ".cfg"; 1049 1050 // If config file starts with architecture name and command line options 1051 // redefine architecture (with options like -m32 -LE etc), try finding new 1052 // config file with that architecture. 1053 SmallString<128> FixedConfigFile; 1054 size_t FixedArchPrefixLen = 0; 1055 if (ArchPrefixLen) { 1056 // Get architecture name from config file name like 'i386.cfg' or 1057 // 'armv7l-clang.cfg'. 1058 // Check if command line options changes effective triple. 1059 llvm::Triple EffectiveTriple = computeTargetTriple(*this, 1060 CfgTriple.getTriple(), *CLOptions); 1061 if (CfgTriple.getArch() != EffectiveTriple.getArch()) { 1062 FixedConfigFile = EffectiveTriple.getArchName(); 1063 FixedArchPrefixLen = FixedConfigFile.size(); 1064 // Append the rest of original file name so that file name transforms 1065 // like: i386-clang.cfg -> x86_64-clang.cfg. 1066 if (ArchPrefixLen < CfgFileName.size()) 1067 FixedConfigFile += CfgFileName.substr(ArchPrefixLen); 1068 } 1069 } 1070 1071 // Prepare list of directories where config file is searched for. 1072 StringRef CfgFileSearchDirs[] = {UserConfigDir, SystemConfigDir, Dir}; 1073 1074 // Try to find config file. First try file with corrected architecture. 1075 llvm::SmallString<128> CfgFilePath; 1076 if (!FixedConfigFile.empty()) { 1077 if (searchForFile(CfgFilePath, CfgFileSearchDirs, FixedConfigFile)) 1078 return readConfigFile(CfgFilePath); 1079 // If 'x86_64-clang.cfg' was not found, try 'x86_64.cfg'. 1080 FixedConfigFile.resize(FixedArchPrefixLen); 1081 FixedConfigFile.append(".cfg"); 1082 if (searchForFile(CfgFilePath, CfgFileSearchDirs, FixedConfigFile)) 1083 return readConfigFile(CfgFilePath); 1084 } 1085 1086 // Then try original file name. 1087 if (searchForFile(CfgFilePath, CfgFileSearchDirs, CfgFileName)) 1088 return readConfigFile(CfgFilePath); 1089 1090 // Finally try removing driver mode part: 'x86_64-clang.cfg' -> 'x86_64.cfg'. 1091 if (!ClangNameParts.ModeSuffix.empty() && 1092 !ClangNameParts.TargetPrefix.empty()) { 1093 CfgFileName.assign(ClangNameParts.TargetPrefix); 1094 CfgFileName.append(".cfg"); 1095 if (searchForFile(CfgFilePath, CfgFileSearchDirs, CfgFileName)) 1096 return readConfigFile(CfgFilePath); 1097 } 1098 1099 // Report error but only if config file was specified explicitly, by option 1100 // --config. If it was deduced from executable name, it is not an error. 1101 if (FileSpecifiedExplicitly) { 1102 Diag(diag::err_drv_config_file_not_found) << CfgFileName; 1103 for (const StringRef &SearchDir : CfgFileSearchDirs) 1104 if (!SearchDir.empty()) 1105 Diag(diag::note_drv_config_file_searched_in) << SearchDir; 1106 return true; 1107 } 1108 1109 return false; 1110 } 1111 1112 Compilation *Driver::BuildCompilation(ArrayRef<const char *> ArgList) { 1113 llvm::PrettyStackTraceString CrashInfo("Compilation construction"); 1114 1115 // FIXME: Handle environment options which affect driver behavior, somewhere 1116 // (client?). GCC_EXEC_PREFIX, LPATH, CC_PRINT_OPTIONS. 1117 1118 // We look for the driver mode option early, because the mode can affect 1119 // how other options are parsed. 1120 1121 auto DriverMode = getDriverMode(ClangExecutable, ArgList.slice(1)); 1122 if (!DriverMode.empty()) 1123 setDriverMode(DriverMode); 1124 1125 // FIXME: What are we going to do with -V and -b? 1126 1127 // Arguments specified in command line. 1128 bool ContainsError; 1129 CLOptions = std::make_unique<InputArgList>( 1130 ParseArgStrings(ArgList.slice(1), IsCLMode(), ContainsError)); 1131 1132 // Try parsing configuration file. 1133 if (!ContainsError) 1134 ContainsError = loadConfigFile(); 1135 bool HasConfigFile = !ContainsError && (CfgOptions.get() != nullptr); 1136 1137 // All arguments, from both config file and command line. 1138 InputArgList Args = std::move(HasConfigFile ? std::move(*CfgOptions) 1139 : std::move(*CLOptions)); 1140 1141 // The args for config files or /clang: flags belong to different InputArgList 1142 // objects than Args. This copies an Arg from one of those other InputArgLists 1143 // to the ownership of Args. 1144 auto appendOneArg = [&Args](const Arg *Opt, const Arg *BaseArg) { 1145 unsigned Index = Args.MakeIndex(Opt->getSpelling()); 1146 Arg *Copy = new llvm::opt::Arg(Opt->getOption(), Args.getArgString(Index), 1147 Index, BaseArg); 1148 Copy->getValues() = Opt->getValues(); 1149 if (Opt->isClaimed()) 1150 Copy->claim(); 1151 Copy->setOwnsValues(Opt->getOwnsValues()); 1152 Opt->setOwnsValues(false); 1153 Args.append(Copy); 1154 }; 1155 1156 if (HasConfigFile) 1157 for (auto *Opt : *CLOptions) { 1158 if (Opt->getOption().matches(options::OPT_config)) 1159 continue; 1160 const Arg *BaseArg = &Opt->getBaseArg(); 1161 if (BaseArg == Opt) 1162 BaseArg = nullptr; 1163 appendOneArg(Opt, BaseArg); 1164 } 1165 1166 // In CL mode, look for any pass-through arguments 1167 if (IsCLMode() && !ContainsError) { 1168 SmallVector<const char *, 16> CLModePassThroughArgList; 1169 for (const auto *A : Args.filtered(options::OPT__SLASH_clang)) { 1170 A->claim(); 1171 CLModePassThroughArgList.push_back(A->getValue()); 1172 } 1173 1174 if (!CLModePassThroughArgList.empty()) { 1175 // Parse any pass through args using default clang processing rather 1176 // than clang-cl processing. 1177 auto CLModePassThroughOptions = std::make_unique<InputArgList>( 1178 ParseArgStrings(CLModePassThroughArgList, false, ContainsError)); 1179 1180 if (!ContainsError) 1181 for (auto *Opt : *CLModePassThroughOptions) { 1182 appendOneArg(Opt, nullptr); 1183 } 1184 } 1185 } 1186 1187 // Check for working directory option before accessing any files 1188 if (Arg *WD = Args.getLastArg(options::OPT_working_directory)) 1189 if (VFS->setCurrentWorkingDirectory(WD->getValue())) 1190 Diag(diag::err_drv_unable_to_set_working_directory) << WD->getValue(); 1191 1192 // FIXME: This stuff needs to go into the Compilation, not the driver. 1193 bool CCCPrintPhases; 1194 1195 // Silence driver warnings if requested 1196 Diags.setIgnoreAllWarnings(Args.hasArg(options::OPT_w)); 1197 1198 // -canonical-prefixes, -no-canonical-prefixes are used very early in main. 1199 Args.ClaimAllArgs(options::OPT_canonical_prefixes); 1200 Args.ClaimAllArgs(options::OPT_no_canonical_prefixes); 1201 1202 // f(no-)integated-cc1 is also used very early in main. 1203 Args.ClaimAllArgs(options::OPT_fintegrated_cc1); 1204 Args.ClaimAllArgs(options::OPT_fno_integrated_cc1); 1205 1206 // Ignore -pipe. 1207 Args.ClaimAllArgs(options::OPT_pipe); 1208 1209 // Extract -ccc args. 1210 // 1211 // FIXME: We need to figure out where this behavior should live. Most of it 1212 // should be outside in the client; the parts that aren't should have proper 1213 // options, either by introducing new ones or by overloading gcc ones like -V 1214 // or -b. 1215 CCCPrintPhases = Args.hasArg(options::OPT_ccc_print_phases); 1216 CCCPrintBindings = Args.hasArg(options::OPT_ccc_print_bindings); 1217 if (const Arg *A = Args.getLastArg(options::OPT_ccc_gcc_name)) 1218 CCCGenericGCCName = A->getValue(); 1219 1220 // Process -fproc-stat-report options. 1221 if (const Arg *A = Args.getLastArg(options::OPT_fproc_stat_report_EQ)) { 1222 CCPrintProcessStats = true; 1223 CCPrintStatReportFilename = A->getValue(); 1224 } 1225 if (Args.hasArg(options::OPT_fproc_stat_report)) 1226 CCPrintProcessStats = true; 1227 1228 // FIXME: TargetTriple is used by the target-prefixed calls to as/ld 1229 // and getToolChain is const. 1230 if (IsCLMode()) { 1231 // clang-cl targets MSVC-style Win32. 1232 llvm::Triple T(TargetTriple); 1233 T.setOS(llvm::Triple::Win32); 1234 T.setVendor(llvm::Triple::PC); 1235 T.setEnvironment(llvm::Triple::MSVC); 1236 T.setObjectFormat(llvm::Triple::COFF); 1237 TargetTriple = T.str(); 1238 } else if (IsDXCMode()) { 1239 // Build TargetTriple from target_profile option for clang-dxc. 1240 if (const Arg *A = Args.getLastArg(options::OPT_target_profile)) { 1241 StringRef TargetProfile = A->getValue(); 1242 if (auto Triple = 1243 toolchains::HLSLToolChain::parseTargetProfile(TargetProfile)) 1244 TargetTriple = Triple.getValue(); 1245 else 1246 Diag(diag::err_drv_invalid_directx_shader_module) << TargetProfile; 1247 1248 A->claim(); 1249 } else { 1250 Diag(diag::err_drv_dxc_missing_target_profile); 1251 } 1252 } 1253 1254 if (const Arg *A = Args.getLastArg(options::OPT_target)) 1255 TargetTriple = A->getValue(); 1256 if (const Arg *A = Args.getLastArg(options::OPT_ccc_install_dir)) 1257 Dir = InstalledDir = A->getValue(); 1258 for (const Arg *A : Args.filtered(options::OPT_B)) { 1259 A->claim(); 1260 PrefixDirs.push_back(A->getValue(0)); 1261 } 1262 if (Optional<std::string> CompilerPathValue = 1263 llvm::sys::Process::GetEnv("COMPILER_PATH")) { 1264 StringRef CompilerPath = *CompilerPathValue; 1265 while (!CompilerPath.empty()) { 1266 std::pair<StringRef, StringRef> Split = 1267 CompilerPath.split(llvm::sys::EnvPathSeparator); 1268 PrefixDirs.push_back(std::string(Split.first)); 1269 CompilerPath = Split.second; 1270 } 1271 } 1272 if (const Arg *A = Args.getLastArg(options::OPT__sysroot_EQ)) 1273 SysRoot = A->getValue(); 1274 if (const Arg *A = Args.getLastArg(options::OPT__dyld_prefix_EQ)) 1275 DyldPrefix = A->getValue(); 1276 1277 if (const Arg *A = Args.getLastArg(options::OPT_resource_dir)) 1278 ResourceDir = A->getValue(); 1279 1280 if (const Arg *A = Args.getLastArg(options::OPT_save_temps_EQ)) { 1281 SaveTemps = llvm::StringSwitch<SaveTempsMode>(A->getValue()) 1282 .Case("cwd", SaveTempsCwd) 1283 .Case("obj", SaveTempsObj) 1284 .Default(SaveTempsCwd); 1285 } 1286 1287 setLTOMode(Args); 1288 1289 // Process -fembed-bitcode= flags. 1290 if (Arg *A = Args.getLastArg(options::OPT_fembed_bitcode_EQ)) { 1291 StringRef Name = A->getValue(); 1292 unsigned Model = llvm::StringSwitch<unsigned>(Name) 1293 .Case("off", EmbedNone) 1294 .Case("all", EmbedBitcode) 1295 .Case("bitcode", EmbedBitcode) 1296 .Case("marker", EmbedMarker) 1297 .Default(~0U); 1298 if (Model == ~0U) { 1299 Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) 1300 << Name; 1301 } else 1302 BitcodeEmbed = static_cast<BitcodeEmbedMode>(Model); 1303 } 1304 1305 // Setting up the jobs for some precompile cases depends on whether we are 1306 // treating them as PCH, implicit modules or C++20 ones. 1307 // TODO: inferring the mode like this seems fragile (it meets the objective 1308 // of not requiring anything new for operation, however). 1309 const Arg *Std = Args.getLastArg(options::OPT_std_EQ); 1310 ModulesModeCXX20 = 1311 !Args.hasArg(options::OPT_fmodules) && Std && 1312 (Std->containsValue("c++20") || Std->containsValue("c++2b") || 1313 Std->containsValue("c++2a") || Std->containsValue("c++latest")); 1314 1315 // Process -fmodule-header{=} flags. 1316 if (Arg *A = Args.getLastArg(options::OPT_fmodule_header_EQ, 1317 options::OPT_fmodule_header)) { 1318 // These flags force C++20 handling of headers. 1319 ModulesModeCXX20 = true; 1320 if (A->getOption().matches(options::OPT_fmodule_header)) 1321 CXX20HeaderType = HeaderMode_Default; 1322 else { 1323 StringRef ArgName = A->getValue(); 1324 unsigned Kind = llvm::StringSwitch<unsigned>(ArgName) 1325 .Case("user", HeaderMode_User) 1326 .Case("system", HeaderMode_System) 1327 .Default(~0U); 1328 if (Kind == ~0U) { 1329 Diags.Report(diag::err_drv_invalid_value) 1330 << A->getAsString(Args) << ArgName; 1331 } else 1332 CXX20HeaderType = static_cast<ModuleHeaderMode>(Kind); 1333 } 1334 } 1335 1336 std::unique_ptr<llvm::opt::InputArgList> UArgs = 1337 std::make_unique<InputArgList>(std::move(Args)); 1338 1339 // Perform the default argument translations. 1340 DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs); 1341 1342 // Owned by the host. 1343 const ToolChain &TC = getToolChain( 1344 *UArgs, computeTargetTriple(*this, TargetTriple, *UArgs)); 1345 1346 // The compilation takes ownership of Args. 1347 Compilation *C = new Compilation(*this, TC, UArgs.release(), TranslatedArgs, 1348 ContainsError); 1349 1350 if (!HandleImmediateArgs(*C)) 1351 return C; 1352 1353 // Construct the list of inputs. 1354 InputList Inputs; 1355 BuildInputs(C->getDefaultToolChain(), *TranslatedArgs, Inputs); 1356 1357 // Populate the tool chains for the offloading devices, if any. 1358 CreateOffloadingDeviceToolChains(*C, Inputs); 1359 1360 // Construct the list of abstract actions to perform for this compilation. On 1361 // MachO targets this uses the driver-driver and universal actions. 1362 if (TC.getTriple().isOSBinFormatMachO()) 1363 BuildUniversalActions(*C, C->getDefaultToolChain(), Inputs); 1364 else 1365 BuildActions(*C, C->getArgs(), Inputs, C->getActions()); 1366 1367 if (CCCPrintPhases) { 1368 PrintActions(*C); 1369 return C; 1370 } 1371 1372 BuildJobs(*C); 1373 1374 return C; 1375 } 1376 1377 static void printArgList(raw_ostream &OS, const llvm::opt::ArgList &Args) { 1378 llvm::opt::ArgStringList ASL; 1379 for (const auto *A : Args) { 1380 // Use user's original spelling of flags. For example, use 1381 // `/source-charset:utf-8` instead of `-finput-charset=utf-8` if the user 1382 // wrote the former. 1383 while (A->getAlias()) 1384 A = A->getAlias(); 1385 A->render(Args, ASL); 1386 } 1387 1388 for (auto I = ASL.begin(), E = ASL.end(); I != E; ++I) { 1389 if (I != ASL.begin()) 1390 OS << ' '; 1391 llvm::sys::printArg(OS, *I, true); 1392 } 1393 OS << '\n'; 1394 } 1395 1396 bool Driver::getCrashDiagnosticFile(StringRef ReproCrashFilename, 1397 SmallString<128> &CrashDiagDir) { 1398 using namespace llvm::sys; 1399 assert(llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin() && 1400 "Only knows about .crash files on Darwin"); 1401 1402 // The .crash file can be found on at ~/Library/Logs/DiagnosticReports/ 1403 // (or /Library/Logs/DiagnosticReports for root) and has the filename pattern 1404 // clang-<VERSION>_<YYYY-MM-DD-HHMMSS>_<hostname>.crash. 1405 path::home_directory(CrashDiagDir); 1406 if (CrashDiagDir.startswith("/var/root")) 1407 CrashDiagDir = "/"; 1408 path::append(CrashDiagDir, "Library/Logs/DiagnosticReports"); 1409 int PID = 1410 #if LLVM_ON_UNIX 1411 getpid(); 1412 #else 1413 0; 1414 #endif 1415 std::error_code EC; 1416 fs::file_status FileStatus; 1417 TimePoint<> LastAccessTime; 1418 SmallString<128> CrashFilePath; 1419 // Lookup the .crash files and get the one generated by a subprocess spawned 1420 // by this driver invocation. 1421 for (fs::directory_iterator File(CrashDiagDir, EC), FileEnd; 1422 File != FileEnd && !EC; File.increment(EC)) { 1423 StringRef FileName = path::filename(File->path()); 1424 if (!FileName.startswith(Name)) 1425 continue; 1426 if (fs::status(File->path(), FileStatus)) 1427 continue; 1428 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CrashFile = 1429 llvm::MemoryBuffer::getFile(File->path()); 1430 if (!CrashFile) 1431 continue; 1432 // The first line should start with "Process:", otherwise this isn't a real 1433 // .crash file. 1434 StringRef Data = CrashFile.get()->getBuffer(); 1435 if (!Data.startswith("Process:")) 1436 continue; 1437 // Parse parent process pid line, e.g: "Parent Process: clang-4.0 [79141]" 1438 size_t ParentProcPos = Data.find("Parent Process:"); 1439 if (ParentProcPos == StringRef::npos) 1440 continue; 1441 size_t LineEnd = Data.find_first_of("\n", ParentProcPos); 1442 if (LineEnd == StringRef::npos) 1443 continue; 1444 StringRef ParentProcess = Data.slice(ParentProcPos+15, LineEnd).trim(); 1445 int OpenBracket = -1, CloseBracket = -1; 1446 for (size_t i = 0, e = ParentProcess.size(); i < e; ++i) { 1447 if (ParentProcess[i] == '[') 1448 OpenBracket = i; 1449 if (ParentProcess[i] == ']') 1450 CloseBracket = i; 1451 } 1452 // Extract the parent process PID from the .crash file and check whether 1453 // it matches this driver invocation pid. 1454 int CrashPID; 1455 if (OpenBracket < 0 || CloseBracket < 0 || 1456 ParentProcess.slice(OpenBracket + 1, CloseBracket) 1457 .getAsInteger(10, CrashPID) || CrashPID != PID) { 1458 continue; 1459 } 1460 1461 // Found a .crash file matching the driver pid. To avoid getting an older 1462 // and misleading crash file, continue looking for the most recent. 1463 // FIXME: the driver can dispatch multiple cc1 invocations, leading to 1464 // multiple crashes poiting to the same parent process. Since the driver 1465 // does not collect pid information for the dispatched invocation there's 1466 // currently no way to distinguish among them. 1467 const auto FileAccessTime = FileStatus.getLastModificationTime(); 1468 if (FileAccessTime > LastAccessTime) { 1469 CrashFilePath.assign(File->path()); 1470 LastAccessTime = FileAccessTime; 1471 } 1472 } 1473 1474 // If found, copy it over to the location of other reproducer files. 1475 if (!CrashFilePath.empty()) { 1476 EC = fs::copy_file(CrashFilePath, ReproCrashFilename); 1477 if (EC) 1478 return false; 1479 return true; 1480 } 1481 1482 return false; 1483 } 1484 1485 // When clang crashes, produce diagnostic information including the fully 1486 // preprocessed source file(s). Request that the developer attach the 1487 // diagnostic information to a bug report. 1488 void Driver::generateCompilationDiagnostics( 1489 Compilation &C, const Command &FailingCommand, 1490 StringRef AdditionalInformation, CompilationDiagnosticReport *Report) { 1491 if (C.getArgs().hasArg(options::OPT_fno_crash_diagnostics)) 1492 return; 1493 1494 // Don't try to generate diagnostics for link or dsymutil jobs. 1495 if (FailingCommand.getCreator().isLinkJob() || 1496 FailingCommand.getCreator().isDsymutilJob()) 1497 return; 1498 1499 // Print the version of the compiler. 1500 PrintVersion(C, llvm::errs()); 1501 1502 // Suppress driver output and emit preprocessor output to temp file. 1503 CCGenDiagnostics = true; 1504 1505 // Save the original job command(s). 1506 Command Cmd = FailingCommand; 1507 1508 // Keep track of whether we produce any errors while trying to produce 1509 // preprocessed sources. 1510 DiagnosticErrorTrap Trap(Diags); 1511 1512 // Suppress tool output. 1513 C.initCompilationForDiagnostics(); 1514 1515 // Construct the list of inputs. 1516 InputList Inputs; 1517 BuildInputs(C.getDefaultToolChain(), C.getArgs(), Inputs); 1518 1519 for (InputList::iterator it = Inputs.begin(), ie = Inputs.end(); it != ie;) { 1520 bool IgnoreInput = false; 1521 1522 // Ignore input from stdin or any inputs that cannot be preprocessed. 1523 // Check type first as not all linker inputs have a value. 1524 if (types::getPreprocessedType(it->first) == types::TY_INVALID) { 1525 IgnoreInput = true; 1526 } else if (!strcmp(it->second->getValue(), "-")) { 1527 Diag(clang::diag::note_drv_command_failed_diag_msg) 1528 << "Error generating preprocessed source(s) - " 1529 "ignoring input from stdin."; 1530 IgnoreInput = true; 1531 } 1532 1533 if (IgnoreInput) { 1534 it = Inputs.erase(it); 1535 ie = Inputs.end(); 1536 } else { 1537 ++it; 1538 } 1539 } 1540 1541 if (Inputs.empty()) { 1542 Diag(clang::diag::note_drv_command_failed_diag_msg) 1543 << "Error generating preprocessed source(s) - " 1544 "no preprocessable inputs."; 1545 return; 1546 } 1547 1548 // Don't attempt to generate preprocessed files if multiple -arch options are 1549 // used, unless they're all duplicates. 1550 llvm::StringSet<> ArchNames; 1551 for (const Arg *A : C.getArgs()) { 1552 if (A->getOption().matches(options::OPT_arch)) { 1553 StringRef ArchName = A->getValue(); 1554 ArchNames.insert(ArchName); 1555 } 1556 } 1557 if (ArchNames.size() > 1) { 1558 Diag(clang::diag::note_drv_command_failed_diag_msg) 1559 << "Error generating preprocessed source(s) - cannot generate " 1560 "preprocessed source with multiple -arch options."; 1561 return; 1562 } 1563 1564 // Construct the list of abstract actions to perform for this compilation. On 1565 // Darwin OSes this uses the driver-driver and builds universal actions. 1566 const ToolChain &TC = C.getDefaultToolChain(); 1567 if (TC.getTriple().isOSBinFormatMachO()) 1568 BuildUniversalActions(C, TC, Inputs); 1569 else 1570 BuildActions(C, C.getArgs(), Inputs, C.getActions()); 1571 1572 BuildJobs(C); 1573 1574 // If there were errors building the compilation, quit now. 1575 if (Trap.hasErrorOccurred()) { 1576 Diag(clang::diag::note_drv_command_failed_diag_msg) 1577 << "Error generating preprocessed source(s)."; 1578 return; 1579 } 1580 1581 // Generate preprocessed output. 1582 SmallVector<std::pair<int, const Command *>, 4> FailingCommands; 1583 C.ExecuteJobs(C.getJobs(), FailingCommands); 1584 1585 // If any of the preprocessing commands failed, clean up and exit. 1586 if (!FailingCommands.empty()) { 1587 Diag(clang::diag::note_drv_command_failed_diag_msg) 1588 << "Error generating preprocessed source(s)."; 1589 return; 1590 } 1591 1592 const ArgStringList &TempFiles = C.getTempFiles(); 1593 if (TempFiles.empty()) { 1594 Diag(clang::diag::note_drv_command_failed_diag_msg) 1595 << "Error generating preprocessed source(s)."; 1596 return; 1597 } 1598 1599 Diag(clang::diag::note_drv_command_failed_diag_msg) 1600 << "\n********************\n\n" 1601 "PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:\n" 1602 "Preprocessed source(s) and associated run script(s) are located at:"; 1603 1604 SmallString<128> VFS; 1605 SmallString<128> ReproCrashFilename; 1606 for (const char *TempFile : TempFiles) { 1607 Diag(clang::diag::note_drv_command_failed_diag_msg) << TempFile; 1608 if (Report) 1609 Report->TemporaryFiles.push_back(TempFile); 1610 if (ReproCrashFilename.empty()) { 1611 ReproCrashFilename = TempFile; 1612 llvm::sys::path::replace_extension(ReproCrashFilename, ".crash"); 1613 } 1614 if (StringRef(TempFile).endswith(".cache")) { 1615 // In some cases (modules) we'll dump extra data to help with reproducing 1616 // the crash into a directory next to the output. 1617 VFS = llvm::sys::path::filename(TempFile); 1618 llvm::sys::path::append(VFS, "vfs", "vfs.yaml"); 1619 } 1620 } 1621 1622 // Assume associated files are based off of the first temporary file. 1623 CrashReportInfo CrashInfo(TempFiles[0], VFS); 1624 1625 llvm::SmallString<128> Script(CrashInfo.Filename); 1626 llvm::sys::path::replace_extension(Script, "sh"); 1627 std::error_code EC; 1628 llvm::raw_fd_ostream ScriptOS(Script, EC, llvm::sys::fs::CD_CreateNew, 1629 llvm::sys::fs::FA_Write, 1630 llvm::sys::fs::OF_Text); 1631 if (EC) { 1632 Diag(clang::diag::note_drv_command_failed_diag_msg) 1633 << "Error generating run script: " << Script << " " << EC.message(); 1634 } else { 1635 ScriptOS << "# Crash reproducer for " << getClangFullVersion() << "\n" 1636 << "# Driver args: "; 1637 printArgList(ScriptOS, C.getInputArgs()); 1638 ScriptOS << "# Original command: "; 1639 Cmd.Print(ScriptOS, "\n", /*Quote=*/true); 1640 Cmd.Print(ScriptOS, "\n", /*Quote=*/true, &CrashInfo); 1641 if (!AdditionalInformation.empty()) 1642 ScriptOS << "\n# Additional information: " << AdditionalInformation 1643 << "\n"; 1644 if (Report) 1645 Report->TemporaryFiles.push_back(std::string(Script.str())); 1646 Diag(clang::diag::note_drv_command_failed_diag_msg) << Script; 1647 } 1648 1649 // On darwin, provide information about the .crash diagnostic report. 1650 if (llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin()) { 1651 SmallString<128> CrashDiagDir; 1652 if (getCrashDiagnosticFile(ReproCrashFilename, CrashDiagDir)) { 1653 Diag(clang::diag::note_drv_command_failed_diag_msg) 1654 << ReproCrashFilename.str(); 1655 } else { // Suggest a directory for the user to look for .crash files. 1656 llvm::sys::path::append(CrashDiagDir, Name); 1657 CrashDiagDir += "_<YYYY-MM-DD-HHMMSS>_<hostname>.crash"; 1658 Diag(clang::diag::note_drv_command_failed_diag_msg) 1659 << "Crash backtrace is located in"; 1660 Diag(clang::diag::note_drv_command_failed_diag_msg) 1661 << CrashDiagDir.str(); 1662 Diag(clang::diag::note_drv_command_failed_diag_msg) 1663 << "(choose the .crash file that corresponds to your crash)"; 1664 } 1665 } 1666 1667 for (const auto &A : C.getArgs().filtered(options::OPT_frewrite_map_file_EQ)) 1668 Diag(clang::diag::note_drv_command_failed_diag_msg) << A->getValue(); 1669 1670 Diag(clang::diag::note_drv_command_failed_diag_msg) 1671 << "\n\n********************"; 1672 } 1673 1674 void Driver::setUpResponseFiles(Compilation &C, Command &Cmd) { 1675 // Since commandLineFitsWithinSystemLimits() may underestimate system's 1676 // capacity if the tool does not support response files, there is a chance/ 1677 // that things will just work without a response file, so we silently just 1678 // skip it. 1679 if (Cmd.getResponseFileSupport().ResponseKind == 1680 ResponseFileSupport::RF_None || 1681 llvm::sys::commandLineFitsWithinSystemLimits(Cmd.getExecutable(), 1682 Cmd.getArguments())) 1683 return; 1684 1685 std::string TmpName = GetTemporaryPath("response", "txt"); 1686 Cmd.setResponseFile(C.addTempFile(C.getArgs().MakeArgString(TmpName))); 1687 } 1688 1689 int Driver::ExecuteCompilation( 1690 Compilation &C, 1691 SmallVectorImpl<std::pair<int, const Command *>> &FailingCommands) { 1692 // Just print if -### was present. 1693 if (C.getArgs().hasArg(options::OPT__HASH_HASH_HASH)) { 1694 C.getJobs().Print(llvm::errs(), "\n", true); 1695 return 0; 1696 } 1697 1698 // If there were errors building the compilation, quit now. 1699 if (Diags.hasErrorOccurred()) 1700 return 1; 1701 1702 // Set up response file names for each command, if necessary. 