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/StringExtras.h" |
20 | #include "llvm/Option/ArgList.h" |
21 | #include "llvm/Support/FileSystem.h" |
22 | #include "llvm/Support/FormatAdapters.h" |
23 | #include "llvm/Support/FormatVariadic.h" |
24 | #include "llvm/Support/Path.h" |
25 | #include "llvm/Support/Process.h" |
26 | #include "llvm/Support/Program.h" |
27 | #include "llvm/Support/VirtualFileSystem.h" |
28 | #include "llvm/TargetParser/Host.h" |
29 | #include "llvm/TargetParser/TargetParser.h" |
30 | #include <system_error> |
31 | |
32 | using namespace clang::driver; |
33 | using namespace clang::driver::toolchains; |
34 | using namespace clang::driver::tools; |
35 | using namespace clang; |
36 | using namespace llvm::opt; |
37 | |
38 | namespace { |
39 | |
40 | CudaVersion getCudaVersion(uint32_t raw_version) { |
41 | if (raw_version < 7050) |
42 | return CudaVersion::CUDA_70; |
43 | if (raw_version < 8000) |
44 | return CudaVersion::CUDA_75; |
45 | if (raw_version < 9000) |
46 | return CudaVersion::CUDA_80; |
47 | if (raw_version < 9010) |
48 | return CudaVersion::CUDA_90; |
49 | if (raw_version < 9020) |
50 | return CudaVersion::CUDA_91; |
51 | if (raw_version < 10000) |
52 | return CudaVersion::CUDA_92; |
53 | if (raw_version < 10010) |
54 | return CudaVersion::CUDA_100; |
55 | if (raw_version < 10020) |
56 | return CudaVersion::CUDA_101; |
57 | if (raw_version < 11000) |
58 | return CudaVersion::CUDA_102; |
59 | if (raw_version < 11010) |
60 | return CudaVersion::CUDA_110; |
61 | if (raw_version < 11020) |
62 | return CudaVersion::CUDA_111; |
63 | if (raw_version < 11030) |
64 | return CudaVersion::CUDA_112; |
65 | if (raw_version < 11040) |
66 | return CudaVersion::CUDA_113; |
67 | if (raw_version < 11050) |
68 | return CudaVersion::CUDA_114; |
69 | if (raw_version < 11060) |
70 | return CudaVersion::CUDA_115; |
71 | if (raw_version < 11070) |
72 | return CudaVersion::CUDA_116; |
73 | if (raw_version < 11080) |
74 | return CudaVersion::CUDA_117; |
75 | if (raw_version < 11090) |
76 | return CudaVersion::CUDA_118; |
77 | if (raw_version < 12010) |
78 | return CudaVersion::CUDA_120; |
79 | if (raw_version < 12020) |
80 | return CudaVersion::CUDA_121; |
81 | if (raw_version < 12030) |
82 | return CudaVersion::CUDA_122; |
83 | if (raw_version < 12040) |
84 | return CudaVersion::CUDA_123; |
85 | if (raw_version < 12050) |
86 | return CudaVersion::CUDA_124; |
87 | if (raw_version < 12060) |
88 | return CudaVersion::CUDA_125; |
89 | return CudaVersion::NEW; |
90 | } |
91 | |
92 | CudaVersion parseCudaHFile(llvm::StringRef Input) { |
93 | // Helper lambda which skips the words if the line starts with them or returns |
94 | // std::nullopt otherwise. |
95 | auto StartsWithWords = |
96 | [](llvm::StringRef Line, |
97 | const SmallVector<StringRef, 3> words) -> std::optional<StringRef> { |
98 | for (StringRef word : words) { |
99 | if (!Line.consume_front(Prefix: word)) |
100 | return {}; |
101 | Line = Line.ltrim(); |
102 | } |
103 | return Line; |
104 | }; |
105 | |
106 | Input = Input.ltrim(); |
107 | while (!Input.empty()) { |
108 | if (auto Line = |
109 | StartsWithWords(Input.ltrim(), {"#" , "define" , "CUDA_VERSION" })) { |
110 | uint32_t RawVersion; |
111 | Line->consumeInteger(Radix: 10, Result&: RawVersion); |
112 | return getCudaVersion(raw_version: RawVersion); |
113 | } |
114 | // Find next non-empty line. |
115 | Input = Input.drop_front(N: Input.find_first_of(Chars: "\n\r" )).ltrim(); |
116 | } |
117 | return CudaVersion::UNKNOWN; |
118 | } |
119 | } // namespace |
120 | |
121 | void CudaInstallationDetector::WarnIfUnsupportedVersion() { |
122 | if (Version > CudaVersion::PARTIALLY_SUPPORTED) { |
123 | std::string VersionString = CudaVersionToString(V: Version); |
124 | if (!VersionString.empty()) |
125 | VersionString.insert(pos: 0, s: " " ); |
126 | D.Diag(DiagID: diag::warn_drv_new_cuda_version) |
127 | << VersionString |
128 | << (CudaVersion::PARTIALLY_SUPPORTED != CudaVersion::FULLY_SUPPORTED) |
129 | << CudaVersionToString(V: CudaVersion::PARTIALLY_SUPPORTED); |
130 | } else if (Version > CudaVersion::FULLY_SUPPORTED) |
131 | D.Diag(DiagID: diag::warn_drv_partially_supported_cuda_version) |
132 | << CudaVersionToString(V: Version); |
133 | } |
134 | |
135 | CudaInstallationDetector::CudaInstallationDetector( |
136 | const Driver &D, const llvm::Triple &HostTriple, |
137 | const llvm::opt::ArgList &Args) |
138 | : D(D) { |
139 | struct Candidate { |
140 | std::string Path; |
141 | bool StrictChecking; |
142 | |
143 | Candidate(std::string Path, bool StrictChecking = false) |
144 | : Path(Path), StrictChecking(StrictChecking) {} |
145 | }; |
146 | SmallVector<Candidate, 4> Candidates; |
147 | |
148 | // In decreasing order so we prefer newer versions to older versions. |
149 | std::initializer_list<const char *> Versions = {"8.0" , "7.5" , "7.0" }; |
150 | auto &FS = D.getVFS(); |
151 | |
152 | if (Args.hasArg(Ids: clang::driver::options::OPT_cuda_path_EQ)) { |
153 | Candidates.emplace_back( |
154 | Args: Args.getLastArgValue(Id: clang::driver::options::OPT_cuda_path_EQ).str()); |
155 | } else if (HostTriple.isOSWindows()) { |
156 | for (const char *Ver : Versions) |
157 | Candidates.emplace_back( |
158 | Args: D.SysRoot + "/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v" + |
159 | Ver); |
160 | } else { |
161 | if (!Args.hasArg(Ids: clang::driver::options::OPT_cuda_path_ignore_env)) { |
162 | // Try to find ptxas binary. If the executable is located in a directory |
163 | // called 'bin/', its parent directory might be a good guess for a valid |
164 | // CUDA installation. |
165 | // However, some distributions might installs 'ptxas' to /usr/bin. In that |
