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 if (raw_version < 11060)
69 return CudaVersion::CUDA_115;
70 return CudaVersion::NEW;
71 }
72
parseCudaHFile(llvm::StringRef Input)73 CudaVersion parseCudaHFile(llvm::StringRef Input) {
74 // Helper lambda which skips the words if the line starts with them or returns
75 // None otherwise.
76 auto StartsWithWords =
77 [](llvm::StringRef Line,
78 const SmallVector<StringRef, 3> words) -> llvm::Optional<StringRef> {
79 for (StringRef word : words) {
80 if (!Line.consume_front(word))
81 return {};
82 Line = Line.ltrim();
83 }
84 return Line;
85 };
86
87 Input = Input.ltrim();
88 while (!Input.empty()) {
89 if (auto Line =
90 StartsWithWords(Input.ltrim(), {"#", "define", "CUDA_VERSION"})) {
91 uint32_t RawVersion;
92 Line->consumeInteger(10, RawVersion);
93 return getCudaVersion(RawVersion);
94 }
95 // Find next non-empty line.
96 Input = Input.drop_front(Input.find_first_of("\n\r")).ltrim();
97 }
98 return CudaVersion::UNKNOWN;
99 }
100 } // namespace
101
WarnIfUnsupportedVersion()102 void CudaInstallationDetector::WarnIfUnsupportedVersion() {
103 if (Version > CudaVersion::PARTIALLY_SUPPORTED) {
104 std::string VersionString = CudaVersionToString(Version);
105 if (!VersionString.empty())
106 VersionString.insert(0, " ");
107 D.Diag(diag::warn_drv_new_cuda_version)
108 << VersionString
109 << (CudaVersion::PARTIALLY_SUPPORTED != CudaVersion::FULLY_SUPPORTED)
110 << CudaVersionToString(CudaVersion::PARTIALLY_SUPPORTED);
111 } else if (Version > CudaVersion::FULLY_SUPPORTED)
112 D.Diag(diag::warn_drv_partially_supported_cuda_version)
113 << CudaVersionToString(Version);
114 }
115
CudaInstallationDetector(const Driver & D,const llvm::Triple & HostTriple,const llvm::opt::ArgList & Args)116 CudaInstallationDetector::CudaInstallationDetector(
117 const Driver &D, const llvm::Triple &HostTriple,
118 const llvm::opt::ArgList &Args)
119 : D(D) {
120 struct Candidate {
121 std::string Path;
122 bool StrictChecking;
123
124 Candidate(std::string Path, bool StrictChecking = false)
125 : Path(Path), StrictChecking(StrictChecking) {}
126 };
127 SmallVector<Candidate, 4> Candidates;
128
129 // In decreasing order so we prefer newer versions to older versions.
130 std::initializer_list<const char *> Versions = {"8.0", "7.5", "7.0"};
131 auto &FS = D.getVFS();
132
133 if (Args.hasArg(clang::driver::options::OPT_cuda_path_EQ)) {
134 Candidates.emplace_back(
135 Args.getLastArgValue(clang::driver::options::OPT_cuda_path_EQ).str());
136 } else if (HostTriple.isOSWindows()) {
137 for (const char *Ver : Versions)
138 Candidates.emplace_back(
139 D.SysRoot + "/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v" +
140 Ver);
141 } else {
142 if (!Args.hasArg(clang::driver::options::OPT_cuda_path_ignore_env)) {
143 // Try to find ptxas binary. If the executable is located in a directory
144 // called 'bin/', its parent directory might be a good guess for a valid
145 // CUDA installation.
146 // However, some distributions might installs 'ptxas' to /usr/bin. In that
147 // case the candidate would be '/usr' which passes the following checks
148 // because '/usr/include' exists as well. To avoid this case, we always
149 // check for the directory potentially containing files for libdevice,
150 // even if the user passes -nocudalib.
151 if (llvm::ErrorOr<std::string> ptxas =
152 llvm::sys::findProgramByName("ptxas")) {
153 SmallString<256> ptxasAbsolutePath;
154 llvm::sys::fs::real_path(*ptxas, ptxasAbsolutePath);
155
156 StringRef ptxasDir = llvm::sys::path::parent_path(ptxasAbsolutePath);
157 if (llvm::sys::path::filename(ptxasDir) == "bin")
158 Candidates.emplace_back(
159 std::string(llvm::sys::path::parent_path(ptxasDir)),
160 /*StrictChecking=*/true);
161 }
162 }
163
164 Candidates.emplace_back(D.SysRoot + "/usr/local/cuda");
165 for (const char *Ver : Versions)
166 Candidates.emplace_back(D.SysRoot + "/usr/local/cuda-" + Ver);
167
168 Distro Dist(FS, llvm::Triple(llvm::sys::getProcessTriple()));
169 if (Dist.IsDebian() || Dist.IsUbuntu())
170 // Special case for Debian to have nvidia-cuda-toolkit work
171 // out of the box. More info on http://bugs.debian.org/882505
172 Candidates.emplace_back(D.SysRoot + "/usr/lib/cuda");
173 }
174
175 bool NoCudaLib = Args.hasArg(options::OPT_nogpulib);
176
177 for (const auto &Candidate : Candidates) {
178 InstallPath = Candidate.Path;
179 if (InstallPath.empty() || !FS.exists(InstallPath))
180 continue;
181
182 BinPath = InstallPath + "/bin";
183 IncludePath = InstallPath + "/include";
184 LibDevicePath = InstallPath + "/nvvm/libdevice";
185
186 if (!(FS.exists(IncludePath) && FS.exists(BinPath)))
187 continue;
188 bool CheckLibDevice = (!NoCudaLib || Candidate.StrictChecking);
189 if (CheckLibDevice && !FS.exists(LibDevicePath))
190 continue;
191
192 // On Linux, we have both lib and lib64 directories, and we need to choose
193 // based on our triple. On MacOS, we have only a lib directory.
