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 "clang/Basic/Cuda.h"
12 #include "clang/Config/config.h"
13 #include "clang/Driver/Compilation.h"
14 #include "clang/Driver/Distro.h"
15 #include "clang/Driver/Driver.h"
16 #include "clang/Driver/DriverDiagnostic.h"
17 #include "clang/Driver/InputInfo.h"
18 #include "clang/Driver/Options.h"
19 #include "llvm/ADT/Optional.h"
20 #include "llvm/ADT/StringExtras.h"
21 #include "llvm/Option/ArgList.h"
22 #include "llvm/Support/FileSystem.h"
23 #include "llvm/Support/Host.h"
24 #include "llvm/Support/Path.h"
25 #include "llvm/Support/Process.h"
26 #include "llvm/Support/Program.h"
27 #include "llvm/Support/TargetParser.h"
28 #include "llvm/Support/VirtualFileSystem.h"
29 #include <system_error>
30 
31 using namespace clang::driver;
32 using namespace clang::driver::toolchains;
33 using namespace clang::driver::tools;
34 using namespace clang;
35 using namespace llvm::opt;
36 
37 namespace {
38 
getCudaVersion(uint32_t raw_version)39 CudaVersion getCudaVersion(uint32_t raw_version) {
40   if (raw_version < 7050)
41     return CudaVersion::CUDA_70;
42   if (raw_version < 8000)
43     return CudaVersion::CUDA_75;
44   if (raw_version < 9000)
45     return CudaVersion::CUDA_80;
46   if (raw_version < 9010)
47     return CudaVersion::CUDA_90;
48   if (raw_version < 9020)
49     return CudaVersion::CUDA_91;
50   if (raw_version < 10000)
51     return CudaVersion::CUDA_92;
52   if (raw_version < 10010)
53     return CudaVersion::CUDA_100;
54   if (raw_version < 10020)
55     return CudaVersion::CUDA_101;
56   if (raw_version < 11000)
57     return CudaVersion::CUDA_102;
58   if (raw_version < 11010)
59     return CudaVersion::CUDA_110;
60   if (raw_version < 11020)
61     return CudaVersion::CUDA_111;
62   if (raw_version < 11030)
63     return CudaVersion::CUDA_112;
64   if (raw_version < 11040)
65     return CudaVersion::CUDA_113;
66   if (raw_version < 11050)
67     return CudaVersion::CUDA_114;
68   return CudaVersion::NEW;
69 }
70 
parseCudaHFile(llvm::StringRef Input)71 CudaVersion parseCudaHFile(llvm::StringRef Input) {
72   // Helper lambda which skips the words if the line starts with them or returns
73   // None otherwise.
74   auto StartsWithWords =
75       [](llvm::StringRef Line,
76          const SmallVector<StringRef, 3> words) -> llvm::Optional<StringRef> {
77     for (StringRef word : words) {
78       if (!Line.consume_front(word))
79         return {};
80       Line = Line.ltrim();
81     }
82     return Line;
83   };
84 
85   Input = Input.ltrim();
86   while (!Input.empty()) {
87     if (auto Line =
88             StartsWithWords(Input.ltrim(), {"#", "define", "CUDA_VERSION"})) {
89       uint32_t RawVersion;
90       Line->consumeInteger(10, RawVersion);
91       return getCudaVersion(RawVersion);
92     }
93     // Find next non-empty line.
94     Input = Input.drop_front(Input.find_first_of("\n\r")).ltrim();
95   }
96   return CudaVersion::UNKNOWN;
97 }
98 } // namespace
99 
WarnIfUnsupportedVersion()100 void CudaInstallationDetector::WarnIfUnsupportedVersion() {
101   if (Version > CudaVersion::PARTIALLY_SUPPORTED) {
102     std::string VersionString = CudaVersionToString(Version);
103     if (!VersionString.empty())
104       VersionString.insert(0, " ");
105     D.Diag(diag::warn_drv_new_cuda_version)
106         << VersionString
107         << (CudaVersion::PARTIALLY_SUPPORTED != CudaVersion::FULLY_SUPPORTED)
108         << CudaVersionToString(CudaVersion::PARTIALLY_SUPPORTED);
109   } else if (Version > CudaVersion::FULLY_SUPPORTED)
110     D.Diag(diag::warn_drv_partially_supported_cuda_version)
111         << CudaVersionToString(Version);
112 }
113 
CudaInstallationDetector(const Driver & D,const llvm::Triple & HostTriple,const llvm::opt::ArgList & Args)114 CudaInstallationDetector::CudaInstallationDetector(
115     const Driver &D, const llvm::Triple &HostTriple,
116     const llvm::opt::ArgList &Args)
117     : D(D) {
118   struct Candidate {
119     std::string Path;
120     bool StrictChecking;
121 
122     Candidate(std::string Path, bool StrictChecking = false)
123         : Path(Path), StrictChecking(StrictChecking) {}
124   };
125   SmallVector<Candidate, 4> Candidates;
126 
127   // In decreasing order so we prefer newer versions to older versions.
128   std::initializer_list<const char *> Versions = {"8.0", "7.5", "7.0"};
129   auto &FS = D.getVFS();
130 
131   if (Args.hasArg(clang::driver::options::OPT_cuda_path_EQ)) {
132     Candidates.emplace_back(
133         Args.getLastArgValue(clang::driver::options::OPT_cuda_path_EQ).str());
134   } else if (HostTriple.isOSWindows()) {
135     for (const char *Ver : Versions)
136       Candidates.emplace_back(
137           D.SysRoot + "/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v" +
138           Ver);
139   } else {
140     if (!Args.hasArg(clang::driver::options::OPT_cuda_path_ignore_env)) {
141       // Try to find ptxas binary. If the executable is located in a directory
142       // called 'bin/', its parent directory might be a good guess for a valid
143       // CUDA installation.