166 | // case the candidate would be '/usr' which passes the following checks |
167 | // because '/usr/include' exists as well. To avoid this case, we always |
168 | // check for the directory potentially containing files for libdevice, |
169 | // even if the user passes -nocudalib. |
170 | if (llvm::ErrorOr<std::string> ptxas = |
171 | llvm::sys::findProgramByName(Name: "ptxas" )) { |
172 | SmallString<256> ptxasAbsolutePath; |
173 | llvm::sys::fs::real_path(path: *ptxas, output&: ptxasAbsolutePath); |
174 | |
175 | StringRef ptxasDir = llvm::sys::path::parent_path(path: ptxasAbsolutePath); |
176 | if (llvm::sys::path::filename(path: ptxasDir) == "bin" ) |
177 | Candidates.emplace_back( |
178 | Args: std::string(llvm::sys::path::parent_path(path: ptxasDir)), |
179 | /*StrictChecking=*/Args: true); |
180 | } |
181 | } |
182 | |
183 | Candidates.emplace_back(Args: D.SysRoot + "/usr/local/cuda" ); |
184 | for (const char *Ver : Versions) |
185 | Candidates.emplace_back(Args: D.SysRoot + "/usr/local/cuda-" + Ver); |
186 | |
187 | Distro Dist(FS, llvm::Triple(llvm::sys::getProcessTriple())); |
188 | if (Dist.IsDebian() || Dist.IsUbuntu()) |
189 | // Special case for Debian to have nvidia-cuda-toolkit work |
190 | // out of the box. More info on http://bugs.debian.org/882505 |
191 | Candidates.emplace_back(Args: D.SysRoot + "/usr/lib/cuda" ); |
192 | } |
193 | |
194 | bool NoCudaLib = Args.hasArg(Ids: options::OPT_nogpulib); |
195 | |
196 | for (const auto &Candidate : Candidates) { |
197 | InstallPath = Candidate.Path; |
198 | if (InstallPath.empty() || !FS.exists(Path: InstallPath)) |
199 | continue; |
200 | |
201 | BinPath = InstallPath + "/bin" ; |
202 | IncludePath = InstallPath + "/include" ; |
203 | LibDevicePath = InstallPath + "/nvvm/libdevice" ; |
204 | |
205 | if (!(FS.exists(Path: IncludePath) && FS.exists(Path: BinPath))) |
206 | continue; |
207 | bool CheckLibDevice = (!NoCudaLib || Candidate.StrictChecking); |
208 | if (CheckLibDevice && !FS.exists(Path: LibDevicePath)) |
209 | continue; |
210 | |
211 | Version = CudaVersion::UNKNOWN; |
212 | if (auto CudaHFile = FS.getBufferForFile(Name: InstallPath + "/include/cuda.h" )) |
213 | Version = parseCudaHFile(Input: (*CudaHFile)->getBuffer()); |
214 | // As the last resort, make an educated guess between CUDA-7.0, which had |
215 | // old-style libdevice bitcode, and an unknown recent CUDA version. |
216 | if (Version == CudaVersion::UNKNOWN) { |
217 | Version = FS.exists(Path: LibDevicePath + "/libdevice.10.bc" ) |
218 | ? CudaVersion::NEW |
219 | : CudaVersion::CUDA_70; |
220 | } |
221 | |
222 | if (Version >= CudaVersion::CUDA_90) { |
223 | // CUDA-9+ uses single libdevice file for all GPU variants. |
224 | std::string FilePath = LibDevicePath + "/libdevice.10.bc" ; |
225 | if (FS.exists(Path: FilePath)) { |
226 | for (int Arch = (int)OffloadArch::SM_30, E = (int)OffloadArch::LAST; |
227 | Arch < E; ++Arch) { |
228 | OffloadArch OA = static_cast<OffloadArch>(Arch); |
229 | if (!IsNVIDIAOffloadArch(A: OA)) |
230 | continue; |
231 | std::string OffloadArchName(OffloadArchToString(A: OA)); |
232 | LibDeviceMap[OffloadArchName] = FilePath; |
233 | } |
234 | } |
235 | } else { |
236 | std::error_code EC; |
237 | for (llvm::vfs::directory_iterator LI = FS.dir_begin(Dir: LibDevicePath, EC), |
238 | LE; |
239 | !EC && LI != LE; LI = LI.increment(EC)) { |
240 | StringRef FilePath = LI->path(); |
241 | StringRef FileName = llvm::sys::path::filename(path: FilePath); |
242 | // Process all bitcode filenames that look like |
243 | // libdevice.compute_XX.YY.bc |
244 | const StringRef LibDeviceName = "libdevice." ; |
245 | if (!(FileName.starts_with(Prefix: LibDeviceName) && FileName.ends_with(Suffix: ".bc" ))) |
246 | continue; |
247 | StringRef GpuArch = FileName.slice( |
248 | Start: LibDeviceName.size(), End: FileName.find(C: '.', From: LibDeviceName.size())); |
249 | LibDeviceMap[GpuArch] = FilePath.str(); |
250 | // Insert map entries for specific devices with this compute |
251 | // capability. NVCC's choice of the libdevice library version is |
252 | // rather peculiar and depends on the CUDA version. |
253 | if (GpuArch == "compute_20" ) { |
254 | LibDeviceMap["sm_20" ] = std::string(FilePath); |
255 | LibDeviceMap["sm_21" ] = std::string(FilePath); |
256 | LibDeviceMap["sm_32" ] = std::string(FilePath); |
257 | } else if (GpuArch == "compute_30" ) { |
258 | LibDeviceMap["sm_30" ] = std::string(FilePath); |
259 | if (Version < CudaVersion::CUDA_80) { |
260 | LibDeviceMap["sm_50" ] = std::string(FilePath); |
261 | LibDeviceMap["sm_52" ] = std::string(FilePath); |
262 | LibDeviceMap["sm_53" ] = std::string(FilePath); |
263 | } |
264 | LibDeviceMap["sm_60" ] = std::string(FilePath); |
265 | LibDeviceMap["sm_61" ] = std::string(FilePath); |
266 | LibDeviceMap["sm_62" ] = std::string(FilePath); |
267 | } else if (GpuArch == "compute_35" ) { |
268 | LibDeviceMap["sm_35" ] = std::string(FilePath); |
269 | LibDeviceMap["sm_37" ] = std::string(FilePath); |
270 | } else if (GpuArch == "compute_50" ) { |
271 | if (Version >= CudaVersion::CUDA_80) { |
272 | LibDeviceMap["sm_50" ] = std::string(FilePath); |
273 | LibDeviceMap["sm_52" ] = std::string(FilePath); |
274 | LibDeviceMap["sm_53" ] = std::string(FilePath); |
275 | } |
276 | } |
277 | } |
278 | } |
279 | |
280 | // Check that we have found at least one libdevice that we can link in if |
281 | // -nocudalib hasn't been specified. |
282 | if (LibDeviceMap.empty() && !NoCudaLib) |
283 | continue; |
284 | |
285 | IsValid = true; |
286 | break; |
287 | } |
288 | } |
289 | |
290 | void CudaInstallationDetector::AddCudaIncludeArgs( |