1703 for (auto &Job : C.getJobs()) 1704 setUpResponseFiles(C, Job); 1705 1706 C.ExecuteJobs(C.getJobs(), FailingCommands); 1707 1708 // If the command succeeded, we are done. 1709 if (FailingCommands.empty()) 1710 return 0; 1711 1712 // Otherwise, remove result files and print extra information about abnormal 1713 // failures. 1714 int Res = 0; 1715 for (const auto &CmdPair : FailingCommands) { 1716 int CommandRes = CmdPair.first; 1717 const Command *FailingCommand = CmdPair.second; 1718 1719 // Remove result files if we're not saving temps. 1720 if (!isSaveTempsEnabled()) { 1721 const JobAction *JA = cast<JobAction>(&FailingCommand->getSource()); 1722 C.CleanupFileMap(C.getResultFiles(), JA, true); 1723 1724 // Failure result files are valid unless we crashed. 1725 if (CommandRes < 0) 1726 C.CleanupFileMap(C.getFailureResultFiles(), JA, true); 1727 } 1728 1729 #if LLVM_ON_UNIX 1730 // llvm/lib/Support/Unix/Signals.inc will exit with a special return code 1731 // for SIGPIPE. Do not print diagnostics for this case. 1732 if (CommandRes == EX_IOERR) { 1733 Res = CommandRes; 1734 continue; 1735 } 1736 #endif 1737 1738 // Print extra information about abnormal failures, if possible. 1739 // 1740 // This is ad-hoc, but we don't want to be excessively noisy. If the result 1741 // status was 1, assume the command failed normally. In particular, if it 1742 // was the compiler then assume it gave a reasonable error code. Failures 1743 // in other tools are less common, and they generally have worse 1744 // diagnostics, so always print the diagnostic there. 1745 const Tool &FailingTool = FailingCommand->getCreator(); 1746 1747 if (!FailingCommand->getCreator().hasGoodDiagnostics() || CommandRes != 1) { 1748 // FIXME: See FIXME above regarding result code interpretation. 1749 if (CommandRes < 0) 1750 Diag(clang::diag::err_drv_command_signalled) 1751 << FailingTool.getShortName(); 1752 else 1753 Diag(clang::diag::err_drv_command_failed) 1754 << FailingTool.getShortName() << CommandRes; 1755 } 1756 } 1757 return Res; 1758 } 1759 1760 void Driver::PrintHelp(bool ShowHidden) const { 1761 unsigned IncludedFlagsBitmask; 1762 unsigned ExcludedFlagsBitmask; 1763 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 1764 getIncludeExcludeOptionFlagMasks(IsCLMode()); 1765 1766 ExcludedFlagsBitmask |= options::NoDriverOption; 1767 if (!ShowHidden) 1768 ExcludedFlagsBitmask |= HelpHidden; 1769 1770 if (IsFlangMode()) 1771 IncludedFlagsBitmask |= options::FlangOption; 1772 else 1773 ExcludedFlagsBitmask |= options::FlangOnlyOption; 1774 1775 std::string Usage = llvm::formatv("{0} [options] file...", Name).str(); 1776 getOpts().printHelp(llvm::outs(), Usage.c_str(), DriverTitle.c_str(), 1777 IncludedFlagsBitmask, ExcludedFlagsBitmask, 1778 /*ShowAllAliases=*/false); 1779 } 1780 1781 void Driver::PrintVersion(const Compilation &C, raw_ostream &OS) const { 1782 if (IsFlangMode()) { 1783 OS << getClangToolFullVersion("flang-new") << '\n'; 1784 } else { 1785 // FIXME: The following handlers should use a callback mechanism, we don't 1786 // know what the client would like to do. 1787 OS << getClangFullVersion() << '\n'; 1788 } 1789 const ToolChain &TC = C.getDefaultToolChain(); 1790 OS << "Target: " << TC.getTripleString() << '\n'; 1791 1792 // Print the threading model. 1793 if (Arg *A = C.getArgs().getLastArg(options::OPT_mthread_model)) { 1794 // Don't print if the ToolChain would have barfed on it already 1795 if (TC.isThreadModelSupported(A->getValue())) 1796 OS << "Thread model: " << A->getValue(); 1797 } else 1798 OS << "Thread model: " << TC.getThreadModel(); 1799 OS << '\n'; 1800 1801 // Print out the install directory. 1802 OS << "InstalledDir: " << InstalledDir << '\n'; 1803 1804 // If configuration file was used, print its path. 1805 if (!ConfigFile.empty()) 1806 OS << "Configuration file: " << ConfigFile << '\n'; 1807 } 1808 1809 /// PrintDiagnosticCategories - Implement the --print-diagnostic-categories 1810 /// option. 1811 static void PrintDiagnosticCategories(raw_ostream &OS) { 1812 // Skip the empty category. 1813 for (unsigned i = 1, max = DiagnosticIDs::getNumberOfCategories(); i != max; 1814 ++i) 1815 OS << i << ',' << DiagnosticIDs::getCategoryNameFromID(i) << '\n'; 1816 } 1817 1818 void Driver::HandleAutocompletions(StringRef PassedFlags) const { 1819 if (PassedFlags == "") 1820 return; 1821 // Print out all options that start with a given argument. This is used for 1822 // shell autocompletion. 1823 std::vector<std::string> SuggestedCompletions; 1824 std::vector<std::string> Flags; 1825 1826 unsigned int DisableFlags = 1827 options::NoDriverOption | options::Unsupported | options::Ignored; 1828 1829 // Make sure that Flang-only options don't pollute the Clang output 1830 // TODO: Make sure that Clang-only options don't pollute Flang output 1831 if (!IsFlangMode()) 1832 DisableFlags |= options::FlangOnlyOption; 1833 1834 // Distinguish "--autocomplete=-someflag" and "--autocomplete=-someflag," 1835 // because the latter indicates that the user put space before pushing tab 1836 // which should end up in a file completion. 1837 const bool HasSpace = PassedFlags.endswith(","); 1838 1839 // Parse PassedFlags by "," as all the command-line flags are passed to this 1840 // function separated by "," 1841 StringRef TargetFlags = PassedFlags; 1842 while (TargetFlags != "") { 1843 StringRef CurFlag; 1844 std::tie(CurFlag, TargetFlags) = TargetFlags.split(","); 1845 Flags.push_back(std::string(CurFlag)); 1846 } 1847 1848 // We want to show cc1-only options only when clang is invoked with -cc1 or 1849 // -Xclang. 1850 if (llvm::is_contained(Flags, "-Xclang") || llvm::is_contained(Flags, "-cc1")) 1851 DisableFlags &= ~options::NoDriverOption; 1852 1853 const llvm::opt::OptTable &Opts = getOpts(); 1854 StringRef Cur; 1855 Cur = Flags.at(Flags.size() - 1); 1856 StringRef Prev; 1857 if (Flags.size() >= 2) { 1858 Prev = Flags.at(Flags.size() - 2); 1859 SuggestedCompletions = Opts.suggestValueCompletions(Prev, Cur); 1860 } 1861 1862 if (SuggestedCompletions.empty()) 1863 SuggestedCompletions = Opts.suggestValueCompletions(Cur, ""); 1864 1865 // If Flags were empty, it means the user typed `clang [tab]` where we should 1866 // list all possible flags. If there was no value completion and the user 1867 // pressed tab after a space, we should fall back to a file completion. 1868 // We're printing a newline to be consistent with what we print at the end of 1869 // this function. 1870 if (SuggestedCompletions.empty() && HasSpace && !Flags.empty()) { 1871 llvm::outs() << '\n'; 1872 return; 1873 } 1874 1875 // When flag ends with '=' and there was no value completion, return empty 1876 // string and fall back to the file autocompletion. 1877 if (SuggestedCompletions.empty() && !Cur.endswith("=")) { 1878 // If the flag is in the form of "--autocomplete=-foo", 1879 // we were requested to print out all option names that start with "-foo". 1880 // For example, "--autocomplete=-fsyn" is expanded to "-fsyntax-only". 1881 SuggestedCompletions = Opts.findByPrefix(Cur, DisableFlags); 1882 1883 // We have to query the -W flags manually as they're not in the OptTable. 1884 // TODO: Find a good way to add them to OptTable instead and them remove 1885 // this code. 1886 for (StringRef S : DiagnosticIDs::getDiagnosticFlags()) 1887 if (S.startswith(Cur)) 1888 SuggestedCompletions.push_back(std::string(S)); 1889 } 1890 1891 // Sort the autocomplete candidates so that shells print them out in a 1892 // deterministic order. We could sort in any way, but we chose 1893 // case-insensitive sorting for consistency with the -help option 1894 // which prints out options in the case-insensitive alphabetical order. 1895 llvm::sort(SuggestedCompletions, [](StringRef A, StringRef B) { 1896 if (int X = A.compare_insensitive(B)) 1897 return X < 0; 1898 return A.compare(B) > 0; 1899 }); 1900 1901 llvm::outs() << llvm::join(SuggestedCompletions, "\n") << '\n'; 1902 } 1903 1904 bool Driver::HandleImmediateArgs(const Compilation &C) { 1905 // The order these options are handled in gcc is all over the place, but we 1906 // don't expect inconsistencies w.r.t. that to matter in practice. 1907 1908 if (C.getArgs().hasArg(options::OPT_dumpmachine)) { 1909 llvm::outs() << C.getDefaultToolChain().getTripleString() << '\n'; 1910 return false; 1911 } 1912 1913 if (C.getArgs().hasArg(options::OPT_dumpversion)) { 1914 // Since -dumpversion is only implemented for pedantic GCC compatibility, we 1915 // return an answer which matches our definition of __VERSION__. 1916 llvm::outs() << CLANG_VERSION_STRING << "\n"; 1917 return false; 1918 } 1919 1920 if (C.getArgs().hasArg(options::OPT__print_diagnostic_categories)) { 1921 PrintDiagnosticCategories(llvm::outs()); 1922 return false; 1923 } 1924 1925 if (C.getArgs().hasArg(options::OPT_help) || 1926 C.getArgs().hasArg(options::OPT__help_hidden)) { 1927 PrintHelp(C.getArgs().hasArg(options::OPT__help_hidden)); 1928 return false; 1929 } 1930 1931 if (C.getArgs().hasArg(options::OPT__version)) { 1932 // Follow gcc behavior and use stdout for --version and stderr for -v. 1933 PrintVersion(C, llvm::outs()); 1934 return false; 1935 } 1936 1937 if (C.getArgs().hasArg(options::OPT_v) || 1938 C.getArgs().hasArg(options::OPT__HASH_HASH_HASH) || 1939 C.getArgs().hasArg(options::OPT_print_supported_cpus)) { 1940 PrintVersion(C, llvm::errs()); 1941 SuppressMissingInputWarning = true; 1942 } 1943 1944 if (C.getArgs().hasArg(options::OPT_v)) { 1945 if (!SystemConfigDir.empty()) 1946 llvm::errs() << "System configuration file directory: " 1947 << SystemConfigDir << "\n"; 1948 if (!UserConfigDir.empty()) 1949 llvm::errs() << "User configuration file directory: " 1950 << UserConfigDir << "\n"; 1951 } 1952 1953 const ToolChain &TC = C.getDefaultToolChain(); 1954 1955 if (C.getArgs().hasArg(options::OPT_v)) 1956 TC.printVerboseInfo(llvm::errs()); 1957 1958 if (C.getArgs().hasArg(options::OPT_print_resource_dir)) { 1959 llvm::outs() << ResourceDir << '\n'; 1960 return false; 1961 } 1962 1963 if (C.getArgs().hasArg(options::OPT_print_search_dirs)) { 1964 llvm::outs() << "programs: ="; 1965 bool separator = false; 1966 // Print -B and COMPILER_PATH. 1967 for (const std::string &Path : PrefixDirs) { 1968 if (separator) 1969 llvm::outs() << llvm::sys::EnvPathSeparator; 1970 llvm::outs() << Path; 1971 separator = true; 1972 } 1973 for (const std::string &Path : TC.getProgramPaths()) { 1974 if (separator) 1975 llvm::outs() << llvm::sys::EnvPathSeparator; 1976 llvm::outs() << Path; 1977 separator = true; 1978 } 1979 llvm::outs() << "\n"; 1980 llvm::outs() << "libraries: =" << ResourceDir; 1981 1982 StringRef sysroot = C.getSysRoot(); 1983 1984 for (const std::string &Path : TC.getFilePaths()) { 1985 // Always print a separator. ResourceDir was the first item shown. 1986 llvm::outs() << llvm::sys::EnvPathSeparator; 1987 // Interpretation of leading '=' is needed only for NetBSD. 1988 if (Path[0] == '=') 1989 llvm::outs() << sysroot << Path.substr(1); 1990 else 1991 llvm::outs() << Path; 1992 } 1993 llvm::outs() << "\n"; 1994 return false; 1995 } 1996 1997 if (C.getArgs().hasArg(options::OPT_print_runtime_dir)) { 1998 std::string RuntimePath; 1999 // Get the first existing path, if any. 2000 for (auto Path : TC.getRuntimePaths()) { 2001 if (getVFS().exists(Path)) { 2002 RuntimePath = Path; 2003 break; 2004 } 2005 } 2006 if (!RuntimePath.empty()) 2007 llvm::outs() << RuntimePath << '\n'; 2008 else 2009 llvm::outs() << TC.getCompilerRTPath() << '\n'; 2010 return false; 2011 } 2012 2013 if (C.getArgs().hasArg(options::OPT_print_diagnostic_options)) { 2014 std::vector<std::string> Flags = DiagnosticIDs::getDiagnosticFlags(); 2015 for (std::size_t I = 0; I != Flags.size(); I += 2) 2016 llvm::outs() << " " << Flags[I] << "\n " << Flags[I + 1] << "\n\n"; 2017 return false; 2018 } 2019 2020 // FIXME: The following handlers should use a callback mechanism, we don't 2021 // know what the client would like to do. 2022 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_file_name_EQ)) { 2023 llvm::outs() << GetFilePath(A->getValue(), TC) << "\n"; 2024 return false; 2025 } 2026 2027 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_prog_name_EQ)) { 2028 StringRef ProgName = A->getValue(); 2029 2030 // Null program name cannot have a path. 2031 if (! ProgName.empty()) 2032 llvm::outs() << GetProgramPath(ProgName, TC); 2033 2034 llvm::outs() << "\n"; 2035 return false; 2036 } 2037 2038 if (Arg *A = C.getArgs().getLastArg(options::OPT_autocomplete)) { 2039 StringRef PassedFlags = A->getValue(); 2040 HandleAutocompletions(PassedFlags); 2041 return false; 2042 } 2043 2044 if (C.getArgs().hasArg(options::OPT_print_libgcc_file_name)) { 2045 ToolChain::RuntimeLibType RLT = TC.GetRuntimeLibType(C.getArgs()); 2046 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2047 RegisterEffectiveTriple TripleRAII(TC, Triple); 2048 switch (RLT) { 2049 case ToolChain::RLT_CompilerRT: 2050 llvm::outs() << TC.getCompilerRT(C.getArgs(), "builtins") << "\n"; 2051 break; 2052 case ToolChain::RLT_Libgcc: 2053 llvm::outs() << GetFilePath("libgcc.a", TC) << "\n"; 2054 break; 2055 } 2056 return false; 2057 } 2058 2059 if (C.getArgs().hasArg(options::OPT_print_multi_lib)) { 2060 for (const Multilib &Multilib : TC.getMultilibs()) 2061 llvm::outs() << Multilib << "\n"; 2062 return false; 2063 } 2064 2065 if (C.getArgs().hasArg(options::OPT_print_multi_directory)) { 2066 const Multilib &Multilib = TC.getMultilib(); 2067 if (Multilib.gccSuffix().empty()) 2068 llvm::outs() << ".\n"; 2069 else { 2070 StringRef Suffix(Multilib.gccSuffix()); 2071 assert(Suffix.front() == '/'); 2072 llvm::outs() << Suffix.substr(1) << "\n"; 2073 } 2074 return false; 2075 } 2076 2077 if (C.getArgs().hasArg(options::OPT_print_target_triple)) { 2078 llvm::outs() << TC.getTripleString() << "\n"; 2079 return false; 2080 } 2081 2082 if (C.getArgs().hasArg(options::OPT_print_effective_triple)) { 2083 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2084 llvm::outs() << Triple.getTriple() << "\n"; 2085 return false; 2086 } 2087 2088 if (C.getArgs().hasArg(options::OPT_print_multiarch)) { 2089 llvm::outs() << TC.getMultiarchTriple(*this, TC.getTriple(), SysRoot) 2090 << "\n"; 2091 return false; 2092 } 2093 2094 if (C.getArgs().hasArg(options::OPT_print_targets)) { 2095 llvm::TargetRegistry::printRegisteredTargetsForVersion(llvm::outs()); 2096 return false; 2097 } 2098 2099 return true; 2100 } 2101 2102 enum { 2103 TopLevelAction = 0, 2104 HeadSibAction = 1, 2105 OtherSibAction = 2, 2106 }; 2107 2108 // Display an action graph human-readably. Action A is the "sink" node 2109 // and latest-occuring action. Traversal is in pre-order, visiting the 2110 // inputs to each action before printing the action itself. 2111 static unsigned PrintActions1(const Compilation &C, Action *A, 2112 std::map<Action *, unsigned> &Ids, 2113 Twine Indent = {}, int Kind = TopLevelAction) { 2114 if (Ids.count(A)) // A was already visited. 2115 return Ids[A]; 2116 2117 std::string str; 2118 llvm::raw_string_ostream os(str); 2119 2120 auto getSibIndent = [](int K) -> Twine { 2121 return (K == HeadSibAction) ? " " : (K == OtherSibAction) ? "| " : ""; 2122 }; 2123 2124 Twine SibIndent = Indent + getSibIndent(Kind); 2125 int SibKind = HeadSibAction; 2126 os << Action::getClassName(A->getKind()) << ", "; 2127 if (InputAction *IA = dyn_cast<InputAction>(A)) { 2128 os << "\"" << IA->getInputArg().getValue() << "\""; 2129 } else if (BindArchAction *BIA = dyn_cast<BindArchAction>(A)) { 2130 os << '"' << BIA->getArchName() << '"' << ", {" 2131 << PrintActions1(C, *BIA->input_begin(), Ids, SibIndent, SibKind) << "}"; 2132 } else if (OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 2133 bool IsFirst = true; 2134 OA->doOnEachDependence( 2135 [&](Action *A, const ToolChain *TC, const char *BoundArch) { 2136 assert(TC && "Unknown host toolchain"); 2137 // E.g. for two CUDA device dependences whose bound arch is sm_20 and 2138 // sm_35 this will generate: 2139 // "cuda-device" (nvptx64-nvidia-cuda:sm_20) {#ID}, "cuda-device" 2140 // (nvptx64-nvidia-cuda:sm_35) {#ID} 2141 if (!IsFirst) 2142 os << ", "; 2143 os << '"'; 2144 os << A->getOffloadingKindPrefix(); 2145 os << " ("; 2146 os << TC->getTriple().normalize(); 2147 if (BoundArch) 2148 os << ":" << BoundArch; 2149 os << ")"; 2150 os << '"'; 2151 os << " {" << PrintActions1(C, A, Ids, SibIndent, SibKind) << "}"; 2152 IsFirst = false; 2153 SibKind = OtherSibAction; 2154 }); 2155 } else { 2156 const ActionList *AL = &A->getInputs(); 2157 2158 if (AL->size()) { 2159 const char *Prefix = "{"; 2160 for (Action *PreRequisite : *AL) { 2161 os << Prefix << PrintActions1(C, PreRequisite, Ids, SibIndent, SibKind); 2162 Prefix = ", "; 2163 SibKind = OtherSibAction; 2164 } 2165 os << "}"; 2166 } else 2167 os << "{}"; 2168 } 2169 2170 // Append offload info for all options other than the offloading action 2171 // itself (e.g. (cuda-device, sm_20) or (cuda-host)). 2172 std::string offload_str; 2173 llvm::raw_string_ostream offload_os(offload_str); 2174 if (!isa<OffloadAction>(A)) { 2175 auto S = A->getOffloadingKindPrefix(); 2176 if (!S.empty()) { 2177 offload_os << ", (" << S; 2178 if (A->getOffloadingArch()) 2179 offload_os << ", " << A->getOffloadingArch(); 2180 offload_os << ")"; 2181 } 2182 } 2183 2184 auto getSelfIndent = [](int K) -> Twine { 2185 return (K == HeadSibAction) ? "+- " : (K == OtherSibAction) ? "|- " : ""; 2186 }; 2187 2188 unsigned Id = Ids.size(); 2189 Ids[A] = Id; 2190 llvm::errs() << Indent + getSelfIndent(Kind) << Id << ": " << os.str() << ", " 2191 << types::getTypeName(A->getType()) << offload_os.str() << "\n"; 2192 2193 return Id; 2194 } 2195 2196 // Print the action graphs in a compilation C. 2197 // For example "clang -c file1.c file2.c" is composed of two subgraphs. 2198 void Driver::PrintActions(const Compilation &C) const { 2199 std::map<Action *, unsigned> Ids; 2200 for (Action *A : C.getActions()) 2201 PrintActions1(C, A, Ids); 2202 } 2203 2204 /// Check whether the given input tree contains any compilation or 2205 /// assembly actions. 2206 static bool ContainsCompileOrAssembleAction(const Action *A) { 2207 if (isa<CompileJobAction>(A) || isa<BackendJobAction>(A) || 2208 isa<AssembleJobAction>(A)) 2209 return true; 2210 2211 return llvm::any_of(A->inputs(), ContainsCompileOrAssembleAction); 2212 } 2213 2214 void Driver::BuildUniversalActions(Compilation &C, const ToolChain &TC, 2215 const InputList &BAInputs) const { 2216 DerivedArgList &Args = C.getArgs(); 2217 ActionList &Actions = C.getActions(); 2218 llvm::PrettyStackTraceString CrashInfo("Building universal build actions"); 2219 // Collect the list of architectures. Duplicates are allowed, but should only 2220 // be handled once (in the order seen). 2221 llvm::StringSet<> ArchNames; 2222 SmallVector<const char *, 4> Archs; 2223 for (Arg *A : Args) { 2224 if (A->getOption().matches(options::OPT_arch)) { 2225 // Validate the option here; we don't save the type here because its 2226 // particular spelling may participate in other driver choices. 2227 llvm::Triple::ArchType Arch = 2228 tools::darwin::getArchTypeForMachOArchName(A->getValue()); 2229 if (Arch == llvm::Triple::UnknownArch) { 2230 Diag(clang::diag::err_drv_invalid_arch_name) << A->getAsString(Args); 2231 continue; 2232 } 2233 2234 A->claim(); 2235 if (ArchNames.insert(A->getValue()).second) 2236 Archs.push_back(A->getValue()); 2237 } 2238 } 2239 2240 // When there is no explicit arch for this platform, make sure we still bind 2241 // the architecture (to the default) so that -Xarch_ is handled correctly. 2242 if (!Archs.size()) 2243 Archs.push_back(Args.MakeArgString(TC.getDefaultUniversalArchName())); 2244 2245 ActionList SingleActions; 2246 BuildActions(C, Args, BAInputs, SingleActions); 2247 2248 // Add in arch bindings for every top level action, as well as lipo and 2249 // dsymutil steps if needed. 2250 for (Action* Act : SingleActions) { 2251 // Make sure we can lipo this kind of output. If not (and it is an actual 2252 // output) then we disallow, since we can't create an output file with the 2253 // right name without overwriting it. We could remove this oddity by just 2254 // changing the output names to include the arch, which would also fix 2255 // -save-temps. Compatibility wins for now. 2256 2257 if (Archs.size() > 1 && !types::canLipoType(Act->getType())) 2258 Diag(clang::diag::err_drv_invalid_output_with_multiple_archs) 2259 << types::getTypeName(Act->getType()); 2260 2261 ActionList Inputs; 2262 for (unsigned i = 0, e = Archs.size(); i != e; ++i) 2263 Inputs.push_back(C.MakeAction<BindArchAction>(Act, Archs[i])); 2264 2265 // Lipo if necessary, we do it this way because we need to set the arch flag 2266 // so that -Xarch_ gets overwritten. 2267 if (Inputs.size() == 1 || Act->getType() == types::TY_Nothing) 2268 Actions.append(Inputs.begin(), Inputs.end()); 2269 else 2270 Actions.push_back(C.MakeAction<LipoJobAction>(Inputs, Act->getType())); 2271 2272 // Handle debug info queries. 2273 Arg *A = Args.getLastArg(options::OPT_g_Group); 2274 bool enablesDebugInfo = A && !A->getOption().matches(options::OPT_g0) && 2275 !A->getOption().matches(options::OPT_gstabs); 2276 if ((enablesDebugInfo || willEmitRemarks(Args)) && 2277 ContainsCompileOrAssembleAction(Actions.back())) { 2278 2279 // Add a 'dsymutil' step if necessary, when debug info is enabled and we 2280 // have a compile input. We need to run 'dsymutil' ourselves in such cases 2281 // because the debug info will refer to a temporary object file which 2282 // will be removed at the end of the compilation process. 2283 if (Act->getType() == types::TY_Image) { 2284 ActionList Inputs; 2285 Inputs.push_back(Actions.back()); 2286 Actions.pop_back(); 2287 Actions.push_back( 2288 C.MakeAction<DsymutilJobAction>(Inputs, types::TY_dSYM)); 2289 } 2290 2291 // Verify the debug info output. 2292 if (Args.hasArg(options::OPT_verify_debug_info)) { 2293 Action* LastAction = Actions.back(); 2294 Actions.pop_back(); 2295 Actions.push_back(C.MakeAction<VerifyDebugInfoJobAction>( 2296 LastAction, types::TY_Nothing)); 2297 } 2298 } 2299 } 2300 } 2301 2302 bool Driver::DiagnoseInputExistence(const DerivedArgList &Args, StringRef Value, 2303 types::ID Ty, bool TypoCorrect) const { 2304 if (!getCheckInputsExist()) 2305 return true; 2306 2307 // stdin always exists. 2308 if (Value == "-") 2309 return true; 2310 2311 // If it's a header to be found in the system or user search path, then defer 2312 // complaints about its absence until those searches can be done. When we 2313 // are definitely processing headers for C++20 header units, extend this to 2314 // allow the user to put "-fmodule-header -xc++-header vector" for example. 2315 if (Ty == types::TY_CXXSHeader || Ty == types::TY_CXXUHeader || 2316 (ModulesModeCXX20 && Ty == types::TY_CXXHeader)) 2317 return true; 2318 2319 if (getVFS().exists(Value)) 2320 return true; 2321 2322 if (TypoCorrect) { 2323 // Check if the filename is a typo for an option flag. OptTable thinks 2324 // that all args that are not known options and that start with / are 2325 // filenames, but e.g. `/diagnostic:caret` is more likely a typo for 2326 // the option `/diagnostics:caret` than a reference to a file in the root 2327 // directory. 2328 unsigned IncludedFlagsBitmask; 2329 unsigned ExcludedFlagsBitmask; 2330 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 2331 getIncludeExcludeOptionFlagMasks(IsCLMode()); 2332 std::string Nearest; 2333 if (getOpts().findNearest(Value, Nearest, IncludedFlagsBitmask, 2334 ExcludedFlagsBitmask) <= 1) { 2335 Diag(clang::diag::err_drv_no_such_file_with_suggestion) 2336 << Value << Nearest; 2337 return false; 2338 } 2339 } 2340 2341 // In CL mode, don't error on apparently non-existent linker inputs, because 2342 // they can be influenced by linker flags the clang driver might not 2343 // understand. 2344 // Examples: 2345 // - `clang-cl main.cc ole32.lib` in a a non-MSVC shell will make the driver 2346 // module look for an MSVC installation in the registry. (We could ask 2347 // the MSVCToolChain object if it can find `ole32.lib`, but the logic to 2348 // look in the registry might move into lld-link in the future so that 2349 // lld-link invocations in non-MSVC shells just work too.) 2350 // - `clang-cl ... /link ...` can pass arbitrary flags to the linker, 2351 // including /libpath:, which is used to find .lib and .obj files. 2352 // So do not diagnose this on the driver level. Rely on the linker diagnosing 2353 // it. (If we don't end up invoking the linker, this means we'll emit a 2354 // "'linker' input unused [-Wunused-command-line-argument]" warning instead 2355 // of an error.) 2356 // 2357 // Only do this skip after the typo correction step above. `/Brepo` is treated 2358 // as TY_Object, but it's clearly a typo for `/Brepro`. It seems fine to emit 2359 // an error if we have a flag that's within an edit distance of 1 from a 2360 // flag. (Users can use `-Wl,` or `/linker` to launder the flag past the 2361 // driver in the unlikely case they run into this.) 2362 // 2363 // Don't do this for inputs that start with a '/', else we'd pass options 2364 // like /libpath: through to the linker silently. 