144       // However, some distributions might installs 'ptxas' to /usr/bin. In that
145       // case the candidate would be '/usr' which passes the following checks
146       // because '/usr/include' exists as well. To avoid this case, we always
147       // check for the directory potentially containing files for libdevice,
148       // even if the user passes -nocudalib.
149       if (llvm::ErrorOr<std::string> ptxas =
150               llvm::sys::findProgramByName("ptxas")) {
151         SmallString<256> ptxasAbsolutePath;
152         llvm::sys::fs::real_path(*ptxas, ptxasAbsolutePath);
153 
154         StringRef ptxasDir = llvm::sys::path::parent_path(ptxasAbsolutePath);
155         if (llvm::sys::path::filename(ptxasDir) == "bin")
156           Candidates.emplace_back(
157               std::string(llvm::sys::path::parent_path(ptxasDir)),
158               /*StrictChecking=*/true);
159       }
160     }
161 
162     Candidates.emplace_back(D.SysRoot + "/usr/local/cuda");
163     for (const char *Ver : Versions)
164       Candidates.emplace_back(D.SysRoot + "/usr/local/cuda-" + Ver);
165 
166     Distro Dist(FS, llvm::Triple(llvm::sys::getProcessTriple()));
167     if (Dist.IsDebian() || Dist.IsUbuntu())
168       // Special case for Debian to have nvidia-cuda-toolkit work
169       // out of the box. More info on http://bugs.debian.org/882505
170       Candidates.emplace_back(D.SysRoot + "/usr/lib/cuda");
171   }
172 
173   bool NoCudaLib = Args.hasArg(options::OPT_nogpulib);
174 
175   for (const auto &Candidate : Candidates) {
176     InstallPath = Candidate.Path;
177     if (InstallPath.empty() || !FS.exists(InstallPath))
178       continue;
179 
180     BinPath = InstallPath + "/bin";
181     IncludePath = InstallPath + "/include";
182     LibDevicePath = InstallPath + "/nvvm/libdevice";
183 
184     if (!(FS.exists(IncludePath) && FS.exists(BinPath)))
185       continue;
186     bool CheckLibDevice = (!NoCudaLib || Candidate.StrictChecking);
187     if (CheckLibDevice && !FS.exists(LibDevicePath))
188       continue;
189 
190     // On Linux, we have both lib and lib64 directories, and we need to choose
191     // based on our triple.  On MacOS, we have only a lib directory.
192     //
193     // It's sufficient for our purposes to be flexible: If both lib and lib64
194     // exist, we choose whichever one matches our triple.  Otherwise, if only
195     // lib exists, we use it.
196     if (HostTriple.isArch64Bit() && FS.exists(InstallPath + "/lib64"))
197       LibPath = InstallPath + "/lib64";
198     else if (FS.exists(InstallPath + "/lib"))
199       LibPath = InstallPath + "/lib";
200     else
201       continue;
202 
203     Version = CudaVersion::UNKNOWN;
204     if (auto CudaHFile = FS.getBufferForFile(InstallPath + "/include/cuda.h"))
205       Version = parseCudaHFile((*CudaHFile)->getBuffer());
206     // As the last resort, make an educated guess between CUDA-7.0, which had
207     // old-style libdevice bitcode, and an unknown recent CUDA version.
208     if (Version == CudaVersion::UNKNOWN) {
209       Version = FS.exists(LibDevicePath + "/libdevice.10.bc")
210                     ? CudaVersion::NEW
211                     : CudaVersion::CUDA_70;
212     }
213 
214     if (Version >= CudaVersion::CUDA_90) {
215       // CUDA-9+ uses single libdevice file for all GPU variants.
216       std::string FilePath = LibDevicePath + "/libdevice.10.bc";
217       if (FS.exists(FilePath)) {
218         for (int Arch = (int)CudaArch::SM_30, E = (int)CudaArch::LAST; Arch < E;
219              ++Arch) {
220           CudaArch GpuArch = static_cast<CudaArch>(Arch);
221           if (!IsNVIDIAGpuArch(GpuArch))
222             continue;
223           std::string GpuArchName(CudaArchToString(GpuArch));
224           LibDeviceMap[GpuArchName] = FilePath;
225         }
226       }
227     } else {
228       std::error_code EC;
229       for (llvm::vfs::directory_iterator LI = FS.dir_begin(LibDevicePath, EC),
230                                          LE;
231            !EC && LI != LE; LI = LI.increment(EC)) {
232         StringRef FilePath = LI->path();
233         StringRef FileName = llvm::sys::path::filename(FilePath);
234         // Process all bitcode filenames that look like
235         // libdevice.compute_XX.YY.bc
236         const StringRef LibDeviceName = "libdevice.";
237         if (!(FileName.startswith(LibDeviceName) && FileName.endswith(".bc")))
238           continue;
239         StringRef GpuArch = FileName.slice(
240             LibDeviceName.size(), FileName.find('.', LibDeviceName.size()));
241         LibDeviceMap[GpuArch] = FilePath.str();
242         // Insert map entries for specific devices with this compute
243         // capability. NVCC's choice of the libdevice library version is
244         // rather peculiar and depends on the CUDA version.