291 | const ArgList &DriverArgs, ArgStringList &CC1Args) const { |
292 | if (!DriverArgs.hasArg(Ids: options::OPT_nobuiltininc)) { |
293 | // Add cuda_wrappers/* to our system include path. This lets us wrap |
294 | // standard library headers. |
295 | SmallString<128> P(D.ResourceDir); |
296 | llvm::sys::path::append(path&: P, a: "include" ); |
297 | llvm::sys::path::append(path&: P, a: "cuda_wrappers" ); |
298 | CC1Args.push_back(Elt: "-internal-isystem" ); |
299 | CC1Args.push_back(Elt: DriverArgs.MakeArgString(Str: P)); |
300 | } |
301 | |
302 | if (DriverArgs.hasArg(Ids: options::OPT_nogpuinc)) |
303 | return; |
304 | |
305 | if (!isValid()) { |
306 | D.Diag(DiagID: diag::err_drv_no_cuda_installation); |
307 | return; |
308 | } |
309 | |
310 | CC1Args.push_back(Elt: "-include" ); |
311 | CC1Args.push_back(Elt: "__clang_cuda_runtime_wrapper.h" ); |
312 | } |
313 | |
314 | void CudaInstallationDetector::CheckCudaVersionSupportsArch( |
315 | OffloadArch Arch) const { |
316 | if (Arch == OffloadArch::UNKNOWN || Version == CudaVersion::UNKNOWN || |
317 | ArchsWithBadVersion[(int)Arch]) |
318 | return; |
319 | |
320 | auto MinVersion = MinVersionForOffloadArch(A: Arch); |
321 | auto MaxVersion = MaxVersionForOffloadArch(A: Arch); |
322 | if (Version < MinVersion || Version > MaxVersion) { |
323 | ArchsWithBadVersion[(int)Arch] = true; |
324 | D.Diag(DiagID: diag::err_drv_cuda_version_unsupported) |
325 | << OffloadArchToString(A: Arch) << CudaVersionToString(V: MinVersion) |
326 | << CudaVersionToString(V: MaxVersion) << InstallPath |
327 | << CudaVersionToString(V: Version); |
328 | } |
329 | } |
330 | |
331 | void CudaInstallationDetector::print(raw_ostream &OS) const { |
332 | if (isValid()) |
333 | OS << "Found CUDA installation: " << InstallPath << ", version " |
334 | << CudaVersionToString(V: Version) << "\n" ; |
335 | } |
336 | |
337 | namespace { |
338 | /// Debug info level for the NVPTX devices. We may need to emit different debug |
339 | /// info level for the host and for the device itselfi. This type controls |
340 | /// emission of the debug info for the devices. It either prohibits disable info |
341 | /// emission completely, or emits debug directives only, or emits same debug |
342 | /// info as for the host. |
343 | enum DeviceDebugInfoLevel { |
344 | DisableDebugInfo, /// Do not emit debug info for the devices. |
345 | DebugDirectivesOnly, /// Emit only debug directives. |
346 | EmitSameDebugInfoAsHost, /// Use the same debug info level just like for the |
347 | /// host. |
348 | }; |
349 | } // anonymous namespace |
350 | |
351 | /// Define debug info level for the NVPTX devices. If the debug info for both |
352 | /// the host and device are disabled (-g0/-ggdb0 or no debug options at all). If |
353 | /// only debug directives are requested for the both host and device |
354 | /// (-gline-directvies-only), or the debug info only for the device is disabled |
355 | /// (optimization is on and --cuda-noopt-device-debug was not specified), the |
356 | /// debug directves only must be emitted for the device. Otherwise, use the same |
357 | /// debug info level just like for the host (with the limitations of only |
358 | /// supported DWARF2 standard). |
359 | static DeviceDebugInfoLevel mustEmitDebugInfo(const ArgList &Args) { |
360 | const Arg *A = Args.getLastArg(Ids: options::OPT_O_Group); |
361 | bool IsDebugEnabled = !A || A->getOption().matches(ID: options::OPT_O0) || |
362 | Args.hasFlag(Pos: options::OPT_cuda_noopt_device_debug, |
363 | Neg: options::OPT_no_cuda_noopt_device_debug, |
364 | /*Default=*/false); |
365 | if (const Arg *A = Args.getLastArg(Ids: options::OPT_g_Group)) { |
366 | const Option &Opt = A->getOption(); |
367 | if (Opt.matches(ID: options::OPT_gN_Group)) { |
368 | if (Opt.matches(ID: options::OPT_g0) || Opt.matches(ID: options::OPT_ggdb0)) |
369 | return DisableDebugInfo; |
370 | if (Opt.matches(ID: options::OPT_gline_directives_only)) |
371 | return DebugDirectivesOnly; |
372 | } |
373 | return IsDebugEnabled ? EmitSameDebugInfoAsHost : DebugDirectivesOnly; |
374 | } |
375 | return willEmitRemarks(Args) ? DebugDirectivesOnly : DisableDebugInfo; |
376 | } |
377 | |
378 | void NVPTX::Assembler::ConstructJob(Compilation &C, const JobAction &JA, |
379 | const InputInfo &Output, |
380 | const InputInfoList &Inputs, |
381 | const ArgList &Args, |
382 | const char *LinkingOutput) const { |
383 | const auto &TC = |
384 | static_cast<const toolchains::NVPTXToolChain &>(getToolChain()); |
385 | assert(TC.getTriple().isNVPTX() && "Wrong platform" ); |
386 | |
387 | StringRef GPUArchName; |
388 | // If this is a CUDA action we need to extract the device architecture |
389 | // from the Job's associated architecture, otherwise use the -march=arch |
390 | // option. This option may come from -Xopenmp-target flag or the default |
391 | // value. |
392 | if (JA.isDeviceOffloading(OKind: Action::OFK_Cuda)) { |
393 | GPUArchName = JA.getOffloadingArch(); |
394 | } else { |
395 | GPUArchName = Args.getLastArgValue(Id: options::OPT_march_EQ); |
396 | if (GPUArchName.empty()) { |
397 | C.getDriver().Diag(DiagID: diag::err_drv_offload_missing_gpu_arch) |
398 | << getToolChain().getArchName() << getShortName(); |
399 | return; |
400 | } |
401 | } |
402 | |
403 | // Obtain architecture from the action. |
404 | OffloadArch gpu_arch = StringToOffloadArch(S: GPUArchName); |
405 | assert(gpu_arch != OffloadArch::UNKNOWN && |
406 | "Device action expected to have an architecture." ); |
407 | |
408 | // Check that our installation's ptxas supports gpu_arch. |
409 | if (!Args.hasArg(Ids: options::OPT_no_cuda_version_check)) { |
