1 //===--- Driver.cpp - Clang GCC Compatible Driver -------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "clang/Driver/Driver.h" 10 #include "ToolChains/AIX.h" 11 #include "ToolChains/AMDGPU.h" 12 #include "ToolChains/AMDGPUOpenMP.h" 13 #include "ToolChains/AVR.h" 14 #include "ToolChains/Ananas.h" 15 #include "ToolChains/BareMetal.h" 16 #include "ToolChains/CSKYToolChain.h" 17 #include "ToolChains/Clang.h" 18 #include "ToolChains/CloudABI.h" 19 #include "ToolChains/Contiki.h" 20 #include "ToolChains/CrossWindows.h" 21 #include "ToolChains/Cuda.h" 22 #include "ToolChains/Darwin.h" 23 #include "ToolChains/DragonFly.h" 24 #include "ToolChains/FreeBSD.h" 25 #include "ToolChains/Fuchsia.h" 26 #include "ToolChains/Gnu.h" 27 #include "ToolChains/HIPAMD.h" 28 #include "ToolChains/HIPSPV.h" 29 #include "ToolChains/HLSL.h" 30 #include "ToolChains/Haiku.h" 31 #include "ToolChains/Hexagon.h" 32 #include "ToolChains/Hurd.h" 33 #include "ToolChains/Lanai.h" 34 #include "ToolChains/Linux.h" 35 #include "ToolChains/MSP430.h" 36 #include "ToolChains/MSVC.h" 37 #include "ToolChains/MinGW.h" 38 #include "ToolChains/Minix.h" 39 #include "ToolChains/MipsLinux.h" 40 #include "ToolChains/Myriad.h" 41 #include "ToolChains/NaCl.h" 42 #include "ToolChains/NetBSD.h" 43 #include "ToolChains/OpenBSD.h" 44 #include "ToolChains/PPCFreeBSD.h" 45 #include "ToolChains/PPCLinux.h" 46 #include "ToolChains/PS4CPU.h" 47 #include "ToolChains/RISCVToolchain.h" 48 #include "ToolChains/SPIRV.h" 49 #include "ToolChains/Solaris.h" 50 #include "ToolChains/TCE.h" 51 #include "ToolChains/VEToolchain.h" 52 #include "ToolChains/WebAssembly.h" 53 #include "ToolChains/XCore.h" 54 #include "ToolChains/ZOS.h" 55 #include "clang/Basic/TargetID.h" 56 #include "clang/Basic/Version.h" 57 #include "clang/Config/config.h" 58 #include "clang/Driver/Action.h" 59 #include "clang/Driver/Compilation.h" 60 #include "clang/Driver/DriverDiagnostic.h" 61 #include "clang/Driver/InputInfo.h" 62 #include "clang/Driver/Job.h" 63 #include "clang/Driver/Options.h" 64 #include "clang/Driver/Phases.h" 65 #include "clang/Driver/SanitizerArgs.h" 66 #include "clang/Driver/Tool.h" 67 #include "clang/Driver/ToolChain.h" 68 #include "clang/Driver/Types.h" 69 #include "llvm/ADT/ArrayRef.h" 70 #include "llvm/ADT/STLExtras.h" 71 #include "llvm/ADT/SmallSet.h" 72 #include "llvm/ADT/StringExtras.h" 73 #include "llvm/ADT/StringRef.h" 74 #include "llvm/ADT/StringSet.h" 75 #include "llvm/ADT/StringSwitch.h" 76 #include "llvm/Config/llvm-config.h" 77 #include "llvm/MC/TargetRegistry.h" 78 #include "llvm/Option/Arg.h" 79 #include "llvm/Option/ArgList.h" 80 #include "llvm/Option/OptSpecifier.h" 81 #include "llvm/Option/OptTable.h" 82 #include "llvm/Option/Option.h" 83 #include "llvm/Support/CommandLine.h" 84 #include "llvm/Support/ErrorHandling.h" 85 #include "llvm/Support/ExitCodes.h" 86 #include "llvm/Support/FileSystem.h" 87 #include "llvm/Support/FormatVariadic.h" 88 #include "llvm/Support/Host.h" 89 #include "llvm/Support/MD5.h" 90 #include "llvm/Support/Path.h" 91 #include "llvm/Support/PrettyStackTrace.h" 92 #include "llvm/Support/Process.h" 93 #include "llvm/Support/Program.h" 94 #include "llvm/Support/StringSaver.h" 95 #include "llvm/Support/VirtualFileSystem.h" 96 #include "llvm/Support/raw_ostream.h" 97 #include <map> 98 #include <memory> 99 #include <utility> 100 #if LLVM_ON_UNIX 101 #include <unistd.h> // getpid 102 #endif 103 104 using namespace clang::driver; 105 using namespace clang; 106 using namespace llvm::opt; 107 108 static llvm::Optional<llvm::Triple> 109 getOffloadTargetTriple(const Driver &D, const ArgList &Args) { 110 auto OffloadTargets = Args.getAllArgValues(options::OPT_offload_EQ); 111 // Offload compilation flow does not support multiple targets for now. We 112 // need the HIPActionBuilder (and possibly the CudaActionBuilder{,Base}too) 113 // to support multiple tool chains first. 114 switch (OffloadTargets.size()) { 115 default: 116 D.Diag(diag::err_drv_only_one_offload_target_supported); 117 return llvm::None; 118 case 0: 119 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << ""; 120 return llvm::None; 121 case 1: 122 break; 123 } 124 return llvm::Triple(OffloadTargets[0]); 125 } 126 127 static llvm::Optional<llvm::Triple> 128 getNVIDIAOffloadTargetTriple(const Driver &D, const ArgList &Args, 129 const llvm::Triple &HostTriple) { 130 if (!Args.hasArg(options::OPT_offload_EQ)) { 131 return llvm::Triple(HostTriple.isArch64Bit() ? "nvptx64-nvidia-cuda" 132 : "nvptx-nvidia-cuda"); 133 } 134 auto TT = getOffloadTargetTriple(D, Args); 135 if (TT && (TT->getArch() == llvm::Triple::spirv32 || 136 TT->getArch() == llvm::Triple::spirv64)) { 137 if (Args.hasArg(options::OPT_emit_llvm)) 138 return TT; 139 D.Diag(diag::err_drv_cuda_offload_only_emit_bc); 140 return llvm::None; 141 } 142 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 143 return llvm::None; 144 } 145 static llvm::Optional<llvm::Triple> 146 getHIPOffloadTargetTriple(const Driver &D, const ArgList &Args) { 147 if (!Args.hasArg(options::OPT_offload_EQ)) { 148 return llvm::Triple("amdgcn-amd-amdhsa"); // Default HIP triple. 149 } 150 auto TT = getOffloadTargetTriple(D, Args); 151 if (!TT) 152 return llvm::None; 153 if (TT->getArch() == llvm::Triple::amdgcn && 154 TT->getVendor() == llvm::Triple::AMD && 155 TT->getOS() == llvm::Triple::AMDHSA) 156 return TT; 157 if (TT->getArch() == llvm::Triple::spirv64) 158 return TT; 159 D.Diag(diag::err_drv_invalid_or_unsupported_offload_target) << TT->str(); 160 return llvm::None; 161 } 162 163 // static 164 std::string Driver::GetResourcesPath(StringRef BinaryPath, 165 StringRef CustomResourceDir) { 166 // Since the resource directory is embedded in the module hash, it's important 167 // that all places that need it call this function, so that they get the 168 // exact same string ("a/../b/" and "b/" get different hashes, for example). 169 170 // Dir is bin/ or lib/, depending on where BinaryPath is. 171 std::string Dir = std::string(llvm::sys::path::parent_path(BinaryPath)); 172 173 SmallString<128> P(Dir); 174 if (CustomResourceDir != "") { 175 llvm::sys::path::append(P, CustomResourceDir); 176 } else { 177 // On Windows, libclang.dll is in bin/. 178 // On non-Windows, libclang.so/.dylib is in lib/. 179 // With a static-library build of libclang, LibClangPath will contain the 180 // path of the embedding binary, which for LLVM binaries will be in bin/. 181 // ../lib gets us to lib/ in both cases. 182 P = llvm::sys::path::parent_path(Dir); 183 llvm::sys::path::append(P, Twine("lib") + CLANG_LIBDIR_SUFFIX, "clang", 184 CLANG_VERSION_STRING); 185 } 186 187 return std::string(P.str()); 188 } 189 190 Driver::Driver(StringRef ClangExecutable, StringRef TargetTriple, 191 DiagnosticsEngine &Diags, std::string Title, 192 IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS) 193 : Diags(Diags), VFS(std::move(VFS)), Mode(GCCMode), 194 SaveTemps(SaveTempsNone), BitcodeEmbed(EmbedNone), 195 CXX20HeaderType(HeaderMode_None), ModulesModeCXX20(false), 196 LTOMode(LTOK_None), ClangExecutable(ClangExecutable), 197 SysRoot(DEFAULT_SYSROOT), DriverTitle(Title), CCCPrintBindings(false), 198 CCPrintOptions(false), CCPrintHeaders(false), CCLogDiagnostics(false), 199 CCGenDiagnostics(false), CCPrintProcessStats(false), 200 TargetTriple(TargetTriple), Saver(Alloc), CheckInputsExist(true), 201 ProbePrecompiled(true), SuppressMissingInputWarning(false) { 202 // Provide a sane fallback if no VFS is specified. 203 if (!this->VFS) 204 this->VFS = llvm::vfs::getRealFileSystem(); 205 206 Name = std::string(llvm::sys::path::filename(ClangExecutable)); 207 Dir = std::string(llvm::sys::path::parent_path(ClangExecutable)); 208 InstalledDir = Dir; // Provide a sensible default installed dir. 209 210 if ((!SysRoot.empty()) && llvm::sys::path::is_relative(SysRoot)) { 211 // Prepend InstalledDir if SysRoot is relative 212 SmallString<128> P(InstalledDir); 213 llvm::sys::path::append(P, SysRoot); 214 SysRoot = std::string(P); 215 } 216 217 #if defined(CLANG_CONFIG_FILE_SYSTEM_DIR) 218 SystemConfigDir = CLANG_CONFIG_FILE_SYSTEM_DIR; 219 #endif 220 #if defined(CLANG_CONFIG_FILE_USER_DIR) 221 UserConfigDir = CLANG_CONFIG_FILE_USER_DIR; 222 #endif 223 224 // Compute the path to the resource directory. 225 ResourceDir = GetResourcesPath(ClangExecutable, CLANG_RESOURCE_DIR); 226 } 227 228 void Driver::setDriverMode(StringRef Value) { 229 static const std::string OptName = 230 getOpts().getOption(options::OPT_driver_mode).getPrefixedName(); 231 if (auto M = llvm::StringSwitch<llvm::Optional<DriverMode>>(Value) 232 .Case("gcc", GCCMode) 233 .Case("g++", GXXMode) 234 .Case("cpp", CPPMode) 235 .Case("cl", CLMode) 236 .Case("flang", FlangMode) 237 .Case("dxc", DXCMode) 238 .Default(None)) 239 Mode = *M; 240 else 241 Diag(diag::err_drv_unsupported_option_argument) << OptName << Value; 242 } 243 244 InputArgList Driver::ParseArgStrings(ArrayRef<const char *> ArgStrings, 245 bool IsClCompatMode, 246 bool &ContainsError) { 247 llvm::PrettyStackTraceString CrashInfo("Command line argument parsing"); 248 ContainsError = false; 249 250 unsigned IncludedFlagsBitmask; 251 unsigned ExcludedFlagsBitmask; 252 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 253 getIncludeExcludeOptionFlagMasks(IsClCompatMode); 254 255 // Make sure that Flang-only options don't pollute the Clang output 256 // TODO: Make sure that Clang-only options don't pollute Flang output 257 if (!IsFlangMode()) 258 ExcludedFlagsBitmask |= options::FlangOnlyOption; 259 260 unsigned MissingArgIndex, MissingArgCount; 261 InputArgList Args = 262 getOpts().ParseArgs(ArgStrings, MissingArgIndex, MissingArgCount, 263 IncludedFlagsBitmask, ExcludedFlagsBitmask); 264 265 // Check for missing argument error. 266 if (MissingArgCount) { 267 Diag(diag::err_drv_missing_argument) 268 << Args.getArgString(MissingArgIndex) << MissingArgCount; 269 ContainsError |= 270 Diags.getDiagnosticLevel(diag::err_drv_missing_argument, 271 SourceLocation()) > DiagnosticsEngine::Warning; 272 } 273 274 // Check for unsupported options. 275 for (const Arg *A : Args) { 276 if (A->getOption().hasFlag(options::Unsupported)) { 277 unsigned DiagID; 278 auto ArgString = A->getAsString(Args); 279 std::string Nearest; 280 if (getOpts().findNearest( 281 ArgString, Nearest, IncludedFlagsBitmask, 282 ExcludedFlagsBitmask | options::Unsupported) > 1) { 283 DiagID = diag::err_drv_unsupported_opt; 284 Diag(DiagID) << ArgString; 285 } else { 286 DiagID = diag::err_drv_unsupported_opt_with_suggestion; 287 Diag(DiagID) << ArgString << Nearest; 288 } 289 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 290 DiagnosticsEngine::Warning; 291 continue; 292 } 293 294 // Warn about -mcpu= without an argument. 295 if (A->getOption().matches(options::OPT_mcpu_EQ) && A->containsValue("")) { 296 Diag(diag::warn_drv_empty_joined_argument) << A->getAsString(Args); 297 ContainsError |= Diags.getDiagnosticLevel( 298 diag::warn_drv_empty_joined_argument, 299 SourceLocation()) > DiagnosticsEngine::Warning; 300 } 301 } 302 303 for (const Arg *A : Args.filtered(options::OPT_UNKNOWN)) { 304 unsigned DiagID; 305 auto ArgString = A->getAsString(Args); 306 std::string Nearest; 307 if (getOpts().findNearest( 308 ArgString, Nearest, IncludedFlagsBitmask, ExcludedFlagsBitmask) > 1) { 309 DiagID = IsCLMode() ? diag::warn_drv_unknown_argument_clang_cl 310 : diag::err_drv_unknown_argument; 311 Diags.Report(DiagID) << ArgString; 312 } else { 313 DiagID = IsCLMode() 314 ? diag::warn_drv_unknown_argument_clang_cl_with_suggestion 315 : diag::err_drv_unknown_argument_with_suggestion; 316 Diags.Report(DiagID) << ArgString << Nearest; 317 } 318 ContainsError |= Diags.getDiagnosticLevel(DiagID, SourceLocation()) > 319 DiagnosticsEngine::Warning; 320 } 321 322 return Args; 323 } 324 325 // Determine which compilation mode we are in. We look for options which 326 // affect the phase, starting with the earliest phases, and record which 327 // option we used to determine the final phase. 328 phases::ID Driver::getFinalPhase(const DerivedArgList &DAL, 329 Arg **FinalPhaseArg) const { 330 Arg *PhaseArg = nullptr; 331 phases::ID FinalPhase; 332 333 // -{E,EP,P,M,MM} only run the preprocessor. 334 if (CCCIsCPP() || (PhaseArg = DAL.getLastArg(options::OPT_E)) || 335 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_EP)) || 336 (PhaseArg = DAL.getLastArg(options::OPT_M, options::OPT_MM)) || 337 (PhaseArg = DAL.getLastArg(options::OPT__SLASH_P)) || 338 CCGenDiagnostics) { 339 FinalPhase = phases::Preprocess; 340 341 // --precompile only runs up to precompilation. 342 // Options that cause the output of C++20 compiled module interfaces or 343 // header units have the same effect. 344 } else if ((PhaseArg = DAL.getLastArg(options::OPT__precompile)) || 345 (PhaseArg = DAL.getLastArg(options::OPT_extract_api)) || 346 (PhaseArg = DAL.getLastArg(options::OPT_fmodule_header, 347 options::OPT_fmodule_header_EQ))) { 348 FinalPhase = phases::Precompile; 349 // -{fsyntax-only,-analyze,emit-ast} only run up to the compiler. 350 } else if ((PhaseArg = DAL.getLastArg(options::OPT_fsyntax_only)) || 351 (PhaseArg = DAL.getLastArg(options::OPT_print_supported_cpus)) || 352 (PhaseArg = DAL.getLastArg(options::OPT_module_file_info)) || 353 (PhaseArg = DAL.getLastArg(options::OPT_verify_pch)) || 354 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_objc)) || 355 (PhaseArg = DAL.getLastArg(options::OPT_rewrite_legacy_objc)) || 356 (PhaseArg = DAL.getLastArg(options::OPT__migrate)) || 357 (PhaseArg = DAL.getLastArg(options::OPT__analyze)) || 358 (PhaseArg = DAL.getLastArg(options::OPT_emit_ast))) { 359 FinalPhase = phases::Compile; 360 361 // -S only runs up to the backend. 362 } else if ((PhaseArg = DAL.getLastArg(options::OPT_S))) { 363 FinalPhase = phases::Backend; 364 365 // -c compilation only runs up to the assembler. 366 } else if ((PhaseArg = DAL.getLastArg(options::OPT_c))) { 367 FinalPhase = phases::Assemble; 368 369 } else if ((PhaseArg = DAL.getLastArg(options::OPT_emit_interface_stubs))) { 370 FinalPhase = phases::IfsMerge; 371 372 // Otherwise do everything. 373 } else 374 FinalPhase = phases::Link; 375 376 if (FinalPhaseArg) 377 *FinalPhaseArg = PhaseArg; 378 379 return FinalPhase; 380 } 381 382 static Arg *MakeInputArg(DerivedArgList &Args, const OptTable &Opts, 383 StringRef Value, bool Claim = true) { 384 Arg *A = new Arg(Opts.getOption(options::OPT_INPUT), Value, 385 Args.getBaseArgs().MakeIndex(Value), Value.data()); 386 Args.AddSynthesizedArg(A); 387 if (Claim) 388 A->claim(); 389 return A; 390 } 391 392 DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args) const { 393 const llvm::opt::OptTable &Opts = getOpts(); 394 DerivedArgList *DAL = new DerivedArgList(Args); 395 396 bool HasNostdlib = Args.hasArg(options::OPT_nostdlib); 397 bool HasNostdlibxx = Args.hasArg(options::OPT_nostdlibxx); 398 bool HasNodefaultlib = Args.hasArg(options::OPT_nodefaultlibs); 399 bool IgnoreUnused = false; 400 for (Arg *A : Args) { 401 if (IgnoreUnused) 402 A->claim(); 403 404 if (A->getOption().matches(options::OPT_start_no_unused_arguments)) { 405 IgnoreUnused = true; 406 continue; 407 } 408 if (A->getOption().matches(options::OPT_end_no_unused_arguments)) { 409 IgnoreUnused = false; 410 continue; 411 } 412 413 // Unfortunately, we have to parse some forwarding options (-Xassembler, 414 // -Xlinker, -Xpreprocessor) because we either integrate their functionality 415 // (assembler and preprocessor), or bypass a previous driver ('collect2'). 416 417 // Rewrite linker options, to replace --no-demangle with a custom internal 418 // option. 419 if ((A->getOption().matches(options::OPT_Wl_COMMA) || 420 A->getOption().matches(options::OPT_Xlinker)) && 421 A->containsValue("--no-demangle")) { 422 // Add the rewritten no-demangle argument. 423 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_Xlinker__no_demangle)); 424 425 // Add the remaining values as Xlinker arguments. 426 for (StringRef Val : A->getValues()) 427 if (Val != "--no-demangle") 428 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_Xlinker), Val); 429 430 continue; 431 } 432 433 // Rewrite preprocessor options, to replace -Wp,-MD,FOO which is used by 434 // some build systems. We don't try to be complete here because we don't 435 // care to encourage this usage model. 436 if (A->getOption().matches(options::OPT_Wp_COMMA) && 437 (A->getValue(0) == StringRef("-MD") || 438 A->getValue(0) == StringRef("-MMD"))) { 439 // Rewrite to -MD/-MMD along with -MF. 440 if (A->getValue(0) == StringRef("-MD")) 441 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MD)); 442 else 443 DAL->AddFlagArg(A, Opts.getOption(options::OPT_MMD)); 444 if (A->getNumValues() == 2) 445 DAL->AddSeparateArg(A, Opts.getOption(options::OPT_MF), A->getValue(1)); 446 continue; 447 } 448 449 // Rewrite reserved library names. 450 if (A->getOption().matches(options::OPT_l)) { 451 StringRef Value = A->getValue(); 452 453 // Rewrite unless -nostdlib is present. 454 if (!HasNostdlib && !HasNodefaultlib && !HasNostdlibxx && 455 Value == "stdc++") { 456 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_stdcxx)); 457 continue; 458 } 459 460 // Rewrite unconditionally. 461 if (Value == "cc_kext") { 462 DAL->AddFlagArg(A, Opts.getOption(options::OPT_Z_reserved_lib_cckext)); 463 continue; 464 } 465 } 466 467 // Pick up inputs via the -- option. 468 if (A->getOption().matches(options::OPT__DASH_DASH)) { 469 A->claim(); 470 for (StringRef Val : A->getValues()) 471 DAL->append(MakeInputArg(*DAL, Opts, Val, false)); 472 continue; 473 } 474 475 DAL->append(A); 476 } 477 478 // Enforce -static if -miamcu is present. 479 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) 480 DAL->AddFlagArg(nullptr, Opts.getOption(options::OPT_static)); 481 482 // Add a default value of -mlinker-version=, if one was given and the user 483 // didn't specify one. 484 #if defined(HOST_LINK_VERSION) 485 if (!Args.hasArg(options::OPT_mlinker_version_EQ) && 486 strlen(HOST_LINK_VERSION) > 0) { 487 DAL->AddJoinedArg(0, Opts.getOption(options::OPT_mlinker_version_EQ), 488 HOST_LINK_VERSION); 489 DAL->getLastArg(options::OPT_mlinker_version_EQ)->claim(); 490 } 491 #endif 492 493 return DAL; 494 } 495 496 /// Compute target triple from args. 497 /// 498 /// This routine provides the logic to compute a target triple from various 499 /// args passed to the driver and the default triple string. 500 static llvm::Triple computeTargetTriple(const Driver &D, 501 StringRef TargetTriple, 502 const ArgList &Args, 503 StringRef DarwinArchName = "") { 504 // FIXME: Already done in Compilation *Driver::BuildCompilation 505 if (const Arg *A = Args.getLastArg(options::OPT_target)) 506 TargetTriple = A->getValue(); 507 508 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 509 510 // GNU/Hurd's triples should have been -hurd-gnu*, but were historically made 511 // -gnu* only, and we can not change this, so we have to detect that case as 512 // being the Hurd OS. 513 if (TargetTriple.contains("-unknown-gnu") || TargetTriple.contains("-pc-gnu")) 514 Target.setOSName("hurd"); 515 516 // Handle Apple-specific options available here. 517 if (Target.isOSBinFormatMachO()) { 518 // If an explicit Darwin arch name is given, that trumps all. 519 if (!DarwinArchName.empty()) { 520 tools::darwin::setTripleTypeForMachOArchName(Target, DarwinArchName); 521 return Target; 522 } 523 524 // Handle the Darwin '-arch' flag. 525 if (Arg *A = Args.getLastArg(options::OPT_arch)) { 526 StringRef ArchName = A->getValue(); 527 tools::darwin::setTripleTypeForMachOArchName(Target, ArchName); 528 } 529 } 530 531 // Handle pseudo-target flags '-mlittle-endian'/'-EL' and 532 // '-mbig-endian'/'-EB'. 533 if (Arg *A = Args.getLastArg(options::OPT_mlittle_endian, 534 options::OPT_mbig_endian)) { 535 if (A->getOption().matches(options::OPT_mlittle_endian)) { 536 llvm::Triple LE = Target.getLittleEndianArchVariant(); 537 if (LE.getArch() != llvm::Triple::UnknownArch) 538 Target = std::move(LE); 539 } else { 540 llvm::Triple BE = Target.getBigEndianArchVariant(); 541 if (BE.getArch() != llvm::Triple::UnknownArch) 542 Target = std::move(BE); 543 } 544 } 545 546 // Skip further flag support on OSes which don't support '-m32' or '-m64'. 547 if (Target.getArch() == llvm::Triple::tce || 548 Target.getOS() == llvm::Triple::Minix) 549 return Target; 550 551 // On AIX, the env OBJECT_MODE may affect the resulting arch variant. 552 if (Target.isOSAIX()) { 553 if (Optional<std::string> ObjectModeValue = 554 llvm::sys::Process::GetEnv("OBJECT_MODE")) { 555 StringRef ObjectMode = *ObjectModeValue; 556 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 557 558 if (ObjectMode.equals("64")) { 559 AT = Target.get64BitArchVariant().getArch(); 560 } else if (ObjectMode.equals("32")) { 561 AT = Target.get32BitArchVariant().getArch(); 562 } else { 563 D.Diag(diag::err_drv_invalid_object_mode) << ObjectMode; 564 } 565 566 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) 567 Target.setArch(AT); 568 } 569 } 570 571 // Handle pseudo-target flags '-m64', '-mx32', '-m32' and '-m16'. 572 Arg *A = Args.getLastArg(options::OPT_m64, options::OPT_mx32, 573 options::OPT_m32, options::OPT_m16); 574 if (A) { 575 llvm::Triple::ArchType AT = llvm::Triple::UnknownArch; 576 577 if (A->getOption().matches(options::OPT_m64)) { 578 AT = Target.get64BitArchVariant().getArch(); 579 if (Target.getEnvironment() == llvm::Triple::GNUX32) 580 Target.setEnvironment(llvm::Triple::GNU); 581 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 582 Target.setEnvironment(llvm::Triple::Musl); 583 } else if (A->getOption().matches(options::OPT_mx32) && 584 Target.get64BitArchVariant().getArch() == llvm::Triple::x86_64) { 585 AT = llvm::Triple::x86_64; 586 if (Target.getEnvironment() == llvm::Triple::Musl) 587 Target.setEnvironment(llvm::Triple::MuslX32); 588 else 589 Target.setEnvironment(llvm::Triple::GNUX32); 590 } else if (A->getOption().matches(options::OPT_m32)) { 591 AT = Target.get32BitArchVariant().getArch(); 592 if (Target.getEnvironment() == llvm::Triple::GNUX32) 593 Target.setEnvironment(llvm::Triple::GNU); 594 else if (Target.getEnvironment() == llvm::Triple::MuslX32) 595 Target.setEnvironment(llvm::Triple::Musl); 596 } else if (A->getOption().matches(options::OPT_m16) && 597 Target.get32BitArchVariant().getArch() == llvm::Triple::x86) { 598 AT = llvm::Triple::x86; 599 Target.setEnvironment(llvm::Triple::CODE16); 600 } 601 602 if (AT != llvm::Triple::UnknownArch && AT != Target.getArch()) { 603 Target.setArch(AT); 604 if (Target.isWindowsGNUEnvironment()) 605 toolchains::MinGW::fixTripleArch(D, Target, Args); 606 } 607 } 608 609 // Handle -miamcu flag. 610 if (Args.hasFlag(options::OPT_miamcu, options::OPT_mno_iamcu, false)) { 611 if (Target.get32BitArchVariant().getArch() != llvm::Triple::x86) 612 D.Diag(diag::err_drv_unsupported_opt_for_target) << "-miamcu" 613 << Target.str(); 614 615 if (A && !A->getOption().matches(options::OPT_m32)) 616 D.Diag(diag::err_drv_argument_not_allowed_with) 617 << "-miamcu" << A->getBaseArg().getAsString(Args); 618 619 Target.setArch(llvm::Triple::x86); 620 Target.setArchName("i586"); 621 Target.setEnvironment(llvm::Triple::UnknownEnvironment); 622 Target.setEnvironmentName(""); 623 Target.setOS(llvm::Triple::ELFIAMCU); 624 Target.setVendor(llvm::Triple::UnknownVendor); 625 Target.setVendorName("intel"); 626 } 627 628 // If target is MIPS adjust the target triple 629 // accordingly to provided ABI name. 630 A = Args.getLastArg(options::OPT_mabi_EQ); 631 if (A && Target.isMIPS()) { 632 StringRef ABIName = A->getValue(); 633 if (ABIName == "32") { 634 Target = Target.get32BitArchVariant(); 635 if (Target.getEnvironment() == llvm::Triple::GNUABI64 || 636 Target.getEnvironment() == llvm::Triple::GNUABIN32) 637 Target.setEnvironment(llvm::Triple::GNU); 638 } else if (ABIName == "n32") { 639 Target = Target.get64BitArchVariant(); 640 if (Target.getEnvironment() == llvm::Triple::GNU || 641 Target.getEnvironment() == llvm::Triple::GNUABI64) 642 Target.setEnvironment(llvm::Triple::GNUABIN32); 643 } else if (ABIName == "64") { 644 Target = Target.get64BitArchVariant(); 645 if (Target.getEnvironment() == llvm::Triple::GNU || 646 Target.getEnvironment() == llvm::Triple::GNUABIN32) 647 Target.setEnvironment(llvm::Triple::GNUABI64); 648 } 649 } 650 651 // If target is RISC-V adjust the target triple according to 652 // provided architecture name 653 A = Args.getLastArg(options::OPT_march_EQ); 654 if (A && Target.isRISCV()) { 655 StringRef ArchName = A->getValue(); 656 if (ArchName.startswith_insensitive("rv32")) 657 Target.setArch(llvm::Triple::riscv32); 658 else if (ArchName.startswith_insensitive("rv64")) 659 Target.setArch(llvm::Triple::riscv64); 660 } 661 662 return Target; 663 } 664 665 // Parse the LTO options and record the type of LTO compilation 666 // based on which -f(no-)?lto(=.*)? or -f(no-)?offload-lto(=.*)? 667 // option occurs last. 668 static driver::LTOKind parseLTOMode(Driver &D, const llvm::opt::ArgList &Args, 669 OptSpecifier OptEq, OptSpecifier OptNeg) { 670 if (!Args.hasFlag(OptEq, OptNeg, false)) 671 return LTOK_None; 672 673 const Arg *A = Args.getLastArg(OptEq); 674 StringRef LTOName = A->getValue(); 675 676 driver::LTOKind LTOMode = llvm::StringSwitch<LTOKind>(LTOName) 677 .Case("full", LTOK_Full) 678 .Case("thin", LTOK_Thin) 679 .Default(LTOK_Unknown); 680 681 if (LTOMode == LTOK_Unknown) { 682 D.Diag(diag::err_drv_unsupported_option_argument) 683 << A->getOption().getName() << A->getValue(); 684 return LTOK_None; 685 } 686 return LTOMode; 687 } 688 689 // Parse the LTO options. 690 void Driver::setLTOMode(const llvm::opt::ArgList &Args) { 691 LTOMode = 692 parseLTOMode(*this, Args, options::OPT_flto_EQ, options::OPT_fno_lto); 693 694 OffloadLTOMode = parseLTOMode(*this, Args, options::OPT_foffload_lto_EQ, 695 options::OPT_fno_offload_lto); 696 } 697 698 /// Compute the desired OpenMP runtime from the flags provided. 699 Driver::OpenMPRuntimeKind Driver::getOpenMPRuntime(const ArgList &Args) const { 700 StringRef RuntimeName(CLANG_DEFAULT_OPENMP_RUNTIME); 701 702 const Arg *A = Args.getLastArg(options::OPT_fopenmp_EQ); 703 if (A) 704 RuntimeName = A->getValue(); 705 706 auto RT = llvm::StringSwitch<OpenMPRuntimeKind>(RuntimeName) 707 .Case("libomp", OMPRT_OMP) 708 .Case("libgomp", OMPRT_GOMP) 709 .Case("libiomp5", OMPRT_IOMP5) 710 .Default(OMPRT_Unknown); 711 712 if (RT == OMPRT_Unknown) { 713 if (A) 714 Diag(diag::err_drv_unsupported_option_argument) 715 << A->getOption().getName() << A->getValue(); 716 else 717 // FIXME: We could use a nicer diagnostic here. 718 Diag(diag::err_drv_unsupported_opt) << "-fopenmp"; 719 } 720 721 return RT; 722 } 723 724 void Driver::CreateOffloadingDeviceToolChains(Compilation &C, 725 InputList &Inputs) { 726 727 // 728 // CUDA/HIP 729 // 730 // We need to generate a CUDA/HIP toolchain if any of the inputs has a CUDA 731 // or HIP type. However, mixed CUDA/HIP compilation is not supported. 732 bool IsCuda = 733 llvm::any_of(Inputs, [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 734 return types::isCuda(I.first); 735 }); 736 bool IsHIP = 737 llvm::any_of(Inputs, 738 [](std::pair<types::ID, const llvm::opt::Arg *> &I) { 739 return types::isHIP(I.first); 740 }) || 741 C.getInputArgs().hasArg(options::OPT_hip_link); 742 if (IsCuda && IsHIP) { 743 Diag(clang::diag::err_drv_mix_cuda_hip); 744 return; 745 } 746 if (IsCuda) { 747 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 748 const llvm::Triple &HostTriple = HostTC->getTriple(); 749 auto OFK = Action::OFK_Cuda; 750 auto CudaTriple = 751 getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), HostTriple); 752 if (!CudaTriple) 753 return; 754 // Use the CUDA and host triples as the key into the ToolChains map, 755 // because the device toolchain we create depends on both. 756 auto &CudaTC = ToolChains[CudaTriple->str() + "/" + HostTriple.str()]; 757 if (!CudaTC) { 758 CudaTC = std::make_unique<toolchains::CudaToolChain>( 759 *this, *CudaTriple, *HostTC, C.getInputArgs(), OFK); 760 } 761 C.addOffloadDeviceToolChain(CudaTC.get(), OFK); 762 } else if (IsHIP) { 763 if (auto *OMPTargetArg = 764 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 765 Diag(clang::diag::err_drv_unsupported_opt_for_language_mode) 766 << OMPTargetArg->getSpelling() << "HIP"; 767 return; 768 } 769 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 770 auto OFK = Action::OFK_HIP; 771 auto HIPTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 772 if (!HIPTriple) 773 return; 774 auto *HIPTC = &getOffloadingDeviceToolChain(C.getInputArgs(), *HIPTriple, 775 *HostTC, OFK); 776 assert(HIPTC && "Could not create offloading device tool chain."); 777 C.addOffloadDeviceToolChain(HIPTC, OFK); 778 } 779 780 // 781 // OpenMP 782 // 783 // We need to generate an OpenMP toolchain if the user specified targets with 784 // the -fopenmp-targets option or used --offload-arch with OpenMP enabled. 785 bool IsOpenMPOffloading = 786 C.getInputArgs().hasFlag(options::OPT_fopenmp, options::OPT_fopenmp_EQ, 787 options::OPT_fno_openmp, false) && 788 (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ) || 789 C.getInputArgs().hasArg(options::OPT_offload_arch_EQ)); 790 if (IsOpenMPOffloading) { 791 // We expect that -fopenmp-targets is always used in conjunction with the 792 // option -fopenmp specifying a valid runtime with offloading support, i.e. 793 // libomp or libiomp. 794 OpenMPRuntimeKind RuntimeKind = getOpenMPRuntime(C.getInputArgs()); 795 if (RuntimeKind != OMPRT_OMP && RuntimeKind != OMPRT_IOMP5) { 796 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 797 return; 798 } 799 800 llvm::StringMap<llvm::DenseSet<StringRef>> DerivedArchs; 801 llvm::StringMap<StringRef> FoundNormalizedTriples; 802 llvm::SmallVector<StringRef, 4> OpenMPTriples; 803 804 // If the user specified -fopenmp-targets= we create a toolchain for each 805 // valid triple. Otherwise, if only --offload-arch= was specified we instead 806 // attempt to derive the appropriate toolchains from the arguments. 