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 <cstdlib> // ::getenv 98 #include <map> 99 #include <memory> 100 #include <optional> 101 #include <utility> 102 #if LLVM_ON_UNIX 103 #include <unistd.h> // getpid 104 #endif 105 106 using namespace clang::driver; 107 using namespace clang; 108 using namespace llvm::opt; 109 110 static std::optional<llvm::Triple> getOffloadTargetTriple(const Driver &D, 111 const ArgList &Args) { 112 auto OffloadTargets = Args.getAllArgValues(options::OPT_offload_EQ); 113 // Offload compilation flow does not support multiple targets for now. We 114 // need the HIPActionBuilder (and possibly the CudaActionBuilder{,Base}too) 115 // to support multiple tool chains first. 116 switch (OffloadTargets.size()) { 117 default: 118 D.Diag(diag::err_drv_only_one_offload_target_supported); 119 return std::nullopt; 120 case 0: 121 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << ""; 122 return std::nullopt; 123 case 1: 124 break; 125 } 126 return llvm::Triple(OffloadTargets[0]); 127 } 128 129 static std::optional<llvm::Triple> 130 getNVIDIAOffloadTargetTriple(const Driver &D, const ArgList &Args, 131 const llvm::Triple &HostTriple) { 132 if (!Args.hasArg(options::OPT_offload_EQ)) { 133 return llvm::Triple(HostTriple.isArch64Bit() ? "nvptx64-nvidia-cuda" 134 : "nvptx-nvidia-cuda"); 135 } 136 auto TT = getOffloadTargetTriple(D, Args); 137 if (TT && (TT->getArch() == llvm::Triple::spirv32 || 138 TT->getArch() == llvm::Triple::spirv64)) { 139 if (Args.hasArg(options::OPT_emit_llvm)) 140 return TT; 141 D.Diag(diag::err_drv_cuda_offload_only_emit_bc); 142 return std::nullopt; 143 } 144 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 145 return std::nullopt; 146 } 147 static std::optional<llvm::Triple> 148 getHIPOffloadTargetTriple(const Driver &D, const ArgList &Args) { 149 if (!Args.hasArg(options::OPT_offload_EQ)) { 150 return llvm::Triple("amdgcn-amd-amdhsa"); // Default HIP triple. 151 } 152 auto TT = getOffloadTargetTriple(D, Args); 153 if (!TT) 154 return std::nullopt; 155 if (TT->getArch() == llvm::Triple::amdgcn && 156 TT->getVendor() == llvm::Triple::AMD && 157 TT->getOS() == llvm::Triple::AMDHSA) 158 return TT; 159 if (TT->getArch() == llvm::Triple::spirv64) 160 return TT; 161 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 162 return std::nullopt; 163 } 164 165 // static 166 std::string Driver::GetResourcesPath(StringRef BinaryPath, 167 StringRef CustomResourceDir) { 168 // Since the resource directory is embedded in the module hash, it's important 169 // that all places that need it call this function, so that they get the 170 // exact same string ("a/../b/" and "b/" get different hashes, for example). 171 172 // Dir is bin/ or lib/, depending on where BinaryPath is. 173 std::string Dir = std::string(llvm::sys::path::parent_path(BinaryPath)); 174 175 SmallString<128> P(Dir); 176 if (CustomResourceDir != "") { 177 llvm::sys::path::append(P, CustomResourceDir); 178 } else { 179 // On Windows, libclang.dll is in bin/. 180 // On non-Windows, libclang.so/.dylib is in lib/. 181 // With a static-library build of libclang, LibClangPath will contain the 182 // path of the embedding binary, which for LLVM binaries will be in bin/. 183 // ../lib gets us to lib/ in both cases. 184 P = llvm::sys::path::parent_path(Dir); 185 llvm::sys::path::append(P, CLANG_INSTALL_LIBDIR_BASENAME, "clang", 186 CLANG_VERSION_MAJOR_STRING); 187 } 188 189 return std::string(P.str()); 190 } 191 192 Driver::Driver(StringRef ClangExecutable, StringRef TargetTriple, 193 DiagnosticsEngine &Diags, std::string Title, 194 IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS) 195 : Diags(Diags), VFS(std::move(VFS)), Mode(GCCMode), 196 SaveTemps(SaveTempsNone), BitcodeEmbed(EmbedNone), 197 Offload(OffloadHostDevice), CXX20HeaderType(HeaderMode_None), 198 ModulesModeCXX20(false), LTOMode(LTOK_None), 199 ClangExecutable(ClangExecutable), SysRoot(DEFAULT_SYSROOT), 200 DriverTitle(Title), CCCPrintBindings(false), CCPrintOptions(false), 201 CCLogDiagnostics(false), CCGenDiagnostics(false), 202 CCPrintProcessStats(false), TargetTriple(TargetTriple), Saver(Alloc), 203 CheckInputsExist(true), ProbePrecompiled(true), 204 SuppressMissingInputWarning(false) { 205 // Provide a sane fallback if no VFS is specified. 206 if (!this->VFS) 207 this->VFS = llvm::vfs::getRealFileSystem(); 208 209 Name = std::string(llvm::sys::path::filename(ClangExecutable)); 210 Dir = std::string(llvm::sys::path::parent_path(ClangExecutable)); 211 InstalledDir = Dir; // Provide a sensible default installed dir. 212 213 if ((!SysRoot.empty()) && llvm::sys::path::is_relative(SysRoot)) { 214 // Prepend InstalledDir if SysRoot is relative 215 SmallString<128> P(InstalledDir); 216 llvm::sys::path::append(P, SysRoot); 217 SysRoot = std::string(P); 218 } 219 220 #if defined(CLANG_CONFIG_FILE_SYSTEM_DIR) 221 SystemConfigDir = CLANG_CONFIG_FILE_SYSTEM_DIR; 222 #endif 223 #if defined(CLANG_CONFIG_FILE_USER_DIR) 224 { 225 SmallString<128> P; 226 llvm::sys::fs::expand_tilde(CLANG_CONFIG_FILE_USER_DIR, P); 227 UserConfigDir = static_cast<std::string>(P); 228 } 229 #endif 230 231 // Compute the path to the resource directory. 232 ResourceDir = GetResourcesPath(ClangExecutable, CLANG_RESOURCE_DIR); 233 } 234 235 void Driver::setDriverMode(StringRef Value) { 236 static const std::string OptName = 237 getOpts().getOption(options::OPT_driver_mode).getPrefixedName(); 238 if (auto M = llvm::StringSwitch<std::optional<DriverMode>>(Value) 239 .Case("gcc", GCCMode) 240 .Case("g++", GXXMode) 241 .Case("cpp", CPPMode) 242 .Case("cl", CLMode) 243 .Case("flang", FlangMode) 244 .Case("dxc", DXCMode) 245 .Default(std::nullopt)) 246 Mode = *M; 247 else 248 Diag(diag::err_drv_unsupported_option_argument) << OptName << Value; 249 } 250 251 InputArgList Driver::ParseArgStrings(ArrayRef<const char *> ArgStrings, 252 bool IsClCompatMode, 253 bool &ContainsError) { 254 llvm::PrettyStackTraceString CrashInfo("Command line argument parsing"); 255 ContainsError = false; 256 257 unsigned IncludedFlagsBitmask; 258 unsigned ExcludedFlagsBitmask; 259 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 260 getIncludeExcludeOptionFlagMasks(IsClCompatMode); 261 262 // Make sure that Flang-only options don't pollute the Clang output 263 // TODO: Make sure that Clang-only options don't pollute Flang output 264 if (!IsFlangMode()) 265 ExcludedFlagsBitmask |= options::FlangOnlyOption; 266 267 unsigned MissingArgIndex, MissingArgCount; 268 InputArgList Args = 269 getOpts().ParseArgs(ArgStrings, MissingArgIndex, MissingArgCount, 270 IncludedFlagsBitmask, ExcludedFlagsBitmask); 271 272 // Check for missing argument error. 273 if (MissingArgCount) { 274 Diag(diag::err_drv_missing_argument) 275 << Args.getArgString(MissingArgIndex) << MissingArgCount; 276 ContainsError |= 277 Diags.getDiagnosticLevel(diag::err_drv_missing_argument, 278 SourceLocation()) > DiagnosticsEngine::Warning; 279 } 280 281 // Check for unsupported options. 282 for (const Arg *A : Args) { 283 if (A->getOption().hasFlag(options::Unsupported)) { 284 unsigned DiagID; 285 auto ArgString = A->getAsString(Args); 286 std::string Nearest; 287 if (getOpts().findNearest( 288 ArgString, Nearest, IncludedFlagsBitmask, 289 ExcludedFlagsBitmask | options::Unsupported) > 1) { 290 DiagID = diag::err_drv_unsupported_opt; 291 Diag(DiagID) << ArgString; 292 } else { 293 DiagID = diag::err_drv_unsupported_opt_with_suggestion; 294 Diag(DiagID) << ArgString << Nearest; 295 } 296 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 297 DiagnosticsEngine::Warning; 298 continue; 299 } 300 301 // Warn about -mcpu= without an argument. 302 if (A->getOption().matches(options::OPT_mcpu_EQ) && A->containsValue("")) { 303 Diag(diag::warn_drv_empty_joined_argument) << A->getAsString(Args); 304 ContainsError |= Diags.getDiagnosticLevel( 305 diag::warn_drv_empty_joined_argument, 306 SourceLocation()) > DiagnosticsEngine::Warning; 307 } 308 } 309 310 for (const Arg *A : Args.filtered(options::OPT_UNKNOWN)) { 311 unsigned DiagID; 312 auto ArgString = A->getAsString(Args); 313 std::string Nearest; 314 if (getOpts().findNearest(ArgString, Nearest, IncludedFlagsBitmask, 315 ExcludedFlagsBitmask) > 1) { 316 if (!IsCLMode() && 317 getOpts().findExact(ArgString, Nearest, options::CC1Option)) { 318 DiagID = diag::err_drv_unknown_argument_with_suggestion; 319 Diags.Report(DiagID) << ArgString << "-Xclang " + Nearest; 320 } else { 321 DiagID = IsCLMode() ? diag::warn_drv_unknown_argument_clang_cl 322 : diag::err_drv_unknown_argument; 323 Diags.Report(DiagID) << ArgString; 324 } 325 } else { 326 DiagID = IsCLMode() 327 ? diag::warn_drv_unknown_argument_clang_cl_with_suggestion 328 : diag::err_drv_unknown_argument_with_suggestion; 329 Diags.Report(DiagID) << ArgString << Nearest; 330 } 331 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 332 DiagnosticsEngine::Warning; 333 } 334 335 for (const Arg *A : Args.filtered(options::OPT_o)) { 336 if (ArgStrings[A->getIndex()] == A->getSpelling()) 337 continue; 338 339 // Warn on joined arguments that are similar to a long argument. 340 std::string ArgString = ArgStrings[A->getIndex()]; 341 std::string Nearest; 342 if (getOpts().findExact("-" + ArgString, Nearest, IncludedFlagsBitmask, 343 ExcludedFlagsBitmask)) 344 Diags.Report(diag::warn_drv_potentially_misspelled_joined_argument) 345 << A->getAsString(Args) << Nearest; 346 } 347 348 return Args; 349 } 350 351 // Determine which compilation mode we are in. We look for options which 352 // affect the phase, starting with the earliest phases, and record which 353 // option we used to determine the final phase. 354 phases::ID Driver::getFinalPhase(const DerivedArgList &DAL, 355 Arg **FinalPhaseArg) const { 356 Arg *PhaseArg = nullptr; 357 phases::ID FinalPhase; 358 359 // -{E,EP,P,M,MM} only run the preprocessor. 360 if (CCCIsCPP() || (PhaseArg = DAL.getLastArg(options::OPT_E)) || 361 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_EP)) || 362 (PhaseArg = DAL.getLastArg(options::OPT_M, options::OPT_MM)) || 363 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_P)) || 364 CCGenDiagnostics) { 365 FinalPhase = phases::Preprocess; 366 367 // --precompile only runs up to precompilation. 368 // Options that cause the output of C++20 compiled module interfaces or 369 // header units have the same effect. 370 } else if ((PhaseArg = DAL.getLastArg(options::OPT__precompile)) || 371 (PhaseArg = DAL.getLastArg(options::OPT_extract_api)) || 372 (PhaseArg = DAL.getLastArg(options::OPT_fmodule_header, 373 options::OPT_fmodule_header_EQ))) { 374 FinalPhase = phases::Precompile; 375 // -{fsyntax-only,-analyze,emit-ast} only run up to the compiler. 376 } else if ((PhaseArg = DAL.getLastArg(options::OPT_fsyntax_only)) || 377 (PhaseArg = DAL.getLastArg(options::OPT_print_supported_cpus)) || 378 (PhaseArg = DAL.getLastArg(options::OPT_module_file_info)) || 379 (PhaseArg = DAL.getLastArg(options::OPT_verify_pch)) || 380 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_objc)) || 381 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_legacy_objc)) || 382 (PhaseArg = DAL.getLastArg(options::OPT__migrate)) || 383 (PhaseArg = DAL.getLastArg(options::OPT__analyze)) || 384 (PhaseArg = DAL.getLastArg(options::OPT_emit_ast))) { 385 FinalPhase = phases::Compile; 386 387 // -S only runs up to the backend. 388 } else if ((PhaseArg = DAL.getLastArg(options::OPT_S))) { 389 FinalPhase = phases::Backend; 390 391 // -c compilation only runs up to the assembler. 392 } else if ((PhaseArg = DAL.getLastArg(options::OPT_c))) { 393 FinalPhase = phases::Assemble; 394 395 } else if ((PhaseArg = DAL.getLastArg(options::OPT_emit_interface_stubs))) { 396 FinalPhase = phases::IfsMerge; 397 398 // Otherwise do everything. 399 } else 400 FinalPhase = phases::Link; 401 402 if (FinalPhaseArg) 403 *FinalPhaseArg = PhaseArg; 404 405 return FinalPhase; 406 } 407 408 static Arg *MakeInputArg(DerivedArgList &Args, const OptTable &Opts, 409 StringRef Value, bool Claim = true) { 410 Arg *A = new Arg(Opts.getOption(options::OPT_INPUT), Value, 411 Args.getBaseArgs().MakeIndex(Value), Value.data()); 412 Args.AddSynthesizedArg(A); 413 if (Claim) 414 A->claim(); 415 return A; 416 } 417 418 DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args) const { 419 const llvm::opt::OptTable &Opts = getOpts(); 420 DerivedArgList *DAL = new DerivedArgList(Args); 421 422 bool HasNostdlib = Args.hasArg(options::OPT_nostdlib); 423 bool HasNostdlibxx = Args.hasArg(options::OPT_nostdlibxx); 424 bool HasNodefaultlib = Args.hasArg(options::OPT_nodefaultlibs); 425 bool IgnoreUnused = false; 426 for (Arg *A : Args) { 427 if (IgnoreUnused) 428 A->claim(); 429 430 if (A->getOption().matches(options::OPT_start_no_unused_arguments)) { 431 IgnoreUnused = true; 432 continue; 433 } 434 if (A->getOption().matches(options::OPT_end_no_unused_arguments)) { 435 IgnoreUnused = false; 436 continue; 437 } 438 439 // Unfortunately, we have to parse some forwarding options (-Xassembler, 440 // -Xlinker, -Xpreprocessor) because we either integrate their functionality 441 // (assembler and preprocessor), or bypass a previous driver ('collect2'). 442 443 // Rewrite linker options, to replace --no-demangle with a custom internal 444 // option. 445 if ((A->getOption().matches(options::OPT_Wl_COMMA) || 446 A->getOption().matches(options::OPT_Xlinker)) && 447 A->containsValue("--no-demangle")) { 448 // Add the rewritten no-demangle argument. 449 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_Xlinker__no_demangle)); 450 451 // Add the remaining values as Xlinker arguments. 452 for (StringRef Val : A->getValues()) 453 if (Val != "--no-demangle") 454 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_Xlinker), Val); 455 456 continue; 457 } 458 459 // Rewrite preprocessor options, to replace -Wp,-MD,FOO which is used by 460 // some build systems. We don't try to be complete here because we don't 461 // care to encourage this usage model. 462 if (A->getOption().matches(options::OPT_Wp_COMMA) && 463 (A->getValue(0) == StringRef("-MD") || 464 A->getValue(0) == StringRef("-MMD"))) { 465 // Rewrite to -MD/-MMD along with -MF. 466 if (A->getValue(0) == StringRef("-MD")) 467 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MD)); 468 else 469 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MMD)); 470 if (A->getNumValues() == 2) 471 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_MF), A->getValue(1)); 472 continue; 473 } 474 475 // Rewrite reserved library names. 476 if (A->getOption().matches(options::OPT_l)) { 477 StringRef Value = A->getValue(); 478 479 // Rewrite unless -nostdlib is present. 480 if (!HasNostdlib && !HasNodefaultlib && !HasNostdlibxx && 481 Value == "stdc++") { 482 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_stdcxx)); 483 continue; 484 } 485 486 // Rewrite unconditionally. 487 if (Value == "cc_kext") { 488 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_cckext)); 489 continue; 490 } 491 } 492 493 // Pick up inputs via the -- option. 494 if (A->getOption().matches(options::OPT__DASH_DASH)) { 495 A->claim(); 496 for (StringRef Val : A->getValues()) 497 DAL->append(MakeInputArg(*DAL, Opts, Val, false)); 498 continue; 499 } 500 501 DAL->append(A); 502 } 503 504 // Enforce -static if -miamcu is present. 505 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) 506 DAL->AddFlagArg(nullptr, Opts.getOption(options::OPT_static)); 507 508 // Add a default value of -mlinker-version=, if one was given and the user 509 // didn't specify one. 510 #if defined(HOST_LINK_VERSION) 511 if (!Args.hasArg(options::OPT_mlinker_version_EQ) && 512 strlen(HOST_LINK_VERSION) > 0) { 513 DAL->AddJoinedArg(0, Opts.getOption(options::OPT_mlinker_version_EQ), 514 HOST_LINK_VERSION); 515 DAL->getLastArg(options::OPT_mlinker_version_EQ)->claim(); 516 } 517 #endif 518 519 return DAL; 520 } 521 522 /// Compute target triple from args. 523 /// 524 /// This routine provides the logic to compute a target triple from various 525 /// args passed to the driver and the default triple string. 526 static llvm::Triple computeTargetTriple(const Driver &D, 527 StringRef TargetTriple, 528 const ArgList &Args, 529 StringRef DarwinArchName = "") { 530 // FIXME: Already done in Compilation *Driver::BuildCompilation 531 if (const Arg *A = Args.getLastArg(options::OPT_target)) 532 TargetTriple = A->getValue(); 533 534 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 535 536 // GNU/Hurd's triples should have been -hurd-gnu*, but were historically made 537 // -gnu* only, and we can not change this, so we have to detect that case as 538 // being the Hurd OS. 539 if (TargetTriple.contains("-unknown-gnu") || TargetTriple.contains("-pc-gnu")) 540 Target.setOSName("hurd"); 541 542 // Handle Apple-specific options available here. 543 if (Target.isOSBinFormatMachO()) { 544 // If an explicit Darwin arch name is given, that trumps all. 545 if (!DarwinArchName.empty()) { 546 tools::darwin::setTripleTypeForMachOArchName(Target, DarwinArchName); 547 return Target; 548 } 549 550 // Handle the Darwin '-arch' flag. 551 if (Arg *A = Args.getLastArg(options::OPT_arch)) { 552 StringRef ArchName = A->getValue(); 553 tools::darwin::setTripleTypeForMachOArchName(Target, ArchName); 554 } 555 } 556 557 // Handle pseudo-target flags '-mlittle-endian'/'-EL' and 558 // '-mbig-endian'/'-EB'. 559 if (Arg *A = Args.getLastArg(options::OPT_mlittle_endian, 560 options::OPT_mbig_endian)) { 561 if (A->getOption().matches(options::OPT_mlittle_endian)) { 562 llvm::Triple LE = Target.getLittleEndianArchVariant(); 563 if (LE.getArch() != llvm::Triple::UnknownArch) 564 Target = std::move(LE); 565 } else { 566 llvm::Triple BE = Target.getBigEndianArchVariant(); 567 if (BE.getArch() != llvm::Triple::UnknownArch) 568 Target = std::move(BE); 569 } 570 } 571 572 // Skip further flag support on OSes which don't support '-m32' or '-m64'. 573 if (Target.getArch() == llvm::Triple::tce || 574 Target.getOS() == llvm::Triple::Minix) 575 return Target; 576 577 // On AIX, the env OBJECT_MODE may affect the resulting arch variant. 578 if (Target.isOSAIX()) { 579 if (std::optional<std::string> ObjectModeValue = 580 llvm::sys::Process::GetEnv("OBJECT_MODE")) { 581 StringRef ObjectMode = *ObjectModeValue; 582 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 583 584 if (ObjectMode.equals("64")) { 585 AT = Target.get64BitArchVariant().getArch(); 586 } else if (ObjectMode.equals("32")) { 587 AT = Target.get32BitArchVariant().getArch(); 588 } else { 589 D.Diag(diag::err_drv_invalid_object_mode) << ObjectMode; 590 } 591 592 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) 593 Target.setArch(AT); 594 } 595 } 596 597 // Handle pseudo-target flags '-m64', '-mx32', '-m32' and '-m16'. 598 Arg *A = Args.getLastArg(options::OPT_m64, options::OPT_mx32, 599 options::OPT_m32, options::OPT_m16); 600 if (A) { 601 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 602 603 if (A->getOption().matches(options::OPT_m64)) { 604 AT = Target.get64BitArchVariant().getArch(); 605 if (Target.getEnvironment() == llvm::Triple::GNUX32) 606 Target.setEnvironment(llvm::Triple::GNU); 607 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 608 Target.setEnvironment(llvm::Triple::Musl); 609 } else if (A->getOption().matches(options::OPT_mx32) && 610 Target.get64BitArchVariant().getArch() == llvm::Triple::x86_64) { 611 AT = llvm::Triple::x86_64; 612 if (Target.getEnvironment() == llvm::Triple::Musl) 613 Target.setEnvironment(llvm::Triple::MuslX32); 614 else 615 Target.setEnvironment(llvm::Triple::GNUX32); 616 } else if (A->getOption().matches(options::OPT_m32)) { 617 AT = Target.get32BitArchVariant().getArch(); 618 if (Target.getEnvironment() == llvm::Triple::GNUX32) 619 Target.setEnvironment(llvm::Triple::GNU); 620 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 621 Target.setEnvironment(llvm::Triple::Musl); 622 } else if (A->getOption().matches(options::OPT_m16) && 623 Target.get32BitArchVariant().getArch() == llvm::Triple::x86) { 624 AT = llvm::Triple::x86; 625 Target.setEnvironment(llvm::Triple::CODE16); 626 } 627 628 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) { 629 Target.setArch(AT); 630 if (Target.isWindowsGNUEnvironment()) 631 toolchains::MinGW::fixTripleArch(D, Target, Args); 632 } 633 } 634 635 // Handle -miamcu flag. 636 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) { 637 if (Target.get32BitArchVariant().getArch() != llvm::Triple::x86) 638 D.Diag(diag::err_drv_unsupported_opt_for_target) << "-miamcu" 639 << Target.str(); 640 641 if (A && !A->getOption().matches(options::OPT_m32)) 642 D.Diag(diag::err_drv_argument_not_allowed_with) 643 << "-miamcu" << A->getBaseArg().getAsString(Args); 644 645 Target.setArch(llvm::Triple::x86); 646 Target.setArchName("i586"); 647 Target.setEnvironment(llvm::Triple::UnknownEnvironment); 648 Target.setEnvironmentName(""); 649 Target.setOS(llvm::Triple::ELFIAMCU); 650 Target.setVendor(llvm::Triple::UnknownVendor); 651 Target.setVendorName("intel"); 652 } 653 654 // If target is MIPS adjust the target triple 655 // accordingly to provided ABI name. 656 if (Target.isMIPS()) { 657 if ((A = Args.getLastArg(options::OPT_mabi_EQ))) { 658 StringRef ABIName = A->getValue(); 659 if (ABIName == "32") { 660 Target = Target.get32BitArchVariant(); 661 if (Target.getEnvironment() == llvm::Triple::GNUABI64 || 662 Target.getEnvironment() == llvm::Triple::GNUABIN32) 663 Target.setEnvironment(llvm::Triple::GNU); 664 } else if (ABIName == "n32") { 665 Target = Target.get64BitArchVariant(); 666 if (Target.getEnvironment() == llvm::Triple::GNU || 667 Target.getEnvironment() == llvm::Triple::GNUABI64) 668 Target.setEnvironment(llvm::Triple::GNUABIN32); 669 } else if (ABIName == "64") { 670 Target = Target.get64BitArchVariant(); 671 if (Target.getEnvironment() == llvm::Triple::GNU || 672 Target.getEnvironment() == llvm::Triple::GNUABIN32) 673 Target.setEnvironment(llvm::Triple::GNUABI64); 674 } 675 } 676 } 677 678 // If target is RISC-V adjust the target triple according to 679 // provided architecture name 680 if (Target.isRISCV()) { 681 if ((A = Args.getLastArg(options::OPT_march_EQ))) { 682 StringRef ArchName = A->getValue(); 683 if (ArchName.startswith_insensitive("rv32")) 684 Target.setArch(llvm::Triple::riscv32); 685 else if (ArchName.startswith_insensitive("rv64")) 686 Target.setArch(llvm::Triple::riscv64); 687 } 688 } 689 690 return Target; 691 } 692 693 // Parse the LTO options and record the type of LTO compilation 694 // based on which -f(no-)?lto(=.*)? or -f(no-)?offload-lto(=.*)? 695 // option occurs last. 696 static driver::LTOKind parseLTOMode(Driver &D, const llvm::opt::ArgList &Args, 697 OptSpecifier OptEq, OptSpecifier OptNeg) { 698 if (!Args.hasFlag(OptEq, OptNeg, false)) 699 return LTOK_None; 700 701 const Arg *A = Args.getLastArg(OptEq); 702 StringRef LTOName = A->getValue(); 703 704 driver::LTOKind LTOMode = llvm::StringSwitch<LTOKind>(LTOName) 705 .Case("full", LTOK_Full) 706 .Case("thin", LTOK_Thin) 707 .Default(LTOK_Unknown); 708 709 if (LTOMode == LTOK_Unknown) { 710 D.Diag(diag::err_drv_unsupported_option_argument) 711 << A->getSpelling() << A->getValue(); 712 return LTOK_None; 713 } 714 return LTOMode; 715 } 716 717 // Parse the LTO options. 718 void Driver::setLTOMode(const llvm::opt::ArgList &Args) { 719 LTOMode = 720 parseLTOMode(*this, Args, options::OPT_flto_EQ, options::OPT_fno_lto); 721 722 OffloadLTOMode = parseLTOMode(*this, Args, options::OPT_foffload_lto_EQ, 723 options::OPT_fno_offload_lto); 724 725 // Try to enable `-foffload-lto=full` if `-fopenmp-target-jit` is on. 726 if (Args.hasFlag(options::OPT_fopenmp_target_jit, 727 options::OPT_fno_openmp_target_jit, false)) { 728 if (Arg *A = Args.getLastArg(options::OPT_foffload_lto_EQ, 729 options::OPT_fno_offload_lto)) 730 if (OffloadLTOMode != LTOK_Full) 731 Diag(diag::err_drv_incompatible_options) 732 << A->getSpelling() << "-fopenmp-target-jit"; 733 OffloadLTOMode = LTOK_Full; 734 } 735 } 736 737 /// Compute the desired OpenMP runtime from the flags provided. 738 Driver::OpenMPRuntimeKind Driver::getOpenMPRuntime(const ArgList &Args) const { 739 StringRef RuntimeName(CLANG_DEFAULT_OPENMP_RUNTIME); 740 741 const Arg *A = Args.getLastArg(options::OPT_fopenmp_EQ); 742 if (A) 743 RuntimeName = A->getValue(); 744 745 auto RT = llvm::StringSwitch<OpenMPRuntimeKind>(RuntimeName) 746 .Case("libomp", OMPRT_OMP) 747 .Case("libgomp", OMPRT_GOMP) 748 .Case("libiomp5", OMPRT_IOMP5) 749 .Default(OMPRT_Unknown); 750 751 if (RT == OMPRT_Unknown) { 752 if (A) 753 Diag(diag::err_drv_unsupported_option_argument) 754 << A->getSpelling() << A->getValue(); 755 else 756 // FIXME: We could use a nicer diagnostic here. 757 Diag(diag::err_drv_unsupported_opt) << "-fopenmp"; 758 } 759 760 return RT; 761 } 762 763 void Driver::CreateOffloadingDeviceToolChains(Compilation &C, 764 InputList &Inputs) { 765 766 // 767 // CUDA/HIP 768 // 769 // We need to generate a CUDA/HIP toolchain if any of the inputs has a CUDA 770 // or HIP type. However, mixed CUDA/HIP compilation is not supported. 771 bool IsCuda = 772 llvm::any_of(Inputs, [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 773 return types::isCuda(I.first); 774 }); 775 bool IsHIP = 776 llvm::any_of(Inputs, 777 [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 778 return types::isHIP(I.first); 779 }) || 780 C.getInputArgs().hasArg(options::OPT_hip_link); 781 if (IsCuda && IsHIP) { 782 Diag(clang::diag::err_drv_mix_cuda_hip); 783 return; 784 } 785 if (IsCuda) { 786 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 787 const llvm::Triple &HostTriple = HostTC->getTriple(); 788 auto OFK = Action::OFK_Cuda; 789 auto CudaTriple = 790 getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), HostTriple); 791 if (!CudaTriple) 792 return; 793 // Use the CUDA and host triples as the key into the ToolChains map, 794 // because the device toolchain we create depends on both. 795 auto &CudaTC = ToolChains[CudaTriple->str() + "/" + HostTriple.str()]; 796 if (!CudaTC) { 797 CudaTC = std::make_unique<toolchains::CudaToolChain>( 798 *this, *CudaTriple, *HostTC, C.getInputArgs()); 799 } 800 C.addOffloadDeviceToolChain(CudaTC.get(), OFK); 801 } else if (IsHIP) { 802 if (auto *OMPTargetArg = 803 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 804 Diag(clang::diag::err_drv_unsupported_opt_for_language_mode) 805 << OMPTargetArg->getSpelling() << "HIP"; 806 return; 807 } 808 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 809 auto OFK = Action::OFK_HIP; 810 auto HIPTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 811 if (!HIPTriple) 812 return; 813 auto *HIPTC = &getOffloadingDeviceToolChain(C.getInputArgs(), *HIPTriple, 814 *HostTC, OFK); 815 assert(HIPTC && "Could not create offloading device tool chain."); 816 C.addOffloadDeviceToolChain(HIPTC, OFK); 817 } 818 819 // 820 // OpenMP 821 // 822 // We need to generate an OpenMP toolchain if the user specified targets with 823 // the -fopenmp-targets option or used --offload-arch with OpenMP enabled. 824 bool IsOpenMPOffloading = 825 C.getInputArgs().hasFlag(options::OPT_fopenmp, options::OPT_fopenmp_EQ, 826 options::OPT_fno_openmp, false) && 827 (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ) || 828 C.getInputArgs().hasArg(options::OPT_offload_arch_EQ)); 829 if (IsOpenMPOffloading) { 830 // We expect that -fopenmp-targets is always used in conjunction with the 831 // option -fopenmp specifying a valid runtime with offloading support, i.e. 832 // libomp or libiomp. 833 OpenMPRuntimeKind RuntimeKind = getOpenMPRuntime(C.getInputArgs()); 834 if (RuntimeKind != OMPRT_OMP && RuntimeKind != OMPRT_IOMP5) { 835 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 836 return; 837 } 838 839 llvm::StringMap<llvm::DenseSet<StringRef>> DerivedArchs; 840 llvm::StringMap<StringRef> FoundNormalizedTriples; 841 llvm::SmallVector<StringRef, 4> OpenMPTriples; 842 843 // If the user specified -fopenmp-targets= we create a toolchain for each 844 // valid triple. Otherwise, if only --offload-arch= was specified we instead 845 // attempt to derive the appropriate toolchains from the arguments. 846 if (Arg *OpenMPTargets = 847 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 848 if (OpenMPTargets && !OpenMPTargets->getNumValues()) { 849 Diag(clang::diag::warn_drv_empty_joined_argument) 850 << OpenMPTargets->getAsString(C.getInputArgs()); 851 return; 852 } 853 llvm::copy(OpenMPTargets->getValues(), std::back_inserter(OpenMPTriples)); 854 } else if (C.getInputArgs().hasArg(options::OPT_offload_arch_EQ) && 855 !IsHIP && !IsCuda) { 856 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 857 auto AMDTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 858 auto NVPTXTriple = getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), 859 HostTC->getTriple()); 860 861 // Attempt to deduce the offloading triple from the set of architectures. 862 // We can only correctly deduce NVPTX / AMDGPU triples currently. We need 863 // to temporarily create these toolchains so that we can access tools for 864 // inferring architectures. 865 llvm::DenseSet<StringRef> Archs; 866 if (NVPTXTriple) { 867 auto TempTC = std::make_unique<toolchains::CudaToolChain>( 868 *this, *NVPTXTriple, *HostTC, C.getInputArgs()); 869 for (StringRef Arch : getOffloadArchs( 870 C, C.getArgs(), Action::OFK_OpenMP, &*TempTC, true)) 871 Archs.insert(Arch); 872 } 873 if (AMDTriple) { 874 auto TempTC = std::make_unique<toolchains::AMDGPUOpenMPToolChain>( 875 *this, *AMDTriple, *HostTC, C.getInputArgs()); 876 for (StringRef Arch : getOffloadArchs( 877 C, C.getArgs(), Action::OFK_OpenMP, &*TempTC, true)) 878 Archs.insert(Arch); 879 } 880 if (!AMDTriple && !NVPTXTriple) { 881 for (StringRef Arch : 882 getOffloadArchs(C, C.getArgs(), Action::OFK_OpenMP, nullptr, true)) 883 Archs.insert(Arch); 884 } 885 886 for (StringRef Arch : Archs) { 887 if (NVPTXTriple && IsNVIDIAGpuArch(StringToCudaArch( 888 getProcessorFromTargetID(*NVPTXTriple, Arch)))) { 889 DerivedArchs[NVPTXTriple->getTriple()].insert(Arch); 890 } else if (AMDTriple && 891 IsAMDGpuArch(StringToCudaArch( 892 getProcessorFromTargetID(*AMDTriple, Arch)))) { 893 DerivedArchs[AMDTriple->getTriple()].insert(Arch); 894 } else { 895 Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch) << Arch; 896 return; 897 } 898 } 899 900 // If the set is empty then we failed to find a native architecture. 