410 | TC.CudaInstallation.CheckCudaVersionSupportsArch(Arch: gpu_arch); |
411 | } |
412 | |
413 | ArgStringList CmdArgs; |
414 | CmdArgs.push_back(Elt: TC.getTriple().isArch64Bit() ? "-m64" : "-m32" ); |
415 | DeviceDebugInfoLevel DIKind = mustEmitDebugInfo(Args); |
416 | if (DIKind == EmitSameDebugInfoAsHost) { |
417 | // ptxas does not accept -g option if optimization is enabled, so |
418 | // we ignore the compiler's -O* options if we want debug info. |
419 | CmdArgs.push_back(Elt: "-g" ); |
420 | CmdArgs.push_back(Elt: "--dont-merge-basicblocks" ); |
421 | CmdArgs.push_back(Elt: "--return-at-end" ); |
422 | } else if (Arg *A = Args.getLastArg(Ids: options::OPT_O_Group)) { |
423 | // Map the -O we received to -O{0,1,2,3}. |
424 | // |
425 | // TODO: Perhaps we should map host -O2 to ptxas -O3. -O3 is ptxas's |
426 | // default, so it may correspond more closely to the spirit of clang -O2. |
427 | |
428 | // -O3 seems like the least-bad option when -Osomething is specified to |
429 | // clang but it isn't handled below. |
430 | StringRef OOpt = "3" ; |
431 | if (A->getOption().matches(ID: options::OPT_O4) || |
432 | A->getOption().matches(ID: options::OPT_Ofast)) |
433 | OOpt = "3" ; |
434 | else if (A->getOption().matches(ID: options::OPT_O0)) |
435 | OOpt = "0" ; |
436 | else if (A->getOption().matches(ID: options::OPT_O)) { |
437 | // -Os, -Oz, and -O(anything else) map to -O2, for lack of better options. |
438 | OOpt = llvm::StringSwitch<const char *>(A->getValue()) |
439 | .Case(S: "1" , Value: "1" ) |
440 | .Case(S: "2" , Value: "2" ) |
441 | .Case(S: "3" , Value: "3" ) |
442 | .Case(S: "s" , Value: "2" ) |
443 | .Case(S: "z" , Value: "2" ) |
444 | .Default(Value: "2" ); |
445 | } |
446 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: llvm::Twine("-O" ) + OOpt)); |
447 | } else { |
448 | // If no -O was passed, pass -O0 to ptxas -- no opt flag should correspond |
449 | // to no optimizations, but ptxas's default is -O3. |
450 | CmdArgs.push_back(Elt: "-O0" ); |
451 | } |
452 | if (DIKind == DebugDirectivesOnly) |
453 | CmdArgs.push_back(Elt: "-lineinfo" ); |
454 | |
455 | // Pass -v to ptxas if it was passed to the driver. |
456 | if (Args.hasArg(Ids: options::OPT_v)) |
457 | CmdArgs.push_back(Elt: "-v" ); |
458 | |
459 | CmdArgs.push_back(Elt: "--gpu-name" ); |
460 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: OffloadArchToString(A: gpu_arch))); |
461 | CmdArgs.push_back(Elt: "--output-file" ); |
462 | std::string OutputFileName = TC.getInputFilename(Input: Output); |
463 | |
464 | if (Output.isFilename() && OutputFileName != Output.getFilename()) |
465 | C.addTempFile(Name: Args.MakeArgString(Str: OutputFileName)); |
466 | |
467 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: OutputFileName)); |
468 | for (const auto &II : Inputs) |
469 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: II.getFilename())); |
470 | |
471 | for (const auto &A : Args.getAllArgValues(Id: options::OPT_Xcuda_ptxas)) |
472 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: A)); |
473 | |
474 | bool Relocatable; |
475 | if (JA.isOffloading(OKind: Action::OFK_OpenMP)) |
476 | // In OpenMP we need to generate relocatable code. |
477 | Relocatable = Args.hasFlag(Pos: options::OPT_fopenmp_relocatable_target, |
478 | Neg: options::OPT_fnoopenmp_relocatable_target, |
479 | /*Default=*/true); |
480 | else if (JA.isOffloading(OKind: Action::OFK_Cuda)) |
481 | // In CUDA we generate relocatable code by default. |
482 | Relocatable = Args.hasFlag(Pos: options::OPT_fgpu_rdc, Neg: options::OPT_fno_gpu_rdc, |
483 | /*Default=*/false); |
484 | else |
485 | // Otherwise, we are compiling directly and should create linkable output. |
486 | Relocatable = true; |
487 | |
488 | if (Relocatable) |
489 | CmdArgs.push_back(Elt: "-c" ); |
490 | |
491 | const char *Exec; |
492 | if (Arg *A = Args.getLastArg(Ids: options::OPT_ptxas_path_EQ)) |
493 | Exec = A->getValue(); |
494 | else |
495 | Exec = Args.MakeArgString(Str: TC.GetProgramPath(Name: "ptxas" )); |
496 | C.addCommand(C: std::make_unique<Command>( |
497 | args: JA, args: *this, |
498 | args: ResponseFileSupport{.ResponseKind: ResponseFileSupport::RF_Full, .ResponseEncoding: llvm::sys::WEM_UTF8, |
499 | .ResponseFlag: "--options-file" }, |
500 | args&: Exec, args&: CmdArgs, args: Inputs, args: Output)); |
501 | } |
502 | |
503 | static bool shouldIncludePTX(const ArgList &Args, StringRef InputArch) { |
504 | // The new driver does not include PTX by default to avoid overhead. |
505 | bool includePTX = !Args.hasFlag(Pos: options::OPT_offload_new_driver, |
506 | Neg: options::OPT_no_offload_new_driver, Default: false); |
507 | for (Arg *A : Args.filtered(Ids: options::OPT_cuda_include_ptx_EQ, |
508 | Ids: options::OPT_no_cuda_include_ptx_EQ)) { |
509 | A->claim(); |
510 | const StringRef ArchStr = A->getValue(); |
511 | if (A->getOption().matches(ID: options::OPT_cuda_include_ptx_EQ) && |
512 | (ArchStr == "all" || ArchStr == InputArch)) |
513 | includePTX = true; |
514 | else if (A->getOption().matches(ID: options::OPT_no_cuda_include_ptx_EQ) && |
515 | (ArchStr == "all" || ArchStr == InputArch)) |
516 | includePTX = false; |
517 | } |
518 | return includePTX; |
519 | } |
520 | |
521 | // All inputs to this linker must be from CudaDeviceActions, as we need to look |
522 | // at the Inputs' Actions in order to figure out which GPU architecture they |
523 | // correspond to. |
524 | void NVPTX::FatBinary::ConstructJob(Compilation &C, const JobAction &JA, |
525 | const InputInfo &Output, |
526 | const InputInfoList &Inputs, |
527 | const ArgList &Args, |
528 | const char *LinkingOutput) const { |
529 | const auto &TC = |