194 //
195 // It's sufficient for our purposes to be flexible: If both lib and lib64
196 // exist, we choose whichever one matches our triple. Otherwise, if only
197 // lib exists, we use it.
198 if (HostTriple.isArch64Bit() && FS.exists(InstallPath + "/lib64"))
199 LibPath = InstallPath + "/lib64";
200 else if (FS.exists(InstallPath + "/lib"))
201 LibPath = InstallPath + "/lib";
202 else
203 continue;
204
205 Version = CudaVersion::UNKNOWN;
206 if (auto CudaHFile = FS.getBufferForFile(InstallPath + "/include/cuda.h"))
207 Version = parseCudaHFile((*CudaHFile)->getBuffer());
208 // As the last resort, make an educated guess between CUDA-7.0, which had
209 // old-style libdevice bitcode, and an unknown recent CUDA version.
210 if (Version == CudaVersion::UNKNOWN) {
211 Version = FS.exists(LibDevicePath + "/libdevice.10.bc")
212 ? CudaVersion::NEW
213 : CudaVersion::CUDA_70;
214 }
215
216 if (Version >= CudaVersion::CUDA_90) {
217 // CUDA-9+ uses single libdevice file for all GPU variants.
218 std::string FilePath = LibDevicePath + "/libdevice.10.bc";
219 if (FS.exists(FilePath)) {
220 for (int Arch = (int)CudaArch::SM_30, E = (int)CudaArch::LAST; Arch < E;
221 ++Arch) {
222 CudaArch GpuArch = static_cast<CudaArch>(Arch);
223 if (!IsNVIDIAGpuArch(GpuArch))
224 continue;
225 std::string GpuArchName(CudaArchToString(GpuArch));
226 LibDeviceMap[GpuArchName] = FilePath;
227 }
228 }
229 } else {
230 std::error_code EC;
231 for (llvm::vfs::directory_iterator LI = FS.dir_begin(LibDevicePath, EC),
232 LE;
233 !EC && LI != LE; LI = LI.increment(EC)) {
234 StringRef FilePath = LI->path();
235 StringRef FileName = llvm::sys::path::filename(FilePath);
236 // Process all bitcode filenames that look like
237 // libdevice.compute_XX.YY.bc
238 const StringRef LibDeviceName = "libdevice.";
239 if (!(FileName.startswith(LibDeviceName) && FileName.endswith(".bc")))
240 continue;
241 StringRef GpuArch = FileName.slice(
242 LibDeviceName.size(), FileName.find('.', LibDeviceName.size()));
243 LibDeviceMap[GpuArch] = FilePath.str();
244 // Insert map entries for specific devices with this compute
245 // capability. NVCC's choice of the libdevice library version is
246 // rather peculiar and depends on the CUDA version.
247 if (GpuArch == "compute_20") {
248 LibDeviceMap["sm_20"] = std::string(FilePath);
249 LibDeviceMap["sm_21"] = std::string(FilePath);
250 LibDeviceMap["sm_32"] = std::string(FilePath);
251 } else if (GpuArch == "compute_30") {
252 LibDeviceMap["sm_30"] = std::string(FilePath);
253 if (Version < CudaVersion::CUDA_80) {
254 LibDeviceMap["sm_50"] = std::string(FilePath);
255 LibDeviceMap["sm_52"] = std::string(FilePath);
256 LibDeviceMap["sm_53"] = std::string(FilePath);
257 }
258 LibDeviceMap["sm_60"] = std::string(FilePath);
259 LibDeviceMap["sm_61"] = std::string(FilePath);
260 LibDeviceMap["sm_62"] = std::string(FilePath);
261 } else if (GpuArch == "compute_35") {
262 LibDeviceMap["sm_35"] = std::string(FilePath);
263 LibDeviceMap["sm_37"] = std::string(FilePath);
264 } else if (GpuArch == "compute_50") {
265 if (Version >= CudaVersion::CUDA_80) {
266 LibDeviceMap["sm_50"] = std::string(FilePath);
267 LibDeviceMap["sm_52"] = std::string(FilePath);
268 LibDeviceMap["sm_53"] = std::string(FilePath);
269 }
270 }
271 }
272 }
273
274 // Check that we have found at least one libdevice that we can link in if
275 // -nocudalib hasn't been specified.
276 if (LibDeviceMap.empty() && !NoCudaLib)
277 continue;
278
279 IsValid = true;
280 break;
281 }
282 }
283
AddCudaIncludeArgs(const ArgList & DriverArgs,ArgStringList & CC1Args) const284 void CudaInstallationDetector::AddCudaIncludeArgs(
285 const ArgList &DriverArgs, ArgStringList &CC1Args) const {
286 if (!DriverArgs.hasArg(options::OPT_nobuiltininc)) {
287 // Add cuda_wrappers/* to our system include path. This lets us wrap
288 // standard library headers.