245         if (GpuArch == "compute_20") {
246           LibDeviceMap["sm_20"] = std::string(FilePath);
247           LibDeviceMap["sm_21"] = std::string(FilePath);
248           LibDeviceMap["sm_32"] = std::string(FilePath);
249         } else if (GpuArch == "compute_30") {
250           LibDeviceMap["sm_30"] = std::string(FilePath);
251           if (Version < CudaVersion::CUDA_80) {
252             LibDeviceMap["sm_50"] = std::string(FilePath);
253             LibDeviceMap["sm_52"] = std::string(FilePath);
254             LibDeviceMap["sm_53"] = std::string(FilePath);
255           }
256           LibDeviceMap["sm_60"] = std::string(FilePath);
257           LibDeviceMap["sm_61"] = std::string(FilePath);
258           LibDeviceMap["sm_62"] = std::string(FilePath);
259         } else if (GpuArch == "compute_35") {
260           LibDeviceMap["sm_35"] = std::string(FilePath);
261           LibDeviceMap["sm_37"] = std::string(FilePath);
262         } else if (GpuArch == "compute_50") {
263           if (Version >= CudaVersion::CUDA_80) {
264             LibDeviceMap["sm_50"] = std::string(FilePath);
265             LibDeviceMap["sm_52"] = std::string(FilePath);
266             LibDeviceMap["sm_53"] = std::string(FilePath);
267           }
268         }
269       }
270     }
271 
272     // Check that we have found at least one libdevice that we can link in if
273     // -nocudalib hasn't been specified.
274     if (LibDeviceMap.empty() && !NoCudaLib)
275       continue;
276 
277     IsValid = true;
278     break;
279   }
280 }
281 
AddCudaIncludeArgs(const ArgList & DriverArgs,ArgStringList & CC1Args) const282 void CudaInstallationDetector::AddCudaIncludeArgs(
283     const ArgList &DriverArgs, ArgStringList &CC1Args) const {
284   if (!DriverArgs.hasArg(options::OPT_nobuiltininc)) {
285     // Add cuda_wrappers/* to our system include path.  This lets us wrap
286     // standard library headers.
287     SmallString<128> P(D.ResourceDir);
288     llvm::sys::path::append(P, "include");
289     llvm::sys::path::append(P, "cuda_wrappers");
290     CC1Args.push_back("-internal-isystem");
291     CC1Args.push_back(DriverArgs.MakeArgString(P));
292   }
293 
294   if (DriverArgs.hasArg(options::OPT_nogpuinc))
295     return;
296 
297   if (!isValid()) {
298     D.Diag(diag::err_drv_no_cuda_installation);
299     return;
300   }
301 
302   CC1Args.push_back("-include");
303   CC1Args.push_back("__clang_cuda_runtime_wrapper.h");
304 }
305 
CheckCudaVersionSupportsArch(CudaArch Arch) const306 void CudaInstallationDetector::CheckCudaVersionSupportsArch(
307     CudaArch Arch) const {
308   if (Arch == CudaArch::UNKNOWN || Version == CudaVersion::UNKNOWN ||
309       ArchsWithBadVersion[(int)Arch])
310     return;
311 
312   auto MinVersion = MinVersionForCudaArch(Arch);
313   auto MaxVersion = MaxVersionForCudaArch(Arch);
314   if (Version < MinVersion || Version > MaxVersion) {
315     ArchsWithBadVersion[(int)Arch] = true;
316     D.Diag(diag::err_drv_cuda_version_unsupported)
317         << CudaArchToString(Arch) << CudaVersionToString(MinVersion)
318         << CudaVersionToString(MaxVersion) << InstallPath
319         << CudaVersionToString(Version);
320   }
321 }
322 
print(raw_ostream & OS) const323 void CudaInstallationDetector::print(raw_ostream &OS) const {
324   if (isValid())
325     OS << "Found CUDA installation: " << InstallPath << ", version "
326        << CudaVersionToString(Version) << "\n";
327 }
328 
329 namespace {
330 /// Debug info level for the NVPTX devices. We may need to emit different debug
331 /// info level for the host and for the device itselfi. This type controls
332 /// emission of the debug info for the devices. It either prohibits disable info
333 /// emission completely, or emits debug directives only, or emits same debug
334 /// info as for the host.
335 enum DeviceDebugInfoLevel {
336   DisableDebugInfo,        /// Do not emit debug info for the devices.
337   DebugDirectivesOnly,     /// Emit only debug directives.
338   EmitSameDebugInfoAsHost, /// Use the same debug info level just like for the
339                            /// host.
340 };
341 } // anonymous namespace
342 
343 /// Define debug info level for the NVPTX devices. If the debug info for both
344 /// the host and device are disabled (-g0/-ggdb0 or no debug options at all). If
345 /// only debug directives are requested for the both host and device
346 /// (-gline-directvies-only), or the debug info only for the device is disabled
347 /// (optimization is on and --cuda-noopt-device-debug was not specified), the
348 /// debug directves only must be emitted for the device. Otherwise, use the same
349 /// debug info level just like for the host (with the limitations of only
350 /// supported DWARF2 standard).