2365 // 2366 // Emitting an error for linker inputs can also cause incorrect diagnostics 2367 // with the gcc driver. The command 2368 // clang -fuse-ld=lld -Wl,--chroot,some/dir /file.o 2369 // will make lld look for some/dir/file.o, while we will diagnose here that 2370 // `/file.o` does not exist. However, configure scripts check if 2371 // `clang /GR-` compiles without error to see if the compiler is cl.exe, 2372 // so we can't downgrade diagnostics for `/GR-` from an error to a warning 2373 // in cc mode. (We can in cl mode because cl.exe itself only warns on 2374 // unknown flags.) 2375 if (IsCLMode() && Ty == types::TY_Object && !Value.startswith("/")) 2376 return true; 2377 2378 Diag(clang::diag::err_drv_no_such_file) << Value; 2379 return false; 2380 } 2381 2382 // Get the C++20 Header Unit type corresponding to the input type. 2383 static types::ID CXXHeaderUnitType(ModuleHeaderMode HM) { 2384 switch (HM) { 2385 case HeaderMode_User: 2386 return types::TY_CXXUHeader; 2387 case HeaderMode_System: 2388 return types::TY_CXXSHeader; 2389 case HeaderMode_Default: 2390 break; 2391 case HeaderMode_None: 2392 llvm_unreachable("should not be called in this case"); 2393 } 2394 return types::TY_CXXHUHeader; 2395 } 2396 2397 // Construct a the list of inputs and their types. 2398 void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args, 2399 InputList &Inputs) const { 2400 const llvm::opt::OptTable &Opts = getOpts(); 2401 // Track the current user specified (-x) input. We also explicitly track the 2402 // argument used to set the type; we only want to claim the type when we 2403 // actually use it, so we warn about unused -x arguments. 2404 types::ID InputType = types::TY_Nothing; 2405 Arg *InputTypeArg = nullptr; 2406 2407 // The last /TC or /TP option sets the input type to C or C++ globally. 2408 if (Arg *TCTP = Args.getLastArgNoClaim(options::OPT__SLASH_TC, 2409 options::OPT__SLASH_TP)) { 2410 InputTypeArg = TCTP; 2411 InputType = TCTP->getOption().matches(options::OPT__SLASH_TC) 2412 ? types::TY_C 2413 : types::TY_CXX; 2414 2415 Arg *Previous = nullptr; 2416 bool ShowNote = false; 2417 for (Arg *A : 2418 Args.filtered(options::OPT__SLASH_TC, options::OPT__SLASH_TP)) { 2419 if (Previous) { 2420 Diag(clang::diag::warn_drv_overriding_flag_option) 2421 << Previous->getSpelling() << A->getSpelling(); 2422 ShowNote = true; 2423 } 2424 Previous = A; 2425 } 2426 if (ShowNote) 2427 Diag(clang::diag::note_drv_t_option_is_global); 2428 2429 // No driver mode exposes -x and /TC or /TP; we don't support mixing them. 2430 assert(!Args.hasArg(options::OPT_x) && "-x and /TC or /TP is not allowed"); 2431 } 2432 2433 // Warn -x after last input file has no effect 2434 { 2435 Arg *LastXArg = Args.getLastArgNoClaim(options::OPT_x); 2436 Arg *LastInputArg = Args.getLastArgNoClaim(options::OPT_INPUT); 2437 if (LastXArg && LastInputArg && LastInputArg->getIndex() < LastXArg->getIndex()) 2438 Diag(clang::diag::warn_drv_unused_x) << LastXArg->getValue(); 2439 } 2440 2441 for (Arg *A : Args) { 2442 if (A->getOption().getKind() == Option::InputClass) { 2443 const char *Value = A->getValue(); 2444 types::ID Ty = types::TY_INVALID; 2445 2446 // Infer the input type if necessary. 2447 if (InputType == types::TY_Nothing) { 2448 // If there was an explicit arg for this, claim it. 2449 if (InputTypeArg) 2450 InputTypeArg->claim(); 2451 2452 // stdin must be handled specially. 2453 if (memcmp(Value, "-", 2) == 0) { 2454 if (IsFlangMode()) { 2455 Ty = types::TY_Fortran; 2456 } else { 2457 // If running with -E, treat as a C input (this changes the 2458 // builtin macros, for example). This may be overridden by -ObjC 2459 // below. 2460 // 2461 // Otherwise emit an error but still use a valid type to avoid 2462 // spurious errors (e.g., no inputs). 2463 assert(!CCGenDiagnostics && "stdin produces no crash reproducer"); 2464 if (!Args.hasArgNoClaim(options::OPT_E) && !CCCIsCPP()) 2465 Diag(IsCLMode() ? clang::diag::err_drv_unknown_stdin_type_clang_cl 2466 : clang::diag::err_drv_unknown_stdin_type); 2467 Ty = types::TY_C; 2468 } 2469 } else { 2470 // Otherwise lookup by extension. 2471 // Fallback is C if invoked as C preprocessor, C++ if invoked with 2472 // clang-cl /E, or Object otherwise. 2473 // We use a host hook here because Darwin at least has its own 2474 // idea of what .s is. 2475 if (const char *Ext = strrchr(Value, '.')) 2476 Ty = TC.LookupTypeForExtension(Ext + 1); 2477 2478 if (Ty == types::TY_INVALID) { 2479 if (IsCLMode() && (Args.hasArgNoClaim(options::OPT_E) || CCGenDiagnostics)) 2480 Ty = types::TY_CXX; 2481 else if (CCCIsCPP() || CCGenDiagnostics) 2482 Ty = types::TY_C; 2483 else 2484 Ty = types::TY_Object; 2485 } 2486 2487 // If the driver is invoked as C++ compiler (like clang++ or c++) it 2488 // should autodetect some input files as C++ for g++ compatibility. 2489 if (CCCIsCXX()) { 2490 types::ID OldTy = Ty; 2491 Ty = types::lookupCXXTypeForCType(Ty); 2492 2493 // Do not complain about foo.h, when we are known to be processing 2494 // it as a C++20 header unit. 2495 if (Ty != OldTy && !(OldTy == types::TY_CHeader && hasHeaderMode())) 2496 Diag(clang::diag::warn_drv_treating_input_as_cxx) 2497 << getTypeName(OldTy) << getTypeName(Ty); 2498 } 2499 2500 // If running with -fthinlto-index=, extensions that normally identify 2501 // native object files actually identify LLVM bitcode files. 2502 if (Args.hasArgNoClaim(options::OPT_fthinlto_index_EQ) && 2503 Ty == types::TY_Object) 2504 Ty = types::TY_LLVM_BC; 2505 } 2506 2507 // -ObjC and -ObjC++ override the default language, but only for "source 2508 // files". We just treat everything that isn't a linker input as a 2509 // source file. 2510 // 2511 // FIXME: Clean this up if we move the phase sequence into the type. 2512 if (Ty != types::TY_Object) { 2513 if (Args.hasArg(options::OPT_ObjC)) 2514 Ty = types::TY_ObjC; 2515 else if (Args.hasArg(options::OPT_ObjCXX)) 2516 Ty = types::TY_ObjCXX; 2517 } 2518 2519 // Disambiguate headers that are meant to be header units from those 2520 // intended to be PCH. Avoid missing '.h' cases that are counted as 2521 // C headers by default - we know we are in C++ mode and we do not 2522 // want to issue a complaint about compiling things in the wrong mode. 2523 if ((Ty == types::TY_CXXHeader || Ty == types::TY_CHeader) && 2524 hasHeaderMode()) 2525 Ty = CXXHeaderUnitType(CXX20HeaderType); 2526 } else { 2527 assert(InputTypeArg && "InputType set w/o InputTypeArg"); 2528 if (!InputTypeArg->getOption().matches(options::OPT_x)) { 2529 // If emulating cl.exe, make sure that /TC and /TP don't affect input 2530 // object files. 2531 const char *Ext = strrchr(Value, '.'); 2532 if (Ext && TC.LookupTypeForExtension(Ext + 1) == types::TY_Object) 2533 Ty = types::TY_Object; 2534 } 2535 if (Ty == types::TY_INVALID) { 2536 Ty = InputType; 2537 InputTypeArg->claim(); 2538 } 2539 } 2540 2541 if (DiagnoseInputExistence(Args, Value, Ty, /*TypoCorrect=*/true)) 2542 Inputs.push_back(std::make_pair(Ty, A)); 2543 2544 } else if (A->getOption().matches(options::OPT__SLASH_Tc)) { 2545 StringRef Value = A->getValue(); 2546 if (DiagnoseInputExistence(Args, Value, types::TY_C, 2547 /*TypoCorrect=*/false)) { 2548 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2549 Inputs.push_back(std::make_pair(types::TY_C, InputArg)); 2550 } 2551 A->claim(); 2552 } else if (A->getOption().matches(options::OPT__SLASH_Tp)) { 2553 StringRef Value = A->getValue(); 2554 if (DiagnoseInputExistence(Args, Value, types::TY_CXX, 2555 /*TypoCorrect=*/false)) { 2556 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2557 Inputs.push_back(std::make_pair(types::TY_CXX, InputArg)); 2558 } 2559 A->claim(); 2560 } else if (A->getOption().hasFlag(options::LinkerInput)) { 2561 // Just treat as object type, we could make a special type for this if 2562 // necessary. 2563 Inputs.push_back(std::make_pair(types::TY_Object, A)); 2564 2565 } else if (A->getOption().matches(options::OPT_x)) { 2566 InputTypeArg = A; 2567 InputType = types::lookupTypeForTypeSpecifier(A->getValue()); 2568 A->claim(); 2569 2570 // Follow gcc behavior and treat as linker input for invalid -x 2571 // options. Its not clear why we shouldn't just revert to unknown; but 2572 // this isn't very important, we might as well be bug compatible. 2573 if (!InputType) { 2574 Diag(clang::diag::err_drv_unknown_language) << A->getValue(); 2575 InputType = types::TY_Object; 2576 } 2577 2578 // If the user has put -fmodule-header{,=} then we treat C++ headers as 2579 // header unit inputs. So we 'promote' -xc++-header appropriately. 2580 if (InputType == types::TY_CXXHeader && hasHeaderMode()) 2581 InputType = CXXHeaderUnitType(CXX20HeaderType); 2582 } else if (A->getOption().getID() == options::OPT_U) { 2583 assert(A->getNumValues() == 1 && "The /U option has one value."); 2584 StringRef Val = A->getValue(0); 2585 if (Val.find_first_of("/\\") != StringRef::npos) { 2586 // Warn about e.g. "/Users/me/myfile.c". 2587 Diag(diag::warn_slash_u_filename) << Val; 2588 Diag(diag::note_use_dashdash); 2589 } 2590 } 2591 } 2592 if (CCCIsCPP() && Inputs.empty()) { 2593 // If called as standalone preprocessor, stdin is processed 2594 // if no other input is present. 2595 Arg *A = MakeInputArg(Args, Opts, "-"); 2596 Inputs.push_back(std::make_pair(types::TY_C, A)); 2597 } 2598 } 2599 2600 namespace { 2601 /// Provides a convenient interface for different programming models to generate 2602 /// the required device actions. 2603 class OffloadingActionBuilder final { 2604 /// Flag used to trace errors in the builder. 2605 bool IsValid = false; 2606 2607 /// The compilation that is using this builder. 2608 Compilation &C; 2609 2610 /// Map between an input argument and the offload kinds used to process it. 2611 std::map<const Arg *, unsigned> InputArgToOffloadKindMap; 2612 2613 /// Map between a host action and its originating input argument. 2614 std::map<Action *, const Arg *> HostActionToInputArgMap; 2615 2616 /// Builder interface. It doesn't build anything or keep any state. 2617 class DeviceActionBuilder { 2618 public: 2619 typedef const llvm::SmallVectorImpl<phases::ID> PhasesTy; 2620 2621 enum ActionBuilderReturnCode { 2622 // The builder acted successfully on the current action. 2623 ABRT_Success, 2624 // The builder didn't have to act on the current action. 2625 ABRT_Inactive, 2626 // The builder was successful and requested the host action to not be 2627 // generated. 2628 ABRT_Ignore_Host, 2629 }; 2630 2631 protected: 2632 /// Compilation associated with this builder. 2633 Compilation &C; 2634 2635 /// Tool chains associated with this builder. The same programming 2636 /// model may have associated one or more tool chains. 2637 SmallVector<const ToolChain *, 2> ToolChains; 2638 2639 /// The derived arguments associated with this builder. 2640 DerivedArgList &Args; 2641 2642 /// The inputs associated with this builder. 2643 const Driver::InputList &Inputs; 2644 2645 /// The associated offload kind. 2646 Action::OffloadKind AssociatedOffloadKind = Action::OFK_None; 2647 2648 public: 2649 DeviceActionBuilder(Compilation &C, DerivedArgList &Args, 2650 const Driver::InputList &Inputs, 2651 Action::OffloadKind AssociatedOffloadKind) 2652 : C(C), Args(Args), Inputs(Inputs), 2653 AssociatedOffloadKind(AssociatedOffloadKind) {} 2654 virtual ~DeviceActionBuilder() {} 2655 2656 /// Fill up the array \a DA with all the device dependences that should be 2657 /// added to the provided host action \a HostAction. By default it is 2658 /// inactive. 2659 virtual ActionBuilderReturnCode 2660 getDeviceDependences(OffloadAction::DeviceDependences &DA, 2661 phases::ID CurPhase, phases::ID FinalPhase, 2662 PhasesTy &Phases) { 2663 return ABRT_Inactive; 2664 } 2665 2666 /// Update the state to include the provided host action \a HostAction as a 2667 /// dependency of the current device action. By default it is inactive. 2668 virtual ActionBuilderReturnCode addDeviceDepences(Action *HostAction) { 2669 return ABRT_Inactive; 2670 } 2671 2672 /// Append top level actions generated by the builder. 2673 virtual void appendTopLevelActions(ActionList &AL) {} 2674 2675 /// Append linker device actions generated by the builder. 2676 virtual void appendLinkDeviceActions(ActionList &AL) {} 2677 2678 /// Append linker host action generated by the builder. 2679 virtual Action* appendLinkHostActions(ActionList &AL) { return nullptr; } 2680 2681 /// Append linker actions generated by the builder. 2682 virtual void appendLinkDependences(OffloadAction::DeviceDependences &DA) {} 2683 2684 /// Initialize the builder. Return true if any initialization errors are 2685 /// found. 2686 virtual bool initialize() { return false; } 2687 2688 /// Return true if the builder can use bundling/unbundling. 2689 virtual bool canUseBundlerUnbundler() const { return false; } 2690 2691 /// Return true if this builder is valid. We have a valid builder if we have 2692 /// associated device tool chains. 2693 bool isValid() { return !ToolChains.empty(); } 2694 2695 /// Return the associated offload kind. 2696 Action::OffloadKind getAssociatedOffloadKind() { 2697 return AssociatedOffloadKind; 2698 } 2699 }; 2700 2701 /// Base class for CUDA/HIP action builder. It injects device code in 2702 /// the host backend action. 2703 class CudaActionBuilderBase : public DeviceActionBuilder { 2704 protected: 2705 /// Flags to signal if the user requested host-only or device-only 2706 /// compilation. 2707 bool CompileHostOnly = false; 2708 bool CompileDeviceOnly = false; 2709 bool EmitLLVM = false; 2710 bool EmitAsm = false; 2711 2712 /// ID to identify each device compilation. For CUDA it is simply the 2713 /// GPU arch string. For HIP it is either the GPU arch string or GPU 2714 /// arch string plus feature strings delimited by a plus sign, e.g. 2715 /// gfx906+xnack. 2716 struct TargetID { 2717 /// Target ID string which is persistent throughout the compilation. 2718 const char *ID; 2719 TargetID(CudaArch Arch) { ID = CudaArchToString(Arch); } 2720 TargetID(const char *ID) : ID(ID) {} 2721 operator const char *() { return ID; } 2722 operator StringRef() { return StringRef(ID); } 2723 }; 2724 /// List of GPU architectures to use in this compilation. 2725 SmallVector<TargetID, 4> GpuArchList; 2726 2727 /// The CUDA actions for the current input. 2728 ActionList CudaDeviceActions; 2729 2730 /// The CUDA fat binary if it was generated for the current input. 2731 Action *CudaFatBinary = nullptr; 2732 2733 /// Flag that is set to true if this builder acted on the current input. 2734 bool IsActive = false; 2735 2736 /// Flag for -fgpu-rdc. 2737 bool Relocatable = false; 2738 2739 /// Default GPU architecture if there's no one specified. 2740 CudaArch DefaultCudaArch = CudaArch::UNKNOWN; 2741 2742 /// Method to generate compilation unit ID specified by option 2743 /// '-fuse-cuid='. 2744 enum UseCUIDKind { CUID_Hash, CUID_Random, CUID_None, CUID_Invalid }; 2745 UseCUIDKind UseCUID = CUID_Hash; 2746 2747 /// Compilation unit ID specified by option '-cuid='. 2748 StringRef FixedCUID; 2749 2750 public: 2751 CudaActionBuilderBase(Compilation &C, DerivedArgList &Args, 2752 const Driver::InputList &Inputs, 2753 Action::OffloadKind OFKind) 2754 : DeviceActionBuilder(C, Args, Inputs, OFKind) {} 2755 2756 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 2757 // While generating code for CUDA, we only depend on the host input action 2758 // to trigger the creation of all the CUDA device actions. 2759 2760 // If we are dealing with an input action, replicate it for each GPU 2761 // architecture. If we are in host-only mode we return 'success' so that 2762 // the host uses the CUDA offload kind. 2763 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 2764 assert(!GpuArchList.empty() && 2765 "We should have at least one GPU architecture."); 2766 2767 // If the host input is not CUDA or HIP, we don't need to bother about 2768 // this input. 2769 if (!(IA->getType() == types::TY_CUDA || 2770 IA->getType() == types::TY_HIP || 2771 IA->getType() == types::TY_PP_HIP)) { 2772 // The builder will ignore this input. 2773 IsActive = false; 2774 return ABRT_Inactive; 2775 } 2776 2777 // Set the flag to true, so that the builder acts on the current input. 2778 IsActive = true; 2779 2780 if (CompileHostOnly) 2781 return ABRT_Success; 2782 2783 // Replicate inputs for each GPU architecture. 2784 auto Ty = IA->getType() == types::TY_HIP ? types::TY_HIP_DEVICE 2785 : types::TY_CUDA_DEVICE; 2786 std::string CUID = FixedCUID.str(); 2787 if (CUID.empty()) { 2788 if (UseCUID == CUID_Random) 2789 CUID = llvm::utohexstr(llvm::sys::Process::GetRandomNumber(), 2790 /*LowerCase=*/true); 2791 else if (UseCUID == CUID_Hash) { 2792 llvm::MD5 Hasher; 2793 llvm::MD5::MD5Result Hash; 2794 SmallString<256> RealPath; 2795 llvm::sys::fs::real_path(IA->getInputArg().getValue(), RealPath, 2796 /*expand_tilde=*/true); 2797 Hasher.update(RealPath); 2798 for (auto *A : Args) { 2799 if (A->getOption().matches(options::OPT_INPUT)) 2800 continue; 2801 Hasher.update(A->getAsString(Args)); 2802 } 2803 Hasher.final(Hash); 2804 CUID = llvm::utohexstr(Hash.low(), /*LowerCase=*/true); 2805 } 2806 } 2807 IA->setId(CUID); 2808 2809 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 2810 CudaDeviceActions.push_back( 2811 C.MakeAction<InputAction>(IA->getInputArg(), Ty, IA->getId())); 2812 } 2813 2814 return ABRT_Success; 2815 } 2816 2817 // If this is an unbundling action use it as is for each CUDA toolchain. 2818 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 2819 2820 // If -fgpu-rdc is disabled, should not unbundle since there is no 2821 // device code to link. 2822 if (UA->getType() == types::TY_Object && !Relocatable) 2823 return ABRT_Inactive; 2824 2825 CudaDeviceActions.clear(); 2826 auto *IA = cast<InputAction>(UA->getInputs().back()); 2827 std::string FileName = IA->getInputArg().getAsString(Args); 2828 // Check if the type of the file is the same as the action. Do not 2829 // unbundle it if it is not. Do not unbundle .so files, for example, 2830 // which are not object files. 2831 if (IA->getType() == types::TY_Object && 2832 (!llvm::sys::path::has_extension(FileName) || 2833 types::lookupTypeForExtension( 2834 llvm::sys::path::extension(FileName).drop_front()) != 2835 types::TY_Object)) 2836 return ABRT_Inactive; 2837 2838 for (auto Arch : GpuArchList) { 2839 CudaDeviceActions.push_back(UA); 2840 UA->registerDependentActionInfo(ToolChains[0], Arch, 2841 AssociatedOffloadKind); 2842 } 2843 IsActive = true; 2844 return ABRT_Success; 2845 } 2846 2847 return IsActive ? ABRT_Success : ABRT_Inactive; 2848 } 2849 2850 void appendTopLevelActions(ActionList &AL) override { 2851 // Utility to append actions to the top level list. 2852 auto AddTopLevel = [&](Action *A, TargetID TargetID) { 2853 OffloadAction::DeviceDependences Dep; 2854 Dep.add(*A, *ToolChains.front(), TargetID, AssociatedOffloadKind); 2855 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 2856 }; 2857 2858 // If we have a fat binary, add it to the list. 2859 if (CudaFatBinary) { 2860 AddTopLevel(CudaFatBinary, CudaArch::UNUSED); 2861 CudaDeviceActions.clear(); 2862 CudaFatBinary = nullptr; 2863 return; 2864 } 2865 2866 if (CudaDeviceActions.empty()) 2867 return; 2868 2869 // If we have CUDA actions at this point, that's because we have a have 2870 // partial compilation, so we should have an action for each GPU 2871 // architecture. 2872 assert(CudaDeviceActions.size() == GpuArchList.size() && 2873 "Expecting one action per GPU architecture."); 2874 assert(ToolChains.size() == 1 && 2875 "Expecting to have a single CUDA toolchain."); 2876 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) 2877 AddTopLevel(CudaDeviceActions[I], GpuArchList[I]); 2878 2879 CudaDeviceActions.clear(); 2880 } 2881 2882 /// Get canonicalized offload arch option. \returns empty StringRef if the 2883 /// option is invalid. 2884 virtual StringRef getCanonicalOffloadArch(StringRef Arch) = 0; 2885 2886 virtual llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 2887 getConflictOffloadArchCombination(const std::set<StringRef> &GpuArchs) = 0; 2888 2889 bool initialize() override { 2890 assert(AssociatedOffloadKind == Action::OFK_Cuda || 2891 AssociatedOffloadKind == Action::OFK_HIP); 2892 2893 // We don't need to support CUDA. 2894 if (AssociatedOffloadKind == Action::OFK_Cuda && 2895 !C.hasOffloadToolChain<Action::OFK_Cuda>()) 2896 return false; 2897 2898 // We don't need to support HIP. 2899 if (AssociatedOffloadKind == Action::OFK_HIP && 2900 !C.hasOffloadToolChain<Action::OFK_HIP>()) 2901 return false; 2902 2903 Relocatable = Args.hasFlag(options::OPT_fgpu_rdc, 2904 options::OPT_fno_gpu_rdc, /*Default=*/false); 2905 2906 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 2907 assert(HostTC && "No toolchain for host compilation."); 2908 if (HostTC->getTriple().isNVPTX() || 2909 HostTC->getTriple().getArch() == llvm::Triple::amdgcn) { 2910 // We do not support targeting NVPTX/AMDGCN for host compilation. Throw 2911 // an error and abort pipeline construction early so we don't trip 2912 // asserts that assume device-side compilation. 2913 C.getDriver().Diag(diag::err_drv_cuda_host_arch) 2914 << HostTC->getTriple().getArchName(); 2915 return true; 2916 } 2917 2918 ToolChains.push_back( 2919 AssociatedOffloadKind == Action::OFK_Cuda 2920 ? C.getSingleOffloadToolChain<Action::OFK_Cuda>() 2921 : C.getSingleOffloadToolChain<Action::OFK_HIP>()); 2922 2923 Arg *PartialCompilationArg = Args.getLastArg( 2924 options::OPT_offload_host_only, options::OPT_offload_device_only, 2925 options::OPT_offload_host_device); 2926 CompileHostOnly = 2927 PartialCompilationArg && PartialCompilationArg->getOption().matches( 2928 options::OPT_offload_host_only); 2929 CompileDeviceOnly = 2930 PartialCompilationArg && PartialCompilationArg->getOption().matches( 2931 options::OPT_offload_device_only); 2932 EmitLLVM = Args.getLastArg(options::OPT_emit_llvm); 2933 EmitAsm = Args.getLastArg(options::OPT_S); 2934 FixedCUID = Args.getLastArgValue(options::OPT_cuid_EQ); 2935 if (Arg *A = Args.getLastArg(options::OPT_fuse_cuid_EQ)) { 2936 StringRef UseCUIDStr = A->getValue(); 2937 UseCUID = llvm::StringSwitch<UseCUIDKind>(UseCUIDStr) 2938 .Case("hash", CUID_Hash) 2939 .Case("random", CUID_Random) 2940 .Case("none", CUID_None) 2941 .Default(CUID_Invalid); 2942 if (UseCUID == CUID_Invalid) { 2943 C.getDriver().Diag(diag::err_drv_invalid_value) 2944 << A->getAsString(Args) << UseCUIDStr; 2945 C.setContainsError(); 2946 return true; 2947 } 2948 } 2949 2950 // --offload and --offload-arch options are mutually exclusive. 2951 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 2952 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 2953 options::OPT_no_offload_arch_EQ)) { 2954 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) << "--offload-arch" 2955 << "--offload"; 2956 } 2957 2958 // Collect all cuda_gpu_arch parameters, removing duplicates. 2959 std::set<StringRef> GpuArchs; 2960 bool Error = false; 2961 for (Arg *A : Args) { 2962 if (!(A->getOption().matches(options::OPT_offload_arch_EQ) || 2963 A->getOption().matches(options::OPT_no_offload_arch_EQ))) 2964 continue; 2965 A->claim(); 2966 2967 StringRef ArchStr = A->getValue(); 2968 if (A->getOption().matches(options::OPT_no_offload_arch_EQ) && 2969 ArchStr == "all") { 2970 GpuArchs.clear(); 2971 continue; 2972 } 2973 ArchStr = getCanonicalOffloadArch(ArchStr); 2974 if (ArchStr.empty()) { 2975 Error = true; 2976 } else if (A->getOption().matches(options::OPT_offload_arch_EQ)) 2977 GpuArchs.insert(ArchStr); 2978 else if (A->getOption().matches(options::OPT_no_offload_arch_EQ)) 2979 GpuArchs.erase(ArchStr); 2980 else 2981 llvm_unreachable("Unexpected option."); 2982 } 2983 2984 auto &&ConflictingArchs = getConflictOffloadArchCombination(GpuArchs); 2985 if (ConflictingArchs) { 2986 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 2987 << ConflictingArchs.getValue().first 2988 << ConflictingArchs.getValue().second; 2989 C.setContainsError(); 2990 return true; 2991 } 2992 2993 // Collect list of GPUs remaining in the set. 2994 for (auto Arch : GpuArchs) 2995 GpuArchList.push_back(Arch.data()); 2996 2997 // Default to sm_20 which is the lowest common denominator for 2998 // supported GPUs. sm_20 code should work correctly, if 2999 // suboptimally, on all newer GPUs. 3000 if (GpuArchList.empty()) { 3001 if (ToolChains.front()->getTriple().isSPIRV()) 3002 GpuArchList.push_back(CudaArch::Generic); 3003 else 3004 GpuArchList.push_back(DefaultCudaArch); 3005 } 3006 3007 return Error; 3008 } 3009 }; 3010 3011 /// \brief CUDA action builder. It injects device code in the host backend 3012 /// action. 