289 SmallString<128> P(D.ResourceDir);
290 llvm::sys::path::append(P, "include");
291 llvm::sys::path::append(P, "cuda_wrappers");
292 CC1Args.push_back("-internal-isystem");
293 CC1Args.push_back(DriverArgs.MakeArgString(P));
294 }
295
296 if (DriverArgs.hasArg(options::OPT_nogpuinc))
297 return;
298
299 if (!isValid()) {
300 D.Diag(diag::err_drv_no_cuda_installation);
301 return;
302 }
303
304 CC1Args.push_back("-include");
305 CC1Args.push_back("__clang_cuda_runtime_wrapper.h");
306 }
307
CheckCudaVersionSupportsArch(CudaArch Arch) const308 void CudaInstallationDetector::CheckCudaVersionSupportsArch(
309 CudaArch Arch) const {
310 if (Arch == CudaArch::UNKNOWN || Version == CudaVersion::UNKNOWN ||
311 ArchsWithBadVersion[(int)Arch])
312 return;
313
314 auto MinVersion = MinVersionForCudaArch(Arch);
315 auto MaxVersion = MaxVersionForCudaArch(Arch);
316 if (Version < MinVersion || Version > MaxVersion) {
317 ArchsWithBadVersion[(int)Arch] = true;
318 D.Diag(diag::err_drv_cuda_version_unsupported)
319 << CudaArchToString(Arch) << CudaVersionToString(MinVersion)
320 << CudaVersionToString(MaxVersion) << InstallPath
321 << CudaVersionToString(Version);
322 }
323 }
324
print(raw_ostream & OS) const325 void CudaInstallationDetector::print(raw_ostream &OS) const {
326 if (isValid())
327 OS << "Found CUDA installation: " << InstallPath << ", version "
328 << CudaVersionToString(Version) << "\n";
329 }
330
331 namespace {
332 /// Debug info level for the NVPTX devices. We may need to emit different debug
333 /// info level for the host and for the device itselfi. This type controls
334 /// emission of the debug info for the devices. It either prohibits disable info
335 /// emission completely, or emits debug directives only, or emits same debug
336 /// info as for the host.
337 enum DeviceDebugInfoLevel {
338 DisableDebugInfo, /// Do not emit debug info for the devices.
339 DebugDirectivesOnly, /// Emit only debug directives.
340 EmitSameDebugInfoAsHost, /// Use the same debug info level just like for the
341 /// host.
342 };
343 } // anonymous namespace
344
345 /// Define debug info level for the NVPTX devices. If the debug info for both
346 /// the host and device are disabled (-g0/-ggdb0 or no debug options at all). If
347 /// only debug directives are requested for the both host and device
348 /// (-gline-directvies-only), or the debug info only for the device is disabled
349 /// (optimization is on and --cuda-noopt-device-debug was not specified), the
350 /// debug directves only must be emitted for the device. Otherwise, use the same
351 /// debug info level just like for the host (with the limitations of only
352 /// supported DWARF2 standard).
mustEmitDebugInfo(const ArgList & Args)353 static DeviceDebugInfoLevel mustEmitDebugInfo(const ArgList &Args) {
354 const Arg *A = Args.getLastArg(options::OPT_O_Group);
355 bool IsDebugEnabled = !A || A->getOption().matches(options::OPT_O0) ||
356 Args.hasFlag(options::OPT_cuda_noopt_device_debug,
357 options::OPT_no_cuda_noopt_device_debug,
358 /*Default=*/false);
359 if (const Arg *A = Args.getLastArg(options::OPT_g_Group)) {
360 const Option &Opt = A->getOption();
361 if (Opt.matches(options::OPT_gN_Group)) {
362 if (Opt.matches(options::OPT_g0) || Opt.matches(options::OPT_ggdb0))
363 return DisableDebugInfo;
364 if (Opt.matches(options::OPT_gline_directives_only))
365 return DebugDirectivesOnly;
366 }
367 return IsDebugEnabled ? EmitSameDebugInfoAsHost : DebugDirectivesOnly;
368 }
369 return willEmitRemarks(Args) ? DebugDirectivesOnly : DisableDebugInfo;
370 }
371
ConstructJob(Compilation & C,const JobAction & JA,const InputInfo & Output,const InputInfoList & Inputs,const ArgList & Args,const char * LinkingOutput) const372 void NVPTX::Assembler::ConstructJob(Compilation &C, const JobAction &JA,
373 const InputInfo &Output,
374 const InputInfoList &Inputs,
375 const ArgList &Args,
376 const char *LinkingOutput) const {
377 const auto &TC =
378 static_cast<const toolchains::CudaToolChain &>(getToolChain());
379 assert(TC.getTriple().isNVPTX() && "Wrong platform");
380
381 StringRef GPUArchName;
382 // If this is an OpenMP action we need to extract the device architecture
383 // from the -march=arch option. This option may come from -Xopenmp-target
384 // flag or the default value.
385 if (JA.isDeviceOffloading(Action::OFK_OpenMP)) {
386 GPUArchName = Args.getLastArgValue(options::OPT_march_EQ);
387 assert(!GPUArchName.empty() && "Must have an architecture passed in.");
388 } else
389 GPUArchName = JA.getOffloadingArch();
390
391 // Obtain architecture from the action.
