xref: /freebsd/contrib/llvm-project/clang/lib/Driver/ToolChains/Cuda.cpp (revision f9fd7337f63698f33239c58c07bf430198235a22)
1 //===--- Cuda.cpp - Cuda Tool and ToolChain Implementations -----*- C++ -*-===//
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 "Cuda.h"
10 #include "CommonArgs.h"
11 #include "InputInfo.h"
12 #include "clang/Basic/Cuda.h"
13 #include "clang/Config/config.h"
14 #include "clang/Driver/Compilation.h"
15 #include "clang/Driver/Distro.h"
16 #include "clang/Driver/Driver.h"
17 #include "clang/Driver/DriverDiagnostic.h"
18 #include "clang/Driver/Options.h"
19 #include "llvm/Option/ArgList.h"
20 #include "llvm/Support/FileSystem.h"
21 #include "llvm/Support/Host.h"
22 #include "llvm/Support/Path.h"
23 #include "llvm/Support/Process.h"
24 #include "llvm/Support/Program.h"
25 #include "llvm/Support/TargetParser.h"
26 #include "llvm/Support/VirtualFileSystem.h"
27 #include <system_error>
28 
29 using namespace clang::driver;
30 using namespace clang::driver::toolchains;
31 using namespace clang::driver::tools;
32 using namespace clang;
33 using namespace llvm::opt;
34 
35 // Parses the contents of version.txt in an CUDA installation.  It should
36 // contain one line of the from e.g. "CUDA Version 7.5.2".
37 void CudaInstallationDetector::ParseCudaVersionFile(llvm::StringRef V) {
38   Version = CudaVersion::UNKNOWN;
39   if (!V.startswith("CUDA Version "))
40     return;
41   V = V.substr(strlen("CUDA Version "));
42   SmallVector<StringRef,4> VersionParts;
43   V.split(VersionParts, '.');
44   if (VersionParts.size() < 2)
45     return;
46   DetectedVersion = join_items(".", VersionParts[0], VersionParts[1]);
47   Version = CudaStringToVersion(DetectedVersion);
48   if (Version != CudaVersion::UNKNOWN) {
49     // TODO(tra): remove the warning once we have all features of 10.2 and 11.0
50     // implemented.
51     DetectedVersionIsNotSupported = Version > CudaVersion::LATEST_SUPPORTED;
52     return;
53   }
54 
55   Version = CudaVersion::LATEST_SUPPORTED;
56   DetectedVersionIsNotSupported = true;
57 }
58 
59 void CudaInstallationDetector::WarnIfUnsupportedVersion() {
60   if (DetectedVersionIsNotSupported)
61     D.Diag(diag::warn_drv_unknown_cuda_version)
62         << DetectedVersion
63         << CudaVersionToString(CudaVersion::LATEST_SUPPORTED);
64 }
65 
66 CudaInstallationDetector::CudaInstallationDetector(
67     const Driver &D, const llvm::Triple &HostTriple,
68     const llvm::opt::ArgList &Args)
69     : D(D) {
70   struct Candidate {
71     std::string Path;
72     bool StrictChecking;
73 
74     Candidate(std::string Path, bool StrictChecking = false)
75         : Path(Path), StrictChecking(StrictChecking) {}
76   };
77   SmallVector<Candidate, 4> Candidates;
78 
79   // In decreasing order so we prefer newer versions to older versions.
80   std::initializer_list<const char *> Versions = {"8.0", "7.5", "7.0"};
81   auto &FS = D.getVFS();
82 
83   if (Args.hasArg(clang::driver::options::OPT_cuda_path_EQ)) {
84     Candidates.emplace_back(
85         Args.getLastArgValue(clang::driver::options::OPT_cuda_path_EQ).str());
86   } else if (HostTriple.isOSWindows()) {
87     for (const char *Ver : Versions)
88       Candidates.emplace_back(
89           D.SysRoot + "/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v" +
90           Ver);
91   } else {
92     if (!Args.hasArg(clang::driver::options::OPT_cuda_path_ignore_env)) {
93       // Try to find ptxas binary. If the executable is located in a directory
94       // called 'bin/', its parent directory might be a good guess for a valid
95       // CUDA installation.
96       // However, some distributions might installs 'ptxas' to /usr/bin. In that
97       // case the candidate would be '/usr' which passes the following checks
98       // because '/usr/include' exists as well. To avoid this case, we always
99       // check for the directory potentially containing files for libdevice,
100       // even if the user passes -nocudalib.