3013 class CudaActionBuilder final : public CudaActionBuilderBase { 3014 public: 3015 CudaActionBuilder(Compilation &C, DerivedArgList &Args, 3016 const Driver::InputList &Inputs) 3017 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) { 3018 DefaultCudaArch = CudaArch::SM_35; 3019 } 3020 3021 StringRef getCanonicalOffloadArch(StringRef ArchStr) override { 3022 CudaArch Arch = StringToCudaArch(ArchStr); 3023 if (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch)) { 3024 C.getDriver().Diag(clang::diag::err_drv_cuda_bad_gpu_arch) << ArchStr; 3025 return StringRef(); 3026 } 3027 return CudaArchToString(Arch); 3028 } 3029 3030 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3031 getConflictOffloadArchCombination( 3032 const std::set<StringRef> &GpuArchs) override { 3033 return llvm::None; 3034 } 3035 3036 ActionBuilderReturnCode 3037 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3038 phases::ID CurPhase, phases::ID FinalPhase, 3039 PhasesTy &Phases) override { 3040 if (!IsActive) 3041 return ABRT_Inactive; 3042 3043 // If we don't have more CUDA actions, we don't have any dependences to 3044 // create for the host. 3045 if (CudaDeviceActions.empty()) 3046 return ABRT_Success; 3047 3048 assert(CudaDeviceActions.size() == GpuArchList.size() && 3049 "Expecting one action per GPU architecture."); 3050 assert(!CompileHostOnly && 3051 "Not expecting CUDA actions in host-only compilation."); 3052 3053 // If we are generating code for the device or we are in a backend phase, 3054 // we attempt to generate the fat binary. We compile each arch to ptx and 3055 // assemble to cubin, then feed the cubin *and* the ptx into a device 3056 // "link" action, which uses fatbinary to combine these cubins into one 3057 // fatbin. The fatbin is then an input to the host action if not in 3058 // device-only mode. 3059 if (CompileDeviceOnly || CurPhase == phases::Backend) { 3060 ActionList DeviceActions; 3061 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3062 // Produce the device action from the current phase up to the assemble 3063 // phase. 3064 for (auto Ph : Phases) { 3065 // Skip the phases that were already dealt with. 3066 if (Ph < CurPhase) 3067 continue; 3068 // We have to be consistent with the host final phase. 3069 if (Ph > FinalPhase) 3070 break; 3071 3072 CudaDeviceActions[I] = C.getDriver().ConstructPhaseAction( 3073 C, Args, Ph, CudaDeviceActions[I], Action::OFK_Cuda); 3074 3075 if (Ph == phases::Assemble) 3076 break; 3077 } 3078 3079 // If we didn't reach the assemble phase, we can't generate the fat 3080 // binary. We don't need to generate the fat binary if we are not in 3081 // device-only mode. 3082 if (!isa<AssembleJobAction>(CudaDeviceActions[I]) || 3083 CompileDeviceOnly) 3084 continue; 3085 3086 Action *AssembleAction = CudaDeviceActions[I]; 3087 assert(AssembleAction->getType() == types::TY_Object); 3088 assert(AssembleAction->getInputs().size() == 1); 3089 3090 Action *BackendAction = AssembleAction->getInputs()[0]; 3091 assert(BackendAction->getType() == types::TY_PP_Asm); 3092 3093 for (auto &A : {AssembleAction, BackendAction}) { 3094 OffloadAction::DeviceDependences DDep; 3095 DDep.add(*A, *ToolChains.front(), GpuArchList[I], Action::OFK_Cuda); 3096 DeviceActions.push_back( 3097 C.MakeAction<OffloadAction>(DDep, A->getType())); 3098 } 3099 } 3100 3101 // We generate the fat binary if we have device input actions. 3102 if (!DeviceActions.empty()) { 3103 CudaFatBinary = 3104 C.MakeAction<LinkJobAction>(DeviceActions, types::TY_CUDA_FATBIN); 3105 3106 if (!CompileDeviceOnly) { 3107 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3108 Action::OFK_Cuda); 3109 // Clear the fat binary, it is already a dependence to an host 3110 // action. 3111 CudaFatBinary = nullptr; 3112 } 3113 3114 // Remove the CUDA actions as they are already connected to an host 3115 // action or fat binary. 3116 CudaDeviceActions.clear(); 3117 } 3118 3119 // We avoid creating host action in device-only mode. 3120 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3121 } else if (CurPhase > phases::Backend) { 3122 // If we are past the backend phase and still have a device action, we 3123 // don't have to do anything as this action is already a device 3124 // top-level action. 3125 return ABRT_Success; 3126 } 3127 3128 assert(CurPhase < phases::Backend && "Generating single CUDA " 3129 "instructions should only occur " 3130 "before the backend phase!"); 3131 3132 // By default, we produce an action for each device arch. 3133 for (Action *&A : CudaDeviceActions) 3134 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3135 3136 return ABRT_Success; 3137 } 3138 }; 3139 /// \brief HIP action builder. It injects device code in the host backend 3140 /// action. 3141 class HIPActionBuilder final : public CudaActionBuilderBase { 3142 /// The linker inputs obtained for each device arch. 3143 SmallVector<ActionList, 8> DeviceLinkerInputs; 3144 // The default bundling behavior depends on the type of output, therefore 3145 // BundleOutput needs to be tri-value: None, true, or false. 3146 // Bundle code objects except --no-gpu-output is specified for device 3147 // only compilation. Bundle other type of output files only if 3148 // --gpu-bundle-output is specified for device only compilation. 3149 Optional<bool> BundleOutput; 3150 3151 public: 3152 HIPActionBuilder(Compilation &C, DerivedArgList &Args, 3153 const Driver::InputList &Inputs) 3154 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) { 3155 DefaultCudaArch = CudaArch::GFX803; 3156 if (Args.hasArg(options::OPT_gpu_bundle_output, 3157 options::OPT_no_gpu_bundle_output)) 3158 BundleOutput = Args.hasFlag(options::OPT_gpu_bundle_output, 3159 options::OPT_no_gpu_bundle_output, true); 3160 } 3161 3162 bool canUseBundlerUnbundler() const override { return true; } 3163 3164 StringRef getCanonicalOffloadArch(StringRef IdStr) override { 3165 llvm::StringMap<bool> Features; 3166 // getHIPOffloadTargetTriple() is known to return valid value as it has 3167 // been called successfully in the CreateOffloadingDeviceToolChains(). 3168 auto ArchStr = parseTargetID( 3169 *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()), IdStr, 3170 &Features); 3171 if (!ArchStr) { 3172 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << IdStr; 3173 C.setContainsError(); 3174 return StringRef(); 3175 } 3176 auto CanId = getCanonicalTargetID(ArchStr.getValue(), Features); 3177 return Args.MakeArgStringRef(CanId); 3178 }; 3179 3180 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3181 getConflictOffloadArchCombination( 3182 const std::set<StringRef> &GpuArchs) override { 3183 return getConflictTargetIDCombination(GpuArchs); 3184 } 3185 3186 ActionBuilderReturnCode 3187 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3188 phases::ID CurPhase, phases::ID FinalPhase, 3189 PhasesTy &Phases) override { 3190 if (!IsActive) 3191 return ABRT_Inactive; 3192 3193 // amdgcn does not support linking of object files, therefore we skip 3194 // backend and assemble phases to output LLVM IR. Except for generating 3195 // non-relocatable device code, where we generate fat binary for device 3196 // code and pass to host in Backend phase. 3197 if (CudaDeviceActions.empty()) 3198 return ABRT_Success; 3199 3200 assert(((CurPhase == phases::Link && Relocatable) || 3201 CudaDeviceActions.size() == GpuArchList.size()) && 3202 "Expecting one action per GPU architecture."); 3203 assert(!CompileHostOnly && 3204 "Not expecting HIP actions in host-only compilation."); 3205 3206 if (!Relocatable && CurPhase == phases::Backend && !EmitLLVM && 3207 !EmitAsm) { 3208 // If we are in backend phase, we attempt to generate the fat binary. 3209 // We compile each arch to IR and use a link action to generate code 3210 // object containing ISA. Then we use a special "link" action to create 3211 // a fat binary containing all the code objects for different GPU's. 3212 // The fat binary is then an input to the host action. 3213 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3214 if (C.getDriver().isUsingLTO(/*IsOffload=*/true)) { 3215 // When LTO is enabled, skip the backend and assemble phases and 3216 // use lld to link the bitcode. 3217 ActionList AL; 3218 AL.push_back(CudaDeviceActions[I]); 3219 // Create a link action to link device IR with device library 3220 // and generate ISA. 3221 CudaDeviceActions[I] = 3222 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3223 } else { 3224 // When LTO is not enabled, we follow the conventional 3225 // compiler phases, including backend and assemble phases. 3226 ActionList AL; 3227 Action *BackendAction = nullptr; 3228 if (ToolChains.front()->getTriple().isSPIRV()) { 3229 // Emit LLVM bitcode for SPIR-V targets. SPIR-V device tool chain 3230 // (HIPSPVToolChain) runs post-link LLVM IR passes. 3231 types::ID Output = Args.hasArg(options::OPT_S) 3232 ? types::TY_LLVM_IR 3233 : types::TY_LLVM_BC; 3234 BackendAction = 3235 C.MakeAction<BackendJobAction>(CudaDeviceActions[I], Output); 3236 } else 3237 BackendAction = C.getDriver().ConstructPhaseAction( 3238 C, Args, phases::Backend, CudaDeviceActions[I], 3239 AssociatedOffloadKind); 3240 auto AssembleAction = C.getDriver().ConstructPhaseAction( 3241 C, Args, phases::Assemble, BackendAction, 3242 AssociatedOffloadKind); 3243 AL.push_back(AssembleAction); 3244 // Create a link action to link device IR with device library 3245 // and generate ISA. 3246 CudaDeviceActions[I] = 3247 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3248 } 3249 3250 // OffloadingActionBuilder propagates device arch until an offload 3251 // action. Since the next action for creating fatbin does 3252 // not have device arch, whereas the above link action and its input 3253 // have device arch, an offload action is needed to stop the null 3254 // device arch of the next action being propagated to the above link 3255 // action. 3256 OffloadAction::DeviceDependences DDep; 3257 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3258 AssociatedOffloadKind); 3259 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3260 DDep, CudaDeviceActions[I]->getType()); 3261 } 3262 3263 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3264 BundleOutput.getValue()) { 3265 // Create HIP fat binary with a special "link" action. 3266 CudaFatBinary = C.MakeAction<LinkJobAction>(CudaDeviceActions, 3267 types::TY_HIP_FATBIN); 3268 3269 if (!CompileDeviceOnly) { 3270 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3271 AssociatedOffloadKind); 3272 // Clear the fat binary, it is already a dependence to an host 3273 // action. 3274 CudaFatBinary = nullptr; 3275 } 3276 3277 // Remove the CUDA actions as they are already connected to an host 3278 // action or fat binary. 3279 CudaDeviceActions.clear(); 3280 } 3281 3282 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3283 } else if (CurPhase == phases::Link) { 3284 // Save CudaDeviceActions to DeviceLinkerInputs for each GPU subarch. 3285 // This happens to each device action originated from each input file. 3286 // Later on, device actions in DeviceLinkerInputs are used to create 3287 // device link actions in appendLinkDependences and the created device 3288 // link actions are passed to the offload action as device dependence. 3289 DeviceLinkerInputs.resize(CudaDeviceActions.size()); 3290 auto LI = DeviceLinkerInputs.begin(); 3291 for (auto *A : CudaDeviceActions) { 3292 LI->push_back(A); 3293 ++LI; 3294 } 3295 3296 // We will pass the device action as a host dependence, so we don't 3297 // need to do anything else with them. 3298 CudaDeviceActions.clear(); 3299 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3300 } 3301 3302 // By default, we produce an action for each device arch. 3303 for (Action *&A : CudaDeviceActions) 3304 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A, 3305 AssociatedOffloadKind); 3306 3307 if (CompileDeviceOnly && CurPhase == FinalPhase && 3308 BundleOutput.hasValue() && BundleOutput.getValue()) { 3309 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3310 OffloadAction::DeviceDependences DDep; 3311 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3312 AssociatedOffloadKind); 3313 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3314 DDep, CudaDeviceActions[I]->getType()); 3315 } 3316 CudaFatBinary = 3317 C.MakeAction<OffloadBundlingJobAction>(CudaDeviceActions); 3318 CudaDeviceActions.clear(); 3319 } 3320 3321 return (CompileDeviceOnly && CurPhase == FinalPhase) ? ABRT_Ignore_Host 3322 : ABRT_Success; 3323 } 3324 3325 void appendLinkDeviceActions(ActionList &AL) override { 3326 if (DeviceLinkerInputs.size() == 0) 3327 return; 3328 3329 assert(DeviceLinkerInputs.size() == GpuArchList.size() && 3330 "Linker inputs and GPU arch list sizes do not match."); 3331 3332 ActionList Actions; 3333 unsigned I = 0; 3334 // Append a new link action for each device. 3335 // Each entry in DeviceLinkerInputs corresponds to a GPU arch. 3336 for (auto &LI : DeviceLinkerInputs) { 3337 3338 types::ID Output = Args.hasArg(options::OPT_emit_llvm) 3339 ? types::TY_LLVM_BC 3340 : types::TY_Image; 3341 3342 auto *DeviceLinkAction = C.MakeAction<LinkJobAction>(LI, Output); 3343 // Linking all inputs for the current GPU arch. 3344 // LI contains all the inputs for the linker. 3345 OffloadAction::DeviceDependences DeviceLinkDeps; 3346 DeviceLinkDeps.add(*DeviceLinkAction, *ToolChains[0], 3347 GpuArchList[I], AssociatedOffloadKind); 3348 Actions.push_back(C.MakeAction<OffloadAction>( 3349 DeviceLinkDeps, DeviceLinkAction->getType())); 3350 ++I; 3351 } 3352 DeviceLinkerInputs.clear(); 3353 3354 // If emitting LLVM, do not generate final host/device compilation action 3355 if (Args.hasArg(options::OPT_emit_llvm)) { 3356 AL.append(Actions); 3357 return; 3358 } 3359 3360 // Create a host object from all the device images by embedding them 3361 // in a fat binary for mixed host-device compilation. For device-only 3362 // compilation, creates a fat binary. 3363 OffloadAction::DeviceDependences DDeps; 3364 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3365 BundleOutput.getValue()) { 3366 auto *TopDeviceLinkAction = C.MakeAction<LinkJobAction>( 3367 Actions, 3368 CompileDeviceOnly ? types::TY_HIP_FATBIN : types::TY_Object); 3369 DDeps.add(*TopDeviceLinkAction, *ToolChains[0], nullptr, 3370 AssociatedOffloadKind); 3371 // Offload the host object to the host linker. 3372 AL.push_back( 3373 C.MakeAction<OffloadAction>(DDeps, TopDeviceLinkAction->getType())); 3374 } else { 3375 AL.append(Actions); 3376 } 3377 } 3378 3379 Action* appendLinkHostActions(ActionList &AL) override { return AL.back(); } 3380 3381 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3382 }; 3383 3384 /// OpenMP action builder. The host bitcode is passed to the device frontend 3385 /// and all the device linked images are passed to the host link phase. 3386 class OpenMPActionBuilder final : public DeviceActionBuilder { 3387 /// The OpenMP actions for the current input. 3388 ActionList OpenMPDeviceActions; 3389 3390 /// The linker inputs obtained for each toolchain. 3391 SmallVector<ActionList, 8> DeviceLinkerInputs; 3392 3393 public: 3394 OpenMPActionBuilder(Compilation &C, DerivedArgList &Args, 3395 const Driver::InputList &Inputs) 3396 : DeviceActionBuilder(C, Args, Inputs, Action::OFK_OpenMP) {} 3397 3398 ActionBuilderReturnCode 3399 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3400 phases::ID CurPhase, phases::ID FinalPhase, 3401 PhasesTy &Phases) override { 3402 if (OpenMPDeviceActions.empty()) 3403 return ABRT_Inactive; 3404 3405 // We should always have an action for each input. 3406 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3407 "Number of OpenMP actions and toolchains do not match."); 3408 3409 // The host only depends on device action in the linking phase, when all 3410 // the device images have to be embedded in the host image. 3411 if (CurPhase == phases::Link) { 3412 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3413 "Toolchains and linker inputs sizes do not match."); 3414 auto LI = DeviceLinkerInputs.begin(); 3415 for (auto *A : OpenMPDeviceActions) { 3416 LI->push_back(A); 3417 ++LI; 3418 } 3419 3420 // We passed the device action as a host dependence, so we don't need to 3421 // do anything else with them. 3422 OpenMPDeviceActions.clear(); 3423 return ABRT_Success; 3424 } 3425 3426 // By default, we produce an action for each device arch. 3427 for (Action *&A : OpenMPDeviceActions) 3428 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3429 3430 return ABRT_Success; 3431 } 3432 3433 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 3434 3435 // If this is an input action replicate it for each OpenMP toolchain. 3436 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 3437 OpenMPDeviceActions.clear(); 3438 for (unsigned I = 0; I < ToolChains.size(); ++I) 3439 OpenMPDeviceActions.push_back( 3440 C.MakeAction<InputAction>(IA->getInputArg(), IA->getType())); 3441 return ABRT_Success; 3442 } 3443 3444 // If this is an unbundling action use it as is for each OpenMP toolchain. 3445 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 3446 OpenMPDeviceActions.clear(); 3447 auto *IA = cast<InputAction>(UA->getInputs().back()); 3448 std::string FileName = IA->getInputArg().getAsString(Args); 3449 // Check if the type of the file is the same as the action. Do not 3450 // unbundle it if it is not. Do not unbundle .so files, for example, 3451 // which are not object files. 3452 if (IA->getType() == types::TY_Object && 3453 (!llvm::sys::path::has_extension(FileName) || 3454 types::lookupTypeForExtension( 3455 llvm::sys::path::extension(FileName).drop_front()) != 3456 types::TY_Object)) 3457 return ABRT_Inactive; 3458 for (unsigned I = 0; I < ToolChains.size(); ++I) { 3459 OpenMPDeviceActions.push_back(UA); 3460 UA->registerDependentActionInfo( 3461 ToolChains[I], /*BoundArch=*/StringRef(), Action::OFK_OpenMP); 3462 } 3463 return ABRT_Success; 3464 } 3465 3466 // When generating code for OpenMP we use the host compile phase result as 3467 // a dependence to the device compile phase so that it can learn what 3468 // declarations should be emitted. However, this is not the only use for 3469 // the host action, so we prevent it from being collapsed. 3470 if (isa<CompileJobAction>(HostAction)) { 3471 HostAction->setCannotBeCollapsedWithNextDependentAction(); 3472 assert(ToolChains.size() == OpenMPDeviceActions.size() && 3473 "Toolchains and device action sizes do not match."); 3474 OffloadAction::HostDependence HDep( 3475 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3476 /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3477 auto TC = ToolChains.begin(); 3478 for (Action *&A : OpenMPDeviceActions) { 3479 assert(isa<CompileJobAction>(A)); 3480 OffloadAction::DeviceDependences DDep; 3481 DDep.add(*A, **TC, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3482 A = C.MakeAction<OffloadAction>(HDep, DDep); 3483 ++TC; 3484 } 3485 } 3486 return ABRT_Success; 3487 } 3488 3489 void appendTopLevelActions(ActionList &AL) override { 3490 if (OpenMPDeviceActions.empty()) 3491 return; 3492 3493 // We should always have an action for each input. 3494 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3495 "Number of OpenMP actions and toolchains do not match."); 3496 3497 // Append all device actions followed by the proper offload action. 3498 auto TI = ToolChains.begin(); 3499 for (auto *A : OpenMPDeviceActions) { 3500 OffloadAction::DeviceDependences Dep; 3501 Dep.add(*A, **TI, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3502 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 3503 ++TI; 3504 } 3505 // We no longer need the action stored in this builder. 3506 OpenMPDeviceActions.clear(); 3507 } 3508 3509 void appendLinkDeviceActions(ActionList &AL) override { 3510 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3511 "Toolchains and linker inputs sizes do not match."); 3512 3513 // Append a new link action for each device. 3514 auto TC = ToolChains.begin(); 3515 for (auto &LI : DeviceLinkerInputs) { 3516 auto *DeviceLinkAction = 3517 C.MakeAction<LinkJobAction>(LI, types::TY_Image); 3518 OffloadAction::DeviceDependences DeviceLinkDeps; 3519 DeviceLinkDeps.add(*DeviceLinkAction, **TC, /*BoundArch=*/nullptr, 3520 Action::OFK_OpenMP); 3521 AL.push_back(C.MakeAction<OffloadAction>(DeviceLinkDeps, 3522 DeviceLinkAction->getType())); 3523 ++TC; 3524 } 3525 DeviceLinkerInputs.clear(); 3526 } 3527 3528 Action* appendLinkHostActions(ActionList &AL) override { 3529 // Create wrapper bitcode from the result of device link actions and compile 3530 // it to an object which will be added to the host link command. 3531 auto *BC = C.MakeAction<OffloadWrapperJobAction>(AL, types::TY_LLVM_BC); 3532 auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm); 3533 return C.MakeAction<AssembleJobAction>(ASM, types::TY_Object); 3534 } 3535 3536 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3537 3538 bool initialize() override { 3539 // Get the OpenMP toolchains. If we don't get any, the action builder will 3540 // know there is nothing to do related to OpenMP offloading. 3541 auto OpenMPTCRange = C.getOffloadToolChains<Action::OFK_OpenMP>(); 3542 for (auto TI = OpenMPTCRange.first, TE = OpenMPTCRange.second; TI != TE; 3543 ++TI) 3544 ToolChains.push_back(TI->second); 3545 3546 DeviceLinkerInputs.resize(ToolChains.size()); 3547 return false; 3548 } 3549 3550 bool canUseBundlerUnbundler() const override { 3551 // OpenMP should use bundled files whenever possible. 3552 return true; 3553 } 3554 }; 3555 3556 /// 3557 /// TODO: Add the implementation for other specialized builders here. 3558 /// 3559 3560 /// Specialized builders being used by this offloading action builder. 3561 SmallVector<DeviceActionBuilder *, 4> SpecializedBuilders; 3562 3563 /// Flag set to true if all valid builders allow file bundling/unbundling. 3564 bool CanUseBundler; 3565 3566 public: 3567 OffloadingActionBuilder(Compilation &C, DerivedArgList &Args, 3568 const Driver::InputList &Inputs) 3569 : C(C) { 3570 // Create a specialized builder for each device toolchain. 3571 3572 IsValid = true; 3573 3574 // Create a specialized builder for CUDA. 3575 SpecializedBuilders.push_back(new CudaActionBuilder(C, Args, Inputs)); 3576 3577 // Create a specialized builder for HIP. 3578 SpecializedBuilders.push_back(new HIPActionBuilder(C, Args, Inputs)); 3579 3580 // Create a specialized builder for OpenMP. 3581 SpecializedBuilders.push_back(new OpenMPActionBuilder(C, Args, Inputs)); 3582 3583 // 3584 // TODO: Build other specialized builders here. 3585 // 3586 3587 // Initialize all the builders, keeping track of errors. If all valid 3588 // builders agree that we can use bundling, set the flag to true. 3589 unsigned ValidBuilders = 0u; 3590 unsigned ValidBuildersSupportingBundling = 0u; 3591 for (auto *SB : SpecializedBuilders) { 3592 IsValid = IsValid && !SB->initialize(); 3593 3594 // Update the counters if the builder is valid. 3595 if (SB->isValid()) { 3596 ++ValidBuilders; 3597 if (SB->canUseBundlerUnbundler()) 3598 ++ValidBuildersSupportingBundling; 3599 } 3600 } 3601 CanUseBundler = 3602 ValidBuilders && ValidBuilders == ValidBuildersSupportingBundling; 3603 } 3604 3605 ~OffloadingActionBuilder() { 3606 for (auto *SB : SpecializedBuilders) 3607 delete SB; 3608 } 3609 3610 /// Record a host action and its originating input argument. 3611 void recordHostAction(Action *HostAction, const Arg *InputArg) { 3612 assert(HostAction && "Invalid host action"); 3613 assert(InputArg && "Invalid input argument"); 3614 auto Loc = HostActionToInputArgMap.find(HostAction); 3615 if (Loc == HostActionToInputArgMap.end()) 3616 HostActionToInputArgMap[HostAction] = InputArg; 3617 assert(HostActionToInputArgMap[HostAction] == InputArg && 3618 "host action mapped to multiple input arguments"); 3619 } 3620 3621 /// Generate an action that adds device dependences (if any) to a host action. 3622 /// If no device dependence actions exist, just return the host action \a 3623 /// HostAction. If an error is found or if no builder requires the host action 3624 /// to be generated, return nullptr. 3625 Action * 3626 addDeviceDependencesToHostAction(Action *HostAction, const Arg *InputArg, 3627 phases::ID CurPhase, phases::ID FinalPhase, 3628 DeviceActionBuilder::PhasesTy &Phases) { 3629 if (!IsValid) 3630 return nullptr; 3631 3632 if (SpecializedBuilders.empty()) 3633 return HostAction; 3634 3635 assert(HostAction && "Invalid host action!"); 3636 recordHostAction(HostAction, InputArg); 3637 3638 OffloadAction::DeviceDependences DDeps; 3639 // Check if all the programming models agree we should not emit the host 3640 // action. Also, keep track of the offloading kinds employed. 3641 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3642 unsigned InactiveBuilders = 0u; 3643 unsigned IgnoringBuilders = 0u; 3644 for (auto *SB : SpecializedBuilders) { 3645 if (!SB->isValid()) { 3646 ++InactiveBuilders; 3647 continue; 3648 } 3649 3650 auto RetCode = 3651 SB->getDeviceDependences(DDeps, CurPhase, FinalPhase, Phases); 3652 3653 // If the builder explicitly says the host action should be ignored, 3654 // we need to increment the variable that tracks the builders that request 3655 // the host object to be ignored. 