mustEmitDebugInfo(const ArgList & Args)351 static DeviceDebugInfoLevel mustEmitDebugInfo(const ArgList &Args) {
352   const Arg *A = Args.getLastArg(options::OPT_O_Group);
353   bool IsDebugEnabled = !A || A->getOption().matches(options::OPT_O0) ||
354                         Args.hasFlag(options::OPT_cuda_noopt_device_debug,
355                                      options::OPT_no_cuda_noopt_device_debug,
356                                      /*Default=*/false);
357   if (const Arg *A = Args.getLastArg(options::OPT_g_Group)) {
358     const Option &Opt = A->getOption();
359     if (Opt.matches(options::OPT_gN_Group)) {
360       if (Opt.matches(options::OPT_g0) || Opt.matches(options::OPT_ggdb0))
361         return DisableDebugInfo;
362       if (Opt.matches(options::OPT_gline_directives_only))
363         return DebugDirectivesOnly;
364     }
365     return IsDebugEnabled ? EmitSameDebugInfoAsHost : DebugDirectivesOnly;
366   }
367   return willEmitRemarks(Args) ? DebugDirectivesOnly : DisableDebugInfo;
368 }
369 
ConstructJob(Compilation & C,const JobAction & JA,const InputInfo & Output,const InputInfoList & Inputs,const ArgList & Args,const char * LinkingOutput) const370 void NVPTX::Assembler::ConstructJob(Compilation &C, const JobAction &JA,
371                                     const InputInfo &Output,
372                                     const InputInfoList &Inputs,
373                                     const ArgList &Args,
374                                     const char *LinkingOutput) const {
375   const auto &TC =
376       static_cast<const toolchains::CudaToolChain &>(getToolChain());
377   assert(TC.getTriple().isNVPTX() && "Wrong platform");
378 
379   StringRef GPUArchName;
380   // If this is an OpenMP action we need to extract the device architecture
381   // from the -march=arch option. This option may come from -Xopenmp-target
382   // flag or the default value.
383   if (JA.isDeviceOffloading(Action::OFK_OpenMP)) {
384     GPUArchName = Args.getLastArgValue(options::OPT_march_EQ);
385     assert(!GPUArchName.empty() && "Must have an architecture passed in.");
386   } else
387     GPUArchName = JA.getOffloadingArch();
388 
389   // Obtain architecture from the action.
390   CudaArch gpu_arch = StringToCudaArch(GPUArchName);
391   assert(gpu_arch != CudaArch::UNKNOWN &&
392          "Device action expected to have an architecture.");
393 
394   // Check that our installation's ptxas supports gpu_arch.
395   if (!Args.hasArg(options::OPT_no_cuda_version_check)) {
396     TC.CudaInstallation.CheckCudaVersionSupportsArch(gpu_arch);
397   }
398 
399   ArgStringList CmdArgs;
400   CmdArgs.push_back(TC.getTriple().isArch64Bit() ? "-m64" : "-m32");
401   DeviceDebugInfoLevel DIKind = mustEmitDebugInfo(Args);
402   if (DIKind == EmitSameDebugInfoAsHost) {
403     // ptxas does not accept -g option if optimization is enabled, so
404     // we ignore the compiler's -O* options if we want debug info.
405     CmdArgs.push_back("-g");
406     CmdArgs.push_back("--dont-merge-basicblocks");
407     CmdArgs.push_back("--return-at-end");
408   } else if (Arg *A = Args.getLastArg(options::OPT_O_Group)) {
409     // Map the -O we received to -O{0,1,2,3}.
410     //
411     // TODO: Perhaps we should map host -O2 to ptxas -O3. -O3 is ptxas's
412     // default, so it may correspond more closely to the spirit of clang -O2.
413 
414     // -O3 seems like the least-bad option when -Osomething is specified to
415     // clang but it isn't handled below.
416     StringRef OOpt = "3";
417     if (A->getOption().matches(options::OPT_O4) ||
418         A->getOption().matches(options::OPT_Ofast))
419       OOpt = "3";
420     else if (A->getOption().matches(options::OPT_O0))
421       OOpt = "0";
422     else if (A->getOption().matches(options::OPT_O)) {
423       // -Os, -Oz, and -O(anything else) map to -O2, for lack of better options.
424       OOpt = llvm::StringSwitch<const char *>(A->getValue())
425                  .Case("1", "1")
426                  .Case("2", "2")
427                  .Case("3", "3")
428                  .Case("s", "2")
429                  .Case("z", "2")
430                  .Default("2");
431     }
432     CmdArgs.push_back(Args.MakeArgString(llvm::Twine("-O") + OOpt));
433   } else {
434     // If no -O was passed, pass -O0 to ptxas -- no opt flag should correspond
435     // to no optimizations, but ptxas's default is -O3.
436     CmdArgs.push_back("-O0");
437   }
438   if (DIKind == DebugDirectivesOnly)
439     CmdArgs.push_back("-lineinfo");
440 
441   // Pass -v to ptxas if it was passed to the driver.
442   if (Args.hasArg(options::OPT_v))
443     CmdArgs.push_back("-v");
444 
445   CmdArgs.push_back("--gpu-name");
446   CmdArgs.push_back(Args.MakeArgString(CudaArchToString(gpu_arch)));
447   CmdArgs.push_back("--output-file");
448   CmdArgs.push_back(Args.MakeArgString(TC.getInputFilename(Output)));
449   for (const auto& II : Inputs)
450     CmdArgs.push_back(Args.MakeArgString(II.getFilename()));
451 
452   for (const auto& A : Args.getAllArgValues(options::OPT_Xcuda_ptxas))
453     CmdArgs.push_back(Args.MakeArgString(A));
454 
455   bool Relocatable = false;
456   if (JA.isOffloading(Action::OFK_OpenMP))
457     // In OpenMP we need to generate relocatable code.