101       if (llvm::ErrorOr<std::string> ptxas =
102               llvm::sys::findProgramByName("ptxas")) {
103         SmallString<256> ptxasAbsolutePath;
104         llvm::sys::fs::real_path(*ptxas, ptxasAbsolutePath);
105 
106         StringRef ptxasDir = llvm::sys::path::parent_path(ptxasAbsolutePath);
107         if (llvm::sys::path::filename(ptxasDir) == "bin")
108           Candidates.emplace_back(
109               std::string(llvm::sys::path::parent_path(ptxasDir)),
110               /*StrictChecking=*/true);
111       }
112     }
113 
114     Candidates.emplace_back(D.SysRoot + "/usr/local/cuda");
115     for (const char *Ver : Versions)
116       Candidates.emplace_back(D.SysRoot + "/usr/local/cuda-" + Ver);
117 
118     Distro Dist(FS, llvm::Triple(llvm::sys::getProcessTriple()));
119     if (Dist.IsDebian() || Dist.IsUbuntu())
120       // Special case for Debian to have nvidia-cuda-toolkit work
121       // out of the box. More info on http://bugs.debian.org/882505
122       Candidates.emplace_back(D.SysRoot + "/usr/lib/cuda");
123   }
124 
125   bool NoCudaLib = Args.hasArg(options::OPT_nogpulib);
126 
127   for (const auto &Candidate : Candidates) {
128     InstallPath = Candidate.Path;
129     if (InstallPath.empty() || !FS.exists(InstallPath))
130       continue;
131 
132     BinPath = InstallPath + "/bin";
133     IncludePath = InstallPath + "/include";
134     LibDevicePath = InstallPath + "/nvvm/libdevice";
135 
136     if (!(FS.exists(IncludePath) && FS.exists(BinPath)))
137       continue;
138     bool CheckLibDevice = (!NoCudaLib || Candidate.StrictChecking);
139     if (CheckLibDevice && !FS.exists(LibDevicePath))
140       continue;
141 
142     // On Linux, we have both lib and lib64 directories, and we need to choose
143     // based on our triple.  On MacOS, we have only a lib directory.
144     //
145     // It's sufficient for our purposes to be flexible: If both lib and lib64
146     // exist, we choose whichever one matches our triple.  Otherwise, if only
147     // lib exists, we use it.
148     if (HostTriple.isArch64Bit() && FS.exists(InstallPath + "/lib64"))
149       LibPath = InstallPath + "/lib64";
150     else if (FS.exists(InstallPath + "/lib"))
151       LibPath = InstallPath + "/lib";
152     else
153       continue;
154 
155     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> VersionFile =
156         FS.getBufferForFile(InstallPath + "/version.txt");
157     if (!VersionFile) {
158       // CUDA 7.0 doesn't have a version.txt, so guess that's our version if
159       // version.txt isn't present.
160       Version = CudaVersion::CUDA_70;
161     } else {
162       ParseCudaVersionFile((*VersionFile)->getBuffer());
163     }
164 
165     if (Version >= CudaVersion::CUDA_90) {
166       // CUDA-9+ uses single libdevice file for all GPU variants.
167       std::string FilePath = LibDevicePath + "/libdevice.10.bc";
168       if (FS.exists(FilePath)) {
169         for (int Arch = (int)CudaArch::SM_30, E = (int)CudaArch::LAST; Arch < E;
170              ++Arch) {
171           CudaArch GpuArch = static_cast<CudaArch>(Arch);
172           if (!IsNVIDIAGpuArch(GpuArch))
173             continue;
174           std::string GpuArchName(CudaArchToString(GpuArch));
175           LibDeviceMap[GpuArchName] = FilePath;
176         }
177       }
178     } else {
179       std::error_code EC;
180       for (llvm::vfs::directory_iterator LI = FS.dir_begin(LibDevicePath, EC),
181                                          LE;
182            !EC && LI != LE; LI = LI.increment(EC)) {
183         StringRef FilePath = LI->path();
184         StringRef FileName = llvm::sys::path::filename(FilePath);
185         // Process all bitcode filenames that look like
186         // libdevice.compute_XX.YY.bc
187         const StringRef LibDeviceName = "libdevice.";
188         if (!(FileName.startswith(LibDeviceName) && FileName.endswith(".bc")))
189           continue;
190         StringRef GpuArch = FileName.slice(
191             LibDeviceName.size(), FileName.find('.', LibDeviceName.size()));
192         LibDeviceMap[GpuArch] = FilePath.str();
193         // Insert map entries for specific devices with this compute
194         // capability. NVCC's choice of the libdevice library version is
195         // rather peculiar and depends on the CUDA version.
196         if (GpuArch == "compute_20") {
197           LibDeviceMap["sm_20"] = std::string(FilePath);
198           LibDeviceMap["sm_21"] = std::string(FilePath);
199           LibDeviceMap["sm_32"] = std::string(FilePath);
200         } else if (GpuArch == "compute_30") {
201           LibDeviceMap["sm_30"] = std::string(FilePath);
202           if (Version < CudaVersion::CUDA_80) {
203             LibDeviceMap["sm_50"] = std::string(FilePath);
204             LibDeviceMap["sm_52"] = std::string(FilePath);
205             LibDeviceMap["sm_53"] = std::string(FilePath);
206           }
207           LibDeviceMap["sm_60"] = std::string(FilePath);
208           LibDeviceMap["sm_61"] = std::string(FilePath);
209           LibDeviceMap["sm_62"] = std::string(FilePath);
210         } else if (GpuArch == "compute_35") {
211           LibDeviceMap["sm_35"] = std::string(FilePath);
212           LibDeviceMap["sm_37"] = std::string(FilePath);
213         } else if (GpuArch == "compute_50") {
214           if (Version >= CudaVersion::CUDA_80) {
215             LibDeviceMap["sm_50"] = std::string(FilePath);
216             LibDeviceMap["sm_52"] = std::string(FilePath);
217             LibDeviceMap["sm_53"] = std::string(FilePath);
218           }
219         }
220       }
221     }
222 
223     // Check that we have found at least one libdevice that we can link in if
224     // -nocudalib hasn't been specified.