3656 if (RetCode == DeviceActionBuilder::ABRT_Ignore_Host) 3657 ++IgnoringBuilders; 3658 3659 // Unless the builder was inactive for this action, we have to record the 3660 // offload kind because the host will have to use it. 3661 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3662 OffloadKind |= SB->getAssociatedOffloadKind(); 3663 } 3664 3665 // If all builders agree that the host object should be ignored, just return 3666 // nullptr. 3667 if (IgnoringBuilders && 3668 SpecializedBuilders.size() == (InactiveBuilders + IgnoringBuilders)) 3669 return nullptr; 3670 3671 if (DDeps.getActions().empty()) 3672 return HostAction; 3673 3674 // We have dependences we need to bundle together. We use an offload action 3675 // for that. 3676 OffloadAction::HostDependence HDep( 3677 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3678 /*BoundArch=*/nullptr, DDeps); 3679 return C.MakeAction<OffloadAction>(HDep, DDeps); 3680 } 3681 3682 /// Generate an action that adds a host dependence to a device action. The 3683 /// results will be kept in this action builder. Return true if an error was 3684 /// found. 3685 bool addHostDependenceToDeviceActions(Action *&HostAction, 3686 const Arg *InputArg) { 3687 if (!IsValid) 3688 return true; 3689 3690 recordHostAction(HostAction, InputArg); 3691 3692 // If we are supporting bundling/unbundling and the current action is an 3693 // input action of non-source file, we replace the host action by the 3694 // unbundling action. The bundler tool has the logic to detect if an input 3695 // is a bundle or not and if the input is not a bundle it assumes it is a 3696 // host file. Therefore it is safe to create an unbundling action even if 3697 // the input is not a bundle. 3698 if (CanUseBundler && isa<InputAction>(HostAction) && 3699 InputArg->getOption().getKind() == llvm::opt::Option::InputClass && 3700 (!types::isSrcFile(HostAction->getType()) || 3701 HostAction->getType() == types::TY_PP_HIP)) { 3702 auto UnbundlingHostAction = 3703 C.MakeAction<OffloadUnbundlingJobAction>(HostAction); 3704 UnbundlingHostAction->registerDependentActionInfo( 3705 C.getSingleOffloadToolChain<Action::OFK_Host>(), 3706 /*BoundArch=*/StringRef(), Action::OFK_Host); 3707 HostAction = UnbundlingHostAction; 3708 recordHostAction(HostAction, InputArg); 3709 } 3710 3711 assert(HostAction && "Invalid host action!"); 3712 3713 // Register the offload kinds that are used. 3714 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3715 for (auto *SB : SpecializedBuilders) { 3716 if (!SB->isValid()) 3717 continue; 3718 3719 auto RetCode = SB->addDeviceDepences(HostAction); 3720 3721 // Host dependences for device actions are not compatible with that same 3722 // action being ignored. 3723 assert(RetCode != DeviceActionBuilder::ABRT_Ignore_Host && 3724 "Host dependence not expected to be ignored.!"); 3725 3726 // Unless the builder was inactive for this action, we have to record the 3727 // offload kind because the host will have to use it. 3728 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3729 OffloadKind |= SB->getAssociatedOffloadKind(); 3730 } 3731 3732 // Do not use unbundler if the Host does not depend on device action. 3733 if (OffloadKind == Action::OFK_None && CanUseBundler) 3734 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) 3735 HostAction = UA->getInputs().back(); 3736 3737 return false; 3738 } 3739 3740 /// Add the offloading top level actions to the provided action list. This 3741 /// function can replace the host action by a bundling action if the 3742 /// programming models allow it. 3743 bool appendTopLevelActions(ActionList &AL, Action *HostAction, 3744 const Arg *InputArg) { 3745 if (HostAction) 3746 recordHostAction(HostAction, InputArg); 3747 3748 // Get the device actions to be appended. 3749 ActionList OffloadAL; 3750 for (auto *SB : SpecializedBuilders) { 3751 if (!SB->isValid()) 3752 continue; 3753 SB->appendTopLevelActions(OffloadAL); 3754 } 3755 3756 // If we can use the bundler, replace the host action by the bundling one in 3757 // the resulting list. Otherwise, just append the device actions. For 3758 // device only compilation, HostAction is a null pointer, therefore only do 3759 // this when HostAction is not a null pointer. 3760 if (CanUseBundler && HostAction && 3761 HostAction->getType() != types::TY_Nothing && !OffloadAL.empty()) { 3762 // Add the host action to the list in order to create the bundling action. 3763 OffloadAL.push_back(HostAction); 3764 3765 // We expect that the host action was just appended to the action list 3766 // before this method was called. 3767 assert(HostAction == AL.back() && "Host action not in the list??"); 3768 HostAction = C.MakeAction<OffloadBundlingJobAction>(OffloadAL); 3769 recordHostAction(HostAction, InputArg); 3770 AL.back() = HostAction; 3771 } else 3772 AL.append(OffloadAL.begin(), OffloadAL.end()); 3773 3774 // Propagate to the current host action (if any) the offload information 3775 // associated with the current input. 3776 if (HostAction) 3777 HostAction->propagateHostOffloadInfo(InputArgToOffloadKindMap[InputArg], 3778 /*BoundArch=*/nullptr); 3779 return false; 3780 } 3781 3782 void appendDeviceLinkActions(ActionList &AL) { 3783 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3784 if (!SB->isValid()) 3785 continue; 3786 SB->appendLinkDeviceActions(AL); 3787 } 3788 } 3789 3790 Action *makeHostLinkAction() { 3791 // Build a list of device linking actions. 3792 ActionList DeviceAL; 3793 appendDeviceLinkActions(DeviceAL); 3794 if (DeviceAL.empty()) 3795 return nullptr; 3796 3797 // Let builders add host linking actions. 3798 Action* HA = nullptr; 3799 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3800 if (!SB->isValid()) 3801 continue; 3802 HA = SB->appendLinkHostActions(DeviceAL); 3803 // This created host action has no originating input argument, therefore 3804 // needs to set its offloading kind directly. 3805 if (HA) 3806 HA->propagateHostOffloadInfo(SB->getAssociatedOffloadKind(), 3807 /*BoundArch=*/nullptr); 3808 } 3809 return HA; 3810 } 3811 3812 /// Processes the host linker action. This currently consists of replacing it 3813 /// with an offload action if there are device link objects and propagate to 3814 /// the host action all the offload kinds used in the current compilation. The 3815 /// resulting action is returned. 3816 Action *processHostLinkAction(Action *HostAction) { 3817 // Add all the dependences from the device linking actions. 3818 OffloadAction::DeviceDependences DDeps; 3819 for (auto *SB : SpecializedBuilders) { 3820 if (!SB->isValid()) 3821 continue; 3822 3823 SB->appendLinkDependences(DDeps); 3824 } 3825 3826 // Calculate all the offload kinds used in the current compilation. 3827 unsigned ActiveOffloadKinds = 0u; 3828 for (auto &I : InputArgToOffloadKindMap) 3829 ActiveOffloadKinds |= I.second; 3830 3831 // If we don't have device dependencies, we don't have to create an offload 3832 // action. 3833 if (DDeps.getActions().empty()) { 3834 // Set all the active offloading kinds to the link action. Given that it 3835 // is a link action it is assumed to depend on all actions generated so 3836 // far. 3837 HostAction->setHostOffloadInfo(ActiveOffloadKinds, 3838 /*BoundArch=*/nullptr); 3839 // Propagate active offloading kinds for each input to the link action. 3840 // Each input may have different active offloading kind. 3841 for (auto A : HostAction->inputs()) { 3842 auto ArgLoc = HostActionToInputArgMap.find(A); 3843 if (ArgLoc == HostActionToInputArgMap.end()) 3844 continue; 3845 auto OFKLoc = InputArgToOffloadKindMap.find(ArgLoc->second); 3846 if (OFKLoc == InputArgToOffloadKindMap.end()) 3847 continue; 3848 A->propagateHostOffloadInfo(OFKLoc->second, /*BoundArch=*/nullptr); 3849 } 3850 return HostAction; 3851 } 3852 3853 // Create the offload action with all dependences. When an offload action 3854 // is created the kinds are propagated to the host action, so we don't have 3855 // to do that explicitly here. 3856 OffloadAction::HostDependence HDep( 3857 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3858 /*BoundArch*/ nullptr, ActiveOffloadKinds); 3859 return C.MakeAction<OffloadAction>(HDep, DDeps); 3860 } 3861 }; 3862 } // anonymous namespace. 3863 3864 void Driver::handleArguments(Compilation &C, DerivedArgList &Args, 3865 const InputList &Inputs, 3866 ActionList &Actions) const { 3867 3868 // Ignore /Yc/Yu if both /Yc and /Yu passed but with different filenames. 3869 Arg *YcArg = Args.getLastArg(options::OPT__SLASH_Yc); 3870 Arg *YuArg = Args.getLastArg(options::OPT__SLASH_Yu); 3871 if (YcArg && YuArg && strcmp(YcArg->getValue(), YuArg->getValue()) != 0) { 3872 Diag(clang::diag::warn_drv_ycyu_different_arg_clang_cl); 3873 Args.eraseArg(options::OPT__SLASH_Yc); 3874 Args.eraseArg(options::OPT__SLASH_Yu); 3875 YcArg = YuArg = nullptr; 3876 } 3877 if (YcArg && Inputs.size() > 1) { 3878 Diag(clang::diag::warn_drv_yc_multiple_inputs_clang_cl); 3879 Args.eraseArg(options::OPT__SLASH_Yc); 3880 YcArg = nullptr; 3881 } 3882 3883 Arg *FinalPhaseArg; 3884 phases::ID FinalPhase = getFinalPhase(Args, &FinalPhaseArg); 3885 3886 if (FinalPhase == phases::Link) { 3887 // Emitting LLVM while linking disabled except in HIPAMD Toolchain 3888 if (Args.hasArg(options::OPT_emit_llvm) && !Args.hasArg(options::OPT_hip_link)) 3889 Diag(clang::diag::err_drv_emit_llvm_link); 3890 if (IsCLMode() && LTOMode != LTOK_None && 3891 !Args.getLastArgValue(options::OPT_fuse_ld_EQ) 3892 .equals_insensitive("lld")) 3893 Diag(clang::diag::err_drv_lto_without_lld); 3894 } 3895 3896 if (FinalPhase == phases::Preprocess || Args.hasArg(options::OPT__SLASH_Y_)) { 3897 // If only preprocessing or /Y- is used, all pch handling is disabled. 3898 // Rather than check for it everywhere, just remove clang-cl pch-related 3899 // flags here. 3900 Args.eraseArg(options::OPT__SLASH_Fp); 3901 Args.eraseArg(options::OPT__SLASH_Yc); 3902 Args.eraseArg(options::OPT__SLASH_Yu); 3903 YcArg = YuArg = nullptr; 3904 } 3905 3906 unsigned LastPLSize = 0; 3907 for (auto &I : Inputs) { 3908 types::ID InputType = I.first; 3909 const Arg *InputArg = I.second; 3910 3911 auto PL = types::getCompilationPhases(InputType); 3912 LastPLSize = PL.size(); 3913 3914 // If the first step comes after the final phase we are doing as part of 3915 // this compilation, warn the user about it. 3916 phases::ID InitialPhase = PL[0]; 3917 if (InitialPhase > FinalPhase) { 3918 if (InputArg->isClaimed()) 3919 continue; 3920 3921 // Claim here to avoid the more general unused warning. 3922 InputArg->claim(); 3923 3924 // Suppress all unused style warnings with -Qunused-arguments 3925 if (Args.hasArg(options::OPT_Qunused_arguments)) 3926 continue; 3927 3928 // Special case when final phase determined by binary name, rather than 3929 // by a command-line argument with a corresponding Arg. 3930 if (CCCIsCPP()) 3931 Diag(clang::diag::warn_drv_input_file_unused_by_cpp) 3932 << InputArg->getAsString(Args) << getPhaseName(InitialPhase); 3933 // Special case '-E' warning on a previously preprocessed file to make 3934 // more sense. 3935 else if (InitialPhase == phases::Compile && 3936 (Args.getLastArg(options::OPT__SLASH_EP, 3937 options::OPT__SLASH_P) || 3938 Args.getLastArg(options::OPT_E) || 3939 Args.getLastArg(options::OPT_M, options::OPT_MM)) && 3940 getPreprocessedType(InputType) == types::TY_INVALID) 3941 Diag(clang::diag::warn_drv_preprocessed_input_file_unused) 3942 << InputArg->getAsString(Args) << !!FinalPhaseArg 3943 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3944 else 3945 Diag(clang::diag::warn_drv_input_file_unused) 3946 << InputArg->getAsString(Args) << getPhaseName(InitialPhase) 3947 << !!FinalPhaseArg 3948 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3949 continue; 3950 } 3951 3952 if (YcArg) { 3953 // Add a separate precompile phase for the compile phase. 3954 if (FinalPhase >= phases::Compile) { 3955 const types::ID HeaderType = lookupHeaderTypeForSourceType(InputType); 3956 // Build the pipeline for the pch file. 3957 Action *ClangClPch = C.MakeAction<InputAction>(*InputArg, HeaderType); 3958 for (phases::ID Phase : types::getCompilationPhases(HeaderType)) 3959 ClangClPch = ConstructPhaseAction(C, Args, Phase, ClangClPch); 3960 assert(ClangClPch); 3961 Actions.push_back(ClangClPch); 3962 // The driver currently exits after the first failed command. This 3963 // relies on that behavior, to make sure if the pch generation fails, 3964 // the main compilation won't run. 3965 // FIXME: If the main compilation fails, the PCH generation should 3966 // probably not be considered successful either. 3967 } 3968 } 3969 } 3970 3971 // If we are linking, claim any options which are obviously only used for 3972 // compilation. 3973 // FIXME: Understand why the last Phase List length is used here. 3974 if (FinalPhase == phases::Link && LastPLSize == 1) { 3975 Args.ClaimAllArgs(options::OPT_CompileOnly_Group); 3976 Args.ClaimAllArgs(options::OPT_cl_compile_Group); 3977 } 3978 } 3979 3980 void Driver::BuildActions(Compilation &C, DerivedArgList &Args, 3981 const InputList &Inputs, ActionList &Actions) const { 3982 llvm::PrettyStackTraceString CrashInfo("Building compilation actions"); 3983 3984 if (!SuppressMissingInputWarning && Inputs.empty()) { 3985 Diag(clang::diag::err_drv_no_input_files); 3986 return; 3987 } 3988 3989 // Reject -Z* at the top level, these options should never have been exposed 3990 // by gcc. 3991 if (Arg *A = Args.getLastArg(options::OPT_Z_Joined)) 3992 Diag(clang::diag::err_drv_use_of_Z_option) << A->getAsString(Args); 3993 3994 // Diagnose misuse of /Fo. 3995 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fo)) { 3996 StringRef V = A->getValue(); 3997 if (Inputs.size() > 1 && !V.empty() && 3998 !llvm::sys::path::is_separator(V.back())) { 3999 // Check whether /Fo tries to name an output file for multiple inputs. 4000 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4001 << A->getSpelling() << V; 4002 Args.eraseArg(options::OPT__SLASH_Fo); 4003 } 4004 } 4005 4006 // Diagnose misuse of /Fa. 4007 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fa)) { 4008 StringRef V = A->getValue(); 4009 if (Inputs.size() > 1 && !V.empty() && 4010 !llvm::sys::path::is_separator(V.back())) { 4011 // Check whether /Fa tries to name an asm file for multiple inputs. 4012 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4013 << A->getSpelling() << V; 4014 Args.eraseArg(options::OPT__SLASH_Fa); 4015 } 4016 } 4017 4018 // Diagnose misuse of /o. 4019 if (Arg *A = Args.getLastArg(options::OPT__SLASH_o)) { 4020 if (A->getValue()[0] == '\0') { 4021 // It has to have a value. 4022 Diag(clang::diag::err_drv_missing_argument) << A->getSpelling() << 1; 4023 Args.eraseArg(options::OPT__SLASH_o); 4024 } 4025 } 4026 4027 handleArguments(C, Args, Inputs, Actions); 4028 4029 // Builder to be used to build offloading actions. 4030 OffloadingActionBuilder OffloadBuilder(C, Args, Inputs); 4031 4032 bool UseNewOffloadingDriver = 4033 (C.isOffloadingHostKind(Action::OFK_OpenMP) && 4034 Args.hasFlag(options::OPT_fopenmp_new_driver, 4035 options::OPT_no_offload_new_driver, true)) || 4036 Args.hasFlag(options::OPT_offload_new_driver, 4037 options::OPT_no_offload_new_driver, false); 4038 4039 // Construct the actions to perform. 4040 HeaderModulePrecompileJobAction *HeaderModuleAction = nullptr; 4041 ExtractAPIJobAction *ExtractAPIAction = nullptr; 4042 ActionList LinkerInputs; 4043 ActionList MergerInputs; 4044 4045 for (auto &I : Inputs) { 4046 types::ID InputType = I.first; 4047 const Arg *InputArg = I.second; 4048 4049 auto PL = types::getCompilationPhases(*this, Args, InputType); 4050 if (PL.empty()) 4051 continue; 4052 4053 auto FullPL = types::getCompilationPhases(InputType); 4054 4055 // Build the pipeline for this file. 4056 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4057 4058 // Use the current host action in any of the offloading actions, if 4059 // required. 4060 if (!UseNewOffloadingDriver) 4061 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4062 break; 4063 4064 for (phases::ID Phase : PL) { 4065 4066 // Add any offload action the host action depends on. 4067 if (!UseNewOffloadingDriver) 4068 Current = OffloadBuilder.addDeviceDependencesToHostAction( 4069 Current, InputArg, Phase, PL.back(), FullPL); 4070 if (!Current) 4071 break; 4072 4073 // Queue linker inputs. 4074 if (Phase == phases::Link) { 4075 assert(Phase == PL.back() && "linking must be final compilation step."); 4076 // We don't need to generate additional link commands if emitting AMD bitcode 4077 if (!(C.getInputArgs().hasArg(options::OPT_hip_link) && 4078 (C.getInputArgs().hasArg(options::OPT_emit_llvm)))) 4079 LinkerInputs.push_back(Current); 4080 Current = nullptr; 4081 break; 4082 } 4083 4084 // TODO: Consider removing this because the merged may not end up being 4085 // the final Phase in the pipeline. Perhaps the merged could just merge 4086 // and then pass an artifact of some sort to the Link Phase. 4087 // Queue merger inputs. 4088 if (Phase == phases::IfsMerge) { 4089 assert(Phase == PL.back() && "merging must be final compilation step."); 4090 MergerInputs.push_back(Current); 4091 Current = nullptr; 4092 break; 4093 } 4094 4095 // Each precompiled header file after a module file action is a module 4096 // header of that same module file, rather than being compiled to a 4097 // separate PCH. 4098 if (Phase == phases::Precompile && HeaderModuleAction && 4099 getPrecompiledType(InputType) == types::TY_PCH) { 4100 HeaderModuleAction->addModuleHeaderInput(Current); 4101 Current = nullptr; 4102 break; 4103 } 4104 4105 if (Phase == phases::Precompile && ExtractAPIAction) { 4106 ExtractAPIAction->addHeaderInput(Current); 4107 Current = nullptr; 4108 break; 4109 } 4110 4111 // FIXME: Should we include any prior module file outputs as inputs of 4112 // later actions in the same command line? 4113 4114 // Otherwise construct the appropriate action. 4115 Action *NewCurrent = ConstructPhaseAction(C, Args, Phase, Current); 4116 4117 // We didn't create a new action, so we will just move to the next phase. 4118 if (NewCurrent == Current) 4119 continue; 4120 4121 if (auto *HMA = dyn_cast<HeaderModulePrecompileJobAction>(NewCurrent)) 4122 HeaderModuleAction = HMA; 4123 else if (auto *EAA = dyn_cast<ExtractAPIJobAction>(NewCurrent)) 4124 ExtractAPIAction = EAA; 4125 4126 Current = NewCurrent; 4127 4128 // Use the current host action in any of the offloading actions, if 4129 // required. 4130 if (!UseNewOffloadingDriver) 4131 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4132 break; 4133 4134 // Try to build the offloading actions and add the result as a dependency 4135 // to the host. 4136 if (UseNewOffloadingDriver) 4137 Current = BuildOffloadingActions(C, Args, I, Current); 4138 4139 if (Current->getType() == types::TY_Nothing) 4140 break; 4141 } 4142 4143 // If we ended with something, add to the output list. 4144 if (Current) 4145 Actions.push_back(Current); 4146 4147 // Add any top level actions generated for offloading. 4148 if (!UseNewOffloadingDriver) 4149 OffloadBuilder.appendTopLevelActions(Actions, Current, InputArg); 4150 else if (Current) 4151 Current->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4152 /*BoundArch=*/nullptr); 4153 } 4154 4155 // Add a link action if necessary. 4156 4157 if (LinkerInputs.empty()) { 4158 Arg *FinalPhaseArg; 4159 if (getFinalPhase(Args, &FinalPhaseArg) == phases::Link) 4160 if (!UseNewOffloadingDriver) 4161 OffloadBuilder.appendDeviceLinkActions(Actions); 4162 } 4163 4164 if (!LinkerInputs.empty()) { 4165 if (!UseNewOffloadingDriver) 4166 if (Action *Wrapper = OffloadBuilder.makeHostLinkAction()) 4167 LinkerInputs.push_back(Wrapper); 4168 Action *LA; 4169 // Check if this Linker Job should emit a static library. 4170 if (ShouldEmitStaticLibrary(Args)) { 4171 LA = C.MakeAction<StaticLibJobAction>(LinkerInputs, types::TY_Image); 4172 } else if (UseNewOffloadingDriver || 4173 Args.hasArg(options::OPT_offload_link)) { 4174 LA = C.MakeAction<LinkerWrapperJobAction>(LinkerInputs, types::TY_Image); 4175 LA->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4176 /*BoundArch=*/nullptr); 4177 } else { 4178 LA = C.MakeAction<LinkJobAction>(LinkerInputs, types::TY_Image); 4179 } 4180 if (!UseNewOffloadingDriver) 4181 LA = OffloadBuilder.processHostLinkAction(LA); 4182 Actions.push_back(LA); 4183 } 4184 4185 // Add an interface stubs merge action if necessary. 4186 if (!MergerInputs.empty()) 4187 Actions.push_back( 4188 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4189 4190 if (Args.hasArg(options::OPT_emit_interface_stubs)) { 4191 auto PhaseList = types::getCompilationPhases( 4192 types::TY_IFS_CPP, 4193 Args.hasArg(options::OPT_c) ? phases::Compile : phases::IfsMerge); 4194 4195 ActionList MergerInputs; 4196 4197 for (auto &I : Inputs) { 4198 types::ID InputType = I.first; 4199 const Arg *InputArg = I.second; 4200 4201 // Currently clang and the llvm assembler do not support generating symbol 4202 // stubs from assembly, so we skip the input on asm files. For ifs files 4203 // we rely on the normal pipeline setup in the pipeline setup code above. 4204 if (InputType == types::TY_IFS || InputType == types::TY_PP_Asm || 4205 InputType == types::TY_Asm) 4206 continue; 4207 4208 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4209 4210 for (auto Phase : PhaseList) { 4211 switch (Phase) { 4212 default: 4213 llvm_unreachable( 4214 "IFS Pipeline can only consist of Compile followed by IfsMerge."); 4215 case phases::Compile: { 4216 // Only IfsMerge (llvm-ifs) can handle .o files by looking for ifs 4217 // files where the .o file is located. The compile action can not 4218 // handle this. 4219 if (InputType == types::TY_Object) 4220 break; 4221 4222 Current = C.MakeAction<CompileJobAction>(Current, types::TY_IFS_CPP); 4223 break; 4224 } 4225 case phases::IfsMerge: { 4226 assert(Phase == PhaseList.back() && 4227 "merging must be final compilation step."); 4228 MergerInputs.push_back(Current); 4229 Current = nullptr; 4230 break; 4231 } 4232 } 4233 } 4234 4235 // If we ended with something, add to the output list. 4236 if (Current) 4237 Actions.push_back(Current); 4238 } 4239 4240 // Add an interface stubs merge action if necessary. 4241 if (!MergerInputs.empty()) 4242 Actions.push_back( 4243 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4244 } 4245 4246 // If --print-supported-cpus, -mcpu=? or -mtune=? is specified, build a custom 4247 // Compile phase that prints out supported cpu models and quits. 4248 if (Arg *A = Args.getLastArg(options::OPT_print_supported_cpus)) { 4249 // Use the -mcpu=? flag as the dummy input to cc1. 4250 Actions.clear(); 4251 Action *InputAc = C.MakeAction<InputAction>(*A, types::TY_C); 4252 Actions.push_back( 4253 C.MakeAction<PrecompileJobAction>(InputAc, types::TY_Nothing)); 4254 for (auto &I : Inputs) 4255 I.second->claim(); 4256 } 4257 4258 // Claim ignored clang-cl options. 4259 Args.ClaimAllArgs(options::OPT_cl_ignored_Group); 4260 4261 // Claim --offload-host-only and --offload-compile-host-device, which may be 4262 // passed to non-CUDA compilations and should not trigger warnings there. 4263 Args.ClaimAllArgs(options::OPT_offload_host_only); 4264 Args.ClaimAllArgs(options::OPT_offload_host_device); 4265 } 4266 4267 /// Returns the canonical name for the offloading architecture when using a HIP 4268 /// or CUDA architecture. 4269 static StringRef getCanonicalArchString(Compilation &C, 4270 const llvm::opt::DerivedArgList &Args, 4271 StringRef ArchStr, 4272 const llvm::Triple &Triple) { 4273 // Lookup the CUDA / HIP architecture string. Only report an error if we were 4274 // expecting the triple to be only NVPTX / AMDGPU. 4275 CudaArch Arch = StringToCudaArch(getProcessorFromTargetID(Triple, ArchStr)); 4276 if (Triple.isNVPTX() && 4277 (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch))) { 4278 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4279 << "CUDA" << ArchStr; 4280 return StringRef(); 4281 } else if (Triple.isAMDGPU() && 4282 (Arch == CudaArch::UNKNOWN || !IsAMDGpuArch(Arch))) { 4283 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4284 << "HIP" << ArchStr; 4285 return StringRef(); 4286 } 4287 4288 if (IsNVIDIAGpuArch(Arch)) 4289 return Args.MakeArgStringRef(CudaArchToString(Arch)); 4290 4291 if (IsAMDGpuArch(Arch)) { 4292 llvm::StringMap<bool> Features; 4293 auto HIPTriple = getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()); 4294 if (!HIPTriple) 4295 return StringRef(); 4296 auto Arch = parseTargetID(*HIPTriple, ArchStr, &Features); 4297 if (!Arch) { 4298 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << ArchStr; 4299 C.setContainsError(); 4300 return StringRef(); 4301 } 4302 return Args.MakeArgStringRef( 4303 getCanonicalTargetID(Arch.getValue(), Features)); 4304 } 4305 4306 // If the input isn't CUDA or HIP just return the architecture. 4307 return ArchStr; 4308 } 4309 4310 /// Checks if the set offloading architectures does not conflict. Returns the 4311 /// incompatible pair if a conflict occurs. 