392 CudaArch gpu_arch = StringToCudaArch(GPUArchName);
393 assert(gpu_arch != CudaArch::UNKNOWN &&
394 "Device action expected to have an architecture.");
395
396 // Check that our installation's ptxas supports gpu_arch.
397 if (!Args.hasArg(options::OPT_no_cuda_version_check)) {
398 TC.CudaInstallation.CheckCudaVersionSupportsArch(gpu_arch);
399 }
400
401 ArgStringList CmdArgs;
402 CmdArgs.push_back(TC.getTriple().isArch64Bit() ? "-m64" : "-m32");
403 DeviceDebugInfoLevel DIKind = mustEmitDebugInfo(Args);
404 if (DIKind == EmitSameDebugInfoAsHost) {
405 // ptxas does not accept -g option if optimization is enabled, so
406 // we ignore the compiler's -O* options if we want debug info.
407 CmdArgs.push_back("-g");
408 CmdArgs.push_back("--dont-merge-basicblocks");
409 CmdArgs.push_back("--return-at-end");
410 } else if (Arg *A = Args.getLastArg(options::OPT_O_Group)) {
411 // Map the -O we received to -O{0,1,2,3}.
412 //
413 // TODO: Perhaps we should map host -O2 to ptxas -O3. -O3 is ptxas's
414 // default, so it may correspond more closely to the spirit of clang -O2.
415
416 // -O3 seems like the least-bad option when -Osomething is specified to
417 // clang but it isn't handled below.
418 StringRef OOpt = "3";
419 if (A->getOption().matches(options::OPT_O4) ||
420 A->getOption().matches(options::OPT_Ofast))
421 OOpt = "3";
422 else if (A->getOption().matches(options::OPT_O0))
423 OOpt = "0";
424 else if (A->getOption().matches(options::OPT_O)) {
425 // -Os, -Oz, and -O(anything else) map to -O2, for lack of better options.
426 OOpt = llvm::StringSwitch<const char *>(A->getValue())
427 .Case("1", "1")
428 .Case("2", "2")
429 .Case("3", "3")
430 .Case("s", "2")
431 .Case("z", "2")
432 .Default("2");
433 }
434 CmdArgs.push_back(Args.MakeArgString(llvm::Twine("-O") + OOpt));
435 } else {
436 // If no -O was passed, pass -O0 to ptxas -- no opt flag should correspond
437 // to no optimizations, but ptxas's default is -O3.
438 CmdArgs.push_back("-O0");
439 }
440 if (DIKind == DebugDirectivesOnly)
441 CmdArgs.push_back("-lineinfo");
442
443 // Pass -v to ptxas if it was passed to the driver.
444 if (Args.hasArg(options::OPT_v))
445 CmdArgs.push_back("-v");
446
447 CmdArgs.push_back("--gpu-name");
448 CmdArgs.push_back(Args.MakeArgString(CudaArchToString(gpu_arch)));
449 CmdArgs.push_back("--output-file");
450 const char *OutputFileName = Args.MakeArgString(TC.getInputFilename(Output));
451 if (std::string(OutputFileName) != std::string(Output.getFilename()))
452 C.addTempFile(OutputFileName);
453 CmdArgs.push_back(OutputFileName);
454 for (const auto& II : Inputs)
455 CmdArgs.push_back(Args.MakeArgString(II.getFilename()));
456
457 for (const auto& A : Args.getAllArgValues(options::OPT_Xcuda_ptxas))
458 CmdArgs.push_back(Args.MakeArgString(A));
459
460 bool Relocatable = false;
461 if (JA.isOffloading(Action::OFK_OpenMP))
462 // In OpenMP we need to generate relocatable code.
463 Relocatable = Args.hasFlag(options::OPT_fopenmp_relocatable_target,
464 options::OPT_fnoopenmp_relocatable_target,
465 /*Default=*/true);
466 else if (JA.isOffloading(Action::OFK_Cuda))
467 Relocatable = Args.hasFlag(options::OPT_fgpu_rdc,
468 options::OPT_fno_gpu_rdc, /*Default=*/false);
469
470 if (Relocatable)
471 CmdArgs.push_back("-c");
472
473 const char *Exec;
474 if (Arg *A = Args.getLastArg(options::OPT_ptxas_path_EQ))
475 Exec = A->getValue();
476 else
477 Exec = Args.MakeArgString(TC.GetProgramPath("ptxas"));
478 C.addCommand(std::make_unique<Command>(
479 JA, *this,
480 ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
481 "--options-file"},
482 Exec, CmdArgs, Inputs, Output));
483 }
484
shouldIncludePTX(const ArgList & Args,const char * gpu_arch)485 static bool shouldIncludePTX(const ArgList &Args, const char *gpu_arch) {
486 bool includePTX = true;
487 for (Arg *A : Args) {
488 if (!(A->getOption().matches(options::OPT_cuda_include_ptx_EQ) ||
489 A->getOption().matches(options::OPT_no_cuda_include_ptx_EQ)))
490 continue;
491 A->claim();
492 const StringRef ArchStr = A->getValue();
493 if (ArchStr == "all" || ArchStr == gpu_arch) {
494 includePTX = A->getOption().matches(options::OPT_cuda_include_ptx_EQ);
495 continue;
496 }
497 }
498 return includePTX;
499 }
500
501 // All inputs to this linker must be from CudaDeviceActions, as we need to look
502 // at the Inputs' Actions in order to figure out which GPU architecture they
503 // correspond to.