458     Relocatable = Args.hasFlag(options::OPT_fopenmp_relocatable_target,
459                                options::OPT_fnoopenmp_relocatable_target,
460                                /*Default=*/true);
461   else if (JA.isOffloading(Action::OFK_Cuda))
462     Relocatable = Args.hasFlag(options::OPT_fgpu_rdc,
463                                options::OPT_fno_gpu_rdc, /*Default=*/false);
464 
465   if (Relocatable)
466     CmdArgs.push_back("-c");
467 
468   const char *Exec;
469   if (Arg *A = Args.getLastArg(options::OPT_ptxas_path_EQ))
470     Exec = A->getValue();
471   else
472     Exec = Args.MakeArgString(TC.GetProgramPath("ptxas"));
473   C.addCommand(std::make_unique<Command>(
474       JA, *this,
475       ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
476                           "--options-file"},
477       Exec, CmdArgs, Inputs, Output));
478 }
479 
shouldIncludePTX(const ArgList & Args,const char * gpu_arch)480 static bool shouldIncludePTX(const ArgList &Args, const char *gpu_arch) {
481   bool includePTX = true;
482   for (Arg *A : Args) {
483     if (!(A->getOption().matches(options::OPT_cuda_include_ptx_EQ) ||
484           A->getOption().matches(options::OPT_no_cuda_include_ptx_EQ)))
485       continue;
486     A->claim();
487     const StringRef ArchStr = A->getValue();
488     if (ArchStr == "all" || ArchStr == gpu_arch) {
489       includePTX = A->getOption().matches(options::OPT_cuda_include_ptx_EQ);
490       continue;
491     }
492   }
493   return includePTX;
494 }
495 
496 // All inputs to this linker must be from CudaDeviceActions, as we need to look
497 // at the Inputs' Actions in order to figure out which GPU architecture they
498 // correspond to.
ConstructJob(Compilation & C,const JobAction & JA,const InputInfo & Output,const InputInfoList & Inputs,const ArgList & Args,const char * LinkingOutput) const499 void NVPTX::Linker::ConstructJob(Compilation &C, const JobAction &JA,
500                                  const InputInfo &Output,
501                                  const InputInfoList &Inputs,
502                                  const ArgList &Args,
503                                  const char *LinkingOutput) const {
504   const auto &TC =
505       static_cast<const toolchains::CudaToolChain &>(getToolChain());
506   assert(TC.getTriple().isNVPTX() && "Wrong platform");
507 
508   ArgStringList CmdArgs;
509   if (TC.CudaInstallation.version() <= CudaVersion::CUDA_100)
510     CmdArgs.push_back("--cuda");
511   CmdArgs.push_back(TC.getTriple().isArch64Bit() ? "-64" : "-32");
512   CmdArgs.push_back(Args.MakeArgString("--create"));
513   CmdArgs.push_back(Args.MakeArgString(Output.getFilename()));
514   if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost)
515     CmdArgs.push_back("-g");
516 
517   for (const auto& II : Inputs) {
518     auto *A = II.getAction();
519     assert(A->getInputs().size() == 1 &&
520            "Device offload action is expected to have a single input");
521     const char *gpu_arch_str = A->getOffloadingArch();
522     assert(gpu_arch_str &&
523            "Device action expected to have associated a GPU architecture!");
524     CudaArch gpu_arch = StringToCudaArch(gpu_arch_str);
525 
526     if (II.getType() == types::TY_PP_Asm &&
527         !shouldIncludePTX(Args, gpu_arch_str))
528       continue;
529     // We need to pass an Arch of the form "sm_XX" for cubin files and
530     // "compute_XX" for ptx.
531     const char *Arch = (II.getType() == types::TY_PP_Asm)
532                            ? CudaArchToVirtualArchString(gpu_arch)
533                            : gpu_arch_str;
534     CmdArgs.push_back(Args.MakeArgString(llvm::Twine("--image=profile=") +
535                                          Arch + ",file=" + II.getFilename()));
536   }
537 
538   for (const auto& A : Args.getAllArgValues(options::OPT_Xcuda_fatbinary))
539     CmdArgs.push_back(Args.MakeArgString(A));
540 
541   const char *Exec = Args.MakeArgString(TC.GetProgramPath("fatbinary"));
542   C.addCommand(std::make_unique<Command>(
543       JA, *this,
544       ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
545                           "--options-file"},
546       Exec, CmdArgs, Inputs, Output));
547 }
548 
ConstructJob(Compilation & C,const JobAction & JA,const InputInfo & Output,const InputInfoList & Inputs,const ArgList & Args,const char * LinkingOutput) const549 void NVPTX::OpenMPLinker::ConstructJob(Compilation &C, const JobAction &JA,
550                                        const InputInfo &Output,
551                                        const InputInfoList &Inputs,
552                                        const ArgList &Args,
553                                        const char *LinkingOutput) const {
554   const auto &TC =
555       static_cast<const toolchains::CudaToolChain &>(getToolChain());
556   assert(TC.getTriple().isNVPTX() && "Wrong platform");
557 
558   ArgStringList CmdArgs;
559 
560   // OpenMP uses nvlink to link cubin files. The result will be embedded in the
561   // host binary by the host linker.
562   assert(!JA.isHostOffloading(Action::OFK_OpenMP) &&
563          "CUDA toolchain not expected for an OpenMP host device.");
564 
565   if (Output.isFilename()) {
566     CmdArgs.push_back("-o");
567     CmdArgs.push_back(Output.getFilename());
568   } else
569     assert(Output.isNothing() && "Invalid output.");
570   if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost)
571     CmdArgs.push_back("-g");
572 
573   if (Args.hasArg(options::OPT_v))
574     CmdArgs.push_back("-v");
575 
576   StringRef GPUArch =
577       Args.getLastArgValue(options::OPT_march_EQ);
578   assert(!GPUArch.empty() && "At least one GPU Arch required for ptxas.");
579 
580   CmdArgs.push_back("-arch");
581   CmdArgs.push_back(Args.MakeArgString(GPUArch));
582 
583   // Add paths specified in LIBRARY_PATH environment variable as -L options.