901 if (Archs.empty()) { 902 Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch) 903 << "native"; 904 return; 905 } 906 907 for (const auto &TripleAndArchs : DerivedArchs) 908 OpenMPTriples.push_back(TripleAndArchs.first()); 909 } 910 911 for (StringRef Val : OpenMPTriples) { 912 llvm::Triple TT(ToolChain::getOpenMPTriple(Val)); 913 std::string NormalizedName = TT.normalize(); 914 915 // Make sure we don't have a duplicate triple. 916 auto Duplicate = FoundNormalizedTriples.find(NormalizedName); 917 if (Duplicate != FoundNormalizedTriples.end()) { 918 Diag(clang::diag::warn_drv_omp_offload_target_duplicate) 919 << Val << Duplicate->second; 920 continue; 921 } 922 923 // Store the current triple so that we can check for duplicates in the 924 // following iterations. 925 FoundNormalizedTriples[NormalizedName] = Val; 926 927 // If the specified target is invalid, emit a diagnostic. 928 if (TT.getArch() == llvm::Triple::UnknownArch) 929 Diag(clang::diag::err_drv_invalid_omp_target) << Val; 930 else { 931 const ToolChain *TC; 932 // Device toolchains have to be selected differently. They pair host 933 // and device in their implementation. 934 if (TT.isNVPTX() || TT.isAMDGCN()) { 935 const ToolChain *HostTC = 936 C.getSingleOffloadToolChain<Action::OFK_Host>(); 937 assert(HostTC && "Host toolchain should be always defined."); 938 auto &DeviceTC = 939 ToolChains[TT.str() + "/" + HostTC->getTriple().normalize()]; 940 if (!DeviceTC) { 941 if (TT.isNVPTX()) 942 DeviceTC = std::make_unique<toolchains::CudaToolChain>( 943 *this, TT, *HostTC, C.getInputArgs()); 944 else if (TT.isAMDGCN()) 945 DeviceTC = std::make_unique<toolchains::AMDGPUOpenMPToolChain>( 946 *this, TT, *HostTC, C.getInputArgs()); 947 else 948 assert(DeviceTC && "Device toolchain not defined."); 949 } 950 951 TC = DeviceTC.get(); 952 } else 953 TC = &getToolChain(C.getInputArgs(), TT); 954 C.addOffloadDeviceToolChain(TC, Action::OFK_OpenMP); 955 if (DerivedArchs.find(TT.getTriple()) != DerivedArchs.end()) 956 KnownArchs[TC] = DerivedArchs[TT.getTriple()]; 957 } 958 } 959 } else if (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ)) { 960 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 961 return; 962 } 963 964 // 965 // TODO: Add support for other offloading programming models here. 966 // 967 } 968 969 static void appendOneArg(InputArgList &Args, const Arg *Opt, 970 const Arg *BaseArg) { 971 // The args for config files or /clang: flags belong to different InputArgList 972 // objects than Args. This copies an Arg from one of those other InputArgLists 973 // to the ownership of Args. 974 unsigned Index = Args.MakeIndex(Opt->getSpelling()); 975 Arg *Copy = new llvm::opt::Arg(Opt->getOption(), Args.getArgString(Index), 976 Index, BaseArg); 977 Copy->getValues() = Opt->getValues(); 978 if (Opt->isClaimed()) 979 Copy->claim(); 980 Copy->setOwnsValues(Opt->getOwnsValues()); 981 Opt->setOwnsValues(false); 982 Args.append(Copy); 983 } 984 985 bool Driver::readConfigFile(StringRef FileName, 986 llvm::cl::ExpansionContext &ExpCtx) { 987 // Try opening the given file. 988 auto Status = getVFS().status(FileName); 989 if (!Status) { 990 Diag(diag::err_drv_cannot_open_config_file) 991 << FileName << Status.getError().message(); 992 return true; 993 } 994 if (Status->getType() != llvm::sys::fs::file_type::regular_file) { 995 Diag(diag::err_drv_cannot_open_config_file) 996 << FileName << "not a regular file"; 997 return true; 998 } 999 1000 // Try reading the given file. 1001 SmallVector<const char *, 32> NewCfgArgs; 1002 if (llvm::Error Err = ExpCtx.readConfigFile(FileName, NewCfgArgs)) { 1003 Diag(diag::err_drv_cannot_read_config_file) 1004 << FileName << toString(std::move(Err)); 1005 return true; 1006 } 1007 1008 // Read options from config file. 1009 llvm::SmallString<128> CfgFileName(FileName); 1010 llvm::sys::path::native(CfgFileName); 1011 bool ContainErrors; 1012 std::unique_ptr<InputArgList> NewOptions = std::make_unique<InputArgList>( 1013 ParseArgStrings(NewCfgArgs, IsCLMode(), ContainErrors)); 1014 if (ContainErrors) 1015 return true; 1016 1017 // Claim all arguments that come from a configuration file so that the driver 1018 // does not warn on any that is unused. 1019 for (Arg *A : *NewOptions) 1020 A->claim(); 1021 1022 if (!CfgOptions) 1023 CfgOptions = std::move(NewOptions); 1024 else { 1025 // If this is a subsequent config file, append options to the previous one. 1026 for (auto *Opt : *NewOptions) { 1027 const Arg *BaseArg = &Opt->getBaseArg(); 1028 if (BaseArg == Opt) 1029 BaseArg = nullptr; 1030 appendOneArg(*CfgOptions, Opt, BaseArg); 1031 } 1032 } 1033 ConfigFiles.push_back(std::string(CfgFileName)); 1034 return false; 1035 } 1036 1037 bool Driver::loadConfigFiles() { 1038 llvm::cl::ExpansionContext ExpCtx(Saver.getAllocator(), 1039 llvm::cl::tokenizeConfigFile); 1040 ExpCtx.setVFS(&getVFS()); 1041 1042 // Process options that change search path for config files. 1043 if (CLOptions) { 1044 if (CLOptions->hasArg(options::OPT_config_system_dir_EQ)) { 1045 SmallString<128> CfgDir; 1046 CfgDir.append( 1047 CLOptions->getLastArgValue(options::OPT_config_system_dir_EQ)); 1048 if (CfgDir.empty() || getVFS().makeAbsolute(CfgDir)) 1049 SystemConfigDir.clear(); 1050 else 1051 SystemConfigDir = static_cast<std::string>(CfgDir); 1052 } 1053 if (CLOptions->hasArg(options::OPT_config_user_dir_EQ)) { 1054 SmallString<128> CfgDir; 1055 llvm::sys::fs::expand_tilde( 1056 CLOptions->getLastArgValue(options::OPT_config_user_dir_EQ), CfgDir); 1057 if (CfgDir.empty() || getVFS().makeAbsolute(CfgDir)) 1058 UserConfigDir.clear(); 1059 else 1060 UserConfigDir = static_cast<std::string>(CfgDir); 1061 } 1062 } 1063 1064 // Prepare list of directories where config file is searched for. 1065 StringRef CfgFileSearchDirs[] = {UserConfigDir, SystemConfigDir, Dir}; 1066 ExpCtx.setSearchDirs(CfgFileSearchDirs); 1067 1068 // First try to load configuration from the default files, return on error. 1069 if (loadDefaultConfigFiles(ExpCtx)) 1070 return true; 1071 1072 // Then load configuration files specified explicitly. 1073 SmallString<128> CfgFilePath; 1074 if (CLOptions) { 1075 for (auto CfgFileName : CLOptions->getAllArgValues(options::OPT_config)) { 1076 // If argument contains directory separator, treat it as a path to 1077 // configuration file. 1078 if (llvm::sys::path::has_parent_path(CfgFileName)) { 1079 CfgFilePath.assign(CfgFileName); 1080 if (llvm::sys::path::is_relative(CfgFilePath)) { 1081 if (getVFS().makeAbsolute(CfgFilePath)) { 1082 Diag(diag::err_drv_cannot_open_config_file) 1083 << CfgFilePath << "cannot get absolute path"; 1084 return true; 1085 } 1086 } 1087 } else if (!ExpCtx.findConfigFile(CfgFileName, CfgFilePath)) { 1088 // Report an error that the config file could not be found. 1089 Diag(diag::err_drv_config_file_not_found) << CfgFileName; 1090 for (const StringRef &SearchDir : CfgFileSearchDirs) 1091 if (!SearchDir.empty()) 1092 Diag(diag::note_drv_config_file_searched_in) << SearchDir; 1093 return true; 1094 } 1095 1096 // Try to read the config file, return on error. 1097 if (readConfigFile(CfgFilePath, ExpCtx)) 1098 return true; 1099 } 1100 } 1101 1102 // No error occurred. 1103 return false; 1104 } 1105 1106 bool Driver::loadDefaultConfigFiles(llvm::cl::ExpansionContext &ExpCtx) { 1107 // Disable default config if CLANG_NO_DEFAULT_CONFIG is set to a non-empty 1108 // value. 1109 if (const char *NoConfigEnv = ::getenv("CLANG_NO_DEFAULT_CONFIG")) { 1110 if (*NoConfigEnv) 1111 return false; 1112 } 1113 if (CLOptions && CLOptions->hasArg(options::OPT_no_default_config)) 1114 return false; 1115 1116 std::string RealMode = getExecutableForDriverMode(Mode); 1117 std::string Triple; 1118 1119 // If name prefix is present, no --target= override was passed via CLOptions 1120 // and the name prefix is not a valid triple, force it for backwards 1121 // compatibility. 1122 if (!ClangNameParts.TargetPrefix.empty() && 1123 computeTargetTriple(*this, "/invalid/", *CLOptions).str() == 1124 "/invalid/") { 1125 llvm::Triple PrefixTriple{ClangNameParts.TargetPrefix}; 1126 if (PrefixTriple.getArch() == llvm::Triple::UnknownArch || 1127 PrefixTriple.isOSUnknown()) 1128 Triple = PrefixTriple.str(); 1129 } 1130 1131 // Otherwise, use the real triple as used by the driver. 1132 if (Triple.empty()) { 1133 llvm::Triple RealTriple = 1134 computeTargetTriple(*this, TargetTriple, *CLOptions); 1135 Triple = RealTriple.str(); 1136 assert(!Triple.empty()); 1137 } 1138 1139 // Search for config files in the following order: 1140 // 1. <triple>-<mode>.cfg using real driver mode 1141 // (e.g. i386-pc-linux-gnu-clang++.cfg). 1142 // 2. <triple>-<mode>.cfg using executable suffix 1143 // (e.g. i386-pc-linux-gnu-clang-g++.cfg for *clang-g++). 1144 // 3. <triple>.cfg + <mode>.cfg using real driver mode 1145 // (e.g. i386-pc-linux-gnu.cfg + clang++.cfg). 1146 // 4. <triple>.cfg + <mode>.cfg using executable suffix 1147 // (e.g. i386-pc-linux-gnu.cfg + clang-g++.cfg for *clang-g++). 1148 1149 // Try loading <triple>-<mode>.cfg, and return if we find a match. 1150 SmallString<128> CfgFilePath; 1151 std::string CfgFileName = Triple + '-' + RealMode + ".cfg"; 1152 if (ExpCtx.findConfigFile(CfgFileName, CfgFilePath)) 1153 return readConfigFile(CfgFilePath, ExpCtx); 1154 1155 bool TryModeSuffix = !ClangNameParts.ModeSuffix.empty() && 1156 ClangNameParts.ModeSuffix != RealMode; 1157 if (TryModeSuffix) { 1158 CfgFileName = Triple + '-' + ClangNameParts.ModeSuffix + ".cfg"; 1159 if (ExpCtx.findConfigFile(CfgFileName, CfgFilePath)) 1160 return readConfigFile(CfgFilePath, ExpCtx); 1161 } 1162 1163 // Try loading <mode>.cfg, and return if loading failed. If a matching file 1164 // was not found, still proceed on to try <triple>.cfg. 1165 CfgFileName = RealMode + ".cfg"; 1166 if (ExpCtx.findConfigFile(CfgFileName, CfgFilePath)) { 1167 if (readConfigFile(CfgFilePath, ExpCtx)) 1168 return true; 1169 } else if (TryModeSuffix) { 1170 CfgFileName = ClangNameParts.ModeSuffix + ".cfg"; 1171 if (ExpCtx.findConfigFile(CfgFileName, CfgFilePath) && 1172 readConfigFile(CfgFilePath, ExpCtx)) 1173 return true; 1174 } 1175 1176 // Try loading <triple>.cfg and return if we find a match. 1177 CfgFileName = Triple + ".cfg"; 1178 if (ExpCtx.findConfigFile(CfgFileName, CfgFilePath)) 1179 return readConfigFile(CfgFilePath, ExpCtx); 1180 1181 // If we were unable to find a config file deduced from executable name, 1182 // that is not an error. 1183 return false; 1184 } 1185 1186 Compilation *Driver::BuildCompilation(ArrayRef<const char *> ArgList) { 1187 llvm::PrettyStackTraceString CrashInfo("Compilation construction"); 1188 1189 // FIXME: Handle environment options which affect driver behavior, somewhere 1190 // (client?). GCC_EXEC_PREFIX, LPATH, CC_PRINT_OPTIONS. 1191 1192 // We look for the driver mode option early, because the mode can affect 1193 // how other options are parsed. 1194 1195 auto DriverMode = getDriverMode(ClangExecutable, ArgList.slice(1)); 1196 if (!DriverMode.empty()) 1197 setDriverMode(DriverMode); 1198 1199 // FIXME: What are we going to do with -V and -b? 1200 1201 // Arguments specified in command line. 1202 bool ContainsError; 1203 CLOptions = std::make_unique<InputArgList>( 1204 ParseArgStrings(ArgList.slice(1), IsCLMode(), ContainsError)); 1205 1206 // Try parsing configuration file. 1207 if (!ContainsError) 1208 ContainsError = loadConfigFiles(); 1209 bool HasConfigFile = !ContainsError && (CfgOptions.get() != nullptr); 1210 1211 // All arguments, from both config file and command line. 1212 InputArgList Args = std::move(HasConfigFile ? std::move(*CfgOptions) 1213 : std::move(*CLOptions)); 1214 1215 if (HasConfigFile) 1216 for (auto *Opt : *CLOptions) { 1217 if (Opt->getOption().matches(options::OPT_config)) 1218 continue; 1219 const Arg *BaseArg = &Opt->getBaseArg(); 1220 if (BaseArg == Opt) 1221 BaseArg = nullptr; 1222 appendOneArg(Args, Opt, BaseArg); 1223 } 1224 1225 // In CL mode, look for any pass-through arguments 1226 if (IsCLMode() && !ContainsError) { 1227 SmallVector<const char *, 16> CLModePassThroughArgList; 1228 for (const auto *A : Args.filtered(options::OPT__SLASH_clang)) { 1229 A->claim(); 1230 CLModePassThroughArgList.push_back(A->getValue()); 1231 } 1232 1233 if (!CLModePassThroughArgList.empty()) { 1234 // Parse any pass through args using default clang processing rather 1235 // than clang-cl processing. 1236 auto CLModePassThroughOptions = std::make_unique<InputArgList>( 1237 ParseArgStrings(CLModePassThroughArgList, false, ContainsError)); 1238 1239 if (!ContainsError) 1240 for (auto *Opt : *CLModePassThroughOptions) { 1241 appendOneArg(Args, Opt, nullptr); 1242 } 1243 } 1244 } 1245 1246 // Check for working directory option before accessing any files 1247 if (Arg *WD = Args.getLastArg(options::OPT_working_directory)) 1248 if (VFS->setCurrentWorkingDirectory(WD->getValue())) 1249 Diag(diag::err_drv_unable_to_set_working_directory) << WD->getValue(); 1250 1251 // FIXME: This stuff needs to go into the Compilation, not the driver. 1252 bool CCCPrintPhases; 1253 1254 // -canonical-prefixes, -no-canonical-prefixes are used very early in main. 1255 Args.ClaimAllArgs(options::OPT_canonical_prefixes); 1256 Args.ClaimAllArgs(options::OPT_no_canonical_prefixes); 1257 1258 // f(no-)integated-cc1 is also used very early in main. 1259 Args.ClaimAllArgs(options::OPT_fintegrated_cc1); 1260 Args.ClaimAllArgs(options::OPT_fno_integrated_cc1); 1261 1262 // Ignore -pipe. 1263 Args.ClaimAllArgs(options::OPT_pipe); 1264 1265 // Extract -ccc args. 1266 // 1267 // FIXME: We need to figure out where this behavior should live. Most of it 1268 // should be outside in the client; the parts that aren't should have proper 1269 // options, either by introducing new ones or by overloading gcc ones like -V 1270 // or -b. 1271 CCCPrintPhases = Args.hasArg(options::OPT_ccc_print_phases); 1272 CCCPrintBindings = Args.hasArg(options::OPT_ccc_print_bindings); 1273 if (const Arg *A = Args.getLastArg(options::OPT_ccc_gcc_name)) 1274 CCCGenericGCCName = A->getValue(); 1275 1276 // Process -fproc-stat-report options. 1277 if (const Arg *A = Args.getLastArg(options::OPT_fproc_stat_report_EQ)) { 1278 CCPrintProcessStats = true; 1279 CCPrintStatReportFilename = A->getValue(); 1280 } 1281 if (Args.hasArg(options::OPT_fproc_stat_report)) 1282 CCPrintProcessStats = true; 1283 1284 // FIXME: TargetTriple is used by the target-prefixed calls to as/ld 1285 // and getToolChain is const. 1286 if (IsCLMode()) { 1287 // clang-cl targets MSVC-style Win32. 1288 llvm::Triple T(TargetTriple); 1289 T.setOS(llvm::Triple::Win32); 1290 T.setVendor(llvm::Triple::PC); 1291 T.setEnvironment(llvm::Triple::MSVC); 1292 T.setObjectFormat(llvm::Triple::COFF); 1293 if (Args.hasArg(options::OPT__SLASH_arm64EC)) 1294 T.setArch(llvm::Triple::aarch64, llvm::Triple::AArch64SubArch_arm64ec); 1295 TargetTriple = T.str(); 1296 } else if (IsDXCMode()) { 1297 // Build TargetTriple from target_profile option for clang-dxc. 1298 if (const Arg *A = Args.getLastArg(options::OPT_target_profile)) { 1299 StringRef TargetProfile = A->getValue(); 1300 if (auto Triple = 1301 toolchains::HLSLToolChain::parseTargetProfile(TargetProfile)) 1302 TargetTriple = *Triple; 1303 else 1304 Diag(diag::err_drv_invalid_directx_shader_module) << TargetProfile; 1305 1306 A->claim(); 1307 } else { 1308 Diag(diag::err_drv_dxc_missing_target_profile); 1309 } 1310 } 1311 1312 if (const Arg *A = Args.getLastArg(options::OPT_target)) 1313 TargetTriple = A->getValue(); 1314 if (const Arg *A = Args.getLastArg(options::OPT_ccc_install_dir)) 1315 Dir = InstalledDir = A->getValue(); 1316 for (const Arg *A : Args.filtered(options::OPT_B)) { 1317 A->claim(); 1318 PrefixDirs.push_back(A->getValue(0)); 1319 } 1320 if (std::optional<std::string> CompilerPathValue = 1321 llvm::sys::Process::GetEnv("COMPILER_PATH")) { 1322 StringRef CompilerPath = *CompilerPathValue; 1323 while (!CompilerPath.empty()) { 1324 std::pair<StringRef, StringRef> Split = 1325 CompilerPath.split(llvm::sys::EnvPathSeparator); 1326 PrefixDirs.push_back(std::string(Split.first)); 1327 CompilerPath = Split.second; 1328 } 1329 } 1330 if (const Arg *A = Args.getLastArg(options::OPT__sysroot_EQ)) 1331 SysRoot = A->getValue(); 1332 if (const Arg *A = Args.getLastArg(options::OPT__dyld_prefix_EQ)) 1333 DyldPrefix = A->getValue(); 1334 1335 if (const Arg *A = Args.getLastArg(options::OPT_resource_dir)) 1336 ResourceDir = A->getValue(); 1337 1338 if (const Arg *A = Args.getLastArg(options::OPT_save_temps_EQ)) { 1339 SaveTemps = llvm::StringSwitch<SaveTempsMode>(A->getValue()) 1340 .Case("cwd", SaveTempsCwd) 1341 .Case("obj", SaveTempsObj) 1342 .Default(SaveTempsCwd); 1343 } 1344 1345 if (const Arg *A = Args.getLastArg(options::OPT_offload_host_only, 1346 options::OPT_offload_device_only, 1347 options::OPT_offload_host_device)) { 1348 if (A->getOption().matches(options::OPT_offload_host_only)) 1349 Offload = OffloadHost; 1350 else if (A->getOption().matches(options::OPT_offload_device_only)) 1351 Offload = OffloadDevice; 1352 else 1353 Offload = OffloadHostDevice; 1354 } 1355 1356 setLTOMode(Args); 1357 1358 // Process -fembed-bitcode= flags. 1359 if (Arg *A = Args.getLastArg(options::OPT_fembed_bitcode_EQ)) { 1360 StringRef Name = A->getValue(); 1361 unsigned Model = llvm::StringSwitch<unsigned>(Name) 1362 .Case("off", EmbedNone) 1363 .Case("all", EmbedBitcode) 1364 .Case("bitcode", EmbedBitcode) 1365 .Case("marker", EmbedMarker) 1366 .Default(~0U); 1367 if (Model == ~0U) { 1368 Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) 1369 << Name; 1370 } else 1371 BitcodeEmbed = static_cast<BitcodeEmbedMode>(Model); 1372 } 1373 1374 // Remove existing compilation database so that each job can append to it. 1375 if (Arg *A = Args.getLastArg(options::OPT_MJ)) 1376 llvm::sys::fs::remove(A->getValue()); 1377 1378 // Setting up the jobs for some precompile cases depends on whether we are 1379 // treating them as PCH, implicit modules or C++20 ones. 1380 // TODO: inferring the mode like this seems fragile (it meets the objective 1381 // of not requiring anything new for operation, however). 1382 const Arg *Std = Args.getLastArg(options::OPT_std_EQ); 1383 ModulesModeCXX20 = 1384 !Args.hasArg(options::OPT_fmodules) && Std && 1385 (Std->containsValue("c++20") || Std->containsValue("c++2b") || 1386 Std->containsValue("c++2a") || Std->containsValue("c++latest")); 1387 1388 // Process -fmodule-header{=} flags. 1389 if (Arg *A = Args.getLastArg(options::OPT_fmodule_header_EQ, 1390 options::OPT_fmodule_header)) { 1391 // These flags force C++20 handling of headers. 1392 ModulesModeCXX20 = true; 1393 if (A->getOption().matches(options::OPT_fmodule_header)) 1394 CXX20HeaderType = HeaderMode_Default; 1395 else { 1396 StringRef ArgName = A->getValue(); 1397 unsigned Kind = llvm::StringSwitch<unsigned>(ArgName) 1398 .Case("user", HeaderMode_User) 1399 .Case("system", HeaderMode_System) 1400 .Default(~0U); 1401 if (Kind == ~0U) { 1402 Diags.Report(diag::err_drv_invalid_value) 1403 << A->getAsString(Args) << ArgName; 1404 } else 1405 CXX20HeaderType = static_cast<ModuleHeaderMode>(Kind); 1406 } 1407 } 1408 1409 std::unique_ptr<llvm::opt::InputArgList> UArgs = 1410 std::make_unique<InputArgList>(std::move(Args)); 1411 1412 // Perform the default argument translations. 1413 DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs); 1414 1415 // Owned by the host. 1416 const ToolChain &TC = getToolChain( 1417 *UArgs, computeTargetTriple(*this, TargetTriple, *UArgs)); 1418 1419 // Report warning when arm64EC option is overridden by specified target 1420 if ((TC.getTriple().getArch() != llvm::Triple::aarch64 || 1421 TC.getTriple().getSubArch() != llvm::Triple::AArch64SubArch_arm64ec) && 1422 UArgs->hasArg(options::OPT__SLASH_arm64EC)) { 1423 getDiags().Report(clang::diag::warn_target_override_arm64ec) 1424 << TC.getTriple().str(); 1425 } 1426 1427 // The compilation takes ownership of Args. 1428 Compilation *C = new Compilation(*this, TC, UArgs.release(), TranslatedArgs, 1429 ContainsError); 1430 1431 if (!HandleImmediateArgs(*C)) 1432 return C; 1433 1434 // Construct the list of inputs. 1435 InputList Inputs; 1436 BuildInputs(C->getDefaultToolChain(), *TranslatedArgs, Inputs); 1437 1438 // Populate the tool chains for the offloading devices, if any. 1439 CreateOffloadingDeviceToolChains(*C, Inputs); 1440 1441 // Construct the list of abstract actions to perform for this compilation. On 1442 // MachO targets this uses the driver-driver and universal actions. 1443 if (TC.getTriple().isOSBinFormatMachO()) 1444 BuildUniversalActions(*C, C->getDefaultToolChain(), Inputs); 1445 else 1446 BuildActions(*C, C->getArgs(), Inputs, C->getActions()); 1447 1448 if (CCCPrintPhases) { 1449 PrintActions(*C); 1450 return C; 1451 } 1452 1453 BuildJobs(*C); 1454 1455 return C; 1456 } 1457 1458 static void printArgList(raw_ostream &OS, const llvm::opt::ArgList &Args) { 1459 llvm::opt::ArgStringList ASL; 1460 for (const auto *A : Args) { 1461 // Use user's original spelling of flags. For example, use 1462 // `/source-charset:utf-8` instead of `-finput-charset=utf-8` if the user 1463 // wrote the former. 1464 while (A->getAlias()) 1465 A = A->getAlias(); 1466 A->render(Args, ASL); 1467 } 1468 1469 for (auto I = ASL.begin(), E = ASL.end(); I != E; ++I) { 1470 if (I != ASL.begin()) 1471 OS << ' '; 1472 llvm::sys::printArg(OS, *I, true); 1473 } 1474 OS << '\n'; 1475 } 1476 1477 bool Driver::getCrashDiagnosticFile(StringRef ReproCrashFilename, 1478 SmallString<128> &CrashDiagDir) { 1479 using namespace llvm::sys; 1480 assert(llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin() && 1481 "Only knows about .crash files on Darwin"); 1482 1483 // The .crash file can be found on at ~/Library/Logs/DiagnosticReports/ 1484 // (or /Library/Logs/DiagnosticReports for root) and has the filename pattern 1485 // clang-<VERSION>_<YYYY-MM-DD-HHMMSS>_<hostname>.crash. 1486 path::home_directory(CrashDiagDir); 1487 if (CrashDiagDir.startswith("/var/root")) 1488 CrashDiagDir = "/"; 1489 path::append(CrashDiagDir, "Library/Logs/DiagnosticReports"); 1490 int PID = 1491 #if LLVM_ON_UNIX 1492 getpid(); 1493 #else 1494 0; 1495 #endif 1496 std::error_code EC; 1497 fs::file_status FileStatus; 1498 TimePoint<> LastAccessTime; 1499 SmallString<128> CrashFilePath; 1500 // Lookup the .crash files and get the one generated by a subprocess spawned 1501 // by this driver invocation. 1502 for (fs::directory_iterator File(CrashDiagDir, EC), FileEnd; 1503 File != FileEnd && !EC; File.increment(EC)) { 1504 StringRef FileName = path::filename(File->path()); 1505 if (!FileName.startswith(Name)) 1506 continue; 1507 if (fs::status(File->path(), FileStatus)) 1508 continue; 1509 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CrashFile = 1510 llvm::MemoryBuffer::getFile(File->path()); 1511 if (!CrashFile) 1512 continue; 1513 // The first line should start with "Process:", otherwise this isn't a real 1514 // .crash file. 1515 StringRef Data = CrashFile.get()->getBuffer(); 1516 if (!Data.startswith("Process:")) 1517 continue; 1518 // Parse parent process pid line, e.g: "Parent Process: clang-4.0 [79141]" 1519 size_t ParentProcPos = Data.find("Parent Process:"); 1520 if (ParentProcPos == StringRef::npos) 1521 continue; 1522 size_t LineEnd = Data.find_first_of("\n", ParentProcPos); 1523 if (LineEnd == StringRef::npos) 1524 continue; 1525 StringRef ParentProcess = Data.slice(ParentProcPos+15, LineEnd).trim(); 1526 int OpenBracket = -1, CloseBracket = -1; 1527 for (size_t i = 0, e = ParentProcess.size(); i < e; ++i) { 1528 if (ParentProcess[i] == '[') 1529 OpenBracket = i; 1530 if (ParentProcess[i] == ']') 1531 CloseBracket = i; 1532 } 1533 // Extract the parent process PID from the .crash file and check whether 1534 // it matches this driver invocation pid. 1535 int CrashPID; 1536 if (OpenBracket < 0 || CloseBracket < 0 || 1537 ParentProcess.slice(OpenBracket + 1, CloseBracket) 1538 .getAsInteger(10, CrashPID) || CrashPID != PID) { 1539 continue; 1540 } 1541 1542 // Found a .crash file matching the driver pid. To avoid getting an older 1543 // and misleading crash file, continue looking for the most recent. 1544 // FIXME: the driver can dispatch multiple cc1 invocations, leading to 1545 // multiple crashes poiting to the same parent process. Since the driver 1546 // does not collect pid information for the dispatched invocation there's 1547 // currently no way to distinguish among them. 1548 const auto FileAccessTime = FileStatus.getLastModificationTime(); 1549 if (FileAccessTime > LastAccessTime) { 1550 CrashFilePath.assign(File->path()); 1551 LastAccessTime = FileAccessTime; 1552 } 1553 } 1554 1555 // If found, copy it over to the location of other reproducer files. 1556 if (!CrashFilePath.empty()) { 1557 EC = fs::copy_file(CrashFilePath, ReproCrashFilename); 1558 if (EC) 1559 return false; 1560 return true; 1561 } 1562 1563 return false; 1564 } 1565 1566 static const char BugReporMsg[] = 1567 "\n********************\n\n" 1568 "PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:\n" 1569 "Preprocessed source(s) and associated run script(s) are located at:"; 1570 1571 // When clang crashes, produce diagnostic information including the fully 1572 // preprocessed source file(s). Request that the developer attach the 1573 // diagnostic information to a bug report. 1574 void Driver::generateCompilationDiagnostics( 1575 Compilation &C, const Command &FailingCommand, 1576 StringRef AdditionalInformation, CompilationDiagnosticReport *Report) { 1577 if (C.getArgs().hasArg(options::OPT_fno_crash_diagnostics)) 1578 return; 1579 1580 unsigned Level = 1; 1581 if (Arg *A = C.getArgs().getLastArg(options::OPT_fcrash_diagnostics_EQ)) { 1582 Level = llvm::StringSwitch<unsigned>(A->getValue()) 1583 .Case("off", 0) 1584 .Case("compiler", 1) 1585 .Case("all", 2) 1586 .Default(1); 1587 } 1588 if (!Level) 1589 return; 1590 1591 // Don't try to generate diagnostics for dsymutil jobs. 1592 if (FailingCommand.getCreator().isDsymutilJob()) 1593 return; 1594 1595 bool IsLLD = false; 1596 ArgStringList SavedTemps; 1597 if (FailingCommand.getCreator().isLinkJob()) { 1598 C.getDefaultToolChain().GetLinkerPath(&IsLLD); 1599 if (!IsLLD || Level < 2) 1600 return; 1601 1602 // If lld crashed, we will re-run the same command with the input it used 1603 // to have. In that case we should not remove temp files in 1604 // initCompilationForDiagnostics yet. They will be added back and removed 1605 // later. 1606 SavedTemps = std::move(C.getTempFiles()); 1607 assert(!C.getTempFiles().size()); 1608 } 1609 1610 // Print the version of the compiler. 1611 PrintVersion(C, llvm::errs()); 1612 1613 // Suppress driver output and emit preprocessor output to temp file. 1614 CCGenDiagnostics = true; 1615 1616 // Save the original job command(s). 1617 Command Cmd = FailingCommand; 1618 1619 // Keep track of whether we produce any errors while trying to produce 1620 // preprocessed sources. 1621 DiagnosticErrorTrap Trap(Diags); 1622 1623 // Suppress tool output. 1624 C.initCompilationForDiagnostics(); 1625 1626 // If lld failed, rerun it again with --reproduce. 1627 if (IsLLD) { 1628 const char *TmpName = CreateTempFile(C, "linker-crash", "tar"); 1629 Command NewLLDInvocation = Cmd; 1630 llvm::opt::ArgStringList ArgList = NewLLDInvocation.getArguments(); 1631 StringRef ReproduceOption = 1632 C.getDefaultToolChain().getTriple().isWindowsMSVCEnvironment() 1633 ? "/reproduce:" 1634 : "--reproduce="; 1635 ArgList.push_back(Saver.save(Twine(ReproduceOption) + TmpName).data()); 1636 NewLLDInvocation.replaceArguments(std::move(ArgList)); 1637 1638 // Redirect stdout/stderr to /dev/null. 1639 NewLLDInvocation.Execute({std::nullopt, {""}, {""}}, nullptr, nullptr); 1640 Diag(clang::diag::note_drv_command_failed_diag_msg) << BugReporMsg; 1641 Diag(clang::diag::note_drv_command_failed_diag_msg) << TmpName; 1642 Diag(clang::diag::note_drv_command_failed_diag_msg) 1643 << "\n\n********************"; 1644 if (Report) 1645 Report->TemporaryFiles.push_back(TmpName); 1646 return; 1647 } 1648 1649 // Construct the list of inputs. 1650 InputList Inputs; 1651 BuildInputs(C.getDefaultToolChain(), C.getArgs(), Inputs); 1652 1653 for (InputList::iterator it = Inputs.begin(), ie = Inputs.end(); it != ie;) { 1654 bool IgnoreInput = false; 1655 1656 // Ignore input from stdin or any inputs that cannot be preprocessed. 1657 // Check type first as not all linker inputs have a value. 1658 if (types::getPreprocessedType(it->first) == types::TY_INVALID) { 1659 IgnoreInput = true; 1660 } else if (!strcmp(it->second->getValue(), "-")) { 1661 Diag(clang::diag::note_drv_command_failed_diag_msg) 1662 << "Error generating preprocessed source(s) - " 1663 "ignoring input from stdin."; 1664 IgnoreInput = true; 1665 } 1666 1667 if (IgnoreInput) { 1668 it = Inputs.erase(it); 1669 ie = Inputs.end(); 1670 } else { 1671 ++it; 1672 } 1673 } 1674 1675 if (Inputs.empty()) { 1676 Diag(clang::diag::note_drv_command_failed_diag_msg) 1677 << "Error generating preprocessed source(s) - " 1678 "no preprocessable inputs."; 1679 return; 1680 } 1681 1682 // Don't attempt to generate preprocessed files if multiple -arch options are 1683 // used, unless they're all duplicates. 1684 llvm::StringSet<> ArchNames; 1685 for (const Arg *A : C.getArgs()) { 1686 if (A->getOption().matches(options::OPT_arch)) { 1687 StringRef ArchName = A->getValue(); 1688 ArchNames.insert(ArchName); 1689 } 1690 } 1691 if (ArchNames.size() > 1) { 1692 Diag(clang::diag::note_drv_command_failed_diag_msg) 1693 << "Error generating preprocessed source(s) - cannot generate " 1694 "preprocessed source with multiple -arch options."; 1695 return; 1696 } 1697 1698 // Construct the list of abstract actions to perform for this compilation. On 1699 // Darwin OSes this uses the driver-driver and builds universal actions. 