ConstructJob(Compilation & C,const JobAction & JA,const InputInfo & Output,const InputInfoList & Inputs,const ArgList & Args,const char * LinkingOutput) const504 void NVPTX::Linker::ConstructJob(Compilation &C, const JobAction &JA,
505 const InputInfo &Output,
506 const InputInfoList &Inputs,
507 const ArgList &Args,
508 const char *LinkingOutput) const {
509 const auto &TC =
510 static_cast<const toolchains::CudaToolChain &>(getToolChain());
511 assert(TC.getTriple().isNVPTX() && "Wrong platform");
512
513 ArgStringList CmdArgs;
514 if (TC.CudaInstallation.version() <= CudaVersion::CUDA_100)
515 CmdArgs.push_back("--cuda");
516 CmdArgs.push_back(TC.getTriple().isArch64Bit() ? "-64" : "-32");
517 CmdArgs.push_back(Args.MakeArgString("--create"));
518 CmdArgs.push_back(Args.MakeArgString(Output.getFilename()));
519 if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost)
520 CmdArgs.push_back("-g");
521
522 for (const auto& II : Inputs) {
523 auto *A = II.getAction();
524 assert(A->getInputs().size() == 1 &&
525 "Device offload action is expected to have a single input");
526 const char *gpu_arch_str = A->getOffloadingArch();
527 assert(gpu_arch_str &&
528 "Device action expected to have associated a GPU architecture!");
529 CudaArch gpu_arch = StringToCudaArch(gpu_arch_str);
530
531 if (II.getType() == types::TY_PP_Asm &&
532 !shouldIncludePTX(Args, gpu_arch_str))
533 continue;
534 // We need to pass an Arch of the form "sm_XX" for cubin files and
535 // "compute_XX" for ptx.
536 const char *Arch = (II.getType() == types::TY_PP_Asm)
537 ? CudaArchToVirtualArchString(gpu_arch)
538 : gpu_arch_str;
539 CmdArgs.push_back(
540 Args.MakeArgString(llvm::Twine("--image=profile=") + Arch +
541 ",file=" + getToolChain().getInputFilename(II)));
542 }
543
544 for (const auto& A : Args.getAllArgValues(options::OPT_Xcuda_fatbinary))
545 CmdArgs.push_back(Args.MakeArgString(A));
546
547 const char *Exec = Args.MakeArgString(TC.GetProgramPath("fatbinary"));
548 C.addCommand(std::make_unique<Command>(
549 JA, *this,
550 ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
551 "--options-file"},
552 Exec, CmdArgs, Inputs, Output));
553 }
554
ConstructJob(Compilation & C,const JobAction & JA,const InputInfo & Output,const InputInfoList & Inputs,const ArgList & Args,const char * LinkingOutput) const555 void NVPTX::OpenMPLinker::ConstructJob(Compilation &C, const JobAction &JA,
556 const InputInfo &Output,
557 const InputInfoList &Inputs,
558 const ArgList &Args,
559 const char *LinkingOutput) const {
560 const auto &TC =
561 static_cast<const toolchains::CudaToolChain &>(getToolChain());
562 assert(TC.getTriple().isNVPTX() && "Wrong platform");
563
564 ArgStringList CmdArgs;
565
566 // OpenMP uses nvlink to link cubin files. The result will be embedded in the
567 // host binary by the host linker.
568 assert(!JA.isHostOffloading(Action::OFK_OpenMP) &&
569 "CUDA toolchain not expected for an OpenMP host device.");
570
571 if (Output.isFilename()) {
572 CmdArgs.push_back("-o");
573 CmdArgs.push_back(Output.getFilename());
574 } else
575 assert(Output.isNothing() && "Invalid output.");
576 if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost)
577 CmdArgs.push_back("-g");
578
579 if (Args.hasArg(options::OPT_v))
580 CmdArgs.push_back("-v");
581
582 StringRef GPUArch =
583 Args.getLastArgValue(options::OPT_march_EQ);
584 assert(!GPUArch.empty() && "At least one GPU Arch required for ptxas.");
585
586 CmdArgs.push_back("-arch");
587 CmdArgs.push_back(Args.MakeArgString(GPUArch));
588
589 // Add paths specified in LIBRARY_PATH environment variable as -L options.
590 addDirectoryList(Args, CmdArgs, "-L", "LIBRARY_PATH");
591
592 // Add paths for the default clang library path.
593 SmallString<256> DefaultLibPath =
594 llvm::sys::path::parent_path(TC.getDriver().Dir);
595 llvm::sys::path::append(DefaultLibPath, "lib" CLANG_LIBDIR_SUFFIX);
596 CmdArgs.push_back(Args.MakeArgString(Twine("-L") + DefaultLibPath));
597
598 for (const auto &II : Inputs) {
599 if (II.getType() == types::TY_LLVM_IR ||
600 II.getType() == types::TY_LTO_IR ||
601 II.getType() == types::TY_LTO_BC ||
602 II.getType() == types::TY_LLVM_BC) {
603 C.getDriver().Diag(diag::err_drv_no_linker_llvm_support)
604 << getToolChain().getTripleString();
605 continue;
606 }
607
608 // Currently, we only pass the input files to the linker, we do not pass
609 // any libraries that may be valid only for the host.