4312 static llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 4313 getConflictOffloadArchCombination(const llvm::DenseSet<StringRef> &Archs, 4314 Action::OffloadKind Kind) { 4315 if (Kind != Action::OFK_HIP) 4316 return None; 4317 4318 std::set<StringRef> ArchSet; 4319 llvm::copy(Archs, std::inserter(ArchSet, ArchSet.begin())); 4320 return getConflictTargetIDCombination(ArchSet); 4321 } 4322 4323 llvm::DenseSet<StringRef> 4324 Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args, 4325 Action::OffloadKind Kind, const ToolChain *TC) const { 4326 if (!TC) 4327 TC = &C.getDefaultToolChain(); 4328 4329 // --offload and --offload-arch options are mutually exclusive. 4330 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 4331 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 4332 options::OPT_no_offload_arch_EQ)) { 4333 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) 4334 << "--offload" 4335 << (Args.hasArgNoClaim(options::OPT_offload_arch_EQ) 4336 ? "--offload-arch" 4337 : "--no-offload-arch"); 4338 } 4339 4340 if (KnownArchs.find(TC) != KnownArchs.end()) 4341 return KnownArchs.lookup(TC); 4342 4343 llvm::DenseSet<StringRef> Archs; 4344 for (auto &Arg : Args) { 4345 if (Arg->getOption().matches(options::OPT_offload_arch_EQ)) { 4346 Archs.insert( 4347 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4348 } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) { 4349 if (Arg->getValue() == StringRef("all")) 4350 Archs.clear(); 4351 else 4352 Archs.erase( 4353 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4354 } 4355 } 4356 4357 if (auto ConflictingArchs = getConflictOffloadArchCombination(Archs, Kind)) { 4358 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 4359 << ConflictingArchs.getValue().first 4360 << ConflictingArchs.getValue().second; 4361 C.setContainsError(); 4362 } 4363 4364 if (Archs.empty()) { 4365 if (Kind == Action::OFK_Cuda) 4366 Archs.insert(CudaArchToString(CudaArch::CudaDefault)); 4367 else if (Kind == Action::OFK_HIP) 4368 Archs.insert(CudaArchToString(CudaArch::HIPDefault)); 4369 else if (Kind == Action::OFK_OpenMP) 4370 Archs.insert(StringRef()); 4371 } else { 4372 Args.ClaimAllArgs(options::OPT_offload_arch_EQ); 4373 Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ); 4374 } 4375 4376 return Archs; 4377 } 4378 4379 Action *Driver::BuildOffloadingActions(Compilation &C, 4380 llvm::opt::DerivedArgList &Args, 4381 const InputTy &Input, 4382 Action *HostAction) const { 4383 const Arg *Mode = Args.getLastArg(options::OPT_offload_host_only, 4384 options::OPT_offload_device_only, 4385 options::OPT_offload_host_device); 4386 const bool HostOnly = 4387 Mode && Mode->getOption().matches(options::OPT_offload_host_only); 4388 const bool DeviceOnly = 4389 Mode && Mode->getOption().matches(options::OPT_offload_device_only); 4390 4391 // Don't build offloading actions if explicitly disabled or we do not have a 4392 // valid source input and compile action to embed it in. If preprocessing only 4393 // ignore embedding. 4394 if (HostOnly || !types::isSrcFile(Input.first) || 4395 !(isa<CompileJobAction>(HostAction) || 4396 getFinalPhase(Args) == phases::Preprocess)) 4397 return HostAction; 4398 4399 ActionList OffloadActions; 4400 OffloadAction::DeviceDependences DDeps; 4401 4402 const Action::OffloadKind OffloadKinds[] = { 4403 Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP}; 4404 4405 for (Action::OffloadKind Kind : OffloadKinds) { 4406 SmallVector<const ToolChain *, 2> ToolChains; 4407 ActionList DeviceActions; 4408 4409 auto TCRange = C.getOffloadToolChains(Kind); 4410 for (auto TI = TCRange.first, TE = TCRange.second; TI != TE; ++TI) 4411 ToolChains.push_back(TI->second); 4412 4413 if (ToolChains.empty()) 4414 continue; 4415 4416 types::ID InputType = Input.first; 4417 const Arg *InputArg = Input.second; 4418 4419 // Get the product of all bound architectures and toolchains. 4420 SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs; 4421 for (const ToolChain *TC : ToolChains) 4422 for (StringRef Arch : getOffloadArchs( 4423 C, C.getArgsForToolChain(TC, "generic", Kind), Kind, TC)) 4424 TCAndArchs.push_back(std::make_pair(TC, Arch)); 4425 4426 for (unsigned I = 0, E = TCAndArchs.size(); I != E; ++I) 4427 DeviceActions.push_back(C.MakeAction<InputAction>(*InputArg, InputType)); 4428 4429 if (DeviceActions.empty()) 4430 return HostAction; 4431 4432 auto PL = types::getCompilationPhases(*this, Args, InputType); 4433 4434 for (phases::ID Phase : PL) { 4435 if (Phase == phases::Link) { 4436 assert(Phase == PL.back() && "linking must be final compilation step."); 4437 break; 4438 } 4439 4440 auto TCAndArch = TCAndArchs.begin(); 4441 for (Action *&A : DeviceActions) { 4442 A = ConstructPhaseAction(C, Args, Phase, A, Kind); 4443 4444 if (isa<CompileJobAction>(A) && isa<CompileJobAction>(HostAction) && 4445 Kind == Action::OFK_OpenMP) { 4446 // OpenMP offloading has a dependency on the host compile action to 4447 // identify which declarations need to be emitted. This shouldn't be 4448 // collapsed with any other actions so we can use it in the device. 4449 HostAction->setCannotBeCollapsedWithNextDependentAction(); 4450 OffloadAction::HostDependence HDep( 4451 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4452 TCAndArch->second.data(), Kind); 4453 OffloadAction::DeviceDependences DDep; 4454 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4455 A = C.MakeAction<OffloadAction>(HDep, DDep); 4456 } else if (isa<AssembleJobAction>(A) && Kind == Action::OFK_Cuda) { 4457 // The Cuda toolchain uses fatbinary as the linker phase to bundle the 4458 // PTX and Cubin output. 4459 ActionList FatbinActions; 4460 for (Action *A : {A, A->getInputs()[0]}) { 4461 OffloadAction::DeviceDependences DDep; 4462 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4463 FatbinActions.emplace_back( 4464 C.MakeAction<OffloadAction>(DDep, A->getType())); 4465 } 4466 A = C.MakeAction<LinkJobAction>(FatbinActions, types::TY_CUDA_FATBIN); 4467 } 4468 ++TCAndArch; 4469 } 4470 } 4471 4472 auto TCAndArch = TCAndArchs.begin(); 4473 for (Action *A : DeviceActions) { 4474 DDeps.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4475 OffloadAction::DeviceDependences DDep; 4476 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4477 OffloadActions.push_back(C.MakeAction<OffloadAction>(DDep, A->getType())); 4478 ++TCAndArch; 4479 } 4480 } 4481 4482 if (DeviceOnly) 4483 return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing); 4484 4485 Action *OffloadPackager = 4486 C.MakeAction<OffloadPackagerJobAction>(OffloadActions, types::TY_Image); 4487 OffloadAction::DeviceDependences DDep; 4488 DDep.add(*OffloadPackager, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4489 nullptr, Action::OFK_None); 4490 OffloadAction::HostDependence HDep( 4491 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4492 /*BoundArch=*/nullptr, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4493 return C.MakeAction<OffloadAction>( 4494 HDep, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4495 } 4496 4497 Action *Driver::ConstructPhaseAction( 4498 Compilation &C, const ArgList &Args, phases::ID Phase, Action *Input, 4499 Action::OffloadKind TargetDeviceOffloadKind) const { 4500 llvm::PrettyStackTraceString CrashInfo("Constructing phase actions"); 4501 4502 // Some types skip the assembler phase (e.g., llvm-bc), but we can't 4503 // encode this in the steps because the intermediate type depends on 4504 // arguments. Just special case here. 4505 if (Phase == phases::Assemble && Input->getType() != types::TY_PP_Asm) 4506 return Input; 4507 4508 // Build the appropriate action. 4509 switch (Phase) { 4510 case phases::Link: 4511 llvm_unreachable("link action invalid here."); 4512 case phases::IfsMerge: 4513 llvm_unreachable("ifsmerge action invalid here."); 4514 case phases::Preprocess: { 4515 types::ID OutputTy; 4516 // -M and -MM specify the dependency file name by altering the output type, 4517 // -if -MD and -MMD are not specified. 4518 if (Args.hasArg(options::OPT_M, options::OPT_MM) && 4519 !Args.hasArg(options::OPT_MD, options::OPT_MMD)) { 4520 OutputTy = types::TY_Dependencies; 4521 } else { 4522 OutputTy = Input->getType(); 4523 // For these cases, the preprocessor is only translating forms, the Output 4524 // still needs preprocessing. 4525 if (!Args.hasFlag(options::OPT_frewrite_includes, 4526 options::OPT_fno_rewrite_includes, false) && 4527 !Args.hasFlag(options::OPT_frewrite_imports, 4528 options::OPT_fno_rewrite_imports, false) && 4529 !Args.hasFlag(options::OPT_fdirectives_only, 4530 options::OPT_fno_directives_only, false) && 4531 !CCGenDiagnostics) 4532 OutputTy = types::getPreprocessedType(OutputTy); 4533 assert(OutputTy != types::TY_INVALID && 4534 "Cannot preprocess this input type!"); 4535 } 4536 return C.MakeAction<PreprocessJobAction>(Input, OutputTy); 4537 } 4538 case phases::Precompile: { 4539 // API extraction should not generate an actual precompilation action. 4540 if (Args.hasArg(options::OPT_extract_api)) 4541 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4542 4543 types::ID OutputTy = getPrecompiledType(Input->getType()); 4544 assert(OutputTy != types::TY_INVALID && 4545 "Cannot precompile this input type!"); 4546 4547 // If we're given a module name, precompile header file inputs as a 4548 // module, not as a precompiled header. 4549 const char *ModName = nullptr; 4550 if (OutputTy == types::TY_PCH) { 4551 if (Arg *A = Args.getLastArg(options::OPT_fmodule_name_EQ)) 4552 ModName = A->getValue(); 4553 if (ModName) 4554 OutputTy = types::TY_ModuleFile; 4555 } 4556 4557 if (Args.hasArg(options::OPT_fsyntax_only)) { 4558 // Syntax checks should not emit a PCH file 4559 OutputTy = types::TY_Nothing; 4560 } 4561 4562 if (ModName) 4563 return C.MakeAction<HeaderModulePrecompileJobAction>(Input, OutputTy, 4564 ModName); 4565 return C.MakeAction<PrecompileJobAction>(Input, OutputTy); 4566 } 4567 case phases::Compile: { 4568 if (Args.hasArg(options::OPT_fsyntax_only)) 4569 return C.MakeAction<CompileJobAction>(Input, types::TY_Nothing); 4570 if (Args.hasArg(options::OPT_rewrite_objc)) 4571 return C.MakeAction<CompileJobAction>(Input, types::TY_RewrittenObjC); 4572 if (Args.hasArg(options::OPT_rewrite_legacy_objc)) 4573 return C.MakeAction<CompileJobAction>(Input, 4574 types::TY_RewrittenLegacyObjC); 4575 if (Args.hasArg(options::OPT__analyze)) 4576 return C.MakeAction<AnalyzeJobAction>(Input, types::TY_Plist); 4577 if (Args.hasArg(options::OPT__migrate)) 4578 return C.MakeAction<MigrateJobAction>(Input, types::TY_Remap); 4579 if (Args.hasArg(options::OPT_emit_ast)) 4580 return C.MakeAction<CompileJobAction>(Input, types::TY_AST); 4581 if (Args.hasArg(options::OPT_module_file_info)) 4582 return C.MakeAction<CompileJobAction>(Input, types::TY_ModuleFile); 4583 if (Args.hasArg(options::OPT_verify_pch)) 4584 return C.MakeAction<VerifyPCHJobAction>(Input, types::TY_Nothing); 4585 if (Args.hasArg(options::OPT_extract_api)) 4586 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4587 return C.MakeAction<CompileJobAction>(Input, types::TY_LLVM_BC); 4588 } 4589 case phases::Backend: { 4590 if (isUsingLTO() && TargetDeviceOffloadKind == Action::OFK_None) { 4591 types::ID Output = 4592 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4593 return C.MakeAction<BackendJobAction>(Input, Output); 4594 } 4595 if (isUsingLTO(/* IsOffload */ true) && 4596 TargetDeviceOffloadKind != Action::OFK_None) { 4597 types::ID Output = 4598 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4599 return C.MakeAction<BackendJobAction>(Input, Output); 4600 } 4601 if (Args.hasArg(options::OPT_emit_llvm) || 4602 (TargetDeviceOffloadKind == Action::OFK_HIP && 4603 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4604 false))) { 4605 types::ID Output = 4606 Args.hasArg(options::OPT_S) ? types::TY_LLVM_IR : types::TY_LLVM_BC; 4607 return C.MakeAction<BackendJobAction>(Input, Output); 4608 } 4609 return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm); 4610 } 4611 case phases::Assemble: 4612 return C.MakeAction<AssembleJobAction>(std::move(Input), types::TY_Object); 4613 } 4614 4615 llvm_unreachable("invalid phase in ConstructPhaseAction"); 4616 } 4617 4618 void Driver::BuildJobs(Compilation &C) const { 4619 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 4620 4621 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 4622 4623 // It is an error to provide a -o option if we are making multiple output 4624 // files. There are exceptions: 4625 // 4626 // IfsMergeJob: when generating interface stubs enabled we want to be able to 4627 // generate the stub file at the same time that we generate the real 4628 // library/a.out. So when a .o, .so, etc are the output, with clang interface 4629 // stubs there will also be a .ifs and .ifso at the same location. 4630 // 4631 // CompileJob of type TY_IFS_CPP: when generating interface stubs is enabled 4632 // and -c is passed, we still want to be able to generate a .ifs file while 4633 // we are also generating .o files. So we allow more than one output file in 4634 // this case as well. 4635 // 4636 if (FinalOutput) { 4637 unsigned NumOutputs = 0; 4638 unsigned NumIfsOutputs = 0; 4639 for (const Action *A : C.getActions()) 4640 if (A->getType() != types::TY_Nothing && 4641 !(A->getKind() == Action::IfsMergeJobClass || 4642 (A->getType() == clang::driver::types::TY_IFS_CPP && 4643 A->getKind() == clang::driver::Action::CompileJobClass && 4644 0 == NumIfsOutputs++) || 4645 (A->getKind() == Action::BindArchClass && A->getInputs().size() && 4646 A->getInputs().front()->getKind() == Action::IfsMergeJobClass))) 4647 ++NumOutputs; 4648 4649 if (NumOutputs > 1) { 4650 Diag(clang::diag::err_drv_output_argument_with_multiple_files); 4651 FinalOutput = nullptr; 4652 } 4653 } 4654 4655 const llvm::Triple &RawTriple = C.getDefaultToolChain().getTriple(); 4656 if (RawTriple.isOSAIX()) { 4657 if (Arg *A = C.getArgs().getLastArg(options::OPT_G)) 4658 Diag(diag::err_drv_unsupported_opt_for_target) 4659 << A->getSpelling() << RawTriple.str(); 4660 if (LTOMode == LTOK_Thin) 4661 Diag(diag::err_drv_clang_unsupported) << "thinLTO on AIX"; 4662 } 4663 4664 // Collect the list of architectures. 4665 llvm::StringSet<> ArchNames; 4666 if (RawTriple.isOSBinFormatMachO()) 4667 for (const Arg *A : C.getArgs()) 4668 if (A->getOption().matches(options::OPT_arch)) 4669 ArchNames.insert(A->getValue()); 4670 4671 // Set of (Action, canonical ToolChain triple) pairs we've built jobs for. 4672 std::map<std::pair<const Action *, std::string>, InputInfoList> CachedResults; 4673 for (Action *A : C.getActions()) { 4674 // If we are linking an image for multiple archs then the linker wants 4675 // -arch_multiple and -final_output <final image name>. Unfortunately, this 4676 // doesn't fit in cleanly because we have to pass this information down. 4677 // 4678 // FIXME: This is a hack; find a cleaner way to integrate this into the 4679 // process. 4680 const char *LinkingOutput = nullptr; 4681 if (isa<LipoJobAction>(A)) { 4682 if (FinalOutput) 4683 LinkingOutput = FinalOutput->getValue(); 4684 else 4685 LinkingOutput = getDefaultImageName(); 4686 } 4687 4688 BuildJobsForAction(C, A, &C.getDefaultToolChain(), 4689 /*BoundArch*/ StringRef(), 4690 /*AtTopLevel*/ true, 4691 /*MultipleArchs*/ ArchNames.size() > 1, 4692 /*LinkingOutput*/ LinkingOutput, CachedResults, 4693 /*TargetDeviceOffloadKind*/ Action::OFK_None); 4694 } 4695 4696 // If we have more than one job, then disable integrated-cc1 for now. Do this 4697 // also when we need to report process execution statistics. 4698 if (C.getJobs().size() > 1 || CCPrintProcessStats) 4699 for (auto &J : C.getJobs()) 4700 J.InProcess = false; 4701 4702 if (CCPrintProcessStats) { 4703 C.setPostCallback([=](const Command &Cmd, int Res) { 4704 Optional<llvm::sys::ProcessStatistics> ProcStat = 4705 Cmd.getProcessStatistics(); 4706 if (!ProcStat) 4707 return; 4708 4709 const char *LinkingOutput = nullptr; 4710 if (FinalOutput) 4711 LinkingOutput = FinalOutput->getValue(); 4712 else if (!Cmd.getOutputFilenames().empty()) 4713 LinkingOutput = Cmd.getOutputFilenames().front().c_str(); 4714 else 4715 LinkingOutput = getDefaultImageName(); 4716 4717 if (CCPrintStatReportFilename.empty()) { 4718 using namespace llvm; 4719 // Human readable output. 4720 outs() << sys::path::filename(Cmd.getExecutable()) << ": " 4721 << "output=" << LinkingOutput; 4722 outs() << ", total=" 4723 << format("%.3f", ProcStat->TotalTime.count() / 1000.) << " ms" 4724 << ", user=" 4725 << format("%.3f", ProcStat->UserTime.count() / 1000.) << " ms" 4726 << ", mem=" << ProcStat->PeakMemory << " Kb\n"; 4727 } else { 4728 // CSV format. 4729 std::string Buffer; 4730 llvm::raw_string_ostream Out(Buffer); 4731 llvm::sys::printArg(Out, llvm::sys::path::filename(Cmd.getExecutable()), 4732 /*Quote*/ true); 4733 Out << ','; 4734 llvm::sys::printArg(Out, LinkingOutput, true); 4735 Out << ',' << ProcStat->TotalTime.count() << ',' 4736 << ProcStat->UserTime.count() << ',' << ProcStat->PeakMemory 4737 << '\n'; 4738 Out.flush(); 4739 std::error_code EC; 4740 llvm::raw_fd_ostream OS(CCPrintStatReportFilename, EC, 4741 llvm::sys::fs::OF_Append | 4742 llvm::sys::fs::OF_Text); 4743 if (EC) 4744 return; 4745 auto L = OS.lock(); 4746 if (!L) { 4747 llvm::errs() << "ERROR: Cannot lock file " 4748 << CCPrintStatReportFilename << ": " 4749 << toString(L.takeError()) << "\n"; 4750 return; 4751 } 4752 OS << Buffer; 4753 OS.flush(); 4754 } 4755 }); 4756 } 4757 4758 // If the user passed -Qunused-arguments or there were errors, don't warn 4759 // about any unused arguments. 4760 if (Diags.hasErrorOccurred() || 4761 C.getArgs().hasArg(options::OPT_Qunused_arguments)) 4762 return; 4763 4764 // Claim -### here. 4765 (void)C.getArgs().hasArg(options::OPT__HASH_HASH_HASH); 4766 4767 // Claim --driver-mode, --rsp-quoting, it was handled earlier. 4768 (void)C.getArgs().hasArg(options::OPT_driver_mode); 4769 (void)C.getArgs().hasArg(options::OPT_rsp_quoting); 4770 4771 for (Arg *A : C.getArgs()) { 4772 // FIXME: It would be nice to be able to send the argument to the 4773 // DiagnosticsEngine, so that extra values, position, and so on could be 4774 // printed. 4775 if (!A->isClaimed()) { 4776 if (A->getOption().hasFlag(options::NoArgumentUnused)) 4777 continue; 4778 4779 // Suppress the warning automatically if this is just a flag, and it is an 4780 // instance of an argument we already claimed. 4781 const Option &Opt = A->getOption(); 4782 if (Opt.getKind() == Option::FlagClass) { 4783 bool DuplicateClaimed = false; 4784 4785 for (const Arg *AA : C.getArgs().filtered(&Opt)) { 4786 if (AA->isClaimed()) { 4787 DuplicateClaimed = true; 4788 break; 4789 } 4790 } 4791 4792 if (DuplicateClaimed) 4793 continue; 4794 } 4795 4796 // In clang-cl, don't mention unknown arguments here since they have 4797 // already been warned about. 4798 if (!IsCLMode() || !A->getOption().matches(options::OPT_UNKNOWN)) 4799 Diag(clang::diag::warn_drv_unused_argument) 4800 << A->getAsString(C.getArgs()); 4801 } 4802 } 4803 } 4804 4805 namespace { 4806 /// Utility class to control the collapse of dependent actions and select the 4807 /// tools accordingly. 4808 class ToolSelector final { 4809 /// The tool chain this selector refers to. 4810 const ToolChain &TC; 4811 4812 /// The compilation this selector refers to. 4813 const Compilation &C; 4814 4815 /// The base action this selector refers to. 4816 const JobAction *BaseAction; 4817 4818 /// Set to true if the current toolchain refers to host actions. 4819 bool IsHostSelector; 4820 4821 /// Set to true if save-temps and embed-bitcode functionalities are active. 4822 bool SaveTemps; 4823 bool EmbedBitcode; 4824 4825 /// Get previous dependent action or null if that does not exist. If 4826 /// \a CanBeCollapsed is false, that action must be legal to collapse or 4827 /// null will be returned. 4828 const JobAction *getPrevDependentAction(const ActionList &Inputs, 4829 ActionList &SavedOffloadAction, 4830 bool CanBeCollapsed = true) { 4831 // An option can be collapsed only if it has a single input. 4832 if (Inputs.size() != 1) 4833 return nullptr; 4834 4835 Action *CurAction = *Inputs.begin(); 4836 if (CanBeCollapsed && 4837 !CurAction->isCollapsingWithNextDependentActionLegal()) 4838 return nullptr; 4839 4840 // If the input action is an offload action. Look through it and save any 4841 // offload action that can be dropped in the event of a collapse. 4842 if (auto *OA = dyn_cast<OffloadAction>(CurAction)) { 4843 // If the dependent action is a device action, we will attempt to collapse 4844 // only with other device actions. Otherwise, we would do the same but 4845 // with host actions only. 4846 if (!IsHostSelector) { 4847 if (OA->hasSingleDeviceDependence(/*DoNotConsiderHostActions=*/true)) { 4848 CurAction = 4849 OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true); 4850 if (CanBeCollapsed && 4851 !CurAction->isCollapsingWithNextDependentActionLegal()) 4852 return nullptr; 4853 SavedOffloadAction.push_back(OA); 4854 return dyn_cast<JobAction>(CurAction); 4855 } 4856 } else if (OA->hasHostDependence()) { 4857 CurAction = OA->getHostDependence(); 4858 if (CanBeCollapsed && 4859 !CurAction->isCollapsingWithNextDependentActionLegal()) 4860 return nullptr; 4861 SavedOffloadAction.push_back(OA); 4862 return dyn_cast<JobAction>(CurAction); 4863 } 4864 return nullptr; 4865 } 4866 4867 return dyn_cast<JobAction>(CurAction); 4868 } 4869 4870 /// Return true if an assemble action can be collapsed. 4871 bool canCollapseAssembleAction() const { 4872 return TC.useIntegratedAs() && !SaveTemps && 4873 !C.getArgs().hasArg(options::OPT_via_file_asm) && 4874 !C.getArgs().hasArg(options::OPT__SLASH_FA) && 4875 !C.getArgs().hasArg(options::OPT__SLASH_Fa); 4876 } 4877 4878 /// Return true if a preprocessor action can be collapsed. 4879 bool canCollapsePreprocessorAction() const { 4880 return !C.getArgs().hasArg(options::OPT_no_integrated_cpp) && 4881 !C.getArgs().hasArg(options::OPT_traditional_cpp) && !SaveTemps && 4882 !C.getArgs().hasArg(options::OPT_rewrite_objc); 4883 } 4884 4885 /// Struct that relates an action with the offload actions that would be 4886 /// collapsed with it. 4887 struct JobActionInfo final { 4888 /// The action this info refers to. 4889 const JobAction *JA = nullptr; 4890 /// The offload actions we need to take care off if this action is 4891 /// collapsed. 4892 ActionList SavedOffloadAction; 4893 }; 4894 4895 /// Append collapsed offload actions from the give nnumber of elements in the 4896 /// action info array. 4897 static void AppendCollapsedOffloadAction(ActionList &CollapsedOffloadAction, 4898 ArrayRef<JobActionInfo> &ActionInfo, 4899 unsigned ElementNum) { 4900 assert(ElementNum <= ActionInfo.size() && "Invalid number of elements."); 4901 for (unsigned I = 0; I < ElementNum; ++I) 4902 CollapsedOffloadAction.append(ActionInfo[I].SavedOffloadAction.begin(), 4903 ActionInfo[I].SavedOffloadAction.end()); 4904 } 4905 4906 /// Functions that attempt to perform the combining. They detect if that is 4907 /// legal, and if so they update the inputs \a Inputs and the offload action 4908 /// that were collapsed in \a CollapsedOffloadAction. A tool that deals with 4909 /// the combined action is returned. If the combining is not legal or if the 4910 /// tool does not exist, null is returned. 4911 /// Currently three kinds of collapsing are supported: 4912 /// - Assemble + Backend + Compile; 4913 /// - Assemble + Backend ; 4914 /// - Backend + Compile. 4915 const Tool * 4916 combineAssembleBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4917 ActionList &Inputs, 4918 ActionList &CollapsedOffloadAction) { 4919 if (ActionInfo.size() < 3 || !canCollapseAssembleAction()) 4920 return nullptr; 4921 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4922 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4923 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[2].JA); 4924 if (!AJ || !BJ || !CJ) 4925 return nullptr; 4926 4927 // Get compiler tool. 4928 const Tool *T = TC.SelectTool(*CJ); 4929 if (!T) 4930 return nullptr; 4931 4932 // Can't collapse if we don't have codegen support unless we are 4933 // emitting LLVM IR. 4934 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 4935 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 4936 return nullptr; 4937 4938 // When using -fembed-bitcode, it is required to have the same tool (clang) 4939 // for both CompilerJA and BackendJA. Otherwise, combine two stages. 