807 if (Arg *OpenMPTargets = 808 C.getInputArgs().getLastArg(options::OPT_fopenmp_targets_EQ)) { 809 if (OpenMPTargets && !OpenMPTargets->getNumValues()) { 810 Diag(clang::diag::warn_drv_empty_joined_argument) 811 << OpenMPTargets->getAsString(C.getInputArgs()); 812 return; 813 } 814 llvm::copy(OpenMPTargets->getValues(), std::back_inserter(OpenMPTriples)); 815 } else if (C.getInputArgs().hasArg(options::OPT_offload_arch_EQ) && 816 !IsHIP && !IsCuda) { 817 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 818 auto AMDTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs()); 819 auto NVPTXTriple = getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(), 820 HostTC->getTriple()); 821 822 // Attempt to deduce the offloading triple from the set of architectures. 823 // We can only correctly deduce NVPTX / AMDGPU triples currently. 824 llvm::DenseSet<StringRef> Archs = 825 getOffloadArchs(C, C.getArgs(), Action::OFK_OpenMP, nullptr); 826 for (StringRef Arch : Archs) { 827 if (NVPTXTriple && IsNVIDIAGpuArch(StringToCudaArch( 828 getProcessorFromTargetID(*NVPTXTriple, Arch)))) { 829 DerivedArchs[NVPTXTriple->getTriple()].insert(Arch); 830 } else if (AMDTriple && 831 IsAMDGpuArch(StringToCudaArch( 832 getProcessorFromTargetID(*AMDTriple, Arch)))) { 833 DerivedArchs[AMDTriple->getTriple()].insert(Arch); 834 } else { 835 Diag(clang::diag::err_drv_failed_to_deduce_target_from_arch) << Arch; 836 return; 837 } 838 } 839 840 for (const auto &TripleAndArchs : DerivedArchs) 841 OpenMPTriples.push_back(TripleAndArchs.first()); 842 } 843 844 for (StringRef Val : OpenMPTriples) { 845 llvm::Triple TT(ToolChain::getOpenMPTriple(Val)); 846 std::string NormalizedName = TT.normalize(); 847 848 // Make sure we don't have a duplicate triple. 849 auto Duplicate = FoundNormalizedTriples.find(NormalizedName); 850 if (Duplicate != FoundNormalizedTriples.end()) { 851 Diag(clang::diag::warn_drv_omp_offload_target_duplicate) 852 << Val << Duplicate->second; 853 continue; 854 } 855 856 // Store the current triple so that we can check for duplicates in the 857 // following iterations. 858 FoundNormalizedTriples[NormalizedName] = Val; 859 860 // If the specified target is invalid, emit a diagnostic. 861 if (TT.getArch() == llvm::Triple::UnknownArch) 862 Diag(clang::diag::err_drv_invalid_omp_target) << Val; 863 else { 864 const ToolChain *TC; 865 // Device toolchains have to be selected differently. They pair host 866 // and device in their implementation. 867 if (TT.isNVPTX() || TT.isAMDGCN()) { 868 const ToolChain *HostTC = 869 C.getSingleOffloadToolChain<Action::OFK_Host>(); 870 assert(HostTC && "Host toolchain should be always defined."); 871 auto &DeviceTC = 872 ToolChains[TT.str() + "/" + HostTC->getTriple().normalize()]; 873 if (!DeviceTC) { 874 if (TT.isNVPTX()) 875 DeviceTC = std::make_unique<toolchains::CudaToolChain>( 876 *this, TT, *HostTC, C.getInputArgs(), Action::OFK_OpenMP); 877 else if (TT.isAMDGCN()) 878 DeviceTC = std::make_unique<toolchains::AMDGPUOpenMPToolChain>( 879 *this, TT, *HostTC, C.getInputArgs()); 880 else 881 assert(DeviceTC && "Device toolchain not defined."); 882 } 883 884 TC = DeviceTC.get(); 885 } else 886 TC = &getToolChain(C.getInputArgs(), TT); 887 C.addOffloadDeviceToolChain(TC, Action::OFK_OpenMP); 888 if (DerivedArchs.find(TT.getTriple()) != DerivedArchs.end()) 889 KnownArchs[TC] = DerivedArchs[TT.getTriple()]; 890 } 891 } 892 } else if (C.getInputArgs().hasArg(options::OPT_fopenmp_targets_EQ)) { 893 Diag(clang::diag::err_drv_expecting_fopenmp_with_fopenmp_targets); 894 return; 895 } 896 897 // 898 // TODO: Add support for other offloading programming models here. 899 // 900 } 901 902 /// Looks the given directories for the specified file. 903 /// 904 /// \param[out] FilePath File path, if the file was found. 905 /// \param[in] Dirs Directories used for the search. 906 /// \param[in] FileName Name of the file to search for. 907 /// \return True if file was found. 908 /// 909 /// Looks for file specified by FileName sequentially in directories specified 910 /// by Dirs. 911 /// 912 static bool searchForFile(SmallVectorImpl<char> &FilePath, 913 ArrayRef<StringRef> Dirs, StringRef FileName) { 914 SmallString<128> WPath; 915 for (const StringRef &Dir : Dirs) { 916 if (Dir.empty()) 917 continue; 918 WPath.clear(); 919 llvm::sys::path::append(WPath, Dir, FileName); 920 llvm::sys::path::native(WPath); 921 if (llvm::sys::fs::is_regular_file(WPath)) { 922 FilePath = std::move(WPath); 923 return true; 924 } 925 } 926 return false; 927 } 928 929 bool Driver::readConfigFile(StringRef FileName) { 930 // Try reading the given file. 931 SmallVector<const char *, 32> NewCfgArgs; 932 if (!llvm::cl::readConfigFile(FileName, Saver, NewCfgArgs)) { 933 Diag(diag::err_drv_cannot_read_config_file) << FileName; 934 return true; 935 } 936 937 // Read options from config file. 938 llvm::SmallString<128> CfgFileName(FileName); 939 llvm::sys::path::native(CfgFileName); 940 ConfigFile = std::string(CfgFileName); 941 bool ContainErrors; 942 CfgOptions = std::make_unique<InputArgList>( 943 ParseArgStrings(NewCfgArgs, IsCLMode(), ContainErrors)); 944 if (ContainErrors) { 945 CfgOptions.reset(); 946 return true; 947 } 948 949 if (CfgOptions->hasArg(options::OPT_config)) { 950 CfgOptions.reset(); 951 Diag(diag::err_drv_nested_config_file); 952 return true; 953 } 954 955 // Claim all arguments that come from a configuration file so that the driver 956 // does not warn on any that is unused. 957 for (Arg *A : *CfgOptions) 958 A->claim(); 959 return false; 960 } 961 962 bool Driver::loadConfigFile() { 963 std::string CfgFileName; 964 bool FileSpecifiedExplicitly = false; 965 966 // Process options that change search path for config files. 967 if (CLOptions) { 968 if (CLOptions->hasArg(options::OPT_config_system_dir_EQ)) { 969 SmallString<128> CfgDir; 970 CfgDir.append( 971 CLOptions->getLastArgValue(options::OPT_config_system_dir_EQ)); 972 if (!CfgDir.empty()) { 973 if (llvm::sys::fs::make_absolute(CfgDir).value() != 0) 974 SystemConfigDir.clear(); 975 else 976 SystemConfigDir = static_cast<std::string>(CfgDir); 977 } 978 } 979 if (CLOptions->hasArg(options::OPT_config_user_dir_EQ)) { 980 SmallString<128> CfgDir; 981 CfgDir.append( 982 CLOptions->getLastArgValue(options::OPT_config_user_dir_EQ)); 983 if (!CfgDir.empty()) { 984 if (llvm::sys::fs::make_absolute(CfgDir).value() != 0) 985 UserConfigDir.clear(); 986 else 987 UserConfigDir = static_cast<std::string>(CfgDir); 988 } 989 } 990 } 991 992 // First try to find config file specified in command line. 993 if (CLOptions) { 994 std::vector<std::string> ConfigFiles = 995 CLOptions->getAllArgValues(options::OPT_config); 996 if (ConfigFiles.size() > 1) { 997 if (!llvm::all_of(ConfigFiles, [ConfigFiles](const std::string &s) { 998 return s == ConfigFiles[0]; 999 })) { 1000 Diag(diag::err_drv_duplicate_config); 1001 return true; 1002 } 1003 } 1004 1005 if (!ConfigFiles.empty()) { 1006 CfgFileName = ConfigFiles.front(); 1007 assert(!CfgFileName.empty()); 1008 1009 // If argument contains directory separator, treat it as a path to 1010 // configuration file. 1011 if (llvm::sys::path::has_parent_path(CfgFileName)) { 1012 SmallString<128> CfgFilePath; 1013 if (llvm::sys::path::is_relative(CfgFileName)) 1014 llvm::sys::fs::current_path(CfgFilePath); 1015 llvm::sys::path::append(CfgFilePath, CfgFileName); 1016 if (!llvm::sys::fs::is_regular_file(CfgFilePath)) { 1017 Diag(diag::err_drv_config_file_not_exist) << CfgFilePath; 1018 return true; 1019 } 1020 return readConfigFile(CfgFilePath); 1021 } 1022 1023 FileSpecifiedExplicitly = true; 1024 } 1025 } 1026 1027 // If config file is not specified explicitly, try to deduce configuration 1028 // from executable name. For instance, an executable 'armv7l-clang' will 1029 // search for config file 'armv7l-clang.cfg'. 1030 if (CfgFileName.empty() && !ClangNameParts.TargetPrefix.empty()) 1031 CfgFileName = ClangNameParts.TargetPrefix + '-' + ClangNameParts.ModeSuffix; 1032 1033 if (CfgFileName.empty()) 1034 return false; 1035 1036 // Determine architecture part of the file name, if it is present. 1037 StringRef CfgFileArch = CfgFileName; 1038 size_t ArchPrefixLen = CfgFileArch.find('-'); 1039 if (ArchPrefixLen == StringRef::npos) 1040 ArchPrefixLen = CfgFileArch.size(); 1041 llvm::Triple CfgTriple; 1042 CfgFileArch = CfgFileArch.take_front(ArchPrefixLen); 1043 CfgTriple = llvm::Triple(llvm::Triple::normalize(CfgFileArch)); 1044 if (CfgTriple.getArch() == llvm::Triple::ArchType::UnknownArch) 1045 ArchPrefixLen = 0; 1046 1047 if (!StringRef(CfgFileName).endswith(".cfg")) 1048 CfgFileName += ".cfg"; 1049 1050 // If config file starts with architecture name and command line options 1051 // redefine architecture (with options like -m32 -LE etc), try finding new 1052 // config file with that architecture. 1053 SmallString<128> FixedConfigFile; 1054 size_t FixedArchPrefixLen = 0; 1055 if (ArchPrefixLen) { 1056 // Get architecture name from config file name like 'i386.cfg' or 1057 // 'armv7l-clang.cfg'. 1058 // Check if command line options changes effective triple. 1059 llvm::Triple EffectiveTriple = computeTargetTriple(*this, 1060 CfgTriple.getTriple(), *CLOptions); 1061 if (CfgTriple.getArch() != EffectiveTriple.getArch()) { 1062 FixedConfigFile = EffectiveTriple.getArchName(); 1063 FixedArchPrefixLen = FixedConfigFile.size(); 1064 // Append the rest of original file name so that file name transforms 1065 // like: i386-clang.cfg -> x86_64-clang.cfg. 1066 if (ArchPrefixLen < CfgFileName.size()) 1067 FixedConfigFile += CfgFileName.substr(ArchPrefixLen); 1068 } 1069 } 1070 1071 // Prepare list of directories where config file is searched for. 1072 StringRef CfgFileSearchDirs[] = {UserConfigDir, SystemConfigDir, Dir}; 1073 1074 // Try to find config file. First try file with corrected architecture. 1075 llvm::SmallString<128> CfgFilePath; 1076 if (!FixedConfigFile.empty()) { 1077 if (searchForFile(CfgFilePath, CfgFileSearchDirs, FixedConfigFile)) 1078 return readConfigFile(CfgFilePath); 1079 // If 'x86_64-clang.cfg' was not found, try 'x86_64.cfg'. 1080 FixedConfigFile.resize(FixedArchPrefixLen); 1081 FixedConfigFile.append(".cfg"); 1082 if (searchForFile(CfgFilePath, CfgFileSearchDirs, FixedConfigFile)) 1083 return readConfigFile(CfgFilePath); 1084 } 1085 1086 // Then try original file name. 1087 if (searchForFile(CfgFilePath, CfgFileSearchDirs, CfgFileName)) 1088 return readConfigFile(CfgFilePath); 1089 1090 // Finally try removing driver mode part: 'x86_64-clang.cfg' -> 'x86_64.cfg'. 1091 if (!ClangNameParts.ModeSuffix.empty() && 1092 !ClangNameParts.TargetPrefix.empty()) { 1093 CfgFileName.assign(ClangNameParts.TargetPrefix); 1094 CfgFileName.append(".cfg"); 1095 if (searchForFile(CfgFilePath, CfgFileSearchDirs, CfgFileName)) 1096 return readConfigFile(CfgFilePath); 1097 } 1098 1099 // Report error but only if config file was specified explicitly, by option 1100 // --config. If it was deduced from executable name, it is not an error. 1101 if (FileSpecifiedExplicitly) { 1102 Diag(diag::err_drv_config_file_not_found) << CfgFileName; 1103 for (const StringRef &SearchDir : CfgFileSearchDirs) 1104 if (!SearchDir.empty()) 1105 Diag(diag::note_drv_config_file_searched_in) << SearchDir; 1106 return true; 1107 } 1108 1109 return false; 1110 } 1111 1112 Compilation *Driver::BuildCompilation(ArrayRef<const char *> ArgList) { 1113 llvm::PrettyStackTraceString CrashInfo("Compilation construction"); 1114 1115 // FIXME: Handle environment options which affect driver behavior, somewhere 1116 // (client?). GCC_EXEC_PREFIX, LPATH, CC_PRINT_OPTIONS. 1117 1118 // We look for the driver mode option early, because the mode can affect 1119 // how other options are parsed. 1120 1121 auto DriverMode = getDriverMode(ClangExecutable, ArgList.slice(1)); 1122 if (!DriverMode.empty()) 1123 setDriverMode(DriverMode); 1124 1125 // FIXME: What are we going to do with -V and -b? 1126 1127 // Arguments specified in command line. 1128 bool ContainsError; 1129 CLOptions = std::make_unique<InputArgList>( 1130 ParseArgStrings(ArgList.slice(1), IsCLMode(), ContainsError)); 1131 1132 // Try parsing configuration file. 1133 if (!ContainsError) 1134 ContainsError = loadConfigFile(); 1135 bool HasConfigFile = !ContainsError && (CfgOptions.get() != nullptr); 1136 1137 // All arguments, from both config file and command line. 1138 InputArgList Args = std::move(HasConfigFile ? std::move(*CfgOptions) 1139 : std::move(*CLOptions)); 1140 1141 // The args for config files or /clang: flags belong to different InputArgList 1142 // objects than Args. This copies an Arg from one of those other InputArgLists 1143 // to the ownership of Args. 1144 auto appendOneArg = [&Args](const Arg *Opt, const Arg *BaseArg) { 1145 unsigned Index = Args.MakeIndex(Opt->getSpelling()); 1146 Arg *Copy = new llvm::opt::Arg(Opt->getOption(), Args.getArgString(Index), 1147 Index, BaseArg); 1148 Copy->getValues() = Opt->getValues(); 1149 if (Opt->isClaimed()) 1150 Copy->claim(); 1151 Copy->setOwnsValues(Opt->getOwnsValues()); 1152 Opt->setOwnsValues(false); 1153 Args.append(Copy); 1154 }; 1155 1156 if (HasConfigFile) 1157 for (auto *Opt : *CLOptions) { 1158 if (Opt->getOption().matches(options::OPT_config)) 1159 continue; 1160 const Arg *BaseArg = &Opt->getBaseArg(); 1161 if (BaseArg == Opt) 1162 BaseArg = nullptr; 1163 appendOneArg(Opt, BaseArg); 1164 } 1165 1166 // In CL mode, look for any pass-through arguments 1167 if (IsCLMode() && !ContainsError) { 1168 SmallVector<const char *, 16> CLModePassThroughArgList; 1169 for (const auto *A : Args.filtered(options::OPT__SLASH_clang)) { 1170 A->claim(); 1171 CLModePassThroughArgList.push_back(A->getValue()); 1172 } 1173 1174 if (!CLModePassThroughArgList.empty()) { 1175 // Parse any pass through args using default clang processing rather 1176 // than clang-cl processing. 1177 auto CLModePassThroughOptions = std::make_unique<InputArgList>( 1178 ParseArgStrings(CLModePassThroughArgList, false, ContainsError)); 1179 1180 if (!ContainsError) 1181 for (auto *Opt : *CLModePassThroughOptions) { 1182 appendOneArg(Opt, nullptr); 1183 } 1184 } 1185 } 1186 1187 // Check for working directory option before accessing any files 1188 if (Arg *WD = Args.getLastArg(options::OPT_working_directory)) 1189 if (VFS->setCurrentWorkingDirectory(WD->getValue())) 1190 Diag(diag::err_drv_unable_to_set_working_directory) << WD->getValue(); 1191 1192 // FIXME: This stuff needs to go into the Compilation, not the driver. 1193 bool CCCPrintPhases; 1194 1195 // Silence driver warnings if requested 1196 Diags.setIgnoreAllWarnings(Args.hasArg(options::OPT_w)); 1197 1198 // -canonical-prefixes, -no-canonical-prefixes are used very early in main. 1199 Args.ClaimAllArgs(options::OPT_canonical_prefixes); 1200 Args.ClaimAllArgs(options::OPT_no_canonical_prefixes); 1201 1202 // f(no-)integated-cc1 is also used very early in main. 1203 Args.ClaimAllArgs(options::OPT_fintegrated_cc1); 1204 Args.ClaimAllArgs(options::OPT_fno_integrated_cc1); 1205 1206 // Ignore -pipe. 1207 Args.ClaimAllArgs(options::OPT_pipe); 1208 1209 // Extract -ccc args. 1210 // 1211 // FIXME: We need to figure out where this behavior should live. Most of it 1212 // should be outside in the client; the parts that aren't should have proper 1213 // options, either by introducing new ones or by overloading gcc ones like -V 1214 // or -b. 1215 CCCPrintPhases = Args.hasArg(options::OPT_ccc_print_phases); 1216 CCCPrintBindings = Args.hasArg(options::OPT_ccc_print_bindings); 1217 if (const Arg *A = Args.getLastArg(options::OPT_ccc_gcc_name)) 1218 CCCGenericGCCName = A->getValue(); 1219 1220 // Process -fproc-stat-report options. 1221 if (const Arg *A = Args.getLastArg(options::OPT_fproc_stat_report_EQ)) { 1222 CCPrintProcessStats = true; 1223 CCPrintStatReportFilename = A->getValue(); 1224 } 1225 if (Args.hasArg(options::OPT_fproc_stat_report)) 1226 CCPrintProcessStats = true; 1227 1228 // FIXME: TargetTriple is used by the target-prefixed calls to as/ld 1229 // and getToolChain is const. 1230 if (IsCLMode()) { 1231 // clang-cl targets MSVC-style Win32. 1232 llvm::Triple T(TargetTriple); 1233 T.setOS(llvm::Triple::Win32); 1234 T.setVendor(llvm::Triple::PC); 1235 T.setEnvironment(llvm::Triple::MSVC); 1236 T.setObjectFormat(llvm::Triple::COFF); 1237 TargetTriple = T.str(); 1238 } else if (IsDXCMode()) { 1239 // Build TargetTriple from target_profile option for clang-dxc. 1240 if (const Arg *A = Args.getLastArg(options::OPT_target_profile)) { 1241 StringRef TargetProfile = A->getValue(); 1242 if (auto Triple = 1243 toolchains::HLSLToolChain::parseTargetProfile(TargetProfile)) 1244 TargetTriple = Triple.getValue(); 1245 else 1246 Diag(diag::err_drv_invalid_directx_shader_module) << TargetProfile; 1247 1248 A->claim(); 1249 } else { 1250 Diag(diag::err_drv_dxc_missing_target_profile); 1251 } 1252 } 1253 1254 if (const Arg *A = Args.getLastArg(options::OPT_target)) 1255 TargetTriple = A->getValue(); 1256 if (const Arg *A = Args.getLastArg(options::OPT_ccc_install_dir)) 1257 Dir = InstalledDir = A->getValue(); 1258 for (const Arg *A : Args.filtered(options::OPT_B)) { 1259 A->claim(); 1260 PrefixDirs.push_back(A->getValue(0)); 1261 } 1262 if (Optional<std::string> CompilerPathValue = 1263 llvm::sys::Process::GetEnv("COMPILER_PATH")) { 1264 StringRef CompilerPath = *CompilerPathValue; 1265 while (!CompilerPath.empty()) { 1266 std::pair<StringRef, StringRef> Split = 1267 CompilerPath.split(llvm::sys::EnvPathSeparator); 1268 PrefixDirs.push_back(std::string(Split.first)); 1269 CompilerPath = Split.second; 1270 } 1271 } 1272 if (const Arg *A = Args.getLastArg(options::OPT__sysroot_EQ)) 1273 SysRoot = A->getValue(); 1274 if (const Arg *A = Args.getLastArg(options::OPT__dyld_prefix_EQ)) 1275 DyldPrefix = A->getValue(); 1276 1277 if (const Arg *A = Args.getLastArg(options::OPT_resource_dir)) 1278 ResourceDir = A->getValue(); 1279 1280 if (const Arg *A = Args.getLastArg(options::OPT_save_temps_EQ)) { 1281 SaveTemps = llvm::StringSwitch<SaveTempsMode>(A->getValue()) 1282 .Case("cwd", SaveTempsCwd) 1283 .Case("obj", SaveTempsObj) 1284 .Default(SaveTempsCwd); 1285 } 1286 1287 setLTOMode(Args); 1288 1289 // Process -fembed-bitcode= flags. 1290 if (Arg *A = Args.getLastArg(options::OPT_fembed_bitcode_EQ)) { 1291 StringRef Name = A->getValue(); 1292 unsigned Model = llvm::StringSwitch<unsigned>(Name) 1293 .Case("off", EmbedNone) 1294 .Case("all", EmbedBitcode) 1295 .Case("bitcode", EmbedBitcode) 1296 .Case("marker", EmbedMarker) 1297 .Default(~0U); 1298 if (Model == ~0U) { 1299 Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) 1300 << Name; 1301 } else 1302 BitcodeEmbed = static_cast<BitcodeEmbedMode>(Model); 1303 } 1304 1305 // Setting up the jobs for some precompile cases depends on whether we are 1306 // treating them as PCH, implicit modules or C++20 ones. 1307 // TODO: inferring the mode like this seems fragile (it meets the objective 1308 // of not requiring anything new for operation, however). 1309 const Arg *Std = Args.getLastArg(options::OPT_std_EQ); 1310 ModulesModeCXX20 = 1311 !Args.hasArg(options::OPT_fmodules) && Std && 1312 (Std->containsValue("c++20") || Std->containsValue("c++2b") || 1313 Std->containsValue("c++2a") || Std->containsValue("c++latest")); 1314 1315 // Process -fmodule-header{=} flags. 1316 if (Arg *A = Args.getLastArg(options::OPT_fmodule_header_EQ, 1317 options::OPT_fmodule_header)) { 1318 // These flags force C++20 handling of headers. 1319 ModulesModeCXX20 = true; 1320 if (A->getOption().matches(options::OPT_fmodule_header)) 1321 CXX20HeaderType = HeaderMode_Default; 1322 else { 1323 StringRef ArgName = A->getValue(); 1324 unsigned Kind = llvm::StringSwitch<unsigned>(ArgName) 1325 .Case("user", HeaderMode_User) 1326 .Case("system", HeaderMode_System) 1327 .Default(~0U); 1328 if (Kind == ~0U) { 1329 Diags.Report(diag::err_drv_invalid_value) 1330 << A->getAsString(Args) << ArgName; 1331 } else 1332 CXX20HeaderType = static_cast<ModuleHeaderMode>(Kind); 1333 } 1334 } 1335 1336 std::unique_ptr<llvm::opt::InputArgList> UArgs = 1337 std::make_unique<InputArgList>(std::move(Args)); 1338 1339 // Perform the default argument translations. 1340 DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs); 1341 1342 // Owned by the host. 1343 const ToolChain &TC = getToolChain( 1344 *UArgs, computeTargetTriple(*this, TargetTriple, *UArgs)); 1345 1346 // The compilation takes ownership of Args. 1347 Compilation *C = new Compilation(*this, TC, UArgs.release(), TranslatedArgs, 1348 ContainsError); 1349 1350 if (!HandleImmediateArgs(*C)) 1351 return C; 1352 1353 // Construct the list of inputs. 1354 InputList Inputs; 1355 BuildInputs(C->getDefaultToolChain(), *TranslatedArgs, Inputs); 1356 1357 // Populate the tool chains for the offloading devices, if any. 1358 CreateOffloadingDeviceToolChains(*C, Inputs); 1359 1360 // Construct the list of abstract actions to perform for this compilation. On 1361 // MachO targets this uses the driver-driver and universal actions. 1362 if (TC.getTriple().isOSBinFormatMachO()) 1363 BuildUniversalActions(*C, C->getDefaultToolChain(), Inputs); 1364 else 1365 BuildActions(*C, C->getArgs(), Inputs, C->getActions()); 1366 1367 if (CCCPrintPhases) { 1368 PrintActions(*C); 1369 return C; 1370 } 1371 1372 BuildJobs(*C); 1373 1374 return C; 1375 } 1376 1377 static void printArgList(raw_ostream &OS, const llvm::opt::ArgList &Args) { 1378 llvm::opt::ArgStringList ASL; 1379 for (const auto *A : Args) { 1380 // Use user's original spelling of flags. For example, use 1381 // `/source-charset:utf-8` instead of `-finput-charset=utf-8` if the user 1382 // wrote the former. 1383 while (A->getAlias()) 1384 A = A->getAlias(); 1385 A->render(Args, ASL); 1386 } 1387 1388 for (auto I = ASL.begin(), E = ASL.end(); I != E; ++I) { 1389 if (I != ASL.begin()) 1390 OS << ' '; 1391 llvm::sys::printArg(OS, *I, true); 1392 } 1393 OS << '\n'; 1394 } 1395 1396 bool Driver::getCrashDiagnosticFile(StringRef ReproCrashFilename, 1397 SmallString<128> &CrashDiagDir) { 1398 using namespace llvm::sys; 1399 assert(llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin() && 1400 "Only knows about .crash files on Darwin"); 1401 1402 // The .crash file can be found on at ~/Library/Logs/DiagnosticReports/ 1403 // (or /Library/Logs/DiagnosticReports for root) and has the filename pattern 1404 // clang-<VERSION>_<YYYY-MM-DD-HHMMSS>_<hostname>.crash. 1405 path::home_directory(CrashDiagDir); 1406 if (CrashDiagDir.startswith("/var/root")) 1407 CrashDiagDir = "/"; 1408 path::append(CrashDiagDir, "Library/Logs/DiagnosticReports"); 1409 int PID = 1410 #if LLVM_ON_UNIX 1411 getpid(); 1412 #else 1413 0; 1414 #endif 1415 std::error_code EC; 1416 fs::file_status FileStatus; 1417 TimePoint<> LastAccessTime; 1418 SmallString<128> CrashFilePath; 1419 // Lookup the .crash files and get the one generated by a subprocess spawned 1420 // by this driver invocation. 1421 for (fs::directory_iterator File(CrashDiagDir, EC), FileEnd; 1422 File != FileEnd && !EC; File.increment(EC)) { 1423 StringRef FileName = path::filename(File->path()); 1424 if (!FileName.startswith(Name)) 1425 continue; 1426 if (fs::status(File->path(), FileStatus)) 1427 continue; 1428 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CrashFile = 1429 llvm::MemoryBuffer::getFile(File->path()); 1430 if (!CrashFile) 1431 continue; 1432 // The first line should start with "Process:", otherwise this isn't a real 1433 // .crash file. 1434 StringRef Data = CrashFile.get()->getBuffer(); 1435 if (!Data.startswith("Process:")) 1436 continue; 1437 // Parse parent process pid line, e.g: "Parent Process: clang-4.0 [79141]" 1438 size_t ParentProcPos = Data.find("Parent Process:"); 1439 if (ParentProcPos == StringRef::npos) 1440 continue; 1441 size_t LineEnd = Data.find_first_of("\n", ParentProcPos); 1442 if (LineEnd == StringRef::npos) 1443 continue; 1444 StringRef ParentProcess = Data.slice(ParentProcPos+15, LineEnd).trim(); 1445 int OpenBracket = -1, CloseBracket = -1; 1446 for (size_t i = 0, e = ParentProcess.size(); i < e; ++i) { 1447 if (ParentProcess[i] == '[') 1448 OpenBracket = i; 1449 if (ParentProcess[i] == ']') 1450 CloseBracket = i; 1451 } 1452 // Extract the parent process PID from the .crash file and check whether 1453 // it matches this driver invocation pid. 1454 int CrashPID; 1455 if (OpenBracket < 0 || CloseBracket < 0 || 1456 ParentProcess.slice(OpenBracket + 1, CloseBracket) 1457 .getAsInteger(10, CrashPID) || CrashPID != PID) { 1458 continue; 1459 } 1460 1461 // Found a .crash file matching the driver pid. To avoid getting an older 1462 // and misleading crash file, continue looking for the most recent. 1463 // FIXME: the driver can dispatch multiple cc1 invocations, leading to 1464 // multiple crashes poiting to the same parent process. Since the driver 1465 // does not collect pid information for the dispatched invocation there's 1466 // currently no way to distinguish among them. 1467 const auto FileAccessTime = FileStatus.getLastModificationTime(); 1468 if (FileAccessTime > LastAccessTime) { 1469 CrashFilePath.assign(File->path()); 1470 LastAccessTime = FileAccessTime; 1471 } 1472 } 1473 1474 // If found, copy it over to the location of other reproducer files. 1475 if (!CrashFilePath.empty()) { 1476 EC = fs::copy_file(CrashFilePath, ReproCrashFilename); 1477 if (EC) 1478 return false; 1479 return true; 1480 } 1481 1482 return false; 1483 } 1484 1485 // When clang crashes, produce diagnostic information including the fully 1486 // preprocessed source file(s). Request that the developer attach the 1487 // diagnostic information to a bug report. 1488 void Driver::generateCompilationDiagnostics( 1489 Compilation &C, const Command &FailingCommand, 1490 StringRef AdditionalInformation, CompilationDiagnosticReport *Report) { 1491 if (C.getArgs().hasArg(options::OPT_fno_crash_diagnostics)) 1492 return; 1493 1494 // Don't try to generate diagnostics for link or dsymutil jobs. 1495 if (FailingCommand.getCreator().isLinkJob() || 1496 FailingCommand.getCreator().isDsymutilJob()) 1497 return; 1498 1499 // Print the version of the compiler. 1500 PrintVersion(C, llvm::errs()); 1501 1502 // Suppress driver output and emit preprocessor output to temp file. 1503 CCGenDiagnostics = true; 1504 1505 // Save the original job command(s). 1506 Command Cmd = FailingCommand; 1507 1508 // Keep track of whether we produce any errors while trying to produce 1509 // preprocessed sources. 1510 DiagnosticErrorTrap Trap(Diags); 1511 1512 // Suppress tool output. 1513 C.initCompilationForDiagnostics(); 1514 1515 // Construct the list of inputs. 1516 InputList Inputs; 1517 BuildInputs(C.getDefaultToolChain(), C.getArgs(), Inputs); 1518 1519 for (InputList::iterator it = Inputs.begin(), ie = Inputs.end(); it != ie;) { 1520 bool IgnoreInput = false; 1521 1522 // Ignore input from stdin or any inputs that cannot be preprocessed. 1523 // Check type first as not all linker inputs have a value. 1524 if (types::getPreprocessedType(it->first) == types::TY_INVALID) { 1525 IgnoreInput = true; 1526 } else if (!strcmp(it->second->getValue(), "-")) { 1527 Diag(clang::diag::note_drv_command_failed_diag_msg) 1528 << "Error generating preprocessed source(s) - " 1529 "ignoring input from stdin."; 1530 IgnoreInput = true; 1531 } 1532 1533 if (IgnoreInput) { 1534 it = Inputs.erase(it); 1535 ie = Inputs.end(); 1536 } else { 1537 ++it; 1538 } 1539 } 1540 1541 if (Inputs.empty()) { 1542 Diag(clang::diag::note_drv_command_failed_diag_msg) 1543 << "Error generating preprocessed source(s) - " 1544 "no preprocessable inputs."; 1545 return; 1546 } 1547 1548 // Don't attempt to generate preprocessed files if multiple -arch options are 1549 // used, unless they're all duplicates. 1550 llvm::StringSet<> ArchNames; 1551 for (const Arg *A : C.getArgs()) { 1552 if (A->getOption().matches(options::OPT_arch)) { 1553 StringRef ArchName = A->getValue(); 1554 ArchNames.insert(ArchName); 1555 } 1556 } 1557 if (ArchNames.size() > 1) { 1558 Diag(clang::diag::note_drv_command_failed_diag_msg) 1559 << "Error generating preprocessed source(s) - cannot generate " 1560 "preprocessed source with multiple -arch options."; 1561 return; 1562 } 1563 1564 // Construct the list of abstract actions to perform for this compilation. On 1565 // Darwin OSes this uses the driver-driver and builds universal actions. 1566 const ToolChain &TC = C.getDefaultToolChain(); 1567 if (TC.getTriple().isOSBinFormatMachO()) 1568 BuildUniversalActions(C, TC, Inputs); 1569 else 1570 BuildActions(C, C.getArgs(), Inputs, C.getActions()); 1571 1572 BuildJobs(C); 1573 1574 // If there were errors building the compilation, quit now. 1575 if (Trap.hasErrorOccurred()) { 1576 Diag(clang::diag::note_drv_command_failed_diag_msg) 1577 << "Error generating preprocessed source(s)."; 1578 return; 1579 } 1580 1581 // Generate preprocessed output. 1582 SmallVector<std::pair<int, const Command *>, 4> FailingCommands; 1583 C.ExecuteJobs(C.getJobs(), FailingCommands); 1584 1585 // If any of the preprocessing commands failed, clean up and exit. 1586 if (!FailingCommands.empty()) { 1587 Diag(clang::diag::note_drv_command_failed_diag_msg) 1588 << "Error generating preprocessed source(s)."; 1589 return; 1590 } 1591 1592 const ArgStringList &TempFiles = C.getTempFiles(); 1593 if (TempFiles.empty()) { 1594 Diag(clang::diag::note_drv_command_failed_diag_msg) 1595 << "Error generating preprocessed source(s)."; 1596 return; 1597 } 1598 1599 Diag(clang::diag::note_drv_command_failed_diag_msg) 1600 << "\n********************\n\n" 1601 "PLEASE ATTACH THE FOLLOWING FILES TO THE BUG REPORT:\n" 1602 "Preprocessed source(s) and associated run script(s) are located at:"; 1603 1604 SmallString<128> VFS; 1605 SmallString<128> ReproCrashFilename; 1606 for (const char *TempFile : TempFiles) { 1607 Diag(clang::diag::note_drv_command_failed_diag_msg) << TempFile; 1608 if (Report) 1609 Report->TemporaryFiles.push_back(TempFile); 1610 if (ReproCrashFilename.empty()) { 1611 ReproCrashFilename = TempFile; 1612 llvm::sys::path::replace_extension(ReproCrashFilename, ".crash"); 1613 } 1614 if (StringRef(TempFile).endswith(".cache")) { 1615 // In some cases (modules) we'll dump extra data to help with reproducing 1616 // the crash into a directory next to the output. 