225     if (LibDeviceMap.empty() && !NoCudaLib)
226       continue;
227 
228     IsValid = true;
229     break;
230   }
231 }
232 
233 void CudaInstallationDetector::AddCudaIncludeArgs(
234     const ArgList &DriverArgs, ArgStringList &CC1Args) const {
235   if (!DriverArgs.hasArg(options::OPT_nobuiltininc)) {
236     // Add cuda_wrappers/* to our system include path.  This lets us wrap
237     // standard library headers.
238     SmallString<128> P(D.ResourceDir);
239     llvm::sys::path::append(P, "include");
240     llvm::sys::path::append(P, "cuda_wrappers");
241     CC1Args.push_back("-internal-isystem");
242     CC1Args.push_back(DriverArgs.MakeArgString(P));
243   }
244 
245   if (DriverArgs.hasArg(options::OPT_nogpuinc))
246     return;
247 
248   if (!isValid()) {
249     D.Diag(diag::err_drv_no_cuda_installation);
250     return;
251   }
252 
253   CC1Args.push_back("-internal-isystem");
254   CC1Args.push_back(DriverArgs.MakeArgString(getIncludePath()));
255   CC1Args.push_back("-include");
256   CC1Args.push_back("__clang_cuda_runtime_wrapper.h");
257 }
258 
259 void CudaInstallationDetector::CheckCudaVersionSupportsArch(
260     CudaArch Arch) const {
261   if (Arch == CudaArch::UNKNOWN || Version == CudaVersion::UNKNOWN ||
262       ArchsWithBadVersion.count(Arch) > 0)
263     return;
264 
265   auto MinVersion = MinVersionForCudaArch(Arch);
266   auto MaxVersion = MaxVersionForCudaArch(Arch);
267   if (Version < MinVersion || Version > MaxVersion) {
268     ArchsWithBadVersion.insert(Arch);
269     D.Diag(diag::err_drv_cuda_version_unsupported)
270         << CudaArchToString(Arch) << CudaVersionToString(MinVersion)
271         << CudaVersionToString(MaxVersion) << InstallPath
272         << CudaVersionToString(Version);
273   }
274 }
275 
276 void CudaInstallationDetector::print(raw_ostream &OS) const {
277   if (isValid())
278     OS << "Found CUDA installation: " << InstallPath << ", version "
279        << CudaVersionToString(Version) << "\n";
280 }
281 
282 namespace {
283 /// Debug info level for the NVPTX devices. We may need to emit different debug
284 /// info level for the host and for the device itselfi. This type controls
285 /// emission of the debug info for the devices. It either prohibits disable info
286 /// emission completely, or emits debug directives only, or emits same debug
287 /// info as for the host.
288 enum DeviceDebugInfoLevel {
289   DisableDebugInfo,        /// Do not emit debug info for the devices.
290   DebugDirectivesOnly,     /// Emit only debug directives.
291   EmitSameDebugInfoAsHost, /// Use the same debug info level just like for the
292                            /// host.
293 };
294 } // anonymous namespace
295 
296 /// Define debug info level for the NVPTX devices. If the debug info for both
297 /// the host and device are disabled (-g0/-ggdb0 or no debug options at all). If
298 /// only debug directives are requested for the both host and device
299 /// (-gline-directvies-only), or the debug info only for the device is disabled
300 /// (optimization is on and --cuda-noopt-device-debug was not specified), the
301 /// debug directves only must be emitted for the device. Otherwise, use the same
302 /// debug info level just like for the host (with the limitations of only
303 /// supported DWARF2 standard).
304 static DeviceDebugInfoLevel mustEmitDebugInfo(const ArgList &Args) {
305   const Arg *A = Args.getLastArg(options::OPT_O_Group);
306   bool IsDebugEnabled = !A || A->getOption().matches(options::OPT_O0) ||
307                         Args.hasFlag(options::OPT_cuda_noopt_device_debug,
308                                      options::OPT_no_cuda_noopt_device_debug,
309                                      /*Default=*/false);
310   if (const Arg *A = Args.getLastArg(options::OPT_g_Group)) {
311     const Option &Opt = A->getOption();
312     if (Opt.matches(options::OPT_gN_Group)) {
313       if (Opt.matches(options::OPT_g0) || Opt.matches(options::OPT_ggdb0))
314         return DisableDebugInfo;
315       if (Opt.matches(options::OPT_gline_directives_only))
316         return DebugDirectivesOnly;
317     }
318     return IsDebugEnabled ? EmitSameDebugInfoAsHost : DebugDirectivesOnly;
319   }
320   return DisableDebugInfo;
321 }
322 
323 void NVPTX::Assembler::ConstructJob(Compilation &C, const JobAction &JA,
324                                     const InputInfo &Output,
325                                     const InputInfoList &Inputs,
326                                     const ArgList &Args,
327                                     const char *LinkingOutput) const {
328   const auto &TC =
329       static_cast<const toolchains::CudaToolChain &>(getToolChain());
330   assert(TC.getTriple().isNVPTX() && "Wrong platform");
331 
332   StringRef GPUArchName;
333   // If this is an OpenMP action we need to extract the device architecture
334   // from the -march=arch option. This option may come from -Xopenmp-target
335   // flag or the default value.