1700 const ToolChain &TC = C.getDefaultToolChain(); 1701 if (TC.getTriple().isOSBinFormatMachO()) 1702 BuildUniversalActions(C, TC, Inputs); 1703 else 1704 BuildActions(C, C.getArgs(), Inputs, C.getActions()); 1705 1706 BuildJobs(C); 1707 1708 // If there were errors building the compilation, quit now. 1709 if (Trap.hasErrorOccurred()) { 1710 Diag(clang::diag::note_drv_command_failed_diag_msg) 1711 << "Error generating preprocessed source(s)."; 1712 return; 1713 } 1714 1715 // Generate preprocessed output. 1716 SmallVector<std::pair<int, const Command *>, 4> FailingCommands; 1717 C.ExecuteJobs(C.getJobs(), FailingCommands); 1718 1719 // If any of the preprocessing commands failed, clean up and exit. 1720 if (!FailingCommands.empty()) { 1721 Diag(clang::diag::note_drv_command_failed_diag_msg) 1722 << "Error generating preprocessed source(s)."; 1723 return; 1724 } 1725 1726 const ArgStringList &TempFiles = C.getTempFiles(); 1727 if (TempFiles.empty()) { 1728 Diag(clang::diag::note_drv_command_failed_diag_msg) 1729 << "Error generating preprocessed source(s)."; 1730 return; 1731 } 1732 1733 Diag(clang::diag::note_drv_command_failed_diag_msg) << BugReporMsg; 1734 1735 SmallString<128> VFS; 1736 SmallString<128> ReproCrashFilename; 1737 for (const char *TempFile : TempFiles) { 1738 Diag(clang::diag::note_drv_command_failed_diag_msg) << TempFile; 1739 if (Report) 1740 Report->TemporaryFiles.push_back(TempFile); 1741 if (ReproCrashFilename.empty()) { 1742 ReproCrashFilename = TempFile; 1743 llvm::sys::path::replace_extension(ReproCrashFilename, ".crash"); 1744 } 1745 if (StringRef(TempFile).endswith(".cache")) { 1746 // In some cases (modules) we'll dump extra data to help with reproducing 1747 // the crash into a directory next to the output. 1748 VFS = llvm::sys::path::filename(TempFile); 1749 llvm::sys::path::append(VFS, "vfs", "vfs.yaml"); 1750 } 1751 } 1752 1753 for (const char *TempFile : SavedTemps) 1754 C.addTempFile(TempFile); 1755 1756 // Assume associated files are based off of the first temporary file. 1757 CrashReportInfo CrashInfo(TempFiles[0], VFS); 1758 1759 llvm::SmallString<128> Script(CrashInfo.Filename); 1760 llvm::sys::path::replace_extension(Script, "sh"); 1761 std::error_code EC; 1762 llvm::raw_fd_ostream ScriptOS(Script, EC, llvm::sys::fs::CD_CreateNew, 1763 llvm::sys::fs::FA_Write, 1764 llvm::sys::fs::OF_Text); 1765 if (EC) { 1766 Diag(clang::diag::note_drv_command_failed_diag_msg) 1767 << "Error generating run script: " << Script << " " << EC.message(); 1768 } else { 1769 ScriptOS << "# Crash reproducer for " << getClangFullVersion() << "\n" 1770 << "# Driver args: "; 1771 printArgList(ScriptOS, C.getInputArgs()); 1772 ScriptOS << "# Original command: "; 1773 Cmd.Print(ScriptOS, "\n", /*Quote=*/true); 1774 Cmd.Print(ScriptOS, "\n", /*Quote=*/true, &CrashInfo); 1775 if (!AdditionalInformation.empty()) 1776 ScriptOS << "\n# Additional information: " << AdditionalInformation 1777 << "\n"; 1778 if (Report) 1779 Report->TemporaryFiles.push_back(std::string(Script.str())); 1780 Diag(clang::diag::note_drv_command_failed_diag_msg) << Script; 1781 } 1782 1783 // On darwin, provide information about the .crash diagnostic report. 1784 if (llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin()) { 1785 SmallString<128> CrashDiagDir; 1786 if (getCrashDiagnosticFile(ReproCrashFilename, CrashDiagDir)) { 1787 Diag(clang::diag::note_drv_command_failed_diag_msg) 1788 << ReproCrashFilename.str(); 1789 } else { // Suggest a directory for the user to look for .crash files. 1790 llvm::sys::path::append(CrashDiagDir, Name); 1791 CrashDiagDir += "_<YYYY-MM-DD-HHMMSS>_<hostname>.crash"; 1792 Diag(clang::diag::note_drv_command_failed_diag_msg) 1793 << "Crash backtrace is located in"; 1794 Diag(clang::diag::note_drv_command_failed_diag_msg) 1795 << CrashDiagDir.str(); 1796 Diag(clang::diag::note_drv_command_failed_diag_msg) 1797 << "(choose the .crash file that corresponds to your crash)"; 1798 } 1799 } 1800 1801 for (const auto &A : C.getArgs().filtered(options::OPT_frewrite_map_file_EQ)) 1802 Diag(clang::diag::note_drv_command_failed_diag_msg) << A->getValue(); 1803 1804 Diag(clang::diag::note_drv_command_failed_diag_msg) 1805 << "\n\n********************"; 1806 } 1807 1808 void Driver::setUpResponseFiles(Compilation &C, Command &Cmd) { 1809 // Since commandLineFitsWithinSystemLimits() may underestimate system's 1810 // capacity if the tool does not support response files, there is a chance/ 1811 // that things will just work without a response file, so we silently just 1812 // skip it. 1813 if (Cmd.getResponseFileSupport().ResponseKind == 1814 ResponseFileSupport::RF_None || 1815 llvm::sys::commandLineFitsWithinSystemLimits(Cmd.getExecutable(), 1816 Cmd.getArguments())) 1817 return; 1818 1819 std::string TmpName = GetTemporaryPath("response", "txt"); 1820 Cmd.setResponseFile(C.addTempFile(C.getArgs().MakeArgString(TmpName))); 1821 } 1822 1823 int Driver::ExecuteCompilation( 1824 Compilation &C, 1825 SmallVectorImpl<std::pair<int, const Command *>> &FailingCommands) { 1826 if (C.getArgs().hasArg(options::OPT_fdriver_only)) { 1827 if (C.getArgs().hasArg(options::OPT_v)) 1828 C.getJobs().Print(llvm::errs(), "\n", true); 1829 1830 C.ExecuteJobs(C.getJobs(), FailingCommands, /*LogOnly=*/true); 1831 1832 // If there were errors building the compilation, quit now. 1833 if (!FailingCommands.empty() || Diags.hasErrorOccurred()) 1834 return 1; 1835 1836 return 0; 1837 } 1838 1839 // Just print if -### was present. 1840 if (C.getArgs().hasArg(options::OPT__HASH_HASH_HASH)) { 1841 C.getJobs().Print(llvm::errs(), "\n", true); 1842 return 0; 1843 } 1844 1845 // If there were errors building the compilation, quit now. 1846 if (Diags.hasErrorOccurred()) 1847 return 1; 1848 1849 // Set up response file names for each command, if necessary. 1850 for (auto &Job : C.getJobs()) 1851 setUpResponseFiles(C, Job); 1852 1853 C.ExecuteJobs(C.getJobs(), FailingCommands); 1854 1855 // If the command succeeded, we are done. 1856 if (FailingCommands.empty()) 1857 return 0; 1858 1859 // Otherwise, remove result files and print extra information about abnormal 1860 // failures. 1861 int Res = 0; 1862 for (const auto &CmdPair : FailingCommands) { 1863 int CommandRes = CmdPair.first; 1864 const Command *FailingCommand = CmdPair.second; 1865 1866 // Remove result files if we're not saving temps. 1867 if (!isSaveTempsEnabled()) { 1868 const JobAction *JA = cast<JobAction>(&FailingCommand->getSource()); 1869 C.CleanupFileMap(C.getResultFiles(), JA, true); 1870 1871 // Failure result files are valid unless we crashed. 1872 if (CommandRes < 0) 1873 C.CleanupFileMap(C.getFailureResultFiles(), JA, true); 1874 } 1875 1876 // llvm/lib/Support/*/Signals.inc will exit with a special return code 1877 // for SIGPIPE. Do not print diagnostics for this case. 1878 if (CommandRes == EX_IOERR) { 1879 Res = CommandRes; 1880 continue; 1881 } 1882 1883 // Print extra information about abnormal failures, if possible. 1884 // 1885 // This is ad-hoc, but we don't want to be excessively noisy. If the result 1886 // status was 1, assume the command failed normally. In particular, if it 1887 // was the compiler then assume it gave a reasonable error code. Failures 1888 // in other tools are less common, and they generally have worse 1889 // diagnostics, so always print the diagnostic there. 1890 const Tool &FailingTool = FailingCommand->getCreator(); 1891 1892 if (!FailingCommand->getCreator().hasGoodDiagnostics() || CommandRes != 1) { 1893 // FIXME: See FIXME above regarding result code interpretation. 1894 if (CommandRes < 0) 1895 Diag(clang::diag::err_drv_command_signalled) 1896 << FailingTool.getShortName(); 1897 else 1898 Diag(clang::diag::err_drv_command_failed) 1899 << FailingTool.getShortName() << CommandRes; 1900 } 1901 } 1902 return Res; 1903 } 1904 1905 void Driver::PrintHelp(bool ShowHidden) const { 1906 unsigned IncludedFlagsBitmask; 1907 unsigned ExcludedFlagsBitmask; 1908 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 1909 getIncludeExcludeOptionFlagMasks(IsCLMode()); 1910 1911 ExcludedFlagsBitmask |= options::NoDriverOption; 1912 if (!ShowHidden) 1913 ExcludedFlagsBitmask |= HelpHidden; 1914 1915 if (IsFlangMode()) 1916 IncludedFlagsBitmask |= options::FlangOption; 1917 else 1918 ExcludedFlagsBitmask |= options::FlangOnlyOption; 1919 1920 std::string Usage = llvm::formatv("{0} [options] file...", Name).str(); 1921 getOpts().printHelp(llvm::outs(), Usage.c_str(), DriverTitle.c_str(), 1922 IncludedFlagsBitmask, ExcludedFlagsBitmask, 1923 /*ShowAllAliases=*/false); 1924 } 1925 1926 void Driver::PrintVersion(const Compilation &C, raw_ostream &OS) const { 1927 if (IsFlangMode()) { 1928 OS << getClangToolFullVersion("flang-new") << '\n'; 1929 } else { 1930 // FIXME: The following handlers should use a callback mechanism, we don't 1931 // know what the client would like to do. 1932 OS << getClangFullVersion() << '\n'; 1933 } 1934 const ToolChain &TC = C.getDefaultToolChain(); 1935 OS << "Target: " << TC.getTripleString() << '\n'; 1936 1937 // Print the threading model. 1938 if (Arg *A = C.getArgs().getLastArg(options::OPT_mthread_model)) { 1939 // Don't print if the ToolChain would have barfed on it already 1940 if (TC.isThreadModelSupported(A->getValue())) 1941 OS << "Thread model: " << A->getValue(); 1942 } else 1943 OS << "Thread model: " << TC.getThreadModel(); 1944 OS << '\n'; 1945 1946 // Print out the install directory. 1947 OS << "InstalledDir: " << InstalledDir << '\n'; 1948 1949 // If configuration files were used, print their paths. 1950 for (auto ConfigFile : ConfigFiles) 1951 OS << "Configuration file: " << ConfigFile << '\n'; 1952 } 1953 1954 /// PrintDiagnosticCategories - Implement the --print-diagnostic-categories 1955 /// option. 1956 static void PrintDiagnosticCategories(raw_ostream &OS) { 1957 // Skip the empty category. 1958 for (unsigned i = 1, max = DiagnosticIDs::getNumberOfCategories(); i != max; 1959 ++i) 1960 OS << i << ',' << DiagnosticIDs::getCategoryNameFromID(i) << '\n'; 1961 } 1962 1963 void Driver::HandleAutocompletions(StringRef PassedFlags) const { 1964 if (PassedFlags == "") 1965 return; 1966 // Print out all options that start with a given argument. This is used for 1967 // shell autocompletion. 1968 std::vector<std::string> SuggestedCompletions; 1969 std::vector<std::string> Flags; 1970 1971 unsigned int DisableFlags = 1972 options::NoDriverOption | options::Unsupported | options::Ignored; 1973 1974 // Make sure that Flang-only options don't pollute the Clang output 1975 // TODO: Make sure that Clang-only options don't pollute Flang output 1976 if (!IsFlangMode()) 1977 DisableFlags |= options::FlangOnlyOption; 1978 1979 // Distinguish "--autocomplete=-someflag" and "--autocomplete=-someflag," 1980 // because the latter indicates that the user put space before pushing tab 1981 // which should end up in a file completion. 1982 const bool HasSpace = PassedFlags.endswith(","); 1983 1984 // Parse PassedFlags by "," as all the command-line flags are passed to this 1985 // function separated by "," 1986 StringRef TargetFlags = PassedFlags; 1987 while (TargetFlags != "") { 1988 StringRef CurFlag; 1989 std::tie(CurFlag, TargetFlags) = TargetFlags.split(","); 1990 Flags.push_back(std::string(CurFlag)); 1991 } 1992 1993 // We want to show cc1-only options only when clang is invoked with -cc1 or 1994 // -Xclang. 1995 if (llvm::is_contained(Flags, "-Xclang") || llvm::is_contained(Flags, "-cc1")) 1996 DisableFlags &= ~options::NoDriverOption; 1997 1998 const llvm::opt::OptTable &Opts = getOpts(); 1999 StringRef Cur; 2000 Cur = Flags.at(Flags.size() - 1); 2001 StringRef Prev; 2002 if (Flags.size() >= 2) { 2003 Prev = Flags.at(Flags.size() - 2); 2004 SuggestedCompletions = Opts.suggestValueCompletions(Prev, Cur); 2005 } 2006 2007 if (SuggestedCompletions.empty()) 2008 SuggestedCompletions = Opts.suggestValueCompletions(Cur, ""); 2009 2010 // If Flags were empty, it means the user typed `clang [tab]` where we should 2011 // list all possible flags. If there was no value completion and the user 2012 // pressed tab after a space, we should fall back to a file completion. 2013 // We're printing a newline to be consistent with what we print at the end of 2014 // this function. 2015 if (SuggestedCompletions.empty() && HasSpace && !Flags.empty()) { 2016 llvm::outs() << '\n'; 2017 return; 2018 } 2019 2020 // When flag ends with '=' and there was no value completion, return empty 2021 // string and fall back to the file autocompletion. 2022 if (SuggestedCompletions.empty() && !Cur.endswith("=")) { 2023 // If the flag is in the form of "--autocomplete=-foo", 2024 // we were requested to print out all option names that start with "-foo". 2025 // For example, "--autocomplete=-fsyn" is expanded to "-fsyntax-only". 2026 SuggestedCompletions = Opts.findByPrefix(Cur, DisableFlags); 2027 2028 // We have to query the -W flags manually as they're not in the OptTable. 2029 // TODO: Find a good way to add them to OptTable instead and them remove 2030 // this code. 2031 for (StringRef S : DiagnosticIDs::getDiagnosticFlags()) 2032 if (S.startswith(Cur)) 2033 SuggestedCompletions.push_back(std::string(S)); 2034 } 2035 2036 // Sort the autocomplete candidates so that shells print them out in a 2037 // deterministic order. We could sort in any way, but we chose 2038 // case-insensitive sorting for consistency with the -help option 2039 // which prints out options in the case-insensitive alphabetical order. 2040 llvm::sort(SuggestedCompletions, [](StringRef A, StringRef B) { 2041 if (int X = A.compare_insensitive(B)) 2042 return X < 0; 2043 return A.compare(B) > 0; 2044 }); 2045 2046 llvm::outs() << llvm::join(SuggestedCompletions, "\n") << '\n'; 2047 } 2048 2049 bool Driver::HandleImmediateArgs(const Compilation &C) { 2050 // The order these options are handled in gcc is all over the place, but we 2051 // don't expect inconsistencies w.r.t. that to matter in practice. 2052 2053 if (C.getArgs().hasArg(options::OPT_dumpmachine)) { 2054 llvm::outs() << C.getDefaultToolChain().getTripleString() << '\n'; 2055 return false; 2056 } 2057 2058 if (C.getArgs().hasArg(options::OPT_dumpversion)) { 2059 // Since -dumpversion is only implemented for pedantic GCC compatibility, we 2060 // return an answer which matches our definition of __VERSION__. 2061 llvm::outs() << CLANG_VERSION_STRING << "\n"; 2062 return false; 2063 } 2064 2065 if (C.getArgs().hasArg(options::OPT__print_diagnostic_categories)) { 2066 PrintDiagnosticCategories(llvm::outs()); 2067 return false; 2068 } 2069 2070 if (C.getArgs().hasArg(options::OPT_help) || 2071 C.getArgs().hasArg(options::OPT__help_hidden)) { 2072 PrintHelp(C.getArgs().hasArg(options::OPT__help_hidden)); 2073 return false; 2074 } 2075 2076 if (C.getArgs().hasArg(options::OPT__version)) { 2077 // Follow gcc behavior and use stdout for --version and stderr for -v. 2078 PrintVersion(C, llvm::outs()); 2079 return false; 2080 } 2081 2082 if (C.getArgs().hasArg(options::OPT_v) || 2083 C.getArgs().hasArg(options::OPT__HASH_HASH_HASH) || 2084 C.getArgs().hasArg(options::OPT_print_supported_cpus)) { 2085 PrintVersion(C, llvm::errs()); 2086 SuppressMissingInputWarning = true; 2087 } 2088 2089 if (C.getArgs().hasArg(options::OPT_v)) { 2090 if (!SystemConfigDir.empty()) 2091 llvm::errs() << "System configuration file directory: " 2092 << SystemConfigDir << "\n"; 2093 if (!UserConfigDir.empty()) 2094 llvm::errs() << "User configuration file directory: " 2095 << UserConfigDir << "\n"; 2096 } 2097 2098 const ToolChain &TC = C.getDefaultToolChain(); 2099 2100 if (C.getArgs().hasArg(options::OPT_v)) 2101 TC.printVerboseInfo(llvm::errs()); 2102 2103 if (C.getArgs().hasArg(options::OPT_print_resource_dir)) { 2104 llvm::outs() << ResourceDir << '\n'; 2105 return false; 2106 } 2107 2108 if (C.getArgs().hasArg(options::OPT_print_search_dirs)) { 2109 llvm::outs() << "programs: ="; 2110 bool separator = false; 2111 // Print -B and COMPILER_PATH. 2112 for (const std::string &Path : PrefixDirs) { 2113 if (separator) 2114 llvm::outs() << llvm::sys::EnvPathSeparator; 2115 llvm::outs() << Path; 2116 separator = true; 2117 } 2118 for (const std::string &Path : TC.getProgramPaths()) { 2119 if (separator) 2120 llvm::outs() << llvm::sys::EnvPathSeparator; 2121 llvm::outs() << Path; 2122 separator = true; 2123 } 2124 llvm::outs() << "\n"; 2125 llvm::outs() << "libraries: =" << ResourceDir; 2126 2127 StringRef sysroot = C.getSysRoot(); 2128 2129 for (const std::string &Path : TC.getFilePaths()) { 2130 // Always print a separator. ResourceDir was the first item shown. 2131 llvm::outs() << llvm::sys::EnvPathSeparator; 2132 // Interpretation of leading '=' is needed only for NetBSD. 2133 if (Path[0] == '=') 2134 llvm::outs() << sysroot << Path.substr(1); 2135 else 2136 llvm::outs() << Path; 2137 } 2138 llvm::outs() << "\n"; 2139 return false; 2140 } 2141 2142 if (C.getArgs().hasArg(options::OPT_print_runtime_dir)) { 2143 std::string RuntimePath; 2144 // Get the first existing path, if any. 2145 for (auto Path : TC.getRuntimePaths()) { 2146 if (getVFS().exists(Path)) { 2147 RuntimePath = Path; 2148 break; 2149 } 2150 } 2151 if (!RuntimePath.empty()) 2152 llvm::outs() << RuntimePath << '\n'; 2153 else 2154 llvm::outs() << TC.getCompilerRTPath() << '\n'; 2155 return false; 2156 } 2157 2158 if (C.getArgs().hasArg(options::OPT_print_diagnostic_options)) { 2159 std::vector<std::string> Flags = DiagnosticIDs::getDiagnosticFlags(); 2160 for (std::size_t I = 0; I != Flags.size(); I += 2) 2161 llvm::outs() << " " << Flags[I] << "\n " << Flags[I + 1] << "\n\n"; 2162 return false; 2163 } 2164 2165 // FIXME: The following handlers should use a callback mechanism, we don't 2166 // know what the client would like to do. 2167 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_file_name_EQ)) { 2168 llvm::outs() << GetFilePath(A->getValue(), TC) << "\n"; 2169 return false; 2170 } 2171 2172 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_prog_name_EQ)) { 2173 StringRef ProgName = A->getValue(); 2174 2175 // Null program name cannot have a path. 2176 if (! ProgName.empty()) 2177 llvm::outs() << GetProgramPath(ProgName, TC); 2178 2179 llvm::outs() << "\n"; 2180 return false; 2181 } 2182 2183 if (Arg *A = C.getArgs().getLastArg(options::OPT_autocomplete)) { 2184 StringRef PassedFlags = A->getValue(); 2185 HandleAutocompletions(PassedFlags); 2186 return false; 2187 } 2188 2189 if (C.getArgs().hasArg(options::OPT_print_libgcc_file_name)) { 2190 ToolChain::RuntimeLibType RLT = TC.GetRuntimeLibType(C.getArgs()); 2191 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2192 RegisterEffectiveTriple TripleRAII(TC, Triple); 2193 switch (RLT) { 2194 case ToolChain::RLT_CompilerRT: 2195 llvm::outs() << TC.getCompilerRT(C.getArgs(), "builtins") << "\n"; 2196 break; 2197 case ToolChain::RLT_Libgcc: 2198 llvm::outs() << GetFilePath("libgcc.a", TC) << "\n"; 2199 break; 2200 } 2201 return false; 2202 } 2203 2204 if (C.getArgs().hasArg(options::OPT_print_multi_lib)) { 2205 for (const Multilib &Multilib : TC.getMultilibs()) 2206 llvm::outs() << Multilib << "\n"; 2207 return false; 2208 } 2209 2210 if (C.getArgs().hasArg(options::OPT_print_multi_directory)) { 2211 const Multilib &Multilib = TC.getMultilib(); 2212 if (Multilib.gccSuffix().empty()) 2213 llvm::outs() << ".\n"; 2214 else { 2215 StringRef Suffix(Multilib.gccSuffix()); 2216 assert(Suffix.front() == '/'); 2217 llvm::outs() << Suffix.substr(1) << "\n"; 2218 } 2219 return false; 2220 } 2221 2222 if (C.getArgs().hasArg(options::OPT_print_target_triple)) { 2223 llvm::outs() << TC.getTripleString() << "\n"; 2224 return false; 2225 } 2226 2227 if (C.getArgs().hasArg(options::OPT_print_effective_triple)) { 2228 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2229 llvm::outs() << Triple.getTriple() << "\n"; 2230 return false; 2231 } 2232 2233 if (C.getArgs().hasArg(options::OPT_print_targets)) { 2234 llvm::TargetRegistry::printRegisteredTargetsForVersion(llvm::outs()); 2235 return false; 2236 } 2237 2238 return true; 2239 } 2240 2241 enum { 2242 TopLevelAction = 0, 2243 HeadSibAction = 1, 2244 OtherSibAction = 2, 2245 }; 2246 2247 // Display an action graph human-readably. Action A is the "sink" node 2248 // and latest-occuring action. Traversal is in pre-order, visiting the 2249 // inputs to each action before printing the action itself. 2250 static unsigned PrintActions1(const Compilation &C, Action *A, 2251 std::map<Action *, unsigned> &Ids, 2252 Twine Indent = {}, int Kind = TopLevelAction) { 2253 if (Ids.count(A)) // A was already visited. 2254 return Ids[A]; 2255 2256 std::string str; 2257 llvm::raw_string_ostream os(str); 2258 2259 auto getSibIndent = [](int K) -> Twine { 2260 return (K == HeadSibAction) ? " " : (K == OtherSibAction) ? "| " : ""; 2261 }; 2262 2263 Twine SibIndent = Indent + getSibIndent(Kind); 2264 int SibKind = HeadSibAction; 2265 os << Action::getClassName(A->getKind()) << ", "; 2266 if (InputAction *IA = dyn_cast<InputAction>(A)) { 2267 os << "\"" << IA->getInputArg().getValue() << "\""; 2268 } else if (BindArchAction *BIA = dyn_cast<BindArchAction>(A)) { 2269 os << '"' << BIA->getArchName() << '"' << ", {" 2270 << PrintActions1(C, *BIA->input_begin(), Ids, SibIndent, SibKind) << "}"; 2271 } else if (OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 2272 bool IsFirst = true; 2273 OA->doOnEachDependence( 2274 [&](Action *A, const ToolChain *TC, const char *BoundArch) { 2275 assert(TC && "Unknown host toolchain"); 2276 // E.g. for two CUDA device dependences whose bound arch is sm_20 and 2277 // sm_35 this will generate: 2278 // "cuda-device" (nvptx64-nvidia-cuda:sm_20) {#ID}, "cuda-device" 2279 // (nvptx64-nvidia-cuda:sm_35) {#ID} 2280 if (!IsFirst) 2281 os << ", "; 2282 os << '"'; 2283 os << A->getOffloadingKindPrefix(); 2284 os << " ("; 2285 os << TC->getTriple().normalize(); 2286 if (BoundArch) 2287 os << ":" << BoundArch; 2288 os << ")"; 2289 os << '"'; 2290 os << " {" << PrintActions1(C, A, Ids, SibIndent, SibKind) << "}"; 2291 IsFirst = false; 2292 SibKind = OtherSibAction; 2293 }); 2294 } else { 2295 const ActionList *AL = &A->getInputs(); 2296 2297 if (AL->size()) { 2298 const char *Prefix = "{"; 2299 for (Action *PreRequisite : *AL) { 2300 os << Prefix << PrintActions1(C, PreRequisite, Ids, SibIndent, SibKind); 2301 Prefix = ", "; 2302 SibKind = OtherSibAction; 2303 } 2304 os << "}"; 2305 } else 2306 os << "{}"; 2307 } 2308 2309 // Append offload info for all options other than the offloading action 2310 // itself (e.g. (cuda-device, sm_20) or (cuda-host)). 2311 std::string offload_str; 2312 llvm::raw_string_ostream offload_os(offload_str); 2313 if (!isa<OffloadAction>(A)) { 2314 auto S = A->getOffloadingKindPrefix(); 2315 if (!S.empty()) { 2316 offload_os << ", (" << S; 2317 if (A->getOffloadingArch()) 2318 offload_os << ", " << A->getOffloadingArch(); 2319 offload_os << ")"; 2320 } 2321 } 2322 2323 auto getSelfIndent = [](int K) -> Twine { 2324 return (K == HeadSibAction) ? "+- " : (K == OtherSibAction) ? "|- " : ""; 2325 }; 2326 2327 unsigned Id = Ids.size(); 2328 Ids[A] = Id; 2329 llvm::errs() << Indent + getSelfIndent(Kind) << Id << ": " << os.str() << ", " 2330 << types::getTypeName(A->getType()) << offload_os.str() << "\n"; 2331 2332 return Id; 2333 } 2334 2335 // Print the action graphs in a compilation C. 2336 // For example "clang -c file1.c file2.c" is composed of two subgraphs. 2337 void Driver::PrintActions(const Compilation &C) const { 2338 std::map<Action *, unsigned> Ids; 2339 for (Action *A : C.getActions()) 2340 PrintActions1(C, A, Ids); 2341 } 2342 2343 /// Check whether the given input tree contains any compilation or 2344 /// assembly actions. 2345 static bool ContainsCompileOrAssembleAction(const Action *A) { 2346 if (isa<CompileJobAction>(A) || isa<BackendJobAction>(A) || 2347 isa<AssembleJobAction>(A)) 2348 return true; 2349 2350 return llvm::any_of(A->inputs(), ContainsCompileOrAssembleAction); 2351 } 2352 2353 void Driver::BuildUniversalActions(Compilation &C, const ToolChain &TC, 2354 const InputList &BAInputs) const { 2355 DerivedArgList &Args = C.getArgs(); 2356 ActionList &Actions = C.getActions(); 2357 llvm::PrettyStackTraceString CrashInfo("Building universal build actions"); 2358 // Collect the list of architectures. Duplicates are allowed, but should only 2359 // be handled once (in the order seen). 2360 llvm::StringSet<> ArchNames; 2361 SmallVector<const char *, 4> Archs; 2362 for (Arg *A : Args) { 2363 if (A->getOption().matches(options::OPT_arch)) { 2364 // Validate the option here; we don't save the type here because its 2365 // particular spelling may participate in other driver choices. 2366 llvm::Triple::ArchType Arch = 2367 tools::darwin::getArchTypeForMachOArchName(A->getValue()); 2368 if (Arch == llvm::Triple::UnknownArch) { 2369 Diag(clang::diag::err_drv_invalid_arch_name) << A->getAsString(Args); 2370 continue; 2371 } 2372 2373 A->claim(); 2374 if (ArchNames.insert(A->getValue()).second) 2375 Archs.push_back(A->getValue()); 2376 } 2377 } 2378 2379 // When there is no explicit arch for this platform, make sure we still bind 2380 // the architecture (to the default) so that -Xarch_ is handled correctly. 2381 if (!Archs.size()) 2382 Archs.push_back(Args.MakeArgString(TC.getDefaultUniversalArchName())); 2383 2384 ActionList SingleActions; 2385 BuildActions(C, Args, BAInputs, SingleActions); 2386 2387 // Add in arch bindings for every top level action, as well as lipo and 2388 // dsymutil steps if needed. 2389 for (Action* Act : SingleActions) { 2390 // Make sure we can lipo this kind of output. If not (and it is an actual 2391 // output) then we disallow, since we can't create an output file with the 2392 // right name without overwriting it. We could remove this oddity by just 2393 // changing the output names to include the arch, which would also fix 2394 // -save-temps. Compatibility wins for now. 2395 2396 if (Archs.size() > 1 && !types::canLipoType(Act->getType())) 2397 Diag(clang::diag::err_drv_invalid_output_with_multiple_archs) 2398 << types::getTypeName(Act->getType()); 2399 2400 ActionList Inputs; 2401 for (unsigned i = 0, e = Archs.size(); i != e; ++i) 2402 Inputs.push_back(C.MakeAction<BindArchAction>(Act, Archs[i])); 2403 2404 // Lipo if necessary, we do it this way because we need to set the arch flag 2405 // so that -Xarch_ gets overwritten. 2406 if (Inputs.size() == 1 || Act->getType() == types::TY_Nothing) 2407 Actions.append(Inputs.begin(), Inputs.end()); 2408 else 2409 Actions.push_back(C.MakeAction<LipoJobAction>(Inputs, Act->getType())); 2410 2411 // Handle debug info queries. 2412 Arg *A = Args.getLastArg(options::OPT_g_Group); 2413 bool enablesDebugInfo = A && !A->getOption().matches(options::OPT_g0) && 2414 !A->getOption().matches(options::OPT_gstabs); 2415 if ((enablesDebugInfo || willEmitRemarks(Args)) && 2416 ContainsCompileOrAssembleAction(Actions.back())) { 2417 2418 // Add a 'dsymutil' step if necessary, when debug info is enabled and we 2419 // have a compile input. We need to run 'dsymutil' ourselves in such cases 2420 // because the debug info will refer to a temporary object file which 2421 // will be removed at the end of the compilation process. 2422 if (Act->getType() == types::TY_Image) { 2423 ActionList Inputs; 2424 Inputs.push_back(Actions.back()); 2425 Actions.pop_back(); 2426 Actions.push_back( 2427 C.MakeAction<DsymutilJobAction>(Inputs, types::TY_dSYM)); 2428 } 2429 2430 // Verify the debug info output. 2431 if (Args.hasArg(options::OPT_verify_debug_info)) { 2432 Action* LastAction = Actions.back(); 2433 Actions.pop_back(); 2434 Actions.push_back(C.MakeAction<VerifyDebugInfoJobAction>( 2435 LastAction, types::TY_Nothing)); 2436 } 2437 } 2438 } 2439 } 2440 2441 bool Driver::DiagnoseInputExistence(const DerivedArgList &Args, StringRef Value, 2442 types::ID Ty, bool TypoCorrect) const { 2443 if (!getCheckInputsExist()) 2444 return true; 2445 2446 // stdin always exists. 2447 if (Value == "-") 2448 return true; 2449 2450 // If it's a header to be found in the system or user search path, then defer 2451 // complaints about its absence until those searches can be done. When we 2452 // are definitely processing headers for C++20 header units, extend this to 2453 // allow the user to put "-fmodule-header -xc++-header vector" for example. 2454 if (Ty == types::TY_CXXSHeader || Ty == types::TY_CXXUHeader || 2455 (ModulesModeCXX20 && Ty == types::TY_CXXHeader)) 2456 return true; 2457 2458 if (getVFS().exists(Value)) 2459 return true; 2460 2461 if (TypoCorrect) { 2462 // Check if the filename is a typo for an option flag. OptTable thinks 2463 // that all args that are not known options and that start with / are 2464 // filenames, but e.g. `/diagnostic:caret` is more likely a typo for 2465 // the option `/diagnostics:caret` than a reference to a file in the root 2466 // directory. 2467 unsigned IncludedFlagsBitmask; 2468 unsigned ExcludedFlagsBitmask; 2469 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 2470 getIncludeExcludeOptionFlagMasks(IsCLMode()); 2471 std::string Nearest; 2472 if (getOpts().findNearest(Value, Nearest, IncludedFlagsBitmask, 2473 ExcludedFlagsBitmask) <= 1) { 2474 Diag(clang::diag::err_drv_no_such_file_with_suggestion) 2475 << Value << Nearest; 2476 return false; 2477 } 2478 } 2479 2480 // In CL mode, don't error on apparently non-existent linker inputs, because 2481 // they can be influenced by linker flags the clang driver might not 2482 // understand. 2483 // Examples: 2484 // - `clang-cl main.cc ole32.lib` in a non-MSVC shell will make the driver 2485 // module look for an MSVC installation in the registry. (We could ask 2486 // the MSVCToolChain object if it can find `ole32.lib`, but the logic to 2487 // look in the registry might move into lld-link in the future so that 2488 // lld-link invocations in non-MSVC shells just work too.) 2489 // - `clang-cl ... /link ...