1617 VFS = llvm::sys::path::filename(TempFile); 1618 llvm::sys::path::append(VFS, "vfs", "vfs.yaml"); 1619 } 1620 } 1621 1622 // Assume associated files are based off of the first temporary file. 1623 CrashReportInfo CrashInfo(TempFiles[0], VFS); 1624 1625 llvm::SmallString<128> Script(CrashInfo.Filename); 1626 llvm::sys::path::replace_extension(Script, "sh"); 1627 std::error_code EC; 1628 llvm::raw_fd_ostream ScriptOS(Script, EC, llvm::sys::fs::CD_CreateNew, 1629 llvm::sys::fs::FA_Write, 1630 llvm::sys::fs::OF_Text); 1631 if (EC) { 1632 Diag(clang::diag::note_drv_command_failed_diag_msg) 1633 << "Error generating run script: " << Script << " " << EC.message(); 1634 } else { 1635 ScriptOS << "# Crash reproducer for " << getClangFullVersion() << "\n" 1636 << "# Driver args: "; 1637 printArgList(ScriptOS, C.getInputArgs()); 1638 ScriptOS << "# Original command: "; 1639 Cmd.Print(ScriptOS, "\n", /*Quote=*/true); 1640 Cmd.Print(ScriptOS, "\n", /*Quote=*/true, &CrashInfo); 1641 if (!AdditionalInformation.empty()) 1642 ScriptOS << "\n# Additional information: " << AdditionalInformation 1643 << "\n"; 1644 if (Report) 1645 Report->TemporaryFiles.push_back(std::string(Script.str())); 1646 Diag(clang::diag::note_drv_command_failed_diag_msg) << Script; 1647 } 1648 1649 // On darwin, provide information about the .crash diagnostic report. 1650 if (llvm::Triple(llvm::sys::getProcessTriple()).isOSDarwin()) { 1651 SmallString<128> CrashDiagDir; 1652 if (getCrashDiagnosticFile(ReproCrashFilename, CrashDiagDir)) { 1653 Diag(clang::diag::note_drv_command_failed_diag_msg) 1654 << ReproCrashFilename.str(); 1655 } else { // Suggest a directory for the user to look for .crash files. 1656 llvm::sys::path::append(CrashDiagDir, Name); 1657 CrashDiagDir += "_<YYYY-MM-DD-HHMMSS>_<hostname>.crash"; 1658 Diag(clang::diag::note_drv_command_failed_diag_msg) 1659 << "Crash backtrace is located in"; 1660 Diag(clang::diag::note_drv_command_failed_diag_msg) 1661 << CrashDiagDir.str(); 1662 Diag(clang::diag::note_drv_command_failed_diag_msg) 1663 << "(choose the .crash file that corresponds to your crash)"; 1664 } 1665 } 1666 1667 for (const auto &A : C.getArgs().filtered(options::OPT_frewrite_map_file_EQ)) 1668 Diag(clang::diag::note_drv_command_failed_diag_msg) << A->getValue(); 1669 1670 Diag(clang::diag::note_drv_command_failed_diag_msg) 1671 << "\n\n********************"; 1672 } 1673 1674 void Driver::setUpResponseFiles(Compilation &C, Command &Cmd) { 1675 // Since commandLineFitsWithinSystemLimits() may underestimate system's 1676 // capacity if the tool does not support response files, there is a chance/ 1677 // that things will just work without a response file, so we silently just 1678 // skip it. 1679 if (Cmd.getResponseFileSupport().ResponseKind == 1680 ResponseFileSupport::RF_None || 1681 llvm::sys::commandLineFitsWithinSystemLimits(Cmd.getExecutable(), 1682 Cmd.getArguments())) 1683 return; 1684 1685 std::string TmpName = GetTemporaryPath("response", "txt"); 1686 Cmd.setResponseFile(C.addTempFile(C.getArgs().MakeArgString(TmpName))); 1687 } 1688 1689 int Driver::ExecuteCompilation( 1690 Compilation &C, 1691 SmallVectorImpl<std::pair<int, const Command *>> &FailingCommands) { 1692 if (C.getArgs().hasArg(options::OPT_fdriver_only)) { 1693 if (C.getArgs().hasArg(options::OPT_v)) 1694 C.getJobs().Print(llvm::errs(), "\n", true); 1695 1696 C.ExecuteJobs(C.getJobs(), FailingCommands, /*LogOnly=*/true); 1697 1698 // If there were errors building the compilation, quit now. 1699 if (!FailingCommands.empty() || Diags.hasErrorOccurred()) 1700 return 1; 1701 1702 return 0; 1703 } 1704 1705 // Just print if -### was present. 1706 if (C.getArgs().hasArg(options::OPT__HASH_HASH_HASH)) { 1707 C.getJobs().Print(llvm::errs(), "\n", true); 1708 return 0; 1709 } 1710 1711 // If there were errors building the compilation, quit now. 1712 if (Diags.hasErrorOccurred()) 1713 return 1; 1714 1715 // Set up response file names for each command, if necessary. 1716 for (auto &Job : C.getJobs()) 1717 setUpResponseFiles(C, Job); 1718 1719 C.ExecuteJobs(C.getJobs(), FailingCommands); 1720 1721 // If the command succeeded, we are done. 1722 if (FailingCommands.empty()) 1723 return 0; 1724 1725 // Otherwise, remove result files and print extra information about abnormal 1726 // failures. 1727 int Res = 0; 1728 for (const auto &CmdPair : FailingCommands) { 1729 int CommandRes = CmdPair.first; 1730 const Command *FailingCommand = CmdPair.second; 1731 1732 // Remove result files if we're not saving temps. 1733 if (!isSaveTempsEnabled()) { 1734 const JobAction *JA = cast<JobAction>(&FailingCommand->getSource()); 1735 C.CleanupFileMap(C.getResultFiles(), JA, true); 1736 1737 // Failure result files are valid unless we crashed. 1738 if (CommandRes < 0) 1739 C.CleanupFileMap(C.getFailureResultFiles(), JA, true); 1740 } 1741 1742 #if LLVM_ON_UNIX 1743 // llvm/lib/Support/Unix/Signals.inc will exit with a special return code 1744 // for SIGPIPE. Do not print diagnostics for this case. 1745 if (CommandRes == EX_IOERR) { 1746 Res = CommandRes; 1747 continue; 1748 } 1749 #endif 1750 1751 // Print extra information about abnormal failures, if possible. 1752 // 1753 // This is ad-hoc, but we don't want to be excessively noisy. If the result 1754 // status was 1, assume the command failed normally. In particular, if it 1755 // was the compiler then assume it gave a reasonable error code. Failures 1756 // in other tools are less common, and they generally have worse 1757 // diagnostics, so always print the diagnostic there. 1758 const Tool &FailingTool = FailingCommand->getCreator(); 1759 1760 if (!FailingCommand->getCreator().hasGoodDiagnostics() || CommandRes != 1) { 1761 // FIXME: See FIXME above regarding result code interpretation. 1762 if (CommandRes < 0) 1763 Diag(clang::diag::err_drv_command_signalled) 1764 << FailingTool.getShortName(); 1765 else 1766 Diag(clang::diag::err_drv_command_failed) 1767 << FailingTool.getShortName() << CommandRes; 1768 } 1769 } 1770 return Res; 1771 } 1772 1773 void Driver::PrintHelp(bool ShowHidden) const { 1774 unsigned IncludedFlagsBitmask; 1775 unsigned ExcludedFlagsBitmask; 1776 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 1777 getIncludeExcludeOptionFlagMasks(IsCLMode()); 1778 1779 ExcludedFlagsBitmask |= options::NoDriverOption; 1780 if (!ShowHidden) 1781 ExcludedFlagsBitmask |= HelpHidden; 1782 1783 if (IsFlangMode()) 1784 IncludedFlagsBitmask |= options::FlangOption; 1785 else 1786 ExcludedFlagsBitmask |= options::FlangOnlyOption; 1787 1788 std::string Usage = llvm::formatv("{0} [options] file...", Name).str(); 1789 getOpts().printHelp(llvm::outs(), Usage.c_str(), DriverTitle.c_str(), 1790 IncludedFlagsBitmask, ExcludedFlagsBitmask, 1791 /*ShowAllAliases=*/false); 1792 } 1793 1794 void Driver::PrintVersion(const Compilation &C, raw_ostream &OS) const { 1795 if (IsFlangMode()) { 1796 OS << getClangToolFullVersion("flang-new") << '\n'; 1797 } else { 1798 // FIXME: The following handlers should use a callback mechanism, we don't 1799 // know what the client would like to do. 1800 OS << getClangFullVersion() << '\n'; 1801 } 1802 const ToolChain &TC = C.getDefaultToolChain(); 1803 OS << "Target: " << TC.getTripleString() << '\n'; 1804 1805 // Print the threading model. 1806 if (Arg *A = C.getArgs().getLastArg(options::OPT_mthread_model)) { 1807 // Don't print if the ToolChain would have barfed on it already 1808 if (TC.isThreadModelSupported(A->getValue())) 1809 OS << "Thread model: " << A->getValue(); 1810 } else 1811 OS << "Thread model: " << TC.getThreadModel(); 1812 OS << '\n'; 1813 1814 // Print out the install directory. 1815 OS << "InstalledDir: " << InstalledDir << '\n'; 1816 1817 // If configuration file was used, print its path. 1818 if (!ConfigFile.empty()) 1819 OS << "Configuration file: " << ConfigFile << '\n'; 1820 } 1821 1822 /// PrintDiagnosticCategories - Implement the --print-diagnostic-categories 1823 /// option. 1824 static void PrintDiagnosticCategories(raw_ostream &OS) { 1825 // Skip the empty category. 1826 for (unsigned i = 1, max = DiagnosticIDs::getNumberOfCategories(); i != max; 1827 ++i) 1828 OS << i << ',' << DiagnosticIDs::getCategoryNameFromID(i) << '\n'; 1829 } 1830 1831 void Driver::HandleAutocompletions(StringRef PassedFlags) const { 1832 if (PassedFlags == "") 1833 return; 1834 // Print out all options that start with a given argument. This is used for 1835 // shell autocompletion. 1836 std::vector<std::string> SuggestedCompletions; 1837 std::vector<std::string> Flags; 1838 1839 unsigned int DisableFlags = 1840 options::NoDriverOption | options::Unsupported | options::Ignored; 1841 1842 // Make sure that Flang-only options don't pollute the Clang output 1843 // TODO: Make sure that Clang-only options don't pollute Flang output 1844 if (!IsFlangMode()) 1845 DisableFlags |= options::FlangOnlyOption; 1846 1847 // Distinguish "--autocomplete=-someflag" and "--autocomplete=-someflag," 1848 // because the latter indicates that the user put space before pushing tab 1849 // which should end up in a file completion. 1850 const bool HasSpace = PassedFlags.endswith(","); 1851 1852 // Parse PassedFlags by "," as all the command-line flags are passed to this 1853 // function separated by "," 1854 StringRef TargetFlags = PassedFlags; 1855 while (TargetFlags != "") { 1856 StringRef CurFlag; 1857 std::tie(CurFlag, TargetFlags) = TargetFlags.split(","); 1858 Flags.push_back(std::string(CurFlag)); 1859 } 1860 1861 // We want to show cc1-only options only when clang is invoked with -cc1 or 1862 // -Xclang. 1863 if (llvm::is_contained(Flags, "-Xclang") || llvm::is_contained(Flags, "-cc1")) 1864 DisableFlags &= ~options::NoDriverOption; 1865 1866 const llvm::opt::OptTable &Opts = getOpts(); 1867 StringRef Cur; 1868 Cur = Flags.at(Flags.size() - 1); 1869 StringRef Prev; 1870 if (Flags.size() >= 2) { 1871 Prev = Flags.at(Flags.size() - 2); 1872 SuggestedCompletions = Opts.suggestValueCompletions(Prev, Cur); 1873 } 1874 1875 if (SuggestedCompletions.empty()) 1876 SuggestedCompletions = Opts.suggestValueCompletions(Cur, ""); 1877 1878 // If Flags were empty, it means the user typed `clang [tab]` where we should 1879 // list all possible flags. If there was no value completion and the user 1880 // pressed tab after a space, we should fall back to a file completion. 1881 // We're printing a newline to be consistent with what we print at the end of 1882 // this function. 1883 if (SuggestedCompletions.empty() && HasSpace && !Flags.empty()) { 1884 llvm::outs() << '\n'; 1885 return; 1886 } 1887 1888 // When flag ends with '=' and there was no value completion, return empty 1889 // string and fall back to the file autocompletion. 1890 if (SuggestedCompletions.empty() && !Cur.endswith("=")) { 1891 // If the flag is in the form of "--autocomplete=-foo", 1892 // we were requested to print out all option names that start with "-foo". 1893 // For example, "--autocomplete=-fsyn" is expanded to "-fsyntax-only". 1894 SuggestedCompletions = Opts.findByPrefix(Cur, DisableFlags); 1895 1896 // We have to query the -W flags manually as they're not in the OptTable. 1897 // TODO: Find a good way to add them to OptTable instead and them remove 1898 // this code. 1899 for (StringRef S : DiagnosticIDs::getDiagnosticFlags()) 1900 if (S.startswith(Cur)) 1901 SuggestedCompletions.push_back(std::string(S)); 1902 } 1903 1904 // Sort the autocomplete candidates so that shells print them out in a 1905 // deterministic order. We could sort in any way, but we chose 1906 // case-insensitive sorting for consistency with the -help option 1907 // which prints out options in the case-insensitive alphabetical order. 1908 llvm::sort(SuggestedCompletions, [](StringRef A, StringRef B) { 1909 if (int X = A.compare_insensitive(B)) 1910 return X < 0; 1911 return A.compare(B) > 0; 1912 }); 1913 1914 llvm::outs() << llvm::join(SuggestedCompletions, "\n") << '\n'; 1915 } 1916 1917 bool Driver::HandleImmediateArgs(const Compilation &C) { 1918 // The order these options are handled in gcc is all over the place, but we 1919 // don't expect inconsistencies w.r.t. that to matter in practice. 1920 1921 if (C.getArgs().hasArg(options::OPT_dumpmachine)) { 1922 llvm::outs() << C.getDefaultToolChain().getTripleString() << '\n'; 1923 return false; 1924 } 1925 1926 if (C.getArgs().hasArg(options::OPT_dumpversion)) { 1927 // Since -dumpversion is only implemented for pedantic GCC compatibility, we 1928 // return an answer which matches our definition of __VERSION__. 1929 llvm::outs() << CLANG_VERSION_STRING << "\n"; 1930 return false; 1931 } 1932 1933 if (C.getArgs().hasArg(options::OPT__print_diagnostic_categories)) { 1934 PrintDiagnosticCategories(llvm::outs()); 1935 return false; 1936 } 1937 1938 if (C.getArgs().hasArg(options::OPT_help) || 1939 C.getArgs().hasArg(options::OPT__help_hidden)) { 1940 PrintHelp(C.getArgs().hasArg(options::OPT__help_hidden)); 1941 return false; 1942 } 1943 1944 if (C.getArgs().hasArg(options::OPT__version)) { 1945 // Follow gcc behavior and use stdout for --version and stderr for -v. 1946 PrintVersion(C, llvm::outs()); 1947 return false; 1948 } 1949 1950 if (C.getArgs().hasArg(options::OPT_v) || 1951 C.getArgs().hasArg(options::OPT__HASH_HASH_HASH) || 1952 C.getArgs().hasArg(options::OPT_print_supported_cpus)) { 1953 PrintVersion(C, llvm::errs()); 1954 SuppressMissingInputWarning = true; 1955 } 1956 1957 if (C.getArgs().hasArg(options::OPT_v)) { 1958 if (!SystemConfigDir.empty()) 1959 llvm::errs() << "System configuration file directory: " 1960 << SystemConfigDir << "\n"; 1961 if (!UserConfigDir.empty()) 1962 llvm::errs() << "User configuration file directory: " 1963 << UserConfigDir << "\n"; 1964 } 1965 1966 const ToolChain &TC = C.getDefaultToolChain(); 1967 1968 if (C.getArgs().hasArg(options::OPT_v)) 1969 TC.printVerboseInfo(llvm::errs()); 1970 1971 if (C.getArgs().hasArg(options::OPT_print_resource_dir)) { 1972 llvm::outs() << ResourceDir << '\n'; 1973 return false; 1974 } 1975 1976 if (C.getArgs().hasArg(options::OPT_print_search_dirs)) { 1977 llvm::outs() << "programs: ="; 1978 bool separator = false; 1979 // Print -B and COMPILER_PATH. 1980 for (const std::string &Path : PrefixDirs) { 1981 if (separator) 1982 llvm::outs() << llvm::sys::EnvPathSeparator; 1983 llvm::outs() << Path; 1984 separator = true; 1985 } 1986 for (const std::string &Path : TC.getProgramPaths()) { 1987 if (separator) 1988 llvm::outs() << llvm::sys::EnvPathSeparator; 1989 llvm::outs() << Path; 1990 separator = true; 1991 } 1992 llvm::outs() << "\n"; 1993 llvm::outs() << "libraries: =" << ResourceDir; 1994 1995 StringRef sysroot = C.getSysRoot(); 1996 1997 for (const std::string &Path : TC.getFilePaths()) { 1998 // Always print a separator. ResourceDir was the first item shown. 1999 llvm::outs() << llvm::sys::EnvPathSeparator; 2000 // Interpretation of leading '=' is needed only for NetBSD. 2001 if (Path[0] == '=') 2002 llvm::outs() << sysroot << Path.substr(1); 2003 else 2004 llvm::outs() << Path; 2005 } 2006 llvm::outs() << "\n"; 2007 return false; 2008 } 2009 2010 if (C.getArgs().hasArg(options::OPT_print_runtime_dir)) { 2011 std::string RuntimePath; 2012 // Get the first existing path, if any. 2013 for (auto Path : TC.getRuntimePaths()) { 2014 if (getVFS().exists(Path)) { 2015 RuntimePath = Path; 2016 break; 2017 } 2018 } 2019 if (!RuntimePath.empty()) 2020 llvm::outs() << RuntimePath << '\n'; 2021 else 2022 llvm::outs() << TC.getCompilerRTPath() << '\n'; 2023 return false; 2024 } 2025 2026 if (C.getArgs().hasArg(options::OPT_print_diagnostic_options)) { 2027 std::vector<std::string> Flags = DiagnosticIDs::getDiagnosticFlags(); 2028 for (std::size_t I = 0; I != Flags.size(); I += 2) 2029 llvm::outs() << " " << Flags[I] << "\n " << Flags[I + 1] << "\n\n"; 2030 return false; 2031 } 2032 2033 // FIXME: The following handlers should use a callback mechanism, we don't 2034 // know what the client would like to do. 2035 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_file_name_EQ)) { 2036 llvm::outs() << GetFilePath(A->getValue(), TC) << "\n"; 2037 return false; 2038 } 2039 2040 if (Arg *A = C.getArgs().getLastArg(options::OPT_print_prog_name_EQ)) { 2041 StringRef ProgName = A->getValue(); 2042 2043 // Null program name cannot have a path. 2044 if (! ProgName.empty()) 2045 llvm::outs() << GetProgramPath(ProgName, TC); 2046 2047 llvm::outs() << "\n"; 2048 return false; 2049 } 2050 2051 if (Arg *A = C.getArgs().getLastArg(options::OPT_autocomplete)) { 2052 StringRef PassedFlags = A->getValue(); 2053 HandleAutocompletions(PassedFlags); 2054 return false; 2055 } 2056 2057 if (C.getArgs().hasArg(options::OPT_print_libgcc_file_name)) { 2058 ToolChain::RuntimeLibType RLT = TC.GetRuntimeLibType(C.getArgs()); 2059 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2060 RegisterEffectiveTriple TripleRAII(TC, Triple); 2061 switch (RLT) { 2062 case ToolChain::RLT_CompilerRT: 2063 llvm::outs() << TC.getCompilerRT(C.getArgs(), "builtins") << "\n"; 2064 break; 2065 case ToolChain::RLT_Libgcc: 2066 llvm::outs() << GetFilePath("libgcc.a", TC) << "\n"; 2067 break; 2068 } 2069 return false; 2070 } 2071 2072 if (C.getArgs().hasArg(options::OPT_print_multi_lib)) { 2073 for (const Multilib &Multilib : TC.getMultilibs()) 2074 llvm::outs() << Multilib << "\n"; 2075 return false; 2076 } 2077 2078 if (C.getArgs().hasArg(options::OPT_print_multi_directory)) { 2079 const Multilib &Multilib = TC.getMultilib(); 2080 if (Multilib.gccSuffix().empty()) 2081 llvm::outs() << ".\n"; 2082 else { 2083 StringRef Suffix(Multilib.gccSuffix()); 2084 assert(Suffix.front() == '/'); 2085 llvm::outs() << Suffix.substr(1) << "\n"; 2086 } 2087 return false; 2088 } 2089 2090 if (C.getArgs().hasArg(options::OPT_print_target_triple)) { 2091 llvm::outs() << TC.getTripleString() << "\n"; 2092 return false; 2093 } 2094 2095 if (C.getArgs().hasArg(options::OPT_print_effective_triple)) { 2096 const llvm::Triple Triple(TC.ComputeEffectiveClangTriple(C.getArgs())); 2097 llvm::outs() << Triple.getTriple() << "\n"; 2098 return false; 2099 } 2100 2101 if (C.getArgs().hasArg(options::OPT_print_multiarch)) { 2102 llvm::outs() << TC.getMultiarchTriple(*this, TC.getTriple(), SysRoot) 2103 << "\n"; 2104 return false; 2105 } 2106 2107 if (C.getArgs().hasArg(options::OPT_print_targets)) { 2108 llvm::TargetRegistry::printRegisteredTargetsForVersion(llvm::outs()); 2109 return false; 2110 } 2111 2112 return true; 2113 } 2114 2115 enum { 2116 TopLevelAction = 0, 2117 HeadSibAction = 1, 2118 OtherSibAction = 2, 2119 }; 2120 2121 // Display an action graph human-readably. Action A is the "sink" node 2122 // and latest-occuring action. Traversal is in pre-order, visiting the 2123 // inputs to each action before printing the action itself. 2124 static unsigned PrintActions1(const Compilation &C, Action *A, 2125 std::map<Action *, unsigned> &Ids, 2126 Twine Indent = {}, int Kind = TopLevelAction) { 2127 if (Ids.count(A)) // A was already visited. 2128 return Ids[A]; 2129 2130 std::string str; 2131 llvm::raw_string_ostream os(str); 2132 2133 auto getSibIndent = [](int K) -> Twine { 2134 return (K == HeadSibAction) ? " " : (K == OtherSibAction) ? "| " : ""; 2135 }; 2136 2137 Twine SibIndent = Indent + getSibIndent(Kind); 2138 int SibKind = HeadSibAction; 2139 os << Action::getClassName(A->getKind()) << ", "; 2140 if (InputAction *IA = dyn_cast<InputAction>(A)) { 2141 os << "\"" << IA->getInputArg().getValue() << "\""; 2142 } else if (BindArchAction *BIA = dyn_cast<BindArchAction>(A)) { 2143 os << '"' << BIA->getArchName() << '"' << ", {" 2144 << PrintActions1(C, *BIA->input_begin(), Ids, SibIndent, SibKind) << "}"; 2145 } else if (OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 2146 bool IsFirst = true; 2147 OA->doOnEachDependence( 2148 [&](Action *A, const ToolChain *TC, const char *BoundArch) { 2149 assert(TC && "Unknown host toolchain"); 2150 // E.g. for two CUDA device dependences whose bound arch is sm_20 and 2151 // sm_35 this will generate: 2152 // "cuda-device" (nvptx64-nvidia-cuda:sm_20) {#ID}, "cuda-device" 2153 // (nvptx64-nvidia-cuda:sm_35) {#ID} 2154 if (!IsFirst) 2155 os << ", "; 2156 os << '"'; 2157 os << A->getOffloadingKindPrefix(); 2158 os << " ("; 2159 os << TC->getTriple().normalize(); 2160 if (BoundArch) 2161 os << ":" << BoundArch; 2162 os << ")"; 2163 os << '"'; 2164 os << " {" << PrintActions1(C, A, Ids, SibIndent, SibKind) << "}"; 2165 IsFirst = false; 2166 SibKind = OtherSibAction; 2167 }); 2168 } else { 2169 const ActionList *AL = &A->getInputs(); 2170 2171 if (AL->size()) { 2172 const char *Prefix = "{"; 2173 for (Action *PreRequisite : *AL) { 2174 os << Prefix << PrintActions1(C, PreRequisite, Ids, SibIndent, SibKind); 2175 Prefix = ", "; 2176 SibKind = OtherSibAction; 2177 } 2178 os << "}"; 2179 } else 2180 os << "{}"; 2181 } 2182 2183 // Append offload info for all options other than the offloading action 2184 // itself (e.g. (cuda-device, sm_20) or (cuda-host)). 2185 std::string offload_str; 2186 llvm::raw_string_ostream offload_os(offload_str); 2187 if (!isa<OffloadAction>(A)) { 2188 auto S = A->getOffloadingKindPrefix(); 2189 if (!S.empty()) { 2190 offload_os << ", (" << S; 2191 if (A->getOffloadingArch()) 2192 offload_os << ", " << A->getOffloadingArch(); 2193 offload_os << ")"; 2194 } 2195 } 2196 2197 auto getSelfIndent = [](int K) -> Twine { 2198 return (K == HeadSibAction) ? "+- " : (K == OtherSibAction) ? "|- " : ""; 2199 }; 2200 2201 unsigned Id = Ids.size(); 2202 Ids[A] = Id; 2203 llvm::errs() << Indent + getSelfIndent(Kind) << Id << ": " << os.str() << ", " 2204 << types::getTypeName(A->getType()) << offload_os.str() << "\n"; 2205 2206 return Id; 2207 } 2208 2209 // Print the action graphs in a compilation C. 2210 // For example "clang -c file1.c file2.c" is composed of two subgraphs. 2211 void Driver::PrintActions(const Compilation &C) const { 2212 std::map<Action *, unsigned> Ids; 2213 for (Action *A : C.getActions()) 2214 PrintActions1(C, A, Ids); 2215 } 2216 2217 /// Check whether the given input tree contains any compilation or 2218 /// assembly actions. 2219 static bool ContainsCompileOrAssembleAction(const Action *A) { 2220 if (isa<CompileJobAction>(A) || isa<BackendJobAction>(A) || 2221 isa<AssembleJobAction>(A)) 2222 return true; 2223 2224 return llvm::any_of(A->inputs(), ContainsCompileOrAssembleAction); 2225 } 2226 2227 void Driver::BuildUniversalActions(Compilation &C, const ToolChain &TC, 2228 const InputList &BAInputs) const { 2229 DerivedArgList &Args = C.getArgs(); 2230 ActionList &Actions = C.getActions(); 2231 llvm::PrettyStackTraceString CrashInfo("Building universal build actions"); 2232 // Collect the list of architectures. Duplicates are allowed, but should only 2233 // be handled once (in the order seen). 2234 llvm::StringSet<> ArchNames; 2235 SmallVector<const char *, 4> Archs; 2236 for (Arg *A : Args) { 2237 if (A->getOption().matches(options::OPT_arch)) { 2238 // Validate the option here; we don't save the type here because its 2239 // particular spelling may participate in other driver choices. 2240 llvm::Triple::ArchType Arch = 2241 tools::darwin::getArchTypeForMachOArchName(A->getValue()); 2242 if (Arch == llvm::Triple::UnknownArch) { 2243 Diag(clang::diag::err_drv_invalid_arch_name) << A->getAsString(Args); 2244 continue; 2245 } 2246 2247 A->claim(); 2248 if (ArchNames.insert(A->getValue()).second) 2249 Archs.push_back(A->getValue()); 2250 } 2251 } 2252 2253 // When there is no explicit arch for this platform, make sure we still bind 2254 // the architecture (to the default) so that -Xarch_ is handled correctly. 2255 if (!Archs.size()) 2256 Archs.push_back(Args.MakeArgString(TC.getDefaultUniversalArchName())); 2257 2258 ActionList SingleActions; 2259 BuildActions(C, Args, BAInputs, SingleActions); 2260 2261 // Add in arch bindings for every top level action, as well as lipo and 2262 // dsymutil steps if needed. 2263 for (Action* Act : SingleActions) { 2264 // Make sure we can lipo this kind of output. If not (and it is an actual 2265 // output) then we disallow, since we can't create an output file with the 2266 // right name without overwriting it. We could remove this oddity by just 2267 // changing the output names to include the arch, which would also fix 2268 // -save-temps. Compatibility wins for now. 2269 2270 if (Archs.size() > 1 && !types::canLipoType(Act->getType())) 2271 Diag(clang::diag::err_drv_invalid_output_with_multiple_archs) 2272 << types::getTypeName(Act->getType()); 2273 2274 ActionList Inputs; 2275 for (unsigned i = 0, e = Archs.size(); i != e; ++i) 2276 Inputs.push_back(C.MakeAction<BindArchAction>(Act, Archs[i])); 2277 2278 // Lipo if necessary, we do it this way because we need to set the arch flag 2279 // so that -Xarch_ gets overwritten. 2280 if (Inputs.size() == 1 || Act->getType() == types::TY_Nothing) 2281 Actions.append(Inputs.begin(), Inputs.end()); 2282 else 2283 Actions.push_back(C.MakeAction<LipoJobAction>(Inputs, Act->getType())); 2284 2285 // Handle debug info queries. 2286 Arg *A = Args.getLastArg(options::OPT_g_Group); 2287 bool enablesDebugInfo = A && !A->getOption().matches(options::OPT_g0) && 2288 !A->getOption().matches(options::OPT_gstabs); 2289 if ((enablesDebugInfo || willEmitRemarks(Args)) && 2290 ContainsCompileOrAssembleAction(Actions.back())) { 2291 2292 // Add a 'dsymutil' step if necessary, when debug info is enabled and we 2293 // have a compile input. We need to run 'dsymutil' ourselves in such cases 2294 // because the debug info will refer to a temporary object file which 2295 // will be removed at the end of the compilation process. 2296 if (Act->getType() == types::TY_Image) { 2297 ActionList Inputs; 2298 Inputs.push_back(Actions.back()); 2299 Actions.pop_back(); 2300 Actions.push_back( 2301 C.MakeAction<DsymutilJobAction>(Inputs, types::TY_dSYM)); 2302 } 2303 2304 // Verify the debug info output. 2305 if (Args.hasArg(options::OPT_verify_debug_info)) { 2306 Action* LastAction = Actions.back(); 2307 Actions.pop_back(); 2308 Actions.push_back(C.MakeAction<VerifyDebugInfoJobAction>( 2309 LastAction, types::TY_Nothing)); 2310 } 2311 } 2312 } 2313 } 2314 2315 bool Driver::DiagnoseInputExistence(const DerivedArgList &Args, StringRef Value, 2316 types::ID Ty, bool TypoCorrect) const { 2317 if (!getCheckInputsExist()) 2318 return true; 2319 2320 // stdin always exists. 2321 if (Value == "-") 2322 return true; 2323 2324 // If it's a header to be found in the system or user search path, then defer 2325 // complaints about its absence until those searches can be done. When we 2326 // are definitely processing headers for C++20 header units, extend this to 2327 // allow the user to put "-fmodule-header -xc++-header vector" for example. 2328 if (Ty == types::TY_CXXSHeader || Ty == types::TY_CXXUHeader || 2329 (ModulesModeCXX20 && Ty == types::TY_CXXHeader)) 2330 return true; 2331 2332 if (getVFS().exists(Value)) 2333 return true; 2334 2335 if (TypoCorrect) { 2336 // Check if the filename is a typo for an option flag. OptTable thinks 2337 // that all args that are not known options and that start with / are 2338 // filenames, but e.g. `/diagnostic:caret` is more likely a typo for 2339 // the option `/diagnostics:caret` than a reference to a file in the root 2340 // directory. 2341 unsigned IncludedFlagsBitmask; 2342 unsigned ExcludedFlagsBitmask; 2343 std::tie(IncludedFlagsBitmask, ExcludedFlagsBitmask) = 2344 getIncludeExcludeOptionFlagMasks(IsCLMode()); 2345 std::string Nearest; 2346 if (getOpts().findNearest(Value, Nearest, IncludedFlagsBitmask, 2347 ExcludedFlagsBitmask) <= 1) { 2348 Diag(clang::diag::err_drv_no_such_file_with_suggestion) 2349 << Value << Nearest; 2350 return false; 2351 } 2352 } 2353 2354 // In CL mode, don't error on apparently non-existent linker inputs, because 2355 // they can be influenced by linker flags the clang driver might not 2356 // understand. 2357 // Examples: 2358 // - `clang-cl main.cc ole32.lib` in a a non-MSVC shell will make the driver 2359 // module look for an MSVC installation in the registry. (We could ask 2360 // the MSVCToolChain object if it can find `ole32.lib`, but the logic to 2361 // look in the registry might move into lld-link in the future so that 2362 // lld-link invocations in non-MSVC shells just work too.) 2363 // - `clang-cl ... /link ...` can pass arbitrary flags to the linker, 2364 // including /libpath:, which is used to find .lib and .obj files. 2365 // So do not diagnose this on the driver level. Rely on the linker diagnosing 2366 // it. (If we don't end up invoking the linker, this means we'll emit a 2367 // "'linker' input unused [-Wunused-command-line-argument]" warning instead 2368 // of an error.) 2369 // 2370 // Only do this skip after the typo correction step above. `/Brepo` is treated 2371 // as TY_Object, but it's clearly a typo for `/Brepro`. It seems fine to emit 2372 // an error if we have a flag that's within an edit distance of 1 from a 2373 // flag. (Users can use `-Wl,` or `/linker` to launder the flag past the 2374 // driver in the unlikely case they run into this.) 2375 // 2376 // Don't do this for inputs that start with a '/', else we'd pass options 2377 // like /libpath: through to the linker silently. 2378 // 2379 // Emitting an error for linker inputs can also cause incorrect diagnostics 2380 // with the gcc driver. The command 2381 // clang -fuse-ld=lld -Wl,--chroot,some/dir /file.o 2382 // will make lld look for some/dir/file.o, while we will diagnose here that 2383 // `/file.o` does not exist. However, configure scripts check if 2384 // `clang /GR-` compiles without error to see if the compiler is cl.exe, 2385 // so we can't downgrade diagnostics for `/GR-` from an error to a warning 2386 // in cc mode. (We can in cl mode because cl.exe itself only warns on 2387 // unknown flags.) 