336   if (JA.isDeviceOffloading(Action::OFK_OpenMP)) {
337     GPUArchName = Args.getLastArgValue(options::OPT_march_EQ);
338     assert(!GPUArchName.empty() && "Must have an architecture passed in.");
339   } else
340     GPUArchName = JA.getOffloadingArch();
341 
342   // Obtain architecture from the action.
343   CudaArch gpu_arch = StringToCudaArch(GPUArchName);
344   assert(gpu_arch != CudaArch::UNKNOWN &&
345          "Device action expected to have an architecture.");
346 
347   // Check that our installation's ptxas supports gpu_arch.
348   if (!Args.hasArg(options::OPT_no_cuda_version_check)) {
349     TC.CudaInstallation.CheckCudaVersionSupportsArch(gpu_arch);
350   }
351 
352   ArgStringList CmdArgs;
353   CmdArgs.push_back(TC.getTriple().isArch64Bit() ? "-m64" : "-m32");
354   DeviceDebugInfoLevel DIKind = mustEmitDebugInfo(Args);
355   if (DIKind == EmitSameDebugInfoAsHost) {
356     // ptxas does not accept -g option if optimization is enabled, so
357     // we ignore the compiler's -O* options if we want debug info.
358     CmdArgs.push_back("-g");
359     CmdArgs.push_back("--dont-merge-basicblocks");
360     CmdArgs.push_back("--return-at-end");
361   } else if (Arg *A = Args.getLastArg(options::OPT_O_Group)) {
362     // Map the -O we received to -O{0,1,2,3}.
363     //
364     // TODO: Perhaps we should map host -O2 to ptxas -O3. -O3 is ptxas's
365     // default, so it may correspond more closely to the spirit of clang -O2.
366 
367     // -O3 seems like the least-bad option when -Osomething is specified to
368     // clang but it isn't handled below.
369     StringRef OOpt = "3";
370     if (A->getOption().matches(options::OPT_O4) ||
371         A->getOption().matches(options::OPT_Ofast))
372       OOpt = "3";
373     else if (A->getOption().matches(options::OPT_O0))
374       OOpt = "0";
375     else if (A->getOption().matches(options::OPT_O)) {
376       // -Os, -Oz, and -O(anything else) map to -O2, for lack of better options.
377       OOpt = llvm::StringSwitch<const char *>(A->getValue())
378                  .Case("1", "1")
379                  .Case("2", "2")
380                  .Case("3", "3")
381                  .Case("s", "2")
382                  .Case("z", "2")
383                  .Default("2");
384     }
385     CmdArgs.push_back(Args.MakeArgString(llvm::Twine("-O") + OOpt));
386   } else {
387     // If no -O was passed, pass -O0 to ptxas -- no opt flag should correspond
388     // to no optimizations, but ptxas's default is -O3.
389     CmdArgs.push_back("-O0");
390   }
391   if (DIKind == DebugDirectivesOnly)
392     CmdArgs.push_back("-lineinfo");
393 
394   // Pass -v to ptxas if it was passed to the driver.
395   if (Args.hasArg(options::OPT_v))
396     CmdArgs.push_back("-v");
397 
398   CmdArgs.push_back("--gpu-name");
399   CmdArgs.push_back(Args.MakeArgString(CudaArchToString(gpu_arch)));
400   CmdArgs.push_back("--output-file");
401   CmdArgs.push_back(Args.MakeArgString(TC.getInputFilename(Output)));
402   for (const auto& II : Inputs)
403     CmdArgs.push_back(Args.MakeArgString(II.getFilename()));
404 
405   for (const auto& A : Args.getAllArgValues(options::OPT_Xcuda_ptxas))
406     CmdArgs.push_back(Args.MakeArgString(A));
407 
408   bool Relocatable = false;
409   if (JA.isOffloading(Action::OFK_OpenMP))
410     // In OpenMP we need to generate relocatable code.