4940 if (EmbedBitcode) { 4941 const Tool *BT = TC.SelectTool(*BJ); 4942 if (BT == T) 4943 return nullptr; 4944 } 4945 4946 if (!T->hasIntegratedAssembler()) 4947 return nullptr; 4948 4949 Inputs = CJ->getInputs(); 4950 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4951 /*NumElements=*/3); 4952 return T; 4953 } 4954 const Tool *combineAssembleBackend(ArrayRef<JobActionInfo> ActionInfo, 4955 ActionList &Inputs, 4956 ActionList &CollapsedOffloadAction) { 4957 if (ActionInfo.size() < 2 || !canCollapseAssembleAction()) 4958 return nullptr; 4959 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4960 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4961 if (!AJ || !BJ) 4962 return nullptr; 4963 4964 // Get backend tool. 4965 const Tool *T = TC.SelectTool(*BJ); 4966 if (!T) 4967 return nullptr; 4968 4969 if (!T->hasIntegratedAssembler()) 4970 return nullptr; 4971 4972 Inputs = BJ->getInputs(); 4973 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4974 /*NumElements=*/2); 4975 return T; 4976 } 4977 const Tool *combineBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4978 ActionList &Inputs, 4979 ActionList &CollapsedOffloadAction) { 4980 if (ActionInfo.size() < 2) 4981 return nullptr; 4982 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[0].JA); 4983 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[1].JA); 4984 if (!BJ || !CJ) 4985 return nullptr; 4986 4987 // Check if the initial input (to the compile job or its predessor if one 4988 // exists) is LLVM bitcode. In that case, no preprocessor step is required 4989 // and we can still collapse the compile and backend jobs when we have 4990 // -save-temps. I.e. there is no need for a separate compile job just to 4991 // emit unoptimized bitcode. 4992 bool InputIsBitcode = true; 4993 for (size_t i = 1; i < ActionInfo.size(); i++) 4994 if (ActionInfo[i].JA->getType() != types::TY_LLVM_BC && 4995 ActionInfo[i].JA->getType() != types::TY_LTO_BC) { 4996 InputIsBitcode = false; 4997 break; 4998 } 4999 if (!InputIsBitcode && !canCollapsePreprocessorAction()) 5000 return nullptr; 5001 5002 // Get compiler tool. 5003 const Tool *T = TC.SelectTool(*CJ); 5004 if (!T) 5005 return nullptr; 5006 5007 // Can't collapse if we don't have codegen support unless we are 5008 // emitting LLVM IR. 5009 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 5010 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 5011 return nullptr; 5012 5013 if (T->canEmitIR() && ((SaveTemps && !InputIsBitcode) || EmbedBitcode)) 5014 return nullptr; 5015 5016 Inputs = CJ->getInputs(); 5017 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5018 /*NumElements=*/2); 5019 return T; 5020 } 5021 5022 /// Updates the inputs if the obtained tool supports combining with 5023 /// preprocessor action, and the current input is indeed a preprocessor 5024 /// action. If combining results in the collapse of offloading actions, those 5025 /// are appended to \a CollapsedOffloadAction. 5026 void combineWithPreprocessor(const Tool *T, ActionList &Inputs, 5027 ActionList &CollapsedOffloadAction) { 5028 if (!T || !canCollapsePreprocessorAction() || !T->hasIntegratedCPP()) 5029 return; 5030 5031 // Attempt to get a preprocessor action dependence. 5032 ActionList PreprocessJobOffloadActions; 5033 ActionList NewInputs; 5034 for (Action *A : Inputs) { 5035 auto *PJ = getPrevDependentAction({A}, PreprocessJobOffloadActions); 5036 if (!PJ || !isa<PreprocessJobAction>(PJ)) { 5037 NewInputs.push_back(A); 5038 continue; 5039 } 5040 5041 // This is legal to combine. Append any offload action we found and add the 5042 // current input to preprocessor inputs. 5043 CollapsedOffloadAction.append(PreprocessJobOffloadActions.begin(), 5044 PreprocessJobOffloadActions.end()); 5045 NewInputs.append(PJ->input_begin(), PJ->input_end()); 5046 } 5047 Inputs = NewInputs; 5048 } 5049 5050 public: 5051 ToolSelector(const JobAction *BaseAction, const ToolChain &TC, 5052 const Compilation &C, bool SaveTemps, bool EmbedBitcode) 5053 : TC(TC), C(C), BaseAction(BaseAction), SaveTemps(SaveTemps), 5054 EmbedBitcode(EmbedBitcode) { 5055 assert(BaseAction && "Invalid base action."); 5056 IsHostSelector = BaseAction->getOffloadingDeviceKind() == Action::OFK_None; 5057 } 5058 5059 /// Check if a chain of actions can be combined and return the tool that can 5060 /// handle the combination of actions. The pointer to the current inputs \a 5061 /// Inputs and the list of offload actions \a CollapsedOffloadActions 5062 /// connected to collapsed actions are updated accordingly. The latter enables 5063 /// the caller of the selector to process them afterwards instead of just 5064 /// dropping them. If no suitable tool is found, null will be returned. 5065 const Tool *getTool(ActionList &Inputs, 5066 ActionList &CollapsedOffloadAction) { 5067 // 5068 // Get the largest chain of actions that we could combine. 5069 // 5070 5071 SmallVector<JobActionInfo, 5> ActionChain(1); 5072 ActionChain.back().JA = BaseAction; 5073 while (ActionChain.back().JA) { 5074 const Action *CurAction = ActionChain.back().JA; 5075 5076 // Grow the chain by one element. 5077 ActionChain.resize(ActionChain.size() + 1); 5078 JobActionInfo &AI = ActionChain.back(); 5079 5080 // Attempt to fill it with the 5081 AI.JA = 5082 getPrevDependentAction(CurAction->getInputs(), AI.SavedOffloadAction); 5083 } 5084 5085 // Pop the last action info as it could not be filled. 5086 ActionChain.pop_back(); 5087 5088 // 5089 // Attempt to combine actions. If all combining attempts failed, just return 5090 // the tool of the provided action. At the end we attempt to combine the 5091 // action with any preprocessor action it may depend on. 5092 // 5093 5094 const Tool *T = combineAssembleBackendCompile(ActionChain, Inputs, 5095 CollapsedOffloadAction); 5096 if (!T) 5097 T = combineAssembleBackend(ActionChain, Inputs, CollapsedOffloadAction); 5098 if (!T) 5099 T = combineBackendCompile(ActionChain, Inputs, CollapsedOffloadAction); 5100 if (!T) { 5101 Inputs = BaseAction->getInputs(); 5102 T = TC.SelectTool(*BaseAction); 5103 } 5104 5105 combineWithPreprocessor(T, Inputs, CollapsedOffloadAction); 5106 return T; 5107 } 5108 }; 5109 } 5110 5111 /// Return a string that uniquely identifies the result of a job. The bound arch 5112 /// is not necessarily represented in the toolchain's triple -- for example, 5113 /// armv7 and armv7s both map to the same triple -- so we need both in our map. 5114 /// Also, we need to add the offloading device kind, as the same tool chain can 5115 /// be used for host and device for some programming models, e.g. OpenMP. 5116 static std::string GetTriplePlusArchString(const ToolChain *TC, 5117 StringRef BoundArch, 5118 Action::OffloadKind OffloadKind) { 5119 std::string TriplePlusArch = TC->getTriple().normalize(); 5120 if (!BoundArch.empty()) { 5121 TriplePlusArch += "-"; 5122 TriplePlusArch += BoundArch; 5123 } 5124 TriplePlusArch += "-"; 5125 TriplePlusArch += Action::GetOffloadKindName(OffloadKind); 5126 return TriplePlusArch; 5127 } 5128 5129 InputInfoList Driver::BuildJobsForAction( 5130 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5131 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5132 std::map<std::pair<const Action *, std::string>, InputInfoList> 5133 &CachedResults, 5134 Action::OffloadKind TargetDeviceOffloadKind) const { 5135 std::pair<const Action *, std::string> ActionTC = { 5136 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5137 auto CachedResult = CachedResults.find(ActionTC); 5138 if (CachedResult != CachedResults.end()) { 5139 return CachedResult->second; 5140 } 5141 InputInfoList Result = BuildJobsForActionNoCache( 5142 C, A, TC, BoundArch, AtTopLevel, MultipleArchs, LinkingOutput, 5143 CachedResults, TargetDeviceOffloadKind); 5144 CachedResults[ActionTC] = Result; 5145 return Result; 5146 } 5147 5148 InputInfoList Driver::BuildJobsForActionNoCache( 5149 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5150 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5151 std::map<std::pair<const Action *, std::string>, InputInfoList> 5152 &CachedResults, 5153 Action::OffloadKind TargetDeviceOffloadKind) const { 5154 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 5155 5156 InputInfoList OffloadDependencesInputInfo; 5157 bool BuildingForOffloadDevice = TargetDeviceOffloadKind != Action::OFK_None; 5158 if (const OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 5159 // The 'Darwin' toolchain is initialized only when its arguments are 5160 // computed. Get the default arguments for OFK_None to ensure that 5161 // initialization is performed before processing the offload action. 5162 // FIXME: Remove when darwin's toolchain is initialized during construction. 5163 C.getArgsForToolChain(TC, BoundArch, Action::OFK_None); 5164 5165 // The offload action is expected to be used in four different situations. 5166 // 5167 // a) Set a toolchain/architecture/kind for a host action: 5168 // Host Action 1 -> OffloadAction -> Host Action 2 5169 // 5170 // b) Set a toolchain/architecture/kind for a device action; 5171 // Device Action 1 -> OffloadAction -> Device Action 2 5172 // 5173 // c) Specify a device dependence to a host action; 5174 // Device Action 1 _ 5175 // \ 5176 // Host Action 1 ---> OffloadAction -> Host Action 2 5177 // 5178 // d) Specify a host dependence to a device action. 5179 // Host Action 1 _ 5180 // \ 5181 // Device Action 1 ---> OffloadAction -> Device Action 2 5182 // 5183 // For a) and b), we just return the job generated for the dependence. For 5184 // c) and d) we override the current action with the host/device dependence 5185 // if the current toolchain is host/device and set the offload dependences 5186 // info with the jobs obtained from the device/host dependence(s). 5187 5188 // If there is a single device option, just generate the job for it. 5189 if (OA->hasSingleDeviceDependence()) { 5190 InputInfoList DevA; 5191 OA->doOnEachDeviceDependence([&](Action *DepA, const ToolChain *DepTC, 5192 const char *DepBoundArch) { 5193 DevA = 5194 BuildJobsForAction(C, DepA, DepTC, DepBoundArch, AtTopLevel, 5195 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, 5196 CachedResults, DepA->getOffloadingDeviceKind()); 5197 }); 5198 return DevA; 5199 } 5200 5201 // If 'Action 2' is host, we generate jobs for the device dependences and 5202 // override the current action with the host dependence. Otherwise, we 5203 // generate the host dependences and override the action with the device 5204 // dependence. The dependences can't therefore be a top-level action. 5205 OA->doOnEachDependence( 5206 /*IsHostDependence=*/BuildingForOffloadDevice, 5207 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5208 OffloadDependencesInputInfo.append(BuildJobsForAction( 5209 C, DepA, DepTC, DepBoundArch, /*AtTopLevel=*/false, 5210 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, CachedResults, 5211 DepA->getOffloadingDeviceKind())); 5212 }); 5213 5214 A = BuildingForOffloadDevice 5215 ? OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true) 5216 : OA->getHostDependence(); 5217 5218 // We may have already built this action as a part of the offloading 5219 // toolchain, return the cached input if so. 5220 std::pair<const Action *, std::string> ActionTC = { 5221 OA->getHostDependence(), 5222 GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5223 if (CachedResults.find(ActionTC) != CachedResults.end()) { 5224 InputInfoList Inputs = CachedResults[ActionTC]; 5225 Inputs.append(OffloadDependencesInputInfo); 5226 return Inputs; 5227 } 5228 } 5229 5230 if (const InputAction *IA = dyn_cast<InputAction>(A)) { 5231 // FIXME: It would be nice to not claim this here; maybe the old scheme of 5232 // just using Args was better? 5233 const Arg &Input = IA->getInputArg(); 5234 Input.claim(); 5235 if (Input.getOption().matches(options::OPT_INPUT)) { 5236 const char *Name = Input.getValue(); 5237 return {InputInfo(A, Name, /* _BaseInput = */ Name)}; 5238 } 5239 return {InputInfo(A, &Input, /* _BaseInput = */ "")}; 5240 } 5241 5242 if (const BindArchAction *BAA = dyn_cast<BindArchAction>(A)) { 5243 const ToolChain *TC; 5244 StringRef ArchName = BAA->getArchName(); 5245 5246 if (!ArchName.empty()) 5247 TC = &getToolChain(C.getArgs(), 5248 computeTargetTriple(*this, TargetTriple, 5249 C.getArgs(), ArchName)); 5250 else 5251 TC = &C.getDefaultToolChain(); 5252 5253 return BuildJobsForAction(C, *BAA->input_begin(), TC, ArchName, AtTopLevel, 5254 MultipleArchs, LinkingOutput, CachedResults, 5255 TargetDeviceOffloadKind); 5256 } 5257 5258 5259 ActionList Inputs = A->getInputs(); 5260 5261 const JobAction *JA = cast<JobAction>(A); 5262 ActionList CollapsedOffloadActions; 5263 5264 ToolSelector TS(JA, *TC, C, isSaveTempsEnabled(), 5265 embedBitcodeInObject() && !isUsingLTO()); 5266 const Tool *T = TS.getTool(Inputs, CollapsedOffloadActions); 5267 5268 if (!T) 5269 return {InputInfo()}; 5270 5271 if (BuildingForOffloadDevice && 5272 A->getOffloadingDeviceKind() == Action::OFK_OpenMP) { 5273 if (TC->getTriple().isAMDGCN()) { 5274 // AMDGCN treats backend and assemble actions as no-op because 5275 // linker does not support object files. 5276 if (const BackendJobAction *BA = dyn_cast<BackendJobAction>(A)) { 5277 return BuildJobsForAction(C, *BA->input_begin(), TC, BoundArch, 5278 AtTopLevel, MultipleArchs, LinkingOutput, 5279 CachedResults, TargetDeviceOffloadKind); 5280 } 5281 5282 if (const AssembleJobAction *AA = dyn_cast<AssembleJobAction>(A)) { 5283 return BuildJobsForAction(C, *AA->input_begin(), TC, BoundArch, 5284 AtTopLevel, MultipleArchs, LinkingOutput, 5285 CachedResults, TargetDeviceOffloadKind); 5286 } 5287 } 5288 } 5289 5290 // If we've collapsed action list that contained OffloadAction we 5291 // need to build jobs for host/device-side inputs it may have held. 5292 for (const auto *OA : CollapsedOffloadActions) 5293 cast<OffloadAction>(OA)->doOnEachDependence( 5294 /*IsHostDependence=*/BuildingForOffloadDevice, 5295 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5296 OffloadDependencesInputInfo.append(BuildJobsForAction( 5297 C, DepA, DepTC, DepBoundArch, /* AtTopLevel */ false, 5298 /*MultipleArchs=*/!!DepBoundArch, LinkingOutput, CachedResults, 5299 DepA->getOffloadingDeviceKind())); 5300 }); 5301 5302 // Only use pipes when there is exactly one input. 5303 InputInfoList InputInfos; 5304 for (const Action *Input : Inputs) { 5305 // Treat dsymutil and verify sub-jobs as being at the top-level too, they 5306 // shouldn't get temporary output names. 5307 // FIXME: Clean this up. 5308 bool SubJobAtTopLevel = 5309 AtTopLevel && (isa<DsymutilJobAction>(A) || isa<VerifyJobAction>(A)); 5310 InputInfos.append(BuildJobsForAction( 5311 C, Input, TC, BoundArch, SubJobAtTopLevel, MultipleArchs, LinkingOutput, 5312 CachedResults, A->getOffloadingDeviceKind())); 5313 } 5314 5315 // Always use the first file input as the base input. 5316 const char *BaseInput = InputInfos[0].getBaseInput(); 5317 for (auto &Info : InputInfos) { 5318 if (Info.isFilename()) { 5319 BaseInput = Info.getBaseInput(); 5320 break; 5321 } 5322 } 5323 5324 // ... except dsymutil actions, which use their actual input as the base 5325 // input. 5326 if (JA->getType() == types::TY_dSYM) 5327 BaseInput = InputInfos[0].getFilename(); 5328 5329 // ... and in header module compilations, which use the module name. 5330 if (auto *ModuleJA = dyn_cast<HeaderModulePrecompileJobAction>(JA)) 5331 BaseInput = ModuleJA->getModuleName(); 5332 5333 // Append outputs of offload device jobs to the input list 5334 if (!OffloadDependencesInputInfo.empty()) 5335 InputInfos.append(OffloadDependencesInputInfo.begin(), 5336 OffloadDependencesInputInfo.end()); 5337 5338 // Set the effective triple of the toolchain for the duration of this job. 5339 llvm::Triple EffectiveTriple; 5340 const ToolChain &ToolTC = T->getToolChain(); 5341 const ArgList &Args = 5342 C.getArgsForToolChain(TC, BoundArch, A->getOffloadingDeviceKind()); 5343 if (InputInfos.size() != 1) { 5344 EffectiveTriple = llvm::Triple(ToolTC.ComputeEffectiveClangTriple(Args)); 5345 } else { 5346 // Pass along the input type if it can be unambiguously determined. 5347 EffectiveTriple = llvm::Triple( 5348 ToolTC.ComputeEffectiveClangTriple(Args, InputInfos[0].getType())); 5349 } 5350 RegisterEffectiveTriple TripleRAII(ToolTC, EffectiveTriple); 5351 5352 // Determine the place to write output to, if any. 5353 InputInfo Result; 5354 InputInfoList UnbundlingResults; 5355 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(JA)) { 5356 // If we have an unbundling job, we need to create results for all the 5357 // outputs. We also update the results cache so that other actions using 5358 // this unbundling action can get the right results. 5359 for (auto &UI : UA->getDependentActionsInfo()) { 5360 assert(UI.DependentOffloadKind != Action::OFK_None && 5361 "Unbundling with no offloading??"); 5362 5363 // Unbundling actions are never at the top level. When we generate the 5364 // offloading prefix, we also do that for the host file because the 5365 // unbundling action does not change the type of the output which can 5366 // cause a overwrite. 5367 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5368 UI.DependentOffloadKind, 5369 UI.DependentToolChain->getTriple().normalize(), 5370 /*CreatePrefixForHost=*/true); 5371 auto CurI = InputInfo( 5372 UA, 5373 GetNamedOutputPath(C, *UA, BaseInput, UI.DependentBoundArch, 5374 /*AtTopLevel=*/false, 5375 MultipleArchs || 5376 UI.DependentOffloadKind == Action::OFK_HIP, 5377 OffloadingPrefix), 5378 BaseInput); 5379 // Save the unbundling result. 5380 UnbundlingResults.push_back(CurI); 5381 5382 // Get the unique string identifier for this dependence and cache the 5383 // result. 5384 StringRef Arch; 5385 if (TargetDeviceOffloadKind == Action::OFK_HIP) { 5386 if (UI.DependentOffloadKind == Action::OFK_Host) 5387 Arch = StringRef(); 5388 else 5389 Arch = UI.DependentBoundArch; 5390 } else 5391 Arch = BoundArch; 5392 5393 CachedResults[{A, GetTriplePlusArchString(UI.DependentToolChain, Arch, 5394 UI.DependentOffloadKind)}] = { 5395 CurI}; 5396 } 5397 5398 // Now that we have all the results generated, select the one that should be 5399 // returned for the current depending action. 5400 std::pair<const Action *, std::string> ActionTC = { 5401 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5402 assert(CachedResults.find(ActionTC) != CachedResults.end() && 5403 "Result does not exist??"); 5404 Result = CachedResults[ActionTC].front(); 5405 } else if (JA->getType() == types::TY_Nothing) 5406 Result = {InputInfo(A, BaseInput)}; 5407 else { 5408 // We only have to generate a prefix for the host if this is not a top-level 5409 // action. 5410 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5411 A->getOffloadingDeviceKind(), TC->getTriple().normalize(), 5412 /*CreatePrefixForHost=*/!!A->getOffloadingHostActiveKinds() && 5413 !AtTopLevel); 5414 if (isa<OffloadWrapperJobAction>(JA)) { 5415 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5416 BaseInput = FinalOutput->getValue(); 5417 else 5418 BaseInput = getDefaultImageName(); 5419 BaseInput = 5420 C.getArgs().MakeArgString(std::string(BaseInput) + "-wrapper"); 5421 } 5422 Result = InputInfo(A, GetNamedOutputPath(C, *JA, BaseInput, BoundArch, 5423 AtTopLevel, MultipleArchs, 5424 OffloadingPrefix), 5425 BaseInput); 5426 } 5427 5428 if (CCCPrintBindings && !CCGenDiagnostics) { 5429 llvm::errs() << "# \"" << T->getToolChain().getTripleString() << '"' 5430 << " - \"" << T->getName() << "\", inputs: ["; 5431 for (unsigned i = 0, e = InputInfos.size(); i != e; ++i) { 5432 llvm::errs() << InputInfos[i].getAsString(); 5433 if (i + 1 != e) 5434 llvm::errs() << ", "; 5435 } 5436 if (UnbundlingResults.empty()) 5437 llvm::errs() << "], output: " << Result.getAsString() << "\n"; 5438 else { 5439 llvm::errs() << "], outputs: ["; 5440 for (unsigned i = 0, e = UnbundlingResults.size(); i != e; ++i) { 5441 llvm::errs() << UnbundlingResults[i].getAsString(); 5442 if (i + 1 != e) 5443 llvm::errs() << ", "; 5444 } 5445 llvm::errs() << "] \n"; 5446 } 5447 } else { 5448 if (UnbundlingResults.empty()) 5449 T->ConstructJob( 5450 C, *JA, Result, InputInfos, 5451 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5452 LinkingOutput); 5453 else 5454 T->ConstructJobMultipleOutputs( 5455 C, *JA, UnbundlingResults, InputInfos, 5456 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5457 LinkingOutput); 5458 } 5459 return {Result}; 5460 } 5461 5462 const char *Driver::getDefaultImageName() const { 5463 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 5464 return Target.isOSWindows() ? "a.exe" : "a.out"; 5465 } 5466 5467 /// Create output filename based on ArgValue, which could either be a 5468 /// full filename, filename without extension, or a directory. If ArgValue 5469 /// does not provide a filename, then use BaseName, and use the extension 5470 /// suitable for FileType. 5471 static const char *MakeCLOutputFilename(const ArgList &Args, StringRef ArgValue, 5472 StringRef BaseName, 5473 types::ID FileType) { 5474 SmallString<128> Filename = ArgValue; 5475 5476 if (ArgValue.empty()) { 5477 // If the argument is empty, output to BaseName in the current dir. 5478 Filename = BaseName; 5479 } else if (llvm::sys::path::is_separator(Filename.back())) { 5480 // If the argument is a directory, output to BaseName in that dir. 5481 llvm::sys::path::append(Filename, BaseName); 5482 } 5483 5484 if (!llvm::sys::path::has_extension(ArgValue)) { 5485 // If the argument didn't provide an extension, then set it. 5486 const char *Extension = types::getTypeTempSuffix(FileType, true); 5487 5488 if (FileType == types::TY_Image && 5489 Args.hasArg(options::OPT__SLASH_LD, options::OPT__SLASH_LDd)) { 5490 // The output file is a dll. 5491 Extension = "dll"; 5492 } 5493 5494 llvm::sys::path::replace_extension(Filename, Extension); 5495 } 5496 5497 return Args.MakeArgString(Filename.c_str()); 5498 } 5499 5500 static bool HasPreprocessOutput(const Action &JA) { 5501 if (isa<PreprocessJobAction>(JA)) 5502 return true; 5503 if (isa<OffloadAction>(JA) && isa<PreprocessJobAction>(JA.getInputs()[0])) 5504 return true; 5505 if (isa<OffloadBundlingJobAction>(JA) && 5506 HasPreprocessOutput(*(JA.getInputs()[0]))) 5507 return true; 5508 return false; 5509 } 5510 5511 const char *Driver::GetNamedOutputPath(Compilation &C, const JobAction &JA, 5512 const char *BaseInput, 5513 StringRef OrigBoundArch, bool AtTopLevel, 5514 bool MultipleArchs, 5515 StringRef OffloadingPrefix) const { 5516 std::string BoundArch = OrigBoundArch.str(); 5517 if (is_style_windows(llvm::sys::path::Style::native)) { 5518 // BoundArch may contains ':', which is invalid in file names on Windows, 5519 // therefore replace it with '%'. 5520 std::replace(BoundArch.begin(), BoundArch.end(), ':', '@'); 5521 } 5522 5523 llvm::PrettyStackTraceString CrashInfo("Computing output path"); 5524 // Output to a user requested destination? 5525 if (AtTopLevel && !isa<DsymutilJobAction>(JA) && !isa<VerifyJobAction>(JA)) { 5526 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5527 return C.addResultFile(FinalOutput->getValue(), &JA); 5528 } 5529 5530 // For /P, preprocess to file named after BaseInput. 5531 if (C.getArgs().hasArg(options::OPT__SLASH_P)) { 5532 assert(AtTopLevel && isa<PreprocessJobAction>(JA)); 5533 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5534 StringRef NameArg; 5535 if (Arg *A = C.getArgs().getLastArg(options::OPT__SLASH_Fi)) 5536 NameArg = A->getValue(); 5537 return C.addResultFile( 5538 MakeCLOutputFilename(C.getArgs(), NameArg, BaseName, types::TY_PP_C), 5539 &JA); 5540 } 5541 5542 // Default to writing to stdout? 5543 if (AtTopLevel && !CCGenDiagnostics && HasPreprocessOutput(JA)) { 5544 return "-"; 5545 } 5546 5547 if (JA.getType() == types::TY_ModuleFile && 5548 C.getArgs().getLastArg(options::OPT_module_file_info)) { 5549 return "-"; 5550 } 5551 5552 // Is this the assembly listing for /FA? 5553 if (JA.getType() == types::TY_PP_Asm && 5554 (C.getArgs().hasArg(options::OPT__SLASH_FA) || 5555 C.getArgs().hasArg(options::OPT__SLASH_Fa))) { 5556 // Use /Fa and the input filename to determine the asm file name. 5557 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5558 StringRef FaValue = C.getArgs().getLastArgValue(options::OPT__SLASH_Fa); 5559 return C.addResultFile( 5560 MakeCLOutputFilename(C.getArgs(), FaValue, BaseName, JA.getType()), 5561 &JA); 5562 } 5563 5564 // Output to a temporary file? 5565 if ((!AtTopLevel && !isSaveTempsEnabled() && 5566 !C.getArgs().hasArg(options::OPT__SLASH_Fo)) || 5567 CCGenDiagnostics) { 5568 StringRef Name = llvm::sys::path::filename(BaseInput); 5569 std::pair<StringRef, StringRef> Split = Name.split('.'); 5570 SmallString<128> TmpName; 5571 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5572 Arg *A = C.getArgs().getLastArg(options::OPT_fcrash_diagnostics_dir); 5573 if (CCGenDiagnostics && A) { 5574 SmallString<128> CrashDirectory(A->getValue()); 5575 if (!getVFS().exists(CrashDirectory)) 5576 llvm::sys::fs::create_directories(CrashDirectory); 5577 llvm::sys::path::append(CrashDirectory, Split.first); 5578 const char *Middle = Suffix ? "-%%%%%%." : "-%%%%%%"; 5579 std::error_code EC = llvm::sys::fs::createUniqueFile( 5580 CrashDirectory + Middle + Suffix, TmpName); 5581 if (EC) { 5582 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5583 return ""; 5584 } 5585 } else { 5586 if (MultipleArchs && !BoundArch.empty()) { 5587 TmpName = GetTemporaryDirectory(Split.first); 5588 llvm::sys::path::append(TmpName, 5589 Split.first + "-" + BoundArch + "." + Suffix); 5590 } else { 5591 TmpName = GetTemporaryPath(Split.first, Suffix); 5592 } 5593 } 5594 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5595 } 5596 5597 SmallString<128> BasePath(BaseInput); 5598 SmallString<128> ExternalPath(""); 5599 StringRef BaseName; 5600 5601 // Dsymutil actions should use the full path. 5602 if (isa<DsymutilJobAction>(JA) && C.getArgs().hasArg(options::OPT_dsym_dir)) { 5603 ExternalPath += C.getArgs().getLastArg(options::OPT_dsym_dir)->getValue(); 5604 // We use posix style here because the tests (specifically 5605 // darwin-dsymutil.c) demonstrate that posix style paths are acceptable 5606 // even on Windows and if we don't then the similar test covering this 5607 // fails. 