2388 if (IsCLMode() && Ty == types::TY_Object && !Value.startswith("/")) 2389 return true; 2390 2391 Diag(clang::diag::err_drv_no_such_file) << Value; 2392 return false; 2393 } 2394 2395 // Get the C++20 Header Unit type corresponding to the input type. 2396 static types::ID CXXHeaderUnitType(ModuleHeaderMode HM) { 2397 switch (HM) { 2398 case HeaderMode_User: 2399 return types::TY_CXXUHeader; 2400 case HeaderMode_System: 2401 return types::TY_CXXSHeader; 2402 case HeaderMode_Default: 2403 break; 2404 case HeaderMode_None: 2405 llvm_unreachable("should not be called in this case"); 2406 } 2407 return types::TY_CXXHUHeader; 2408 } 2409 2410 // Construct a the list of inputs and their types. 2411 void Driver::BuildInputs(const ToolChain &TC, DerivedArgList &Args, 2412 InputList &Inputs) const { 2413 const llvm::opt::OptTable &Opts = getOpts(); 2414 // Track the current user specified (-x) input. We also explicitly track the 2415 // argument used to set the type; we only want to claim the type when we 2416 // actually use it, so we warn about unused -x arguments. 2417 types::ID InputType = types::TY_Nothing; 2418 Arg *InputTypeArg = nullptr; 2419 2420 // The last /TC or /TP option sets the input type to C or C++ globally. 2421 if (Arg *TCTP = Args.getLastArgNoClaim(options::OPT__SLASH_TC, 2422 options::OPT__SLASH_TP)) { 2423 InputTypeArg = TCTP; 2424 InputType = TCTP->getOption().matches(options::OPT__SLASH_TC) 2425 ? types::TY_C 2426 : types::TY_CXX; 2427 2428 Arg *Previous = nullptr; 2429 bool ShowNote = false; 2430 for (Arg *A : 2431 Args.filtered(options::OPT__SLASH_TC, options::OPT__SLASH_TP)) { 2432 if (Previous) { 2433 Diag(clang::diag::warn_drv_overriding_flag_option) 2434 << Previous->getSpelling() << A->getSpelling(); 2435 ShowNote = true; 2436 } 2437 Previous = A; 2438 } 2439 if (ShowNote) 2440 Diag(clang::diag::note_drv_t_option_is_global); 2441 2442 // No driver mode exposes -x and /TC or /TP; we don't support mixing them. 2443 assert(!Args.hasArg(options::OPT_x) && "-x and /TC or /TP is not allowed"); 2444 } 2445 2446 // Warn -x after last input file has no effect 2447 { 2448 Arg *LastXArg = Args.getLastArgNoClaim(options::OPT_x); 2449 Arg *LastInputArg = Args.getLastArgNoClaim(options::OPT_INPUT); 2450 if (LastXArg && LastInputArg && LastInputArg->getIndex() < LastXArg->getIndex()) 2451 Diag(clang::diag::warn_drv_unused_x) << LastXArg->getValue(); 2452 } 2453 2454 for (Arg *A : Args) { 2455 if (A->getOption().getKind() == Option::InputClass) { 2456 const char *Value = A->getValue(); 2457 types::ID Ty = types::TY_INVALID; 2458 2459 // Infer the input type if necessary. 2460 if (InputType == types::TY_Nothing) { 2461 // If there was an explicit arg for this, claim it. 2462 if (InputTypeArg) 2463 InputTypeArg->claim(); 2464 2465 // stdin must be handled specially. 2466 if (memcmp(Value, "-", 2) == 0) { 2467 if (IsFlangMode()) { 2468 Ty = types::TY_Fortran; 2469 } else { 2470 // If running with -E, treat as a C input (this changes the 2471 // builtin macros, for example). This may be overridden by -ObjC 2472 // below. 2473 // 2474 // Otherwise emit an error but still use a valid type to avoid 2475 // spurious errors (e.g., no inputs). 2476 assert(!CCGenDiagnostics && "stdin produces no crash reproducer"); 2477 if (!Args.hasArgNoClaim(options::OPT_E) && !CCCIsCPP()) 2478 Diag(IsCLMode() ? clang::diag::err_drv_unknown_stdin_type_clang_cl 2479 : clang::diag::err_drv_unknown_stdin_type); 2480 Ty = types::TY_C; 2481 } 2482 } else { 2483 // Otherwise lookup by extension. 2484 // Fallback is C if invoked as C preprocessor, C++ if invoked with 2485 // clang-cl /E, or Object otherwise. 2486 // We use a host hook here because Darwin at least has its own 2487 // idea of what .s is. 2488 if (const char *Ext = strrchr(Value, '.')) 2489 Ty = TC.LookupTypeForExtension(Ext + 1); 2490 2491 if (Ty == types::TY_INVALID) { 2492 if (IsCLMode() && (Args.hasArgNoClaim(options::OPT_E) || CCGenDiagnostics)) 2493 Ty = types::TY_CXX; 2494 else if (CCCIsCPP() || CCGenDiagnostics) 2495 Ty = types::TY_C; 2496 else 2497 Ty = types::TY_Object; 2498 } 2499 2500 // If the driver is invoked as C++ compiler (like clang++ or c++) it 2501 // should autodetect some input files as C++ for g++ compatibility. 2502 if (CCCIsCXX()) { 2503 types::ID OldTy = Ty; 2504 Ty = types::lookupCXXTypeForCType(Ty); 2505 2506 // Do not complain about foo.h, when we are known to be processing 2507 // it as a C++20 header unit. 2508 if (Ty != OldTy && !(OldTy == types::TY_CHeader && hasHeaderMode())) 2509 Diag(clang::diag::warn_drv_treating_input_as_cxx) 2510 << getTypeName(OldTy) << getTypeName(Ty); 2511 } 2512 2513 // If running with -fthinlto-index=, extensions that normally identify 2514 // native object files actually identify LLVM bitcode files. 2515 if (Args.hasArgNoClaim(options::OPT_fthinlto_index_EQ) && 2516 Ty == types::TY_Object) 2517 Ty = types::TY_LLVM_BC; 2518 } 2519 2520 // -ObjC and -ObjC++ override the default language, but only for "source 2521 // files". We just treat everything that isn't a linker input as a 2522 // source file. 2523 // 2524 // FIXME: Clean this up if we move the phase sequence into the type. 2525 if (Ty != types::TY_Object) { 2526 if (Args.hasArg(options::OPT_ObjC)) 2527 Ty = types::TY_ObjC; 2528 else if (Args.hasArg(options::OPT_ObjCXX)) 2529 Ty = types::TY_ObjCXX; 2530 } 2531 2532 // Disambiguate headers that are meant to be header units from those 2533 // intended to be PCH. Avoid missing '.h' cases that are counted as 2534 // C headers by default - we know we are in C++ mode and we do not 2535 // want to issue a complaint about compiling things in the wrong mode. 2536 if ((Ty == types::TY_CXXHeader || Ty == types::TY_CHeader) && 2537 hasHeaderMode()) 2538 Ty = CXXHeaderUnitType(CXX20HeaderType); 2539 } else { 2540 assert(InputTypeArg && "InputType set w/o InputTypeArg"); 2541 if (!InputTypeArg->getOption().matches(options::OPT_x)) { 2542 // If emulating cl.exe, make sure that /TC and /TP don't affect input 2543 // object files. 2544 const char *Ext = strrchr(Value, '.'); 2545 if (Ext && TC.LookupTypeForExtension(Ext + 1) == types::TY_Object) 2546 Ty = types::TY_Object; 2547 } 2548 if (Ty == types::TY_INVALID) { 2549 Ty = InputType; 2550 InputTypeArg->claim(); 2551 } 2552 } 2553 2554 if (DiagnoseInputExistence(Args, Value, Ty, /*TypoCorrect=*/true)) 2555 Inputs.push_back(std::make_pair(Ty, A)); 2556 2557 } else if (A->getOption().matches(options::OPT__SLASH_Tc)) { 2558 StringRef Value = A->getValue(); 2559 if (DiagnoseInputExistence(Args, Value, types::TY_C, 2560 /*TypoCorrect=*/false)) { 2561 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2562 Inputs.push_back(std::make_pair(types::TY_C, InputArg)); 2563 } 2564 A->claim(); 2565 } else if (A->getOption().matches(options::OPT__SLASH_Tp)) { 2566 StringRef Value = A->getValue(); 2567 if (DiagnoseInputExistence(Args, Value, types::TY_CXX, 2568 /*TypoCorrect=*/false)) { 2569 Arg *InputArg = MakeInputArg(Args, Opts, A->getValue()); 2570 Inputs.push_back(std::make_pair(types::TY_CXX, InputArg)); 2571 } 2572 A->claim(); 2573 } else if (A->getOption().hasFlag(options::LinkerInput)) { 2574 // Just treat as object type, we could make a special type for this if 2575 // necessary. 2576 Inputs.push_back(std::make_pair(types::TY_Object, A)); 2577 2578 } else if (A->getOption().matches(options::OPT_x)) { 2579 InputTypeArg = A; 2580 InputType = types::lookupTypeForTypeSpecifier(A->getValue()); 2581 A->claim(); 2582 2583 // Follow gcc behavior and treat as linker input for invalid -x 2584 // options. Its not clear why we shouldn't just revert to unknown; but 2585 // this isn't very important, we might as well be bug compatible. 2586 if (!InputType) { 2587 Diag(clang::diag::err_drv_unknown_language) << A->getValue(); 2588 InputType = types::TY_Object; 2589 } 2590 2591 // If the user has put -fmodule-header{,=} then we treat C++ headers as 2592 // header unit inputs. So we 'promote' -xc++-header appropriately. 2593 if (InputType == types::TY_CXXHeader && hasHeaderMode()) 2594 InputType = CXXHeaderUnitType(CXX20HeaderType); 2595 } else if (A->getOption().getID() == options::OPT_U) { 2596 assert(A->getNumValues() == 1 && "The /U option has one value."); 2597 StringRef Val = A->getValue(0); 2598 if (Val.find_first_of("/\\") != StringRef::npos) { 2599 // Warn about e.g. "/Users/me/myfile.c". 2600 Diag(diag::warn_slash_u_filename) << Val; 2601 Diag(diag::note_use_dashdash); 2602 } 2603 } 2604 } 2605 if (CCCIsCPP() && Inputs.empty()) { 2606 // If called as standalone preprocessor, stdin is processed 2607 // if no other input is present. 2608 Arg *A = MakeInputArg(Args, Opts, "-"); 2609 Inputs.push_back(std::make_pair(types::TY_C, A)); 2610 } 2611 } 2612 2613 namespace { 2614 /// Provides a convenient interface for different programming models to generate 2615 /// the required device actions. 2616 class OffloadingActionBuilder final { 2617 /// Flag used to trace errors in the builder. 2618 bool IsValid = false; 2619 2620 /// The compilation that is using this builder. 2621 Compilation &C; 2622 2623 /// Map between an input argument and the offload kinds used to process it. 2624 std::map<const Arg *, unsigned> InputArgToOffloadKindMap; 2625 2626 /// Map between a host action and its originating input argument. 2627 std::map<Action *, const Arg *> HostActionToInputArgMap; 2628 2629 /// Builder interface. It doesn't build anything or keep any state. 2630 class DeviceActionBuilder { 2631 public: 2632 typedef const llvm::SmallVectorImpl<phases::ID> PhasesTy; 2633 2634 enum ActionBuilderReturnCode { 2635 // The builder acted successfully on the current action. 2636 ABRT_Success, 2637 // The builder didn't have to act on the current action. 2638 ABRT_Inactive, 2639 // The builder was successful and requested the host action to not be 2640 // generated. 2641 ABRT_Ignore_Host, 2642 }; 2643 2644 protected: 2645 /// Compilation associated with this builder. 2646 Compilation &C; 2647 2648 /// Tool chains associated with this builder. The same programming 2649 /// model may have associated one or more tool chains. 2650 SmallVector<const ToolChain *, 2> ToolChains; 2651 2652 /// The derived arguments associated with this builder. 2653 DerivedArgList &Args; 2654 2655 /// The inputs associated with this builder. 2656 const Driver::InputList &Inputs; 2657 2658 /// The associated offload kind. 2659 Action::OffloadKind AssociatedOffloadKind = Action::OFK_None; 2660 2661 public: 2662 DeviceActionBuilder(Compilation &C, DerivedArgList &Args, 2663 const Driver::InputList &Inputs, 2664 Action::OffloadKind AssociatedOffloadKind) 2665 : C(C), Args(Args), Inputs(Inputs), 2666 AssociatedOffloadKind(AssociatedOffloadKind) {} 2667 virtual ~DeviceActionBuilder() {} 2668 2669 /// Fill up the array \a DA with all the device dependences that should be 2670 /// added to the provided host action \a HostAction. By default it is 2671 /// inactive. 2672 virtual ActionBuilderReturnCode 2673 getDeviceDependences(OffloadAction::DeviceDependences &DA, 2674 phases::ID CurPhase, phases::ID FinalPhase, 2675 PhasesTy &Phases) { 2676 return ABRT_Inactive; 2677 } 2678 2679 /// Update the state to include the provided host action \a HostAction as a 2680 /// dependency of the current device action. By default it is inactive. 2681 virtual ActionBuilderReturnCode addDeviceDepences(Action *HostAction) { 2682 return ABRT_Inactive; 2683 } 2684 2685 /// Append top level actions generated by the builder. 2686 virtual void appendTopLevelActions(ActionList &AL) {} 2687 2688 /// Append linker device actions generated by the builder. 2689 virtual void appendLinkDeviceActions(ActionList &AL) {} 2690 2691 /// Append linker host action generated by the builder. 2692 virtual Action* appendLinkHostActions(ActionList &AL) { return nullptr; } 2693 2694 /// Append linker actions generated by the builder. 2695 virtual void appendLinkDependences(OffloadAction::DeviceDependences &DA) {} 2696 2697 /// Initialize the builder. Return true if any initialization errors are 2698 /// found. 2699 virtual bool initialize() { return false; } 2700 2701 /// Return true if the builder can use bundling/unbundling. 2702 virtual bool canUseBundlerUnbundler() const { return false; } 2703 2704 /// Return true if this builder is valid. We have a valid builder if we have 2705 /// associated device tool chains. 2706 bool isValid() { return !ToolChains.empty(); } 2707 2708 /// Return the associated offload kind. 2709 Action::OffloadKind getAssociatedOffloadKind() { 2710 return AssociatedOffloadKind; 2711 } 2712 }; 2713 2714 /// Base class for CUDA/HIP action builder. It injects device code in 2715 /// the host backend action. 2716 class CudaActionBuilderBase : public DeviceActionBuilder { 2717 protected: 2718 /// Flags to signal if the user requested host-only or device-only 2719 /// compilation. 2720 bool CompileHostOnly = false; 2721 bool CompileDeviceOnly = false; 2722 bool EmitLLVM = false; 2723 bool EmitAsm = false; 2724 2725 /// ID to identify each device compilation. For CUDA it is simply the 2726 /// GPU arch string. For HIP it is either the GPU arch string or GPU 2727 /// arch string plus feature strings delimited by a plus sign, e.g. 2728 /// gfx906+xnack. 2729 struct TargetID { 2730 /// Target ID string which is persistent throughout the compilation. 2731 const char *ID; 2732 TargetID(CudaArch Arch) { ID = CudaArchToString(Arch); } 2733 TargetID(const char *ID) : ID(ID) {} 2734 operator const char *() { return ID; } 2735 operator StringRef() { return StringRef(ID); } 2736 }; 2737 /// List of GPU architectures to use in this compilation. 2738 SmallVector<TargetID, 4> GpuArchList; 2739 2740 /// The CUDA actions for the current input. 2741 ActionList CudaDeviceActions; 2742 2743 /// The CUDA fat binary if it was generated for the current input. 2744 Action *CudaFatBinary = nullptr; 2745 2746 /// Flag that is set to true if this builder acted on the current input. 2747 bool IsActive = false; 2748 2749 /// Flag for -fgpu-rdc. 2750 bool Relocatable = false; 2751 2752 /// Default GPU architecture if there's no one specified. 2753 CudaArch DefaultCudaArch = CudaArch::UNKNOWN; 2754 2755 /// Method to generate compilation unit ID specified by option 2756 /// '-fuse-cuid='. 2757 enum UseCUIDKind { CUID_Hash, CUID_Random, CUID_None, CUID_Invalid }; 2758 UseCUIDKind UseCUID = CUID_Hash; 2759 2760 /// Compilation unit ID specified by option '-cuid='. 2761 StringRef FixedCUID; 2762 2763 public: 2764 CudaActionBuilderBase(Compilation &C, DerivedArgList &Args, 2765 const Driver::InputList &Inputs, 2766 Action::OffloadKind OFKind) 2767 : DeviceActionBuilder(C, Args, Inputs, OFKind) {} 2768 2769 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 2770 // While generating code for CUDA, we only depend on the host input action 2771 // to trigger the creation of all the CUDA device actions. 2772 2773 // If we are dealing with an input action, replicate it for each GPU 2774 // architecture. If we are in host-only mode we return 'success' so that 2775 // the host uses the CUDA offload kind. 2776 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 2777 assert(!GpuArchList.empty() && 2778 "We should have at least one GPU architecture."); 2779 2780 // If the host input is not CUDA or HIP, we don't need to bother about 2781 // this input. 2782 if (!(IA->getType() == types::TY_CUDA || 2783 IA->getType() == types::TY_HIP || 2784 IA->getType() == types::TY_PP_HIP)) { 2785 // The builder will ignore this input. 2786 IsActive = false; 2787 return ABRT_Inactive; 2788 } 2789 2790 // Set the flag to true, so that the builder acts on the current input. 2791 IsActive = true; 2792 2793 if (CompileHostOnly) 2794 return ABRT_Success; 2795 2796 // Replicate inputs for each GPU architecture. 2797 auto Ty = IA->getType() == types::TY_HIP ? types::TY_HIP_DEVICE 2798 : types::TY_CUDA_DEVICE; 2799 std::string CUID = FixedCUID.str(); 2800 if (CUID.empty()) { 2801 if (UseCUID == CUID_Random) 2802 CUID = llvm::utohexstr(llvm::sys::Process::GetRandomNumber(), 2803 /*LowerCase=*/true); 2804 else if (UseCUID == CUID_Hash) { 2805 llvm::MD5 Hasher; 2806 llvm::MD5::MD5Result Hash; 2807 SmallString<256> RealPath; 2808 llvm::sys::fs::real_path(IA->getInputArg().getValue(), RealPath, 2809 /*expand_tilde=*/true); 2810 Hasher.update(RealPath); 2811 for (auto *A : Args) { 2812 if (A->getOption().matches(options::OPT_INPUT)) 2813 continue; 2814 Hasher.update(A->getAsString(Args)); 2815 } 2816 Hasher.final(Hash); 2817 CUID = llvm::utohexstr(Hash.low(), /*LowerCase=*/true); 2818 } 2819 } 2820 IA->setId(CUID); 2821 2822 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 2823 CudaDeviceActions.push_back( 2824 C.MakeAction<InputAction>(IA->getInputArg(), Ty, IA->getId())); 2825 } 2826 2827 return ABRT_Success; 2828 } 2829 2830 // If this is an unbundling action use it as is for each CUDA toolchain. 2831 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 2832 2833 // If -fgpu-rdc is disabled, should not unbundle since there is no 2834 // device code to link. 2835 if (UA->getType() == types::TY_Object && !Relocatable) 2836 return ABRT_Inactive; 2837 2838 CudaDeviceActions.clear(); 2839 auto *IA = cast<InputAction>(UA->getInputs().back()); 2840 std::string FileName = IA->getInputArg().getAsString(Args); 2841 // Check if the type of the file is the same as the action. Do not 2842 // unbundle it if it is not. Do not unbundle .so files, for example, 2843 // which are not object files. 2844 if (IA->getType() == types::TY_Object && 2845 (!llvm::sys::path::has_extension(FileName) || 2846 types::lookupTypeForExtension( 2847 llvm::sys::path::extension(FileName).drop_front()) != 2848 types::TY_Object)) 2849 return ABRT_Inactive; 2850 2851 for (auto Arch : GpuArchList) { 2852 CudaDeviceActions.push_back(UA); 2853 UA->registerDependentActionInfo(ToolChains[0], Arch, 2854 AssociatedOffloadKind); 2855 } 2856 IsActive = true; 2857 return ABRT_Success; 2858 } 2859 2860 return IsActive ? ABRT_Success : ABRT_Inactive; 2861 } 2862 2863 void appendTopLevelActions(ActionList &AL) override { 2864 // Utility to append actions to the top level list. 2865 auto AddTopLevel = [&](Action *A, TargetID TargetID) { 2866 OffloadAction::DeviceDependences Dep; 2867 Dep.add(*A, *ToolChains.front(), TargetID, AssociatedOffloadKind); 2868 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 2869 }; 2870 2871 // If we have a fat binary, add it to the list. 2872 if (CudaFatBinary) { 2873 AddTopLevel(CudaFatBinary, CudaArch::UNUSED); 2874 CudaDeviceActions.clear(); 2875 CudaFatBinary = nullptr; 2876 return; 2877 } 2878 2879 if (CudaDeviceActions.empty()) 2880 return; 2881 2882 // If we have CUDA actions at this point, that's because we have a have 2883 // partial compilation, so we should have an action for each GPU 2884 // architecture. 2885 assert(CudaDeviceActions.size() == GpuArchList.size() && 2886 "Expecting one action per GPU architecture."); 2887 assert(ToolChains.size() == 1 && 2888 "Expecting to have a single CUDA toolchain."); 2889 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) 2890 AddTopLevel(CudaDeviceActions[I], GpuArchList[I]); 2891 2892 CudaDeviceActions.clear(); 2893 } 2894 2895 /// Get canonicalized offload arch option. \returns empty StringRef if the 2896 /// option is invalid. 2897 virtual StringRef getCanonicalOffloadArch(StringRef Arch) = 0; 2898 2899 virtual llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 2900 getConflictOffloadArchCombination(const std::set<StringRef> &GpuArchs) = 0; 2901 2902 bool initialize() override { 2903 assert(AssociatedOffloadKind == Action::OFK_Cuda || 2904 AssociatedOffloadKind == Action::OFK_HIP); 2905 2906 // We don't need to support CUDA. 2907 if (AssociatedOffloadKind == Action::OFK_Cuda && 2908 !C.hasOffloadToolChain<Action::OFK_Cuda>()) 2909 return false; 2910 2911 // We don't need to support HIP. 2912 if (AssociatedOffloadKind == Action::OFK_HIP && 2913 !C.hasOffloadToolChain<Action::OFK_HIP>()) 2914 return false; 2915 2916 Relocatable = Args.hasFlag(options::OPT_fgpu_rdc, 2917 options::OPT_fno_gpu_rdc, /*Default=*/false); 2918 2919 const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>(); 2920 assert(HostTC && "No toolchain for host compilation."); 2921 if (HostTC->getTriple().isNVPTX() || 2922 HostTC->getTriple().getArch() == llvm::Triple::amdgcn) { 2923 // We do not support targeting NVPTX/AMDGCN for host compilation. Throw 2924 // an error and abort pipeline construction early so we don't trip 2925 // asserts that assume device-side compilation. 2926 C.getDriver().Diag(diag::err_drv_cuda_host_arch) 2927 << HostTC->getTriple().getArchName(); 2928 return true; 2929 } 2930 2931 ToolChains.push_back( 2932 AssociatedOffloadKind == Action::OFK_Cuda 2933 ? C.getSingleOffloadToolChain<Action::OFK_Cuda>() 2934 : C.getSingleOffloadToolChain<Action::OFK_HIP>()); 2935 2936 Arg *PartialCompilationArg = Args.getLastArg( 2937 options::OPT_offload_host_only, options::OPT_offload_device_only, 2938 options::OPT_offload_host_device); 2939 CompileHostOnly = 2940 PartialCompilationArg && PartialCompilationArg->getOption().matches( 2941 options::OPT_offload_host_only); 2942 CompileDeviceOnly = 2943 PartialCompilationArg && PartialCompilationArg->getOption().matches( 2944 options::OPT_offload_device_only); 2945 EmitLLVM = Args.getLastArg(options::OPT_emit_llvm); 2946 EmitAsm = Args.getLastArg(options::OPT_S); 2947 FixedCUID = Args.getLastArgValue(options::OPT_cuid_EQ); 2948 if (Arg *A = Args.getLastArg(options::OPT_fuse_cuid_EQ)) { 2949 StringRef UseCUIDStr = A->getValue(); 2950 UseCUID = llvm::StringSwitch<UseCUIDKind>(UseCUIDStr) 2951 .Case("hash", CUID_Hash) 2952 .Case("random", CUID_Random) 2953 .Case("none", CUID_None) 2954 .Default(CUID_Invalid); 2955 if (UseCUID == CUID_Invalid) { 2956 C.getDriver().Diag(diag::err_drv_invalid_value) 2957 << A->getAsString(Args) << UseCUIDStr; 2958 C.setContainsError(); 2959 return true; 2960 } 2961 } 2962 2963 // --offload and --offload-arch options are mutually exclusive. 2964 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 2965 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 2966 options::OPT_no_offload_arch_EQ)) { 2967 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) << "--offload-arch" 2968 << "--offload"; 2969 } 2970 2971 // Collect all cuda_gpu_arch parameters, removing duplicates. 2972 std::set<StringRef> GpuArchs; 2973 bool Error = false; 2974 for (Arg *A : Args) { 2975 if (!(A->getOption().matches(options::OPT_offload_arch_EQ) || 2976 A->getOption().matches(options::OPT_no_offload_arch_EQ))) 2977 continue; 2978 A->claim(); 2979 2980 StringRef ArchStr = A->getValue(); 2981 if (A->getOption().matches(options::OPT_no_offload_arch_EQ) && 2982 ArchStr == "all") { 2983 GpuArchs.clear(); 2984 continue; 2985 } 2986 ArchStr = getCanonicalOffloadArch(ArchStr); 2987 if (ArchStr.empty()) { 2988 Error = true; 2989 } else if (A->getOption().matches(options::OPT_offload_arch_EQ)) 2990 GpuArchs.insert(ArchStr); 2991 else if (A->getOption().matches(options::OPT_no_offload_arch_EQ)) 2992 GpuArchs.erase(ArchStr); 2993 else 2994 llvm_unreachable("Unexpected option."); 2995 } 2996 2997 auto &&ConflictingArchs = getConflictOffloadArchCombination(GpuArchs); 2998 if (ConflictingArchs) { 2999 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 3000 << ConflictingArchs.getValue().first 3001 << ConflictingArchs.getValue().second; 3002 C.setContainsError(); 3003 return true; 3004 } 3005 3006 // Collect list of GPUs remaining in the set. 3007 for (auto Arch : GpuArchs) 3008 GpuArchList.push_back(Arch.data()); 3009 3010 // Default to sm_20 which is the lowest common denominator for 3011 // supported GPUs. sm_20 code should work correctly, if 3012 // suboptimally, on all newer GPUs. 3013 if (GpuArchList.empty()) { 3014 if (ToolChains.front()->getTriple().isSPIRV()) 3015 GpuArchList.push_back(CudaArch::Generic); 3016 else 3017 GpuArchList.push_back(DefaultCudaArch); 3018 } 3019 3020 return Error; 3021 } 3022 }; 3023 3024 /// \brief CUDA action builder. It injects device code in the host backend 3025 /// action. 3026 class CudaActionBuilder final : public CudaActionBuilderBase { 3027 public: 3028 CudaActionBuilder(Compilation &C, DerivedArgList &Args, 3029 const Driver::InputList &Inputs) 3030 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) { 3031 DefaultCudaArch = CudaArch::SM_35; 3032 } 3033 3034 StringRef getCanonicalOffloadArch(StringRef ArchStr) override { 3035 CudaArch Arch = StringToCudaArch(ArchStr); 3036 if (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch)) { 3037 C.getDriver().Diag(clang::diag::err_drv_cuda_bad_gpu_arch) << ArchStr; 3038 return StringRef(); 3039 } 3040 return CudaArchToString(Arch); 3041 } 3042 3043 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3044 getConflictOffloadArchCombination( 3045 const std::set<StringRef> &GpuArchs) override { 3046 return llvm::None; 3047 } 3048 3049 ActionBuilderReturnCode 3050 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3051 phases::ID CurPhase, phases::ID FinalPhase, 3052 PhasesTy &Phases) override { 3053 if (!IsActive) 3054 return ABRT_Inactive; 3055 3056 // If we don't have more CUDA actions, we don't have any dependences to 3057 // create for the host. 3058 if (CudaDeviceActions.empty()) 3059 return ABRT_Success; 3060 3061 assert(CudaDeviceActions.size() == GpuArchList.size() && 3062 "Expecting one action per GPU architecture."); 3063 assert(!CompileHostOnly && 3064 "Not expecting CUDA actions in host-only compilation."); 3065 3066 // If we are generating code for the device or we are in a backend phase, 3067 // we attempt to generate the fat binary. We compile each arch to ptx and 3068 // assemble to cubin, then feed the cubin *and* the ptx into a device 3069 // "link" action, which uses fatbinary to combine these cubins into one 3070 // fatbin. The fatbin is then an input to the host action if not in 3071 // device-only mode. 3072 if (CompileDeviceOnly || CurPhase == phases::Backend) { 3073 ActionList DeviceActions; 3074 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3075 // Produce the device action from the current phase up to the assemble 3076 // phase. 3077 for (auto Ph : Phases) { 3078 // Skip the phases that were already dealt with. 3079 if (Ph < CurPhase) 3080 continue; 3081 // We have to be consistent with the host final phase. 3082 if (Ph > FinalPhase) 3083 break; 3084 3085 CudaDeviceActions[I] = C.getDriver().ConstructPhaseAction( 3086 C, Args, Ph, CudaDeviceActions[I], Action::OFK_Cuda); 3087 3088 if (Ph == phases::Assemble) 3089 break; 3090 } 3091 3092 // If we didn't reach the assemble phase, we can't generate the fat 3093 // binary. We don't need to generate the fat binary if we are not in 3094 // device-only mode. 3095 if (!isa<AssembleJobAction>(CudaDeviceActions[I]) || 3096 CompileDeviceOnly) 3097 continue; 3098 3099 Action *AssembleAction = CudaDeviceActions[I]; 3100 assert(AssembleAction->getType() == types::TY_Object); 3101 assert(AssembleAction->getInputs().size() == 1); 3102 3103 Action *BackendAction = AssembleAction->getInputs()[0]; 3104 assert(BackendAction->getType() == types::TY_PP_Asm); 3105 3106 for (auto &A : {AssembleAction, BackendAction}) { 3107 OffloadAction::DeviceDependences DDep; 3108 DDep.add(*A, *ToolChains.front(), GpuArchList[I], Action::OFK_Cuda); 3109 DeviceActions.push_back( 3110 C.MakeAction<OffloadAction>(DDep, A->getType())); 3111 } 3112 } 3113 3114 // We generate the fat binary if we have device input actions. 3115 if (!DeviceActions.empty()) { 3116 CudaFatBinary = 3117 C.MakeAction<LinkJobAction>(DeviceActions, types::TY_CUDA_FATBIN); 3118 3119 if (!CompileDeviceOnly) { 3120 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3121 Action::OFK_Cuda); 3122 // Clear the fat binary, it is already a dependence to an host 3123 // action. 3124 CudaFatBinary = nullptr; 3125 } 3126 3127 // Remove the CUDA actions as they are already connected to an host 3128 // action or fat binary. 3129 CudaDeviceActions.clear(); 3130 } 3131 3132 // We avoid creating host action in device-only mode. 3133 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3134 } else if (CurPhase > phases::Backend) { 3135 // If we are past the backend phase and still have a device action, we 3136 // don't have to do anything as this action is already a device 3137 // top-level action. 3138 return ABRT_Success; 3139 } 3140 3141 assert(CurPhase < phases::Backend && "Generating single CUDA " 3142 "instructions should only occur " 3143 "before the backend phase!"); 3144 3145 // By default, we produce an action for each device arch. 3146 for (Action *&A : CudaDeviceActions) 3147 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3148 3149 return ABRT_Success; 3150 } 3151 }; 3152 /// \brief HIP action builder. It injects device code in the host backend 3153 /// action. 3154 class HIPActionBuilder final : public CudaActionBuilderBase { 3155 /// The linker inputs obtained for each device arch. 3156 SmallVector<ActionList, 8> DeviceLinkerInputs; 3157 // The default bundling behavior depends on the type of output, therefore 3158 // BundleOutput needs to be tri-value: None, true, or false. 3159 // Bundle code objects except --no-gpu-output is specified for device 3160 // only compilation. Bundle other type of output files only if 3161 // --gpu-bundle-output is specified for device only compilation. 3162 Optional<bool> BundleOutput; 3163 3164 public: 3165 HIPActionBuilder(Compilation &C, DerivedArgList &Args, 3166 const Driver::InputList &Inputs) 3167 : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) { 3168 DefaultCudaArch = CudaArch::GFX803; 3169 if (Args.hasArg(options::OPT_gpu_bundle_output, 3170 options::OPT_no_gpu_bundle_output)) 3171 BundleOutput = Args.hasFlag(options::OPT_gpu_bundle_output, 3172 options::OPT_no_gpu_bundle_output, true); 3173 } 3174 3175 bool canUseBundlerUnbundler() const override { return true; } 3176 3177 StringRef getCanonicalOffloadArch(StringRef IdStr) override { 3178 llvm::StringMap<bool> Features; 3179 // getHIPOffloadTargetTriple() is known to return valid value as it has 3180 // been called successfully in the CreateOffloadingDeviceToolChains(). 