584   addDirectoryList(Args, CmdArgs, "-L", "LIBRARY_PATH");
585 
586   // Add paths for the default clang library path.
587   SmallString<256> DefaultLibPath =
588       llvm::sys::path::parent_path(TC.getDriver().Dir);
589   llvm::sys::path::append(DefaultLibPath, "lib" CLANG_LIBDIR_SUFFIX);
590   CmdArgs.push_back(Args.MakeArgString(Twine("-L") + DefaultLibPath));
591 
592   for (const auto &II : Inputs) {
593     if (II.getType() == types::TY_LLVM_IR ||
594         II.getType() == types::TY_LTO_IR ||
595         II.getType() == types::TY_LTO_BC ||
596         II.getType() == types::TY_LLVM_BC) {
597       C.getDriver().Diag(diag::err_drv_no_linker_llvm_support)
598           << getToolChain().getTripleString();
599       continue;
600     }
601 
602     // Currently, we only pass the input files to the linker, we do not pass
603     // any libraries that may be valid only for the host.
604     if (!II.isFilename())
605       continue;
606 
607     const char *CubinF = C.addTempFile(
608         C.getArgs().MakeArgString(getToolChain().getInputFilename(II)));
609 
610     CmdArgs.push_back(CubinF);
611   }
612 
613   AddStaticDeviceLibsLinking(C, *this, JA, Inputs, Args, CmdArgs, "nvptx", GPUArch,
614                       false, false);
615 
616   // Find nvlink and pass it as "--nvlink-path=" argument of
617   // clang-nvlink-wrapper.
618   CmdArgs.push_back(Args.MakeArgString(
619       Twine("--nvlink-path=" + getToolChain().GetProgramPath("nvlink"))));
620 
621   const char *Exec =
622       Args.MakeArgString(getToolChain().GetProgramPath("clang-nvlink-wrapper"));
623   C.addCommand(std::make_unique<Command>(
624       JA, *this,
625       ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
626                           "--options-file"},
627       Exec, CmdArgs, Inputs, Output));
628 }
629 
630 /// CUDA toolchain.  Our assembler is ptxas, and our "linker" is fatbinary,
631 /// which isn't properly a linker but nonetheless performs the step of stitching
632 /// together object files from the assembler into a single blob.
633 
CudaToolChain(const Driver & D,const llvm::Triple & Triple,const ToolChain & HostTC,const ArgList & Args,const Action::OffloadKind OK)634 CudaToolChain::CudaToolChain(const Driver &D, const llvm::Triple &Triple,
635                              const ToolChain &HostTC, const ArgList &Args,
636                              const Action::OffloadKind OK)
637     : ToolChain(D, Triple, Args), HostTC(HostTC),
638       CudaInstallation(D, HostTC.getTriple(), Args), OK(OK) {
639   if (CudaInstallation.isValid()) {
640     CudaInstallation.WarnIfUnsupportedVersion();
641     getProgramPaths().push_back(std::string(CudaInstallation.getBinPath()));
642   }
643   // Lookup binaries into the driver directory, this is used to
644   // discover the clang-offload-bundler executable.
645   getProgramPaths().push_back(getDriver().Dir);
646 }
647 
getInputFilename(const InputInfo & Input) const648 std::string CudaToolChain::getInputFilename(const InputInfo &Input) const {
649   // Only object files are changed, for example assembly files keep their .s
650   // extensions. CUDA also continues to use .o as they don't use nvlink but
651   // fatbinary.
652   if (!(OK == Action::OFK_OpenMP && Input.getType() == types::TY_Object))
653     return ToolChain::getInputFilename(Input);
654 
655   // Replace extension for object files with cubin because nvlink relies on
656   // these particular file names.
657   SmallString<256> Filename(ToolChain::getInputFilename(Input));
658   llvm::sys::path::replace_extension(Filename, "cubin");
659   return std::string(Filename.str());
660 }
661 
addClangTargetOptions(const llvm::opt::ArgList & DriverArgs,llvm::opt::ArgStringList & CC1Args,Action::OffloadKind DeviceOffloadingKind) const662 void CudaToolChain::addClangTargetOptions(
663     const llvm::opt::ArgList &DriverArgs,
664     llvm::opt::ArgStringList &CC1Args,
665     Action::OffloadKind DeviceOffloadingKind) const {
666   HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind);
667 
668   StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
669   assert(!GpuArch.empty() && "Must have an explicit GPU arch.");
670   assert((DeviceOffloadingKind == Action::OFK_OpenMP ||
671           DeviceOffloadingKind == Action::OFK_Cuda) &&
672          "Only OpenMP or CUDA offloading kinds are supported for NVIDIA GPUs.");
673 
674   if (DeviceOffloadingKind == Action::OFK_Cuda) {
675     CC1Args.append(
676         {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
677 
678     if (DriverArgs.hasFlag(options::OPT_fcuda_approx_transcendentals,
679                            options::OPT_fno_cuda_approx_transcendentals, false))
680       CC1Args.push_back("-fcuda-approx-transcendentals");
681   }
682 
683   if (DriverArgs.hasArg(options::OPT_nogpulib))
684     return;
685 
686   if (DeviceOffloadingKind == Action::OFK_OpenMP &&
687       DriverArgs.hasArg(options::OPT_S))
688     return;
689 
690   std::string LibDeviceFile = CudaInstallation.getLibDeviceFile(GpuArch);
691   if (LibDeviceFile.empty()) {
692     getDriver().Diag(diag::err_drv_no_cuda_libdevice) << GpuArch;
693     return;
694   }
695 
696   CC1Args.push_back("-mlink-builtin-bitcode");
697   CC1Args.push_back(DriverArgs.MakeArgString(LibDeviceFile));
698 
699   clang::CudaVersion CudaInstallationVersion = CudaInstallation.version();
700 
701   // New CUDA versions often introduce new instructions that are only supported
702   // by new PTX version, so we need to raise PTX level to enable them in NVPTX
703   // back-end.
