1 //===--- Cuda.cpp - Cuda Tool and ToolChain Implementations -----*- C++ -*-===//
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
7 //===----------------------------------------------------------------------===//
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>
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
;
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
;
73 CudaVersion
parseCudaHFile(llvm::StringRef Input
) {
74 // Helper lambda which skips the words if the line starts with them or returns
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
))
87 Input
= Input
.ltrim();
88 while (!Input
.empty()) {
90 StartsWithWords(Input
.ltrim(), {"#", "define", "CUDA_VERSION"})) {
92 Line
->consumeInteger(10, RawVersion
);
93 return getCudaVersion(RawVersion
);
95 // Find next non-empty line.
96 Input
= Input
.drop_front(Input
.find_first_of("\n\r")).ltrim();
98 return CudaVersion::UNKNOWN
;
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
)
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
);
116 CudaInstallationDetector::CudaInstallationDetector(
117 const Driver
&D
, const llvm::Triple
&HostTriple
,
118 const llvm::opt::ArgList
&Args
)
124 Candidate(std::string Path
, bool StrictChecking
= false)
125 : Path(Path
), StrictChecking(StrictChecking
) {}
127 SmallVector
<Candidate
, 4> Candidates
;
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();
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" +
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
);
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);
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
);
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");
175 bool NoCudaLib
= Args
.hasArg(options::OPT_nogpulib
);
177 for (const auto &Candidate
: Candidates
) {
178 InstallPath
= Candidate
.Path
;
179 if (InstallPath
.empty() || !FS
.exists(InstallPath
))
182 BinPath
= InstallPath
+ "/bin";
183 IncludePath
= InstallPath
+ "/include";
184 LibDevicePath
= InstallPath
+ "/nvvm/libdevice";
186 if (!(FS
.exists(IncludePath
) && FS
.exists(BinPath
)))
188 bool CheckLibDevice
= (!NoCudaLib
|| Candidate
.StrictChecking
);
189 if (CheckLibDevice
&& !FS
.exists(LibDevicePath
))
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.
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";
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")
213 : CudaVersion::CUDA_70
;
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
;
222 CudaArch GpuArch
= static_cast<CudaArch
>(Arch
);
223 if (!IsNVIDIAGpuArch(GpuArch
))
225 std::string
GpuArchName(CudaArchToString(GpuArch
));
226 LibDeviceMap
[GpuArchName
] = FilePath
;
231 for (llvm::vfs::directory_iterator LI
= FS
.dir_begin(LibDevicePath
, EC
),
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")))
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
);
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
);
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
)
284 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
));
296 if (DriverArgs
.hasArg(options::OPT_nogpuinc
))
300 D
.Diag(diag::err_drv_no_cuda_installation
);
304 CC1Args
.push_back("-include");
305 CC1Args
.push_back("__clang_cuda_runtime_wrapper.h");
308 void CudaInstallationDetector::CheckCudaVersionSupportsArch(
309 CudaArch Arch
) const {
310 if (Arch
== CudaArch::UNKNOWN
|| Version
== CudaVersion::UNKNOWN
||
311 ArchsWithBadVersion
[(int)Arch
])
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
);
325 void CudaInstallationDetector::print(raw_ostream
&OS
) const {
327 OS
<< "Found CUDA installation: " << InstallPath
<< ", version "
328 << CudaVersionToString(Version
) << "\n";
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
343 } // anonymous namespace
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).
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
,
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
;
367 return IsDebugEnabled
? EmitSameDebugInfoAsHost
: DebugDirectivesOnly
;
369 return willEmitRemarks(Args
) ? DebugDirectivesOnly
: DisableDebugInfo
;
372 void NVPTX::Assembler::ConstructJob(Compilation
&C
, const JobAction
&JA
,
373 const InputInfo
&Output
,
374 const InputInfoList
&Inputs
,
376 const char *LinkingOutput
) const {
378 static_cast<const toolchains::CudaToolChain
&>(getToolChain());
379 assert(TC
.getTriple().isNVPTX() && "Wrong platform");
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.");
389 GPUArchName
= JA
.getOffloadingArch();
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.");
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
);
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}.