` can pass arbitrary flags to the linker, 2490 // including /libpath:, which is used to find .lib and .obj files. 2491 // So do not diagnose this on the driver level. Rely on the linker diagnosing 2492 // it. (If we don't end up invoking the linker, this means we'll emit a 2493 // "'linker' input unused [-Wunused-command-line-argument]" warning instead 2494 // of an error.) 2495 // 2496 // Only do this skip after the typo correction step above. `/Brepo` is treated 2497 // as TY_Object, but it's clearly a typo for `/Brepro`. It seems fine to emit 2498 // an error if we have a flag that's within an edit distance of 1 from a 2499 // flag. (Users can use `-Wl,` or `/linker` to launder the flag past the 2500 // driver in the unlikely case they run into this.) 2501 // 2502 // Don't do this for inputs that start with a '/', else we'd pass options 2503 // like /libpath: through to the linker silently. 2504 // 2505 // Emitting an error for linker inputs can also cause incorrect diagnostics 2506 // with the gcc driver. The command 2507 // clang -fuse-ld=lld -Wl,--chroot,some/dir /file.o 2508 // will make lld look for some/dir/file.o, while we will diagnose here that 2509 // `/file.o` does not exist. However, configure scripts check if 2510 // `clang /GR-` compiles without error to see if the compiler is cl.exe, 2511 // so we can't downgrade diagnostics for `/GR-` from an error to a warning 2512 // in cc mode. (We can in cl mode because cl.exe itself only warns on 2513 // unknown flags.) 2514 if (IsCLMode() && Ty == types::TY_Object && !Value.startswith("/")) 2515 return true; 2516 2517 Diag(clang::diag::err_drv_no_such_file) << Value; 2518 return false; 2519 } 2520 2521 // Get the C++20 Header Unit type corresponding to the input type. 2522 static types::ID CXXHeaderUnitType(ModuleHeaderMode HM) { 2523 switch (HM) { 2524 case HeaderMode_User: 2525 return types::TY_CXXUHeader; 2526 case HeaderMode_System: 2527 return types::TY_CXXSHeader; 2528 case HeaderMode_Default: 2529 break; 2530 case HeaderMode_None: 2531 llvm_unreachable("should not be called in this case"); 2532 } 2533 return types::TY_CXXHUHeader; 2534 } 2535 2536 // Construct a the list of inputs and their types. 2537 void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args, 2538 InputList &Inputs) const { 2539 const llvm::opt::OptTable &Opts = getOpts(); 2540 // Track the current user specified (-x) input. We also explicitly track the 2541 // argument used to set the type; we only want to claim the type when we 2542 // actually use it, so we warn about unused -x arguments. 2543 types::ID InputType = types::TY_Nothing; 2544 Arg *InputTypeArg = nullptr; 2545 2546 // The last /TC or /TP option sets the input type to C or C++ globally. 2547 if (Arg *TCTP = Args.getLastArgNoClaim(options::OPT__SLASH_TC, 2548 options::OPT__SLASH_TP)) { 2549 InputTypeArg = TCTP; 2550 InputType = TCTP->getOption().matches(options::OPT__SLASH_TC) 2551 ? types::TY_C 2552 : types::TY_CXX; 2553 2554 Arg *Previous = nullptr; 2555 bool ShowNote = false; 2556 for (Arg *A : 2557 Args.filtered(options::OPT__SLASH_TC, options::OPT__SLASH_TP)) { 2558 if (Previous) { 2559 Diag(clang::diag::warn_drv_overriding_flag_option) 2560 << Previous->getSpelling() << A->getSpelling(); 2561 ShowNote = true; 2562 } 2563 Previous = A; 2564 } 2565 if (ShowNote) 2566 Diag(clang::diag::note_drv_t_option_is_global); 2567 2568 // No driver mode exposes -x and /TC or /TP; we don't support mixing them. 2569 assert(!Args.hasArg(options::OPT_x) && "-x and /TC or /TP is not allowed"); 2570 } 2571 2572 // Warn -x after last input file has no effect 2573 { 2574 Arg *LastXArg = Args.getLastArgNoClaim(options::OPT_x); 2575 Arg *LastInputArg = Args.getLastArgNoClaim(options::OPT_INPUT); 2576 if (LastXArg && LastInputArg && LastInputArg->getIndex() < LastXArg->getIndex()) 2577 Diag(clang::diag::warn_drv_unused_x) << LastXArg->getValue(); 2578 } 2579 2580 for (Arg *A : Args) { 2581 if (A->getOption().getKind() == Option::InputClass) { 2582 const char *Value = A->getValue(); 2583 types::ID Ty = types::TY_INVALID; 2584 2585 // Infer the input type if necessary. 2586 if (InputType == types::TY_Nothing) { 2587 // If there was an explicit arg for this, claim it. 2588 if (InputTypeArg) 2589 InputTypeArg->claim(); 2590 2591 // stdin must be handled specially. 2592 if (memcmp(Value, "-", 2) == 0) { 2593 if (IsFlangMode()) { 2594 Ty = types::TY_Fortran; 2595 } else { 2596 // If running with -E, treat as a C input (this changes the 2597 // builtin macros, for example). This may be overridden by -ObjC 2598 // below. 2599 // 2600 // Otherwise emit an error but still use a valid type to avoid 2601 // spurious errors (e.g., no inputs). 2602 assert(!CCGenDiagnostics && "stdin produces no crash reproducer"); 2603 if (!Args.hasArgNoClaim(options::OPT_E) && !CCCIsCPP()) 2604 Diag(IsCLMode() ? clang::diag::err_drv_unknown_stdin_type_clang_cl 2605 : clang::diag::err_drv_unknown_stdin_type); 2606 Ty = types::TY_C; 2607 } 2608 } else { 2609 // Otherwise lookup by extension. 2610 // Fallback is C if invoked as C preprocessor, C++ if invoked with 2611 // clang-cl /E, or Object otherwise. 2612 // We use a host hook here because Darwin at least has its own 2613 // idea of what .s is. 2614 if (const char *Ext = strrchr(Value, '.')) 2615 Ty = TC.LookupTypeForExtension(Ext + 1); 2616 2617 if (Ty == types::TY_INVALID) { 2618 if (IsCLMode() && (Args.hasArgNoClaim(options::OPT_E) || CCGenDiagnostics)) 2619 Ty = types::TY_CXX; 2620 else if (CCCIsCPP() || CCGenDiagnostics) 2621 Ty = types::TY_C; 2622 else 2623 Ty = types::TY_Object; 2624 } 2625 2626 // If the driver is invoked as C++ compiler (like clang++ or c++) it 2627 // should autodetect some input files as C++ for g++ compatibility. 2628 if (CCCIsCXX()) { 2629 types::ID OldTy = Ty; 2630 Ty = types::lookupCXXTypeForCType(Ty); 2631 2632 // Do not complain about foo.h, when we are known to be processing 2633 // it as a C++20 header unit. 2634 if (Ty != OldTy && !(OldTy == types::TY_CHeader && hasHeaderMode())) 2635 Diag(clang::diag::warn_drv_treating_input_as_cxx) 2636 << getTypeName(OldTy) << getTypeName(Ty); 2637 } 2638 2639 // If running with -fthinlto-index=, extensions that normally identify 2640 // native object files actually identify LLVM bitcode files. 2641 if (Args.hasArgNoClaim(options::OPT_fthinlto_index_EQ) && 2642 Ty == types::TY_Object) 2643 Ty = types::TY_LLVM_BC; 2644 } 2645 2646 // -ObjC and -ObjC++ override the default language, but only for "source 2647 // files". We just treat everything that isn't a linker input as a 2648 // source file. 2649 // 2650 // FIXME: Clean this up if we move the phase sequence into the type. 2651 if (Ty != types::TY_Object) { 2652 if (Args.hasArg(options::OPT_ObjC)) 2653 Ty = types::TY_ObjC; 2654 else if (Args.hasArg(options::OPT_ObjCXX)) 2655 Ty = types::TY_ObjCXX; 2656 } 2657 2658 // Disambiguate headers that are meant to be header units from those 2659 // intended to be PCH. Avoid missing '.h' cases that are counted as 2660 // C headers by default - we know we are in C++ mode and we do not 2661 // want to issue a complaint about compiling things in the wrong mode. 2662 if ((Ty == types::TY_CXXHeader || Ty == types::TY_CHeader) && 2663 hasHeaderMode()) 2664 Ty = CXXHeaderUnitType(CXX20HeaderType); 2665 } else { 2666 assert(InputTypeArg && "InputType set w/o InputTypeArg"); 2667 if (!InputTypeArg->getOption().matches(options::OPT_x)) { 2668 // If emulating cl.exe, make sure that /TC and /TP don't affect input 2669 // object files. 2670 const char *Ext = strrchr(Value, '.'); 2671 if (Ext && TC.LookupTypeForExtension(Ext + 1) == types::TY_Object) 2672 Ty = types::TY_Object; 2673 } 2674 if (Ty == types::TY_INVALID) { 2675 Ty = InputType; 2676 InputTypeArg->claim(); 2677 } 2678 } 2679 2680 if (DiagnoseInputExistence(Args, Value, Ty, /*TypoCorrect=*/true)) 2681 Inputs.push_back(std::make_pair(Ty, A)); 2682 2683 } else if (A->getOption().matches(options::OPT__SLASH_Tc)) { 2684 StringRef Value = A->getValue(); 2685 if (DiagnoseInputExistence(Args, Value, types::TY_C, 2686 /*TypoCorrect=*/false)) { 2687 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2688 Inputs.push_back(std::make_pair(types::TY_C, InputArg)); 2689 } 2690 A->claim(); 2691 } else if (A->getOption().matches(options::OPT__SLASH_Tp)) { 2692 StringRef Value = A->getValue(); 2693 if (DiagnoseInputExistence(Args, Value, types::TY_CXX, 2694 /*TypoCorrect=*/false)) { 2695 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2696 Inputs.push_back(std::make_pair(types::TY_CXX, InputArg)); 2697 } 2698 A->claim(); 2699 } else if (A->getOption().hasFlag(options::LinkerInput)) { 2700 // Just treat as object type, we could make a special type for this if 2701 // necessary. 2702 Inputs.push_back(std::make_pair(types::TY_Object, A)); 2703 2704 } else if (A->getOption().matches(options::OPT_x)) { 2705 InputTypeArg = A; 2706 InputType = types::lookupTypeForTypeSpecifier(A->getValue()); 2707 A->claim(); 2708 2709 // Follow gcc behavior and treat as linker input for invalid -x 2710 // options. Its not clear why we shouldn't just revert to unknown; but 2711 // this isn't very important, we might as well be bug compatible. 2712 if (!InputType) { 2713 Diag(clang::diag::err_drv_unknown_language) << A->getValue(); 2714 InputType = types::TY_Object; 2715 } 2716 2717 // If the user has put -fmodule-header{,=} then we treat C++ headers as 2718 // header unit inputs. So we 'promote' -xc++-header appropriately. 2719 if (InputType == types::TY_CXXHeader && hasHeaderMode()) 2720 InputType = CXXHeaderUnitType(CXX20HeaderType); 2721 } else if (A->getOption().getID() == options::OPT_U) { 2722 assert(A->getNumValues() == 1 && "The /U option has one value."); 2723 StringRef Val = A->getValue(0); 2724 if (Val.find_first_of("/\\") != StringRef::npos) { 2725 // Warn about e.g. "/Users/me/myfile.c". 2726 Diag(diag::warn_slash_u_filename) << Val; 2727 Diag(diag::note_use_dashdash); 2728 } 2729 } 2730 } 2731 if (CCCIsCPP() && Inputs.empty()) { 2732 // If called as standalone preprocessor, stdin is processed 2733 // if no other input is present. 2734 Arg *A = MakeInputArg(Args, Opts, "-"); 2735 Inputs.push_back(std::make_pair(types::TY_C, A)); 2736 } 2737 } 2738 2739 namespace { 2740 /// Provides a convenient interface for different programming models to generate 2741 /// the required device actions. 2742 class OffloadingActionBuilder final { 2743 /// Flag used to trace errors in the builder. 2744 bool IsValid = false; 2745 2746 /// The compilation that is using this builder. 2747 Compilation &C; 2748 2749 /// Map between an input argument and the offload kinds used to process it. 2750 std::map<const Arg *, unsigned> InputArgToOffloadKindMap; 2751 2752 /// Map between a host action and its originating input argument. 2753 std::map<Action *, const Arg *> HostActionToInputArgMap; 2754 2755 /// Builder interface. It doesn't build anything or keep any state. 2756 class DeviceActionBuilder { 2757 public: 2758 typedef const llvm::SmallVectorImpl<phases::ID> PhasesTy; 2759 2760 enum ActionBuilderReturnCode { 2761 // The builder acted successfully on the current action. 2762 ABRT_Success, 2763 // The builder didn't have to act on the current action. 2764 ABRT_Inactive, 2765 // The builder was successful and requested the host action to not be 2766 // generated. 2767 ABRT_Ignore_Host, 2768 }; 2769 2770 protected: 2771 /// Compilation associated with this builder. 2772 Compilation &C; 2773 2774 /// Tool chains associated with this builder. The same programming 2775 /// model may have associated one or more tool chains. 2776 SmallVector<const ToolChain *, 2> ToolChains; 2777 2778 /// The derived arguments associated with this builder. 2779 DerivedArgList &Args; 2780 2781 /// The inputs associated with this builder. 2782 const Driver::InputList &Inputs; 2783 2784 /// The associated offload kind. 2785 Action::OffloadKind AssociatedOffloadKind = Action::OFK_None; 2786 2787 public: 2788 DeviceActionBuilder(Compilation &C, DerivedArgList &Args, 2789 const Driver::InputList &Inputs, 2790 Action::OffloadKind AssociatedOffloadKind) 2791 : C(C), Args(Args), Inputs(Inputs), 2792 AssociatedOffloadKind(AssociatedOffloadKind) {} 2793 virtual ~DeviceActionBuilder() {} 2794 2795 /// Fill up the array \a DA with all the device dependences that should be 2796 /// added to the provided host action \a HostAction. By default it is 2797 /// inactive. 2798 virtual ActionBuilderReturnCode 2799 getDeviceDependences(OffloadAction::DeviceDependences &DA, 2800 phases::ID CurPhase, phases::ID FinalPhase, 2801 PhasesTy &Phases) { 2802 return ABRT_Inactive; 2803 } 2804 2805 /// Update the state to include the provided host action \a HostAction as a 2806 /// dependency of the current device action. By default it is inactive. 2807 virtual ActionBuilderReturnCode addDeviceDependences(Action *HostAction) { 2808 return ABRT_Inactive; 2809 } 2810 2811 /// Append top level actions generated by the builder. 2812 virtual void appendTopLevelActions(ActionList &AL) {} 2813 2814 /// Append linker device actions generated by the builder. 2815 virtual void appendLinkDeviceActions(ActionList &AL) {} 2816 2817 /// Append linker host action generated by the builder. 2818 virtual Action* appendLinkHostActions(ActionList &AL) { return nullptr; } 2819 2820 /// Append linker actions generated by the builder. 2821 virtual void appendLinkDependences(OffloadAction::DeviceDependences &DA) {} 2822 2823 /// Initialize the builder. Return true if any initialization errors are 2824 /// found. 2825 virtual bool initialize() { return false; } 2826 2827 /// Return true if the builder can use bundling/unbundling. 2828 virtual bool canUseBundlerUnbundler() const { return false; } 2829 2830 /// Return true if this builder is valid. We have a valid builder if we have 2831 /// associated device tool chains. 2832 bool isValid() { return !ToolChains.empty(); } 2833 2834 /// Return the associated offload kind. 2835 Action::OffloadKind getAssociatedOffloadKind() { 2836 return AssociatedOffloadKind; 2837 } 2838 }; 2839 2840 /// Base class for CUDA/HIP action builder. It injects device code in 2841 /// the host backend action. 2842 class CudaActionBuilderBase : public DeviceActionBuilder { 2843 protected: 2844 /// Flags to signal if the user requested host-only or device-only 2845 /// compilation. 2846 bool CompileHostOnly = false; 2847 bool CompileDeviceOnly = false; 2848 bool EmitLLVM = false; 2849 bool EmitAsm = false; 2850 2851 /// ID to identify each device compilation. For CUDA it is simply the 2852 /// GPU arch string. For HIP it is either the GPU arch string or GPU 2853 /// arch string plus feature strings delimited by a plus sign, e.g. 2854 /// gfx906+xnack. 2855 struct TargetID { 2856 /// Target ID string which is persistent throughout the compilation. 2857 const char *ID; 2858 TargetID(CudaArch Arch) { ID = CudaArchToString(Arch); } 2859 TargetID(const char *ID) : ID(ID) {} 2860 operator const char *() { return ID; } 2861 operator StringRef() { return StringRef(ID); } 2862 }; 2863 /// List of GPU architectures to use in this compilation. 2864 SmallVector<TargetID, 4> GpuArchList; 2865 2866 /// The CUDA actions for the current input. 2867 ActionList CudaDeviceActions; 2868 2869 /// The CUDA fat binary if it was generated for the current input. 2870 Action *CudaFatBinary = nullptr; 2871 2872 /// Flag that is set to true if this builder acted on the current input. 2873 bool IsActive = false; 2874 2875 /// Flag for -fgpu-rdc. 2876 bool Relocatable = false; 2877 2878 /// Default GPU architecture if there's no one specified. 2879 CudaArch DefaultCudaArch = CudaArch::UNKNOWN; 2880 2881 /// Method to generate compilation unit ID specified by option 2882 /// '-fuse-cuid='. 2883 enum UseCUIDKind { CUID_Hash, CUID_Random, CUID_None, CUID_Invalid }; 2884 UseCUIDKind UseCUID = CUID_Hash; 2885 2886 /// Compilation unit ID specified by option '-cuid='. 2887 StringRef FixedCUID; 2888 2889 public: 2890 CudaActionBuilderBase(Compilation &C, DerivedArgList &Args, 2891 const Driver::InputList &Inputs, 2892 Action::OffloadKind OFKind) 2893 : DeviceActionBuilder(C, Args, Inputs, OFKind) {} 2894 2895 ActionBuilderReturnCode addDeviceDependences(Action *HostAction) override { 2896 // While generating code for CUDA, we only depend on the host input action 2897 // to trigger the creation of all the CUDA device actions. 2898 2899 // If we are dealing with an input action, replicate it for each GPU 2900 // architecture. If we are in host-only mode we return 'success' so that 2901 // the host uses the CUDA offload kind. 2902 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 2903 assert(!GpuArchList.empty() && 2904 "We should have at least one GPU architecture."); 2905 2906 // If the host input is not CUDA or HIP, we don't need to bother about 2907 // this input. 2908 if (!(IA->getType() == types::TY_CUDA || 2909 IA->getType() == types::TY_HIP || 2910 IA->getType() == types::TY_PP_HIP)) { 2911 // The builder will ignore this input. 2912 IsActive = false; 2913 return ABRT_Inactive; 2914 } 2915 2916 // Set the flag to true, so that the builder acts on the current input. 2917 IsActive = true; 2918 2919 if (CompileHostOnly) 2920 return ABRT_Success; 2921 2922 // Replicate inputs for each GPU architecture. 2923 auto Ty = IA->getType() == types::TY_HIP ? types::TY_HIP_DEVICE 2924 : types::TY_CUDA_DEVICE; 2925 std::string CUID = FixedCUID.str(); 2926 if (CUID.empty()) { 2927 if (UseCUID == CUID_Random) 2928 CUID = llvm::utohexstr(llvm::sys::Process::GetRandomNumber(), 2929 /*LowerCase=*/true); 2930 else if (UseCUID == CUID_Hash) { 2931 llvm::MD5 Hasher; 2932 llvm::MD5::MD5Result Hash; 2933 SmallString<256> RealPath; 2934 llvm::sys::fs::real_path(IA->getInputArg().getValue(), RealPath, 2935 /*expand_tilde=*/true); 2936 Hasher.update(RealPath); 2937 for (auto *A : Args) { 2938 if (A->getOption().matches(options::OPT_INPUT)) 2939 continue; 2940 Hasher.update(A->getAsString(Args)); 2941 } 2942 Hasher.final(Hash); 2943 CUID = llvm::utohexstr(Hash.low(), /*LowerCase=*/true); 2944 } 2945 } 2946 IA->setId(CUID); 2947 2948 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 2949 CudaDeviceActions.push_back( 2950 C.MakeAction<InputAction>(IA->getInputArg(), Ty, IA->getId())); 2951 } 2952 2953 return ABRT_Success; 2954 } 2955 2956 // If this is an unbundling action use it as is for each CUDA toolchain. 2957 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 2958 2959 // If -fgpu-rdc is disabled, should not unbundle since there is no 2960 // device code to link. 2961 if (UA->getType() == types::TY_Object && !Relocatable) 2962 return ABRT_Inactive; 2963 2964 CudaDeviceActions.clear(); 2965 auto *IA = cast<InputAction>(UA->getInputs().back()); 2966 std::string FileName = IA->getInputArg().getAsString(Args); 2967 // Check if the type of the file is the same as the action. Do not 2968 // unbundle it if it is not. Do not unbundle .so files, for example, 2969 // which are not object files. Files with extension ".lib" is classified 2970 // as TY_Object but they are actually archives, therefore should not be 2971 // unbundled here as objects. They will be handled at other places. 2972 const StringRef LibFileExt = ".lib"; 2973 if (IA->getType() == types::TY_Object && 2974 (!llvm::sys::path::has_extension(FileName) || 2975 types::lookupTypeForExtension( 2976 llvm::sys::path::extension(FileName).drop_front()) != 2977 types::TY_Object || 2978 llvm::sys::path::extension(FileName) == LibFileExt)) 2979 return ABRT_Inactive; 2980 2981 for (auto Arch : GpuArchList) { 2982 CudaDeviceActions.push_back(UA); 2983 UA->registerDependentActionInfo(ToolChains[0], Arch, 2984 AssociatedOffloadKind); 2985 } 2986 IsActive = true; 2987 return ABRT_Success; 2988 } 2989 2990 return IsActive ? ABRT_Success : ABRT_Inactive; 2991 } 2992 2993 void appendTopLevelActions(ActionList &AL) override { 2994 // Utility to append actions to the top level list. 2995 auto AddTopLevel = [&](Action *A, TargetID TargetID) { 2996 OffloadAction::DeviceDependences Dep; 2997 Dep.add(*A, *ToolChains.front(), TargetID, AssociatedOffloadKind); 2998 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 2999 }; 3000 3001 // If we have a fat binary, add it to the list. 3002 if (CudaFatBinary) { 3003 AddTopLevel(CudaFatBinary, CudaArch::UNUSED); 3004 CudaDeviceActions.clear(); 3005 CudaFatBinary = nullptr; 3006 return; 3007 } 3008 3009 if (CudaDeviceActions.empty()) 3010 return; 3011 3012 // If we have CUDA actions at this point, that's because we have a have 3013 // partial compilation, so we should have an action for each GPU 3014 // architecture. 3015 assert(CudaDeviceActions.size() == GpuArchList.size() && 3016 "Expecting one action per GPU architecture."); 3017 assert(ToolChains.size() == 1 && 3018 "Expecting to have a single CUDA toolchain."); 3019 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) 3020 AddTopLevel(CudaDeviceActions[I], GpuArchList[I]); 3021 3022 CudaDeviceActions.clear(); 3023 } 3024 3025 /// Get canonicalized offload arch option. \returns empty StringRef if the 3026 /// option is invalid. 3027 virtual StringRef getCanonicalOffloadArch(StringRef Arch) = 0; 3028 3029 virtual std::optional<std::pair<llvm::StringRef, llvm::StringRef>> 3030 getConflictOffloadArchCombination(const std::set<StringRef> &GpuArchs) = 0; 3031 3032 bool initialize() override { 3033 assert(AssociatedOffloadKind == Action::OFK_Cuda || 3034 AssociatedOffloadKind == Action::OFK_HIP); 3035 3036 // We don't need to support CUDA. 3037 if (AssociatedOffloadKind == Action::OFK_Cuda && 3038 !C.hasOffloadToolChain<Action::OFK_Cuda>()) 3039 return false; 3040 3041 // We don't need to support HIP. 3042 if (AssociatedOffloadKind == Action::OFK_HIP && 3043 !C.hasOffloadToolChain<Action::OFK_HIP>()) 3044 return false; 3045 3046 Relocatable = Args.hasFlag(options::OPT_fgpu_rdc, 3047 options::OPT_fno_gpu_rdc, /*Default=*/false); 3048 3049 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 3050 assert(HostTC && "No toolchain for host compilation."); 3051 if (HostTC->getTriple().isNVPTX() || 3052 HostTC->getTriple().getArch() == llvm::Triple::amdgcn) { 3053 // We do not support targeting NVPTX/AMDGCN for host compilation. Throw 3054 // an error and abort pipeline construction early so we don't trip 3055 // asserts that assume device-side compilation. 3056 C.getDriver().Diag(diag::err_drv_cuda_host_arch) 3057 << HostTC->getTriple().getArchName(); 3058 return true; 3059 } 3060 3061 ToolChains.push_back( 3062 AssociatedOffloadKind == Action::OFK_Cuda 3063 ? C.getSingleOffloadToolChain<Action::OFK_Cuda>() 3064 : C.getSingleOffloadToolChain<Action::OFK_HIP>()); 3065 3066 CompileHostOnly = C.getDriver().offloadHostOnly(); 3067 CompileDeviceOnly = C.getDriver().offloadDeviceOnly(); 3068 EmitLLVM = Args.getLastArg(options::OPT_emit_llvm); 3069 EmitAsm = Args.getLastArg(options::OPT_S); 3070 FixedCUID = Args.getLastArgValue(options::OPT_cuid_EQ); 3071 if (Arg *A = Args.getLastArg(options::OPT_fuse_cuid_EQ)) { 3072 StringRef UseCUIDStr = A->getValue(); 3073 UseCUID = llvm::StringSwitch<UseCUIDKind>(UseCUIDStr) 3074 .Case("hash", CUID_Hash) 3075 .Case("random", CUID_Random) 3076 .Case("none", CUID_None) 3077 .Default(CUID_Invalid); 3078 if (UseCUID == CUID_Invalid) { 3079 C.getDriver().Diag(diag::err_drv_invalid_value) 3080 << A->getAsString(Args) << UseCUIDStr; 3081 C.setContainsError(); 3082 return true; 3083 } 3084 } 3085 3086 // --offload and --offload-arch options are mutually exclusive. 3087 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 3088 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 3089 options::OPT_no_offload_arch_EQ)) { 3090 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) << "--offload-arch" 3091 << "--offload"; 3092 } 3093 3094 // Collect all offload arch parameters, removing duplicates. 3095 std::set<StringRef> GpuArchs; 3096 bool Error = false; 3097 for (Arg *A : Args) { 3098 if (!(A->getOption().matches(options::OPT_offload_arch_EQ) || 3099 A->getOption().matches(options::OPT_no_offload_arch_EQ))) 3100 continue; 3101 A->claim(); 3102 3103 for (StringRef ArchStr : llvm::split(A->getValue(), ",")) { 3104 if (A->getOption().matches(options::OPT_no_offload_arch_EQ) && 3105 ArchStr == "all") { 3106 GpuArchs.clear(); 3107 } else if (ArchStr == "native") { 3108 const ToolChain &TC = *ToolChains.front(); 3109 auto GPUsOrErr = ToolChains.front()->getSystemGPUArchs(Args); 3110 if (!GPUsOrErr) { 3111 TC.getDriver().Diag(diag::err_drv_undetermined_gpu_arch) 3112 << llvm::Triple::getArchTypeName(TC.getArch()) 3113 << llvm::toString(GPUsOrErr.takeError()) << "--offload-arch"; 3114 continue; 3115 } 3116 3117 for (auto GPU : *GPUsOrErr) { 3118 GpuArchs.insert(Args.MakeArgString(GPU)); 3119 } 3120 } else { 3121 ArchStr = getCanonicalOffloadArch(ArchStr); 3122 if (ArchStr.empty()) { 3123 Error = true; 3124 } else if (A->getOption().matches(options::OPT_offload_arch_EQ)) 3125 GpuArchs.insert(ArchStr); 3126 else if (A->getOption().matches(options::OPT_no_offload_arch_EQ)) 3127 GpuArchs.erase(ArchStr); 3128 else 3129 llvm_unreachable("Unexpected option."); 3130 } 3131 } 3132 } 3133 3134 auto &&ConflictingArchs = getConflictOffloadArchCombination(GpuArchs); 3135 if (ConflictingArchs) { 3136 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 3137 << ConflictingArchs->first << ConflictingArchs->second; 3138 C.setContainsError(); 3139 return true; 3140 } 3141 3142 // Collect list of GPUs remaining in the set. 3143 for (auto Arch : GpuArchs) 3144 GpuArchList.push_back(Arch.data()); 3145 3146 // Default to sm_20 which is the lowest common denominator for 3147 // supported GPUs. sm_20 code should work correctly, if 3148 // suboptimally, on all newer GPUs. 3149 if (GpuArchList.empty()) { 3150 if (ToolChains.front()->getTriple().isSPIRV()) 3151 GpuArchList.push_back(CudaArch::Generic); 3152 else 3153 GpuArchList.push_back(DefaultCudaArch); 3154 } 3155 3156 return Error; 3157 } 3158 }; 3159 3160 /// \brief CUDA action builder. It injects device code in the host backend 3161 /// action. 3162 class CudaActionBuilder final : public CudaActionBuilderBase { 3163 public: 3164 CudaActionBuilder(Compilation &C, DerivedArgList &Args, 3165 const Driver::InputList &Inputs) 3166 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) { 3167 DefaultCudaArch = CudaArch::SM_35; 3168 } 3169 3170 StringRef getCanonicalOffloadArch(StringRef ArchStr) override { 3171 CudaArch Arch = StringToCudaArch(ArchStr); 3172 if (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch)) { 3173 C.getDriver().Diag(clang::diag::err_drv_cuda_bad_gpu_arch) << ArchStr; 3174 return StringRef(); 3175 } 3176 return CudaArchToString(Arch); 3177 } 3178 3179 std::optional<std::pair<llvm::StringRef, llvm::StringRef>> 3180 getConflictOffloadArchCombination( 3181 const std::set<StringRef> &GpuArchs) override { 3182 return std::nullopt; 3183 } 3184 3185 ActionBuilderReturnCode 3186 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3187 phases::ID CurPhase, phases::ID FinalPhase, 3188 PhasesTy &Phases) override { 3189 if (!IsActive) 3190 return ABRT_Inactive; 3191 3192 // If we don't have more CUDA actions, we don't have any dependences to 3193 // create for the host. 3194 if (CudaDeviceActions.empty()) 3195 return ABRT_Success; 3196 3197 assert(CudaDeviceActions.size() == GpuArchList.size() && 3198 "Expecting one action per GPU architecture."); 3199 assert(!CompileHostOnly && 3200 "Not expecting CUDA actions in host-only compilation."); 3201 3202 // If we are generating code for the device or we are in a backend phase, 3203 // we attempt to generate the fat binary. We compile each arch to ptx and 3204 // assemble to cubin, then feed the cubin *and* the ptx into a device 3205 // "link" action, which uses fatbinary to combine these cubins into one 3206 // fatbin. The fatbin is then an input to the host action if not in 3207 // device-only mode. 3208 if (CompileDeviceOnly || CurPhase == phases::Backend) { 3209 ActionList DeviceActions; 3210 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3211 // Produce the device action from the current phase up to the assemble 3212 // phase. 3213 for (auto Ph : Phases) { 3214 // Skip the phases that were already dealt with. 3215 if (Ph < CurPhase) 3216 continue; 3217 // We have to be consistent with the host final phase. 3218 if (Ph > FinalPhase) 3219 break; 3220 3221 CudaDeviceActions[I] = C.getDriver().ConstructPhaseAction( 3222 C, Args, Ph, CudaDeviceActions[I], Action::OFK_Cuda); 3223 3224 if (Ph == phases::Assemble) 3225 break; 3226 } 3227 3228 // If we didn't reach the assemble phase, we can't generate the fat 3229 // binary. We don't need to generate the fat binary if we are not in 3230 // device-only mode. 3231 if (!isa<AssembleJobAction>(CudaDeviceActions[I]) || 3232 CompileDeviceOnly) 3233 continue; 3234 3235 Action *AssembleAction = CudaDeviceActions[I]; 3236 assert(AssembleAction->getType() == types::TY_Object); 3237 assert(AssembleAction->getInputs().size() == 1); 3238 3239 Action *BackendAction = AssembleAction->getInputs()[0]; 3240 assert(BackendAction->getType() == types::TY_PP_Asm); 3241 3242 for (auto &A : {AssembleAction, BackendAction}) { 3243 OffloadAction::DeviceDependences DDep; 3244 DDep.add(*A, *ToolChains.front(), GpuArchList[I], Action::OFK_Cuda); 3245 DeviceActions.push_back( 3246 C.MakeAction<OffloadAction>(DDep, A->getType())); 3247 } 3248 } 3249 3250 // We generate the fat binary if we have device input actions. 3251 if (!DeviceActions.empty()) { 3252 CudaFatBinary = 3253 C.MakeAction<LinkJobAction>(DeviceActions, types::TY_CUDA_FATBIN); 3254 3255 if (!CompileDeviceOnly) { 3256 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3257 Action::OFK_Cuda); 3258 // Clear the fat binary, it is already a dependence to an host 3259 // action. 3260 CudaFatBinary = nullptr; 3261 } 3262 3263 // Remove the CUDA actions as they are already connected to an host 3264 // action or fat binary. 