610 if (!II.isFilename())
611 continue;
612
613 const char *CubinF =
614 C.getArgs().MakeArgString(getToolChain().getInputFilename(II));
615
616 CmdArgs.push_back(CubinF);
617 }
618
619 AddStaticDeviceLibsLinking(C, *this, JA, Inputs, Args, CmdArgs, "nvptx",
620 GPUArch, /*isBitCodeSDL=*/false,
621 /*postClangLink=*/false);
622
623 // Find nvlink and pass it as "--nvlink-path=" argument of
624 // clang-nvlink-wrapper.
625 CmdArgs.push_back(Args.MakeArgString(
626 Twine("--nvlink-path=" + getToolChain().GetProgramPath("nvlink"))));
627
628 const char *Exec =
629 Args.MakeArgString(getToolChain().GetProgramPath("clang-nvlink-wrapper"));
630 C.addCommand(std::make_unique<Command>(
631 JA, *this,
632 ResponseFileSupport{ResponseFileSupport::RF_Full, llvm::sys::WEM_UTF8,
633 "--options-file"},
634 Exec, CmdArgs, Inputs, Output));
635 }
636
getNVPTXTargetFeatures(const Driver & D,const llvm::Triple & Triple,const llvm::opt::ArgList & Args,std::vector<StringRef> & Features)637 void NVPTX::getNVPTXTargetFeatures(const Driver &D, const llvm::Triple &Triple,
638 const llvm::opt::ArgList &Args,
639 std::vector<StringRef> &Features) {
640 if (Args.hasArg(options::OPT_cuda_feature_EQ)) {
641 StringRef PtxFeature =
642 Args.getLastArgValue(options::OPT_cuda_feature_EQ, "+ptx42");
643 Features.push_back(Args.MakeArgString(PtxFeature));
644 return;
645 }
646 CudaInstallationDetector CudaInstallation(D, Triple, Args);
647
648 // New CUDA versions often introduce new instructions that are only supported
649 // by new PTX version, so we need to raise PTX level to enable them in NVPTX
650 // back-end.
651 const char *PtxFeature = nullptr;
652 switch (CudaInstallation.version()) {
653 #define CASE_CUDA_VERSION(CUDA_VER, PTX_VER) \
654 case CudaVersion::CUDA_##CUDA_VER: \
655 PtxFeature = "+ptx" #PTX_VER; \
656 break;
657 CASE_CUDA_VERSION(115, 75);
658 CASE_CUDA_VERSION(114, 74);
659 CASE_CUDA_VERSION(113, 73);
660 CASE_CUDA_VERSION(112, 72);
661 CASE_CUDA_VERSION(111, 71);
662 CASE_CUDA_VERSION(110, 70);
663 CASE_CUDA_VERSION(102, 65);
664 CASE_CUDA_VERSION(101, 64);
665 CASE_CUDA_VERSION(100, 63);
666 CASE_CUDA_VERSION(92, 61);
667 CASE_CUDA_VERSION(91, 61);
668 CASE_CUDA_VERSION(90, 60);
669 #undef CASE_CUDA_VERSION
670 default:
671 PtxFeature = "+ptx42";
672 }
673 Features.push_back(PtxFeature);
674 }
675
676 /// CUDA toolchain. Our assembler is ptxas, and our "linker" is fatbinary,
677 /// which isn't properly a linker but nonetheless performs the step of stitching
678 /// together object files from the assembler into a single blob.
679
CudaToolChain(const Driver & D,const llvm::Triple & Triple,const ToolChain & HostTC,const ArgList & Args,const Action::OffloadKind OK)680 CudaToolChain::CudaToolChain(const Driver &D, const llvm::Triple &Triple,
681 const ToolChain &HostTC, const ArgList &Args,
682 const Action::OffloadKind OK)
683 : ToolChain(D, Triple, Args), HostTC(HostTC),
684 CudaInstallation(D, HostTC.getTriple(), Args), OK(OK) {
685 if (CudaInstallation.isValid()) {
686 CudaInstallation.WarnIfUnsupportedVersion();
687 getProgramPaths().push_back(std::string(CudaInstallation.getBinPath()));
688 }
689 // Lookup binaries into the driver directory, this is used to
690 // discover the clang-offload-bundler executable.
691 getProgramPaths().push_back(getDriver().Dir);
692 }
693
getInputFilename(const InputInfo & Input) const694 std::string CudaToolChain::getInputFilename(const InputInfo &Input) const {
695 // Only object files are changed, for example assembly files keep their .s
696 // extensions. If the user requested device-only compilation don't change it.
697 if (Input.getType() != types::TY_Object || getDriver().offloadDeviceOnly())
698 return ToolChain::getInputFilename(Input);
699
700 // Replace extension for object files with cubin because nvlink relies on
701 // these particular file names.