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.
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
))
422 else if (A
->getOption().matches(options::OPT_O0
))
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())
434 CmdArgs
.push_back(Args
.MakeArgString(llvm::Twine("-O") + OOpt
));
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");
440 if (DIKind
== DebugDirectivesOnly
)
441 CmdArgs
.push_back("-lineinfo");
443 // Pass -v to ptxas if it was passed to the driver.
444 if (Args
.hasArg(options::OPT_v
))
445 CmdArgs
.push_back("-v");
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()));
457 for (const auto& A
: Args
.getAllArgValues(options::OPT_Xcuda_ptxas
))
458 CmdArgs
.push_back(Args
.MakeArgString(A
));
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
,
466 else if (JA
.isOffloading(Action::OFK_Cuda
))
467 Relocatable
= Args
.hasFlag(options::OPT_fgpu_rdc
,
468 options::OPT_fno_gpu_rdc
, /*Default=*/false);
471 CmdArgs
.push_back("-c");
474 if (Arg
*A
= Args
.getLastArg(options::OPT_ptxas_path_EQ
))
475 Exec
= A
->getValue();
477 Exec
= Args
.MakeArgString(TC
.GetProgramPath("ptxas"));
478 C
.addCommand(std::make_unique
<Command
>(
480 ResponseFileSupport
{ResponseFileSupport::RF_Full
, llvm::sys::WEM_UTF8
,
482 Exec
, CmdArgs
, Inputs
, Output
));
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
)))
492 const StringRef ArchStr
= A
->getValue();
493 if (ArchStr
== "all" || ArchStr
== gpu_arch
) {
494 includePTX
= A
->getOption().matches(options::OPT_cuda_include_ptx_EQ
);
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
504 void NVPTX::Linker::ConstructJob(Compilation
&C
, const JobAction
&JA
,
505 const InputInfo
&Output
,
506 const InputInfoList
&Inputs
,
508 const char *LinkingOutput
) const {
510 static_cast<const toolchains::CudaToolChain
&>(getToolChain());
511 assert(TC
.getTriple().isNVPTX() && "Wrong platform");
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");
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
);
531 if (II
.getType() == types::TY_PP_Asm
&&
532 !shouldIncludePTX(Args
, gpu_arch_str
))
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
)
540 Args
.MakeArgString(llvm::Twine("--image=profile=") + Arch
+
541 ",file=" + getToolChain().getInputFilename(II
)));
544 for (const auto& A
: Args
.getAllArgValues(options::OPT_Xcuda_fatbinary
))
545 CmdArgs
.push_back(Args
.MakeArgString(A
));
547 const char *Exec
= Args
.MakeArgString(TC
.GetProgramPath("fatbinary"));
548 C
.addCommand(std::make_unique
<Command
>(
550 ResponseFileSupport
{ResponseFileSupport::RF_Full
, llvm::sys::WEM_UTF8
,
552 Exec
, CmdArgs
, Inputs
, Output
));
555 void NVPTX::getNVPTXTargetFeatures(const Driver
&D
, const llvm::Triple
&Triple
,
556 const llvm::opt::ArgList
&Args
,
557 std::vector
<StringRef
> &Features
) {
558 if (Args
.hasArg(options::OPT_cuda_feature_EQ
)) {
559 StringRef PtxFeature
=
560 Args
.getLastArgValue(options::OPT_cuda_feature_EQ
, "+ptx42");
561 Features
.push_back(Args
.MakeArgString(PtxFeature
));
564 CudaInstallationDetector
CudaInstallation(D
, Triple
, Args
);
566 // New CUDA versions often introduce new instructions that are only supported
567 // by new PTX version, so we need to raise PTX level to enable them in NVPTX
569 const char *PtxFeature
= nullptr;
570 switch (CudaInstallation
.version()) {
571 #define CASE_CUDA_VERSION(CUDA_VER, PTX_VER) \
572 case CudaVersion::CUDA_##CUDA_VER: \
573 PtxFeature = "+ptx" #PTX_VER; \
575 CASE_CUDA_VERSION(115, 75);
576 CASE_CUDA_VERSION(114, 74);
577 CASE_CUDA_VERSION(113, 73);
578 CASE_CUDA_VERSION(112, 72);
579 CASE_CUDA_VERSION(111, 71);
580 CASE_CUDA_VERSION(110, 70);
581 CASE_CUDA_VERSION(102, 65);
582 CASE_CUDA_VERSION(101, 64);
583 CASE_CUDA_VERSION(100, 63);
584 CASE_CUDA_VERSION(92, 61);
585 CASE_CUDA_VERSION(91, 61);
586 CASE_CUDA_VERSION(90, 60);
587 #undef CASE_CUDA_VERSION
589 PtxFeature
= "+ptx42";
591 Features
.push_back(PtxFeature
);
594 /// CUDA toolchain. Our assembler is ptxas, and our "linker" is fatbinary,
595 /// which isn't properly a linker but nonetheless performs the step of stitching
596 /// together object files from the assembler into a single blob.