411     Relocatable = Args.hasFlag(options::OPT_fopenmp_relocatable_target,
412                                options::OPT_fnoopenmp_relocatable_target,
413                                /*Default=*/true);
414   else if (JA.isOffloading(Action::OFK_Cuda))
415     Relocatable = Args.hasFlag(options::OPT_fgpu_rdc,
416                                options::OPT_fno_gpu_rdc, /*Default=*/false);
417 
418   if (Relocatable)
419     CmdArgs.push_back("-c");
420 
421   const char *Exec;
422   if (Arg *A = Args.getLastArg(options::OPT_ptxas_path_EQ))
423     Exec = A->getValue();
424   else
425     Exec = Args.MakeArgString(TC.GetProgramPath("ptxas"));
426   C.addCommand(std::make_unique<Command>(
427       JA, *this,
428       ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
429                           "--options-file"},
430       Exec, CmdArgs, Inputs));
431 }
432 
433 static bool shouldIncludePTX(const ArgList &Args, const char *gpu_arch) {
434   bool includePTX = true;
435   for (Arg *A : Args) {
436     if (!(A->getOption().matches(options::OPT_cuda_include_ptx_EQ) ||
437           A->getOption().matches(options::OPT_no_cuda_include_ptx_EQ)))
438       continue;
439     A->claim();
440     const StringRef ArchStr = A->getValue();
441     if (ArchStr == "all" || ArchStr == gpu_arch) {
442       includePTX = A->getOption().matches(options::OPT_cuda_include_ptx_EQ);
443       continue;
444     }
445   }
446   return includePTX;
447 }
448 
449 // All inputs to this linker must be from CudaDeviceActions, as we need to look
450 // at the Inputs' Actions in order to figure out which GPU architecture they
451 // correspond to.
452 void NVPTX::Linker::ConstructJob(Compilation &C, const JobAction &JA,
453                                  const InputInfo &Output,
454                                  const InputInfoList &Inputs,
455                                  const ArgList &Args,
456                                  const char *LinkingOutput) const {
457   const auto &TC =
458       static_cast<const toolchains::CudaToolChain &>(getToolChain());
459   assert(TC.getTriple().isNVPTX() && "Wrong platform");
460 
461   ArgStringList CmdArgs;
462   if (TC.CudaInstallation.version() <= CudaVersion::CUDA_100)
463     CmdArgs.push_back("--cuda");
464   CmdArgs.push_back(TC.getTriple().isArch64Bit() ? "-64" : "-32");
465   CmdArgs.push_back(Args.MakeArgString("--create"));
466   CmdArgs.push_back(Args.MakeArgString(Output.getFilename()));
467   if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost)
468     CmdArgs.push_back("-g");
469 
470   for (const auto& II : Inputs) {
471     auto *A = II.getAction();
472     assert(A->getInputs().size() == 1 &&
473            "Device offload action is expected to have a single input");
474     const char *gpu_arch_str = A->getOffloadingArch();
475     assert(gpu_arch_str &&
476            "Device action expected to have associated a GPU architecture!");
477     CudaArch gpu_arch = StringToCudaArch(gpu_arch_str);
478 
479     if (II.getType() == types::TY_PP_Asm &&
480         !shouldIncludePTX(Args, gpu_arch_str))
481       continue;
482     // We need to pass an Arch of the form "sm_XX" for cubin files and
483     // "compute_XX" for ptx.
484     const char *Arch = (II.getType() == types::TY_PP_Asm)
485                            ? CudaArchToVirtualArchString(gpu_arch)
486                            : gpu_arch_str;
487     CmdArgs.push_back(Args.MakeArgString(llvm::Twine("--image=profile=") +
488                                          Arch + ",file=" + II.getFilename()));
489   }
490 
491   for (const auto& A : Args.getAllArgValues(options::OPT_Xcuda_fatbinary))
492     CmdArgs.push_back(Args.MakeArgString(A));
493 
494   const char *Exec = Args.MakeArgString(TC.GetProgramPath("fatbinary"));
495   C.addCommand(std::make_unique<Command>(
496       JA, *this,
497       ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
498                           "--options-file"},
499       Exec, CmdArgs, Inputs));
500 }
501 
502 void NVPTX::OpenMPLinker::ConstructJob(Compilation &C, const JobAction &JA,
503                                        const InputInfo &Output,
504                                        const InputInfoList &Inputs,
505                                        const ArgList &Args,
506                                        const char *LinkingOutput) const {
507   const auto &TC =
508       static_cast<const toolchains::CudaToolChain &>(getToolChain());
509   assert(TC.getTriple().isNVPTX() && "Wrong platform");
510 
511   ArgStringList CmdArgs;
512 
513   // OpenMP uses nvlink to link cubin files. The result will be embedded in the
514   // host binary by the host linker.
515   assert(!JA.isHostOffloading(Action::OFK_OpenMP) &&
516          "CUDA toolchain not expected for an OpenMP host device.");
517 
518   if (Output.isFilename()) {
519     CmdArgs.push_back("-o");
520     CmdArgs.push_back(Output.getFilename());
521   } else
522     assert(Output.isNothing() && "Invalid output.");
523   if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost)
524     CmdArgs.push_back("-g");
525 
526   if (Args.hasArg(options::OPT_v))
527     CmdArgs.push_back("-v");
528 
529   StringRef GPUArch =
530       Args.getLastArgValue(options::OPT_march_EQ);
531   assert(!GPUArch.empty() && "At least one GPU Arch required for ptxas.");
532 
533   CmdArgs.push_back("-arch");
534   CmdArgs.push_back(Args.MakeArgString(GPUArch));
535 
536   // Assume that the directory specified with --libomptarget_nvptx_path
537   // contains the static library libomptarget-nvptx.a.
538   if (const Arg *A = Args.getLastArg(options::OPT_libomptarget_nvptx_path_EQ))
539     CmdArgs.push_back(Args.MakeArgString(Twine("-L") + A->getValue()));
540 
541   // Add paths specified in LIBRARY_PATH environment variable as -L options.