704   const char *PtxFeature = nullptr;
705   switch (CudaInstallationVersion) {
706 #define CASE_CUDA_VERSION(CUDA_VER, PTX_VER)                                   \
707   case CudaVersion::CUDA_##CUDA_VER:                                           \
708     PtxFeature = "+ptx" #PTX_VER;                                              \
709     break;
710     CASE_CUDA_VERSION(114, 74);
711     CASE_CUDA_VERSION(113, 73);
712     CASE_CUDA_VERSION(112, 72);
713     CASE_CUDA_VERSION(111, 71);
714     CASE_CUDA_VERSION(110, 70);
715     CASE_CUDA_VERSION(102, 65);
716     CASE_CUDA_VERSION(101, 64);
717     CASE_CUDA_VERSION(100, 63);
718     CASE_CUDA_VERSION(92, 61);
719     CASE_CUDA_VERSION(91, 61);
720     CASE_CUDA_VERSION(90, 60);
721 #undef CASE_CUDA_VERSION
722   default:
723     PtxFeature = "+ptx42";
724   }
725   CC1Args.append({"-target-feature", PtxFeature});
726   if (DriverArgs.hasFlag(options::OPT_fcuda_short_ptr,
727                          options::OPT_fno_cuda_short_ptr, false))
728     CC1Args.append({"-mllvm", "--nvptx-short-ptr"});
729 
730   if (CudaInstallationVersion >= CudaVersion::UNKNOWN)
731     CC1Args.push_back(
732         DriverArgs.MakeArgString(Twine("-target-sdk-version=") +
733                                  CudaVersionToString(CudaInstallationVersion)));
734 
735   if (DeviceOffloadingKind == Action::OFK_OpenMP) {
736     if (CudaInstallationVersion < CudaVersion::CUDA_92) {
737       getDriver().Diag(
738           diag::err_drv_omp_offload_target_cuda_version_not_support)
739           << CudaVersionToString(CudaInstallationVersion);
740       return;
741     }
742 
743     std::string BitcodeSuffix;
744     if (DriverArgs.hasFlag(options::OPT_fopenmp_target_new_runtime,
745                            options::OPT_fno_openmp_target_new_runtime, false))
746       BitcodeSuffix = "new-nvptx-" + GpuArch.str();
747     else
748       BitcodeSuffix = "nvptx-" + GpuArch.str();
749 
750     addOpenMPDeviceRTL(getDriver(), DriverArgs, CC1Args, BitcodeSuffix,
751                        getTriple());
752     AddStaticDeviceLibsPostLinking(getDriver(), DriverArgs, CC1Args, "nvptx", GpuArch,
753                         /* bitcode SDL?*/ true, /* PostClang Link? */ true);
754   }
755 }
756 
getDefaultDenormalModeForType(const llvm::opt::ArgList & DriverArgs,const JobAction & JA,const llvm::fltSemantics * FPType) const757 llvm::DenormalMode CudaToolChain::getDefaultDenormalModeForType(
758     const llvm::opt::ArgList &DriverArgs, const JobAction &JA,
759     const llvm::fltSemantics *FPType) const {
760   if (JA.getOffloadingDeviceKind() == Action::OFK_Cuda) {
761     if (FPType && FPType == &llvm::APFloat::IEEEsingle() &&
762         DriverArgs.hasFlag(options::OPT_fgpu_flush_denormals_to_zero,
763                            options::OPT_fno_gpu_flush_denormals_to_zero, false))
764       return llvm::DenormalMode::getPreserveSign();
765   }
766 
767   assert(JA.getOffloadingDeviceKind() != Action::OFK_Host);
768   return llvm::DenormalMode::getIEEE();
769 }
770 
supportsDebugInfoOption(const llvm::opt::Arg * A) const771 bool CudaToolChain::supportsDebugInfoOption(const llvm::opt::Arg *A) const {
772   const Option &O = A->getOption();
773   return (O.matches(options::OPT_gN_Group) &&
774           !O.matches(options::OPT_gmodules)) ||
775          O.matches(options::OPT_g_Flag) ||
776          O.matches(options::OPT_ggdbN_Group) || O.matches(options::OPT_ggdb) ||
777          O.matches(options::OPT_gdwarf) || O.matches(options::OPT_gdwarf_2) ||
778          O.matches(options::OPT_gdwarf_3) || O.matches(options::OPT_gdwarf_4) ||
779          O.matches(options::OPT_gdwarf_5) ||
780          O.matches(options::OPT_gcolumn_info);
781 }
782 
adjustDebugInfoKind(codegenoptions::DebugInfoKind & DebugInfoKind,const ArgList & Args) const783 void CudaToolChain::adjustDebugInfoKind(
784     codegenoptions::DebugInfoKind &DebugInfoKind, const ArgList &Args) const {
785   switch (mustEmitDebugInfo(Args)) {
786   case DisableDebugInfo:
787     DebugInfoKind = codegenoptions::NoDebugInfo;
788     break;
789   case DebugDirectivesOnly:
790     DebugInfoKind = codegenoptions::DebugDirectivesOnly;
791     break;
792   case EmitSameDebugInfoAsHost:
793     // Use same debug info level as the host.