3265 CudaDeviceActions.clear(); 3266 } 3267 3268 // We avoid creating host action in device-only mode. 3269 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3270 } else if (CurPhase > phases::Backend) { 3271 // If we are past the backend phase and still have a device action, we 3272 // don't have to do anything as this action is already a device 3273 // top-level action. 3274 return ABRT_Success; 3275 } 3276 3277 assert(CurPhase < phases::Backend && "Generating single CUDA " 3278 "instructions should only occur " 3279 "before the backend phase!"); 3280 3281 // By default, we produce an action for each device arch. 3282 for (Action *&A : CudaDeviceActions) 3283 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3284 3285 return ABRT_Success; 3286 } 3287 }; 3288 /// \brief HIP action builder. It injects device code in the host backend 3289 /// action. 3290 class HIPActionBuilder final : public CudaActionBuilderBase { 3291 /// The linker inputs obtained for each device arch. 3292 SmallVector<ActionList, 8> DeviceLinkerInputs; 3293 // The default bundling behavior depends on the type of output, therefore 3294 // BundleOutput needs to be tri-value: None, true, or false. 3295 // Bundle code objects except --no-gpu-output is specified for device 3296 // only compilation. Bundle other type of output files only if 3297 // --gpu-bundle-output is specified for device only compilation. 3298 std::optional<bool> BundleOutput; 3299 3300 public: 3301 HIPActionBuilder(Compilation &C, DerivedArgList &Args, 3302 const Driver::InputList &Inputs) 3303 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) { 3304 DefaultCudaArch = CudaArch::GFX906; 3305 if (Args.hasArg(options::OPT_gpu_bundle_output, 3306 options::OPT_no_gpu_bundle_output)) 3307 BundleOutput = Args.hasFlag(options::OPT_gpu_bundle_output, 3308 options::OPT_no_gpu_bundle_output, true); 3309 } 3310 3311 bool canUseBundlerUnbundler() const override { return true; } 3312 3313 StringRef getCanonicalOffloadArch(StringRef IdStr) override { 3314 llvm::StringMap<bool> Features; 3315 // getHIPOffloadTargetTriple() is known to return valid value as it has 3316 // been called successfully in the CreateOffloadingDeviceToolChains(). 3317 auto ArchStr = parseTargetID( 3318 *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()), IdStr, 3319 &Features); 3320 if (!ArchStr) { 3321 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << IdStr; 3322 C.setContainsError(); 3323 return StringRef(); 3324 } 3325 auto CanId = getCanonicalTargetID(*ArchStr, Features); 3326 return Args.MakeArgStringRef(CanId); 3327 }; 3328 3329 std::optional<std::pair<llvm::StringRef, llvm::StringRef>> 3330 getConflictOffloadArchCombination( 3331 const std::set<StringRef> &GpuArchs) override { 3332 return getConflictTargetIDCombination(GpuArchs); 3333 } 3334 3335 ActionBuilderReturnCode 3336 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3337 phases::ID CurPhase, phases::ID FinalPhase, 3338 PhasesTy &Phases) override { 3339 if (!IsActive) 3340 return ABRT_Inactive; 3341 3342 // amdgcn does not support linking of object files, therefore we skip 3343 // backend and assemble phases to output LLVM IR. Except for generating 3344 // non-relocatable device code, where we generate fat binary for device 3345 // code and pass to host in Backend phase. 3346 if (CudaDeviceActions.empty()) 3347 return ABRT_Success; 3348 3349 assert(((CurPhase == phases::Link && Relocatable) || 3350 CudaDeviceActions.size() == GpuArchList.size()) && 3351 "Expecting one action per GPU architecture."); 3352 assert(!CompileHostOnly && 3353 "Not expecting HIP actions in host-only compilation."); 3354 3355 if (!Relocatable && CurPhase == phases::Backend && !EmitLLVM && 3356 !EmitAsm) { 3357 // If we are in backend phase, we attempt to generate the fat binary. 3358 // We compile each arch to IR and use a link action to generate code 3359 // object containing ISA. Then we use a special "link" action to create 3360 // a fat binary containing all the code objects for different GPU's. 3361 // The fat binary is then an input to the host action. 3362 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3363 if (C.getDriver().isUsingLTO(/*IsOffload=*/true)) { 3364 // When LTO is enabled, skip the backend and assemble phases and 3365 // use lld to link the bitcode. 3366 ActionList AL; 3367 AL.push_back(CudaDeviceActions[I]); 3368 // Create a link action to link device IR with device library 3369 // and generate ISA. 3370 CudaDeviceActions[I] = 3371 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3372 } else { 3373 // When LTO is not enabled, we follow the conventional 3374 // compiler phases, including backend and assemble phases. 3375 ActionList AL; 3376 Action *BackendAction = nullptr; 3377 if (ToolChains.front()->getTriple().isSPIRV()) { 3378 // Emit LLVM bitcode for SPIR-V targets. SPIR-V device tool chain 3379 // (HIPSPVToolChain) runs post-link LLVM IR passes. 3380 types::ID Output = Args.hasArg(options::OPT_S) 3381 ? types::TY_LLVM_IR 3382 : types::TY_LLVM_BC; 3383 BackendAction = 3384 C.MakeAction<BackendJobAction>(CudaDeviceActions[I], Output); 3385 } else 3386 BackendAction = C.getDriver().ConstructPhaseAction( 3387 C, Args, phases::Backend, CudaDeviceActions[I], 3388 AssociatedOffloadKind); 3389 auto AssembleAction = C.getDriver().ConstructPhaseAction( 3390 C, Args, phases::Assemble, BackendAction, 3391 AssociatedOffloadKind); 3392 AL.push_back(AssembleAction); 3393 // Create a link action to link device IR with device library 3394 // and generate ISA. 3395 CudaDeviceActions[I] = 3396 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3397 } 3398 3399 // OffloadingActionBuilder propagates device arch until an offload 3400 // action. Since the next action for creating fatbin does 3401 // not have device arch, whereas the above link action and its input 3402 // have device arch, an offload action is needed to stop the null 3403 // device arch of the next action being propagated to the above link 3404 // action. 3405 OffloadAction::DeviceDependences DDep; 3406 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3407 AssociatedOffloadKind); 3408 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3409 DDep, CudaDeviceActions[I]->getType()); 3410 } 3411 3412 if (!CompileDeviceOnly || !BundleOutput || *BundleOutput) { 3413 // Create HIP fat binary with a special "link" action. 3414 CudaFatBinary = C.MakeAction<LinkJobAction>(CudaDeviceActions, 3415 types::TY_HIP_FATBIN); 3416 3417 if (!CompileDeviceOnly) { 3418 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3419 AssociatedOffloadKind); 3420 // Clear the fat binary, it is already a dependence to an host 3421 // action. 3422 CudaFatBinary = nullptr; 3423 } 3424 3425 // Remove the CUDA actions as they are already connected to an host 3426 // action or fat binary. 3427 CudaDeviceActions.clear(); 3428 } 3429 3430 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3431 } else if (CurPhase == phases::Link) { 3432 // Save CudaDeviceActions to DeviceLinkerInputs for each GPU subarch. 3433 // This happens to each device action originated from each input file. 3434 // Later on, device actions in DeviceLinkerInputs are used to create 3435 // device link actions in appendLinkDependences and the created device 3436 // link actions are passed to the offload action as device dependence. 3437 DeviceLinkerInputs.resize(CudaDeviceActions.size()); 3438 auto LI = DeviceLinkerInputs.begin(); 3439 for (auto *A : CudaDeviceActions) { 3440 LI->push_back(A); 3441 ++LI; 3442 } 3443 3444 // We will pass the device action as a host dependence, so we don't 3445 // need to do anything else with them. 3446 CudaDeviceActions.clear(); 3447 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3448 } 3449 3450 // By default, we produce an action for each device arch. 3451 for (Action *&A : CudaDeviceActions) 3452 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A, 3453 AssociatedOffloadKind); 3454 3455 if (CompileDeviceOnly && CurPhase == FinalPhase && BundleOutput && 3456 *BundleOutput) { 3457 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3458 OffloadAction::DeviceDependences DDep; 3459 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3460 AssociatedOffloadKind); 3461 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3462 DDep, CudaDeviceActions[I]->getType()); 3463 } 3464 CudaFatBinary = 3465 C.MakeAction<OffloadBundlingJobAction>(CudaDeviceActions); 3466 CudaDeviceActions.clear(); 3467 } 3468 3469 return (CompileDeviceOnly && CurPhase == FinalPhase) ? ABRT_Ignore_Host 3470 : ABRT_Success; 3471 } 3472 3473 void appendLinkDeviceActions(ActionList &AL) override { 3474 if (DeviceLinkerInputs.size() == 0) 3475 return; 3476 3477 assert(DeviceLinkerInputs.size() == GpuArchList.size() && 3478 "Linker inputs and GPU arch list sizes do not match."); 3479 3480 ActionList Actions; 3481 unsigned I = 0; 3482 // Append a new link action for each device. 3483 // Each entry in DeviceLinkerInputs corresponds to a GPU arch. 3484 for (auto &LI : DeviceLinkerInputs) { 3485 3486 types::ID Output = Args.hasArg(options::OPT_emit_llvm) 3487 ? types::TY_LLVM_BC 3488 : types::TY_Image; 3489 3490 auto *DeviceLinkAction = C.MakeAction<LinkJobAction>(LI, Output); 3491 // Linking all inputs for the current GPU arch. 3492 // LI contains all the inputs for the linker. 3493 OffloadAction::DeviceDependences DeviceLinkDeps; 3494 DeviceLinkDeps.add(*DeviceLinkAction, *ToolChains[0], 3495 GpuArchList[I], AssociatedOffloadKind); 3496 Actions.push_back(C.MakeAction<OffloadAction>( 3497 DeviceLinkDeps, DeviceLinkAction->getType())); 3498 ++I; 3499 } 3500 DeviceLinkerInputs.clear(); 3501 3502 // If emitting LLVM, do not generate final host/device compilation action 3503 if (Args.hasArg(options::OPT_emit_llvm)) { 3504 AL.append(Actions); 3505 return; 3506 } 3507 3508 // Create a host object from all the device images by embedding them 3509 // in a fat binary for mixed host-device compilation. For device-only 3510 // compilation, creates a fat binary. 3511 OffloadAction::DeviceDependences DDeps; 3512 if (!CompileDeviceOnly || !BundleOutput || *BundleOutput) { 3513 auto *TopDeviceLinkAction = C.MakeAction<LinkJobAction>( 3514 Actions, 3515 CompileDeviceOnly ? types::TY_HIP_FATBIN : types::TY_Object); 3516 DDeps.add(*TopDeviceLinkAction, *ToolChains[0], nullptr, 3517 AssociatedOffloadKind); 3518 // Offload the host object to the host linker. 3519 AL.push_back( 3520 C.MakeAction<OffloadAction>(DDeps, TopDeviceLinkAction->getType())); 3521 } else { 3522 AL.append(Actions); 3523 } 3524 } 3525 3526 Action* appendLinkHostActions(ActionList &AL) override { return AL.back(); } 3527 3528 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3529 }; 3530 3531 /// 3532 /// TODO: Add the implementation for other specialized builders here. 3533 /// 3534 3535 /// Specialized builders being used by this offloading action builder. 3536 SmallVector<DeviceActionBuilder *, 4> SpecializedBuilders; 3537 3538 /// Flag set to true if all valid builders allow file bundling/unbundling. 3539 bool CanUseBundler; 3540 3541 public: 3542 OffloadingActionBuilder(Compilation &C, DerivedArgList &Args, 3543 const Driver::InputList &Inputs) 3544 : C(C) { 3545 // Create a specialized builder for each device toolchain. 3546 3547 IsValid = true; 3548 3549 // Create a specialized builder for CUDA. 3550 SpecializedBuilders.push_back(new CudaActionBuilder(C, Args, Inputs)); 3551 3552 // Create a specialized builder for HIP. 3553 SpecializedBuilders.push_back(new HIPActionBuilder(C, Args, Inputs)); 3554 3555 // 3556 // TODO: Build other specialized builders here. 3557 // 3558 3559 // Initialize all the builders, keeping track of errors. If all valid 3560 // builders agree that we can use bundling, set the flag to true. 3561 unsigned ValidBuilders = 0u; 3562 unsigned ValidBuildersSupportingBundling = 0u; 3563 for (auto *SB : SpecializedBuilders) { 3564 IsValid = IsValid && !SB->initialize(); 3565 3566 // Update the counters if the builder is valid. 3567 if (SB->isValid()) { 3568 ++ValidBuilders; 3569 if (SB->canUseBundlerUnbundler()) 3570 ++ValidBuildersSupportingBundling; 3571 } 3572 } 3573 CanUseBundler = 3574 ValidBuilders && ValidBuilders == ValidBuildersSupportingBundling; 3575 } 3576 3577 ~OffloadingActionBuilder() { 3578 for (auto *SB : SpecializedBuilders) 3579 delete SB; 3580 } 3581 3582 /// Record a host action and its originating input argument. 3583 void recordHostAction(Action *HostAction, const Arg *InputArg) { 3584 assert(HostAction && "Invalid host action"); 3585 assert(InputArg && "Invalid input argument"); 3586 auto Loc = HostActionToInputArgMap.find(HostAction); 3587 if (Loc == HostActionToInputArgMap.end()) 3588 HostActionToInputArgMap[HostAction] = InputArg; 3589 assert(HostActionToInputArgMap[HostAction] == InputArg && 3590 "host action mapped to multiple input arguments"); 3591 } 3592 3593 /// Generate an action that adds device dependences (if any) to a host action. 3594 /// If no device dependence actions exist, just return the host action \a 3595 /// HostAction. If an error is found or if no builder requires the host action 3596 /// to be generated, return nullptr. 3597 Action * 3598 addDeviceDependencesToHostAction(Action *HostAction, const Arg *InputArg, 3599 phases::ID CurPhase, phases::ID FinalPhase, 3600 DeviceActionBuilder::PhasesTy &Phases) { 3601 if (!IsValid) 3602 return nullptr; 3603 3604 if (SpecializedBuilders.empty()) 3605 return HostAction; 3606 3607 assert(HostAction && "Invalid host action!"); 3608 recordHostAction(HostAction, InputArg); 3609 3610 OffloadAction::DeviceDependences DDeps; 3611 // Check if all the programming models agree we should not emit the host 3612 // action. Also, keep track of the offloading kinds employed. 3613 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3614 unsigned InactiveBuilders = 0u; 3615 unsigned IgnoringBuilders = 0u; 3616 for (auto *SB : SpecializedBuilders) { 3617 if (!SB->isValid()) { 3618 ++InactiveBuilders; 3619 continue; 3620 } 3621 3622 auto RetCode = 3623 SB->getDeviceDependences(DDeps, CurPhase, FinalPhase, Phases); 3624 3625 // If the builder explicitly says the host action should be ignored, 3626 // we need to increment the variable that tracks the builders that request 3627 // the host object to be ignored. 3628 if (RetCode == DeviceActionBuilder::ABRT_Ignore_Host) 3629 ++IgnoringBuilders; 3630 3631 // Unless the builder was inactive for this action, we have to record the 3632 // offload kind because the host will have to use it. 3633 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3634 OffloadKind |= SB->getAssociatedOffloadKind(); 3635 } 3636 3637 // If all builders agree that the host object should be ignored, just return 3638 // nullptr. 3639 if (IgnoringBuilders && 3640 SpecializedBuilders.size() == (InactiveBuilders + IgnoringBuilders)) 3641 return nullptr; 3642 3643 if (DDeps.getActions().empty()) 3644 return HostAction; 3645 3646 // We have dependences we need to bundle together. We use an offload action 3647 // for that. 3648 OffloadAction::HostDependence HDep( 3649 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3650 /*BoundArch=*/nullptr, DDeps); 3651 return C.MakeAction<OffloadAction>(HDep, DDeps); 3652 } 3653 3654 /// Generate an action that adds a host dependence to a device action. The 3655 /// results will be kept in this action builder. Return true if an error was 3656 /// found. 3657 bool addHostDependenceToDeviceActions(Action *&HostAction, 3658 const Arg *InputArg) { 3659 if (!IsValid) 3660 return true; 3661 3662 recordHostAction(HostAction, InputArg); 3663 3664 // If we are supporting bundling/unbundling and the current action is an 3665 // input action of non-source file, we replace the host action by the 3666 // unbundling action. The bundler tool has the logic to detect if an input 3667 // is a bundle or not and if the input is not a bundle it assumes it is a 3668 // host file. Therefore it is safe to create an unbundling action even if 3669 // the input is not a bundle. 3670 if (CanUseBundler && isa<InputAction>(HostAction) && 3671 InputArg->getOption().getKind() == llvm::opt::Option::InputClass && 3672 (!types::isSrcFile(HostAction->getType()) || 3673 HostAction->getType() == types::TY_PP_HIP)) { 3674 auto UnbundlingHostAction = 3675 C.MakeAction<OffloadUnbundlingJobAction>(HostAction); 3676 UnbundlingHostAction->registerDependentActionInfo( 3677 C.getSingleOffloadToolChain<Action::OFK_Host>(), 3678 /*BoundArch=*/StringRef(), Action::OFK_Host); 3679 HostAction = UnbundlingHostAction; 3680 recordHostAction(HostAction, InputArg); 3681 } 3682 3683 assert(HostAction && "Invalid host action!"); 3684 3685 // Register the offload kinds that are used. 3686 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3687 for (auto *SB : SpecializedBuilders) { 3688 if (!SB->isValid()) 3689 continue; 3690 3691 auto RetCode = SB->addDeviceDependences(HostAction); 3692 3693 // Host dependences for device actions are not compatible with that same 3694 // action being ignored. 3695 assert(RetCode != DeviceActionBuilder::ABRT_Ignore_Host && 3696 "Host dependence not expected to be ignored.!"); 3697 3698 // Unless the builder was inactive for this action, we have to record the 3699 // offload kind because the host will have to use it. 3700 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3701 OffloadKind |= SB->getAssociatedOffloadKind(); 3702 } 3703 3704 // Do not use unbundler if the Host does not depend on device action. 3705 if (OffloadKind == Action::OFK_None && CanUseBundler) 3706 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) 3707 HostAction = UA->getInputs().back(); 3708 3709 return false; 3710 } 3711 3712 /// Add the offloading top level actions to the provided action list. This 3713 /// function can replace the host action by a bundling action if the 3714 /// programming models allow it. 3715 bool appendTopLevelActions(ActionList &AL, Action *HostAction, 3716 const Arg *InputArg) { 3717 if (HostAction) 3718 recordHostAction(HostAction, InputArg); 3719 3720 // Get the device actions to be appended. 3721 ActionList OffloadAL; 3722 for (auto *SB : SpecializedBuilders) { 3723 if (!SB->isValid()) 3724 continue; 3725 SB->appendTopLevelActions(OffloadAL); 3726 } 3727 3728 // If we can use the bundler, replace the host action by the bundling one in 3729 // the resulting list. Otherwise, just append the device actions. For 3730 // device only compilation, HostAction is a null pointer, therefore only do 3731 // this when HostAction is not a null pointer. 3732 if (CanUseBundler && HostAction && 3733 HostAction->getType() != types::TY_Nothing && !OffloadAL.empty()) { 3734 // Add the host action to the list in order to create the bundling action. 3735 OffloadAL.push_back(HostAction); 3736 3737 // We expect that the host action was just appended to the action list 3738 // before this method was called. 3739 assert(HostAction == AL.back() && "Host action not in the list??"); 3740 HostAction = C.MakeAction<OffloadBundlingJobAction>(OffloadAL); 3741 recordHostAction(HostAction, InputArg); 3742 AL.back() = HostAction; 3743 } else 3744 AL.append(OffloadAL.begin(), OffloadAL.end()); 3745 3746 // Propagate to the current host action (if any) the offload information 3747 // associated with the current input. 3748 if (HostAction) 3749 HostAction->propagateHostOffloadInfo(InputArgToOffloadKindMap[InputArg], 3750 /*BoundArch=*/nullptr); 3751 return false; 3752 } 3753 3754 void appendDeviceLinkActions(ActionList &AL) { 3755 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3756 if (!SB->isValid()) 3757 continue; 3758 SB->appendLinkDeviceActions(AL); 3759 } 3760 } 3761 3762 Action *makeHostLinkAction() { 3763 // Build a list of device linking actions. 3764 ActionList DeviceAL; 3765 appendDeviceLinkActions(DeviceAL); 3766 if (DeviceAL.empty()) 3767 return nullptr; 3768 3769 // Let builders add host linking actions. 3770 Action* HA = nullptr; 3771 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3772 if (!SB->isValid()) 3773 continue; 3774 HA = SB->appendLinkHostActions(DeviceAL); 3775 // This created host action has no originating input argument, therefore 3776 // needs to set its offloading kind directly. 3777 if (HA) 3778 HA->propagateHostOffloadInfo(SB->getAssociatedOffloadKind(), 3779 /*BoundArch=*/nullptr); 3780 } 3781 return HA; 3782 } 3783 3784 /// Processes the host linker action. This currently consists of replacing it 3785 /// with an offload action if there are device link objects and propagate to 3786 /// the host action all the offload kinds used in the current compilation. The 3787 /// resulting action is returned. 3788 Action *processHostLinkAction(Action *HostAction) { 3789 // Add all the dependences from the device linking actions. 3790 OffloadAction::DeviceDependences DDeps; 3791 for (auto *SB : SpecializedBuilders) { 3792 if (!SB->isValid()) 3793 continue; 3794 3795 SB->appendLinkDependences(DDeps); 3796 } 3797 3798 // Calculate all the offload kinds used in the current compilation. 3799 unsigned ActiveOffloadKinds = 0u; 3800 for (auto &I : InputArgToOffloadKindMap) 3801 ActiveOffloadKinds |= I.second; 3802 3803 // If we don't have device dependencies, we don't have to create an offload 3804 // action. 3805 if (DDeps.getActions().empty()) { 3806 // Set all the active offloading kinds to the link action. Given that it 3807 // is a link action it is assumed to depend on all actions generated so 3808 // far. 3809 HostAction->setHostOffloadInfo(ActiveOffloadKinds, 3810 /*BoundArch=*/nullptr); 3811 // Propagate active offloading kinds for each input to the link action. 3812 // Each input may have different active offloading kind. 3813 for (auto *A : HostAction->inputs()) { 3814 auto ArgLoc = HostActionToInputArgMap.find(A); 3815 if (ArgLoc == HostActionToInputArgMap.end()) 3816 continue; 3817 auto OFKLoc = InputArgToOffloadKindMap.find(ArgLoc->second); 3818 if (OFKLoc == InputArgToOffloadKindMap.end()) 3819 continue; 3820 A->propagateHostOffloadInfo(OFKLoc->second, /*BoundArch=*/nullptr); 3821 } 3822 return HostAction; 3823 } 3824 3825 // Create the offload action with all dependences. When an offload action 3826 // is created the kinds are propagated to the host action, so we don't have 3827 // to do that explicitly here. 3828 OffloadAction::HostDependence HDep( 3829 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3830 /*BoundArch*/ nullptr, ActiveOffloadKinds); 3831 return C.MakeAction<OffloadAction>(HDep, DDeps); 3832 } 3833 }; 3834 } // anonymous namespace. 3835 3836 void Driver::handleArguments(Compilation &C, DerivedArgList &Args, 3837 const InputList &Inputs, 3838 ActionList &Actions) const { 3839 3840 // Ignore /Yc/Yu if both /Yc and /Yu passed but with different filenames. 3841 Arg *YcArg = Args.getLastArg(options::OPT__SLASH_Yc); 3842 Arg *YuArg = Args.getLastArg(options::OPT__SLASH_Yu); 3843 if (YcArg && YuArg && strcmp(YcArg->getValue(), YuArg->getValue()) != 0) { 3844 Diag(clang::diag::warn_drv_ycyu_different_arg_clang_cl); 3845 Args.eraseArg(options::OPT__SLASH_Yc); 3846 Args.eraseArg(options::OPT__SLASH_Yu); 3847 YcArg = YuArg = nullptr; 3848 } 3849 if (YcArg && Inputs.size() > 1) { 3850 Diag(clang::diag::warn_drv_yc_multiple_inputs_clang_cl); 3851 Args.eraseArg(options::OPT__SLASH_Yc); 3852 YcArg = nullptr; 3853 } 3854 3855 Arg *FinalPhaseArg; 3856 phases::ID FinalPhase = getFinalPhase(Args, &FinalPhaseArg); 3857 3858 if (FinalPhase == phases::Link) { 3859 // Emitting LLVM while linking disabled except in HIPAMD Toolchain 3860 if (Args.hasArg(options::OPT_emit_llvm) && !Args.hasArg(options::OPT_hip_link)) 3861 Diag(clang::diag::err_drv_emit_llvm_link); 3862 if (IsCLMode() && LTOMode != LTOK_None && 3863 !Args.getLastArgValue(options::OPT_fuse_ld_EQ) 3864 .equals_insensitive("lld")) 3865 Diag(clang::diag::err_drv_lto_without_lld); 3866 } 3867 3868 if (FinalPhase == phases::Preprocess || Args.hasArg(options::OPT__SLASH_Y_)) { 3869 // If only preprocessing or /Y- is used, all pch handling is disabled. 3870 // Rather than check for it everywhere, just remove clang-cl pch-related 3871 // flags here. 3872 Args.eraseArg(options::OPT__SLASH_Fp); 3873 Args.eraseArg(options::OPT__SLASH_Yc); 3874 Args.eraseArg(options::OPT__SLASH_Yu); 3875 YcArg = YuArg = nullptr; 3876 } 3877 3878 unsigned LastPLSize = 0; 3879 for (auto &I : Inputs) { 3880 types::ID InputType = I.first; 3881 const Arg *InputArg = I.second; 3882 3883 auto PL = types::getCompilationPhases(InputType); 3884 LastPLSize = PL.size(); 3885 3886 // If the first step comes after the final phase we are doing as part of 3887 // this compilation, warn the user about it. 3888 phases::ID InitialPhase = PL[0]; 3889 if (InitialPhase > FinalPhase) { 3890 if (InputArg->isClaimed()) 3891 continue; 3892 3893 // Claim here to avoid the more general unused warning. 3894 InputArg->claim(); 3895 3896 // Suppress all unused style warnings with -Qunused-arguments 3897 if (Args.hasArg(options::OPT_Qunused_arguments)) 3898 continue; 3899 3900 // Special case when final phase determined by binary name, rather than 3901 // by a command-line argument with a corresponding Arg. 3902 if (CCCIsCPP()) 3903 Diag(clang::diag::warn_drv_input_file_unused_by_cpp) 3904 << InputArg->getAsString(Args) << getPhaseName(InitialPhase); 3905 // Special case '-E' warning on a previously preprocessed file to make 3906 // more sense. 3907 else if (InitialPhase == phases::Compile && 3908 (Args.getLastArg(options::OPT__SLASH_EP, 3909 options::OPT__SLASH_P) || 3910 Args.getLastArg(options::OPT_E) || 3911 Args.getLastArg(options::OPT_M, options::OPT_MM)) && 3912 getPreprocessedType(InputType) == types::TY_INVALID) 3913 Diag(clang::diag::warn_drv_preprocessed_input_file_unused) 3914 << InputArg->getAsString(Args) << !!FinalPhaseArg 3915 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3916 else 3917 Diag(clang::diag::warn_drv_input_file_unused) 3918 << InputArg->getAsString(Args) << getPhaseName(InitialPhase) 3919 << !!FinalPhaseArg 3920 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3921 continue; 3922 } 3923 3924 if (YcArg) { 3925 // Add a separate precompile phase for the compile phase. 3926 if (FinalPhase >= phases::Compile) { 3927 const types::ID HeaderType = lookupHeaderTypeForSourceType(InputType); 3928 // Build the pipeline for the pch file. 3929 Action *ClangClPch = C.MakeAction<InputAction>(*InputArg, HeaderType); 3930 for (phases::ID Phase : types::getCompilationPhases(HeaderType)) 3931 ClangClPch = ConstructPhaseAction(C, Args, Phase, ClangClPch); 3932 assert(ClangClPch); 3933 Actions.push_back(ClangClPch); 3934 // The driver currently exits after the first failed command. This 3935 // relies on that behavior, to make sure if the pch generation fails, 3936 // the main compilation won't run. 3937 // FIXME: If the main compilation fails, the PCH generation should 3938 // probably not be considered successful either. 3939 } 3940 } 3941 } 3942 3943 // If we are linking, claim any options which are obviously only used for 3944 // compilation. 3945 // FIXME: Understand why the last Phase List length is used here. 3946 if (FinalPhase == phases::Link && LastPLSize == 1) { 3947 Args.ClaimAllArgs(options::OPT_CompileOnly_Group); 3948 Args.ClaimAllArgs(options::OPT_cl_compile_Group); 3949 } 3950 } 3951 3952 void Driver::BuildActions(Compilation &C, DerivedArgList &Args, 3953 const InputList &Inputs, ActionList &Actions) const { 3954 llvm::PrettyStackTraceString CrashInfo("Building compilation actions"); 3955 3956 if (!SuppressMissingInputWarning && Inputs.empty()) { 3957 Diag(clang::diag::err_drv_no_input_files); 3958 return; 3959 } 3960 3961 // Diagnose misuse of /Fo. 3962 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fo)) { 3963 StringRef V = A->getValue(); 3964 if (Inputs.size() > 1 && !V.empty() && 3965 !llvm::sys::path::is_separator(V.back())) { 3966 // Check whether /Fo tries to name an output file for multiple inputs. 3967 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 3968 << A->getSpelling() << V; 3969 Args.eraseArg(options::OPT__SLASH_Fo); 3970 } 3971 } 3972 3973 // Diagnose misuse of /Fa. 3974 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fa)) { 3975 StringRef V = A->getValue(); 3976 if (Inputs.size() > 1 && !V.empty() && 3977 !llvm::sys::path::is_separator(V.back())) { 3978 // Check whether /Fa tries to name an asm file for multiple inputs. 3979 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 3980 << A->getSpelling() << V; 3981 Args.eraseArg(options::OPT__SLASH_Fa); 3982 } 3983 } 3984 3985 // Diagnose misuse of /o. 3986 if (Arg *A = Args.getLastArg(options::OPT__SLASH_o)) { 3987 if (A->getValue()[0] == '\0') { 3988 // It has to have a value. 3989 Diag(clang::diag::err_drv_missing_argument) << A->getSpelling() << 1; 3990 Args.eraseArg(options::OPT__SLASH_o); 3991 } 3992 } 3993 3994 handleArguments(C, Args, Inputs, Actions); 3995 3996 bool UseNewOffloadingDriver = 3997 C.isOffloadingHostKind(Action::OFK_OpenMP) || 3998 Args.hasFlag(options::OPT_offload_new_driver, 3999 options::OPT_no_offload_new_driver, false); 4000 4001 // Builder to be used to build offloading actions. 4002 std::unique_ptr<OffloadingActionBuilder> OffloadBuilder = 4003 !UseNewOffloadingDriver 4004 ? std::make_unique<OffloadingActionBuilder>(C, Args, Inputs) 4005 : nullptr; 4006 4007 // Construct the actions to perform. 4008 ExtractAPIJobAction *ExtractAPIAction = nullptr; 4009 ActionList LinkerInputs; 4010 ActionList MergerInputs; 4011 4012 for (auto &I : Inputs) { 4013 types::ID InputType = I.first; 4014 const Arg *InputArg = I.second; 4015 4016 auto PL = types::getCompilationPhases(*this, Args, InputType); 4017 if (PL.empty()) 4018 continue; 4019 4020 auto FullPL = types::getCompilationPhases(InputType); 4021 4022 // Build the pipeline for this file. 4023 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4024 4025 // Use the current host action in any of the offloading actions, if 4026 // required. 