542   addDirectoryList(Args, CmdArgs, "-L", "LIBRARY_PATH");
543 
544   // Add paths for the default clang library path.
545   SmallString<256> DefaultLibPath =
546       llvm::sys::path::parent_path(TC.getDriver().Dir);
547   llvm::sys::path::append(DefaultLibPath, "lib" CLANG_LIBDIR_SUFFIX);
548   CmdArgs.push_back(Args.MakeArgString(Twine("-L") + DefaultLibPath));
549 
550   // Add linking against library implementing OpenMP calls on NVPTX target.
551   CmdArgs.push_back("-lomptarget-nvptx");
552 
553   for (const auto &II : Inputs) {
554     if (II.getType() == types::TY_LLVM_IR ||
555         II.getType() == types::TY_LTO_IR ||
556         II.getType() == types::TY_LTO_BC ||
557         II.getType() == types::TY_LLVM_BC) {
558       C.getDriver().Diag(diag::err_drv_no_linker_llvm_support)
559           << getToolChain().getTripleString();
560       continue;
561     }
562 
563     // Currently, we only pass the input files to the linker, we do not pass
564     // any libraries that may be valid only for the host.
565     if (!II.isFilename())
566       continue;
567 
568     const char *CubinF = C.addTempFile(
569         C.getArgs().MakeArgString(getToolChain().getInputFilename(II)));
570 
571     CmdArgs.push_back(CubinF);
572   }
573 
574   const char *Exec =
575       Args.MakeArgString(getToolChain().GetProgramPath("nvlink"));
576   C.addCommand(std::make_unique<Command>(
577       JA, *this,
578       ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
579                           "--options-file"},
580       Exec, CmdArgs, Inputs));
581 }
582 
583 /// CUDA toolchain.  Our assembler is ptxas, and our "linker" is fatbinary,
584 /// which isn't properly a linker but nonetheless performs the step of stitching
585 /// together object files from the assembler into a single blob.
586 
587 CudaToolChain::CudaToolChain(const Driver &D, const llvm::Triple &Triple,
588                              const ToolChain &HostTC, const ArgList &Args,
589                              const Action::OffloadKind OK)
590     : ToolChain(D, Triple, Args), HostTC(HostTC),
591       CudaInstallation(D, HostTC.getTriple(), Args), OK(OK) {
592   if (CudaInstallation.isValid()) {
593     CudaInstallation.WarnIfUnsupportedVersion();
594     getProgramPaths().push_back(std::string(CudaInstallation.getBinPath()));
595   }
596   // Lookup binaries into the driver directory, this is used to
597   // discover the clang-offload-bundler executable.
598   getProgramPaths().push_back(getDriver().Dir);
599 }
600 
601 std::string CudaToolChain::getInputFilename(const InputInfo &Input) const {
602   // Only object files are changed, for example assembly files keep their .s
603   // extensions. CUDA also continues to use .o as they don't use nvlink but
604   // fatbinary.
605   if (!(OK == Action::OFK_OpenMP && Input.getType() == types::TY_Object))
606     return ToolChain::getInputFilename(Input);
607 
608   // Replace extension for object files with cubin because nvlink relies on
609   // these particular file names.
610   SmallString<256> Filename(ToolChain::getInputFilename(Input));
611   llvm::sys::path::replace_extension(Filename, "cubin");
612   return std::string(Filename.str());
613 }
614 
615 void CudaToolChain::addClangTargetOptions(
616     const llvm::opt::ArgList &DriverArgs,
617     llvm::opt::ArgStringList &CC1Args,
618     Action::OffloadKind DeviceOffloadingKind) const {
619   HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind);
620 
621   StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
622   assert(!GpuArch.empty() && "Must have an explicit GPU arch.");
623   assert((DeviceOffloadingKind == Action::OFK_OpenMP ||
624           DeviceOffloadingKind == Action::OFK_Cuda) &&
625          "Only OpenMP or CUDA offloading kinds are supported for NVIDIA GPUs.");
626 
627   if (DeviceOffloadingKind == Action::OFK_Cuda) {
628     CC1Args.push_back("-fcuda-is-device");
629 
630     if (DriverArgs.hasFlag(options::OPT_fcuda_approx_transcendentals,
631                            options::OPT_fno_cuda_approx_transcendentals, false))
632       CC1Args.push_back("-fcuda-approx-transcendentals");
633 
634     if (DriverArgs.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc,
635                            false))
636       CC1Args.push_back("-fgpu-rdc");
637   }
638 
639   if (DriverArgs.hasArg(options::OPT_nogpulib))
640     return;
641 
642   std::string LibDeviceFile = CudaInstallation.getLibDeviceFile(GpuArch);
643 
644   if (LibDeviceFile.empty()) {
645     if (DeviceOffloadingKind == Action::OFK_OpenMP &&
646         DriverArgs.hasArg(options::OPT_S))
647       return;
648 
649     getDriver().Diag(diag::err_drv_no_cuda_libdevice) << GpuArch;
650     return;
651   }
652 
653   CC1Args.push_back("-mlink-builtin-bitcode");
654   CC1Args.push_back(DriverArgs.MakeArgString(LibDeviceFile));
655 
656   // New CUDA versions often introduce new instructions that are only supported
657   // by new PTX version, so we need to raise PTX level to enable them in NVPTX
658   // back-end.