3181 auto ArchStr = parseTargetID( 3182 *getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()), IdStr, 3183 &Features); 3184 if (!ArchStr) { 3185 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << IdStr; 3186 C.setContainsError(); 3187 return StringRef(); 3188 } 3189 auto CanId = getCanonicalTargetID(ArchStr.getValue(), Features); 3190 return Args.MakeArgStringRef(CanId); 3191 }; 3192 3193 llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 3194 getConflictOffloadArchCombination( 3195 const std::set<StringRef> &GpuArchs) override { 3196 return getConflictTargetIDCombination(GpuArchs); 3197 } 3198 3199 ActionBuilderReturnCode 3200 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3201 phases::ID CurPhase, phases::ID FinalPhase, 3202 PhasesTy &Phases) override { 3203 if (!IsActive) 3204 return ABRT_Inactive; 3205 3206 // amdgcn does not support linking of object files, therefore we skip 3207 // backend and assemble phases to output LLVM IR. Except for generating 3208 // non-relocatable device code, where we generate fat binary for device 3209 // code and pass to host in Backend phase. 3210 if (CudaDeviceActions.empty()) 3211 return ABRT_Success; 3212 3213 assert(((CurPhase == phases::Link && Relocatable) || 3214 CudaDeviceActions.size() == GpuArchList.size()) && 3215 "Expecting one action per GPU architecture."); 3216 assert(!CompileHostOnly && 3217 "Not expecting HIP actions in host-only compilation."); 3218 3219 if (!Relocatable && CurPhase == phases::Backend && !EmitLLVM && 3220 !EmitAsm) { 3221 // If we are in backend phase, we attempt to generate the fat binary. 3222 // We compile each arch to IR and use a link action to generate code 3223 // object containing ISA. Then we use a special "link" action to create 3224 // a fat binary containing all the code objects for different GPU's. 3225 // The fat binary is then an input to the host action. 3226 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3227 if (C.getDriver().isUsingLTO(/*IsOffload=*/true)) { 3228 // When LTO is enabled, skip the backend and assemble phases and 3229 // use lld to link the bitcode. 3230 ActionList AL; 3231 AL.push_back(CudaDeviceActions[I]); 3232 // Create a link action to link device IR with device library 3233 // and generate ISA. 3234 CudaDeviceActions[I] = 3235 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3236 } else { 3237 // When LTO is not enabled, we follow the conventional 3238 // compiler phases, including backend and assemble phases. 3239 ActionList AL; 3240 Action *BackendAction = nullptr; 3241 if (ToolChains.front()->getTriple().isSPIRV()) { 3242 // Emit LLVM bitcode for SPIR-V targets. SPIR-V device tool chain 3243 // (HIPSPVToolChain) runs post-link LLVM IR passes. 3244 types::ID Output = Args.hasArg(options::OPT_S) 3245 ? types::TY_LLVM_IR 3246 : types::TY_LLVM_BC; 3247 BackendAction = 3248 C.MakeAction<BackendJobAction>(CudaDeviceActions[I], Output); 3249 } else 3250 BackendAction = C.getDriver().ConstructPhaseAction( 3251 C, Args, phases::Backend, CudaDeviceActions[I], 3252 AssociatedOffloadKind); 3253 auto AssembleAction = C.getDriver().ConstructPhaseAction( 3254 C, Args, phases::Assemble, BackendAction, 3255 AssociatedOffloadKind); 3256 AL.push_back(AssembleAction); 3257 // Create a link action to link device IR with device library 3258 // and generate ISA. 3259 CudaDeviceActions[I] = 3260 C.MakeAction<LinkJobAction>(AL, types::TY_Image); 3261 } 3262 3263 // OffloadingActionBuilder propagates device arch until an offload 3264 // action. Since the next action for creating fatbin does 3265 // not have device arch, whereas the above link action and its input 3266 // have device arch, an offload action is needed to stop the null 3267 // device arch of the next action being propagated to the above link 3268 // action. 3269 OffloadAction::DeviceDependences DDep; 3270 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3271 AssociatedOffloadKind); 3272 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3273 DDep, CudaDeviceActions[I]->getType()); 3274 } 3275 3276 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3277 BundleOutput.getValue()) { 3278 // Create HIP fat binary with a special "link" action. 3279 CudaFatBinary = C.MakeAction<LinkJobAction>(CudaDeviceActions, 3280 types::TY_HIP_FATBIN); 3281 3282 if (!CompileDeviceOnly) { 3283 DA.add(*CudaFatBinary, *ToolChains.front(), /*BoundArch=*/nullptr, 3284 AssociatedOffloadKind); 3285 // Clear the fat binary, it is already a dependence to an host 3286 // action. 3287 CudaFatBinary = nullptr; 3288 } 3289 3290 // Remove the CUDA actions as they are already connected to an host 3291 // action or fat binary. 3292 CudaDeviceActions.clear(); 3293 } 3294 3295 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3296 } else if (CurPhase == phases::Link) { 3297 // Save CudaDeviceActions to DeviceLinkerInputs for each GPU subarch. 3298 // This happens to each device action originated from each input file. 3299 // Later on, device actions in DeviceLinkerInputs are used to create 3300 // device link actions in appendLinkDependences and the created device 3301 // link actions are passed to the offload action as device dependence. 3302 DeviceLinkerInputs.resize(CudaDeviceActions.size()); 3303 auto LI = DeviceLinkerInputs.begin(); 3304 for (auto *A : CudaDeviceActions) { 3305 LI->push_back(A); 3306 ++LI; 3307 } 3308 3309 // We will pass the device action as a host dependence, so we don't 3310 // need to do anything else with them. 3311 CudaDeviceActions.clear(); 3312 return CompileDeviceOnly ? ABRT_Ignore_Host : ABRT_Success; 3313 } 3314 3315 // By default, we produce an action for each device arch. 3316 for (Action *&A : CudaDeviceActions) 3317 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A, 3318 AssociatedOffloadKind); 3319 3320 if (CompileDeviceOnly && CurPhase == FinalPhase && 3321 BundleOutput.hasValue() && BundleOutput.getValue()) { 3322 for (unsigned I = 0, E = GpuArchList.size(); I != E; ++I) { 3323 OffloadAction::DeviceDependences DDep; 3324 DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I], 3325 AssociatedOffloadKind); 3326 CudaDeviceActions[I] = C.MakeAction<OffloadAction>( 3327 DDep, CudaDeviceActions[I]->getType()); 3328 } 3329 CudaFatBinary = 3330 C.MakeAction<OffloadBundlingJobAction>(CudaDeviceActions); 3331 CudaDeviceActions.clear(); 3332 } 3333 3334 return (CompileDeviceOnly && CurPhase == FinalPhase) ? ABRT_Ignore_Host 3335 : ABRT_Success; 3336 } 3337 3338 void appendLinkDeviceActions(ActionList &AL) override { 3339 if (DeviceLinkerInputs.size() == 0) 3340 return; 3341 3342 assert(DeviceLinkerInputs.size() == GpuArchList.size() && 3343 "Linker inputs and GPU arch list sizes do not match."); 3344 3345 ActionList Actions; 3346 unsigned I = 0; 3347 // Append a new link action for each device. 3348 // Each entry in DeviceLinkerInputs corresponds to a GPU arch. 3349 for (auto &LI : DeviceLinkerInputs) { 3350 3351 types::ID Output = Args.hasArg(options::OPT_emit_llvm) 3352 ? types::TY_LLVM_BC 3353 : types::TY_Image; 3354 3355 auto *DeviceLinkAction = C.MakeAction<LinkJobAction>(LI, Output); 3356 // Linking all inputs for the current GPU arch. 3357 // LI contains all the inputs for the linker. 3358 OffloadAction::DeviceDependences DeviceLinkDeps; 3359 DeviceLinkDeps.add(*DeviceLinkAction, *ToolChains[0], 3360 GpuArchList[I], AssociatedOffloadKind); 3361 Actions.push_back(C.MakeAction<OffloadAction>( 3362 DeviceLinkDeps, DeviceLinkAction->getType())); 3363 ++I; 3364 } 3365 DeviceLinkerInputs.clear(); 3366 3367 // If emitting LLVM, do not generate final host/device compilation action 3368 if (Args.hasArg(options::OPT_emit_llvm)) { 3369 AL.append(Actions); 3370 return; 3371 } 3372 3373 // Create a host object from all the device images by embedding them 3374 // in a fat binary for mixed host-device compilation. For device-only 3375 // compilation, creates a fat binary. 3376 OffloadAction::DeviceDependences DDeps; 3377 if (!CompileDeviceOnly || !BundleOutput.hasValue() || 3378 BundleOutput.getValue()) { 3379 auto *TopDeviceLinkAction = C.MakeAction<LinkJobAction>( 3380 Actions, 3381 CompileDeviceOnly ? types::TY_HIP_FATBIN : types::TY_Object); 3382 DDeps.add(*TopDeviceLinkAction, *ToolChains[0], nullptr, 3383 AssociatedOffloadKind); 3384 // Offload the host object to the host linker. 3385 AL.push_back( 3386 C.MakeAction<OffloadAction>(DDeps, TopDeviceLinkAction->getType())); 3387 } else { 3388 AL.append(Actions); 3389 } 3390 } 3391 3392 Action* appendLinkHostActions(ActionList &AL) override { return AL.back(); } 3393 3394 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3395 }; 3396 3397 /// OpenMP action builder. The host bitcode is passed to the device frontend 3398 /// and all the device linked images are passed to the host link phase. 3399 class OpenMPActionBuilder final : public DeviceActionBuilder { 3400 /// The OpenMP actions for the current input. 3401 ActionList OpenMPDeviceActions; 3402 3403 /// The linker inputs obtained for each toolchain. 3404 SmallVector<ActionList, 8> DeviceLinkerInputs; 3405 3406 public: 3407 OpenMPActionBuilder(Compilation &C, DerivedArgList &Args, 3408 const Driver::InputList &Inputs) 3409 : DeviceActionBuilder(C, Args, Inputs, Action::OFK_OpenMP) {} 3410 3411 ActionBuilderReturnCode 3412 getDeviceDependences(OffloadAction::DeviceDependences &DA, 3413 phases::ID CurPhase, phases::ID FinalPhase, 3414 PhasesTy &Phases) override { 3415 if (OpenMPDeviceActions.empty()) 3416 return ABRT_Inactive; 3417 3418 // We should always have an action for each input. 3419 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3420 "Number of OpenMP actions and toolchains do not match."); 3421 3422 // The host only depends on device action in the linking phase, when all 3423 // the device images have to be embedded in the host image. 3424 if (CurPhase == phases::Link) { 3425 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3426 "Toolchains and linker inputs sizes do not match."); 3427 auto LI = DeviceLinkerInputs.begin(); 3428 for (auto *A : OpenMPDeviceActions) { 3429 LI->push_back(A); 3430 ++LI; 3431 } 3432 3433 // We passed the device action as a host dependence, so we don't need to 3434 // do anything else with them. 3435 OpenMPDeviceActions.clear(); 3436 return ABRT_Success; 3437 } 3438 3439 // By default, we produce an action for each device arch. 3440 for (Action *&A : OpenMPDeviceActions) 3441 A = C.getDriver().ConstructPhaseAction(C, Args, CurPhase, A); 3442 3443 return ABRT_Success; 3444 } 3445 3446 ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override { 3447 3448 // If this is an input action replicate it for each OpenMP toolchain. 3449 if (auto *IA = dyn_cast<InputAction>(HostAction)) { 3450 OpenMPDeviceActions.clear(); 3451 for (unsigned I = 0; I < ToolChains.size(); ++I) 3452 OpenMPDeviceActions.push_back( 3453 C.MakeAction<InputAction>(IA->getInputArg(), IA->getType())); 3454 return ABRT_Success; 3455 } 3456 3457 // If this is an unbundling action use it as is for each OpenMP toolchain. 3458 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) { 3459 OpenMPDeviceActions.clear(); 3460 auto *IA = cast<InputAction>(UA->getInputs().back()); 3461 std::string FileName = IA->getInputArg().getAsString(Args); 3462 // Check if the type of the file is the same as the action. Do not 3463 // unbundle it if it is not. Do not unbundle .so files, for example, 3464 // which are not object files. 3465 if (IA->getType() == types::TY_Object && 3466 (!llvm::sys::path::has_extension(FileName) || 3467 types::lookupTypeForExtension( 3468 llvm::sys::path::extension(FileName).drop_front()) != 3469 types::TY_Object)) 3470 return ABRT_Inactive; 3471 for (unsigned I = 0; I < ToolChains.size(); ++I) { 3472 OpenMPDeviceActions.push_back(UA); 3473 UA->registerDependentActionInfo( 3474 ToolChains[I], /*BoundArch=*/StringRef(), Action::OFK_OpenMP); 3475 } 3476 return ABRT_Success; 3477 } 3478 3479 // When generating code for OpenMP we use the host compile phase result as 3480 // a dependence to the device compile phase so that it can learn what 3481 // declarations should be emitted. However, this is not the only use for 3482 // the host action, so we prevent it from being collapsed. 3483 if (isa<CompileJobAction>(HostAction)) { 3484 HostAction->setCannotBeCollapsedWithNextDependentAction(); 3485 assert(ToolChains.size() == OpenMPDeviceActions.size() && 3486 "Toolchains and device action sizes do not match."); 3487 OffloadAction::HostDependence HDep( 3488 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3489 /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3490 auto TC = ToolChains.begin(); 3491 for (Action *&A : OpenMPDeviceActions) { 3492 assert(isa<CompileJobAction>(A)); 3493 OffloadAction::DeviceDependences DDep; 3494 DDep.add(*A, **TC, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3495 A = C.MakeAction<OffloadAction>(HDep, DDep); 3496 ++TC; 3497 } 3498 } 3499 return ABRT_Success; 3500 } 3501 3502 void appendTopLevelActions(ActionList &AL) override { 3503 if (OpenMPDeviceActions.empty()) 3504 return; 3505 3506 // We should always have an action for each input. 3507 assert(OpenMPDeviceActions.size() == ToolChains.size() && 3508 "Number of OpenMP actions and toolchains do not match."); 3509 3510 // Append all device actions followed by the proper offload action. 3511 auto TI = ToolChains.begin(); 3512 for (auto *A : OpenMPDeviceActions) { 3513 OffloadAction::DeviceDependences Dep; 3514 Dep.add(*A, **TI, /*BoundArch=*/nullptr, Action::OFK_OpenMP); 3515 AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType())); 3516 ++TI; 3517 } 3518 // We no longer need the action stored in this builder. 3519 OpenMPDeviceActions.clear(); 3520 } 3521 3522 void appendLinkDeviceActions(ActionList &AL) override { 3523 assert(ToolChains.size() == DeviceLinkerInputs.size() && 3524 "Toolchains and linker inputs sizes do not match."); 3525 3526 // Append a new link action for each device. 3527 auto TC = ToolChains.begin(); 3528 for (auto &LI : DeviceLinkerInputs) { 3529 auto *DeviceLinkAction = 3530 C.MakeAction<LinkJobAction>(LI, types::TY_Image); 3531 OffloadAction::DeviceDependences DeviceLinkDeps; 3532 DeviceLinkDeps.add(*DeviceLinkAction, **TC, /*BoundArch=*/nullptr, 3533 Action::OFK_OpenMP); 3534 AL.push_back(C.MakeAction<OffloadAction>(DeviceLinkDeps, 3535 DeviceLinkAction->getType())); 3536 ++TC; 3537 } 3538 DeviceLinkerInputs.clear(); 3539 } 3540 3541 Action* appendLinkHostActions(ActionList &AL) override { 3542 // Create wrapper bitcode from the result of device link actions and compile 3543 // it to an object which will be added to the host link command. 3544 auto *BC = C.MakeAction<OffloadWrapperJobAction>(AL, types::TY_LLVM_BC); 3545 auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm); 3546 return C.MakeAction<AssembleJobAction>(ASM, types::TY_Object); 3547 } 3548 3549 void appendLinkDependences(OffloadAction::DeviceDependences &DA) override {} 3550 3551 bool initialize() override { 3552 // Get the OpenMP toolchains. If we don't get any, the action builder will 3553 // know there is nothing to do related to OpenMP offloading. 3554 auto OpenMPTCRange = C.getOffloadToolChains<Action::OFK_OpenMP>(); 3555 for (auto TI = OpenMPTCRange.first, TE = OpenMPTCRange.second; TI != TE; 3556 ++TI) 3557 ToolChains.push_back(TI->second); 3558 3559 DeviceLinkerInputs.resize(ToolChains.size()); 3560 return false; 3561 } 3562 3563 bool canUseBundlerUnbundler() const override { 3564 // OpenMP should use bundled files whenever possible. 3565 return true; 3566 } 3567 }; 3568 3569 /// 3570 /// TODO: Add the implementation for other specialized builders here. 3571 /// 3572 3573 /// Specialized builders being used by this offloading action builder. 3574 SmallVector<DeviceActionBuilder *, 4> SpecializedBuilders; 3575 3576 /// Flag set to true if all valid builders allow file bundling/unbundling. 3577 bool CanUseBundler; 3578 3579 public: 3580 OffloadingActionBuilder(Compilation &C, DerivedArgList &Args, 3581 const Driver::InputList &Inputs) 3582 : C(C) { 3583 // Create a specialized builder for each device toolchain. 3584 3585 IsValid = true; 3586 3587 // Create a specialized builder for CUDA. 3588 SpecializedBuilders.push_back(new CudaActionBuilder(C, Args, Inputs)); 3589 3590 // Create a specialized builder for HIP. 3591 SpecializedBuilders.push_back(new HIPActionBuilder(C, Args, Inputs)); 3592 3593 // Create a specialized builder for OpenMP. 3594 SpecializedBuilders.push_back(new OpenMPActionBuilder(C, Args, Inputs)); 3595 3596 // 3597 // TODO: Build other specialized builders here. 3598 // 3599 3600 // Initialize all the builders, keeping track of errors. If all valid 3601 // builders agree that we can use bundling, set the flag to true. 3602 unsigned ValidBuilders = 0u; 3603 unsigned ValidBuildersSupportingBundling = 0u; 3604 for (auto *SB : SpecializedBuilders) { 3605 IsValid = IsValid && !SB->initialize(); 3606 3607 // Update the counters if the builder is valid. 3608 if (SB->isValid()) { 3609 ++ValidBuilders; 3610 if (SB->canUseBundlerUnbundler()) 3611 ++ValidBuildersSupportingBundling; 3612 } 3613 } 3614 CanUseBundler = 3615 ValidBuilders && ValidBuilders == ValidBuildersSupportingBundling; 3616 } 3617 3618 ~OffloadingActionBuilder() { 3619 for (auto *SB : SpecializedBuilders) 3620 delete SB; 3621 } 3622 3623 /// Record a host action and its originating input argument. 3624 void recordHostAction(Action *HostAction, const Arg *InputArg) { 3625 assert(HostAction && "Invalid host action"); 3626 assert(InputArg && "Invalid input argument"); 3627 auto Loc = HostActionToInputArgMap.find(HostAction); 3628 if (Loc == HostActionToInputArgMap.end()) 3629 HostActionToInputArgMap[HostAction] = InputArg; 3630 assert(HostActionToInputArgMap[HostAction] == InputArg && 3631 "host action mapped to multiple input arguments"); 3632 } 3633 3634 /// Generate an action that adds device dependences (if any) to a host action. 3635 /// If no device dependence actions exist, just return the host action \a 3636 /// HostAction. If an error is found or if no builder requires the host action 3637 /// to be generated, return nullptr. 3638 Action * 3639 addDeviceDependencesToHostAction(Action *HostAction, const Arg *InputArg, 3640 phases::ID CurPhase, phases::ID FinalPhase, 3641 DeviceActionBuilder::PhasesTy &Phases) { 3642 if (!IsValid) 3643 return nullptr; 3644 3645 if (SpecializedBuilders.empty()) 3646 return HostAction; 3647 3648 assert(HostAction && "Invalid host action!"); 3649 recordHostAction(HostAction, InputArg); 3650 3651 OffloadAction::DeviceDependences DDeps; 3652 // Check if all the programming models agree we should not emit the host 3653 // action. Also, keep track of the offloading kinds employed. 3654 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3655 unsigned InactiveBuilders = 0u; 3656 unsigned IgnoringBuilders = 0u; 3657 for (auto *SB : SpecializedBuilders) { 3658 if (!SB->isValid()) { 3659 ++InactiveBuilders; 3660 continue; 3661 } 3662 3663 auto RetCode = 3664 SB->getDeviceDependences(DDeps, CurPhase, FinalPhase, Phases); 3665 3666 // If the builder explicitly says the host action should be ignored, 3667 // we need to increment the variable that tracks the builders that request 3668 // the host object to be ignored. 3669 if (RetCode == DeviceActionBuilder::ABRT_Ignore_Host) 3670 ++IgnoringBuilders; 3671 3672 // Unless the builder was inactive for this action, we have to record the 3673 // offload kind because the host will have to use it. 3674 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3675 OffloadKind |= SB->getAssociatedOffloadKind(); 3676 } 3677 3678 // If all builders agree that the host object should be ignored, just return 3679 // nullptr. 3680 if (IgnoringBuilders && 3681 SpecializedBuilders.size() == (InactiveBuilders + IgnoringBuilders)) 3682 return nullptr; 3683 3684 if (DDeps.getActions().empty()) 3685 return HostAction; 3686 3687 // We have dependences we need to bundle together. We use an offload action 3688 // for that. 3689 OffloadAction::HostDependence HDep( 3690 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3691 /*BoundArch=*/nullptr, DDeps); 3692 return C.MakeAction<OffloadAction>(HDep, DDeps); 3693 } 3694 3695 /// Generate an action that adds a host dependence to a device action. The 3696 /// results will be kept in this action builder. Return true if an error was 3697 /// found. 3698 bool addHostDependenceToDeviceActions(Action *&HostAction, 3699 const Arg *InputArg) { 3700 if (!IsValid) 3701 return true; 3702 3703 recordHostAction(HostAction, InputArg); 3704 3705 // If we are supporting bundling/unbundling and the current action is an 3706 // input action of non-source file, we replace the host action by the 3707 // unbundling action. The bundler tool has the logic to detect if an input 3708 // is a bundle or not and if the input is not a bundle it assumes it is a 3709 // host file. Therefore it is safe to create an unbundling action even if 3710 // the input is not a bundle. 3711 if (CanUseBundler && isa<InputAction>(HostAction) && 3712 InputArg->getOption().getKind() == llvm::opt::Option::InputClass && 3713 (!types::isSrcFile(HostAction->getType()) || 3714 HostAction->getType() == types::TY_PP_HIP)) { 3715 auto UnbundlingHostAction = 3716 C.MakeAction<OffloadUnbundlingJobAction>(HostAction); 3717 UnbundlingHostAction->registerDependentActionInfo( 3718 C.getSingleOffloadToolChain<Action::OFK_Host>(), 3719 /*BoundArch=*/StringRef(), Action::OFK_Host); 3720 HostAction = UnbundlingHostAction; 3721 recordHostAction(HostAction, InputArg); 3722 } 3723 3724 assert(HostAction && "Invalid host action!"); 3725 3726 // Register the offload kinds that are used. 3727 auto &OffloadKind = InputArgToOffloadKindMap[InputArg]; 3728 for (auto *SB : SpecializedBuilders) { 3729 if (!SB->isValid()) 3730 continue; 3731 3732 auto RetCode = SB->addDeviceDepences(HostAction); 3733 3734 // Host dependences for device actions are not compatible with that same 3735 // action being ignored. 3736 assert(RetCode != DeviceActionBuilder::ABRT_Ignore_Host && 3737 "Host dependence not expected to be ignored.!"); 3738 3739 // Unless the builder was inactive for this action, we have to record the 3740 // offload kind because the host will have to use it. 3741 if (RetCode != DeviceActionBuilder::ABRT_Inactive) 3742 OffloadKind |= SB->getAssociatedOffloadKind(); 3743 } 3744 3745 // Do not use unbundler if the Host does not depend on device action. 3746 if (OffloadKind == Action::OFK_None && CanUseBundler) 3747 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction)) 3748 HostAction = UA->getInputs().back(); 3749 3750 return false; 3751 } 3752 3753 /// Add the offloading top level actions to the provided action list. This 3754 /// function can replace the host action by a bundling action if the 3755 /// programming models allow it. 3756 bool appendTopLevelActions(ActionList &AL, Action *HostAction, 3757 const Arg *InputArg) { 3758 if (HostAction) 3759 recordHostAction(HostAction, InputArg); 3760 3761 // Get the device actions to be appended. 3762 ActionList OffloadAL; 3763 for (auto *SB : SpecializedBuilders) { 3764 if (!SB->isValid()) 3765 continue; 3766 SB->appendTopLevelActions(OffloadAL); 3767 } 3768 3769 // If we can use the bundler, replace the host action by the bundling one in 3770 // the resulting list. Otherwise, just append the device actions. For 3771 // device only compilation, HostAction is a null pointer, therefore only do 3772 // this when HostAction is not a null pointer. 3773 if (CanUseBundler && HostAction && 3774 HostAction->getType() != types::TY_Nothing && !OffloadAL.empty()) { 3775 // Add the host action to the list in order to create the bundling action. 3776 OffloadAL.push_back(HostAction); 3777 3778 // We expect that the host action was just appended to the action list 3779 // before this method was called. 3780 assert(HostAction == AL.back() && "Host action not in the list??"); 3781 HostAction = C.MakeAction<OffloadBundlingJobAction>(OffloadAL); 3782 recordHostAction(HostAction, InputArg); 3783 AL.back() = HostAction; 3784 } else 3785 AL.append(OffloadAL.begin(), OffloadAL.end()); 3786 3787 // Propagate to the current host action (if any) the offload information 3788 // associated with the current input. 3789 if (HostAction) 3790 HostAction->propagateHostOffloadInfo(InputArgToOffloadKindMap[InputArg], 3791 /*BoundArch=*/nullptr); 3792 return false; 3793 } 3794 3795 void appendDeviceLinkActions(ActionList &AL) { 3796 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3797 if (!SB->isValid()) 3798 continue; 3799 SB->appendLinkDeviceActions(AL); 3800 } 3801 } 3802 3803 Action *makeHostLinkAction() { 3804 // Build a list of device linking actions. 3805 ActionList DeviceAL; 3806 appendDeviceLinkActions(DeviceAL); 3807 if (DeviceAL.empty()) 3808 return nullptr; 3809 3810 // Let builders add host linking actions. 3811 Action* HA = nullptr; 3812 for (DeviceActionBuilder *SB : SpecializedBuilders) { 3813 if (!SB->isValid()) 3814 continue; 3815 HA = SB->appendLinkHostActions(DeviceAL); 3816 // This created host action has no originating input argument, therefore 3817 // needs to set its offloading kind directly. 3818 if (HA) 3819 HA->propagateHostOffloadInfo(SB->getAssociatedOffloadKind(), 3820 /*BoundArch=*/nullptr); 3821 } 3822 return HA; 3823 } 3824 3825 /// Processes the host linker action. This currently consists of replacing it 3826 /// with an offload action if there are device link objects and propagate to 3827 /// the host action all the offload kinds used in the current compilation. The 3828 /// resulting action is returned. 3829 Action *processHostLinkAction(Action *HostAction) { 3830 // Add all the dependences from the device linking actions. 3831 OffloadAction::DeviceDependences DDeps; 3832 for (auto *SB : SpecializedBuilders) { 3833 if (!SB->isValid()) 3834 continue; 3835 3836 SB->appendLinkDependences(DDeps); 3837 } 3838 3839 // Calculate all the offload kinds used in the current compilation. 3840 unsigned ActiveOffloadKinds = 0u; 3841 for (auto &I : InputArgToOffloadKindMap) 3842 ActiveOffloadKinds |= I.second; 3843 3844 // If we don't have device dependencies, we don't have to create an offload 3845 // action. 3846 if (DDeps.getActions().empty()) { 3847 // Set all the active offloading kinds to the link action. Given that it 3848 // is a link action it is assumed to depend on all actions generated so 3849 // far. 3850 HostAction->setHostOffloadInfo(ActiveOffloadKinds, 3851 /*BoundArch=*/nullptr); 3852 // Propagate active offloading kinds for each input to the link action. 3853 // Each input may have different active offloading kind. 3854 for (auto A : HostAction->inputs()) { 3855 auto ArgLoc = HostActionToInputArgMap.find(A); 3856 if (ArgLoc == HostActionToInputArgMap.end()) 3857 continue; 3858 auto OFKLoc = InputArgToOffloadKindMap.find(ArgLoc->second); 3859 if (OFKLoc == InputArgToOffloadKindMap.end()) 3860 continue; 3861 A->propagateHostOffloadInfo(OFKLoc->second, /*BoundArch=*/nullptr); 3862 } 3863 return HostAction; 3864 } 3865 3866 // Create the offload action with all dependences. When an offload action 3867 // is created the kinds are propagated to the host action, so we don't have 3868 // to do that explicitly here. 3869 OffloadAction::HostDependence HDep( 3870 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 3871 /*BoundArch*/ nullptr, ActiveOffloadKinds); 3872 return C.MakeAction<OffloadAction>(HDep, DDeps); 3873 } 3874 }; 3875 } // anonymous namespace. 3876 3877 void Driver::handleArguments(Compilation &C, DerivedArgList &Args, 3878 const InputList &Inputs, 3879 ActionList &Actions) const { 3880 3881 // Ignore /Yc/Yu if both /Yc and /Yu passed but with different filenames. 3882 Arg *YcArg = Args.getLastArg(options::OPT__SLASH_Yc); 3883 Arg *YuArg = Args.getLastArg(options::OPT__SLASH_Yu); 3884 if (YcArg && YuArg && strcmp(YcArg->getValue(), YuArg->getValue()) != 0) { 3885 Diag(clang::diag::warn_drv_ycyu_different_arg_clang_cl); 3886 Args.eraseArg(options::OPT__SLASH_Yc); 3887 Args.eraseArg(options::OPT__SLASH_Yu); 3888 YcArg = YuArg = nullptr; 3889 } 3890 if (YcArg && Inputs.size() > 1) { 3891 Diag(clang::diag::warn_drv_yc_multiple_inputs_clang_cl); 3892 Args.eraseArg(options::OPT__SLASH_Yc); 3893 YcArg = nullptr; 3894 } 3895 3896 Arg *FinalPhaseArg; 3897 phases::ID FinalPhase = getFinalPhase(Args, &FinalPhaseArg); 3898 3899 if (FinalPhase == phases::Link) { 3900 // Emitting LLVM while linking disabled except in HIPAMD Toolchain 3901 if (Args.hasArg(options::OPT_emit_llvm) && !Args.hasArg(options::OPT_hip_link)) 3902 Diag(clang::diag::err_drv_emit_llvm_link); 3903 if (IsCLMode() && LTOMode != LTOK_None && 3904 !Args.getLastArgValue(options::OPT_fuse_ld_EQ) 3905 .equals_insensitive("lld")) 3906 Diag(clang::diag::err_drv_lto_without_lld); 3907 } 3908 3909 if (FinalPhase == phases::Preprocess || Args.hasArg(options::OPT__SLASH_Y_)) { 3910 // If only preprocessing or /Y- is used, all pch handling is disabled. 3911 // Rather than check for it everywhere, just remove clang-cl pch-related 3912 // flags here. 3913 Args.eraseArg(options::OPT__SLASH_Fp); 3914 Args.eraseArg(options::OPT__SLASH_Yc); 3915 Args.eraseArg(options::OPT__SLASH_Yu); 3916 YcArg = YuArg = nullptr; 3917 } 3918 3919 unsigned LastPLSize = 0; 3920 for (auto &I : Inputs) { 3921 types::ID InputType = I.first; 3922 const Arg *InputArg = I.second; 3923 3924 auto PL = types::getCompilationPhases(InputType); 3925 LastPLSize = PL.size(); 3926 3927 // If the first step comes after the final phase we are doing as part of 3928 // this compilation, warn the user about it. 3929 phases::ID InitialPhase = PL[0]; 3930 if (InitialPhase > FinalPhase) { 3931 if (InputArg->isClaimed()) 3932 continue; 3933 3934 // Claim here to avoid the more general unused warning. 3935 InputArg->claim(); 3936 3937 // Suppress all unused style warnings with -Qunused-arguments 3938 if (Args.hasArg(options::OPT_Qunused_arguments)) 3939 continue; 3940 3941 // Special case when final phase determined by binary name, rather than 3942 // by a command-line argument with a corresponding Arg. 3943 if (CCCIsCPP()) 3944 Diag(clang::diag::warn_drv_input_file_unused_by_cpp) 3945 << InputArg->getAsString(Args) << getPhaseName(InitialPhase); 3946 // Special case '-E' warning on a previously preprocessed file to make 3947 // more sense. 