598 CudaToolChain::CudaToolChain(const Driver
&D
, const llvm::Triple
&Triple
,
599 const ToolChain
&HostTC
, const ArgList
&Args
)
600 : ToolChain(D
, Triple
, Args
), HostTC(HostTC
),
601 CudaInstallation(D
, HostTC
.getTriple(), Args
) {
602 if (CudaInstallation
.isValid()) {
603 CudaInstallation
.WarnIfUnsupportedVersion();
604 getProgramPaths().push_back(std::string(CudaInstallation
.getBinPath()));
606 // Lookup binaries into the driver directory, this is used to
607 // discover the clang-offload-bundler executable.
608 getProgramPaths().push_back(getDriver().Dir
);
611 std::string
CudaToolChain::getInputFilename(const InputInfo
&Input
) const {
612 // Only object files are changed, for example assembly files keep their .s
613 // extensions. If the user requested device-only compilation don't change it.
614 if (Input
.getType() != types::TY_Object
|| getDriver().offloadDeviceOnly())
615 return ToolChain::getInputFilename(Input
);
617 // Replace extension for object files with cubin because nvlink relies on
618 // these particular file names.
619 SmallString
<256> Filename(ToolChain::getInputFilename(Input
));
620 llvm::sys::path::replace_extension(Filename
, "cubin");
621 return std::string(Filename
.str());
624 void CudaToolChain::addClangTargetOptions(
625 const llvm::opt::ArgList
&DriverArgs
,
626 llvm::opt::ArgStringList
&CC1Args
,
627 Action::OffloadKind DeviceOffloadingKind
) const {
628 HostTC
.addClangTargetOptions(DriverArgs
, CC1Args
, DeviceOffloadingKind
);
630 StringRef GpuArch
= DriverArgs
.getLastArgValue(options::OPT_march_EQ
);
631 assert(!GpuArch
.empty() && "Must have an explicit GPU arch.");
632 assert((DeviceOffloadingKind
== Action::OFK_OpenMP
||
633 DeviceOffloadingKind
== Action::OFK_Cuda
) &&
634 "Only OpenMP or CUDA offloading kinds are supported for NVIDIA GPUs.");
636 if (DeviceOffloadingKind
== Action::OFK_Cuda
) {
638 {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});
640 if (DriverArgs
.hasFlag(options::OPT_fcuda_approx_transcendentals
,
641 options::OPT_fno_cuda_approx_transcendentals
, false))
642 CC1Args
.push_back("-fcuda-approx-transcendentals");
645 if (DriverArgs
.hasArg(options::OPT_nogpulib
))
648 if (DeviceOffloadingKind
== Action::OFK_OpenMP
&&
649 DriverArgs
.hasArg(options::OPT_S
))
652 std::string LibDeviceFile
= CudaInstallation
.getLibDeviceFile(GpuArch
);
653 if (LibDeviceFile
.empty()) {
654 getDriver().Diag(diag::err_drv_no_cuda_libdevice
) << GpuArch
;
658 CC1Args
.push_back("-mlink-builtin-bitcode");
659 CC1Args
.push_back(DriverArgs
.MakeArgString(LibDeviceFile
));
661 clang::CudaVersion CudaInstallationVersion
= CudaInstallation
.version();
663 if (DriverArgs
.hasFlag(options::OPT_fcuda_short_ptr
,
664 options::OPT_fno_cuda_short_ptr
, false))
665 CC1Args
.append({"-mllvm", "--nvptx-short-ptr"});
667 if (CudaInstallationVersion
>= CudaVersion::UNKNOWN
)
669 DriverArgs
.MakeArgString(Twine("-target-sdk-version=") +
670 CudaVersionToString(CudaInstallationVersion
)));
672 if (DeviceOffloadingKind
== Action::OFK_OpenMP
) {
673 if (CudaInstallationVersion
< CudaVersion::CUDA_92
) {
675 diag::err_drv_omp_offload_target_cuda_version_not_support
)
676 << CudaVersionToString(CudaInstallationVersion
);
680 // Link the bitcode library late if we're using device LTO.