794     break;
795   }
796 }
797 
AddCudaIncludeArgs(const ArgList & DriverArgs,ArgStringList & CC1Args) const798 void CudaToolChain::AddCudaIncludeArgs(const ArgList &DriverArgs,
799                                        ArgStringList &CC1Args) const {
800   // Check our CUDA version if we're going to include the CUDA headers.
801   if (!DriverArgs.hasArg(options::OPT_nogpuinc) &&
802       !DriverArgs.hasArg(options::OPT_no_cuda_version_check)) {
803     StringRef Arch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
804     assert(!Arch.empty() && "Must have an explicit GPU arch.");
805     CudaInstallation.CheckCudaVersionSupportsArch(StringToCudaArch(Arch));
806   }
807   CudaInstallation.AddCudaIncludeArgs(DriverArgs, CC1Args);
808 }
809 
810 llvm::opt::DerivedArgList *
TranslateArgs(const llvm::opt::DerivedArgList & Args,StringRef BoundArch,Action::OffloadKind DeviceOffloadKind) const811 CudaToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
812                              StringRef BoundArch,
813                              Action::OffloadKind DeviceOffloadKind) const {
814   DerivedArgList *DAL =
815       HostTC.TranslateArgs(Args, BoundArch, DeviceOffloadKind);
816   if (!DAL)
817     DAL = new DerivedArgList(Args.getBaseArgs());
818 
819   const OptTable &Opts = getDriver().getOpts();
820 
821   // For OpenMP device offloading, append derived arguments. Make sure
822   // flags are not duplicated.
823   // Also append the compute capability.
824   if (DeviceOffloadKind == Action::OFK_OpenMP) {
825     for (Arg *A : Args) {
826       bool IsDuplicate = false;
827       for (Arg *DALArg : *DAL) {
828         if (A == DALArg) {
829           IsDuplicate = true;
830           break;
831         }
832       }
833       if (!IsDuplicate)
834         DAL->append(A);
835     }
836 
837     StringRef Arch = DAL->getLastArgValue(options::OPT_march_EQ);
838     if (Arch.empty())
839       DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ),
840                         CLANG_OPENMP_NVPTX_DEFAULT_ARCH);
841 
842     return DAL;
843   }
844 
845   for (Arg *A : Args) {
846     DAL->append(A);
847   }
848 
849   if (!BoundArch.empty()) {
850     DAL->eraseArg(options::OPT_march_EQ);
851     DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ), BoundArch);
852   }
853   return DAL;
854 }
855 
buildAssembler() const856 Tool *CudaToolChain::buildAssembler() const {
857   return new tools::NVPTX::Assembler(*this);
858 }
859 
buildLinker() const860 Tool *CudaToolChain::buildLinker() const {
861   if (OK == Action::OFK_OpenMP)
862     return new tools::NVPTX::OpenMPLinker(*this);
863   return new tools::NVPTX::Linker(*this);
864 }
865 
addClangWarningOptions(ArgStringList & CC1Args) const866 void CudaToolChain::addClangWarningOptions(ArgStringList &CC1Args) const {
867   HostTC.addClangWarningOptions(CC1Args);
868 }
869 
870 ToolChain::CXXStdlibType
GetCXXStdlibType(const ArgList & Args) const871 CudaToolChain::GetCXXStdlibType(const ArgList &Args) const {
872   return HostTC.GetCXXStdlibType(Args);
873 }
874 
AddClangSystemIncludeArgs(const ArgList & DriverArgs,ArgStringList & CC1Args) const875 void CudaToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
876                                               ArgStringList &CC1Args) const {
877   HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args);
878 
879   if (!DriverArgs.hasArg(options::OPT_nogpuinc) && CudaInstallation.isValid())
880     CC1Args.append(
881         {"-internal-isystem",
882          DriverArgs.MakeArgString(CudaInstallation.getIncludePath())});
883 }
884 
AddClangCXXStdlibIncludeArgs(const ArgList & Args,ArgStringList & CC1Args) const885 void CudaToolChain::AddClangCXXStdlibIncludeArgs(const ArgList &Args,
886                                                  ArgStringList &CC1Args) const {
887   HostTC.AddClangCXXStdlibIncludeArgs(Args, CC1Args);
888 }
889 
AddIAMCUIncludeArgs(const ArgList & Args,ArgStringList & CC1Args) const890 void CudaToolChain::AddIAMCUIncludeArgs(const ArgList &Args,
891                                         ArgStringList &CC1Args) const {
892   HostTC.AddIAMCUIncludeArgs(Args, CC1Args);
893 }
894 
getSupportedSanitizers() const895 SanitizerMask CudaToolChain::getSupportedSanitizers() const {
896   // The CudaToolChain only supports sanitizers in the sense that it allows
897   // sanitizer arguments on the command line if they are supported by the host
898   // toolchain. The CudaToolChain will actually ignore any command line
899   // arguments for any of these "supported" sanitizers. That means that no
900   // sanitization of device code is actually supported at this time.
901   //
902   // This behavior is necessary because the host and device toolchains
903   // invocations often share the command line, so the device toolchain must
904   // tolerate flags meant only for the host toolchain.
905   return HostTC.getSupportedSanitizers();
906 }
907 
computeMSVCVersion(const Driver * D,const ArgList & Args) const908 VersionTuple CudaToolChain::computeMSVCVersion(const Driver *D,
909                                                const ArgList &Args) const {
910   return HostTC.computeMSVCVersion(D, Args);
911 }
912