659   const char *PtxFeature = nullptr;
660   switch (CudaInstallation.version()) {
661   case CudaVersion::CUDA_110:
662     PtxFeature = "+ptx70";
663     break;
664   case CudaVersion::CUDA_102:
665     PtxFeature = "+ptx65";
666     break;
667   case CudaVersion::CUDA_101:
668     PtxFeature = "+ptx64";
669     break;
670   case CudaVersion::CUDA_100:
671     PtxFeature = "+ptx63";
672     break;
673   case CudaVersion::CUDA_92:
674     PtxFeature = "+ptx61";
675     break;
676   case CudaVersion::CUDA_91:
677     PtxFeature = "+ptx61";
678     break;
679   case CudaVersion::CUDA_90:
680     PtxFeature = "+ptx60";
681     break;
682   default:
683     PtxFeature = "+ptx42";
684   }
685   CC1Args.append({"-target-feature", PtxFeature});
686   if (DriverArgs.hasFlag(options::OPT_fcuda_short_ptr,
687                          options::OPT_fno_cuda_short_ptr, false))
688     CC1Args.append({"-mllvm", "--nvptx-short-ptr"});
689 
690   if (CudaInstallation.version() >= CudaVersion::UNKNOWN)
691     CC1Args.push_back(DriverArgs.MakeArgString(
692         Twine("-target-sdk-version=") +
693         CudaVersionToString(CudaInstallation.version())));
694 
695   if (DeviceOffloadingKind == Action::OFK_OpenMP) {
696     SmallVector<StringRef, 8> LibraryPaths;
697     if (const Arg *A = DriverArgs.getLastArg(options::OPT_libomptarget_nvptx_path_EQ))
698       LibraryPaths.push_back(A->getValue());
699 
700     // Add user defined library paths from LIBRARY_PATH.
701     llvm::Optional<std::string> LibPath =
702         llvm::sys::Process::GetEnv("LIBRARY_PATH");
703     if (LibPath) {
704       SmallVector<StringRef, 8> Frags;
705       const char EnvPathSeparatorStr[] = {llvm::sys::EnvPathSeparator, '\0'};
706       llvm::SplitString(*LibPath, Frags, EnvPathSeparatorStr);
707       for (StringRef Path : Frags)
708         LibraryPaths.emplace_back(Path.trim());
709     }
710 
711     // Add path to lib / lib64 folder.
712     SmallString<256> DefaultLibPath =
713         llvm::sys::path::parent_path(getDriver().Dir);
714     llvm::sys::path::append(DefaultLibPath, Twine("lib") + CLANG_LIBDIR_SUFFIX);
715     LibraryPaths.emplace_back(DefaultLibPath.c_str());
716 
717     std::string LibOmpTargetName =
718       "libomptarget-nvptx-" + GpuArch.str() + ".bc";
719     bool FoundBCLibrary = false;
720     for (StringRef LibraryPath : LibraryPaths) {
721       SmallString<128> LibOmpTargetFile(LibraryPath);
722       llvm::sys::path::append(LibOmpTargetFile, LibOmpTargetName);
723       if (llvm::sys::fs::exists(LibOmpTargetFile)) {
724         CC1Args.push_back("-mlink-builtin-bitcode");
725         CC1Args.push_back(DriverArgs.MakeArgString(LibOmpTargetFile));
726         FoundBCLibrary = true;
727         break;
728       }
729     }
730     if (!FoundBCLibrary)
731       getDriver().Diag(diag::warn_drv_omp_offload_target_missingbcruntime)
732           << LibOmpTargetName;
733   }
734 }
735 
736 llvm::DenormalMode CudaToolChain::getDefaultDenormalModeForType(
737     const llvm::opt::ArgList &DriverArgs, const JobAction &JA,
738     const llvm::fltSemantics *FPType) const {
739   if (JA.getOffloadingDeviceKind() == Action::OFK_Cuda) {
740     if (FPType && FPType == &llvm::APFloat::IEEEsingle() &&
741         DriverArgs.hasFlag(options::OPT_fcuda_flush_denormals_to_zero,
742                            options::OPT_fno_cuda_flush_denormals_to_zero,
743                            false))
744       return llvm::DenormalMode::getPreserveSign();
745   }
746 
747   assert(JA.getOffloadingDeviceKind() != Action::OFK_Host);
748   return llvm::DenormalMode::getIEEE();
749 }
750 
751 bool CudaToolChain::supportsDebugInfoOption(const llvm::opt::Arg *A) const {
752   const Option &O = A->getOption();
753   return (O.matches(options::OPT_gN_Group) &&
754           !O.matches(options::OPT_gmodules)) ||
755          O.matches(options::OPT_g_Flag) ||
756          O.matches(options::OPT_ggdbN_Group) || O.matches(options::OPT_ggdb) ||
757          O.matches(options::OPT_gdwarf) || O.matches(options::OPT_gdwarf_2) ||
758          O.matches(options::OPT_gdwarf_3) || O.matches(options::OPT_gdwarf_4) ||
759          O.matches(options::OPT_gdwarf_5) ||
760          O.matches(options::OPT_gcolumn_info);
761 }
762 
763 void CudaToolChain::adjustDebugInfoKind(
764     codegenoptions::DebugInfoKind &DebugInfoKind, const ArgList &Args) const {
765   switch (mustEmitDebugInfo(Args)) {
766   case DisableDebugInfo:
767     DebugInfoKind = codegenoptions::NoDebugInfo;
768     break;
769   case DebugDirectivesOnly:
770     DebugInfoKind = codegenoptions::DebugDirectivesOnly;
771     break;
772   case EmitSameDebugInfoAsHost:
773     // Use same debug info level as the host.