681 if (getDriver().isUsingLTO(/* IsOffload */ true))
684 addOpenMPDeviceRTL(getDriver(), DriverArgs
, CC1Args
, GpuArch
.str(),
689 llvm::DenormalMode
CudaToolChain::getDefaultDenormalModeForType(
690 const llvm::opt::ArgList
&DriverArgs
, const JobAction
&JA
,
691 const llvm::fltSemantics
*FPType
) const {
692 if (JA
.getOffloadingDeviceKind() == Action::OFK_Cuda
) {
693 if (FPType
&& FPType
== &llvm::APFloat::IEEEsingle() &&
694 DriverArgs
.hasFlag(options::OPT_fgpu_flush_denormals_to_zero
,
695 options::OPT_fno_gpu_flush_denormals_to_zero
, false))
696 return llvm::DenormalMode::getPreserveSign();
699 assert(JA
.getOffloadingDeviceKind() != Action::OFK_Host
);
700 return llvm::DenormalMode::getIEEE();
703 bool CudaToolChain::supportsDebugInfoOption(const llvm::opt::Arg
*A
) const {
704 const Option
&O
= A
->getOption();
705 return (O
.matches(options::OPT_gN_Group
) &&
706 !O
.matches(options::OPT_gmodules
)) ||
707 O
.matches(options::OPT_g_Flag
) ||
708 O
.matches(options::OPT_ggdbN_Group
) || O
.matches(options::OPT_ggdb
) ||
709 O
.matches(options::OPT_gdwarf
) || O
.matches(options::OPT_gdwarf_2
) ||
710 O
.matches(options::OPT_gdwarf_3
) || O
.matches(options::OPT_gdwarf_4
) ||
711 O
.matches(options::OPT_gdwarf_5
) ||
712 O
.matches(options::OPT_gcolumn_info
);
715 void CudaToolChain::adjustDebugInfoKind(
716 codegenoptions::DebugInfoKind
&DebugInfoKind
, const ArgList
&Args
) const {
717 switch (mustEmitDebugInfo(Args
)) {
718 case DisableDebugInfo
:
719 DebugInfoKind
= codegenoptions::NoDebugInfo
;
721 case DebugDirectivesOnly
:
722 DebugInfoKind
= codegenoptions::DebugDirectivesOnly
;
724 case EmitSameDebugInfoAsHost
:
725 // Use same debug info level as the host.
730 void CudaToolChain::AddCudaIncludeArgs(const ArgList
&DriverArgs
,
731 ArgStringList
&CC1Args
) const {
732 // Check our CUDA version if we're going to include the CUDA headers.