530 | static_cast<const toolchains::CudaToolChain &>(getToolChain()); |
531 | assert(TC.getTriple().isNVPTX() && "Wrong platform" ); |
532 | |
533 | ArgStringList CmdArgs; |
534 | if (TC.CudaInstallation.version() <= CudaVersion::CUDA_100) |
535 | CmdArgs.push_back(Elt: "--cuda" ); |
536 | CmdArgs.push_back(Elt: TC.getTriple().isArch64Bit() ? "-64" : "-32" ); |
537 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: "--create" )); |
538 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: Output.getFilename())); |
539 | if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost) |
540 | CmdArgs.push_back(Elt: "-g" ); |
541 | |
542 | for (const auto &II : Inputs) { |
543 | auto *A = II.getAction(); |
544 | assert(A->getInputs().size() == 1 && |
545 | "Device offload action is expected to have a single input" ); |
546 | const char *gpu_arch_str = A->getOffloadingArch(); |
547 | assert(gpu_arch_str && |
548 | "Device action expected to have associated a GPU architecture!" ); |
549 | OffloadArch gpu_arch = StringToOffloadArch(S: gpu_arch_str); |
550 | |
551 | if (II.getType() == types::TY_PP_Asm && |
552 | !shouldIncludePTX(Args, InputArch: gpu_arch_str)) |
553 | continue; |
554 | // We need to pass an Arch of the form "sm_XX" for cubin files and |
555 | // "compute_XX" for ptx. |
556 | const char *Arch = (II.getType() == types::TY_PP_Asm) |
557 | ? OffloadArchToVirtualArchString(A: gpu_arch) |
558 | : gpu_arch_str; |
559 | CmdArgs.push_back( |
560 | Elt: Args.MakeArgString(Str: llvm::Twine("--image=profile=" ) + Arch + |
561 | ",file=" + getToolChain().getInputFilename(Input: II))); |
562 | } |
563 | |
564 | for (const auto &A : Args.getAllArgValues(Id: options::OPT_Xcuda_fatbinary)) |
565 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: A)); |
566 | |
567 | const char *Exec = Args.MakeArgString(Str: TC.GetProgramPath(Name: "fatbinary" )); |
568 | C.addCommand(C: std::make_unique<Command>( |
569 | args: JA, args: *this, |
570 | args: ResponseFileSupport{.ResponseKind: ResponseFileSupport::RF_Full, .ResponseEncoding: llvm::sys::WEM_UTF8, |
571 | .ResponseFlag: "--options-file" }, |
572 | args&: Exec, args&: CmdArgs, args: Inputs, args: Output)); |
573 | } |
574 | |
575 | void NVPTX::Linker::ConstructJob(Compilation &C, const JobAction &JA, |
576 | const InputInfo &Output, |
577 | const InputInfoList &Inputs, |
578 | const ArgList &Args, |
579 | const char *LinkingOutput) const { |
580 | const auto &TC = |
581 | static_cast<const toolchains::NVPTXToolChain &>(getToolChain()); |
582 | ArgStringList CmdArgs; |
583 | |
584 | assert(TC.getTriple().isNVPTX() && "Wrong platform" ); |
585 | |
586 | assert((Output.isFilename() || Output.isNothing()) && "Invalid output." ); |
587 | if (Output.isFilename()) { |
588 | CmdArgs.push_back(Elt: "-o" ); |
589 | CmdArgs.push_back(Elt: Output.getFilename()); |
590 | } |
591 | |
592 | if (mustEmitDebugInfo(Args) == EmitSameDebugInfoAsHost) |
593 | CmdArgs.push_back(Elt: "-g" ); |
594 | |
595 | if (Args.hasArg(Ids: options::OPT_v)) |
596 | CmdArgs.push_back(Elt: "-v" ); |
597 | |
598 | StringRef GPUArch = Args.getLastArgValue(Id: options::OPT_march_EQ); |
599 | if (GPUArch.empty()) { |
600 | C.getDriver().Diag(DiagID: diag::err_drv_offload_missing_gpu_arch) |
601 | << getToolChain().getArchName() << getShortName(); |
602 | return; |
603 | } |
604 | |
605 | CmdArgs.push_back(Elt: "-arch" ); |
606 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: GPUArch)); |
607 | |
608 | if (Args.hasArg(Ids: options::OPT_ptxas_path_EQ)) |
609 | CmdArgs.push_back(Elt: Args.MakeArgString( |
610 | Str: "--pxtas-path=" + Args.getLastArgValue(Id: options::OPT_ptxas_path_EQ))); |
611 | |
612 | if (Args.hasArg(Ids: options::OPT_cuda_path_EQ)) |
613 | CmdArgs.push_back(Elt: Args.MakeArgString( |
614 | Str: "--cuda-path=" + Args.getLastArgValue(Id: options::OPT_cuda_path_EQ))); |
615 | |
616 | // Add paths specified in LIBRARY_PATH environment variable as -L options. |
617 | addDirectoryList(Args, CmdArgs, ArgName: "-L" , EnvVar: "LIBRARY_PATH" ); |
618 | |
619 | // Add standard library search paths passed on the command line. |
620 | Args.AddAllArgs(Output&: CmdArgs, Id0: options::OPT_L); |
621 | getToolChain().AddFilePathLibArgs(Args, CmdArgs); |
622 | AddLinkerInputs(TC: getToolChain(), Inputs, Args, CmdArgs, JA); |
623 | |
624 | if (C.getDriver().isUsingLTO()) |
625 | addLTOOptions(ToolChain: getToolChain(), Args, CmdArgs, Output, Input: Inputs[0], |
626 | IsThinLTO: C.getDriver().getLTOMode() == LTOK_Thin); |
627 | |
628 | // Add paths for the default clang library path. |
629 | SmallString<256> DefaultLibPath = |
630 | llvm::sys::path::parent_path(path: TC.getDriver().Dir); |
631 | llvm::sys::path::append(path&: DefaultLibPath, CLANG_INSTALL_LIBDIR_BASENAME); |
632 | CmdArgs.push_back(Elt: Args.MakeArgString(Str: Twine("-L" ) + DefaultLibPath)); |
633 | |
634 | C.addCommand(C: std::make_unique<Command>( |
635 | args: JA, args: *this, |
636 | args: ResponseFileSupport{.ResponseKind: ResponseFileSupport::RF_Full, .ResponseEncoding: llvm::sys::WEM_UTF8, |
637 | .ResponseFlag: "--options-file" }, |
638 | args: Args.MakeArgString(Str: getToolChain().GetProgramPath(Name: "clang-nvlink-wrapper" )), |
639 | args&: CmdArgs, args: Inputs, args: Output)); |
640 | } |
641 | |
642 | void NVPTX::getNVPTXTargetFeatures(const Driver &D, const llvm::Triple &Triple, |
643 | const llvm::opt::ArgList &Args, |
644 | std::vector<StringRef> &Features) { |
645 | if (Args.hasArg(Ids: options::OPT_cuda_feature_EQ)) { |
646 | StringRef PtxFeature = |
647 | Args.getLastArgValue(Id: options::OPT_cuda_feature_EQ, Default: "+ptx42" ); |
648 | Features.push_back(x: Args.MakeArgString(Str: PtxFeature)); |