3948 else if (InitialPhase == phases::Compile && 3949 (Args.getLastArg(options::OPT__SLASH_EP, 3950 options::OPT__SLASH_P) || 3951 Args.getLastArg(options::OPT_E) || 3952 Args.getLastArg(options::OPT_M, options::OPT_MM)) && 3953 getPreprocessedType(InputType) == types::TY_INVALID) 3954 Diag(clang::diag::warn_drv_preprocessed_input_file_unused) 3955 << InputArg->getAsString(Args) << !!FinalPhaseArg 3956 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3957 else 3958 Diag(clang::diag::warn_drv_input_file_unused) 3959 << InputArg->getAsString(Args) << getPhaseName(InitialPhase) 3960 << !!FinalPhaseArg 3961 << (FinalPhaseArg ? FinalPhaseArg->getOption().getName() : ""); 3962 continue; 3963 } 3964 3965 if (YcArg) { 3966 // Add a separate precompile phase for the compile phase. 3967 if (FinalPhase >= phases::Compile) { 3968 const types::ID HeaderType = lookupHeaderTypeForSourceType(InputType); 3969 // Build the pipeline for the pch file. 3970 Action *ClangClPch = C.MakeAction<InputAction>(*InputArg, HeaderType); 3971 for (phases::ID Phase : types::getCompilationPhases(HeaderType)) 3972 ClangClPch = ConstructPhaseAction(C, Args, Phase, ClangClPch); 3973 assert(ClangClPch); 3974 Actions.push_back(ClangClPch); 3975 // The driver currently exits after the first failed command. This 3976 // relies on that behavior, to make sure if the pch generation fails, 3977 // the main compilation won't run. 3978 // FIXME: If the main compilation fails, the PCH generation should 3979 // probably not be considered successful either. 3980 } 3981 } 3982 } 3983 3984 // If we are linking, claim any options which are obviously only used for 3985 // compilation. 3986 // FIXME: Understand why the last Phase List length is used here. 3987 if (FinalPhase == phases::Link && LastPLSize == 1) { 3988 Args.ClaimAllArgs(options::OPT_CompileOnly_Group); 3989 Args.ClaimAllArgs(options::OPT_cl_compile_Group); 3990 } 3991 } 3992 3993 void Driver::BuildActions(Compilation &C, DerivedArgList &Args, 3994 const InputList &Inputs, ActionList &Actions) const { 3995 llvm::PrettyStackTraceString CrashInfo("Building compilation actions"); 3996 3997 if (!SuppressMissingInputWarning && Inputs.empty()) { 3998 Diag(clang::diag::err_drv_no_input_files); 3999 return; 4000 } 4001 4002 // Reject -Z* at the top level, these options should never have been exposed 4003 // by gcc. 4004 if (Arg *A = Args.getLastArg(options::OPT_Z_Joined)) 4005 Diag(clang::diag::err_drv_use_of_Z_option) << A->getAsString(Args); 4006 4007 // Diagnose misuse of /Fo. 4008 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fo)) { 4009 StringRef V = A->getValue(); 4010 if (Inputs.size() > 1 && !V.empty() && 4011 !llvm::sys::path::is_separator(V.back())) { 4012 // Check whether /Fo tries to name an output file for multiple inputs. 4013 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4014 << A->getSpelling() << V; 4015 Args.eraseArg(options::OPT__SLASH_Fo); 4016 } 4017 } 4018 4019 // Diagnose misuse of /Fa. 4020 if (Arg *A = Args.getLastArg(options::OPT__SLASH_Fa)) { 4021 StringRef V = A->getValue(); 4022 if (Inputs.size() > 1 && !V.empty() && 4023 !llvm::sys::path::is_separator(V.back())) { 4024 // Check whether /Fa tries to name an asm file for multiple inputs. 4025 Diag(clang::diag::err_drv_out_file_argument_with_multiple_sources) 4026 << A->getSpelling() << V; 4027 Args.eraseArg(options::OPT__SLASH_Fa); 4028 } 4029 } 4030 4031 // Diagnose misuse of /o. 4032 if (Arg *A = Args.getLastArg(options::OPT__SLASH_o)) { 4033 if (A->getValue()[0] == '\0') { 4034 // It has to have a value. 4035 Diag(clang::diag::err_drv_missing_argument) << A->getSpelling() << 1; 4036 Args.eraseArg(options::OPT__SLASH_o); 4037 } 4038 } 4039 4040 handleArguments(C, Args, Inputs, Actions); 4041 4042 // Builder to be used to build offloading actions. 4043 OffloadingActionBuilder OffloadBuilder(C, Args, Inputs); 4044 4045 bool UseNewOffloadingDriver = 4046 (C.isOffloadingHostKind(Action::OFK_OpenMP) && 4047 Args.hasFlag(options::OPT_fopenmp_new_driver, 4048 options::OPT_no_offload_new_driver, true)) || 4049 Args.hasFlag(options::OPT_offload_new_driver, 4050 options::OPT_no_offload_new_driver, false); 4051 4052 // Construct the actions to perform. 4053 HeaderModulePrecompileJobAction *HeaderModuleAction = nullptr; 4054 ExtractAPIJobAction *ExtractAPIAction = nullptr; 4055 ActionList LinkerInputs; 4056 ActionList MergerInputs; 4057 4058 for (auto &I : Inputs) { 4059 types::ID InputType = I.first; 4060 const Arg *InputArg = I.second; 4061 4062 auto PL = types::getCompilationPhases(*this, Args, InputType); 4063 if (PL.empty()) 4064 continue; 4065 4066 auto FullPL = types::getCompilationPhases(InputType); 4067 4068 // Build the pipeline for this file. 4069 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4070 4071 // Use the current host action in any of the offloading actions, if 4072 // required. 4073 if (!UseNewOffloadingDriver) 4074 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4075 break; 4076 4077 for (phases::ID Phase : PL) { 4078 4079 // Add any offload action the host action depends on. 4080 if (!UseNewOffloadingDriver) 4081 Current = OffloadBuilder.addDeviceDependencesToHostAction( 4082 Current, InputArg, Phase, PL.back(), FullPL); 4083 if (!Current) 4084 break; 4085 4086 // Queue linker inputs. 4087 if (Phase == phases::Link) { 4088 assert(Phase == PL.back() && "linking must be final compilation step."); 4089 // We don't need to generate additional link commands if emitting AMD bitcode 4090 if (!(C.getInputArgs().hasArg(options::OPT_hip_link) && 4091 (C.getInputArgs().hasArg(options::OPT_emit_llvm)))) 4092 LinkerInputs.push_back(Current); 4093 Current = nullptr; 4094 break; 4095 } 4096 4097 // TODO: Consider removing this because the merged may not end up being 4098 // the final Phase in the pipeline. Perhaps the merged could just merge 4099 // and then pass an artifact of some sort to the Link Phase. 4100 // Queue merger inputs. 4101 if (Phase == phases::IfsMerge) { 4102 assert(Phase == PL.back() && "merging must be final compilation step."); 4103 MergerInputs.push_back(Current); 4104 Current = nullptr; 4105 break; 4106 } 4107 4108 // Each precompiled header file after a module file action is a module 4109 // header of that same module file, rather than being compiled to a 4110 // separate PCH. 4111 if (Phase == phases::Precompile && HeaderModuleAction && 4112 getPrecompiledType(InputType) == types::TY_PCH) { 4113 HeaderModuleAction->addModuleHeaderInput(Current); 4114 Current = nullptr; 4115 break; 4116 } 4117 4118 if (Phase == phases::Precompile && ExtractAPIAction) { 4119 ExtractAPIAction->addHeaderInput(Current); 4120 Current = nullptr; 4121 break; 4122 } 4123 4124 // FIXME: Should we include any prior module file outputs as inputs of 4125 // later actions in the same command line? 4126 4127 // Otherwise construct the appropriate action. 4128 Action *NewCurrent = ConstructPhaseAction(C, Args, Phase, Current); 4129 4130 // We didn't create a new action, so we will just move to the next phase. 4131 if (NewCurrent == Current) 4132 continue; 4133 4134 if (auto *HMA = dyn_cast<HeaderModulePrecompileJobAction>(NewCurrent)) 4135 HeaderModuleAction = HMA; 4136 else if (auto *EAA = dyn_cast<ExtractAPIJobAction>(NewCurrent)) 4137 ExtractAPIAction = EAA; 4138 4139 Current = NewCurrent; 4140 4141 // Use the current host action in any of the offloading actions, if 4142 // required. 4143 if (!UseNewOffloadingDriver) 4144 if (OffloadBuilder.addHostDependenceToDeviceActions(Current, InputArg)) 4145 break; 4146 4147 // Try to build the offloading actions and add the result as a dependency 4148 // to the host. 4149 if (UseNewOffloadingDriver) 4150 Current = BuildOffloadingActions(C, Args, I, Current); 4151 4152 if (Current->getType() == types::TY_Nothing) 4153 break; 4154 } 4155 4156 // If we ended with something, add to the output list. 4157 if (Current) 4158 Actions.push_back(Current); 4159 4160 // Add any top level actions generated for offloading. 4161 if (!UseNewOffloadingDriver) 4162 OffloadBuilder.appendTopLevelActions(Actions, Current, InputArg); 4163 else if (Current) 4164 Current->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4165 /*BoundArch=*/nullptr); 4166 } 4167 4168 // Add a link action if necessary. 4169 4170 if (LinkerInputs.empty()) { 4171 Arg *FinalPhaseArg; 4172 if (getFinalPhase(Args, &FinalPhaseArg) == phases::Link) 4173 if (!UseNewOffloadingDriver) 4174 OffloadBuilder.appendDeviceLinkActions(Actions); 4175 } 4176 4177 if (!LinkerInputs.empty()) { 4178 if (!UseNewOffloadingDriver) 4179 if (Action *Wrapper = OffloadBuilder.makeHostLinkAction()) 4180 LinkerInputs.push_back(Wrapper); 4181 Action *LA; 4182 // Check if this Linker Job should emit a static library. 4183 if (ShouldEmitStaticLibrary(Args)) { 4184 LA = C.MakeAction<StaticLibJobAction>(LinkerInputs, types::TY_Image); 4185 } else if (UseNewOffloadingDriver || 4186 Args.hasArg(options::OPT_offload_link)) { 4187 LA = C.MakeAction<LinkerWrapperJobAction>(LinkerInputs, types::TY_Image); 4188 LA->propagateHostOffloadInfo(C.getActiveOffloadKinds(), 4189 /*BoundArch=*/nullptr); 4190 } else { 4191 LA = C.MakeAction<LinkJobAction>(LinkerInputs, types::TY_Image); 4192 } 4193 if (!UseNewOffloadingDriver) 4194 LA = OffloadBuilder.processHostLinkAction(LA); 4195 Actions.push_back(LA); 4196 } 4197 4198 // Add an interface stubs merge action if necessary. 4199 if (!MergerInputs.empty()) 4200 Actions.push_back( 4201 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4202 4203 if (Args.hasArg(options::OPT_emit_interface_stubs)) { 4204 auto PhaseList = types::getCompilationPhases( 4205 types::TY_IFS_CPP, 4206 Args.hasArg(options::OPT_c) ? phases::Compile : phases::IfsMerge); 4207 4208 ActionList MergerInputs; 4209 4210 for (auto &I : Inputs) { 4211 types::ID InputType = I.first; 4212 const Arg *InputArg = I.second; 4213 4214 // Currently clang and the llvm assembler do not support generating symbol 4215 // stubs from assembly, so we skip the input on asm files. For ifs files 4216 // we rely on the normal pipeline setup in the pipeline setup code above. 4217 if (InputType == types::TY_IFS || InputType == types::TY_PP_Asm || 4218 InputType == types::TY_Asm) 4219 continue; 4220 4221 Action *Current = C.MakeAction<InputAction>(*InputArg, InputType); 4222 4223 for (auto Phase : PhaseList) { 4224 switch (Phase) { 4225 default: 4226 llvm_unreachable( 4227 "IFS Pipeline can only consist of Compile followed by IfsMerge."); 4228 case phases::Compile: { 4229 // Only IfsMerge (llvm-ifs) can handle .o files by looking for ifs 4230 // files where the .o file is located. The compile action can not 4231 // handle this. 4232 if (InputType == types::TY_Object) 4233 break; 4234 4235 Current = C.MakeAction<CompileJobAction>(Current, types::TY_IFS_CPP); 4236 break; 4237 } 4238 case phases::IfsMerge: { 4239 assert(Phase == PhaseList.back() && 4240 "merging must be final compilation step."); 4241 MergerInputs.push_back(Current); 4242 Current = nullptr; 4243 break; 4244 } 4245 } 4246 } 4247 4248 // If we ended with something, add to the output list. 4249 if (Current) 4250 Actions.push_back(Current); 4251 } 4252 4253 // Add an interface stubs merge action if necessary. 4254 if (!MergerInputs.empty()) 4255 Actions.push_back( 4256 C.MakeAction<IfsMergeJobAction>(MergerInputs, types::TY_Image)); 4257 } 4258 4259 // If --print-supported-cpus, -mcpu=? or -mtune=? is specified, build a custom 4260 // Compile phase that prints out supported cpu models and quits. 4261 if (Arg *A = Args.getLastArg(options::OPT_print_supported_cpus)) { 4262 // Use the -mcpu=? flag as the dummy input to cc1. 4263 Actions.clear(); 4264 Action *InputAc = C.MakeAction<InputAction>(*A, types::TY_C); 4265 Actions.push_back( 4266 C.MakeAction<PrecompileJobAction>(InputAc, types::TY_Nothing)); 4267 for (auto &I : Inputs) 4268 I.second->claim(); 4269 } 4270 4271 // Claim ignored clang-cl options. 4272 Args.ClaimAllArgs(options::OPT_cl_ignored_Group); 4273 4274 // Claim --offload-host-only and --offload-compile-host-device, which may be 4275 // passed to non-CUDA compilations and should not trigger warnings there. 4276 Args.ClaimAllArgs(options::OPT_offload_host_only); 4277 Args.ClaimAllArgs(options::OPT_offload_host_device); 4278 } 4279 4280 /// Returns the canonical name for the offloading architecture when using a HIP 4281 /// or CUDA architecture. 4282 static StringRef getCanonicalArchString(Compilation &C, 4283 const llvm::opt::DerivedArgList &Args, 4284 StringRef ArchStr, 4285 const llvm::Triple &Triple) { 4286 // Lookup the CUDA / HIP architecture string. Only report an error if we were 4287 // expecting the triple to be only NVPTX / AMDGPU. 4288 CudaArch Arch = StringToCudaArch(getProcessorFromTargetID(Triple, ArchStr)); 4289 if (Triple.isNVPTX() && 4290 (Arch == CudaArch::UNKNOWN || !IsNVIDIAGpuArch(Arch))) { 4291 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4292 << "CUDA" << ArchStr; 4293 return StringRef(); 4294 } else if (Triple.isAMDGPU() && 4295 (Arch == CudaArch::UNKNOWN || !IsAMDGpuArch(Arch))) { 4296 C.getDriver().Diag(clang::diag::err_drv_offload_bad_gpu_arch) 4297 << "HIP" << ArchStr; 4298 return StringRef(); 4299 } 4300 4301 if (IsNVIDIAGpuArch(Arch)) 4302 return Args.MakeArgStringRef(CudaArchToString(Arch)); 4303 4304 if (IsAMDGpuArch(Arch)) { 4305 llvm::StringMap<bool> Features; 4306 auto HIPTriple = getHIPOffloadTargetTriple(C.getDriver(), C.getInputArgs()); 4307 if (!HIPTriple) 4308 return StringRef(); 4309 auto Arch = parseTargetID(*HIPTriple, ArchStr, &Features); 4310 if (!Arch) { 4311 C.getDriver().Diag(clang::diag::err_drv_bad_target_id) << ArchStr; 4312 C.setContainsError(); 4313 return StringRef(); 4314 } 4315 return Args.MakeArgStringRef( 4316 getCanonicalTargetID(Arch.getValue(), Features)); 4317 } 4318 4319 // If the input isn't CUDA or HIP just return the architecture. 4320 return ArchStr; 4321 } 4322 4323 /// Checks if the set offloading architectures does not conflict. Returns the 4324 /// incompatible pair if a conflict occurs. 4325 static llvm::Optional<std::pair<llvm::StringRef, llvm::StringRef>> 4326 getConflictOffloadArchCombination(const llvm::DenseSet<StringRef> &Archs, 4327 Action::OffloadKind Kind) { 4328 if (Kind != Action::OFK_HIP) 4329 return None; 4330 4331 std::set<StringRef> ArchSet; 4332 llvm::copy(Archs, std::inserter(ArchSet, ArchSet.begin())); 4333 return getConflictTargetIDCombination(ArchSet); 4334 } 4335 4336 llvm::DenseSet<StringRef> 4337 Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args, 4338 Action::OffloadKind Kind, const ToolChain *TC) const { 4339 if (!TC) 4340 TC = &C.getDefaultToolChain(); 4341 4342 // --offload and --offload-arch options are mutually exclusive. 4343 if (Args.hasArgNoClaim(options::OPT_offload_EQ) && 4344 Args.hasArgNoClaim(options::OPT_offload_arch_EQ, 4345 options::OPT_no_offload_arch_EQ)) { 4346 C.getDriver().Diag(diag::err_opt_not_valid_with_opt) 4347 << "--offload" 4348 << (Args.hasArgNoClaim(options::OPT_offload_arch_EQ) 4349 ? "--offload-arch" 4350 : "--no-offload-arch"); 4351 } 4352 4353 if (KnownArchs.find(TC) != KnownArchs.end()) 4354 return KnownArchs.lookup(TC); 4355 4356 llvm::DenseSet<StringRef> Archs; 4357 for (auto &Arg : Args) { 4358 if (Arg->getOption().matches(options::OPT_offload_arch_EQ)) { 4359 Archs.insert( 4360 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4361 } else if (Arg->getOption().matches(options::OPT_no_offload_arch_EQ)) { 4362 if (Arg->getValue() == StringRef("all")) 4363 Archs.clear(); 4364 else 4365 Archs.erase( 4366 getCanonicalArchString(C, Args, Arg->getValue(), TC->getTriple())); 4367 } 4368 } 4369 4370 if (auto ConflictingArchs = getConflictOffloadArchCombination(Archs, Kind)) { 4371 C.getDriver().Diag(clang::diag::err_drv_bad_offload_arch_combo) 4372 << ConflictingArchs.getValue().first 4373 << ConflictingArchs.getValue().second; 4374 C.setContainsError(); 4375 } 4376 4377 if (Archs.empty()) { 4378 if (Kind == Action::OFK_Cuda) 4379 Archs.insert(CudaArchToString(CudaArch::CudaDefault)); 4380 else if (Kind == Action::OFK_HIP) 4381 Archs.insert(CudaArchToString(CudaArch::HIPDefault)); 4382 else if (Kind == Action::OFK_OpenMP) 4383 Archs.insert(StringRef()); 4384 } else { 4385 Args.ClaimAllArgs(options::OPT_offload_arch_EQ); 4386 Args.ClaimAllArgs(options::OPT_no_offload_arch_EQ); 4387 } 4388 4389 return Archs; 4390 } 4391 4392 Action *Driver::BuildOffloadingActions(Compilation &C, 4393 llvm::opt::DerivedArgList &Args, 4394 const InputTy &Input, 4395 Action *HostAction) const { 4396 const Arg *Mode = Args.getLastArg(options::OPT_offload_host_only, 4397 options::OPT_offload_device_only, 4398 options::OPT_offload_host_device); 4399 const bool HostOnly = 4400 Mode && Mode->getOption().matches(options::OPT_offload_host_only); 4401 const bool DeviceOnly = 4402 Mode && Mode->getOption().matches(options::OPT_offload_device_only); 4403 4404 // Don't build offloading actions if explicitly disabled or we do not have a 4405 // valid source input and compile action to embed it in. If preprocessing only 4406 // ignore embedding. 4407 if (HostOnly || !types::isSrcFile(Input.first) || 4408 !(isa<CompileJobAction>(HostAction) || 4409 getFinalPhase(Args) == phases::Preprocess)) 4410 return HostAction; 4411 4412 ActionList OffloadActions; 4413 OffloadAction::DeviceDependences DDeps; 4414 4415 const Action::OffloadKind OffloadKinds[] = { 4416 Action::OFK_OpenMP, Action::OFK_Cuda, Action::OFK_HIP}; 4417 4418 for (Action::OffloadKind Kind : OffloadKinds) { 4419 SmallVector<const ToolChain *, 2> ToolChains; 4420 ActionList DeviceActions; 4421 4422 auto TCRange = C.getOffloadToolChains(Kind); 4423 for (auto TI = TCRange.first, TE = TCRange.second; TI != TE; ++TI) 4424 ToolChains.push_back(TI->second); 4425 4426 if (ToolChains.empty()) 4427 continue; 4428 4429 types::ID InputType = Input.first; 4430 const Arg *InputArg = Input.second; 4431 4432 // Get the product of all bound architectures and toolchains. 4433 SmallVector<std::pair<const ToolChain *, StringRef>> TCAndArchs; 4434 for (const ToolChain *TC : ToolChains) 4435 for (StringRef Arch : getOffloadArchs( 4436 C, C.getArgsForToolChain(TC, "generic", Kind), Kind, TC)) 4437 TCAndArchs.push_back(std::make_pair(TC, Arch)); 4438 4439 for (unsigned I = 0, E = TCAndArchs.size(); I != E; ++I) 4440 DeviceActions.push_back(C.MakeAction<InputAction>(*InputArg, InputType)); 4441 4442 if (DeviceActions.empty()) 4443 return HostAction; 4444 4445 auto PL = types::getCompilationPhases(*this, Args, InputType); 4446 4447 for (phases::ID Phase : PL) { 4448 if (Phase == phases::Link) { 4449 assert(Phase == PL.back() && "linking must be final compilation step."); 4450 break; 4451 } 4452 4453 auto TCAndArch = TCAndArchs.begin(); 4454 for (Action *&A : DeviceActions) { 4455 A = ConstructPhaseAction(C, Args, Phase, A, Kind); 4456 4457 if (isa<CompileJobAction>(A) && isa<CompileJobAction>(HostAction) && 4458 Kind == Action::OFK_OpenMP) { 4459 // OpenMP offloading has a dependency on the host compile action to 4460 // identify which declarations need to be emitted. This shouldn't be 4461 // collapsed with any other actions so we can use it in the device. 4462 HostAction->setCannotBeCollapsedWithNextDependentAction(); 4463 OffloadAction::HostDependence HDep( 4464 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4465 TCAndArch->second.data(), Kind); 4466 OffloadAction::DeviceDependences DDep; 4467 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4468 A = C.MakeAction<OffloadAction>(HDep, DDep); 4469 } else if (isa<AssembleJobAction>(A) && Kind == Action::OFK_Cuda) { 4470 // The Cuda toolchain uses fatbinary as the linker phase to bundle the 4471 // PTX and Cubin output. 4472 ActionList FatbinActions; 4473 for (Action *A : {A, A->getInputs()[0]}) { 4474 OffloadAction::DeviceDependences DDep; 4475 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4476 FatbinActions.emplace_back( 4477 C.MakeAction<OffloadAction>(DDep, A->getType())); 4478 } 4479 A = C.MakeAction<LinkJobAction>(FatbinActions, types::TY_CUDA_FATBIN); 4480 } 4481 ++TCAndArch; 4482 } 4483 } 4484 4485 auto TCAndArch = TCAndArchs.begin(); 4486 for (Action *A : DeviceActions) { 4487 DDeps.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4488 OffloadAction::DeviceDependences DDep; 4489 DDep.add(*A, *TCAndArch->first, TCAndArch->second.data(), Kind); 4490 OffloadActions.push_back(C.MakeAction<OffloadAction>(DDep, A->getType())); 4491 ++TCAndArch; 4492 } 4493 } 4494 4495 if (DeviceOnly) 4496 return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing); 4497 4498 Action *OffloadPackager = 4499 C.MakeAction<OffloadPackagerJobAction>(OffloadActions, types::TY_Image); 4500 OffloadAction::DeviceDependences DDep; 4501 DDep.add(*OffloadPackager, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4502 nullptr, Action::OFK_None); 4503 OffloadAction::HostDependence HDep( 4504 *HostAction, *C.getSingleOffloadToolChain<Action::OFK_Host>(), 4505 /*BoundArch=*/nullptr, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4506 return C.MakeAction<OffloadAction>( 4507 HDep, isa<CompileJobAction>(HostAction) ? DDep : DDeps); 4508 } 4509 4510 Action *Driver::ConstructPhaseAction( 4511 Compilation &C, const ArgList &Args, phases::ID Phase, Action *Input, 4512 Action::OffloadKind TargetDeviceOffloadKind) const { 4513 llvm::PrettyStackTraceString CrashInfo("Constructing phase actions"); 4514 4515 // Some types skip the assembler phase (e.g., llvm-bc), but we can't 4516 // encode this in the steps because the intermediate type depends on 4517 // arguments. Just special case here. 4518 if (Phase == phases::Assemble && Input->getType() != types::TY_PP_Asm) 4519 return Input; 4520 4521 // Build the appropriate action. 4522 switch (Phase) { 4523 case phases::Link: 4524 llvm_unreachable("link action invalid here."); 4525 case phases::IfsMerge: 4526 llvm_unreachable("ifsmerge action invalid here."); 4527 case phases::Preprocess: { 4528 types::ID OutputTy; 4529 // -M and -MM specify the dependency file name by altering the output type, 4530 // -if -MD and -MMD are not specified. 4531 if (Args.hasArg(options::OPT_M, options::OPT_MM) && 4532 !Args.hasArg(options::OPT_MD, options::OPT_MMD)) { 4533 OutputTy = types::TY_Dependencies; 4534 } else { 4535 OutputTy = Input->getType(); 4536 // For these cases, the preprocessor is only translating forms, the Output 4537 // still needs preprocessing. 4538 if (!Args.hasFlag(options::OPT_frewrite_includes, 4539 options::OPT_fno_rewrite_includes, false) && 4540 !Args.hasFlag(options::OPT_frewrite_imports, 4541 options::OPT_fno_rewrite_imports, false) && 4542 !Args.hasFlag(options::OPT_fdirectives_only, 4543 options::OPT_fno_directives_only, false) && 4544 !CCGenDiagnostics) 4545 OutputTy = types::getPreprocessedType(OutputTy); 4546 assert(OutputTy != types::TY_INVALID && 4547 "Cannot preprocess this input type!"); 4548 } 4549 return C.MakeAction<PreprocessJobAction>(Input, OutputTy); 4550 } 4551 case phases::Precompile: { 4552 // API extraction should not generate an actual precompilation action. 4553 if (Args.hasArg(options::OPT_extract_api)) 4554 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4555 4556 types::ID OutputTy = getPrecompiledType(Input->getType()); 4557 assert(OutputTy != types::TY_INVALID && 4558 "Cannot precompile this input type!"); 4559 4560 // If we're given a module name, precompile header file inputs as a 4561 // module, not as a precompiled header. 4562 const char *ModName = nullptr; 4563 if (OutputTy == types::TY_PCH) { 4564 if (Arg *A = Args.getLastArg(options::OPT_fmodule_name_EQ)) 4565 ModName = A->getValue(); 4566 if (ModName) 4567 OutputTy = types::TY_ModuleFile; 4568 } 4569 4570 if (Args.hasArg(options::OPT_fsyntax_only)) { 4571 // Syntax checks should not emit a PCH file 4572 OutputTy = types::TY_Nothing; 4573 } 4574 4575 if (ModName) 4576 return C.MakeAction<HeaderModulePrecompileJobAction>(Input, OutputTy, 4577 ModName); 4578 return C.MakeAction<PrecompileJobAction>(Input, OutputTy); 4579 } 4580 case phases::Compile: { 4581 if (Args.hasArg(options::OPT_fsyntax_only)) 4582 return C.MakeAction<CompileJobAction>(Input, types::TY_Nothing); 4583 if (Args.hasArg(options::OPT_rewrite_objc)) 4584 return C.MakeAction<CompileJobAction>(Input, types::TY_RewrittenObjC); 4585 if (Args.hasArg(options::OPT_rewrite_legacy_objc)) 4586 return C.MakeAction<CompileJobAction>(Input, 4587 types::TY_RewrittenLegacyObjC); 4588 if (Args.hasArg(options::OPT__analyze)) 4589 return C.MakeAction<AnalyzeJobAction>(Input, types::TY_Plist); 4590 if (Args.hasArg(options::OPT__migrate)) 4591 return C.MakeAction<MigrateJobAction>(Input, types::TY_Remap); 4592 if (Args.hasArg(options::OPT_emit_ast)) 4593 return C.MakeAction<CompileJobAction>(Input, types::TY_AST); 4594 if (Args.hasArg(options::OPT_module_file_info)) 4595 return C.MakeAction<CompileJobAction>(Input, types::TY_ModuleFile); 4596 if (Args.hasArg(options::OPT_verify_pch)) 4597 return C.MakeAction<VerifyPCHJobAction>(Input, types::TY_Nothing); 4598 if (Args.hasArg(options::OPT_extract_api)) 4599 return C.MakeAction<ExtractAPIJobAction>(Input, types::TY_API_INFO); 4600 return C.MakeAction<CompileJobAction>(Input, types::TY_LLVM_BC); 4601 } 4602 case phases::Backend: { 4603 if (isUsingLTO() && TargetDeviceOffloadKind == Action::OFK_None) { 4604 types::ID Output = 4605 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4606 return C.MakeAction<BackendJobAction>(Input, Output); 4607 } 4608 if (isUsingLTO(/* IsOffload */ true) && 4609 TargetDeviceOffloadKind != Action::OFK_None) { 4610 types::ID Output = 4611 Args.hasArg(options::OPT_S) ? types::TY_LTO_IR : types::TY_LTO_BC; 4612 return C.MakeAction<BackendJobAction>(Input, Output); 4613 } 4614 if (Args.hasArg(options::OPT_emit_llvm) || 4615 (TargetDeviceOffloadKind == Action::OFK_HIP && 4616 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 4617 false))) { 4618 types::ID Output = 4619 Args.hasArg(options::OPT_S) ? types::TY_LLVM_IR : types::TY_LLVM_BC; 4620 return C.MakeAction<BackendJobAction>(Input, Output); 4621 } 4622 return C.MakeAction<BackendJobAction>(Input, types::TY_PP_Asm); 4623 } 4624 case phases::Assemble: 4625 return C.MakeAction<AssembleJobAction>(std::move(Input), types::TY_Object); 4626 } 4627 4628 llvm_unreachable("invalid phase in ConstructPhaseAction"); 4629 } 4630 4631 void Driver::BuildJobs(Compilation &C) const { 4632 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 4633 4634 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 4635 4636 // It is an error to provide a -o option if we are making multiple output 4637 // files. There are exceptions: 4638 // 4639 // IfsMergeJob: when generating interface stubs enabled we want to be able to 4640 // generate the stub file at the same time that we generate the real 4641 // library/a.out. So when a .o, .so, etc are the output, with clang interface 4642 // stubs there will also be a .ifs and .ifso at the same location. 4643 // 4644 // CompileJob of type TY_IFS_CPP: when generating interface stubs is enabled 4645 // and -c is passed, we still want to be able to generate a .ifs file while 4646 // we are also generating .o files. So we allow more than one output file in 4647 // this case as well. 4648 // 4649 if (FinalOutput) { 4650 unsigned NumOutputs = 0; 4651 unsigned NumIfsOutputs = 0; 4652 for (const Action *A : C.getActions()) 4653 if (A->getType() != types::TY_Nothing && 4654 !(A->getKind() == Action::IfsMergeJobClass || 4655 (A->getType() == clang::driver::types::TY_IFS_CPP && 4656 A->getKind() == clang::driver::Action::CompileJobClass && 4657 0 == NumIfsOutputs++) || 4658 (A->getKind() == Action::BindArchClass && A->getInputs().size() && 4659 A->getInputs().front()->getKind() == Action::IfsMergeJobClass))) 4660 ++NumOutputs; 4661 4662 if (NumOutputs > 1) { 4663 Diag(clang::diag::err_drv_output_argument_with_multiple_files); 4664 FinalOutput = nullptr; 4665 } 4666 } 4667 4668 const llvm::Triple &RawTriple = C.getDefaultToolChain().getTriple(); 4669 if (RawTriple.isOSAIX()) { 4670 if (Arg *A = C.getArgs().getLastArg(options::OPT_G)) 4671 Diag(diag::err_drv_unsupported_opt_for_target) 4672 << A->getSpelling() << RawTriple.str(); 4673 if (LTOMode == LTOK_Thin) 4674 Diag(diag::err_drv_clang_unsupported) << "thinLTO on AIX"; 4675 } 4676 4677 // Collect the list of architectures. 4678 llvm::StringSet<> ArchNames; 4679 if (RawTriple.isOSBinFormatMachO()) 4680 for (const Arg *A : C.getArgs()) 4681 if (A->getOption().matches(options::OPT_arch)) 4682 ArchNames.insert(A->getValue()); 4683 4684 // Set of (Action, canonical ToolChain triple) pairs we've built jobs for. 4685 std::map<std::pair<const Action *, std::string>, InputInfoList> CachedResults; 4686 for (Action *A : C.getActions()) { 4687 // If we are linking an image for multiple archs then the linker wants 4688 // -arch_multiple and -final_output <final image name>. Unfortunately, this 4689 // doesn't fit in cleanly because we have to pass this information down. 4690 // 4691 // FIXME: This is a hack; find a cleaner way to integrate this into the 4692 // process. 4693 const char *LinkingOutput = nullptr; 4694 if (isa<LipoJobAction>(A)) { 4695 if (FinalOutput) 4696 LinkingOutput = FinalOutput->getValue(); 4697 else 4698 LinkingOutput = getDefaultImageName(); 4699 } 4700 4701 BuildJobsForAction(C, A, &C.getDefaultToolChain(), 4702 /*BoundArch*/ StringRef(), 4703 /*AtTopLevel*/ true, 4704 /*MultipleArchs*/ ArchNames.size() > 1, 4705 /*LinkingOutput*/ LinkingOutput, CachedResults, 4706 /*TargetDeviceOffloadKind*/ Action::OFK_None); 4707 } 4708 4709 // If we have more than one job, then disable integrated-cc1 for now. Do this 4710 // also when we need to report process execution statistics. 