4027 if (!UseNewOffloadingDriver) 4028 if (OffloadBuilder->addHostDependenceToDeviceActions(Current, InputArg)) 4029 break; 4030 4031 for (phases::ID Phase : PL) { 4032 4033 // Add any offload action the host action depends on. 4034 if (!UseNewOffloadingDriver) 4035 Current = OffloadBuilder->addDeviceDependencesToHostAction( 4036 Current, InputArg, Phase, PL.back(), FullPL); 4037 if (!Current) 4038 break; 4039 4040 // Queue linker inputs. 4041 if (Phase == phases::Link) { 4042 assert(Phase == PL.back() && "linking must be final compilation step."); 4043 // We don't need to generate additional link commands if emitting AMD bitcode 4044 if (!(C.getInputArgs().hasArg(options::OPT_hip_link) && 4045 (C.getInputArgs().hasArg(options::OPT_emit_llvm)))) 4046 LinkerInputs.push_back(Current); 4047 Current = nullptr; 4048 break; 4049 } 4050 4051 // TODO: Consider removing this because the merged may not end up being 4052 // the final Phase in the pipeline. Perhaps the merged could just merge 4053 // and then pass an artifact of some sort to the Link Phase. 4054 // Queue merger inputs. 4055 if (Phase == phases::IfsMerge) { 4056 assert(Phase == PL.back() && "merging must be final compilation step."); 4057 MergerInputs.push_back(Current); 4058 Current = nullptr; 4059 break; 4060 } 4061 4062 if (Phase == phases::Precompile && ExtractAPIAction) { 4063 ExtractAPIAction->addHeaderInput(Current); 4064 Current = nullptr; 4065 break; 4066 } 4067 4068 // FIXME: Should we include any prior module file outputs as inputs of 4069 // later actions in the same command line? 4070 4071 // Otherwise construct the appropriate action. 4072 Action *NewCurrent = ConstructPhaseAction(C, Args, Phase, Current); 4073 4074 // We didn't create a new action, so we will just move to the next phase. 4075 if (NewCurrent == Current) 4076 continue; 4077 4078 if (auto *EAA = dyn_cast<ExtractAPIJobAction>(NewCurrent)) 4079 ExtractAPIAction = EAA; 4080 4081 Current = NewCurrent; 4082 4083 // Use the current host action in any of the offloading actions, if 4084 // required. 4085 if (!UseNewOffloadingDriver) 4086 if (OffloadBuilder->addHostDependenceToDeviceActions(Current, InputArg)) 4087 break; 4088 4089 // Try to build the offloading actions and add the result as a dependency 4090 // to the host. 4091 if (UseNewOffloadingDriver) 4092 Current = BuildOffloadingActions(C, Args, I, Current); 4093 4094 if (Current->getType() == types::TY_Nothing) 4095 break; 4096 } 4097 4098 // If we ended with something, add to the output list. 4099 if (Current) 4100 Actions.push_back(Current); 4101 4102 // Add any top level actions generated for offloading. 4103 if (!UseNewOffloadingDriver) 4104 OffloadBuilder->appendTopLevelActions(Actions, Current, InputArg); 4105 else if (Current) 4106 Current->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4107 /*BoundArch=*/nullptr); 4108 } 4109 4110 // Add a link action if necessary. 4111 4112 if (LinkerInputs.empty()) { 4113 Arg *FinalPhaseArg; 4114 if (getFinalPhase(Args, &FinalPhaseArg) == phases::Link) 4115 if (!UseNewOffloadingDriver) 4116 OffloadBuilder->appendDeviceLinkActions(Actions); 4117 } 4118 4119 if (!LinkerInputs.empty()) { 4120 if (!UseNewOffloadingDriver) 4121 if (Action *Wrapper = OffloadBuilder->makeHostLinkAction()) 4122 LinkerInputs.push_back(Wrapper); 4123 Action *LA; 4124 // Check if this Linker Job should emit a static library. 4125 if (ShouldEmitStaticLibrary(Args)) { 4126 LA = C.MakeAction<StaticLibJobAction>(LinkerInputs, types::TY_Image); 4127 } else if (UseNewOffloadingDriver || 4128 Args.hasArg(options::OPT_offload_link)) { 4129 LA = C.MakeAction<LinkerWrapperJobAction>(LinkerInputs, types::TY_Image); 4130 LA->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4131 /*BoundArch=*/nullptr); 4132 } else { 4133 LA = C.MakeAction<LinkJobAction>(LinkerInputs, types::TY_Image); 4134 } 4135 if (!UseNewOffloadingDriver) 4136 LA = OffloadBuilder->processHostLinkAction(LA); 4137 Actions.push_back(LA); 4138 } 4139 4140 // Add an interface stubs merge action if necessary. 4141 if (!MergerInputs.empty()) 4142 Actions.push_back( 4143 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4144 4145 if (Args.hasArg(options::OPT_emit_interface_stubs)) { 4146 auto PhaseList = types::getCompilationPhases( 4147 types::TY_IFS_CPP, 4148 Args.hasArg(options::OPT_c) ? phases::Compile : phases::IfsMerge); 4149 4150 ActionList MergerInputs; 4151 4152 for (auto &I : Inputs) { 4153 types::ID InputType = I.first; 4154 const Arg *InputArg = I.second; 4155 4156 // Currently clang and the llvm assembler do not support generating symbol 4157 // stubs from assembly, so we skip the input on asm files. For ifs files 4158 // we rely on the normal pipeline setup in the pipeline setup code above. 4159 if (InputType == types::TY_IFS || InputType == types::TY_PP_Asm || 4160 InputType == types::TY_Asm) 4161 continue; 4162 4163 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4164 4165 for (auto Phase : PhaseList) { 4166 switch (Phase) { 4167 default: 4168 llvm_unreachable( 4169 "IFS Pipeline can only consist of Compile followed by IfsMerge."); 4170 case phases::Compile: { 4171 // Only IfsMerge (llvm-ifs) can handle .o files by looking for ifs 4172 // files where the .o file is located. The compile action can not 4173 // handle this. 4174 if (InputType == types::TY_Object) 4175 break; 4176 4177 Current = C.MakeAction<CompileJobAction>(Current, types::TY_IFS_CPP); 4178 break; 4179 } 4180 case phases::IfsMerge: { 4181 assert(Phase == PhaseList.back() && 4182 "merging must be final compilation step."); 4183 MergerInputs.push_back(Current); 4184 Current = nullptr; 4185 break; 4186 } 4187 } 4188 } 4189 4190 // If we ended with something, add to the output list. 4191 if (Current) 4192 Actions.push_back(Current); 4193 } 4194 4195 // Add an interface stubs merge action if necessary. 4196 if (!MergerInputs.empty()) 4197 Actions.push_back( 4198 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4199 } 4200 4201 // If --print-supported-cpus, -mcpu=? or -mtune=? is specified, build a custom 4202 // Compile phase that prints out supported cpu models and quits. 4203 if (Arg *A = Args.getLastArg(options::OPT_print_supported_cpus)) { 4204 // Use the -mcpu=? flag as the dummy input to cc1. 4205 Actions.clear(); 4206 Action *InputAc = C.MakeAction<InputAction>(*A, types::TY_C); 4207 Actions.push_back( 4208 C.MakeAction<PrecompileJobAction>(InputAc, types::TY_Nothing)); 4209 for (auto &I : Inputs) 4210 I.second->claim(); 4211 } 4212 4213 // Claim ignored clang-cl options. 4214 Args.ClaimAllArgs(options::OPT_cl_ignored_Group); 4215 } 4216 4217 /// Returns the canonical name for the offloading architecture when using a HIP 4218 /// or CUDA architecture. 4219 static StringRef getCanonicalArchString(Compilation &C, 4220 const llvm::opt::DerivedArgList &Args, 4221 StringRef ArchStr, 4222 const llvm::Triple &Triple, 4223 bool SuppressError = false) { 4224 // Lookup the CUDA / HIP architecture string. Only report an error if we were 4225 // expecting the triple to be only NVPTX / AMDGPU. 4226 CudaArch Arch = StringToCudaArch(getProcessorFromTargetID(Triple, ArchStr)); 4227 if (!SuppressError && Triple.isNVPTX() && 4228 (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch))) { 4229 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4230 << "CUDA" << ArchStr; 4231 return StringRef(); 4232 } else if (!SuppressError && Triple.isAMDGPU() && 4233 (Arch == CudaArch::UNKNOWN || !IsAMDGpuArch(Arch))) { 4234 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4235 << "HIP" << ArchStr; 4236 return StringRef(); 4237 } 4238 4239 if (IsNVIDIAGpuArch(Arch)) 4240 return Args.MakeArgStringRef(CudaArchToString(Arch)); 4241 4242 if (IsAMDGpuArch(Arch)) { 4243 llvm::StringMap<bool> Features; 4244 auto HIPTriple = getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()); 4245 if (!HIPTriple) 4246 return StringRef(); 4247 auto Arch = parseTargetID(*HIPTriple, ArchStr, &Features); 4248 if (!Arch) { 4249 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << ArchStr; 4250 C.setContainsError(); 4251 return StringRef(); 4252 } 4253 return Args.MakeArgStringRef(getCanonicalTargetID(*Arch, Features)); 4254 } 4255 4256 // If the input isn't CUDA or HIP just return the architecture. 4257 return ArchStr; 4258 } 4259 4260 /// Checks if the set offloading architectures does not conflict. Returns the 4261 /// incompatible pair if a conflict occurs. 4262 static std::optional<std::pair<llvm::StringRef, llvm::StringRef>> 4263 getConflictOffloadArchCombination(const llvm::DenseSet<StringRef> &Archs, 4264 Action::OffloadKind Kind) { 4265 if (Kind != Action::OFK_HIP) 4266 return std::nullopt; 4267 4268 std::set<StringRef> ArchSet; 4269 llvm::copy(Archs, std::inserter(ArchSet, ArchSet.begin())); 4270 return getConflictTargetIDCombination(ArchSet); 4271 } 4272 4273 llvm::DenseSet<StringRef> 4274 Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args, 4275 Action::OffloadKind Kind, const ToolChain *TC, 4276 bool SuppressError) const { 4277 if (!TC) 4278 TC = &C.getDefaultToolChain(); 4279 4280 // --offload and --offload-arch options are mutually exclusive. 4281 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 4282 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 4283 options::OPT_no_offload_arch_EQ)) { 4284 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) 4285 << "--offload" 4286 << (Args.hasArgNoClaim(options::OPT_offload_arch_EQ) 4287 ? "--offload-arch" 4288 : "--no-offload-arch"); 4289 } 4290 4291 if (KnownArchs.find(TC) != KnownArchs.end()) 4292 return KnownArchs.lookup(TC); 4293 4294 llvm::DenseSet<StringRef> Archs; 4295 for (auto *Arg : Args) { 4296 // Extract any '--[no-]offload-arch' arguments intended for this toolchain. 4297 std::unique_ptr<llvm::opt::Arg> ExtractedArg = nullptr; 4298 if (Arg->getOption().matches(options::OPT_Xopenmp_target_EQ) && 4299 ToolChain::getOpenMPTriple(Arg->getValue(0)) == TC->getTriple()) { 4300 Arg->claim(); 4301 unsigned Index = Args.getBaseArgs().MakeIndex(Arg->getValue(1)); 4302 ExtractedArg = getOpts().ParseOneArg(Args, Index); 4303 Arg = ExtractedArg.get(); 4304 } 4305 4306 // Add or remove the seen architectures in order of appearance. If an 4307 // invalid architecture is given we simply exit. 4308 if (Arg->getOption().matches(options::OPT_offload_arch_EQ)) { 4309 for (StringRef Arch : llvm::split(Arg->getValue(), ",")) { 4310 if (Arch == "native" || Arch.empty()) { 4311 auto GPUsOrErr = TC->getSystemGPUArchs(Args); 4312 if (!GPUsOrErr) { 4313 if (SuppressError) 4314 llvm::consumeError(GPUsOrErr.takeError()); 4315 else 4316 TC->getDriver().Diag(diag::err_drv_undetermined_gpu_arch) 4317 << llvm::Triple::getArchTypeName(TC->getArch()) 4318 << llvm::toString(GPUsOrErr.takeError()) << "--offload-arch"; 4319 continue; 4320 } 4321 4322 for (auto ArchStr : *GPUsOrErr) { 4323 Archs.insert( 4324 getCanonicalArchString(C, Args, Args.MakeArgString(ArchStr), 4325 TC->getTriple(), SuppressError)); 4326 } 4327 } else { 4328 StringRef ArchStr = getCanonicalArchString( 4329 C, Args, Arch, TC->getTriple(), SuppressError); 4330 if (ArchStr.empty()) 4331 return Archs; 4332 Archs.insert(ArchStr); 4333 } 4334 } 4335 } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) { 4336 for (StringRef Arch : llvm::split(Arg->getValue(), ",")) { 4337 if (Arch == "all") { 4338 Archs.clear(); 4339 } else { 4340 StringRef ArchStr = getCanonicalArchString( 4341 C, Args, Arch, TC->getTriple(), SuppressError); 4342 if (ArchStr.empty()) 4343 return Archs; 4344 Archs.erase(ArchStr); 4345 } 4346 } 4347 } 4348 } 4349 4350 if (auto ConflictingArchs = getConflictOffloadArchCombination(Archs, Kind)) { 4351 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 4352 << ConflictingArchs->first << ConflictingArchs->second; 4353 C.setContainsError(); 4354 } 4355 4356 // Skip filling defaults if we're just querying what is availible. 4357 if (SuppressError) 4358 return Archs; 4359 4360 if (Archs.empty()) { 4361 if (Kind == Action::OFK_Cuda) 4362 Archs.insert(CudaArchToString(CudaArch::CudaDefault)); 4363 else if (Kind == Action::OFK_HIP) 4364 Archs.insert(CudaArchToString(CudaArch::HIPDefault)); 4365 else if (Kind == Action::OFK_OpenMP) 4366 Archs.insert(StringRef()); 4367 } else { 4368 Args.ClaimAllArgs(options::OPT_offload_arch_EQ); 4369 Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ); 4370 } 4371 4372 return Archs; 4373 } 4374 4375 Action *Driver::BuildOffloadingActions(Compilation &C, 4376 llvm::opt::DerivedArgList &Args, 4377 const InputTy &Input, 4378 Action *HostAction) const { 4379 // Don't build offloading actions if explicitly disabled or we do not have a 4380 // valid source input and compile action to embed it in. If preprocessing only 4381 // ignore embedding. 4382 if (offloadHostOnly() || !types::isSrcFile(Input.first) || 4383 !(isa<CompileJobAction>(HostAction) || 4384 getFinalPhase(Args) == phases::Preprocess)) 4385 return HostAction; 4386 4387 ActionList OffloadActions; 4388 OffloadAction::DeviceDependences DDeps; 4389 4390 const Action::OffloadKind OffloadKinds[] = { 4391 Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP}; 4392 4393 for (Action::OffloadKind Kind : OffloadKinds) { 4394 SmallVector<const ToolChain *, 2> ToolChains; 4395 ActionList DeviceActions; 4396 4397 auto TCRange = C.getOffloadToolChains(Kind); 4398 for (auto TI = TCRange.first, TE = TCRange.second; TI != TE; ++TI) 4399 ToolChains.push_back(TI->second); 4400 4401 if (ToolChains.empty()) 4402 continue; 4403 4404 types::ID InputType = Input.first; 4405 const Arg *InputArg = Input.second; 4406 4407 // The toolchain can be active for unsupported file types. 4408 if ((Kind == Action::OFK_Cuda && !types::isCuda(InputType)) || 4409 (Kind == Action::OFK_HIP && !types::isHIP(InputType))) 4410 continue; 4411 4412 // Get the product of all bound architectures and toolchains. 4413 SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs; 4414 for (const ToolChain *TC : ToolChains) 4415 for (StringRef Arch : getOffloadArchs(C, Args, Kind, TC)) 4416 TCAndArchs.push_back(std::make_pair(TC, Arch)); 4417 4418 for (unsigned I = 0, E = TCAndArchs.size(); I != E; ++I) 4419 DeviceActions.push_back(C.MakeAction<InputAction>(*InputArg, InputType)); 4420 4421 if (DeviceActions.empty()) 4422 return HostAction; 4423 4424 auto PL = types::getCompilationPhases(*this, Args, InputType); 4425 4426 for (phases::ID Phase : PL) { 4427 if (Phase == phases::Link) { 4428 assert(Phase == PL.back() && "linking must be final compilation step."); 4429 break; 4430 } 4431 4432 auto TCAndArch = TCAndArchs.begin(); 4433 for (Action *&A : DeviceActions) { 4434 if (A->getType() == types::TY_Nothing) 4435 continue; 4436 4437 // Propagate the ToolChain so we can use it in ConstructPhaseAction. 4438 A->propagateDeviceOffloadInfo(Kind, TCAndArch->second.data(), 4439 TCAndArch->first); 4440 A = ConstructPhaseAction(C, Args, Phase, A, Kind); 4441 4442 if (isa<CompileJobAction>(A) && isa<CompileJobAction>(HostAction) && 4443 Kind == Action::OFK_OpenMP && 4444 HostAction->getType() != types::TY_Nothing) { 4445 // OpenMP offloading has a dependency on the host compile action to 4446 // identify which declarations need to be emitted. This shouldn't be 4447 // collapsed with any other actions so we can use it in the device. 4448 HostAction->setCannotBeCollapsedWithNextDependentAction(); 4449 OffloadAction::HostDependence HDep( 4450 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4451 TCAndArch->second.data(), Kind); 4452 OffloadAction::DeviceDependences DDep; 4453 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4454 A = C.MakeAction<OffloadAction>(HDep, DDep); 4455 } 4456 4457 ++TCAndArch; 4458 } 4459 } 4460 4461 // Compiling HIP in non-RDC mode requires linking each action individually. 4462 for (Action *&A : DeviceActions) { 4463 if ((A->getType() != types::TY_Object && 4464 A->getType() != types::TY_LTO_BC) || 4465 Kind != Action::OFK_HIP || 4466 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false)) 4467 continue; 4468 ActionList LinkerInput = {A}; 4469 A = C.MakeAction<LinkJobAction>(LinkerInput, types::TY_Image); 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 (offloadDeviceOnly()) 4483 return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing); 4484 4485 if (OffloadActions.empty()) 4486 return HostAction; 4487 4488 OffloadAction::DeviceDependences DDep; 4489 if (C.isOffloadingHostKind(Action::OFK_Cuda) && 4490 !Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false)) { 4491 // If we are not in RDC-mode we just emit the final CUDA fatbinary for 4492 // each translation unit without requiring any linking. 4493 Action *FatbinAction = 4494 C.MakeAction<LinkJobAction>(OffloadActions, types::TY_CUDA_FATBIN); 4495 DDep.add(*FatbinAction, *C.getSingleOffloadToolChain<Action::OFK_Cuda>(), 4496 nullptr, Action::OFK_Cuda); 4497 } else if (C.isOffloadingHostKind(Action::OFK_HIP) && 4498 !Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4499 false)) { 4500 // If we are not in RDC-mode we just emit the final HIP fatbinary for each 4501 // translation unit, linking each input individually. 4502 Action *FatbinAction = 4503 C.MakeAction<LinkJobAction>(OffloadActions, types::TY_HIP_FATBIN); 4504 DDep.add(*FatbinAction, *C.getSingleOffloadToolChain<Action::OFK_HIP>(), 4505 nullptr, Action::OFK_HIP); 4506 } else { 4507 // Package all the offloading actions into a single output that can be 4508 // embedded in the host and linked. 4509 Action *PackagerAction = 4510 C.MakeAction<OffloadPackagerJobAction>(OffloadActions, types::TY_Image); 4511 DDep.add(*PackagerAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4512 nullptr, C.getActiveOffloadKinds()); 4513 } 4514 4515 // If we are unable to embed a single device output into the host, we need to 4516 // add each device output as a host dependency to ensure they are still built. 4517 bool SingleDeviceOutput = !llvm::any_of(OffloadActions, [](Action *A) { 4518 return A->getType() == types::TY_Nothing; 4519 }) && isa<CompileJobAction>(HostAction); 4520 OffloadAction::HostDependence HDep( 4521 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4522 /*BoundArch=*/nullptr, SingleDeviceOutput ? DDep : DDeps); 4523 return C.MakeAction<OffloadAction>(HDep, SingleDeviceOutput ? DDep : DDeps); 4524 } 4525 4526 Action *Driver::ConstructPhaseAction( 4527 Compilation &C, const ArgList &Args, phases::ID Phase, Action *Input, 4528 Action::OffloadKind TargetDeviceOffloadKind) const { 4529 llvm::PrettyStackTraceString CrashInfo("Constructing phase actions"); 4530 4531 // Some types skip the assembler phase (e.g., llvm-bc), but we can't 4532 // encode this in the steps because the intermediate type depends on 4533 // arguments. Just special case here. 4534 if (Phase == phases::Assemble && Input->getType() != types::TY_PP_Asm) 4535 return Input; 4536 4537 // Build the appropriate action. 4538 switch (Phase) { 4539 case phases::Link: 4540 llvm_unreachable("link action invalid here."); 4541 case phases::IfsMerge: 4542 llvm_unreachable("ifsmerge action invalid here."); 4543 case phases::Preprocess: { 4544 types::ID OutputTy; 4545 // -M and -MM specify the dependency file name by altering the output type, 4546 // -if -MD and -MMD are not specified. 4547 if (Args.hasArg(options::OPT_M, options::OPT_MM) && 4548 !Args.hasArg(options::OPT_MD, options::OPT_MMD)) { 4549 OutputTy = types::TY_Dependencies; 4550 } else { 4551 OutputTy = Input->getType(); 4552 // For these cases, the preprocessor is only translating forms, the Output 4553 // still needs preprocessing. 4554 if (!Args.hasFlag(options::OPT_frewrite_includes, 4555 options::OPT_fno_rewrite_includes, false) && 4556 !Args.hasFlag(options::OPT_frewrite_imports, 4557 options::OPT_fno_rewrite_imports, false) && 4558 !Args.hasFlag(options::OPT_fdirectives_only, 4559 options::OPT_fno_directives_only, false) && 4560 !CCGenDiagnostics) 4561 OutputTy = types::getPreprocessedType(OutputTy); 4562 assert(OutputTy != types::TY_INVALID && 4563 "Cannot preprocess this input type!"); 4564 } 4565 return C.MakeAction<PreprocessJobAction>(Input, OutputTy); 4566 } 4567 case phases::Precompile: { 4568 // API extraction should not generate an actual precompilation action. 4569 if (Args.hasArg(options::OPT_extract_api)) 4570 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4571 4572 types::ID OutputTy = getPrecompiledType(Input->getType()); 4573 assert(OutputTy != types::TY_INVALID && 4574 "Cannot precompile this input type!"); 4575 4576 // If we're given a module name, precompile header file inputs as a 4577 // module, not as a precompiled header. 4578 const char *ModName = nullptr; 4579 if (OutputTy == types::TY_PCH) { 4580 if (Arg *A = Args.getLastArg(options::OPT_fmodule_name_EQ)) 4581 ModName = A->getValue(); 4582 if (ModName) 4583 OutputTy = types::TY_ModuleFile; 4584 } 4585 4586 if (Args.hasArg(options::OPT_fsyntax_only)) { 4587 // Syntax checks should not emit a PCH file 4588 OutputTy = types::TY_Nothing; 4589 } 4590 4591 return C.MakeAction<PrecompileJobAction>(Input, OutputTy); 4592 } 4593 case phases::Compile: { 4594 if (Args.hasArg(options::OPT_fsyntax_only)) 4595 return C.MakeAction<CompileJobAction>(Input, types::TY_Nothing); 4596 if (Args.hasArg(options::OPT_rewrite_objc)) 4597 return C.MakeAction<CompileJobAction>(Input, types::TY_RewrittenObjC); 4598 if (Args.hasArg(options::OPT_rewrite_legacy_objc)) 4599 return C.MakeAction<CompileJobAction>(Input, 4600 types::TY_RewrittenLegacyObjC); 4601 if (Args.hasArg(options::OPT__analyze)) 4602 return C.MakeAction<AnalyzeJobAction>(Input, types::TY_Plist); 4603 if (Args.hasArg(options::OPT__migrate)) 4604 return C.MakeAction<MigrateJobAction>(Input, types::TY_Remap); 4605 if (Args.hasArg(options::OPT_emit_ast)) 4606 return C.MakeAction<CompileJobAction>(Input, types::TY_AST); 4607 if (Args.hasArg(options::OPT_module_file_info)) 4608 return C.MakeAction<CompileJobAction>(Input, types::TY_ModuleFile); 4609 if (Args.hasArg(options::OPT_verify_pch)) 4610 return C.MakeAction<VerifyPCHJobAction>(Input, types::TY_Nothing); 4611 if (Args.hasArg(options::OPT_extract_api)) 4612 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4613 return C.MakeAction<CompileJobAction>(Input, types::TY_LLVM_BC); 4614 } 4615 case phases::Backend: { 4616 if (isUsingLTO() && TargetDeviceOffloadKind == Action::OFK_None) { 4617 types::ID Output = 4618 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4619 return C.MakeAction<BackendJobAction>(Input, Output); 4620 } 4621 if (isUsingLTO(/* IsOffload */ true) && 4622 TargetDeviceOffloadKind != Action::OFK_None) { 4623 types::ID Output = 4624 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4625 return C.MakeAction<BackendJobAction>(Input, Output); 4626 } 4627 if (Args.hasArg(options::OPT_emit_llvm) || 4628 (((Input->getOffloadingToolChain() && 4629 Input->getOffloadingToolChain()->getTriple().isAMDGPU()) || 4630 TargetDeviceOffloadKind == Action::OFK_HIP) && 4631 (Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4632 false) || 4633 TargetDeviceOffloadKind == Action::OFK_OpenMP))) { 4634 types::ID Output = 4635 Args.hasArg(options::OPT_S) && 4636 (TargetDeviceOffloadKind == Action::OFK_None || 4637 offloadDeviceOnly() || 4638 (TargetDeviceOffloadKind == Action::OFK_HIP && 4639 !Args.hasFlag(options::OPT_offload_new_driver, 4640 options::OPT_no_offload_new_driver, false))) 4641 ? types::TY_LLVM_IR 4642 : types::TY_LLVM_BC; 4643 return C.MakeAction<BackendJobAction>(Input, Output); 4644 } 4645 return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm); 4646 } 4647 case phases::Assemble: 4648 return C.MakeAction<AssembleJobAction>(std::move(Input), types::TY_Object); 4649 } 4650 4651 llvm_unreachable("invalid phase in ConstructPhaseAction"); 4652 } 4653 4654 void Driver::BuildJobs(Compilation &C) const { 4655 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 4656 4657 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 4658 4659 // It is an error to provide a -o option if we are making multiple output 4660 // files. There are exceptions: 4661 // 4662 // IfsMergeJob: when generating interface stubs enabled we want to be able to 4663 // generate the stub file at the same time that we generate the real 4664 // library/a.out. So when a .o, .so, etc are the output, with clang interface 4665 // stubs there will also be a .ifs and .ifso at the same location. 4666 // 4667 // CompileJob of type TY_IFS_CPP: when generating interface stubs is enabled 4668 // and -c is passed, we still want to be able to generate a .ifs file while 4669 // we are also generating .o files. So we allow more than one output file in 4670 // this case as well. 4671 // 4672 // OffloadClass of type TY_Nothing: device-only output will place many outputs 4673 // into a single offloading action. We should count all inputs to the action 4674 // as outputs. Also ignore device-only outputs if we're compiling with 4675 // -fsyntax-only. 4676 if (FinalOutput) { 4677 unsigned NumOutputs = 0; 4678 unsigned NumIfsOutputs = 0; 4679 for (const Action *A : C.getActions()) { 4680 if (A->getType() != types::TY_Nothing && 4681 !(A->getKind() == Action::IfsMergeJobClass || 4682 (A->getType() == clang::driver::types::TY_IFS_CPP && 4683 A->getKind() == clang::driver::Action::CompileJobClass && 4684 0 == NumIfsOutputs++) || 4685 (A->getKind() == Action::BindArchClass && A->getInputs().size() && 4686 A->getInputs().front()->getKind() == Action::IfsMergeJobClass))) 4687 ++NumOutputs; 4688 else if (A->getKind() == Action::OffloadClass && 4689 A->getType() == types::TY_Nothing && 4690 !C.getArgs().hasArg(options::OPT_fsyntax_only)) 4691 NumOutputs += A->size(); 4692 } 4693 4694 if (NumOutputs > 1) { 4695 Diag(clang::diag::err_drv_output_argument_with_multiple_files); 4696 FinalOutput = nullptr; 4697 } 4698 } 4699 4700 const llvm::Triple &RawTriple = C.getDefaultToolChain().getTriple(); 4701 if (RawTriple.isOSAIX()) { 4702 if (Arg *A = C.getArgs().getLastArg(options::OPT_G)) 4703 Diag(diag::err_drv_unsupported_opt_for_target) 4704 << A->getSpelling() << RawTriple.str(); 4705 if (LTOMode == LTOK_Thin) 4706 Diag(diag::err_drv_clang_unsupported) << "thinLTO on AIX"; 4707 } 4708 4709 // Collect the list of architectures. 4710 llvm::StringSet<> ArchNames; 4711 if (RawTriple.isOSBinFormatMachO()) 4712 for (const Arg *A : C.getArgs()) 4713 if (A->getOption().matches(options::OPT_arch)) 4714 ArchNames.insert(A->getValue()); 4715 4716 // Set of (Action, canonical ToolChain triple) pairs we've built jobs for. 4717 std::map<std::pair<const Action *, std::string>, InputInfoList> CachedResults; 4718 for (Action *A : C.getActions()) { 4719 // If we are linking an image for multiple archs then the linker wants 4720 // -arch_multiple and -final_output <final image name>. Unfortunately, this 4721 // doesn't fit in cleanly because we have to pass this information down. 4722 // 4723 // FIXME: This is a hack; find a cleaner way to integrate this into the 4724 // process. 4725 const char *LinkingOutput = nullptr; 4726 if (isa<LipoJobAction>(A)) { 4727 if (FinalOutput) 4728 LinkingOutput = FinalOutput->getValue(); 4729 else 4730 LinkingOutput = getDefaultImageName(); 4731 } 4732 4733 BuildJobsForAction(C, A, &C.getDefaultToolChain(), 4734 /*BoundArch*/ StringRef(), 4735 /*AtTopLevel*/ true, 4736 /*MultipleArchs*/ ArchNames.size() > 1, 4737 /*LinkingOutput*/ LinkingOutput, CachedResults, 4738 /*TargetDeviceOffloadKind*/ Action::OFK_None); 4739 } 4740 4741 // If we have more than one job, then disable integrated-cc1 for now. Do this 4742 // also when we need to report process execution statistics. 4743 if (C.getJobs().size() > 1 || CCPrintProcessStats) 4744 for (auto &J : C.getJobs()) 4745 J.InProcess = false; 4746 4747 if (CCPrintProcessStats) { 4748 C.setPostCallback([=](const Command &Cmd, int Res) { 4749 std::optional<llvm::sys::ProcessStatistics> ProcStat = 4750 Cmd.getProcessStatistics(); 4751 if (!ProcStat) 4752 return; 4753 4754 const char *LinkingOutput = nullptr; 4755 if (FinalOutput) 4756 LinkingOutput = FinalOutput->getValue(); 4757 else if (!Cmd.getOutputFilenames().empty()) 4758 LinkingOutput = Cmd.getOutputFilenames().front().c_str(); 4759 else 4760 LinkingOutput = getDefaultImageName(); 4761 4762 if (CCPrintStatReportFilename.empty()) { 4763 using namespace llvm; 4764 // Human readable output. 4765 outs() << sys::path::filename(Cmd.getExecutable()) << ": " 4766 << "output=" << LinkingOutput; 4767 outs() << ", total=" 4768 << format("%.3f", ProcStat->TotalTime.count() / 1000.) << " ms" 4769 << ", user=" 4770 << format("%.3f", ProcStat->UserTime.count() / 1000.) << " ms" 4771 << ", mem=" << ProcStat->PeakMemory << " Kb\n"; 4772 } else { 4773 // CSV format. 4774 std::string Buffer; 4775 llvm::raw_string_ostream Out(Buffer); 4776 llvm::sys::printArg(Out, llvm::sys::path::filename(Cmd.getExecutable()), 4777 /*Quote*/ true); 4778 Out << ','; 4779 llvm::sys::printArg(Out, LinkingOutput, true); 4780 Out << ',' << ProcStat->TotalTime.count() << ',' 4781 << ProcStat->UserTime.count() << ',' << ProcStat->PeakMemory 4782 << '\n'; 4783 Out.flush(); 4784 std::error_code EC; 4785 llvm::raw_fd_ostream OS(CCPrintStatReportFilename, EC, 4786 llvm::sys::fs::OF_Append | 4787 llvm::sys::fs::OF_Text); 4788 if (EC) 4789 return; 4790 auto L = OS.lock(); 4791 if (!L) { 4792 llvm::errs() << "ERROR: Cannot lock file " 4793 << CCPrintStatReportFilename << ": " 4794 << toString(L.takeError()) << "\n"; 4795 return; 4796 } 4797 OS << Buffer; 4798 OS.flush(); 4799 } 4800 }); 4801 } 4802 4803 // If the user passed -Qunused-arguments or there were errors, don't warn 4804 // about any unused arguments. 4805 if (Diags.hasErrorOccurred() || 4806 C.getArgs().hasArg(options::OPT_Qunused_arguments)) 4807 return; 4808 4809 // Claim -fdriver-only here. 4810 (void)C.getArgs().hasArg(options::OPT_fdriver_only); 4811 // Claim -### here. 4812 (void)C.getArgs().hasArg(options::OPT__HASH_HASH_HASH); 4813 4814 // Claim --driver-mode, --rsp-quoting, it was handled earlier. 4815 (void)C.getArgs().hasArg(options::OPT_driver_mode); 4816 (void)C.getArgs().hasArg(options::OPT_rsp_quoting); 4817 4818 for (Arg *A : C.getArgs()) { 4819 // FIXME: It would be nice to be able to send the argument to the 4820 // DiagnosticsEngine, so that extra values, position, and so on could be 4821 // printed. 