649 | return; |
650 | } |
651 | CudaInstallationDetector CudaInstallation(D, Triple, Args); |
652 | |
653 | // New CUDA versions often introduce new instructions that are only supported |
654 | // by new PTX version, so we need to raise PTX level to enable them in NVPTX |
655 | // back-end. |
656 | const char *PtxFeature = nullptr; |
657 | switch (CudaInstallation.version()) { |
658 | #define CASE_CUDA_VERSION(CUDA_VER, PTX_VER) \ |
659 | case CudaVersion::CUDA_##CUDA_VER: \ |
660 | PtxFeature = "+ptx" #PTX_VER; \ |
661 | break; |
662 | CASE_CUDA_VERSION(125, 85); |
663 | CASE_CUDA_VERSION(124, 84); |
664 | CASE_CUDA_VERSION(123, 83); |
665 | CASE_CUDA_VERSION(122, 82); |
666 | CASE_CUDA_VERSION(121, 81); |
667 | CASE_CUDA_VERSION(120, 80); |
668 | CASE_CUDA_VERSION(118, 78); |
669 | CASE_CUDA_VERSION(117, 77); |
670 | CASE_CUDA_VERSION(116, 76); |
671 | CASE_CUDA_VERSION(115, 75); |
672 | CASE_CUDA_VERSION(114, 74); |
673 | CASE_CUDA_VERSION(113, 73); |
674 | CASE_CUDA_VERSION(112, 72); |
675 | CASE_CUDA_VERSION(111, 71); |
676 | CASE_CUDA_VERSION(110, 70); |
677 | CASE_CUDA_VERSION(102, 65); |
678 | CASE_CUDA_VERSION(101, 64); |
679 | CASE_CUDA_VERSION(100, 63); |
680 | CASE_CUDA_VERSION(92, 61); |
681 | CASE_CUDA_VERSION(91, 61); |
682 | CASE_CUDA_VERSION(90, 60); |
683 | #undef CASE_CUDA_VERSION |
684 | default: |
685 | PtxFeature = "+ptx42" ; |
686 | } |
687 | Features.push_back(x: PtxFeature); |
688 | } |
689 | |
690 | /// NVPTX toolchain. Our assembler is ptxas, and our linker is nvlink. This |
691 | /// operates as a stand-alone version of the NVPTX tools without the host |
692 | /// toolchain. |
693 | NVPTXToolChain::NVPTXToolChain(const Driver &D, const llvm::Triple &Triple, |
694 | const llvm::Triple &HostTriple, |
695 | const ArgList &Args, bool Freestanding = false) |
696 | : ToolChain(D, Triple, Args), CudaInstallation(D, HostTriple, Args), |
697 | Freestanding(Freestanding) { |
698 | if (CudaInstallation.isValid()) |
699 | getProgramPaths().push_back(Elt: std::string(CudaInstallation.getBinPath())); |
700 | // Lookup binaries into the driver directory, this is used to |
701 | // discover the 'nvptx-arch' executable. |
702 | getProgramPaths().push_back(Elt: getDriver().Dir); |
703 | } |
704 | |
705 | /// We only need the host triple to locate the CUDA binary utilities, use the |
706 | /// system's default triple if not provided. |
707 | NVPTXToolChain::NVPTXToolChain(const Driver &D, const llvm::Triple &Triple, |
708 | const ArgList &Args) |
709 | : NVPTXToolChain(D, Triple, llvm::Triple(LLVM_HOST_TRIPLE), Args, |
710 | /*Freestanding=*/true) {} |
711 | |
712 | llvm::opt::DerivedArgList * |
713 | NVPTXToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args, |
714 | StringRef BoundArch, |
715 | Action::OffloadKind OffloadKind) const { |
716 | DerivedArgList *DAL = ToolChain::TranslateArgs(Args, BoundArch, DeviceOffloadKind: OffloadKind); |
717 | if (!DAL) |
718 | DAL = new DerivedArgList(Args.getBaseArgs()); |
719 | |
720 | const OptTable &Opts = getDriver().getOpts(); |
721 | |
722 | for (Arg *A : Args) |
723 | if (!llvm::is_contained(Range&: *DAL, Element: A)) |
724 | DAL->append(A); |
725 | |
726 | if (!DAL->hasArg(Ids: options::OPT_march_EQ) && OffloadKind != Action::OFK_None) { |
727 | DAL->AddJoinedArg(BaseArg: nullptr, Opt: Opts.getOption(Opt: options::OPT_march_EQ), |
728 | Value: OffloadArchToString(A: OffloadArch::CudaDefault)); |
729 | } else if (DAL->getLastArgValue(Id: options::OPT_march_EQ) == "generic" && |
730 | OffloadKind == Action::OFK_None) { |
731 | DAL->eraseArg(Id: options::OPT_march_EQ); |
732 | } else if (DAL->getLastArgValue(Id: options::OPT_march_EQ) == "native" ) { |
733 | auto GPUsOrErr = getSystemGPUArchs(Args); |
734 | if (!GPUsOrErr) { |
735 | getDriver().Diag(DiagID: diag::err_drv_undetermined_gpu_arch) |
736 | << getArchName() << llvm::toString(E: GPUsOrErr.takeError()) << "-march" ; |
737 | } else { |
738 | if (GPUsOrErr->size() > 1) |
739 | getDriver().Diag(DiagID: diag::warn_drv_multi_gpu_arch) |
740 | << getArchName() << llvm::join(R&: *GPUsOrErr, Separator: ", " ) << "-march" ; |
741 | DAL->AddJoinedArg(BaseArg: nullptr, Opt: Opts.getOption(Opt: options::OPT_march_EQ), |
742 | Value: Args.MakeArgString(Str: GPUsOrErr->front())); |
743 | } |
744 | } |
745 | |
746 | return DAL; |
747 | } |
748 | |
749 | void NVPTXToolChain::addClangTargetOptions( |
750 | const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args, |
751 | Action::OffloadKind DeviceOffloadingKind) const { |
752 | // If we are compiling with a standalone NVPTX toolchain we want to try to |
753 | // mimic a standard environment as much as possible. So we enable lowering |
754 | // ctor / dtor functions to global symbols that can be registered. |
755 | if (Freestanding) |
756 | CC1Args.append(IL: {"-mllvm" , "--nvptx-lower-global-ctor-dtor" }); |
757 | } |
758 | |
759 | bool NVPTXToolChain::supportsDebugInfoOption(const llvm::opt::Arg *A) const { |
760 | const Option &O = A->getOption(); |
761 | return (O.matches(ID: options::OPT_gN_Group) && |
762 | !O.matches(ID: options::OPT_gmodules)) || |
763 | O.matches(ID: options::OPT_g_Flag) || |
764 | O.matches(ID: options::OPT_ggdbN_Group) || O.matches(ID: options::OPT_ggdb) || |
765 | O.matches(ID: options::OPT_gdwarf) || O.matches(ID: options::OPT_gdwarf_2) || |
766 | O.matches(ID: options::OPT_gdwarf_3) || O.matches(ID: options::OPT_gdwarf_4) || |
767 | O.matches(ID: options::OPT_gdwarf_5) || |
768 | O.matches(ID: options::OPT_gcolumn_info); |
769 | } |
770 | |
771 | void NVPTXToolChain::adjustDebugInfoKind( |
772 | llvm::codegenoptions::DebugInfoKind &DebugInfoKind, |