4711 if (C.getJobs().size() > 1 || CCPrintProcessStats) 4712 for (auto &J : C.getJobs()) 4713 J.InProcess = false; 4714 4715 if (CCPrintProcessStats) { 4716 C.setPostCallback([=](const Command &Cmd, int Res) { 4717 Optional<llvm::sys::ProcessStatistics> ProcStat = 4718 Cmd.getProcessStatistics(); 4719 if (!ProcStat) 4720 return; 4721 4722 const char *LinkingOutput = nullptr; 4723 if (FinalOutput) 4724 LinkingOutput = FinalOutput->getValue(); 4725 else if (!Cmd.getOutputFilenames().empty()) 4726 LinkingOutput = Cmd.getOutputFilenames().front().c_str(); 4727 else 4728 LinkingOutput = getDefaultImageName(); 4729 4730 if (CCPrintStatReportFilename.empty()) { 4731 using namespace llvm; 4732 // Human readable output. 4733 outs() << sys::path::filename(Cmd.getExecutable()) << ": " 4734 << "output=" << LinkingOutput; 4735 outs() << ", total=" 4736 << format("%.3f", ProcStat->TotalTime.count() / 1000.) << " ms" 4737 << ", user=" 4738 << format("%.3f", ProcStat->UserTime.count() / 1000.) << " ms" 4739 << ", mem=" << ProcStat->PeakMemory << " Kb\n"; 4740 } else { 4741 // CSV format. 4742 std::string Buffer; 4743 llvm::raw_string_ostream Out(Buffer); 4744 llvm::sys::printArg(Out, llvm::sys::path::filename(Cmd.getExecutable()), 4745 /*Quote*/ true); 4746 Out << ','; 4747 llvm::sys::printArg(Out, LinkingOutput, true); 4748 Out << ',' << ProcStat->TotalTime.count() << ',' 4749 << ProcStat->UserTime.count() << ',' << ProcStat->PeakMemory 4750 << '\n'; 4751 Out.flush(); 4752 std::error_code EC; 4753 llvm::raw_fd_ostream OS(CCPrintStatReportFilename, EC, 4754 llvm::sys::fs::OF_Append | 4755 llvm::sys::fs::OF_Text); 4756 if (EC) 4757 return; 4758 auto L = OS.lock(); 4759 if (!L) { 4760 llvm::errs() << "ERROR: Cannot lock file " 4761 << CCPrintStatReportFilename << ": " 4762 << toString(L.takeError()) << "\n"; 4763 return; 4764 } 4765 OS << Buffer; 4766 OS.flush(); 4767 } 4768 }); 4769 } 4770 4771 // If the user passed -Qunused-arguments or there were errors, don't warn 4772 // about any unused arguments. 4773 if (Diags.hasErrorOccurred() || 4774 C.getArgs().hasArg(options::OPT_Qunused_arguments)) 4775 return; 4776 4777 // Claim -fdriver-only here. 4778 (void)C.getArgs().hasArg(options::OPT_fdriver_only); 4779 // Claim -### here. 4780 (void)C.getArgs().hasArg(options::OPT__HASH_HASH_HASH); 4781 4782 // Claim --driver-mode, --rsp-quoting, it was handled earlier. 4783 (void)C.getArgs().hasArg(options::OPT_driver_mode); 4784 (void)C.getArgs().hasArg(options::OPT_rsp_quoting); 4785 4786 for (Arg *A : C.getArgs()) { 4787 // FIXME: It would be nice to be able to send the argument to the 4788 // DiagnosticsEngine, so that extra values, position, and so on could be 4789 // printed. 4790 if (!A->isClaimed()) { 4791 if (A->getOption().hasFlag(options::NoArgumentUnused)) 4792 continue; 4793 4794 // Suppress the warning automatically if this is just a flag, and it is an 4795 // instance of an argument we already claimed. 4796 const Option &Opt = A->getOption(); 4797 if (Opt.getKind() == Option::FlagClass) { 4798 bool DuplicateClaimed = false; 4799 4800 for (const Arg *AA : C.getArgs().filtered(&Opt)) { 4801 if (AA->isClaimed()) { 4802 DuplicateClaimed = true; 4803 break; 4804 } 4805 } 4806 4807 if (DuplicateClaimed) 4808 continue; 4809 } 4810 4811 // In clang-cl, don't mention unknown arguments here since they have 4812 // already been warned about. 4813 if (!IsCLMode() || !A->getOption().matches(options::OPT_UNKNOWN)) 4814 Diag(clang::diag::warn_drv_unused_argument) 4815 << A->getAsString(C.getArgs()); 4816 } 4817 } 4818 } 4819 4820 namespace { 4821 /// Utility class to control the collapse of dependent actions and select the 4822 /// tools accordingly. 4823 class ToolSelector final { 4824 /// The tool chain this selector refers to. 4825 const ToolChain &TC; 4826 4827 /// The compilation this selector refers to. 4828 const Compilation &C; 4829 4830 /// The base action this selector refers to. 4831 const JobAction *BaseAction; 4832 4833 /// Set to true if the current toolchain refers to host actions. 4834 bool IsHostSelector; 4835 4836 /// Set to true if save-temps and embed-bitcode functionalities are active. 4837 bool SaveTemps; 4838 bool EmbedBitcode; 4839 4840 /// Get previous dependent action or null if that does not exist. If 4841 /// \a CanBeCollapsed is false, that action must be legal to collapse or 4842 /// null will be returned. 4843 const JobAction *getPrevDependentAction(const ActionList &Inputs, 4844 ActionList &SavedOffloadAction, 4845 bool CanBeCollapsed = true) { 4846 // An option can be collapsed only if it has a single input. 4847 if (Inputs.size() != 1) 4848 return nullptr; 4849 4850 Action *CurAction = *Inputs.begin(); 4851 if (CanBeCollapsed && 4852 !CurAction->isCollapsingWithNextDependentActionLegal()) 4853 return nullptr; 4854 4855 // If the input action is an offload action. Look through it and save any 4856 // offload action that can be dropped in the event of a collapse. 4857 if (auto *OA = dyn_cast<OffloadAction>(CurAction)) { 4858 // If the dependent action is a device action, we will attempt to collapse 4859 // only with other device actions. Otherwise, we would do the same but 4860 // with host actions only. 4861 if (!IsHostSelector) { 4862 if (OA->hasSingleDeviceDependence(/*DoNotConsiderHostActions=*/true)) { 4863 CurAction = 4864 OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true); 4865 if (CanBeCollapsed && 4866 !CurAction->isCollapsingWithNextDependentActionLegal()) 4867 return nullptr; 4868 SavedOffloadAction.push_back(OA); 4869 return dyn_cast<JobAction>(CurAction); 4870 } 4871 } else if (OA->hasHostDependence()) { 4872 CurAction = OA->getHostDependence(); 4873 if (CanBeCollapsed && 4874 !CurAction->isCollapsingWithNextDependentActionLegal()) 4875 return nullptr; 4876 SavedOffloadAction.push_back(OA); 4877 return dyn_cast<JobAction>(CurAction); 4878 } 4879 return nullptr; 4880 } 4881 4882 return dyn_cast<JobAction>(CurAction); 4883 } 4884 4885 /// Return true if an assemble action can be collapsed. 4886 bool canCollapseAssembleAction() const { 4887 return TC.useIntegratedAs() && !SaveTemps && 4888 !C.getArgs().hasArg(options::OPT_via_file_asm) && 4889 !C.getArgs().hasArg(options::OPT__SLASH_FA) && 4890 !C.getArgs().hasArg(options::OPT__SLASH_Fa); 4891 } 4892 4893 /// Return true if a preprocessor action can be collapsed. 4894 bool canCollapsePreprocessorAction() const { 4895 return !C.getArgs().hasArg(options::OPT_no_integrated_cpp) && 4896 !C.getArgs().hasArg(options::OPT_traditional_cpp) && !SaveTemps && 4897 !C.getArgs().hasArg(options::OPT_rewrite_objc); 4898 } 4899 4900 /// Struct that relates an action with the offload actions that would be 4901 /// collapsed with it. 4902 struct JobActionInfo final { 4903 /// The action this info refers to. 4904 const JobAction *JA = nullptr; 4905 /// The offload actions we need to take care off if this action is 4906 /// collapsed. 4907 ActionList SavedOffloadAction; 4908 }; 4909 4910 /// Append collapsed offload actions from the give nnumber of elements in the 4911 /// action info array. 4912 static void AppendCollapsedOffloadAction(ActionList &CollapsedOffloadAction, 4913 ArrayRef<JobActionInfo> &ActionInfo, 4914 unsigned ElementNum) { 4915 assert(ElementNum <= ActionInfo.size() && "Invalid number of elements."); 4916 for (unsigned I = 0; I < ElementNum; ++I) 4917 CollapsedOffloadAction.append(ActionInfo[I].SavedOffloadAction.begin(), 4918 ActionInfo[I].SavedOffloadAction.end()); 4919 } 4920 4921 /// Functions that attempt to perform the combining. They detect if that is 4922 /// legal, and if so they update the inputs \a Inputs and the offload action 4923 /// that were collapsed in \a CollapsedOffloadAction. A tool that deals with 4924 /// the combined action is returned. If the combining is not legal or if the 4925 /// tool does not exist, null is returned. 4926 /// Currently three kinds of collapsing are supported: 4927 /// - Assemble + Backend + Compile; 4928 /// - Assemble + Backend ; 4929 /// - Backend + Compile. 4930 const Tool * 4931 combineAssembleBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4932 ActionList &Inputs, 4933 ActionList &CollapsedOffloadAction) { 4934 if (ActionInfo.size() < 3 || !canCollapseAssembleAction()) 4935 return nullptr; 4936 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4937 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4938 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[2].JA); 4939 if (!AJ || !BJ || !CJ) 4940 return nullptr; 4941 4942 // Get compiler tool. 4943 const Tool *T = TC.SelectTool(*CJ); 4944 if (!T) 4945 return nullptr; 4946 4947 // Can't collapse if we don't have codegen support unless we are 4948 // emitting LLVM IR. 4949 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 4950 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 4951 return nullptr; 4952 4953 // When using -fembed-bitcode, it is required to have the same tool (clang) 4954 // for both CompilerJA and BackendJA. Otherwise, combine two stages. 4955 if (EmbedBitcode) { 4956 const Tool *BT = TC.SelectTool(*BJ); 4957 if (BT == T) 4958 return nullptr; 4959 } 4960 4961 if (!T->hasIntegratedAssembler()) 4962 return nullptr; 4963 4964 Inputs = CJ->getInputs(); 4965 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4966 /*NumElements=*/3); 4967 return T; 4968 } 4969 const Tool *combineAssembleBackend(ArrayRef<JobActionInfo> ActionInfo, 4970 ActionList &Inputs, 4971 ActionList &CollapsedOffloadAction) { 4972 if (ActionInfo.size() < 2 || !canCollapseAssembleAction()) 4973 return nullptr; 4974 auto *AJ = dyn_cast<AssembleJobAction>(ActionInfo[0].JA); 4975 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[1].JA); 4976 if (!AJ || !BJ) 4977 return nullptr; 4978 4979 // Get backend tool. 4980 const Tool *T = TC.SelectTool(*BJ); 4981 if (!T) 4982 return nullptr; 4983 4984 if (!T->hasIntegratedAssembler()) 4985 return nullptr; 4986 4987 Inputs = BJ->getInputs(); 4988 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 4989 /*NumElements=*/2); 4990 return T; 4991 } 4992 const Tool *combineBackendCompile(ArrayRef<JobActionInfo> ActionInfo, 4993 ActionList &Inputs, 4994 ActionList &CollapsedOffloadAction) { 4995 if (ActionInfo.size() < 2) 4996 return nullptr; 4997 auto *BJ = dyn_cast<BackendJobAction>(ActionInfo[0].JA); 4998 auto *CJ = dyn_cast<CompileJobAction>(ActionInfo[1].JA); 4999 if (!BJ || !CJ) 5000 return nullptr; 5001 5002 // Check if the initial input (to the compile job or its predessor if one 5003 // exists) is LLVM bitcode. In that case, no preprocessor step is required 5004 // and we can still collapse the compile and backend jobs when we have 5005 // -save-temps. I.e. there is no need for a separate compile job just to 5006 // emit unoptimized bitcode. 5007 bool InputIsBitcode = true; 5008 for (size_t i = 1; i < ActionInfo.size(); i++) 5009 if (ActionInfo[i].JA->getType() != types::TY_LLVM_BC && 5010 ActionInfo[i].JA->getType() != types::TY_LTO_BC) { 5011 InputIsBitcode = false; 5012 break; 5013 } 5014 if (!InputIsBitcode && !canCollapsePreprocessorAction()) 5015 return nullptr; 5016 5017 // Get compiler tool. 5018 const Tool *T = TC.SelectTool(*CJ); 5019 if (!T) 5020 return nullptr; 5021 5022 // Can't collapse if we don't have codegen support unless we are 5023 // emitting LLVM IR. 5024 bool OutputIsLLVM = types::isLLVMIR(ActionInfo[0].JA->getType()); 5025 if (!T->hasIntegratedBackend() && !(OutputIsLLVM && T->canEmitIR())) 5026 return nullptr; 5027 5028 if (T->canEmitIR() && ((SaveTemps && !InputIsBitcode) || EmbedBitcode)) 5029 return nullptr; 5030 5031 Inputs = CJ->getInputs(); 5032 AppendCollapsedOffloadAction(CollapsedOffloadAction, ActionInfo, 5033 /*NumElements=*/2); 5034 return T; 5035 } 5036 5037 /// Updates the inputs if the obtained tool supports combining with 5038 /// preprocessor action, and the current input is indeed a preprocessor 5039 /// action. If combining results in the collapse of offloading actions, those 5040 /// are appended to \a CollapsedOffloadAction. 5041 void combineWithPreprocessor(const Tool *T, ActionList &Inputs, 5042 ActionList &CollapsedOffloadAction) { 5043 if (!T || !canCollapsePreprocessorAction() || !T->hasIntegratedCPP()) 5044 return; 5045 5046 // Attempt to get a preprocessor action dependence. 5047 ActionList PreprocessJobOffloadActions; 5048 ActionList NewInputs; 5049 for (Action *A : Inputs) { 5050 auto *PJ = getPrevDependentAction({A}, PreprocessJobOffloadActions); 5051 if (!PJ || !isa<PreprocessJobAction>(PJ)) { 5052 NewInputs.push_back(A); 5053 continue; 5054 } 5055 5056 // This is legal to combine. Append any offload action we found and add the 5057 // current input to preprocessor inputs. 5058 CollapsedOffloadAction.append(PreprocessJobOffloadActions.begin(), 5059 PreprocessJobOffloadActions.end()); 5060 NewInputs.append(PJ->input_begin(), PJ->input_end()); 5061 } 5062 Inputs = NewInputs; 5063 } 5064 5065 public: 5066 ToolSelector(const JobAction *BaseAction, const ToolChain &TC, 5067 const Compilation &C, bool SaveTemps, bool EmbedBitcode) 5068 : TC(TC), C(C), BaseAction(BaseAction), SaveTemps(SaveTemps), 5069 EmbedBitcode(EmbedBitcode) { 5070 assert(BaseAction && "Invalid base action."); 5071 IsHostSelector = BaseAction->getOffloadingDeviceKind() == Action::OFK_None; 5072 } 5073 5074 /// Check if a chain of actions can be combined and return the tool that can 5075 /// handle the combination of actions. The pointer to the current inputs \a 5076 /// Inputs and the list of offload actions \a CollapsedOffloadActions 5077 /// connected to collapsed actions are updated accordingly. The latter enables 5078 /// the caller of the selector to process them afterwards instead of just 5079 /// dropping them. If no suitable tool is found, null will be returned. 5080 const Tool *getTool(ActionList &Inputs, 5081 ActionList &CollapsedOffloadAction) { 5082 // 5083 // Get the largest chain of actions that we could combine. 5084 // 5085 5086 SmallVector<JobActionInfo, 5> ActionChain(1); 5087 ActionChain.back().JA = BaseAction; 5088 while (ActionChain.back().JA) { 5089 const Action *CurAction = ActionChain.back().JA; 5090 5091 // Grow the chain by one element. 5092 ActionChain.resize(ActionChain.size() + 1); 5093 JobActionInfo &AI = ActionChain.back(); 5094 5095 // Attempt to fill it with the 5096 AI.JA = 5097 getPrevDependentAction(CurAction->getInputs(), AI.SavedOffloadAction); 5098 } 5099 5100 // Pop the last action info as it could not be filled. 5101 ActionChain.pop_back(); 5102 5103 // 5104 // Attempt to combine actions. If all combining attempts failed, just return 5105 // the tool of the provided action. At the end we attempt to combine the 5106 // action with any preprocessor action it may depend on. 5107 // 5108 5109 const Tool *T = combineAssembleBackendCompile(ActionChain, Inputs, 5110 CollapsedOffloadAction); 5111 if (!T) 5112 T = combineAssembleBackend(ActionChain, Inputs, CollapsedOffloadAction); 5113 if (!T) 5114 T = combineBackendCompile(ActionChain, Inputs, CollapsedOffloadAction); 5115 if (!T) { 5116 Inputs = BaseAction->getInputs(); 5117 T = TC.SelectTool(*BaseAction); 5118 } 5119 5120 combineWithPreprocessor(T, Inputs, CollapsedOffloadAction); 5121 return T; 5122 } 5123 }; 5124 } 5125 5126 /// Return a string that uniquely identifies the result of a job. The bound arch 5127 /// is not necessarily represented in the toolchain's triple -- for example, 5128 /// armv7 and armv7s both map to the same triple -- so we need both in our map. 5129 /// Also, we need to add the offloading device kind, as the same tool chain can 5130 /// be used for host and device for some programming models, e.g. OpenMP. 5131 static std::string GetTriplePlusArchString(const ToolChain *TC, 5132 StringRef BoundArch, 5133 Action::OffloadKind OffloadKind) { 5134 std::string TriplePlusArch = TC->getTriple().normalize(); 5135 if (!BoundArch.empty()) { 5136 TriplePlusArch += "-"; 5137 TriplePlusArch += BoundArch; 5138 } 5139 TriplePlusArch += "-"; 5140 TriplePlusArch += Action::GetOffloadKindName(OffloadKind); 5141 return TriplePlusArch; 5142 } 5143 5144 InputInfoList Driver::BuildJobsForAction( 5145 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5146 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5147 std::map<std::pair<const Action *, std::string>, InputInfoList> 5148 &CachedResults, 5149 Action::OffloadKind TargetDeviceOffloadKind) const { 5150 std::pair<const Action *, std::string> ActionTC = { 5151 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5152 auto CachedResult = CachedResults.find(ActionTC); 5153 if (CachedResult != CachedResults.end()) { 5154 return CachedResult->second; 5155 } 5156 InputInfoList Result = BuildJobsForActionNoCache( 5157 C, A, TC, BoundArch, AtTopLevel, MultipleArchs, LinkingOutput, 5158 CachedResults, TargetDeviceOffloadKind); 5159 CachedResults[ActionTC] = Result; 5160 return Result; 5161 } 5162 5163 InputInfoList Driver::BuildJobsForActionNoCache( 5164 Compilation &C, const Action *A, const ToolChain *TC, StringRef BoundArch, 5165 bool AtTopLevel, bool MultipleArchs, const char *LinkingOutput, 5166 std::map<std::pair<const Action *, std::string>, InputInfoList> 5167 &CachedResults, 5168 Action::OffloadKind TargetDeviceOffloadKind) const { 5169 llvm::PrettyStackTraceString CrashInfo("Building compilation jobs"); 5170 5171 InputInfoList OffloadDependencesInputInfo; 5172 bool BuildingForOffloadDevice = TargetDeviceOffloadKind != Action::OFK_None; 5173 if (const OffloadAction *OA = dyn_cast<OffloadAction>(A)) { 5174 // The 'Darwin' toolchain is initialized only when its arguments are 5175 // computed. Get the default arguments for OFK_None to ensure that 5176 // initialization is performed before processing the offload action. 5177 // FIXME: Remove when darwin's toolchain is initialized during construction. 5178 C.getArgsForToolChain(TC, BoundArch, Action::OFK_None); 5179 5180 // The offload action is expected to be used in four different situations. 5181 // 5182 // a) Set a toolchain/architecture/kind for a host action: 5183 // Host Action 1 -> OffloadAction -> Host Action 2 5184 // 5185 // b) Set a toolchain/architecture/kind for a device action; 5186 // Device Action 1 -> OffloadAction -> Device Action 2 5187 // 5188 // c) Specify a device dependence to a host action; 5189 // Device Action 1 _ 5190 // \ 5191 // Host Action 1 ---> OffloadAction -> Host Action 2 5192 // 5193 // d) Specify a host dependence to a device action. 5194 // Host Action 1 _ 5195 // \ 5196 // Device Action 1 ---> OffloadAction -> Device Action 2 5197 // 5198 // For a) and b), we just return the job generated for the dependence. For 5199 // c) and d) we override the current action with the host/device dependence 5200 // if the current toolchain is host/device and set the offload dependences 5201 // info with the jobs obtained from the device/host dependence(s). 5202 5203 // If there is a single device option, just generate the job for it. 5204 if (OA->hasSingleDeviceDependence()) { 5205 InputInfoList DevA; 5206 OA->doOnEachDeviceDependence([&](Action *DepA, const ToolChain *DepTC, 5207 const char *DepBoundArch) { 5208 DevA = 5209 BuildJobsForAction(C, DepA, DepTC, DepBoundArch, AtTopLevel, 5210 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, 5211 CachedResults, DepA->getOffloadingDeviceKind()); 5212 }); 5213 return DevA; 5214 } 5215 5216 // If 'Action 2' is host, we generate jobs for the device dependences and 5217 // override the current action with the host dependence. Otherwise, we 5218 // generate the host dependences and override the action with the device 5219 // dependence. The dependences can't therefore be a top-level action. 5220 OA->doOnEachDependence( 5221 /*IsHostDependence=*/BuildingForOffloadDevice, 5222 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5223 OffloadDependencesInputInfo.append(BuildJobsForAction( 5224 C, DepA, DepTC, DepBoundArch, /*AtTopLevel=*/false, 5225 /*MultipleArchs*/ !!DepBoundArch, LinkingOutput, CachedResults, 5226 DepA->getOffloadingDeviceKind())); 5227 }); 5228 5229 A = BuildingForOffloadDevice 5230 ? OA->getSingleDeviceDependence(/*DoNotConsiderHostActions=*/true) 5231 : OA->getHostDependence(); 5232 5233 // We may have already built this action as a part of the offloading 5234 // toolchain, return the cached input if so. 5235 std::pair<const Action *, std::string> ActionTC = { 5236 OA->getHostDependence(), 5237 GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5238 if (CachedResults.find(ActionTC) != CachedResults.end()) { 5239 InputInfoList Inputs = CachedResults[ActionTC]; 5240 Inputs.append(OffloadDependencesInputInfo); 5241 return Inputs; 5242 } 5243 } 5244 5245 if (const InputAction *IA = dyn_cast<InputAction>(A)) { 5246 // FIXME: It would be nice to not claim this here; maybe the old scheme of 5247 // just using Args was better? 5248 const Arg &Input = IA->getInputArg(); 5249 Input.claim(); 5250 if (Input.getOption().matches(options::OPT_INPUT)) { 5251 const char *Name = Input.getValue(); 5252 return {InputInfo(A, Name, /* _BaseInput = */ Name)}; 5253 } 5254 return {InputInfo(A, &Input, /* _BaseInput = */ "")}; 5255 } 5256 5257 if (const BindArchAction *BAA = dyn_cast<BindArchAction>(A)) { 5258 const ToolChain *TC; 5259 StringRef ArchName = BAA->getArchName(); 5260 5261 if (!ArchName.empty()) 5262 TC = &getToolChain(C.getArgs(), 5263 computeTargetTriple(*this, TargetTriple, 5264 C.getArgs(), ArchName)); 5265 else 5266 TC = &C.getDefaultToolChain(); 5267 5268 return BuildJobsForAction(C, *BAA->input_begin(), TC, ArchName, AtTopLevel, 5269 MultipleArchs, LinkingOutput, CachedResults, 5270 TargetDeviceOffloadKind); 5271 } 5272 5273 5274 ActionList Inputs = A->getInputs(); 5275 5276 const JobAction *JA = cast<JobAction>(A); 5277 ActionList CollapsedOffloadActions; 5278 5279 ToolSelector TS(JA, *TC, C, isSaveTempsEnabled(), 5280 embedBitcodeInObject() && !isUsingLTO()); 5281 const Tool *T = TS.getTool(Inputs, CollapsedOffloadActions); 5282 5283 if (!T) 5284 return {InputInfo()}; 5285 5286 if (BuildingForOffloadDevice && 5287 A->getOffloadingDeviceKind() == Action::OFK_OpenMP) { 5288 if (TC->getTriple().isAMDGCN()) { 5289 // AMDGCN treats backend and assemble actions as no-op because 5290 // linker does not support object files. 5291 if (const BackendJobAction *BA = dyn_cast<BackendJobAction>(A)) { 5292 return BuildJobsForAction(C, *BA->input_begin(), TC, BoundArch, 5293 AtTopLevel, MultipleArchs, LinkingOutput, 5294 CachedResults, TargetDeviceOffloadKind); 5295 } 5296 5297 if (const AssembleJobAction *AA = dyn_cast<AssembleJobAction>(A)) { 5298 return BuildJobsForAction(C, *AA->input_begin(), TC, BoundArch, 5299 AtTopLevel, MultipleArchs, LinkingOutput, 5300 CachedResults, TargetDeviceOffloadKind); 5301 } 5302 } 5303 } 5304 5305 // If we've collapsed action list that contained OffloadAction we 5306 // need to build jobs for host/device-side inputs it may have held. 5307 for (const auto *OA : CollapsedOffloadActions) 5308 cast<OffloadAction>(OA)->doOnEachDependence( 5309 /*IsHostDependence=*/BuildingForOffloadDevice, 5310 [&](Action *DepA, const ToolChain *DepTC, const char *DepBoundArch) { 5311 OffloadDependencesInputInfo.append(BuildJobsForAction( 5312 C, DepA, DepTC, DepBoundArch, /* AtTopLevel */ false, 5313 /*MultipleArchs=*/!!DepBoundArch, LinkingOutput, CachedResults, 5314 DepA->getOffloadingDeviceKind())); 5315 }); 5316 5317 // Only use pipes when there is exactly one input. 5318 InputInfoList InputInfos; 5319 for (const Action *Input : Inputs) { 5320 // Treat dsymutil and verify sub-jobs as being at the top-level too, they 5321 // shouldn't get temporary output names. 5322 // FIXME: Clean this up. 5323 bool SubJobAtTopLevel = 5324 AtTopLevel && (isa<DsymutilJobAction>(A) || isa<VerifyJobAction>(A)); 5325 InputInfos.append(BuildJobsForAction( 5326 C, Input, TC, BoundArch, SubJobAtTopLevel, MultipleArchs, LinkingOutput, 5327 CachedResults, A->getOffloadingDeviceKind())); 5328 } 5329 5330 // Always use the first file input as the base input. 5331 const char *BaseInput = InputInfos[0].getBaseInput(); 5332 for (auto &Info : InputInfos) { 5333 if (Info.isFilename()) { 5334 BaseInput = Info.getBaseInput(); 5335 break; 5336 } 5337 } 5338 5339 // ... except dsymutil actions, which use their actual input as the base 5340 // input. 5341 if (JA->getType() == types::TY_dSYM) 5342 BaseInput = InputInfos[0].getFilename(); 5343 5344 // ... and in header module compilations, which use the module name. 5345 if (auto *ModuleJA = dyn_cast<HeaderModulePrecompileJobAction>(JA)) 5346 BaseInput = ModuleJA->getModuleName(); 5347 5348 // Append outputs of offload device jobs to the input list 5349 if (!OffloadDependencesInputInfo.empty()) 5350 InputInfos.append(OffloadDependencesInputInfo.begin(), 5351 OffloadDependencesInputInfo.end()); 5352 5353 // Set the effective triple of the toolchain for the duration of this job. 5354 llvm::Triple EffectiveTriple; 5355 const ToolChain &ToolTC = T->getToolChain(); 5356 const ArgList &Args = 5357 C.getArgsForToolChain(TC, BoundArch, A->getOffloadingDeviceKind()); 5358 if (InputInfos.size() != 1) { 5359 EffectiveTriple = llvm::Triple(ToolTC.ComputeEffectiveClangTriple(Args)); 5360 } else { 5361 // Pass along the input type if it can be unambiguously determined. 5362 EffectiveTriple = llvm::Triple( 5363 ToolTC.ComputeEffectiveClangTriple(Args, InputInfos[0].getType())); 5364 } 5365 RegisterEffectiveTriple TripleRAII(ToolTC, EffectiveTriple); 5366 5367 // Determine the place to write output to, if any. 5368 InputInfo Result; 5369 InputInfoList UnbundlingResults; 5370 if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(JA)) { 5371 // If we have an unbundling job, we need to create results for all the 5372 // outputs. We also update the results cache so that other actions using 5373 // this unbundling action can get the right results. 5374 for (auto &UI : UA->getDependentActionsInfo()) { 5375 assert(UI.DependentOffloadKind != Action::OFK_None && 5376 "Unbundling with no offloading??"); 5377 5378 // Unbundling actions are never at the top level. When we generate the 5379 // offloading prefix, we also do that for the host file because the 5380 // unbundling action does not change the type of the output which can 5381 // cause a overwrite. 5382 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5383 UI.DependentOffloadKind, 5384 UI.DependentToolChain->getTriple().normalize(), 5385 /*CreatePrefixForHost=*/true); 5386 auto CurI = InputInfo( 5387 UA, 5388 GetNamedOutputPath(C, *UA, BaseInput, UI.DependentBoundArch, 5389 /*AtTopLevel=*/false, 5390 MultipleArchs || 5391 UI.DependentOffloadKind == Action::OFK_HIP, 5392 OffloadingPrefix), 5393 BaseInput); 5394 // Save the unbundling result. 5395 UnbundlingResults.push_back(CurI); 5396 5397 // Get the unique string identifier for this dependence and cache the 5398 // result. 5399 StringRef Arch; 5400 if (TargetDeviceOffloadKind == Action::OFK_HIP) { 5401 if (UI.DependentOffloadKind == Action::OFK_Host) 5402 Arch = StringRef(); 5403 else 5404 Arch = UI.DependentBoundArch; 5405 } else 5406 Arch = BoundArch; 5407 5408 CachedResults[{A, GetTriplePlusArchString(UI.DependentToolChain, Arch, 5409 UI.DependentOffloadKind)}] = { 5410 CurI}; 5411 } 5412 5413 // Now that we have all the results generated, select the one that should be 5414 // returned for the current depending action. 5415 std::pair<const Action *, std::string> ActionTC = { 5416 A, GetTriplePlusArchString(TC, BoundArch, TargetDeviceOffloadKind)}; 5417 assert(CachedResults.find(ActionTC) != CachedResults.end() && 5418 "Result does not exist??"); 5419 Result = CachedResults[ActionTC].front(); 5420 } else if (JA->getType() == types::TY_Nothing) 5421 Result = {InputInfo(A, BaseInput)}; 5422 else { 5423 // We only have to generate a prefix for the host if this is not a top-level 5424 // action. 5425 std::string OffloadingPrefix = Action::GetOffloadingFileNamePrefix( 5426 A->getOffloadingDeviceKind(), TC->getTriple().normalize(), 5427 /*CreatePrefixForHost=*/!!A->getOffloadingHostActiveKinds() && 5428 !AtTopLevel); 5429 if (isa<OffloadWrapperJobAction>(JA)) { 5430 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5431 BaseInput = FinalOutput->getValue(); 5432 else 5433 BaseInput = getDefaultImageName(); 5434 BaseInput = 5435 C.getArgs().MakeArgString(std::string(BaseInput) + "-wrapper"); 5436 } 5437 Result = InputInfo(A, GetNamedOutputPath(C, *JA, BaseInput, BoundArch, 5438 AtTopLevel, MultipleArchs, 5439 OffloadingPrefix), 5440 BaseInput); 5441 } 5442 5443 if (CCCPrintBindings && !CCGenDiagnostics) { 5444 llvm::errs() << "# \"" << T->getToolChain().getTripleString() << '"' 5445 << " - \"" << T->getName() << "\", inputs: ["; 5446 for (unsigned i = 0, e = InputInfos.size(); i != e; ++i) { 5447 llvm::errs() << InputInfos[i].getAsString(); 5448 if (i + 1 != e) 5449 llvm::errs() << ", "; 5450 } 5451 if (UnbundlingResults.empty()) 5452 llvm::errs() << "], output: " << Result.getAsString() << "\n"; 5453 else { 5454 llvm::errs() << "], outputs: ["; 5455 for (unsigned i = 0, e = UnbundlingResults.size(); i != e; ++i) { 5456 llvm::errs() << UnbundlingResults[i].getAsString(); 5457 if (i + 1 != e) 5458 llvm::errs() << ", "; 5459 } 5460 llvm::errs() << "] \n"; 5461 } 5462 } else { 5463 if (UnbundlingResults.empty()) 5464 T->ConstructJob( 5465 C, *JA, Result, InputInfos, 5466 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5467 LinkingOutput); 5468 else 5469 T->ConstructJobMultipleOutputs( 5470 C, *JA, UnbundlingResults, InputInfos, 5471 C.getArgsForToolChain(TC, BoundArch, JA->getOffloadingDeviceKind()), 5472 LinkingOutput); 5473 } 5474 return {Result}; 5475 } 5476 5477 const char *Driver::getDefaultImageName() const { 5478 llvm::Triple Target(llvm::Triple::normalize(TargetTriple)); 5479 return Target.isOSWindows() ? "a.exe" : "a.out"; 5480 } 5481 5482 /// Create output filename based on ArgValue, which could either be a 5483 /// full filename, filename without extension, or a directory. If ArgValue 5484 /// does not provide a filename, then use BaseName, and use the extension 5485 /// suitable for FileType. 