4822 if (!A->isClaimed()) { 4823 if (A->getOption().hasFlag(options::NoArgumentUnused)) 4824 continue; 4825 4826 // Suppress the warning automatically if this is just a flag, and it is an 4827 // instance of an argument we already claimed. 4828 const Option &Opt = A->getOption(); 4829 if (Opt.getKind() == Option::FlagClass) { 4830 bool DuplicateClaimed = false; 4831 4832 for (const Arg *AA : C.getArgs().filtered(&Opt)) { 4833 if (AA->isClaimed()) { 4834 DuplicateClaimed = true; 4835 break; 4836 } 4837 } 4838 4839 if (DuplicateClaimed) 4840 continue; 4841 } 4842 4843 // In clang-cl, don't mention unknown arguments here since they have 4844 // already been warned about. 4845 if (!IsCLMode() || !A->getOption().matches(options::OPT_UNKNOWN)) 4846 Diag(clang::diag::warn_drv_unused_argument) 4847 << A->getAsString(C.getArgs()); 4848 } 4849 } 4850 } 4851 4852 namespace { 4853 /// Utility class to control the collapse of dependent actions and select the 4854 /// tools accordingly. 4855 class ToolSelector final { 4856 /// The tool chain this selector refers to. 4857 const ToolChain &TC; 4858 4859 /// The compilation this selector refers to. 4860 const Compilation &C; 4861 4862 /// The base action this selector refers to. 4863 const JobAction *BaseAction; 4864 4865 /// Set to true if the current toolchain refers to host actions. 4866 bool IsHostSelector; 4867 4868 /// Set to true if save-temps and embed-bitcode functionalities are active. 4869 bool SaveTemps; 4870 bool EmbedBitcode; 4871 4872 /// Get previous dependent action or null if that does not exist. If 4873 /// \a CanBeCollapsed is false, that action must be legal to collapse or 4874 /// null will be returned. 4875 const JobAction *getPrevDependentAction(const ActionList &Inputs, 4876 ActionList &SavedOffloadAction, 4877 bool CanBeCollapsed = true) { 4878 // An option can be collapsed only if it has a single input. 4879 if (Inputs.size() != 1) 4880 return nullptr; 4881 4882 Action *CurAction = *Inputs.begin(); 4883 if (CanBeCollapsed && 4884 !CurAction->isCollapsingWithNextDependentActionLegal()) 4885 return nullptr; 4886 4887 // If the input action is an offload action. Look through it and save any 4888 // offload action that can be dropped in the event of a collapse. 4889 if (auto *OA = dyn_cast<OffloadAction>(CurAction)) { 4890 // If the dependent action is a device action, we will attempt to collapse 4891 // only with other device actions. Otherwise, we would do the same but 4892 // with host actions only. 4893 if (!IsHostSelector) { 4894 if (OA->hasSingleDeviceDependence(/*DoNotConsiderHostActions=*/true)) { 4895 CurAction = 4896 OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true); 4897 if (CanBeCollapsed && 4898 !CurAction->isCollapsingWithNextDependentActionLegal()) 4899 return nullptr; 4900 SavedOffloadAction.push_back(OA); 4901 return dyn_cast<JobAction>(CurAction); 4902 } 4903 } else if (OA->hasHostDependence()) { 4904 CurAction = OA->getHostDependence(); 4905 if (CanBeCollapsed && 4906 !CurAction->isCollapsingWithNextDependentActionLegal()) 4907 return nullptr; 4908 SavedOffloadAction.push_back(OA); 4909 return dyn_cast<JobAction>(CurAction); 4910 } 4911 return nullptr; 4912 } 4913 4914 return dyn_cast<JobAction>(CurAction); 4915 } 4916 4917 /// Return true if an assemble action can be collapsed. 4918 bool canCollapseAssembleAction() const { 4919 return TC.useIntegratedAs() && !SaveTemps && 4920 !C.getArgs().hasArg(options::OPT_via_file_asm) && 4921 !C.getArgs().hasArg(options::OPT__SLASH_FA) && 4922 !C.getArgs().hasArg(options::OPT__SLASH_Fa); 4923 } 4924 4925 /// Return true if a preprocessor action can be collapsed. 4926 bool canCollapsePreprocessorAction() const { 4927 return !C.getArgs().hasArg(options::OPT_no_integrated_cpp) && 4928 !C.getArgs().hasArg(options::OPT_traditional_cpp) && !SaveTemps && 4929 !C.getArgs().hasArg(options::OPT_rewrite_objc); 4930 } 4931 4932 /// Struct that relates an action with the offload actions that would be 4933 /// collapsed with it. 4934 struct JobActionInfo final { 4935 /// The action this info refers to. 4936 const JobAction *JA = nullptr; 4937 /// The offload actions we need to take care off if this action is 4938 /// collapsed. 4939 ActionList SavedOffloadAction; 4940 }; 4941 4942 /// Append collapsed offload actions from the give nnumber of elements in the 4943 /// action info array. 4944 static void AppendCollapsedOffloadAction(ActionList &CollapsedOffloadAction, 4945 ArrayRef<JobActionInfo> &ActionInfo, 4946 unsigned ElementNum) { 4947 assert(ElementNum <= ActionInfo.size() && "Invalid number of elements."); 4948 for (unsigned I = 0; I < ElementNum; ++I) 4949 CollapsedOffloadAction.append(ActionInfo[I].SavedOffloadAction.begin(), 4950 ActionInfo[I].SavedOffloadAction.end()); 4951 } 4952 4953 /// Functions that attempt to perform the combining. They detect if that is 4954 /// legal, and if so they update the inputs \a Inputs and the offload action 4955 /// that were collapsed in \a CollapsedOffloadAction. A tool that deals with 4956 /// the combined action is returned. If the combining is not legal or if the 4957 /// tool does not exist, null is returned. 4958 /// Currently three kinds of collapsing are supported: 4959 /// - Assemble + Backend + Compile; 4960 /// - Assemble + Backend ; 4961 /// - Backend + Compile. 4962 const Tool * 4963 combineAssembleBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4964 ActionList &Inputs, 4965 ActionList &CollapsedOffloadAction) { 4966 if (ActionInfo.size() < 3 || !canCollapseAssembleAction()) 4967 return nullptr; 4968 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4969 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4970 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[2].JA); 4971 if (!AJ || !BJ || !CJ) 4972 return nullptr; 4973 4974 // Get compiler tool. 4975 const Tool *T = TC.SelectTool(*CJ); 4976 if (!T) 4977 return nullptr; 4978 4979 // Can't collapse if we don't have codegen support unless we are 4980 // emitting LLVM IR. 4981 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 4982 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 4983 return nullptr; 4984 4985 // When using -fembed-bitcode, it is required to have the same tool (clang) 4986 // for both CompilerJA and BackendJA. Otherwise, combine two stages. 4987 if (EmbedBitcode) { 4988 const Tool *BT = TC.SelectTool(*BJ); 4989 if (BT == T) 4990 return nullptr; 4991 } 4992 4993 if (!T->hasIntegratedAssembler()) 4994 return nullptr; 4995 4996 Inputs = CJ->getInputs(); 4997 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4998 /*NumElements=*/3); 4999 return T; 5000 } 5001 const Tool *combineAssembleBackend(ArrayRef<JobActionInfo> ActionInfo, 5002 ActionList &Inputs, 5003 ActionList &CollapsedOffloadAction) { 5004 if (ActionInfo.size() < 2 || !canCollapseAssembleAction()) 5005 return nullptr; 5006 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 5007 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 5008 if (!AJ || !BJ) 5009 return nullptr; 5010 5011 // Get backend tool. 5012 const Tool *T = TC.SelectTool(*BJ); 5013 if (!T) 5014 return nullptr; 5015 5016 if (!T->hasIntegratedAssembler()) 5017 return nullptr; 5018 5019 Inputs = BJ->getInputs(); 5020 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5021 /*NumElements=*/2); 5022 return T; 5023 } 5024 const Tool *combineBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 5025 ActionList &Inputs, 5026 ActionList &CollapsedOffloadAction) { 5027 if (ActionInfo.size() < 2) 5028 return nullptr; 5029 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[0].JA); 5030 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[1].JA); 5031 if (!BJ || !CJ) 5032 return nullptr; 5033 5034 // Check if the initial input (to the compile job or its predessor if one 5035 // exists) is LLVM bitcode. In that case, no preprocessor step is required 5036 // and we can still collapse the compile and backend jobs when we have 5037 // -save-temps. I.e. there is no need for a separate compile job just to 5038 // emit unoptimized bitcode. 5039 bool InputIsBitcode = true; 5040 for (size_t i = 1; i < ActionInfo.size(); i++) 5041 if (ActionInfo[i].JA->getType() != types::TY_LLVM_BC && 5042 ActionInfo[i].JA->getType() != types::TY_LTO_BC) { 5043 InputIsBitcode = false; 5044 break; 5045 } 5046 if (!InputIsBitcode && !canCollapsePreprocessorAction()) 5047 return nullptr; 5048 5049 // Get compiler tool. 5050 const Tool *T = TC.SelectTool(*CJ); 5051 if (!T) 5052 return nullptr; 5053 5054 // Can't collapse if we don't have codegen support unless we are 5055 // emitting LLVM IR. 5056 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 5057 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 5058 return nullptr; 5059 5060 if (T->canEmitIR() && ((SaveTemps && !InputIsBitcode) || EmbedBitcode)) 5061 return nullptr; 5062 5063 Inputs = CJ->getInputs(); 5064 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5065 /*NumElements=*/2); 5066 return T; 5067 } 5068 5069 /// Updates the inputs if the obtained tool supports combining with 5070 /// preprocessor action, and the current input is indeed a preprocessor 5071 /// action. If combining results in the collapse of offloading actions, those 5072 /// are appended to \a CollapsedOffloadAction. 5073 void combineWithPreprocessor(const Tool *T, ActionList &Inputs, 5074 ActionList &CollapsedOffloadAction) { 5075 if (!T || !canCollapsePreprocessorAction() || !T->hasIntegratedCPP()) 5076 return; 5077 5078 // Attempt to get a preprocessor action dependence. 5079 ActionList PreprocessJobOffloadActions; 5080 ActionList NewInputs; 5081 for (Action *A : Inputs) { 5082 auto *PJ = getPrevDependentAction({A}, PreprocessJobOffloadActions); 5083 if (!PJ || !isa<PreprocessJobAction>(PJ)) { 5084 NewInputs.push_back(A); 5085 continue; 5086 } 5087 5088 // This is legal to combine. Append any offload action we found and add the 5089 // current input to preprocessor inputs. 5090 CollapsedOffloadAction.append(PreprocessJobOffloadActions.begin(), 5091 PreprocessJobOffloadActions.end()); 5092 NewInputs.append(PJ->input_begin(), PJ->input_end()); 5093 } 5094 Inputs = NewInputs; 5095 } 5096 5097 public: 5098 ToolSelector(const JobAction *BaseAction, const ToolChain &TC, 5099 const Compilation &C, bool SaveTemps, bool EmbedBitcode) 5100 : TC(TC), C(C), BaseAction(BaseAction), SaveTemps(SaveTemps), 5101 EmbedBitcode(EmbedBitcode) { 5102 assert(BaseAction && "Invalid base action."); 5103 IsHostSelector = BaseAction->getOffloadingDeviceKind() == Action::OFK_None; 5104 } 5105 5106 /// Check if a chain of actions can be combined and return the tool that can 5107 /// handle the combination of actions. The pointer to the current inputs \a 5108 /// Inputs and the list of offload actions \a CollapsedOffloadActions 5109 /// connected to collapsed actions are updated accordingly. The latter enables 5110 /// the caller of the selector to process them afterwards instead of just 5111 /// dropping them. If no suitable tool is found, null will be returned. 5112 const Tool *getTool(ActionList &Inputs, 5113 ActionList &CollapsedOffloadAction) { 5114 // 5115 // Get the largest chain of actions that we could combine. 5116 // 5117 5118 SmallVector<JobActionInfo, 5> ActionChain(1); 5119 ActionChain.back().JA = BaseAction; 5120 while (ActionChain.back().JA) { 5121 const Action *CurAction = ActionChain.back().JA; 5122 5123 // Grow the chain by one element. 5124 ActionChain.resize(ActionChain.size() + 1); 5125 JobActionInfo &AI = ActionChain.back(); 5126 5127 // Attempt to fill it with the 5128 AI.JA = 5129 getPrevDependentAction(CurAction->getInputs(), AI.SavedOffloadAction); 5130 } 5131 5132 // Pop the last action info as it could not be filled. 5133 ActionChain.pop_back(); 5134 5135 // 5136 // Attempt to combine actions. If all combining attempts failed, just return 5137 // the tool of the provided action. At the end we attempt to combine the 5138 // action with any preprocessor action it may depend on. 5139 // 5140 5141 const Tool *T = combineAssembleBackendCompile(ActionChain, Inputs, 5142 CollapsedOffloadAction); 5143 if (!T) 5144 T = combineAssembleBackend(ActionChain, Inputs, CollapsedOffloadAction); 5145 if (!T) 5146 T = combineBackendCompile(ActionChain, Inputs, CollapsedOffloadAction); 5147 if (!T) { 5148 Inputs = BaseAction->getInputs(); 5149 T = TC.SelectTool(*BaseAction); 5150 } 5151 5152 combineWithPreprocessor(T, Inputs, CollapsedOffloadAction); 5153 return T; 5154 } 5155 }; 5156 } 5157 5158 /// Return a string that uniquely identifies the result of a job. The bound arch 5159 /// is not necessarily represented in the toolchain's triple -- for example, 5160 /// armv7 and armv7s both map to the same triple -- so we need both in our map. 5161 /// Also, we need to add the offloading device kind, as the same tool chain can 5162 /// be used for host and device for some programming models, e.g. OpenMP. 5163 static std::string GetTriplePlusArchString(const ToolChain *TC, 5164 StringRef BoundArch, 5165 Action::OffloadKind OffloadKind) { 5166 std::string TriplePlusArch = TC->getTriple().normalize(); 5167 if (!BoundArch.empty()) { 5168 TriplePlusArch += "-"; 5169 TriplePlusArch += BoundArch; 5170 } 5171 TriplePlusArch += "-"; 5172 TriplePlusArch += Action::GetOffloadKindName(OffloadKind); 5173 return TriplePlusArch; 5174 } 5175 5176 InputInfoList Driver::BuildJobsForAction( 5177 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5178 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5179 std::map<std::pair<const Action *, std::string>, InputInfoList> 5180 &CachedResults, 5181 Action::OffloadKind TargetDeviceOffloadKind) const { 5182 std::pair<const Action *, std::string> ActionTC = { 5183 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5184 auto CachedResult = CachedResults.find(ActionTC); 5185 if (CachedResult != CachedResults.end()) { 5186 return CachedResult->second; 5187 } 5188 InputInfoList Result = BuildJobsForActionNoCache( 5189 C, A, TC, BoundArch, AtTopLevel, MultipleArchs, LinkingOutput, 5190 CachedResults, TargetDeviceOffloadKind); 5191 CachedResults[ActionTC] = Result; 5192 return Result; 5193 } 5194 5195 InputInfoList Driver::BuildJobsForActionNoCache( 5196 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5197 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5198 std::map<std::pair<const Action *, std::string>, InputInfoList> 5199 &CachedResults, 5200 Action::OffloadKind TargetDeviceOffloadKind) const { 5201 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 5202 5203 InputInfoList OffloadDependencesInputInfo; 5204 bool BuildingForOffloadDevice = TargetDeviceOffloadKind != Action::OFK_None; 5205 if (const OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 5206 // The 'Darwin' toolchain is initialized only when its arguments are 5207 // computed. Get the default arguments for OFK_None to ensure that 5208 // initialization is performed before processing the offload action. 5209 // FIXME: Remove when darwin's toolchain is initialized during construction. 5210 C.getArgsForToolChain(TC, BoundArch, Action::OFK_None); 5211 5212 // The offload action is expected to be used in four different situations. 5213 // 5214 // a) Set a toolchain/architecture/kind for a host action: 5215 // Host Action 1 -> OffloadAction -> Host Action 2 5216 // 5217 // b) Set a toolchain/architecture/kind for a device action; 5218 // Device Action 1 -> OffloadAction -> Device Action 2 5219 // 5220 // c) Specify a device dependence to a host action; 5221 // Device Action 1 _ 5222 // \ 5223 // Host Action 1 ---> OffloadAction -> Host Action 2 5224 // 5225 // d) Specify a host dependence to a device action. 5226 // Host Action 1 _ 5227 // \ 5228 // Device Action 1 ---> OffloadAction -> Device Action 2 5229 // 5230 // For a) and b), we just return the job generated for the dependences. For 5231 // c) and d) we override the current action with the host/device dependence 5232 // if the current toolchain is host/device and set the offload dependences 5233 // info with the jobs obtained from the device/host dependence(s). 5234 5235 // If there is a single device option or has no host action, just generate 5236 // the job for it. 5237 if (OA->hasSingleDeviceDependence() || !OA->hasHostDependence()) { 5238 InputInfoList DevA; 5239 OA->doOnEachDeviceDependence([&](Action *DepA, const ToolChain *DepTC, 5240 const char *DepBoundArch) { 5241 DevA.append(BuildJobsForAction(C, DepA, DepTC, DepBoundArch, AtTopLevel, 5242 /*MultipleArchs*/ !!DepBoundArch, 5243 LinkingOutput, CachedResults, 5244 DepA->getOffloadingDeviceKind())); 5245 }); 5246 return DevA; 5247 } 5248 5249 // If 'Action 2' is host, we generate jobs for the device dependences and 5250 // override the current action with the host dependence. Otherwise, we 5251 // generate the host dependences and override the action with the device 5252 // dependence. The dependences can't therefore be a top-level action. 5253 OA->doOnEachDependence( 5254 /*IsHostDependence=*/BuildingForOffloadDevice, 5255 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5256 OffloadDependencesInputInfo.append(BuildJobsForAction( 5257 C, DepA, DepTC, DepBoundArch, /*AtTopLevel=*/false, 5258 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, CachedResults, 5259 DepA->getOffloadingDeviceKind())); 5260 }); 5261 5262 A = BuildingForOffloadDevice 5263 ? OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true) 5264 : OA->getHostDependence(); 5265 5266 // We may have already built this action as a part of the offloading 5267 // toolchain, return the cached input if so. 5268 std::pair<const Action *, std::string> ActionTC = { 5269 OA->getHostDependence(), 5270 GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5271 if (CachedResults.find(ActionTC) != CachedResults.end()) { 5272 InputInfoList Inputs = CachedResults[ActionTC]; 5273 Inputs.append(OffloadDependencesInputInfo); 5274 return Inputs; 5275 } 5276 } 5277 5278 if (const InputAction *IA = dyn_cast<InputAction>(A)) { 5279 // FIXME: It would be nice to not claim this here; maybe the old scheme of 5280 // just using Args was better? 5281 const Arg &Input = IA->getInputArg(); 5282 Input.claim(); 5283 if (Input.getOption().matches(options::OPT_INPUT)) { 5284 const char *Name = Input.getValue(); 5285 return {InputInfo(A, Name, /* _BaseInput = */ Name)}; 5286 } 5287 return {InputInfo(A, &Input, /* _BaseInput = */ "")}; 5288 } 5289 5290 if (const BindArchAction *BAA = dyn_cast<BindArchAction>(A)) { 5291 const ToolChain *TC; 5292 StringRef ArchName = BAA->getArchName(); 5293 5294 if (!ArchName.empty()) 5295 TC = &getToolChain(C.getArgs(), 5296 computeTargetTriple(*this, TargetTriple, 5297 C.getArgs(), ArchName)); 5298 else 5299 TC = &C.getDefaultToolChain(); 5300 5301 return BuildJobsForAction(C, *BAA->input_begin(), TC, ArchName, AtTopLevel, 5302 MultipleArchs, LinkingOutput, CachedResults, 5303 TargetDeviceOffloadKind); 5304 } 5305 5306 5307 ActionList Inputs = A->getInputs(); 5308 5309 const JobAction *JA = cast<JobAction>(A); 5310 ActionList CollapsedOffloadActions; 5311 5312 ToolSelector TS(JA, *TC, C, isSaveTempsEnabled(), 5313 embedBitcodeInObject() && !isUsingLTO()); 5314 const Tool *T = TS.getTool(Inputs, CollapsedOffloadActions); 5315 5316 if (!T) 5317 return {InputInfo()}; 5318 5319 // If we've collapsed action list that contained OffloadAction we 5320 // need to build jobs for host/device-side inputs it may have held. 5321 for (const auto *OA : CollapsedOffloadActions) 5322 cast<OffloadAction>(OA)->doOnEachDependence( 5323 /*IsHostDependence=*/BuildingForOffloadDevice, 5324 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5325 OffloadDependencesInputInfo.append(BuildJobsForAction( 5326 C, DepA, DepTC, DepBoundArch, /* AtTopLevel */ false, 5327 /*MultipleArchs=*/!!DepBoundArch, LinkingOutput, CachedResults, 5328 DepA->getOffloadingDeviceKind())); 5329 }); 5330 5331 // Only use pipes when there is exactly one input. 5332 InputInfoList InputInfos; 5333 for (const Action *Input : Inputs) { 5334 // Treat dsymutil and verify sub-jobs as being at the top-level too, they 5335 // shouldn't get temporary output names. 5336 // FIXME: Clean this up. 5337 bool SubJobAtTopLevel = 5338 AtTopLevel && (isa<DsymutilJobAction>(A) || isa<VerifyJobAction>(A)); 5339 InputInfos.append(BuildJobsForAction( 5340 C, Input, TC, BoundArch, SubJobAtTopLevel, MultipleArchs, LinkingOutput, 5341 CachedResults, A->getOffloadingDeviceKind())); 5342 } 5343 5344 // Always use the first file input as the base input. 5345 const char *BaseInput = InputInfos[0].getBaseInput(); 5346 for (auto &Info : InputInfos) { 5347 if (Info.isFilename()) { 5348 BaseInput = Info.getBaseInput(); 5349 break; 5350 } 5351 } 5352 5353 // ... except dsymutil actions, which use their actual input as the base 5354 // input. 5355 if (JA->getType() == types::TY_dSYM) 5356 BaseInput = InputInfos[0].getFilename(); 5357 5358 // Append outputs of offload device jobs to the input list 5359 if (!OffloadDependencesInputInfo.empty()) 5360 InputInfos.append(OffloadDependencesInputInfo.begin(), 5361 OffloadDependencesInputInfo.end()); 5362 5363 // Set the effective triple of the toolchain for the duration of this job. 5364 llvm::Triple EffectiveTriple; 5365 const ToolChain &ToolTC = T->getToolChain(); 5366 const ArgList &Args = 5367 C.getArgsForToolChain(TC, BoundArch, A->getOffloadingDeviceKind()); 5368 if (InputInfos.size() != 1) { 5369 EffectiveTriple = llvm::Triple(ToolTC.ComputeEffectiveClangTriple(Args)); 5370 } else { 5371 // Pass along the input type if it can be unambiguously determined. 5372 EffectiveTriple = llvm::Triple( 5373 ToolTC.ComputeEffectiveClangTriple(Args, InputInfos[0].getType())); 5374 } 5375 RegisterEffectiveTriple TripleRAII(ToolTC, EffectiveTriple); 5376 5377 // Determine the place to write output to, if any. 5378 InputInfo Result; 5379 InputInfoList UnbundlingResults; 5380 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(JA)) { 5381 // If we have an unbundling job, we need to create results for all the 5382 // outputs. We also update the results cache so that other actions using 5383 // this unbundling action can get the right results. 5384 for (auto &UI : UA->getDependentActionsInfo()) { 5385 assert(UI.DependentOffloadKind != Action::OFK_None && 5386 "Unbundling with no offloading??"); 5387 5388 // Unbundling actions are never at the top level. When we generate the 5389 // offloading prefix, we also do that for the host file because the 5390 // unbundling action does not change the type of the output which can 5391 // cause a overwrite. 5392 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5393 UI.DependentOffloadKind, 5394 UI.DependentToolChain->getTriple().normalize(), 5395 /*CreatePrefixForHost=*/true); 5396 auto CurI = InputInfo( 5397 UA, 5398 GetNamedOutputPath(C, *UA, BaseInput, UI.DependentBoundArch, 5399 /*AtTopLevel=*/false, 5400 MultipleArchs || 5401 UI.DependentOffloadKind == Action::OFK_HIP, 5402 OffloadingPrefix), 5403 BaseInput); 5404 // Save the unbundling result. 5405 UnbundlingResults.push_back(CurI); 5406 5407 // Get the unique string identifier for this dependence and cache the 5408 // result. 5409 StringRef Arch; 5410 if (TargetDeviceOffloadKind == Action::OFK_HIP) { 5411 if (UI.DependentOffloadKind == Action::OFK_Host) 5412 Arch = StringRef(); 5413 else 5414 Arch = UI.DependentBoundArch; 5415 } else 5416 Arch = BoundArch; 5417 5418 CachedResults[{A, GetTriplePlusArchString(UI.DependentToolChain, Arch, 5419 UI.DependentOffloadKind)}] = { 5420 CurI}; 5421 } 5422 5423 // Now that we have all the results generated, select the one that should be 5424 // returned for the current depending action. 5425 std::pair<const Action *, std::string> ActionTC = { 5426 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5427 assert(CachedResults.find(ActionTC) != CachedResults.end() && 5428 "Result does not exist??"); 5429 Result = CachedResults[ActionTC].front(); 5430 } else if (JA->getType() == types::TY_Nothing) 5431 Result = {InputInfo(A, BaseInput)}; 5432 else { 5433 // We only have to generate a prefix for the host if this is not a top-level 5434 // action. 5435 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5436 A->getOffloadingDeviceKind(), TC->getTriple().normalize(), 5437 /*CreatePrefixForHost=*/isa<OffloadPackagerJobAction>(A) || 5438 !(A->getOffloadingHostActiveKinds() == Action::OFK_None || 5439 AtTopLevel)); 5440 Result = InputInfo(A, GetNamedOutputPath(C, *JA, BaseInput, BoundArch, 5441 AtTopLevel, MultipleArchs, 5442 OffloadingPrefix), 5443 BaseInput); 5444 } 5445 5446 if (CCCPrintBindings && !CCGenDiagnostics) { 5447 llvm::errs() << "# \"" << T->getToolChain().getTripleString() << '"' 5448 << " - \"" << T->getName() << "\", inputs: ["; 5449 for (unsigned i = 0, e = InputInfos.size(); i != e; ++i) { 5450 llvm::errs() << InputInfos[i].getAsString(); 5451 if (i + 1 != e) 5452 llvm::errs() << ", "; 5453 } 5454 if (UnbundlingResults.empty()) 5455 llvm::errs() << "], output: " << Result.getAsString() << "\n"; 5456 else { 5457 llvm::errs() << "], outputs: ["; 5458 for (unsigned i = 0, e = UnbundlingResults.size(); i != e; ++i) { 5459 llvm::errs() << UnbundlingResults[i].getAsString(); 5460 if (i + 1 != e) 5461 llvm::errs() << ", "; 5462 } 5463 llvm::errs() << "] \n"; 5464 } 5465 } else { 5466 if (UnbundlingResults.empty()) 5467 T->ConstructJob( 5468 C, *JA, Result, InputInfos, 5469 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5470 LinkingOutput); 5471 else 5472 T->ConstructJobMultipleOutputs( 5473 C, *JA, UnbundlingResults, InputInfos, 5474 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5475 LinkingOutput); 5476 } 5477 return {Result}; 5478 } 5479 5480 const char *Driver::getDefaultImageName() const { 5481 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 5482 return Target.isOSWindows() ? "a.exe" : "a.out"; 5483 } 5484 5485 /// Create output filename based on ArgValue, which could either be a 5486 /// full filename, filename without extension, or a directory. If ArgValue 5487 /// does not provide a filename, then use BaseName, and use the extension 5488 /// suitable for FileType. 5489 static const char *MakeCLOutputFilename(const ArgList &Args, StringRef ArgValue, 5490 StringRef BaseName, 5491 types::ID FileType) { 5492 SmallString<128> Filename = ArgValue; 5493 5494 if (ArgValue.empty()) { 5495 // If the argument is empty, output to BaseName in the current dir. 5496 Filename = BaseName; 5497 } else if (llvm::sys::path::is_separator(Filename.back())) { 5498 // If the argument is a directory, output to BaseName in that dir. 5499 llvm::sys::path::append(Filename, BaseName); 5500 } 5501 5502 if (!llvm::sys::path::has_extension(ArgValue)) { 5503 // If the argument didn't provide an extension, then set it. 5504 const char *Extension = types::getTypeTempSuffix(FileType, true); 5505 5506 if (FileType == types::TY_Image && 5507 Args.hasArg(options::OPT__SLASH_LD, options::OPT__SLASH_LDd)) { 5508 // The output file is a dll. 5509 Extension = "dll"; 5510 } 5511 5512 llvm::sys::path::replace_extension(Filename, Extension); 5513 } 5514 5515 return Args.MakeArgString(Filename.c_str()); 5516 } 5517 5518 static bool HasPreprocessOutput(const Action &JA) { 5519 if (isa<PreprocessJobAction>(JA)) 5520 return true; 5521 if (isa<OffloadAction>(JA) && isa<PreprocessJobAction>(JA.getInputs()[0])) 5522 return true; 5523 if (isa<OffloadBundlingJobAction>(JA) && 5524 HasPreprocessOutput(*(JA.getInputs()[0]))) 5525 return true; 5526 return false; 5527 } 5528 5529 const char *Driver::CreateTempFile(Compilation &C, StringRef Prefix, 5530 StringRef Suffix, bool MultipleArchs, 5531 StringRef BoundArch) const { 5532 SmallString<128> TmpName; 5533 Arg *A = C.getArgs().getLastArg(options::OPT_fcrash_diagnostics_dir); 5534 std::optional<std::string> CrashDirectory = 5535 CCGenDiagnostics && A 5536 ? std::string(A->getValue()) 5537 : llvm::sys::Process::GetEnv("CLANG_CRASH_DIAGNOSTICS_DIR"); 5538 if (CrashDirectory) { 5539 if (!getVFS().exists(*CrashDirectory)) 5540 llvm::sys::fs::create_directories(*CrashDirectory); 5541 SmallString<128> Path(*CrashDirectory); 5542 llvm::sys::path::append(Path, Prefix); 5543 const char *Middle = !Suffix.empty() ? "-%%%%%%." : "-%%%%%%"; 5544 if (std::error_code EC = 5545 llvm::sys::fs::createUniqueFile(Path + Middle + Suffix, TmpName)) { 5546 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5547 return ""; 5548 } 5549 } else { 5550 if (MultipleArchs && !BoundArch.empty()) { 5551 TmpName = GetTemporaryDirectory(Prefix); 5552 llvm::sys::path::append(TmpName, 5553 Twine(Prefix) + "-" + BoundArch + "." + Suffix); 5554 } else { 5555 TmpName = GetTemporaryPath(Prefix, Suffix); 5556 } 5557 } 5558 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5559 } 5560 5561 // Calculate the output path of the module file when compiling a module unit 5562 // with the `-fmodule-output` option or `-fmodule-output=` option specified. 5563 // The behavior is: 5564 // - If `-fmodule-output=` is specfied, then the module file is 5565 // writing to the value. 5566 // - Otherwise if the output object file of the module unit is specified, the 5567 // output path 5568 // of the module file should be the same with the output object file except 5569 // the corresponding suffix. This requires both `-o` and `-c` are specified. 5570 // - Otherwise, the output path of the module file will be the same with the 5571 // input with the corresponding suffix. 