774     break;
775   }
776 }
777 
778 void CudaToolChain::AddCudaIncludeArgs(const ArgList &DriverArgs,
779                                        ArgStringList &CC1Args) const {
780   // Check our CUDA version if we're going to include the CUDA headers.
781   if (!DriverArgs.hasArg(options::OPT_nogpuinc) &&
782       !DriverArgs.hasArg(options::OPT_no_cuda_version_check)) {
783     StringRef Arch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
784     assert(!Arch.empty() && "Must have an explicit GPU arch.");
785     CudaInstallation.CheckCudaVersionSupportsArch(StringToCudaArch(Arch));
786   }
787   CudaInstallation.AddCudaIncludeArgs(DriverArgs, CC1Args);
788 }
789 
790 llvm::opt::DerivedArgList *
791 CudaToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
792                              StringRef BoundArch,
793                              Action::OffloadKind DeviceOffloadKind) const {
794   DerivedArgList *DAL =
795       HostTC.TranslateArgs(Args, BoundArch, DeviceOffloadKind);
796   if (!DAL)
797     DAL = new DerivedArgList(Args.getBaseArgs());
798 
799   const OptTable &Opts = getDriver().getOpts();
800 
801   // For OpenMP device offloading, append derived arguments. Make sure
802   // flags are not duplicated.
803   // Also append the compute capability.
804   if (DeviceOffloadKind == Action::OFK_OpenMP) {
805     for (Arg *A : Args) {
806       bool IsDuplicate = false;
807       for (Arg *DALArg : *DAL) {
808         if (A == DALArg) {
809           IsDuplicate = true;
810           break;
811         }
812       }
813       if (!IsDuplicate)
814         DAL->append(A);
815     }
816 
817     StringRef Arch = DAL->getLastArgValue(options::OPT_march_EQ);
818     if (Arch.empty())
819       DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ),
820                         CLANG_OPENMP_NVPTX_DEFAULT_ARCH);
821 
822     return DAL;
823   }
824 
825   for (Arg *A : Args) {
826     DAL->append(A);
827   }
828 
829   if (!BoundArch.empty()) {
830     DAL->eraseArg(options::OPT_march_EQ);
831     DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ), BoundArch);
832   }
833   return DAL;
834 }
835 
836 Tool *CudaToolChain::buildAssembler() const {
837   return new tools::NVPTX::Assembler(*this);
838 }
839 
840 Tool *CudaToolChain::buildLinker() const {
841   if (OK == Action::OFK_OpenMP)
842     return new tools::NVPTX::OpenMPLinker(*this);
843   return new tools::NVPTX::Linker(*this);
844 }
845 
846 void CudaToolChain::addClangWarningOptions(ArgStringList &CC1Args) const {
847   HostTC.addClangWarningOptions(CC1Args);
848 }
849 
850 ToolChain::CXXStdlibType
851 CudaToolChain::GetCXXStdlibType(const ArgList &Args) const {
852   return HostTC.GetCXXStdlibType(Args);
853 }
854 
855 void CudaToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
856                                               ArgStringList &CC1Args) const {
857   HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args);
858 }
859 
860 void CudaToolChain::AddClangCXXStdlibIncludeArgs(const ArgList &Args,
861                                                  ArgStringList &CC1Args) const {
862   HostTC.AddClangCXXStdlibIncludeArgs(Args, CC1Args);
863 }
864 
865 void CudaToolChain::AddIAMCUIncludeArgs(const ArgList &Args,
866                                         ArgStringList &CC1Args) const {
867   HostTC.AddIAMCUIncludeArgs(Args, CC1Args);
868 }
869 
870 SanitizerMask CudaToolChain::getSupportedSanitizers() const {
871   // The CudaToolChain only supports sanitizers in the sense that it allows
872   // sanitizer arguments on the command line if they are supported by the host
873   // toolchain. The CudaToolChain will actually ignore any command line
874   // arguments for any of these "supported" sanitizers. That means that no
875   // sanitization of device code is actually supported at this time.
876   //
877   // This behavior is necessary because the host and device toolchains
878   // invocations often share the command line, so the device toolchain must
879   // tolerate flags meant only for the host toolchain.
880   return HostTC.getSupportedSanitizers();
881 }
882 
883 VersionTuple CudaToolChain::computeMSVCVersion(const Driver *D,
884                                                const ArgList &Args) const {
885   return HostTC.computeMSVCVersion(D, Args);
886 }
887