702 SmallString<256> Filename(ToolChain::getInputFilename(Input));
703 llvm::sys::path::replace_extension(Filename, "cubin");
704 return std::string(Filename.str());
705 }
706
addClangTargetOptions(const llvm::opt::ArgList & DriverArgs,llvm::opt::ArgStringList & CC1Args,Action::OffloadKind DeviceOffloadingKind) const707 void CudaToolChain::addClangTargetOptions(
708 const llvm::opt::ArgList &DriverArgs,
709 llvm::opt::ArgStringList &CC1Args,
710 Action::OffloadKind DeviceOffloadingKind) const {
711 HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind);
712
713 StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
714 assert(!GpuArch.empty() && "Must have an explicit GPU arch.");
715 assert((DeviceOffloadingKind == Action::OFK_OpenMP ||
716 DeviceOffloadingKind == Action::OFK_Cuda) &&
717 "Only OpenMP or CUDA offloading kinds are supported for NVIDIA GPUs.");
718
719 if (DeviceOffloadingKind == Action::OFK_Cuda) {
720 CC1Args.append(
721 {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
722
723 if (DriverArgs.hasFlag(options::OPT_fcuda_approx_transcendentals,
724 options::OPT_fno_cuda_approx_transcendentals, false))
725 CC1Args.push_back("-fcuda-approx-transcendentals");
726 }
727
728 if (DriverArgs.hasArg(options::OPT_nogpulib))
729 return;
730
731 if (DeviceOffloadingKind == Action::OFK_OpenMP &&
732 DriverArgs.hasArg(options::OPT_S))
733 return;
734
735 std::string LibDeviceFile = CudaInstallation.getLibDeviceFile(GpuArch);
736 if (LibDeviceFile.empty()) {
737 getDriver().Diag(diag::err_drv_no_cuda_libdevice) << GpuArch;
738 return;
739 }
740
741 CC1Args.push_back("-mlink-builtin-bitcode");
742 CC1Args.push_back(DriverArgs.MakeArgString(LibDeviceFile));
743
744 clang::CudaVersion CudaInstallationVersion = CudaInstallation.version();
745
746 if (DriverArgs.hasFlag(options::OPT_fcuda_short_ptr,
747 options::OPT_fno_cuda_short_ptr, false))
748 CC1Args.append({"-mllvm", "--nvptx-short-ptr"});
749
750 if (CudaInstallationVersion >= CudaVersion::UNKNOWN)
751 CC1Args.push_back(
752 DriverArgs.MakeArgString(Twine("-target-sdk-version=") +
753 CudaVersionToString(CudaInstallationVersion)));
754
755 if (DeviceOffloadingKind == Action::OFK_OpenMP) {
756 if (CudaInstallationVersion < CudaVersion::CUDA_92) {
757 getDriver().Diag(
758 diag::err_drv_omp_offload_target_cuda_version_not_support)
759 << CudaVersionToString(CudaInstallationVersion);
760 return;
761 }
762
763 // Link the bitcode library late if we're using device LTO.
764 if (getDriver().isUsingLTO(/* IsOffload */ true))
765 return;
766
767 addOpenMPDeviceRTL(getDriver(), DriverArgs, CC1Args, GpuArch.str(),
768 getTriple());
769 AddStaticDeviceLibsPostLinking(getDriver(), DriverArgs, CC1Args, "nvptx",
770 GpuArch, /*isBitCodeSDL=*/true,
771 /*postClangLink=*/true);
772 }
773 }
774
getDefaultDenormalModeForType(const llvm::opt::ArgList & DriverArgs,const JobAction & JA,const llvm::fltSemantics * FPType) const775 llvm::DenormalMode CudaToolChain::getDefaultDenormalModeForType(
776 const llvm::opt::ArgList &DriverArgs, const JobAction &JA,
777 const llvm::fltSemantics *FPType) const {
778 if (JA.getOffloadingDeviceKind() == Action::OFK_Cuda) {
779 if (FPType && FPType == &llvm::APFloat::IEEEsingle() &&
780 DriverArgs.hasFlag(options::OPT_fgpu_flush_denormals_to_zero,
781 options::OPT_fno_gpu_flush_denormals_to_zero, false))
782 return llvm::DenormalMode::getPreserveSign();
783 }
784
785 assert(JA.getOffloadingDeviceKind() != Action::OFK_Host);
786 return llvm::DenormalMode::getIEEE();
787 }
788
supportsDebugInfoOption(const llvm::opt::Arg * A) const789 bool CudaToolChain::supportsDebugInfoOption(const llvm::opt::Arg *A) const {
790 const Option &O = A->getOption();
791 return (O.matches(options::OPT_gN_Group) &&
792 !O.matches(options::OPT_gmodules)) ||
793 O.matches(options::OPT_g_Flag) ||
794 O.matches(options::OPT_ggdbN_Group) || O.matches(options::OPT_ggdb) ||
795 O.matches(options::OPT_gdwarf) || O.matches(options::OPT_gdwarf_2) ||
796 O.matches(options::OPT_gdwarf_3) || O.matches(options::OPT_gdwarf_4) ||
797 O.matches(options::OPT_gdwarf_5) ||
798 O.matches(options::OPT_gcolumn_info);
799 }
800
adjustDebugInfoKind(codegenoptions::DebugInfoKind & DebugInfoKind,const ArgList & Args) const801 void CudaToolChain::adjustDebugInfoKind(
802 codegenoptions::DebugInfoKind &DebugInfoKind, const ArgList &Args) const {
803 switch (mustEmitDebugInfo(Args)) {
804 case DisableDebugInfo:
805 DebugInfoKind = codegenoptions::NoDebugInfo;
806 break;
807 case DebugDirectivesOnly:
808 DebugInfoKind = codegenoptions::DebugDirectivesOnly;
809 break;
810 case EmitSameDebugInfoAsHost:
811 // Use same debug info level as the host.