5572 static const char *GetModuleOutputPath(Compilation &C, const JobAction &JA, 5573 const char *BaseInput) { 5574 assert(isa<PrecompileJobAction>(JA) && JA.getType() == types::TY_ModuleFile && 5575 (C.getArgs().hasArg(options::OPT_fmodule_output) || 5576 C.getArgs().hasArg(options::OPT_fmodule_output_EQ))); 5577 5578 if (Arg *ModuleOutputEQ = 5579 C.getArgs().getLastArg(options::OPT_fmodule_output_EQ)) 5580 return C.addResultFile(ModuleOutputEQ->getValue(), &JA); 5581 5582 SmallString<64> OutputPath; 5583 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 5584 if (FinalOutput && C.getArgs().hasArg(options::OPT_c)) 5585 OutputPath = FinalOutput->getValue(); 5586 else 5587 OutputPath = BaseInput; 5588 5589 const char *Extension = types::getTypeTempSuffix(JA.getType()); 5590 llvm::sys::path::replace_extension(OutputPath, Extension); 5591 return C.addResultFile(C.getArgs().MakeArgString(OutputPath.c_str()), &JA); 5592 } 5593 5594 const char *Driver::GetNamedOutputPath(Compilation &C, const JobAction &JA, 5595 const char *BaseInput, 5596 StringRef OrigBoundArch, bool AtTopLevel, 5597 bool MultipleArchs, 5598 StringRef OffloadingPrefix) const { 5599 std::string BoundArch = OrigBoundArch.str(); 5600 if (is_style_windows(llvm::sys::path::Style::native)) { 5601 // BoundArch may contains ':', which is invalid in file names on Windows, 5602 // therefore replace it with '%'. 5603 std::replace(BoundArch.begin(), BoundArch.end(), ':', '@'); 5604 } 5605 5606 llvm::PrettyStackTraceString CrashInfo("Computing output path"); 5607 // Output to a user requested destination? 5608 if (AtTopLevel && !isa<DsymutilJobAction>(JA) && !isa<VerifyJobAction>(JA)) { 5609 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5610 return C.addResultFile(FinalOutput->getValue(), &JA); 5611 } 5612 5613 // For /P, preprocess to file named after BaseInput. 5614 if (C.getArgs().hasArg(options::OPT__SLASH_P)) { 5615 assert(AtTopLevel && isa<PreprocessJobAction>(JA)); 5616 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5617 StringRef NameArg; 5618 if (Arg *A = C.getArgs().getLastArg(options::OPT__SLASH_Fi)) 5619 NameArg = A->getValue(); 5620 return C.addResultFile( 5621 MakeCLOutputFilename(C.getArgs(), NameArg, BaseName, types::TY_PP_C), 5622 &JA); 5623 } 5624 5625 // Default to writing to stdout? 5626 if (AtTopLevel && !CCGenDiagnostics && HasPreprocessOutput(JA)) { 5627 return "-"; 5628 } 5629 5630 if (JA.getType() == types::TY_ModuleFile && 5631 C.getArgs().getLastArg(options::OPT_module_file_info)) { 5632 return "-"; 5633 } 5634 5635 if (IsDXCMode() && !C.getArgs().hasArg(options::OPT_o)) 5636 return "-"; 5637 5638 // Is this the assembly listing for /FA? 5639 if (JA.getType() == types::TY_PP_Asm && 5640 (C.getArgs().hasArg(options::OPT__SLASH_FA) || 5641 C.getArgs().hasArg(options::OPT__SLASH_Fa))) { 5642 // Use /Fa and the input filename to determine the asm file name. 5643 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5644 StringRef FaValue = C.getArgs().getLastArgValue(options::OPT__SLASH_Fa); 5645 return C.addResultFile( 5646 MakeCLOutputFilename(C.getArgs(), FaValue, BaseName, JA.getType()), 5647 &JA); 5648 } 5649 5650 bool SpecifiedModuleOutput = 5651 C.getArgs().hasArg(options::OPT_fmodule_output) || 5652 C.getArgs().hasArg(options::OPT_fmodule_output_EQ); 5653 if (MultipleArchs && SpecifiedModuleOutput) 5654 Diag(clang::diag::err_drv_module_output_with_multiple_arch); 5655 5656 // If we're emitting a module output with the specified option 5657 // `-fmodule-output`. 5658 if (!AtTopLevel && isa<PrecompileJobAction>(JA) && 5659 JA.getType() == types::TY_ModuleFile && SpecifiedModuleOutput) 5660 return GetModuleOutputPath(C, JA, BaseInput); 5661 5662 // Output to a temporary file? 5663 if ((!AtTopLevel && !isSaveTempsEnabled() && 5664 !C.getArgs().hasArg(options::OPT__SLASH_Fo)) || 5665 CCGenDiagnostics) { 5666 StringRef Name = llvm::sys::path::filename(BaseInput); 5667 std::pair<StringRef, StringRef> Split = Name.split('.'); 5668 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5669 return CreateTempFile(C, Split.first, Suffix, MultipleArchs, BoundArch); 5670 } 5671 5672 SmallString<128> BasePath(BaseInput); 5673 SmallString<128> ExternalPath(""); 5674 StringRef BaseName; 5675 5676 // Dsymutil actions should use the full path. 5677 if (isa<DsymutilJobAction>(JA) && C.getArgs().hasArg(options::OPT_dsym_dir)) { 5678 ExternalPath += C.getArgs().getLastArg(options::OPT_dsym_dir)->getValue(); 5679 // We use posix style here because the tests (specifically 5680 // darwin-dsymutil.c) demonstrate that posix style paths are acceptable 5681 // even on Windows and if we don't then the similar test covering this 5682 // fails. 5683 llvm::sys::path::append(ExternalPath, llvm::sys::path::Style::posix, 5684 llvm::sys::path::filename(BasePath)); 5685 BaseName = ExternalPath; 5686 } else if (isa<DsymutilJobAction>(JA) || isa<VerifyJobAction>(JA)) 5687 BaseName = BasePath; 5688 else 5689 BaseName = llvm::sys::path::filename(BasePath); 5690 5691 // Determine what the derived output name should be. 5692 const char *NamedOutput; 5693 5694 if ((JA.getType() == types::TY_Object || JA.getType() == types::TY_LTO_BC) && 5695 C.getArgs().hasArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o)) { 5696 // The /Fo or /o flag decides the object filename. 5697 StringRef Val = 5698 C.getArgs() 5699 .getLastArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o) 5700 ->getValue(); 5701 NamedOutput = 5702 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5703 } else if (JA.getType() == types::TY_Image && 5704 C.getArgs().hasArg(options::OPT__SLASH_Fe, 5705 options::OPT__SLASH_o)) { 5706 // The /Fe or /o flag names the linked file. 5707 StringRef Val = 5708 C.getArgs() 5709 .getLastArg(options::OPT__SLASH_Fe, options::OPT__SLASH_o) 5710 ->getValue(); 5711 NamedOutput = 5712 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Image); 5713 } else if (JA.getType() == types::TY_Image) { 5714 if (IsCLMode()) { 5715 // clang-cl uses BaseName for the executable name. 5716 NamedOutput = 5717 MakeCLOutputFilename(C.getArgs(), "", BaseName, types::TY_Image); 5718 } else { 5719 SmallString<128> Output(getDefaultImageName()); 5720 // HIP image for device compilation with -fno-gpu-rdc is per compilation 5721 // unit. 5722 bool IsHIPNoRDC = JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5723 !C.getArgs().hasFlag(options::OPT_fgpu_rdc, 5724 options::OPT_fno_gpu_rdc, false); 5725 bool UseOutExtension = IsHIPNoRDC || isa<OffloadPackagerJobAction>(JA); 5726 if (UseOutExtension) { 5727 Output = BaseName; 5728 llvm::sys::path::replace_extension(Output, ""); 5729 } 5730 Output += OffloadingPrefix; 5731 if (MultipleArchs && !BoundArch.empty()) { 5732 Output += "-"; 5733 Output.append(BoundArch); 5734 } 5735 if (UseOutExtension) 5736 Output += ".out"; 5737 NamedOutput = C.getArgs().MakeArgString(Output.c_str()); 5738 } 5739 } else if (JA.getType() == types::TY_PCH && IsCLMode()) { 5740 NamedOutput = C.getArgs().MakeArgString(GetClPchPath(C, BaseName)); 5741 } else if ((JA.getType() == types::TY_Plist || JA.getType() == types::TY_AST) && 5742 C.getArgs().hasArg(options::OPT__SLASH_o)) { 5743 StringRef Val = 5744 C.getArgs() 5745 .getLastArg(options::OPT__SLASH_o) 5746 ->getValue(); 5747 NamedOutput = 5748 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5749 } else { 5750 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5751 assert(Suffix && "All types used for output should have a suffix."); 5752 5753 std::string::size_type End = std::string::npos; 5754 if (!types::appendSuffixForType(JA.getType())) 5755 End = BaseName.rfind('.'); 5756 SmallString<128> Suffixed(BaseName.substr(0, End)); 5757 Suffixed += OffloadingPrefix; 5758 if (MultipleArchs && !BoundArch.empty()) { 5759 Suffixed += "-"; 5760 Suffixed.append(BoundArch); 5761 } 5762 // When using both -save-temps and -emit-llvm, use a ".tmp.bc" suffix for 5763 // the unoptimized bitcode so that it does not get overwritten by the ".bc" 5764 // optimized bitcode output. 5765 auto IsAMDRDCInCompilePhase = [](const JobAction &JA, 5766 const llvm::opt::DerivedArgList &Args) { 5767 // The relocatable compilation in HIP and OpenMP implies -emit-llvm. 5768 // Similarly, use a ".tmp.bc" suffix for the unoptimized bitcode 5769 // (generated in the compile phase.) 5770 const ToolChain *TC = JA.getOffloadingToolChain(); 5771 return isa<CompileJobAction>(JA) && 5772 ((JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5773 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 5774 false)) || 5775 (JA.getOffloadingDeviceKind() == Action::OFK_OpenMP && TC && 5776 TC->getTriple().isAMDGPU())); 5777 }; 5778 if (!AtTopLevel && JA.getType() == types::TY_LLVM_BC && 5779 (C.getArgs().hasArg(options::OPT_emit_llvm) || 5780 IsAMDRDCInCompilePhase(JA, C.getArgs()))) 5781 Suffixed += ".tmp"; 5782 Suffixed += '.'; 5783 Suffixed += Suffix; 5784 NamedOutput = C.getArgs().MakeArgString(Suffixed.c_str()); 5785 } 5786 5787 // Prepend object file path if -save-temps=obj 5788 if (!AtTopLevel && isSaveTempsObj() && C.getArgs().hasArg(options::OPT_o) && 5789 JA.getType() != types::TY_PCH) { 5790 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 5791 SmallString<128> TempPath(FinalOutput->getValue()); 5792 llvm::sys::path::remove_filename(TempPath); 5793 StringRef OutputFileName = llvm::sys::path::filename(NamedOutput); 5794 llvm::sys::path::append(TempPath, OutputFileName); 5795 NamedOutput = C.getArgs().MakeArgString(TempPath.c_str()); 5796 } 5797 5798 // If we're saving temps and the temp file conflicts with the input file, 5799 // then avoid overwriting input file. 5800 if (!AtTopLevel && isSaveTempsEnabled() && NamedOutput == BaseName) { 5801 bool SameFile = false; 5802 SmallString<256> Result; 5803 llvm::sys::fs::current_path(Result); 5804 llvm::sys::path::append(Result, BaseName); 5805 llvm::sys::fs::equivalent(BaseInput, Result.c_str(), SameFile); 5806 // Must share the same path to conflict. 5807 if (SameFile) { 5808 StringRef Name = llvm::sys::path::filename(BaseInput); 5809 std::pair<StringRef, StringRef> Split = Name.split('.'); 5810 std::string TmpName = GetTemporaryPath( 5811 Split.first, types::getTypeTempSuffix(JA.getType(), IsCLMode())); 5812 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5813 } 5814 } 5815 5816 // As an annoying special case, PCH generation doesn't strip the pathname. 5817 if (JA.getType() == types::TY_PCH && !IsCLMode()) { 5818 llvm::sys::path::remove_filename(BasePath); 5819 if (BasePath.empty()) 5820 BasePath = NamedOutput; 5821 else 5822 llvm::sys::path::append(BasePath, NamedOutput); 5823 return C.addResultFile(C.getArgs().MakeArgString(BasePath.c_str()), &JA); 5824 } 5825 5826 return C.addResultFile(NamedOutput, &JA); 5827 } 5828 5829 std::string Driver::GetFilePath(StringRef Name, const ToolChain &TC) const { 5830 // Search for Name in a list of paths. 5831 auto SearchPaths = [&](const llvm::SmallVectorImpl<std::string> &P) 5832 -> std::optional<std::string> { 5833 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5834 // attempting to use this prefix when looking for file paths. 5835 for (const auto &Dir : P) { 5836 if (Dir.empty()) 5837 continue; 5838 SmallString<128> P(Dir[0] == '=' ? SysRoot + Dir.substr(1) : Dir); 5839 llvm::sys::path::append(P, Name); 5840 if (llvm::sys::fs::exists(Twine(P))) 5841 return std::string(P); 5842 } 5843 return std::nullopt; 5844 }; 5845 5846 if (auto P = SearchPaths(PrefixDirs)) 5847 return *P; 5848 5849 SmallString<128> R(ResourceDir); 5850 llvm::sys::path::append(R, Name); 5851 if (llvm::sys::fs::exists(Twine(R))) 5852 return std::string(R.str()); 5853 5854 SmallString<128> P(TC.getCompilerRTPath()); 5855 llvm::sys::path::append(P, Name); 5856 if (llvm::sys::fs::exists(Twine(P))) 5857 return std::string(P.str()); 5858 5859 SmallString<128> D(Dir); 5860 llvm::sys::path::append(D, "..", Name); 5861 if (llvm::sys::fs::exists(Twine(D))) 5862 return std::string(D.str()); 5863 5864 if (auto P = SearchPaths(TC.getLibraryPaths())) 5865 return *P; 5866 5867 if (auto P = SearchPaths(TC.getFilePaths())) 5868 return *P; 5869 5870 return std::string(Name); 5871 } 5872 5873 void Driver::generatePrefixedToolNames( 5874 StringRef Tool, const ToolChain &TC, 5875 SmallVectorImpl<std::string> &Names) const { 5876 // FIXME: Needs a better variable than TargetTriple 5877 Names.emplace_back((TargetTriple + "-" + Tool).str()); 5878 Names.emplace_back(Tool); 5879 } 5880 5881 static bool ScanDirForExecutable(SmallString<128> &Dir, StringRef Name) { 5882 llvm::sys::path::append(Dir, Name); 5883 if (llvm::sys::fs::can_execute(Twine(Dir))) 5884 return true; 5885 llvm::sys::path::remove_filename(Dir); 5886 return false; 5887 } 5888 5889 std::string Driver::GetProgramPath(StringRef Name, const ToolChain &TC) const { 5890 SmallVector<std::string, 2> TargetSpecificExecutables; 5891 generatePrefixedToolNames(Name, TC, TargetSpecificExecutables); 5892 5893 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5894 // attempting to use this prefix when looking for program paths. 5895 for (const auto &PrefixDir : PrefixDirs) { 5896 if (llvm::sys::fs::is_directory(PrefixDir)) { 5897 SmallString<128> P(PrefixDir); 5898 if (ScanDirForExecutable(P, Name)) 5899 return std::string(P.str()); 5900 } else { 5901 SmallString<128> P((PrefixDir + Name).str()); 5902 if (llvm::sys::fs::can_execute(Twine(P))) 5903 return std::string(P.str()); 5904 } 5905 } 5906 5907 const ToolChain::path_list &List = TC.getProgramPaths(); 5908 for (const auto &TargetSpecificExecutable : TargetSpecificExecutables) { 5909 // For each possible name of the tool look for it in 5910 // program paths first, then the path. 5911 // Higher priority names will be first, meaning that 5912 // a higher priority name in the path will be found 5913 // instead of a lower priority name in the program path. 5914 // E.g. <triple>-gcc on the path will be found instead 5915 // of gcc in the program path 5916 for (const auto &Path : List) { 5917 SmallString<128> P(Path); 5918 if (ScanDirForExecutable(P, TargetSpecificExecutable)) 5919 return std::string(P.str()); 5920 } 5921 5922 // Fall back to the path 5923 if (llvm::ErrorOr<std::string> P = 5924 llvm::sys::findProgramByName(TargetSpecificExecutable)) 5925 return *P; 5926 } 5927 5928 return std::string(Name); 5929 } 5930 5931 std::string Driver::GetTemporaryPath(StringRef Prefix, StringRef Suffix) const { 5932 SmallString<128> Path; 5933 std::error_code EC = llvm::sys::fs::createTemporaryFile(Prefix, Suffix, Path); 5934 if (EC) { 5935 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5936 return ""; 5937 } 5938 5939 return std::string(Path.str()); 5940 } 5941 5942 std::string Driver::GetTemporaryDirectory(StringRef Prefix) const { 5943 SmallString<128> Path; 5944 std::error_code EC = llvm::sys::fs::createUniqueDirectory(Prefix, Path); 5945 if (EC) { 5946 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5947 return ""; 5948 } 5949 5950 return std::string(Path.str()); 5951 } 5952 5953 std::string Driver::GetClPchPath(Compilation &C, StringRef BaseName) const { 5954 SmallString<128> Output; 5955 if (Arg *FpArg = C.getArgs().getLastArg(options::OPT__SLASH_Fp)) { 5956 // FIXME: If anybody needs it, implement this obscure rule: 5957 // "If you specify a directory without a file name, the default file name 5958 // is VCx0.pch., where x is the major version of Visual C++ in use." 5959 Output = FpArg->getValue(); 5960 5961 // "If you do not specify an extension as part of the path name, an 5962 // extension of .pch is assumed. " 5963 if (!llvm::sys::path::has_extension(Output)) 5964 Output += ".pch"; 5965 } else { 5966 if (Arg *YcArg = C.getArgs().getLastArg(options::OPT__SLASH_Yc)) 5967 Output = YcArg->getValue(); 5968 if (Output.empty()) 5969 Output = BaseName; 5970 llvm::sys::path::replace_extension(Output, ".pch"); 5971 } 5972 return std::string(Output.str()); 5973 } 5974 5975 const ToolChain &Driver::getToolChain(const ArgList &Args, 5976 const llvm::Triple &Target) const { 5977 5978 auto &TC = ToolChains[Target.str()]; 5979 if (!TC) { 5980 switch (Target.getOS()) { 5981 case llvm::Triple::AIX: 5982 TC = std::make_unique<toolchains::AIX>(*this, Target, Args); 5983 break; 5984 case llvm::Triple::Haiku: 5985 TC = std::make_unique<toolchains::Haiku>(*this, Target, Args); 5986 break; 5987 case llvm::Triple::Ananas: 5988 TC = std::make_unique<toolchains::Ananas>(*this, Target, Args); 5989 break; 5990 case llvm::Triple::CloudABI: 5991 TC = std::make_unique<toolchains::CloudABI>(*this, Target, Args); 5992 break; 5993 case llvm::Triple::Darwin: 5994 case llvm::Triple::MacOSX: 5995 case llvm::Triple::IOS: 5996 case llvm::Triple::TvOS: 5997 case llvm::Triple::WatchOS: 5998 case llvm::Triple::DriverKit: 5999 TC = std::make_unique<toolchains::DarwinClang>(*this, Target, Args); 6000 break; 6001 case llvm::Triple::DragonFly: 6002 TC = std::make_unique<toolchains::DragonFly>(*this, Target, Args); 6003 break; 6004 case llvm::Triple::OpenBSD: 6005 TC = std::make_unique<toolchains::OpenBSD>(*this, Target, Args); 6006 break; 6007 case llvm::Triple::NetBSD: 6008 TC = std::make_unique<toolchains::NetBSD>(*this, Target, Args); 6009 break; 6010 case llvm::Triple::FreeBSD: 6011 if (Target.isPPC()) 6012 TC = std::make_unique<toolchains::PPCFreeBSDToolChain>(*this, Target, 6013 Args); 6014 else 6015 TC = std::make_unique<toolchains::FreeBSD>(*this, Target, Args); 6016 break; 6017 case llvm::Triple::Minix: 6018 TC = std::make_unique<toolchains::Minix>(*this, Target, Args); 6019 break; 6020 case llvm::Triple::Linux: 6021 case llvm::Triple::ELFIAMCU: 6022 if (Target.getArch() == llvm::Triple::hexagon) 6023 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 6024 Args); 6025 else if ((Target.getVendor() == llvm::Triple::MipsTechnologies) && 6026 !Target.hasEnvironment()) 6027 TC = std::make_unique<toolchains::MipsLLVMToolChain>(*this, Target, 6028 Args); 6029 else if (Target.isPPC()) 6030 TC = std::make_unique<toolchains::PPCLinuxToolChain>(*this, Target, 6031 Args); 6032 else if (Target.getArch() == llvm::Triple::ve) 6033 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6034 6035 else 6036 TC = std::make_unique<toolchains::Linux>(*this, Target, Args); 6037 break; 6038 case llvm::Triple::NaCl: 6039 TC = std::make_unique<toolchains::NaClToolChain>(*this, Target, Args); 6040 break; 6041 case llvm::Triple::Fuchsia: 6042 TC = std::make_unique<toolchains::Fuchsia>(*this, Target, Args); 6043 break; 6044 case llvm::Triple::Solaris: 6045 TC = std::make_unique<toolchains::Solaris>(*this, Target, Args); 6046 break; 6047 case llvm::Triple::CUDA: 6048 TC = std::make_unique<toolchains::NVPTXToolChain>(*this, Target, Args); 6049 break; 6050 case llvm::Triple::AMDHSA: 6051 TC = std::make_unique<toolchains::ROCMToolChain>(*this, Target, Args); 6052 break; 6053 case llvm::Triple::AMDPAL: 6054 case llvm::Triple::Mesa3D: 6055 TC = std::make_unique<toolchains::AMDGPUToolChain>(*this, Target, Args); 6056 break; 6057 case llvm::Triple::Win32: 6058 switch (Target.getEnvironment()) { 6059 default: 6060 if (Target.isOSBinFormatELF()) 6061 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6062 else if (Target.isOSBinFormatMachO()) 6063 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6064 else 6065 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6066 break; 6067 case llvm::Triple::GNU: 6068 TC = std::make_unique<toolchains::MinGW>(*this, Target, Args); 6069 break; 6070 case llvm::Triple::Itanium: 6071 TC = std::make_unique<toolchains::CrossWindowsToolChain>(*this, Target, 6072 Args); 6073 break; 6074 case llvm::Triple::MSVC: 6075 case llvm::Triple::UnknownEnvironment: 6076 if (Args.getLastArgValue(options::OPT_fuse_ld_EQ) 6077 .startswith_insensitive("bfd")) 6078 TC = std::make_unique<toolchains::CrossWindowsToolChain>( 6079 *this, Target, Args); 6080 else 6081 TC = 6082 std::make_unique<toolchains::MSVCToolChain>(*this, Target, Args); 6083 break; 6084 } 6085 break; 6086 case llvm::Triple::PS4: 6087 TC = std::make_unique<toolchains::PS4CPU>(*this, Target, Args); 6088 break; 6089 case llvm::Triple::PS5: 6090 TC = std::make_unique<toolchains::PS5CPU>(*this, Target, Args); 6091 break; 6092 case llvm::Triple::Contiki: 6093 TC = std::make_unique<toolchains::Contiki>(*this, Target, Args); 6094 break; 6095 case llvm::Triple::Hurd: 6096 TC = std::make_unique<toolchains::Hurd>(*this, Target, Args); 6097 break; 6098 case llvm::Triple::ZOS: 6099 TC = std::make_unique<toolchains::ZOS>(*this, Target, Args); 6100 break; 6101 case llvm::Triple::ShaderModel: 6102 TC = std::make_unique<toolchains::HLSLToolChain>(*this, Target, Args); 6103 break; 6104 default: 6105 // Of these targets, Hexagon is the only one that might have 6106 // an OS of Linux, in which case it got handled above already. 6107 switch (Target.getArch()) { 6108 case llvm::Triple::tce: 6109 TC = std::make_unique<toolchains::TCEToolChain>(*this, Target, Args); 6110 break; 6111 case llvm::Triple::tcele: 6112 TC = std::make_unique<toolchains::TCELEToolChain>(*this, Target, Args); 6113 break; 6114 case llvm::Triple::hexagon: 6115 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 6116 Args); 6117 break; 6118 case llvm::Triple::lanai: 6119 TC = std::make_unique<toolchains::LanaiToolChain>(*this, Target, Args); 6120 break; 6121 case llvm::Triple::xcore: 6122 TC = std::make_unique<toolchains::XCoreToolChain>(*this, Target, Args); 6123 break; 6124 case llvm::Triple::wasm32: 6125 case llvm::Triple::wasm64: 6126 TC = std::make_unique<toolchains::WebAssembly>(*this, Target, Args); 6127 break; 6128 case llvm::Triple::avr: 6129 TC = std::make_unique<toolchains::AVRToolChain>(*this, Target, Args); 6130 break; 6131 case llvm::Triple::msp430: 6132 TC = 6133 std::make_unique<toolchains::MSP430ToolChain>(*this, Target, Args); 6134 break; 6135 case llvm::Triple::riscv32: 6136 case llvm::Triple::riscv64: 6137 if (toolchains::RISCVToolChain::hasGCCToolchain(*this, Args)) 6138 TC = 6139 std::make_unique<toolchains::RISCVToolChain>(*this, Target, Args); 6140 else 6141 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6142 break; 6143 case llvm::Triple::ve: 6144 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6145 break; 6146 case llvm::Triple::spirv32: 6147 case llvm::Triple::spirv64: 6148 TC = std::make_unique<toolchains::SPIRVToolChain>(*this, Target, Args); 6149 break; 6150 case llvm::Triple::csky: 6151 TC = std::make_unique<toolchains::CSKYToolChain>(*this, Target, Args); 6152 break; 6153 default: 6154 if (Target.getVendor() == llvm::Triple::Myriad) 6155 TC = std::make_unique<toolchains::MyriadToolChain>(*this, Target, 6156 Args); 6157 else if (toolchains::BareMetal::handlesTarget(Target)) 6158 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6159 else if (Target.isOSBinFormatELF()) 6160 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6161 else if (Target.isOSBinFormatMachO()) 6162 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6163 else 6164 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6165 } 6166 } 6167 } 6168 6169 return *TC; 6170 } 6171 6172 const ToolChain &Driver::getOffloadingDeviceToolChain( 6173 const ArgList &Args, const llvm::Triple &Target, const ToolChain &HostTC, 6174 const Action::OffloadKind &TargetDeviceOffloadKind) const { 6175 // Use device / host triples as the key into the ToolChains map because the 6176 // device ToolChain we create depends on both. 6177 auto &TC = ToolChains[Target.str() + "/" + HostTC.getTriple().str()]; 6178 if (!TC) { 6179 // Categorized by offload kind > arch rather than OS > arch like 6180 // the normal getToolChain call, as it seems a reasonable way to categorize 6181 // things. 6182 switch (TargetDeviceOffloadKind) { 6183 case Action::OFK_HIP: { 6184 if (Target.getArch() == llvm::Triple::amdgcn && 6185 Target.getVendor() == llvm::Triple::AMD && 6186 Target.getOS() == llvm::Triple::AMDHSA) 6187 TC = std::make_unique<toolchains::HIPAMDToolChain>(*this, Target, 6188 HostTC, Args); 6189 else if (Target.getArch() == llvm::Triple::spirv64 && 6190 Target.getVendor() == llvm::Triple::UnknownVendor && 6191 Target.getOS() == llvm::Triple::UnknownOS) 6192 TC = std::make_unique<toolchains::HIPSPVToolChain>(*this, Target, 6193 HostTC, Args); 6194 break; 6195 } 6196 default: 6197 break; 6198 } 6199 } 6200 6201 return *TC; 6202 } 6203 6204 bool Driver::ShouldUseClangCompiler(const JobAction &JA) const { 6205 // Say "no" if there is not exactly one input of a type clang understands. 6206 if (JA.size() != 1 || 6207 !types::isAcceptedByClang((*JA.input_begin())->getType())) 6208 return false; 6209 6210 // And say "no" if this is not a kind of action clang understands. 6211 if (!isa<PreprocessJobAction>(JA) && !isa<PrecompileJobAction>(JA) && 6212 !isa<CompileJobAction>(JA) && !isa<BackendJobAction>(JA) && 6213 !isa<ExtractAPIJobAction>(JA)) 6214 return false; 6215 6216 return true; 6217 } 6218 6219 bool Driver::ShouldUseFlangCompiler(const JobAction &JA) const { 6220 // Say "no" if there is not exactly one input of a type flang understands. 6221 if (JA.size() != 1 || 6222 !types::isAcceptedByFlang((*JA.input_begin())->getType())) 6223 return false; 6224 6225 // And say "no" if this is not a kind of action flang understands. 6226 if (!isa<PreprocessJobAction>(JA) && !isa<CompileJobAction>(JA) && 6227 !isa<BackendJobAction>(JA)) 6228 return false; 6229 6230 return true; 6231 } 6232 6233 bool Driver::ShouldEmitStaticLibrary(const ArgList &Args) const { 6234 // Only emit static library if the flag is set explicitly. 6235 if (Args.hasArg(options::OPT_emit_static_lib)) 6236 return true; 6237 return false; 6238 } 6239 6240 /// GetReleaseVersion - Parse (([0-9]+)(.([0-9]+)(.([0-9]+)?))?)? and return the 6241 /// grouped values as integers. Numbers which are not provided are set to 0. 6242 /// 6243 /// \return True if the entire string was parsed (9.2), or all groups were 6244 /// parsed (10.3.5extrastuff). 6245 bool Driver::GetReleaseVersion(StringRef Str, unsigned &Major, unsigned &Minor, 6246 unsigned &Micro, bool &HadExtra) { 6247 HadExtra = false; 6248 6249 Major = Minor = Micro = 0; 6250 if (Str.empty()) 6251 return false; 6252 6253 if (Str.consumeInteger(10, Major)) 6254 return false; 6255 if (Str.empty()) 6256 return true; 6257 if (Str[0] != '.') 6258 return false; 6259 6260 Str = Str.drop_front(1); 6261 6262 if (Str.consumeInteger(10, Minor)) 6263 return false; 6264 if (Str.empty()) 6265 return true; 6266 if (Str[0] != '.') 6267 return false; 6268 Str = Str.drop_front(1); 6269 6270 if (Str.consumeInteger(10, Micro)) 6271 return false; 6272 if (!Str.empty()) 6273 HadExtra = true; 6274 return true; 6275 } 6276 6277 /// Parse digits from a string \p Str and fulfill \p Digits with 6278 /// the parsed numbers. This method assumes that the max number of 6279 /// digits to look for is equal to Digits.size(). 6280 /// 6281 /// \return True if the entire string was parsed and there are 6282 /// no extra characters remaining at the end. 6283 bool Driver::GetReleaseVersion(StringRef Str, 6284 MutableArrayRef<unsigned> Digits) { 6285 if (Str.empty()) 6286 return false; 6287 6288 unsigned CurDigit = 0; 6289 while (CurDigit < Digits.size()) { 6290 unsigned Digit; 6291 if (Str.consumeInteger(10, Digit)) 6292 return false; 6293 Digits[CurDigit] = Digit; 6294 if (Str.empty()) 6295 return true; 6296 if (Str[0] != '.') 6297 return false; 6298 Str = Str.drop_front(1); 6299 CurDigit++; 6300 } 6301 6302 // More digits than requested, bail out... 6303 return false; 6304 } 6305 6306 std::pair<unsigned, unsigned> 6307 Driver::getIncludeExcludeOptionFlagMasks(bool IsClCompatMode) const { 6308 unsigned IncludedFlagsBitmask = 0; 6309 unsigned ExcludedFlagsBitmask = options::NoDriverOption; 6310 6311 if (IsClCompatMode) { 6312 // Include CL and Core options. 6313 IncludedFlagsBitmask |= options::CLOption; 6314 IncludedFlagsBitmask |= options::CLDXCOption; 6315 IncludedFlagsBitmask |= options::CoreOption; 6316 } else { 6317 ExcludedFlagsBitmask |= options::CLOption; 6318 } 6319 if (IsDXCMode()) { 6320 // Include DXC and Core options. 6321 IncludedFlagsBitmask |= options::DXCOption; 6322 IncludedFlagsBitmask |= options::CLDXCOption; 6323 IncludedFlagsBitmask |= options::CoreOption; 6324 } else { 6325 ExcludedFlagsBitmask |= options::DXCOption; 6326 } 6327 if (!IsClCompatMode && !IsDXCMode()) 6328 ExcludedFlagsBitmask |= options::CLDXCOption; 6329 6330 return std::make_pair(IncludedFlagsBitmask, ExcludedFlagsBitmask); 6331 } 6332 6333 const char *Driver::getExecutableForDriverMode(DriverMode Mode) { 6334 switch (Mode) { 6335 case GCCMode: 6336 return "clang"; 6337 case GXXMode: 6338 return "clang++"; 6339 case CPPMode: 6340 return "clang-cpp"; 6341 case CLMode: 6342 return "clang-cl"; 6343 case FlangMode: 6344 return "flang"; 6345 case DXCMode: 6346 return "clang-dxc"; 6347 } 6348 6349 llvm_unreachable("Unhandled Mode"); 6350 } 6351 6352 bool clang::driver::isOptimizationLevelFast(const ArgList &Args) { 6353 return Args.hasFlag(options::OPT_Ofast, options::OPT_O_Group, false); 6354 } 6355 6356 bool clang::driver::willEmitRemarks(const ArgList &Args) { 6357 // -fsave-optimization-record enables it. 6358 if (Args.hasFlag(options::OPT_fsave_optimization_record, 6359 options::OPT_fno_save_optimization_record, false)) 6360 return true; 6361 6362 // -fsave-optimization-record=<format> enables it as well. 6363 if (Args.hasFlag(options::OPT_fsave_optimization_record_EQ, 6364 options::OPT_fno_save_optimization_record, false)) 6365 return true; 6366 6367 // -foptimization-record-file alone enables it too. 6368 if (Args.hasFlag(options::OPT_foptimization_record_file_EQ, 6369 options::OPT_fno_save_optimization_record, false)) 6370 return true; 6371 6372 // -foptimization-record-passes alone enables it too. 6373 if (Args.hasFlag(options::OPT_foptimization_record_passes_EQ, 6374 options::OPT_fno_save_optimization_record, false)) 6375 return true; 6376 return false; 6377 } 6378 6379 llvm::StringRef clang::driver::getDriverMode(StringRef ProgName, 6380 ArrayRef<const char *> Args) { 6381 static const std::string OptName = 6382 getDriverOptTable().getOption(options::OPT_driver_mode).getPrefixedName(); 6383 llvm::StringRef Opt; 6384 for (StringRef Arg : Args) { 6385 if (!Arg.startswith(OptName)) 6386 continue; 6387 Opt = Arg; 6388 } 6389 if (Opt.empty()) 6390 Opt = ToolChain::getTargetAndModeFromProgramName(ProgName).DriverMode; 6391 return Opt.consume_front(OptName) ? Opt : ""; 6392 } 6393 6394 bool driver::IsClangCL(StringRef DriverMode) { return DriverMode.equals("cl"); } 6395