5486 static const char *MakeCLOutputFilename(const ArgList &Args, StringRef ArgValue, 5487 StringRef BaseName, 5488 types::ID FileType) { 5489 SmallString<128> Filename = ArgValue; 5490 5491 if (ArgValue.empty()) { 5492 // If the argument is empty, output to BaseName in the current dir. 5493 Filename = BaseName; 5494 } else if (llvm::sys::path::is_separator(Filename.back())) { 5495 // If the argument is a directory, output to BaseName in that dir. 5496 llvm::sys::path::append(Filename, BaseName); 5497 } 5498 5499 if (!llvm::sys::path::has_extension(ArgValue)) { 5500 // If the argument didn't provide an extension, then set it. 5501 const char *Extension = types::getTypeTempSuffix(FileType, true); 5502 5503 if (FileType == types::TY_Image && 5504 Args.hasArg(options::OPT__SLASH_LD, options::OPT__SLASH_LDd)) { 5505 // The output file is a dll. 5506 Extension = "dll"; 5507 } 5508 5509 llvm::sys::path::replace_extension(Filename, Extension); 5510 } 5511 5512 return Args.MakeArgString(Filename.c_str()); 5513 } 5514 5515 static bool HasPreprocessOutput(const Action &JA) { 5516 if (isa<PreprocessJobAction>(JA)) 5517 return true; 5518 if (isa<OffloadAction>(JA) && isa<PreprocessJobAction>(JA.getInputs()[0])) 5519 return true; 5520 if (isa<OffloadBundlingJobAction>(JA) && 5521 HasPreprocessOutput(*(JA.getInputs()[0]))) 5522 return true; 5523 return false; 5524 } 5525 5526 const char *Driver::GetNamedOutputPath(Compilation &C, const JobAction &JA, 5527 const char *BaseInput, 5528 StringRef OrigBoundArch, bool AtTopLevel, 5529 bool MultipleArchs, 5530 StringRef OffloadingPrefix) const { 5531 std::string BoundArch = OrigBoundArch.str(); 5532 if (is_style_windows(llvm::sys::path::Style::native)) { 5533 // BoundArch may contains ':', which is invalid in file names on Windows, 5534 // therefore replace it with '%'. 5535 std::replace(BoundArch.begin(), BoundArch.end(), ':', '@'); 5536 } 5537 5538 llvm::PrettyStackTraceString CrashInfo("Computing output path"); 5539 // Output to a user requested destination? 5540 if (AtTopLevel && !isa<DsymutilJobAction>(JA) && !isa<VerifyJobAction>(JA)) { 5541 if (Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o)) 5542 return C.addResultFile(FinalOutput->getValue(), &JA); 5543 } 5544 5545 // For /P, preprocess to file named after BaseInput. 5546 if (C.getArgs().hasArg(options::OPT__SLASH_P)) { 5547 assert(AtTopLevel && isa<PreprocessJobAction>(JA)); 5548 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5549 StringRef NameArg; 5550 if (Arg *A = C.getArgs().getLastArg(options::OPT__SLASH_Fi)) 5551 NameArg = A->getValue(); 5552 return C.addResultFile( 5553 MakeCLOutputFilename(C.getArgs(), NameArg, BaseName, types::TY_PP_C), 5554 &JA); 5555 } 5556 5557 // Default to writing to stdout? 5558 if (AtTopLevel && !CCGenDiagnostics && HasPreprocessOutput(JA)) { 5559 return "-"; 5560 } 5561 5562 if (JA.getType() == types::TY_ModuleFile && 5563 C.getArgs().getLastArg(options::OPT_module_file_info)) { 5564 return "-"; 5565 } 5566 5567 // Is this the assembly listing for /FA? 5568 if (JA.getType() == types::TY_PP_Asm && 5569 (C.getArgs().hasArg(options::OPT__SLASH_FA) || 5570 C.getArgs().hasArg(options::OPT__SLASH_Fa))) { 5571 // Use /Fa and the input filename to determine the asm file name. 5572 StringRef BaseName = llvm::sys::path::filename(BaseInput); 5573 StringRef FaValue = C.getArgs().getLastArgValue(options::OPT__SLASH_Fa); 5574 return C.addResultFile( 5575 MakeCLOutputFilename(C.getArgs(), FaValue, BaseName, JA.getType()), 5576 &JA); 5577 } 5578 5579 // Output to a temporary file? 5580 if ((!AtTopLevel && !isSaveTempsEnabled() && 5581 !C.getArgs().hasArg(options::OPT__SLASH_Fo)) || 5582 CCGenDiagnostics) { 5583 StringRef Name = llvm::sys::path::filename(BaseInput); 5584 std::pair<StringRef, StringRef> Split = Name.split('.'); 5585 SmallString<128> TmpName; 5586 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5587 Arg *A = C.getArgs().getLastArg(options::OPT_fcrash_diagnostics_dir); 5588 if (CCGenDiagnostics && A) { 5589 SmallString<128> CrashDirectory(A->getValue()); 5590 if (!getVFS().exists(CrashDirectory)) 5591 llvm::sys::fs::create_directories(CrashDirectory); 5592 llvm::sys::path::append(CrashDirectory, Split.first); 5593 const char *Middle = Suffix ? "-%%%%%%." : "-%%%%%%"; 5594 std::error_code EC = llvm::sys::fs::createUniqueFile( 5595 CrashDirectory + Middle + Suffix, TmpName); 5596 if (EC) { 5597 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5598 return ""; 5599 } 5600 } else { 5601 if (MultipleArchs && !BoundArch.empty()) { 5602 TmpName = GetTemporaryDirectory(Split.first); 5603 llvm::sys::path::append(TmpName, 5604 Split.first + "-" + BoundArch + "." + Suffix); 5605 } else { 5606 TmpName = GetTemporaryPath(Split.first, Suffix); 5607 } 5608 } 5609 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5610 } 5611 5612 SmallString<128> BasePath(BaseInput); 5613 SmallString<128> ExternalPath(""); 5614 StringRef BaseName; 5615 5616 // Dsymutil actions should use the full path. 5617 if (isa<DsymutilJobAction>(JA) && C.getArgs().hasArg(options::OPT_dsym_dir)) { 5618 ExternalPath += C.getArgs().getLastArg(options::OPT_dsym_dir)->getValue(); 5619 // We use posix style here because the tests (specifically 5620 // darwin-dsymutil.c) demonstrate that posix style paths are acceptable 5621 // even on Windows and if we don't then the similar test covering this 5622 // fails. 5623 llvm::sys::path::append(ExternalPath, llvm::sys::path::Style::posix, 5624 llvm::sys::path::filename(BasePath)); 5625 BaseName = ExternalPath; 5626 } else if (isa<DsymutilJobAction>(JA) || isa<VerifyJobAction>(JA)) 5627 BaseName = BasePath; 5628 else 5629 BaseName = llvm::sys::path::filename(BasePath); 5630 5631 // Determine what the derived output name should be. 5632 const char *NamedOutput; 5633 5634 if ((JA.getType() == types::TY_Object || JA.getType() == types::TY_LTO_BC) && 5635 C.getArgs().hasArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o)) { 5636 // The /Fo or /o flag decides the object filename. 5637 StringRef Val = 5638 C.getArgs() 5639 .getLastArg(options::OPT__SLASH_Fo, options::OPT__SLASH_o) 5640 ->getValue(); 5641 NamedOutput = 5642 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Object); 5643 } else if (JA.getType() == types::TY_Image && 5644 C.getArgs().hasArg(options::OPT__SLASH_Fe, 5645 options::OPT__SLASH_o)) { 5646 // The /Fe or /o flag names the linked file. 5647 StringRef Val = 5648 C.getArgs() 5649 .getLastArg(options::OPT__SLASH_Fe, options::OPT__SLASH_o) 5650 ->getValue(); 5651 NamedOutput = 5652 MakeCLOutputFilename(C.getArgs(), Val, BaseName, types::TY_Image); 5653 } else if (JA.getType() == types::TY_Image) { 5654 if (IsCLMode()) { 5655 // clang-cl uses BaseName for the executable name. 5656 NamedOutput = 5657 MakeCLOutputFilename(C.getArgs(), "", BaseName, types::TY_Image); 5658 } else { 5659 SmallString<128> Output(getDefaultImageName()); 5660 // HIP image for device compilation with -fno-gpu-rdc is per compilation 5661 // unit. 5662 bool IsHIPNoRDC = JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5663 !C.getArgs().hasFlag(options::OPT_fgpu_rdc, 5664 options::OPT_fno_gpu_rdc, false); 5665 if (IsHIPNoRDC) { 5666 Output = BaseName; 5667 llvm::sys::path::replace_extension(Output, ""); 5668 } 5669 Output += OffloadingPrefix; 5670 if (MultipleArchs && !BoundArch.empty()) { 5671 Output += "-"; 5672 Output.append(BoundArch); 5673 } 5674 if (IsHIPNoRDC) 5675 Output += ".out"; 5676 NamedOutput = C.getArgs().MakeArgString(Output.c_str()); 5677 } 5678 } else if (JA.getType() == types::TY_PCH && IsCLMode()) { 5679 NamedOutput = C.getArgs().MakeArgString(GetClPchPath(C, BaseName)); 5680 } else { 5681 const char *Suffix = types::getTypeTempSuffix(JA.getType(), IsCLMode()); 5682 assert(Suffix && "All types used for output should have a suffix."); 5683 5684 std::string::size_type End = std::string::npos; 5685 if (!types::appendSuffixForType(JA.getType())) 5686 End = BaseName.rfind('.'); 5687 SmallString<128> Suffixed(BaseName.substr(0, End)); 5688 Suffixed += OffloadingPrefix; 5689 if (MultipleArchs && !BoundArch.empty()) { 5690 Suffixed += "-"; 5691 Suffixed.append(BoundArch); 5692 } 5693 // When using both -save-temps and -emit-llvm, use a ".tmp.bc" suffix for 5694 // the unoptimized bitcode so that it does not get overwritten by the ".bc" 5695 // optimized bitcode output. 5696 auto IsHIPRDCInCompilePhase = [](const JobAction &JA, 5697 const llvm::opt::DerivedArgList &Args) { 5698 // The relocatable compilation in HIP implies -emit-llvm. Similarly, use a 5699 // ".tmp.bc" suffix for the unoptimized bitcode (generated in the compile 5700 // phase.) 5701 return isa<CompileJobAction>(JA) && 5702 JA.getOffloadingDeviceKind() == Action::OFK_HIP && 5703 Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, 5704 false); 5705 }; 5706 if (!AtTopLevel && JA.getType() == types::TY_LLVM_BC && 5707 (C.getArgs().hasArg(options::OPT_emit_llvm) || 5708 IsHIPRDCInCompilePhase(JA, C.getArgs()))) 5709 Suffixed += ".tmp"; 5710 Suffixed += '.'; 5711 Suffixed += Suffix; 5712 NamedOutput = C.getArgs().MakeArgString(Suffixed.c_str()); 5713 } 5714 5715 // Prepend object file path if -save-temps=obj 5716 if (!AtTopLevel && isSaveTempsObj() && C.getArgs().hasArg(options::OPT_o) && 5717 JA.getType() != types::TY_PCH) { 5718 Arg *FinalOutput = C.getArgs().getLastArg(options::OPT_o); 5719 SmallString<128> TempPath(FinalOutput->getValue()); 5720 llvm::sys::path::remove_filename(TempPath); 5721 StringRef OutputFileName = llvm::sys::path::filename(NamedOutput); 5722 llvm::sys::path::append(TempPath, OutputFileName); 5723 NamedOutput = C.getArgs().MakeArgString(TempPath.c_str()); 5724 } 5725 5726 // If we're saving temps and the temp file conflicts with the input file, 5727 // then avoid overwriting input file. 5728 if (!AtTopLevel && isSaveTempsEnabled() && NamedOutput == BaseName) { 5729 bool SameFile = false; 5730 SmallString<256> Result; 5731 llvm::sys::fs::current_path(Result); 5732 llvm::sys::path::append(Result, BaseName); 5733 llvm::sys::fs::equivalent(BaseInput, Result.c_str(), SameFile); 5734 // Must share the same path to conflict. 5735 if (SameFile) { 5736 StringRef Name = llvm::sys::path::filename(BaseInput); 5737 std::pair<StringRef, StringRef> Split = Name.split('.'); 5738 std::string TmpName = GetTemporaryPath( 5739 Split.first, types::getTypeTempSuffix(JA.getType(), IsCLMode())); 5740 return C.addTempFile(C.getArgs().MakeArgString(TmpName)); 5741 } 5742 } 5743 5744 // As an annoying special case, PCH generation doesn't strip the pathname. 5745 if (JA.getType() == types::TY_PCH && !IsCLMode()) { 5746 llvm::sys::path::remove_filename(BasePath); 5747 if (BasePath.empty()) 5748 BasePath = NamedOutput; 5749 else 5750 llvm::sys::path::append(BasePath, NamedOutput); 5751 return C.addResultFile(C.getArgs().MakeArgString(BasePath.c_str()), &JA); 5752 } else { 5753 return C.addResultFile(NamedOutput, &JA); 5754 } 5755 } 5756 5757 std::string Driver::GetFilePath(StringRef Name, const ToolChain &TC) const { 5758 // Search for Name in a list of paths. 5759 auto SearchPaths = [&](const llvm::SmallVectorImpl<std::string> &P) 5760 -> llvm::Optional<std::string> { 5761 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5762 // attempting to use this prefix when looking for file paths. 5763 for (const auto &Dir : P) { 5764 if (Dir.empty()) 5765 continue; 5766 SmallString<128> P(Dir[0] == '=' ? SysRoot + Dir.substr(1) : Dir); 5767 llvm::sys::path::append(P, Name); 5768 if (llvm::sys::fs::exists(Twine(P))) 5769 return std::string(P); 5770 } 5771 return None; 5772 }; 5773 5774 if (auto P = SearchPaths(PrefixDirs)) 5775 return *P; 5776 5777 SmallString<128> R(ResourceDir); 5778 llvm::sys::path::append(R, Name); 5779 if (llvm::sys::fs::exists(Twine(R))) 5780 return std::string(R.str()); 5781 5782 SmallString<128> P(TC.getCompilerRTPath()); 5783 llvm::sys::path::append(P, Name); 5784 if (llvm::sys::fs::exists(Twine(P))) 5785 return std::string(P.str()); 5786 5787 SmallString<128> D(Dir); 5788 llvm::sys::path::append(D, "..", Name); 5789 if (llvm::sys::fs::exists(Twine(D))) 5790 return std::string(D.str()); 5791 5792 if (auto P = SearchPaths(TC.getLibraryPaths())) 5793 return *P; 5794 5795 if (auto P = SearchPaths(TC.getFilePaths())) 5796 return *P; 5797 5798 return std::string(Name); 5799 } 5800 5801 void Driver::generatePrefixedToolNames( 5802 StringRef Tool, const ToolChain &TC, 5803 SmallVectorImpl<std::string> &Names) const { 5804 // FIXME: Needs a better variable than TargetTriple 5805 Names.emplace_back((TargetTriple + "-" + Tool).str()); 5806 Names.emplace_back(Tool); 5807 } 5808 5809 static bool ScanDirForExecutable(SmallString<128> &Dir, StringRef Name) { 5810 llvm::sys::path::append(Dir, Name); 5811 if (llvm::sys::fs::can_execute(Twine(Dir))) 5812 return true; 5813 llvm::sys::path::remove_filename(Dir); 5814 return false; 5815 } 5816 5817 std::string Driver::GetProgramPath(StringRef Name, const ToolChain &TC) const { 5818 SmallVector<std::string, 2> TargetSpecificExecutables; 5819 generatePrefixedToolNames(Name, TC, TargetSpecificExecutables); 5820 5821 // Respect a limited subset of the '-Bprefix' functionality in GCC by 5822 // attempting to use this prefix when looking for program paths. 5823 for (const auto &PrefixDir : PrefixDirs) { 5824 if (llvm::sys::fs::is_directory(PrefixDir)) { 5825 SmallString<128> P(PrefixDir); 5826 if (ScanDirForExecutable(P, Name)) 5827 return std::string(P.str()); 5828 } else { 5829 SmallString<128> P((PrefixDir + Name).str()); 5830 if (llvm::sys::fs::can_execute(Twine(P))) 5831 return std::string(P.str()); 5832 } 5833 } 5834 5835 const ToolChain::path_list &List = TC.getProgramPaths(); 5836 for (const auto &TargetSpecificExecutable : TargetSpecificExecutables) { 5837 // For each possible name of the tool look for it in 5838 // program paths first, then the path. 5839 // Higher priority names will be first, meaning that 5840 // a higher priority name in the path will be found 5841 // instead of a lower priority name in the program path. 5842 // E.g. <triple>-gcc on the path will be found instead 5843 // of gcc in the program path 5844 for (const auto &Path : List) { 5845 SmallString<128> P(Path); 5846 if (ScanDirForExecutable(P, TargetSpecificExecutable)) 5847 return std::string(P.str()); 5848 } 5849 5850 // Fall back to the path 5851 if (llvm::ErrorOr<std::string> P = 5852 llvm::sys::findProgramByName(TargetSpecificExecutable)) 5853 return *P; 5854 } 5855 5856 return std::string(Name); 5857 } 5858 5859 std::string Driver::GetTemporaryPath(StringRef Prefix, StringRef Suffix) const { 5860 SmallString<128> Path; 5861 std::error_code EC = llvm::sys::fs::createTemporaryFile(Prefix, Suffix, Path); 5862 if (EC) { 5863 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5864 return ""; 5865 } 5866 5867 return std::string(Path.str()); 5868 } 5869 5870 std::string Driver::GetTemporaryDirectory(StringRef Prefix) const { 5871 SmallString<128> Path; 5872 std::error_code EC = llvm::sys::fs::createUniqueDirectory(Prefix, Path); 5873 if (EC) { 5874 Diag(clang::diag::err_unable_to_make_temp) << EC.message(); 5875 return ""; 5876 } 5877 5878 return std::string(Path.str()); 5879 } 5880 5881 std::string Driver::GetClPchPath(Compilation &C, StringRef BaseName) const { 5882 SmallString<128> Output; 5883 if (Arg *FpArg = C.getArgs().getLastArg(options::OPT__SLASH_Fp)) { 5884 // FIXME: If anybody needs it, implement this obscure rule: 5885 // "If you specify a directory without a file name, the default file name 5886 // is VCx0.pch., where x is the major version of Visual C++ in use." 5887 Output = FpArg->getValue(); 5888 5889 // "If you do not specify an extension as part of the path name, an 5890 // extension of .pch is assumed. " 5891 if (!llvm::sys::path::has_extension(Output)) 5892 Output += ".pch"; 5893 } else { 5894 if (Arg *YcArg = C.getArgs().getLastArg(options::OPT__SLASH_Yc)) 5895 Output = YcArg->getValue(); 5896 if (Output.empty()) 5897 Output = BaseName; 5898 llvm::sys::path::replace_extension(Output, ".pch"); 5899 } 5900 return std::string(Output.str()); 5901 } 5902 5903 const ToolChain &Driver::getToolChain(const ArgList &Args, 5904 const llvm::Triple &Target) const { 5905 5906 auto &TC = ToolChains[Target.str()]; 5907 if (!TC) { 5908 switch (Target.getOS()) { 5909 case llvm::Triple::AIX: 5910 TC = std::make_unique<toolchains::AIX>(*this, Target, Args); 5911 break; 5912 case llvm::Triple::Haiku: 5913 TC = std::make_unique<toolchains::Haiku>(*this, Target, Args); 5914 break; 5915 case llvm::Triple::Ananas: 5916 TC = std::make_unique<toolchains::Ananas>(*this, Target, Args); 5917 break; 5918 case llvm::Triple::CloudABI: 5919 TC = std::make_unique<toolchains::CloudABI>(*this, Target, Args); 5920 break; 5921 case llvm::Triple::Darwin: 5922 case llvm::Triple::MacOSX: 5923 case llvm::Triple::IOS: 5924 case llvm::Triple::TvOS: 5925 case llvm::Triple::WatchOS: 5926 case llvm::Triple::DriverKit: 5927 TC = std::make_unique<toolchains::DarwinClang>(*this, Target, Args); 5928 break; 5929 case llvm::Triple::DragonFly: 5930 TC = std::make_unique<toolchains::DragonFly>(*this, Target, Args); 5931 break; 5932 case llvm::Triple::OpenBSD: 5933 TC = std::make_unique<toolchains::OpenBSD>(*this, Target, Args); 5934 break; 5935 case llvm::Triple::NetBSD: 5936 TC = std::make_unique<toolchains::NetBSD>(*this, Target, Args); 5937 break; 5938 case llvm::Triple::FreeBSD: 5939 if (Target.isPPC()) 5940 TC = std::make_unique<toolchains::PPCFreeBSDToolChain>(*this, Target, 5941 Args); 5942 else 5943 TC = std::make_unique<toolchains::FreeBSD>(*this, Target, Args); 5944 break; 5945 case llvm::Triple::Minix: 5946 TC = std::make_unique<toolchains::Minix>(*this, Target, Args); 5947 break; 5948 case llvm::Triple::Linux: 5949 case llvm::Triple::ELFIAMCU: 5950 if (Target.getArch() == llvm::Triple::hexagon) 5951 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 5952 Args); 5953 else if ((Target.getVendor() == llvm::Triple::MipsTechnologies) && 5954 !Target.hasEnvironment()) 5955 TC = std::make_unique<toolchains::MipsLLVMToolChain>(*this, Target, 5956 Args); 5957 else if (Target.isPPC()) 5958 TC = std::make_unique<toolchains::PPCLinuxToolChain>(*this, Target, 5959 Args); 5960 else if (Target.getArch() == llvm::Triple::ve) 5961 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 5962 5963 else 5964 TC = std::make_unique<toolchains::Linux>(*this, Target, Args); 5965 break; 5966 case llvm::Triple::NaCl: 5967 TC = std::make_unique<toolchains::NaClToolChain>(*this, Target, Args); 5968 break; 5969 case llvm::Triple::Fuchsia: 5970 TC = std::make_unique<toolchains::Fuchsia>(*this, Target, Args); 5971 break; 5972 case llvm::Triple::Solaris: 5973 TC = std::make_unique<toolchains::Solaris>(*this, Target, Args); 5974 break; 5975 case llvm::Triple::AMDHSA: 5976 TC = std::make_unique<toolchains::ROCMToolChain>(*this, Target, Args); 5977 break; 5978 case llvm::Triple::AMDPAL: 5979 case llvm::Triple::Mesa3D: 5980 TC = std::make_unique<toolchains::AMDGPUToolChain>(*this, Target, Args); 5981 break; 5982 case llvm::Triple::Win32: 5983 switch (Target.getEnvironment()) { 5984 default: 5985 if (Target.isOSBinFormatELF()) 5986 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 5987 else if (Target.isOSBinFormatMachO()) 5988 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 5989 else 5990 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 5991 break; 5992 case llvm::Triple::GNU: 5993 TC = std::make_unique<toolchains::MinGW>(*this, Target, Args); 5994 break; 5995 case llvm::Triple::Itanium: 5996 TC = std::make_unique<toolchains::CrossWindowsToolChain>(*this, Target, 5997 Args); 5998 break; 5999 case llvm::Triple::MSVC: 6000 case llvm::Triple::UnknownEnvironment: 6001 if (Args.getLastArgValue(options::OPT_fuse_ld_EQ) 6002 .startswith_insensitive("bfd")) 6003 TC = std::make_unique<toolchains::CrossWindowsToolChain>( 6004 *this, Target, Args); 6005 else 6006 TC = 6007 std::make_unique<toolchains::MSVCToolChain>(*this, Target, Args); 6008 break; 6009 } 6010 break; 6011 case llvm::Triple::PS4: 6012 TC = std::make_unique<toolchains::PS4CPU>(*this, Target, Args); 6013 break; 6014 case llvm::Triple::PS5: 6015 TC = std::make_unique<toolchains::PS5CPU>(*this, Target, Args); 6016 break; 6017 case llvm::Triple::Contiki: 6018 TC = std::make_unique<toolchains::Contiki>(*this, Target, Args); 6019 break; 6020 case llvm::Triple::Hurd: 6021 TC = std::make_unique<toolchains::Hurd>(*this, Target, Args); 6022 break; 6023 case llvm::Triple::ZOS: 6024 TC = std::make_unique<toolchains::ZOS>(*this, Target, Args); 6025 break; 6026 case llvm::Triple::ShaderModel: 6027 TC = std::make_unique<toolchains::HLSLToolChain>(*this, Target, Args); 6028 break; 6029 default: 6030 // Of these targets, Hexagon is the only one that might have 6031 // an OS of Linux, in which case it got handled above already. 6032 switch (Target.getArch()) { 6033 case llvm::Triple::tce: 6034 TC = std::make_unique<toolchains::TCEToolChain>(*this, Target, Args); 6035 break; 6036 case llvm::Triple::tcele: 6037 TC = std::make_unique<toolchains::TCELEToolChain>(*this, Target, Args); 6038 break; 6039 case llvm::Triple::hexagon: 6040 TC = std::make_unique<toolchains::HexagonToolChain>(*this, Target, 6041 Args); 6042 break; 6043 case llvm::Triple::lanai: 6044 TC = std::make_unique<toolchains::LanaiToolChain>(*this, Target, Args); 6045 break; 6046 case llvm::Triple::xcore: 6047 TC = std::make_unique<toolchains::XCoreToolChain>(*this, Target, Args); 6048 break; 6049 case llvm::Triple::wasm32: 6050 case llvm::Triple::wasm64: 6051 TC = std::make_unique<toolchains::WebAssembly>(*this, Target, Args); 6052 break; 6053 case llvm::Triple::avr: 6054 TC = std::make_unique<toolchains::AVRToolChain>(*this, Target, Args); 6055 break; 6056 case llvm::Triple::msp430: 6057 TC = 6058 std::make_unique<toolchains::MSP430ToolChain>(*this, Target, Args); 6059 break; 6060 case llvm::Triple::riscv32: 6061 case llvm::Triple::riscv64: 6062 if (toolchains::RISCVToolChain::hasGCCToolchain(*this, Args)) 6063 TC = 6064 std::make_unique<toolchains::RISCVToolChain>(*this, Target, Args); 6065 else 6066 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6067 break; 6068 case llvm::Triple::ve: 6069 TC = std::make_unique<toolchains::VEToolChain>(*this, Target, Args); 6070 break; 6071 case llvm::Triple::spirv32: 6072 case llvm::Triple::spirv64: 6073 TC = std::make_unique<toolchains::SPIRVToolChain>(*this, Target, Args); 6074 break; 6075 case llvm::Triple::csky: 6076 TC = std::make_unique<toolchains::CSKYToolChain>(*this, Target, Args); 6077 break; 6078 default: 6079 if (Target.getVendor() == llvm::Triple::Myriad) 6080 TC = std::make_unique<toolchains::MyriadToolChain>(*this, Target, 6081 Args); 6082 else if (toolchains::BareMetal::handlesTarget(Target)) 6083 TC = std::make_unique<toolchains::BareMetal>(*this, Target, Args); 6084 else if (Target.isOSBinFormatELF()) 6085 TC = std::make_unique<toolchains::Generic_ELF>(*this, Target, Args); 6086 else if (Target.isOSBinFormatMachO()) 6087 TC = std::make_unique<toolchains::MachO>(*this, Target, Args); 6088 else 6089 TC = std::make_unique<toolchains::Generic_GCC>(*this, Target, Args); 6090 } 6091 } 6092 } 6093 6094 // Intentionally omitted from the switch above: llvm::Triple::CUDA. CUDA 6095 // compiles always need two toolchains, the CUDA toolchain and the host 6096 // toolchain. So the only valid way to create a CUDA toolchain is via 6097 // CreateOffloadingDeviceToolChains. 6098 6099 return *TC; 6100 } 6101 6102 const ToolChain &Driver::getOffloadingDeviceToolChain( 6103 const ArgList &Args, const llvm::Triple &Target, const ToolChain &HostTC, 6104 const Action::OffloadKind &TargetDeviceOffloadKind) const { 6105 // Use device / host triples as the key into the ToolChains map because the 6106 // device ToolChain we create depends on both. 6107 auto &TC = ToolChains[Target.str() + "/" + HostTC.getTriple().str()]; 6108 if (!TC) { 6109 // Categorized by offload kind > arch rather than OS > arch like 6110 // the normal getToolChain call, as it seems a reasonable way to categorize 6111 // things. 6112 switch (TargetDeviceOffloadKind) { 6113 case Action::OFK_HIP: { 6114 if (Target.getArch() == llvm::Triple::amdgcn && 6115 Target.getVendor() == llvm::Triple::AMD && 6116 Target.getOS() == llvm::Triple::AMDHSA) 6117 TC = std::make_unique<toolchains::HIPAMDToolChain>(*this, Target, 6118 HostTC, Args); 6119 else if (Target.getArch() == llvm::Triple::spirv64 && 6120 Target.getVendor() == llvm::Triple::UnknownVendor && 6121 Target.getOS() == llvm::Triple::UnknownOS) 6122 TC = std::make_unique<toolchains::HIPSPVToolChain>(*this, Target, 6123 HostTC, Args); 6124 break; 6125 } 6126 default: 6127 break; 6128 } 6129 } 6130 6131 return *TC; 6132 } 6133 6134 bool Driver::ShouldUseClangCompiler(const JobAction &JA) const { 6135 // Say "no" if there is not exactly one input of a type clang understands. 6136 if (JA.size() != 1 || 6137 !types::isAcceptedByClang((*JA.input_begin())->getType())) 6138 return false; 6139 6140 // And say "no" if this is not a kind of action clang understands. 6141 if (!isa<PreprocessJobAction>(JA) && !isa<PrecompileJobAction>(JA) && 6142 !isa<CompileJobAction>(JA) && !isa<BackendJobAction>(JA) && 6143 !isa<ExtractAPIJobAction>(JA)) 6144 return false; 6145 6146 return true; 6147 } 6148 6149 bool Driver::ShouldUseFlangCompiler(const JobAction &JA) const { 6150 // Say "no" if there is not exactly one input of a type flang understands. 6151 if (JA.size() != 1 || 6152 !types::isAcceptedByFlang((*JA.input_begin())->getType())) 6153 return false; 6154 6155 // And say "no" if this is not a kind of action flang understands. 6156 if (!isa<PreprocessJobAction>(JA) && !isa<CompileJobAction>(JA) && 6157 !isa<BackendJobAction>(JA)) 6158 return false; 6159 6160 return true; 6161 } 6162 6163 bool Driver::ShouldEmitStaticLibrary(const ArgList &Args) const { 6164 // Only emit static library if the flag is set explicitly. 6165 if (Args.hasArg(options::OPT_emit_static_lib)) 6166 return true; 6167 return false; 6168 } 6169 6170 /// GetReleaseVersion - Parse (([0-9]+)(.([0-9]+)(.([0-9]+)?))?)? and return the 6171 /// grouped values as integers. Numbers which are not provided are set to 0. 6172 /// 6173 /// \return True if the entire string was parsed (9.2), or all groups were 6174 /// parsed (10.3.5extrastuff). 6175 bool Driver::GetReleaseVersion(StringRef Str, unsigned &Major, unsigned &Minor, 6176 unsigned &Micro, bool &HadExtra) { 6177 HadExtra = false; 6178 6179 Major = Minor = Micro = 0; 6180 if (Str.empty()) 6181 return false; 6182 6183 if (Str.consumeInteger(10, Major)) 6184 return false; 6185 if (Str.empty()) 6186 return true; 6187 if (Str[0] != '.') 6188 return false; 6189 6190 Str = Str.drop_front(1); 6191 6192 if (Str.consumeInteger(10, Minor)) 6193 return false; 6194 if (Str.empty()) 6195 return true; 6196 if (Str[0] != '.') 6197 return false; 6198 Str = Str.drop_front(1); 6199 6200 if (Str.consumeInteger(10, Micro)) 6201 return false; 6202 if (!Str.empty()) 6203 HadExtra = true; 6204 return true; 6205 } 6206 6207 /// Parse digits from a string \p Str and fulfill \p Digits with 6208 /// the parsed numbers. This method assumes that the max number of 6209 /// digits to look for is equal to Digits.size(). 6210 /// 6211 /// \return True if the entire string was parsed and there are 6212 /// no extra characters remaining at the end. 6213 bool Driver::GetReleaseVersion(StringRef Str, 6214 MutableArrayRef<unsigned> Digits) { 6215 if (Str.empty()) 6216 return false; 6217 6218 unsigned CurDigit = 0; 6219 while (CurDigit < Digits.size()) { 6220 unsigned Digit; 6221 if (Str.consumeInteger(10, Digit)) 6222 return false; 6223 Digits[CurDigit] = Digit; 6224 if (Str.empty()) 6225 return true; 6226 if (Str[0] != '.') 6227 return false; 6228 Str = Str.drop_front(1); 6229 CurDigit++; 6230 } 6231 6232 // More digits than requested, bail out... 6233 return false; 6234 } 6235 6236 std::pair<unsigned, unsigned> 6237 Driver::getIncludeExcludeOptionFlagMasks(bool IsClCompatMode) const { 6238 unsigned IncludedFlagsBitmask = 0; 6239 unsigned ExcludedFlagsBitmask = options::NoDriverOption; 6240 6241 if (IsClCompatMode) { 6242 // Include CL and Core options. 6243 IncludedFlagsBitmask |= options::CLOption; 6244 IncludedFlagsBitmask |= options::CoreOption; 6245 } else { 6246 ExcludedFlagsBitmask |= options::CLOption; 6247 } 6248 if (IsDXCMode()) { 6249 // Include DXC and Core options. 6250 IncludedFlagsBitmask |= options::DXCOption; 6251 IncludedFlagsBitmask |= options::CoreOption; 6252 } else { 6253 ExcludedFlagsBitmask |= options::DXCOption; 6254 } 6255 return std::make_pair(IncludedFlagsBitmask, ExcludedFlagsBitmask); 6256 } 6257 6258 bool clang::driver::isOptimizationLevelFast(const ArgList &Args) { 6259 return Args.hasFlag(options::OPT_Ofast, options::OPT_O_Group, false); 6260 } 6261 6262 bool clang::driver::willEmitRemarks(const ArgList &Args) { 6263 // -fsave-optimization-record enables it. 6264 if (Args.hasFlag(options::OPT_fsave_optimization_record, 6265 options::OPT_fno_save_optimization_record, false)) 6266 return true; 6267 6268 // -fsave-optimization-record=<format> enables it as well. 6269 if (Args.hasFlag(options::OPT_fsave_optimization_record_EQ, 6270 options::OPT_fno_save_optimization_record, false)) 6271 return true; 6272 6273 // -foptimization-record-file alone enables it too. 6274 if (Args.hasFlag(options::OPT_foptimization_record_file_EQ, 6275 options::OPT_fno_save_optimization_record, false)) 6276 return true; 6277 6278 // -foptimization-record-passes alone enables it too. 6279 if (Args.hasFlag(options::OPT_foptimization_record_passes_EQ, 6280 options::OPT_fno_save_optimization_record, false)) 6281 return true; 6282 return false; 6283 } 6284 6285 llvm::StringRef clang::driver::getDriverMode(StringRef ProgName, 6286 ArrayRef<const char *> Args) { 6287 static const std::string OptName = 6288 getDriverOptTable().getOption(options::OPT_driver_mode).getPrefixedName(); 6289 llvm::StringRef Opt; 6290 for (StringRef Arg : Args) { 6291 if (!Arg.startswith(OptName)) 6292 continue; 6293 Opt = Arg; 6294 } 6295 if (Opt.empty()) 6296 Opt = ToolChain::getTargetAndModeFromProgramName(ProgName).DriverMode; 6297 return Opt.consume_front(OptName) ? Opt : ""; 6298 } 6299 6300 bool driver::IsClangCL(StringRef DriverMode) { return DriverMode.equals("cl"); } 6301