773 | const ArgList &Args) const { |
774 | switch (mustEmitDebugInfo(Args)) { |
775 | case DisableDebugInfo: |
776 | DebugInfoKind = llvm::codegenoptions::NoDebugInfo; |
777 | break; |
778 | case DebugDirectivesOnly: |
779 | DebugInfoKind = llvm::codegenoptions::DebugDirectivesOnly; |
780 | break; |
781 | case EmitSameDebugInfoAsHost: |
782 | // Use same debug info level as the host. |
783 | break; |
784 | } |
785 | } |
786 | |
787 | Expected<SmallVector<std::string>> |
788 | NVPTXToolChain::getSystemGPUArchs(const ArgList &Args) const { |
789 | // Detect NVIDIA GPUs availible on the system. |
790 | std::string Program; |
791 | if (Arg *A = Args.getLastArg(Ids: options::OPT_nvptx_arch_tool_EQ)) |
792 | Program = A->getValue(); |
793 | else |
794 | Program = GetProgramPath(Name: "nvptx-arch" ); |
795 | |
796 | auto StdoutOrErr = executeToolChainProgram(Executable: Program, /*SecondsToWait=*/10); |
797 | if (!StdoutOrErr) |
798 | return StdoutOrErr.takeError(); |
799 | |
800 | SmallVector<std::string, 1> GPUArchs; |
801 | for (StringRef Arch : llvm::split(Str: (*StdoutOrErr)->getBuffer(), Separator: "\n" )) |
802 | if (!Arch.empty()) |
803 | GPUArchs.push_back(Elt: Arch.str()); |
804 | |
805 | if (GPUArchs.empty()) |
806 | return llvm::createStringError(EC: std::error_code(), |
807 | S: "No NVIDIA GPU detected in the system" ); |
808 | |
809 | return std::move(GPUArchs); |
810 | } |
811 | |
812 | /// CUDA toolchain. Our assembler is ptxas, and our "linker" is fatbinary, |
813 | /// which isn't properly a linker but nonetheless performs the step of stitching |
814 | /// together object files from the assembler into a single blob. |
815 | |
816 | CudaToolChain::CudaToolChain(const Driver &D, const llvm::Triple &Triple, |
817 | const ToolChain &HostTC, const ArgList &Args) |
818 | : NVPTXToolChain(D, Triple, HostTC.getTriple(), Args), HostTC(HostTC) {} |
819 | |
820 | void CudaToolChain::addClangTargetOptions( |
821 | const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args, |
822 | Action::OffloadKind DeviceOffloadingKind) const { |
823 | HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadKind: DeviceOffloadingKind); |
824 | |
825 | StringRef GpuArch = DriverArgs.getLastArgValue(Id: options::OPT_march_EQ); |
826 | assert(!GpuArch.empty() && "Must have an explicit GPU arch." ); |
827 | assert((DeviceOffloadingKind == Action::OFK_OpenMP || |
828 | DeviceOffloadingKind == Action::OFK_Cuda) && |
829 | "Only OpenMP or CUDA offloading kinds are supported for NVIDIA GPUs." ); |
830 | |
831 | if (DeviceOffloadingKind == Action::OFK_Cuda) { |
832 | CC1Args.append( |
833 | IL: {"-fcuda-is-device" , "-mllvm" , "-enable-memcpyopt-without-libcalls" }); |
834 | |
835 | // Unsized function arguments used for variadics were introduced in CUDA-9.0 |
836 | // We still do not support generating code that actually uses variadic |
837 | // arguments yet, but we do need to allow parsing them as recent CUDA |
838 | // headers rely on that. https://github.com/llvm/llvm-project/issues/58410 |
839 | if (CudaInstallation.version() >= CudaVersion::CUDA_90) |
840 | CC1Args.push_back(Elt: "-fcuda-allow-variadic-functions" ); |
841 | } |
842 | |
843 | if (DriverArgs.hasArg(Ids: options::OPT_nogpulib)) |
844 | return; |
845 | |
846 | if (DeviceOffloadingKind == Action::OFK_OpenMP && |
847 | DriverArgs.hasArg(Ids: options::OPT_S)) |
848 | return; |
849 | |
850 | std::string LibDeviceFile = CudaInstallation.getLibDeviceFile(Gpu: GpuArch); |
851 | if (LibDeviceFile.empty()) { |
852 | getDriver().Diag(DiagID: diag::err_drv_no_cuda_libdevice) << GpuArch; |
853 | return; |
854 | } |
855 | |
856 | CC1Args.push_back(Elt: "-mlink-builtin-bitcode" ); |
857 | CC1Args.push_back(Elt: DriverArgs.MakeArgString(Str: LibDeviceFile)); |
858 | |
859 | clang::CudaVersion CudaInstallationVersion = CudaInstallation.version(); |
860 | |
861 | if (DriverArgs.hasFlag(Pos: options::OPT_fcuda_short_ptr, |
862 | Neg: options::OPT_fno_cuda_short_ptr, Default: false)) |
863 | CC1Args.append(IL: {"-mllvm" , "--nvptx-short-ptr" }); |
864 | |
865 | if (CudaInstallationVersion >= CudaVersion::UNKNOWN) |
866 | CC1Args.push_back( |
867 | Elt: DriverArgs.MakeArgString(Str: Twine("-target-sdk-version=" ) + |
868 | CudaVersionToString(V: CudaInstallationVersion))); |
869 | |
870 | if (DeviceOffloadingKind == Action::OFK_OpenMP) { |
871 | if (CudaInstallationVersion < CudaVersion::CUDA_92) { |
872 | getDriver().Diag( |
873 | DiagID: diag::err_drv_omp_offload_target_cuda_version_not_support) |
874 | << CudaVersionToString(V: CudaInstallationVersion); |
875 | return; |
876 | } |
877 | |
878 | // Link the bitcode library late if we're using device LTO. |
879 | if (getDriver().isUsingLTO(/* IsOffload */ true)) |
880 | return; |
881 | |
882 | addOpenMPDeviceRTL(D: getDriver(), DriverArgs, CC1Args, BitcodeSuffix: GpuArch.str(), |
883 | Triple: getTriple(), HostTC); |
884 | } |
885 | } |
886 | |
887 | llvm::DenormalMode CudaToolChain::getDefaultDenormalModeForType( |
888 | const llvm::opt::ArgList &DriverArgs, const JobAction &JA, |
889 | const llvm::fltSemantics *FPType) const { |
890 | if (JA.getOffloadingDeviceKind() == Action::OFK_Cuda) { |
891 | if (FPType && FPType == &llvm::APFloat::IEEEsingle() && |
892 | DriverArgs.hasFlag(Pos: options::OPT_fgpu_flush_denormals_to_zero, |
893 | Neg: options::OPT_fno_gpu_flush_denormals_to_zero, Default: false)) |
894 | return llvm::DenormalMode::getPreserveSign(); |
895 | } |
896 | |
897 | assert(JA.getOffloadingDeviceKind() != Action::OFK_Host); |
898 | return llvm::DenormalMode::getIEEE(); |
899 | } |
900 | |
901 | void CudaToolChain::AddCudaIncludeArgs(const ArgList &DriverArgs, |