812 break;
813 }
814 }
815
AddCudaIncludeArgs(const ArgList & DriverArgs,ArgStringList & CC1Args) const816 void CudaToolChain::AddCudaIncludeArgs(const ArgList &DriverArgs,
817 ArgStringList &CC1Args) const {
818 // Check our CUDA version if we're going to include the CUDA headers.
819 if (!DriverArgs.hasArg(options::OPT_nogpuinc) &&
820 !DriverArgs.hasArg(options::OPT_no_cuda_version_check)) {
821 StringRef Arch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
822 assert(!Arch.empty() && "Must have an explicit GPU arch.");
823 CudaInstallation.CheckCudaVersionSupportsArch(StringToCudaArch(Arch));
824 }
825 CudaInstallation.AddCudaIncludeArgs(DriverArgs, CC1Args);
826 }
827
828 llvm::opt::DerivedArgList *
TranslateArgs(const llvm::opt::DerivedArgList & Args,StringRef BoundArch,Action::OffloadKind DeviceOffloadKind) const829 CudaToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
830 StringRef BoundArch,
831 Action::OffloadKind DeviceOffloadKind) const {
832 DerivedArgList *DAL =
833 HostTC.TranslateArgs(Args, BoundArch, DeviceOffloadKind);
834 if (!DAL)
835 DAL = new DerivedArgList(Args.getBaseArgs());
836
837 const OptTable &Opts = getDriver().getOpts();
838
839 // For OpenMP device offloading, append derived arguments. Make sure
840 // flags are not duplicated.
841 // Also append the compute capability.
842 if (DeviceOffloadKind == Action::OFK_OpenMP) {
843 for (Arg *A : Args)
844 if (!llvm::is_contained(*DAL, A))
845 DAL->append(A);
846
847 if (!DAL->hasArg(options::OPT_march_EQ))
848 DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ),
849 !BoundArch.empty() ? BoundArch
850 : CLANG_OPENMP_NVPTX_DEFAULT_ARCH);
851
852 return DAL;
853 }
854
855 for (Arg *A : Args) {
856 DAL->append(A);
857 }
858
859 if (!BoundArch.empty()) {
860 DAL->eraseArg(options::OPT_march_EQ);
861 DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ), BoundArch);
862 }
863 return DAL;
864 }
865
buildAssembler() const866 Tool *CudaToolChain::buildAssembler() const {
867 return new tools::NVPTX::Assembler(*this);
868 }
869
buildLinker() const870 Tool *CudaToolChain::buildLinker() const {
871 if (OK == Action::OFK_OpenMP)
872 return new tools::NVPTX::OpenMPLinker(*this);
873 return new tools::NVPTX::Linker(*this);
874 }
875
addClangWarningOptions(ArgStringList & CC1Args) const876 void CudaToolChain::addClangWarningOptions(ArgStringList &CC1Args) const {
877 HostTC.addClangWarningOptions(CC1Args);
878 }
879
880 ToolChain::CXXStdlibType
GetCXXStdlibType(const ArgList & Args) const881 CudaToolChain::GetCXXStdlibType(const ArgList &Args) const {
882 return HostTC.GetCXXStdlibType(Args);
883 }
884
AddClangSystemIncludeArgs(const ArgList & DriverArgs,ArgStringList & CC1Args) const885 void CudaToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
886 ArgStringList &CC1Args) const {
887 HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args);
888
889 if (!DriverArgs.hasArg(options::OPT_nogpuinc) && CudaInstallation.isValid())
890 CC1Args.append(
891 {"-internal-isystem",
892 DriverArgs.MakeArgString(CudaInstallation.getIncludePath())});
893 }
894
AddClangCXXStdlibIncludeArgs(const ArgList & Args,ArgStringList & CC1Args) const895 void CudaToolChain::AddClangCXXStdlibIncludeArgs(const ArgList &Args,
896 ArgStringList &CC1Args) const {
897 HostTC.AddClangCXXStdlibIncludeArgs(Args, CC1Args);
898 }
899
AddIAMCUIncludeArgs(const ArgList & Args,ArgStringList & CC1Args) const900 void CudaToolChain::AddIAMCUIncludeArgs(const ArgList &Args,
901 ArgStringList &CC1Args) const {
902 HostTC.AddIAMCUIncludeArgs(Args, CC1Args);
903 }
904
getSupportedSanitizers() const905 SanitizerMask CudaToolChain::getSupportedSanitizers() const {
906 // The CudaToolChain only supports sanitizers in the sense that it allows
907 // sanitizer arguments on the command line if they are supported by the host
908 // toolchain. The CudaToolChain will actually ignore any command line
909 // arguments for any of these "supported" sanitizers. That means that no
910 // sanitization of device code is actually supported at this time.
911 //
912 // This behavior is necessary because the host and device toolchains
913 // invocations often share the command line, so the device toolchain must
914 // tolerate flags meant only for the host toolchain.
915 return HostTC.getSupportedSanitizers();
916 }
917
computeMSVCVersion(const Driver * D,const ArgList & Args) const918 VersionTuple CudaToolChain::computeMSVCVersion(const Driver *D,
919 const ArgList &Args) const {
920 return HostTC.computeMSVCVersion(D, Args);
921 }
922