733 if (!DriverArgs
.hasArg(options::OPT_nogpuinc
) &&
734 !DriverArgs
.hasArg(options::OPT_no_cuda_version_check
)) {
735 StringRef Arch
= DriverArgs
.getLastArgValue(options::OPT_march_EQ
);
736 assert(!Arch
.empty() && "Must have an explicit GPU arch.");
737 CudaInstallation
.CheckCudaVersionSupportsArch(StringToCudaArch(Arch
));
739 CudaInstallation
.AddCudaIncludeArgs(DriverArgs
, CC1Args
);
742 llvm::opt::DerivedArgList
*
743 CudaToolChain::TranslateArgs(const llvm::opt::DerivedArgList
&Args
,
745 Action::OffloadKind DeviceOffloadKind
) const {
746 DerivedArgList
*DAL
=
747 HostTC
.TranslateArgs(Args
, BoundArch
, DeviceOffloadKind
);
749 DAL
= new DerivedArgList(Args
.getBaseArgs());
751 const OptTable
&Opts
= getDriver().getOpts();
753 // For OpenMP device offloading, append derived arguments. Make sure
754 // flags are not duplicated.
755 // Also append the compute capability.
756 if (DeviceOffloadKind
== Action::OFK_OpenMP
) {
758 if (!llvm::is_contained(*DAL
, A
))
761 if (!DAL
->hasArg(options::OPT_march_EQ
))
762 DAL
->AddJoinedArg(nullptr, Opts
.getOption(options::OPT_march_EQ
),
763 !BoundArch
.empty() ? BoundArch
764 : CLANG_OPENMP_NVPTX_DEFAULT_ARCH
);
769 for (Arg
*A
: Args
) {
773 if (!BoundArch
.empty()) {
774 DAL
->eraseArg(options::OPT_march_EQ
);
775 DAL
->AddJoinedArg(nullptr, Opts
.getOption(options::OPT_march_EQ
), BoundArch
);
780 Tool
*CudaToolChain::buildAssembler() const {
781 return new tools::NVPTX::Assembler(*this);
784 Tool
*CudaToolChain::buildLinker() const {
785 return new tools::NVPTX::Linker(*this);
788 void CudaToolChain::addClangWarningOptions(ArgStringList
&CC1Args
) const {
789 HostTC
.addClangWarningOptions(CC1Args
);
792 ToolChain::CXXStdlibType
793 CudaToolChain::GetCXXStdlibType(const ArgList
&Args
) const {
794 return HostTC
.GetCXXStdlibType(Args
);
797 void CudaToolChain::AddClangSystemIncludeArgs(const ArgList
&DriverArgs
,
798 ArgStringList
&CC1Args
) const {
799 HostTC
.AddClangSystemIncludeArgs(DriverArgs
, CC1Args
);
801 if (!DriverArgs
.hasArg(options::OPT_nogpuinc
) && CudaInstallation
.isValid())
803 {"-internal-isystem",
804 DriverArgs
.MakeArgString(CudaInstallation
.getIncludePath())});
807 void CudaToolChain::AddClangCXXStdlibIncludeArgs(const ArgList
&Args
,
808 ArgStringList
&CC1Args
) const {
809 HostTC
.AddClangCXXStdlibIncludeArgs(Args
, CC1Args
);
812 void CudaToolChain::AddIAMCUIncludeArgs(const ArgList
&Args
,
813 ArgStringList
&CC1Args
) const {
814 HostTC
.AddIAMCUIncludeArgs(Args
, CC1Args
);
817 SanitizerMask
CudaToolChain::getSupportedSanitizers() const {
818 // The CudaToolChain only supports sanitizers in the sense that it allows
819 // sanitizer arguments on the command line if they are supported by the host
820 // toolchain. The CudaToolChain will actually ignore any command line
821 // arguments for any of these "supported" sanitizers. That means that no
822 // sanitization of device code is actually supported at this time.
824 // This behavior is necessary because the host and device toolchains
825 // invocations often share the command line, so the device toolchain must
826 // tolerate flags meant only for the host toolchain.
827 return HostTC
.getSupportedSanitizers();
830 VersionTuple
CudaToolChain::computeMSVCVersion(const Driver
*D
,
831 const ArgList
&Args
) const {
832 return HostTC
.computeMSVCVersion(D
, Args
);