902 | ArgStringList &CC1Args) const { |
903 | // Check our CUDA version if we're going to include the CUDA headers. |
904 | if (!DriverArgs.hasArg(Ids: options::OPT_nogpuinc) && |
905 | !DriverArgs.hasArg(Ids: options::OPT_no_cuda_version_check)) { |
906 | StringRef Arch = DriverArgs.getLastArgValue(Id: options::OPT_march_EQ); |
907 | assert(!Arch.empty() && "Must have an explicit GPU arch." ); |
908 | CudaInstallation.CheckCudaVersionSupportsArch(Arch: StringToOffloadArch(S: Arch)); |
909 | } |
910 | CudaInstallation.AddCudaIncludeArgs(DriverArgs, CC1Args); |
911 | } |
912 | |
913 | std::string CudaToolChain::getInputFilename(const InputInfo &Input) const { |
914 | // Only object files are changed, for example assembly files keep their .s |
915 | // extensions. If the user requested device-only compilation don't change it. |
916 | if (Input.getType() != types::TY_Object || getDriver().offloadDeviceOnly()) |
917 | return ToolChain::getInputFilename(Input); |
918 | |
919 | return ToolChain::getInputFilename(Input); |
920 | } |
921 | |
922 | llvm::opt::DerivedArgList * |
923 | CudaToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args, |
924 | StringRef BoundArch, |
925 | Action::OffloadKind DeviceOffloadKind) const { |
926 | DerivedArgList *DAL = |
927 | HostTC.TranslateArgs(Args, BoundArch, DeviceOffloadKind); |
928 | if (!DAL) |
929 | DAL = new DerivedArgList(Args.getBaseArgs()); |
930 | |
931 | const OptTable &Opts = getDriver().getOpts(); |
932 | |
933 | // For OpenMP device offloading, append derived arguments. Make sure |
934 | // flags are not duplicated. |
935 | // Also append the compute capability. |
936 | if (DeviceOffloadKind == Action::OFK_OpenMP) { |
937 | for (Arg *A : Args) |
938 | if (!llvm::is_contained(Range&: *DAL, Element: A)) |
939 | DAL->append(A); |
940 | |
941 | if (!DAL->hasArg(Ids: options::OPT_march_EQ)) { |
942 | StringRef Arch = BoundArch; |
943 | if (Arch.empty()) { |
944 | auto ArchsOrErr = getSystemGPUArchs(Args); |
945 | if (!ArchsOrErr) { |
946 | std::string ErrMsg = |
947 | llvm::formatv(Fmt: "{0}" , Vals: llvm::fmt_consume(Item: ArchsOrErr.takeError())); |
948 | getDriver().Diag(DiagID: diag::err_drv_undetermined_gpu_arch) |
949 | << llvm::Triple::getArchTypeName(Kind: getArch()) << ErrMsg << "-march" ; |
950 | Arch = OffloadArchToString(A: OffloadArch::CudaDefault); |
951 | } else { |
952 | Arch = Args.MakeArgString(Str: ArchsOrErr->front()); |
953 | } |
954 | } |
955 | DAL->AddJoinedArg(BaseArg: nullptr, Opt: Opts.getOption(Opt: options::OPT_march_EQ), Value: Arch); |
956 | } |
957 | |
958 | return DAL; |
959 | } |
960 | |
961 | for (Arg *A : Args) { |
962 | // Make sure flags are not duplicated. |
963 | if (!llvm::is_contained(Range&: *DAL, Element: A)) { |
964 | DAL->append(A); |
965 | } |
966 | } |
967 | |
968 | if (!BoundArch.empty()) { |
969 | DAL->eraseArg(Id: options::OPT_march_EQ); |
970 | DAL->AddJoinedArg(BaseArg: nullptr, Opt: Opts.getOption(Opt: options::OPT_march_EQ), |
971 | Value: BoundArch); |
972 | } |
973 | return DAL; |
974 | } |
975 | |
976 | Tool *NVPTXToolChain::buildAssembler() const { |
977 | return new tools::NVPTX::Assembler(*this); |
978 | } |
979 | |
980 | Tool *NVPTXToolChain::buildLinker() const { |
981 | return new tools::NVPTX::Linker(*this); |
982 | } |
983 | |
984 | Tool *CudaToolChain::buildAssembler() const { |
985 | return new tools::NVPTX::Assembler(*this); |
986 | } |
987 | |
988 | Tool *CudaToolChain::buildLinker() const { |
989 | return new tools::NVPTX::FatBinary(*this); |
990 | } |
991 | |
992 | void CudaToolChain::addClangWarningOptions(ArgStringList &CC1Args) const { |
993 | HostTC.addClangWarningOptions(CC1Args); |
994 | } |
995 | |
996 | ToolChain::CXXStdlibType |
997 | CudaToolChain::GetCXXStdlibType(const ArgList &Args) const { |
998 | return HostTC.GetCXXStdlibType(Args); |
999 | } |
1000 | |
1001 | void CudaToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs, |
1002 | ArgStringList &CC1Args) const { |
1003 | HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args); |
1004 | |
1005 | if (!DriverArgs.hasArg(Ids: options::OPT_nogpuinc) && CudaInstallation.isValid()) |
1006 | CC1Args.append( |
1007 | IL: {"-internal-isystem" , |
1008 | DriverArgs.MakeArgString(Str: CudaInstallation.getIncludePath())}); |
1009 | } |
1010 | |
1011 | void CudaToolChain::AddClangCXXStdlibIncludeArgs(const ArgList &Args, |
1012 | ArgStringList &CC1Args) const { |
1013 | HostTC.AddClangCXXStdlibIncludeArgs(DriverArgs: Args, CC1Args); |
1014 | } |
1015 | |
1016 | void CudaToolChain::AddIAMCUIncludeArgs(const ArgList &Args, |
1017 | ArgStringList &CC1Args) const { |
1018 | HostTC.AddIAMCUIncludeArgs(DriverArgs: Args, CC1Args); |
1019 | } |
1020 | |
1021 | SanitizerMask CudaToolChain::getSupportedSanitizers() const { |
1022 | // The CudaToolChain only supports sanitizers in the sense that it allows |
1023 | // sanitizer arguments on the command line if they are supported by the host |
1024 | // toolchain. The CudaToolChain will actually ignore any command line |
1025 | // arguments for any of these "supported" sanitizers. That means that no |
1026 | // sanitization of device code is actually supported at this time. |
1027 | // |
1028 | // This behavior is necessary because the host and device toolchains |
1029 | // invocations often share the command line, so the device toolchain must |
1030 | // tolerate flags meant only for the host toolchain. |
1031 | return HostTC.getSupportedSanitizers(); |
1032 | } |
1033 | |
1034 | VersionTuple CudaToolChain::computeMSVCVersion(const Driver *D, |
1035 | const ArgList &Args) const { |
1036 | return HostTC.computeMSVCVersion(D, Args); |
1037 | } |
1038 | |