5608 llvm::sys::path::append(ExternalPath, llvm::sys::path::Style::posix, 5609 llvm::sys::path::filename(BasePath)); 5610 BaseName = ExternalPath; 5611 } else if (isa<DsymutilJobAction>(JA) || isa<VerifyJobAction>(JA)) 5612 BaseName = BasePath; 5613 else 5614 BaseName = llvm::sys::path::filename(BasePath); 5615 5616 // Determine what the derived output name should be. 5617 const char *NamedOutput; 5618 5619 if ((JA.getType() == types::TY_Object || JA.getType() == types::TY_LTO_BC) && 5620 C.getArgs().hasArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o)) { 5621 // The /Fo or /o flag decides the object filename. 5622 StringRef Val = 5623 C.getArgs() 5624 .getLastArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o) 5625 ->getValue(); 5626 NamedOutput = 5627 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5628 } else if (JA.getType() == types::TY_Image && 5629 C.getArgs().hasArg(options::OPT__SLASH_Fe, 5630 options::OPT__SLASH_o)) { 5631 // The /Fe or /o flag names the linked file. 5632 StringRef Val = 5633 C.getArgs() 5634 .getLastArg(options::OPT__SLASH_Fe, options::OPT__SLASH_o) 5635 ->getValue(); 5636 NamedOutput = 5637 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Image); 5638 } else if (JA.getType() == types::TY_Image) { 5639 if (IsCLMode()) { 5640 // clang-cl uses BaseName for the executable name. 5641 NamedOutput = 5642 MakeCLOutputFilename(C.getArgs(), "", BaseName, types::TY_Image); 5643 } else { 5644 SmallString<128> Output(getDefaultImageName()); 5645 // HIP image for device compilation with -fno-gpu-rdc is per compilation 5646 // unit. 5647 bool IsHIPNoRDC = JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5648 !C.getArgs().hasFlag(options::OPT_fgpu_rdc, 5649 options::OPT_fno_gpu_rdc, false); 5650 if (IsHIPNoRDC) { 5651 Output = BaseName; 5652 llvm::sys::path::replace_extension(Output, ""); 5653 } 5654 Output += OffloadingPrefix; 5655 if (MultipleArchs && !BoundArch.empty()) { 5656 Output += "-"; 5657 Output.append(BoundArch); 5658 } 5659 if (IsHIPNoRDC) 5660 Output += ".out"; 5661 NamedOutput = C.getArgs().MakeArgString(Output.c_str()); 5662 } 5663 } else if (JA.getType() == types::TY_PCH && IsCLMode()) { 5664 NamedOutput = C.getArgs().MakeArgString(GetClPchPath(C, BaseName)); 5665 } else { 5666 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5667 assert(Suffix && "All types used for output should have a suffix."); 5668 5669 std::string::size_type End = std::string::npos; 5670 if (!types::appendSuffixForType(JA.getType())) 5671 End = BaseName.rfind('.'); 5672 SmallString<128> Suffixed(BaseName.substr(0, End)); 5673 Suffixed += OffloadingPrefix; 5674 if (MultipleArchs && !BoundArch.empty()) { 5675 Suffixed += "-"; 5676 Suffixed.append(BoundArch); 5677 } 5678 // When using both -save-temps and -emit-llvm, use a ".tmp.bc" suffix for 5679 // the unoptimized bitcode so that it does not get overwritten by the ".bc" 5680 // optimized bitcode output. 5681 auto IsHIPRDCInCompilePhase = [](const JobAction &JA, 5682 const llvm::opt::DerivedArgList &Args) { 5683 // The relocatable compilation in HIP implies -emit-llvm. Similarly, use a 5684 // ".tmp.bc" suffix for the unoptimized bitcode (generated in the compile 5685 // phase.) 5686 return isa<CompileJobAction>(JA) && 5687 JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5688 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 5689 false); 5690 }; 5691 if (!AtTopLevel && JA.getType() == types::TY_LLVM_BC && 5692 (C.getArgs().hasArg(options::OPT_emit_llvm) || 5693 IsHIPRDCInCompilePhase(JA, C.getArgs()))) 5694 Suffixed += ".tmp"; 5695 Suffixed += '.'; 5696 Suffixed += Suffix; 5697 NamedOutput = C.getArgs().MakeArgString(Suffixed.c_str()); 5698 } 5699 5700 // Prepend object file path if -save-temps=obj 5701 if (!AtTopLevel && isSaveTempsObj() && C.getArgs().hasArg(options::OPT_o) && 5702 JA.getType() != types::TY_PCH) { 5703 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 5704 SmallString<128> TempPath(FinalOutput->getValue()); 5705 llvm::sys::path::remove_filename(TempPath); 5706 StringRef OutputFileName = llvm::sys::path::filename(NamedOutput); 5707 llvm::sys::path::append(TempPath, OutputFileName); 5708 NamedOutput = C.getArgs().MakeArgString(TempPath.c_str()); 5709 } 5710 5711 // If we're saving temps and the temp file conflicts with the input file, 5712 // then avoid overwriting input file. 5713 if (!AtTopLevel && isSaveTempsEnabled() && NamedOutput == BaseName) { 5714 bool SameFile = false; 5715 SmallString<256> Result; 5716 llvm::sys::fs::current_path(Result); 5717 llvm::sys::path::append(Result, BaseName); 5718 llvm::sys::fs::equivalent(BaseInput, Result.c_str(), SameFile); 5719 // Must share the same path to conflict. 5720 if (SameFile) { 5721 StringRef Name = llvm::sys::path::filename(BaseInput); 5722 std::pair<StringRef, StringRef> Split = Name.split('.'); 5723 std::string TmpName = GetTemporaryPath( 5724 Split.first, types::getTypeTempSuffix(JA.getType(), IsCLMode())); 5725 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5726 } 5727 } 5728 5729 // As an annoying special case, PCH generation doesn't strip the pathname. 5730 if (JA.getType() == types::TY_PCH && !IsCLMode()) { 5731 llvm::sys::path::remove_filename(BasePath); 5732 if (BasePath.empty()) 5733 BasePath = NamedOutput; 5734 else 5735 llvm::sys::path::append(BasePath, NamedOutput); 5736 return C.addResultFile(C.getArgs().MakeArgString(BasePath.c_str()), &JA); 5737 } else { 5738 return C.addResultFile(NamedOutput, &JA); 5739 } 5740 } 5741 5742 std::string Driver::GetFilePath(StringRef Name, const ToolChain &TC) const { 5743 // Search for Name in a list of paths. 5744 auto SearchPaths = [&](const llvm::SmallVectorImpl<std::string> &P) 5745 -> llvm::Optional<std::string> { 5746 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5747 // attempting to use this prefix when looking for file paths. 5748 for (const auto &Dir : P) { 5749 if (Dir.empty()) 5750 continue; 5751 SmallString<128> P(Dir[0] == '=' ? SysRoot + Dir.substr(1) : Dir); 5752 llvm::sys::path::append(P, Name); 5753 if (llvm::sys::fs::exists(Twine(P))) 5754 return std::string(P); 5755 } 5756 return None; 5757 }; 5758 5759 if (auto P = SearchPaths(PrefixDirs)) 5760 return *P; 5761 5762 SmallString<128> R(ResourceDir); 5763 llvm::sys::path::append(R, Name); 5764 if (llvm::sys::fs::exists(Twine(R))) 5765 return std::string(R.str()); 5766 5767 SmallString<128> P(TC.getCompilerRTPath()); 5768 llvm::sys::path::append(P, Name); 5769 if (llvm::sys::fs::exists(Twine(P))) 5770 return std::string(P.str()); 5771 5772 SmallString<128> D(Dir); 5773 llvm::sys::path::append(D, "..", Name); 5774 if (llvm::sys::fs::exists(Twine(D))) 5775 return std::string(D.str()); 5776 5777 if (auto P = SearchPaths(TC.getLibraryPaths())) 5778 return *P; 5779 5780 if (auto P = SearchPaths(TC.getFilePaths())) 5781 return *P; 5782 5783 return std::string(Name); 5784 } 5785 5786 void Driver::generatePrefixedToolNames( 5787 StringRef Tool, const ToolChain &TC, 5788 SmallVectorImpl<std::string> &Names) const { 5789 // FIXME: Needs a better variable than TargetTriple 5790 Names.emplace_back((TargetTriple + "-" + Tool).str()); 5791 Names.emplace_back(Tool); 5792 } 5793 5794 static bool ScanDirForExecutable(SmallString<128> &Dir, StringRef Name) { 5795 llvm::sys::path::append(Dir, Name); 5796 if (llvm::sys::fs::can_execute(Twine(Dir))) 5797 return true; 5798 llvm::sys::path::remove_filename(Dir); 5799 return false; 5800 } 5801 5802 std::string Driver::GetProgramPath(StringRef Name, const ToolChain &TC) const { 5803 SmallVector<std::string, 2> TargetSpecificExecutables; 5804 generatePrefixedToolNames(Name, TC, TargetSpecificExecutables); 5805 5806 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5807 // attempting to use this prefix when looking for program paths. 5808 for (const auto &PrefixDir : PrefixDirs) { 5809 if (llvm::sys::fs::is_directory(PrefixDir)) { 5810 SmallString<128> P(PrefixDir); 5811 if (ScanDirForExecutable(P, Name)) 5812 return std::string(P.str()); 5813 } else { 5814 SmallString<128> P((PrefixDir + Name).str()); 5815 if (llvm::sys::fs::can_execute(Twine(P))) 5816 return std::string(P.str()); 5817 } 5818 } 5819 5820 const ToolChain::path_list &List = TC.getProgramPaths(); 5821 for (const auto &TargetSpecificExecutable : TargetSpecificExecutables) { 5822 // For each possible name of the tool look for it in 5823 // program paths first, then the path. 5824 // Higher priority names will be first, meaning that 5825 // a higher priority name in the path will be found 5826 // instead of a lower priority name in the program path. 5827 // E.g. <triple>-gcc on the path will be found instead 5828 // of gcc in the program path 5829 for (const auto &Path : List) { 5830 SmallString<128> P(Path); 5831 if (ScanDirForExecutable(P, TargetSpecificExecutable)) 5832 return std::string(P.str()); 5833 } 5834 5835 // Fall back to the path 5836 if (llvm::ErrorOr<std::string> P = 5837 llvm::sys::findProgramByName(TargetSpecificExecutable)) 5838 return *P; 5839 } 5840 5841 return std::string(Name); 5842 } 5843 5844 std::string Driver::GetTemporaryPath(StringRef Prefix, StringRef Suffix) const { 5845 SmallString<128> Path; 5846 std::error_code EC = llvm::sys::fs::createTemporaryFile(Prefix, Suffix, Path); 5847 if (EC) { 5848 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5849 return ""; 5850 } 5851 5852 return std::string(Path.str()); 5853 } 5854 5855 std::string Driver::GetTemporaryDirectory(StringRef Prefix) const { 5856 SmallString<128> Path; 5857 std::error_code EC = llvm::sys::fs::createUniqueDirectory(Prefix, Path); 5858 if (EC) { 5859 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5860 return ""; 5861 } 5862 5863 return std::string(Path.str()); 5864 } 5865 5866 std::string Driver::GetClPchPath(Compilation &C, StringRef BaseName) const { 5867 SmallString<128> Output; 5868 if (Arg *FpArg = C.getArgs().getLastArg(options::OPT__SLASH_Fp)) { 5869 // FIXME: If anybody needs it, implement this obscure rule: 5870 // "If you specify a directory without a file name, the default file name 5871 // is VCx0.pch., where x is the major version of Visual C++ in use." 5872 Output = FpArg->getValue(); 5873 5874 // "If you do not specify an extension as part of the path name, an 5875 // extension of .pch is assumed. " 5876 if (!llvm::sys::path::has_extension(Output)) 5877 Output += ".pch"; 5878 } else { 5879 if (Arg *YcArg = C.getArgs().getLastArg(options::OPT__SLASH_Yc)) 5880 Output = YcArg->getValue(); 5881 if (Output.empty()) 5882 Output = BaseName; 5883 llvm::sys::path::replace_extension(Output, ".pch"); 5884 } 5885 return std::string(Output.str()); 5886 } 5887 5888 const ToolChain &Driver::getToolChain(const ArgList &Args, 5889 const llvm::Triple &Target) const { 5890 5891 auto &TC = ToolChains[Target.str()]; 5892 if (!TC) { 5893 switch (Target.getOS()) { 5894 case llvm::Triple::AIX: 5895 TC = std::make_unique<toolchains::AIX>(*this, Target, Args); 5896 break; 5897 case llvm::Triple::Haiku: 5898 TC = std::make_unique<toolchains::Haiku>(*this, Target, Args); 5899 break; 5900 case llvm::Triple::Ananas: 5901 TC = std::make_unique<toolchains::Ananas>(*this, Target, Args); 5902 break; 5903 case llvm::Triple::CloudABI: 5904 TC = std::make_unique<toolchains::CloudABI>(*this, Target, Args); 5905 break; 5906 case llvm::Triple::Darwin: 5907 case llvm::Triple::MacOSX: 5908 case llvm::Triple::IOS: 5909 case llvm::Triple::TvOS: 5910 case llvm::Triple::WatchOS: 5911 case llvm::Triple::DriverKit: 5912 TC = std::make_unique<toolchains::DarwinClang>(*this, Target, Args); 5913 break; 5914 case llvm::Triple::DragonFly: 5915 TC = std::make_unique<toolchains::DragonFly>(*this, Target, Args); 5916 break; 5917 case llvm::Triple::OpenBSD: 5918 TC = std::make_unique<toolchains::OpenBSD>(*this, Target, Args); 5919 break; 5920 case llvm::Triple::NetBSD: 5921 TC = std::make_unique<toolchains::NetBSD>(*this, Target, Args); 5922 break; 5923 case llvm::Triple::FreeBSD: 5924 if (Target.isPPC()) 5925 TC = std::make_unique<toolchains::PPCFreeBSDToolChain>(*this, Target, 5926 Args); 5927 else 5928 TC = std::make_unique<toolchains::FreeBSD>(*this, Target, Args); 5929 break; 5930 case llvm::Triple::Minix: 5931 TC = std::make_unique<toolchains::Minix>(*this, Target, Args); 5932 break; 5933 case llvm::Triple::Linux: 5934 case llvm::Triple::ELFIAMCU: 5935 if (Target.getArch() == llvm::Triple::hexagon) 5936 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 5937 Args); 5938 else if ((Target.getVendor() == llvm::Triple::MipsTechnologies) && 5939 !Target.hasEnvironment()) 5940 TC = std::make_unique<toolchains::MipsLLVMToolChain>(*this, Target, 5941 Args); 5942 else if (Target.isPPC()) 5943 TC = std::make_unique<toolchains::PPCLinuxToolChain>(*this, Target, 5944 Args); 5945 else if (Target.getArch() == llvm::Triple::ve) 5946 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 5947 5948 else 5949 TC = std::make_unique<toolchains::Linux>(*this, Target, Args); 5950 break; 5951 case llvm::Triple::NaCl: 5952 TC = std::make_unique<toolchains::NaClToolChain>(*this, Target, Args); 5953 break; 5954 case llvm::Triple::Fuchsia: 5955 TC = std::make_unique<toolchains::Fuchsia>(*this, Target, Args); 5956 break; 5957 case llvm::Triple::Solaris: 5958 TC = std::make_unique<toolchains::Solaris>(*this, Target, Args); 5959 break; 5960 case llvm::Triple::AMDHSA: 5961 TC = std::make_unique<toolchains::ROCMToolChain>(*this, Target, Args); 5962 break; 5963 case llvm::Triple::AMDPAL: 5964 case llvm::Triple::Mesa3D: 5965 TC = std::make_unique<toolchains::AMDGPUToolChain>(*this, Target, Args); 5966 break; 5967 case llvm::Triple::Win32: 5968 switch (Target.getEnvironment()) { 5969 default: 5970 if (Target.isOSBinFormatELF()) 5971 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 5972 else if (Target.isOSBinFormatMachO()) 5973 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 5974 else 5975 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 5976 break; 5977 case llvm::Triple::GNU: 5978 TC = std::make_unique<toolchains::MinGW>(*this, Target, Args); 5979 break; 5980 case llvm::Triple::Itanium: 5981 TC = std::make_unique<toolchains::CrossWindowsToolChain>(*this, Target, 5982 Args); 5983 break; 5984 case llvm::Triple::MSVC: 5985 case llvm::Triple::UnknownEnvironment: 5986 if (Args.getLastArgValue(options::OPT_fuse_ld_EQ) 5987 .startswith_insensitive("bfd")) 5988 TC = std::make_unique<toolchains::CrossWindowsToolChain>( 5989 *this, Target, Args); 5990 else 5991 TC = 5992 std::make_unique<toolchains::MSVCToolChain>(*this, Target, Args); 5993 break; 5994 } 5995 break; 5996 case llvm::Triple::PS4: 5997 TC = std::make_unique<toolchains::PS4CPU>(*this, Target, Args); 5998 break; 5999 case llvm::Triple::PS5: 6000 TC = std::make_unique<toolchains::PS5CPU>(*this, Target, Args); 6001 break; 6002 case llvm::Triple::Contiki: 6003 TC = std::make_unique<toolchains::Contiki>(*this, Target, Args); 6004 break; 6005 case llvm::Triple::Hurd: 6006 TC = std::make_unique<toolchains::Hurd>(*this, Target, Args); 6007 break; 6008 case llvm::Triple::ZOS: 6009 TC = std::make_unique<toolchains::ZOS>(*this, Target, Args); 6010 break; 6011 case llvm::Triple::ShaderModel: 6012 TC = std::make_unique<toolchains::HLSLToolChain>(*this, Target, Args); 6013 break; 6014 default: 6015 // Of these targets, Hexagon is the only one that might have 6016 // an OS of Linux, in which case it got handled above already. 6017 switch (Target.getArch()) { 6018 case llvm::Triple::tce: 6019 TC = std::make_unique<toolchains::TCEToolChain>(*this, Target, Args); 6020 break; 6021 case llvm::Triple::tcele: 6022 TC = std::make_unique<toolchains::TCELEToolChain>(*this, Target, Args); 6023 break; 6024 case llvm::Triple::hexagon: 6025 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 6026 Args); 6027 break; 6028 case llvm::Triple::lanai: 6029 TC = std::make_unique<toolchains::LanaiToolChain>(*this, Target, Args); 6030 break; 6031 case llvm::Triple::xcore: 6032 TC = std::make_unique<toolchains::XCoreToolChain>(*this, Target, Args); 6033 break; 6034 case llvm::Triple::wasm32: 6035 case llvm::Triple::wasm64: 6036 TC = std::make_unique<toolchains::WebAssembly>(*this, Target, Args); 6037 break; 6038 case llvm::Triple::avr: 6039 TC = std::make_unique<toolchains::AVRToolChain>(*this, Target, Args); 6040 break; 6041 case llvm::Triple::msp430: 6042 TC = 6043 std::make_unique<toolchains::MSP430ToolChain>(*this, Target, Args); 6044 break; 6045 case llvm::Triple::riscv32: 6046 case llvm::Triple::riscv64: 6047 if (toolchains::RISCVToolChain::hasGCCToolchain(*this, Args)) 6048 TC = 6049 std::make_unique<toolchains::RISCVToolChain>(*this, Target, Args); 6050 else 6051 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6052 break; 6053 case llvm::Triple::ve: 6054 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6055 break; 6056 case llvm::Triple::spirv32: 6057 case llvm::Triple::spirv64: 6058 TC = std::make_unique<toolchains::SPIRVToolChain>(*this, Target, Args); 6059 break; 6060 case llvm::Triple::csky: 6061 TC = std::make_unique<toolchains::CSKYToolChain>(*this, Target, Args); 6062 break; 6063 default: 6064 if (Target.getVendor() == llvm::Triple::Myriad) 6065 TC = std::make_unique<toolchains::MyriadToolChain>(*this, Target, 6066 Args); 6067 else if (toolchains::BareMetal::handlesTarget(Target)) 6068 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6069 else if (Target.isOSBinFormatELF()) 6070 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6071 else if (Target.isOSBinFormatMachO()) 6072 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6073 else 6074 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6075 } 6076 } 6077 } 6078 6079 // Intentionally omitted from the switch above: llvm::Triple::CUDA. CUDA 6080 // compiles always need two toolchains, the CUDA toolchain and the host 6081 // toolchain. So the only valid way to create a CUDA toolchain is via 6082 // CreateOffloadingDeviceToolChains. 6083 6084 return *TC; 6085 } 6086 6087 const ToolChain &Driver::getOffloadingDeviceToolChain( 6088 const ArgList &Args, const llvm::Triple &Target, const ToolChain &HostTC, 6089 const Action::OffloadKind &TargetDeviceOffloadKind) const { 6090 // Use device / host triples as the key into the ToolChains map because the 6091 // device ToolChain we create depends on both. 6092 auto &TC = ToolChains[Target.str() + "/" + HostTC.getTriple().str()]; 6093 if (!TC) { 6094 // Categorized by offload kind > arch rather than OS > arch like 6095 // the normal getToolChain call, as it seems a reasonable way to categorize 6096 // things. 6097 switch (TargetDeviceOffloadKind) { 6098 case Action::OFK_HIP: { 6099 if (Target.getArch() == llvm::Triple::amdgcn && 6100 Target.getVendor() == llvm::Triple::AMD && 6101 Target.getOS() == llvm::Triple::AMDHSA) 6102 TC = std::make_unique<toolchains::HIPAMDToolChain>(*this, Target, 6103 HostTC, Args); 6104 else if (Target.getArch() == llvm::Triple::spirv64 && 6105 Target.getVendor() == llvm::Triple::UnknownVendor && 6106 Target.getOS() == llvm::Triple::UnknownOS) 6107 TC = std::make_unique<toolchains::HIPSPVToolChain>(*this, Target, 6108 HostTC, Args); 6109 break; 6110 } 6111 default: 6112 break; 6113 } 6114 } 6115 6116 return *TC; 6117 } 6118 6119 bool Driver::ShouldUseClangCompiler(const JobAction &JA) const { 6120 // Say "no" if there is not exactly one input of a type clang understands. 6121 if (JA.size() != 1 || 6122 !types::isAcceptedByClang((*JA.input_begin())->getType())) 6123 return false; 6124 6125 // And say "no" if this is not a kind of action clang understands. 6126 if (!isa<PreprocessJobAction>(JA) && !isa<PrecompileJobAction>(JA) && 6127 !isa<CompileJobAction>(JA) && !isa<BackendJobAction>(JA) && 6128 !isa<ExtractAPIJobAction>(JA)) 6129 return false; 6130 6131 return true; 6132 } 6133 6134 bool Driver::ShouldUseFlangCompiler(const JobAction &JA) const { 6135 // Say "no" if there is not exactly one input of a type flang understands. 6136 if (JA.size() != 1 || 6137 !types::isAcceptedByFlang((*JA.input_begin())->getType())) 6138 return false; 6139 6140 // And say "no" if this is not a kind of action flang understands. 6141 if (!isa<PreprocessJobAction>(JA) && !isa<CompileJobAction>(JA) && 6142 !isa<BackendJobAction>(JA)) 6143 return false; 6144 6145 return true; 6146 } 6147 6148 bool Driver::ShouldEmitStaticLibrary(const ArgList &Args) const { 6149 // Only emit static library if the flag is set explicitly. 6150 if (Args.hasArg(options::OPT_emit_static_lib)) 6151 return true; 6152 return false; 6153 } 6154 6155 /// GetReleaseVersion - Parse (([0-9]+)(.([0-9]+)(.([0-9]+)?))?)? and return the 6156 /// grouped values as integers. Numbers which are not provided are set to 0. 6157 /// 6158 /// \return True if the entire string was parsed (9.2), or all groups were 6159 /// parsed (10.3.5extrastuff). 6160 bool Driver::GetReleaseVersion(StringRef Str, unsigned &Major, unsigned &Minor, 6161 unsigned &Micro, bool &HadExtra) { 6162 HadExtra = false; 6163 6164 Major = Minor = Micro = 0; 6165 if (Str.empty()) 6166 return false; 6167 6168 if (Str.consumeInteger(10, Major)) 6169 return false; 6170 if (Str.empty()) 6171 return true; 6172 if (Str[0] != '.') 6173 return false; 6174 6175 Str = Str.drop_front(1); 6176 6177 if (Str.consumeInteger(10, Minor)) 6178 return false; 6179 if (Str.empty()) 6180 return true; 6181 if (Str[0] != '.') 6182 return false; 6183 Str = Str.drop_front(1); 6184 6185 if (Str.consumeInteger(10, Micro)) 6186 return false; 6187 if (!Str.empty()) 6188 HadExtra = true; 6189 return true; 6190 } 6191 6192 /// Parse digits from a string \p Str and fulfill \p Digits with 6193 /// the parsed numbers. This method assumes that the max number of 6194 /// digits to look for is equal to Digits.size(). 6195 /// 6196 /// \return True if the entire string was parsed and there are 6197 /// no extra characters remaining at the end. 6198 bool Driver::GetReleaseVersion(StringRef Str, 6199 MutableArrayRef<unsigned> Digits) { 6200 if (Str.empty()) 6201 return false; 6202 6203 unsigned CurDigit = 0; 6204 while (CurDigit < Digits.size()) { 6205 unsigned Digit; 6206 if (Str.consumeInteger(10, Digit)) 6207 return false; 6208 Digits[CurDigit] = Digit; 6209 if (Str.empty()) 6210 return true; 6211 if (Str[0] != '.') 6212 return false; 6213 Str = Str.drop_front(1); 6214 CurDigit++; 6215 } 6216 6217 // More digits than requested, bail out... 6218 return false; 6219 } 6220 6221 std::pair<unsigned, unsigned> 6222 Driver::getIncludeExcludeOptionFlagMasks(bool IsClCompatMode) const { 6223 unsigned IncludedFlagsBitmask = 0; 6224 unsigned ExcludedFlagsBitmask = options::NoDriverOption; 6225 6226 if (IsClCompatMode) { 6227 // Include CL and Core options. 6228 IncludedFlagsBitmask |= options::CLOption; 6229 IncludedFlagsBitmask |= options::CoreOption; 6230 } else { 6231 ExcludedFlagsBitmask |= options::CLOption; 6232 } 6233 if (IsDXCMode()) { 6234 // Include DXC and Core options. 6235 IncludedFlagsBitmask |= options::DXCOption; 6236 IncludedFlagsBitmask |= options::CoreOption; 6237 } else { 6238 ExcludedFlagsBitmask |= options::DXCOption; 6239 } 6240 return std::make_pair(IncludedFlagsBitmask, ExcludedFlagsBitmask); 6241 } 6242 6243 bool clang::driver::isOptimizationLevelFast(const ArgList &Args) { 6244 return Args.hasFlag(options::OPT_Ofast, options::OPT_O_Group, false); 6245 } 6246 6247 bool clang::driver::willEmitRemarks(const ArgList &Args) { 6248 // -fsave-optimization-record enables it. 6249 if (Args.hasFlag(options::OPT_fsave_optimization_record, 6250 options::OPT_fno_save_optimization_record, false)) 6251 return true; 6252 6253 // -fsave-optimization-record=<format> enables it as well. 6254 if (Args.hasFlag(options::OPT_fsave_optimization_record_EQ, 6255 options::OPT_fno_save_optimization_record, false)) 6256 return true; 6257 6258 // -foptimization-record-file alone enables it too. 6259 if (Args.hasFlag(options::OPT_foptimization_record_file_EQ, 6260 options::OPT_fno_save_optimization_record, false)) 6261 return true; 6262 6263 // -foptimization-record-passes alone enables it too. 6264 if (Args.hasFlag(options::OPT_foptimization_record_passes_EQ, 6265 options::OPT_fno_save_optimization_record, false)) 6266 return true; 6267 return false; 6268 } 6269 6270 llvm::StringRef clang::driver::getDriverMode(StringRef ProgName, 6271 ArrayRef<const char *> Args) { 6272 static const std::string OptName = 6273 getDriverOptTable().getOption(options::OPT_driver_mode).getPrefixedName(); 6274 llvm::StringRef Opt; 6275 for (StringRef Arg : Args) { 6276 if (!Arg.startswith(OptName)) 6277 continue; 6278 Opt = Arg; 6279 } 6280 if (Opt.empty()) 6281 Opt = ToolChain::getTargetAndModeFromProgramName(ProgName).DriverMode; 6282 return Opt.consume_front(OptName) ? Opt : ""; 6283 } 6284 6285 bool driver::IsClangCL(StringRef DriverMode) { return DriverMode.equals("cl"); } 6286