1 //===--- AMDGPU.cpp - AMDGPU 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 "AMDGPU.h"
10 #include "CommonArgs.h"
11 #include "InputInfo.h"
12 #include "clang/Basic/TargetID.h"
13 #include "clang/Driver/Compilation.h"
14 #include "clang/Driver/DriverDiagnostic.h"
15 #include "clang/Driver/Options.h"
16 #include "llvm/Option/ArgList.h"
17 #include "llvm/Support/Error.h"
18 #include "llvm/Support/FileUtilities.h"
19 #include "llvm/Support/LineIterator.h"
20 #include "llvm/Support/Path.h"
21 #include "llvm/Support/VirtualFileSystem.h"
22 #include <system_error>
23
24 #define AMDGPU_ARCH_PROGRAM_NAME "amdgpu-arch"
25
26 using namespace clang::driver;
27 using namespace clang::driver::tools;
28 using namespace clang::driver::toolchains;
29 using namespace clang;
30 using namespace llvm::opt;
31
32 // Look for sub-directory starts with PackageName under ROCm candidate path.
33 // If there is one and only one matching sub-directory found, append the
34 // sub-directory to Path. If there is no matching sub-directory or there are
35 // more than one matching sub-directories, diagnose them. Returns the full
36 // path of the package if there is only one matching sub-directory, otherwise
37 // returns an empty string.
38 llvm::SmallString<0>
findSPACKPackage(const Candidate & Cand,StringRef PackageName)39 RocmInstallationDetector::findSPACKPackage(const Candidate &Cand,
40 StringRef PackageName) {
41 if (!Cand.isSPACK())
42 return {};
43 std::error_code EC;
44 std::string Prefix = Twine(PackageName + "-" + Cand.SPACKReleaseStr).str();
45 llvm::SmallVector<llvm::SmallString<0>> SubDirs;
46 for (llvm::vfs::directory_iterator File = D.getVFS().dir_begin(Cand.Path, EC),
47 FileEnd;
48 File != FileEnd && !EC; File.increment(EC)) {
49 llvm::StringRef FileName = llvm::sys::path::filename(File->path());
50 if (FileName.startswith(Prefix)) {
51 SubDirs.push_back(FileName);
52 if (SubDirs.size() > 1)
53 break;
54 }
55 }
56 if (SubDirs.size() == 1) {
57 auto PackagePath = Cand.Path;
58 llvm::sys::path::append(PackagePath, SubDirs[0]);
59 return PackagePath;
60 }
61 if (SubDirs.size() == 0 && Verbose) {
62 llvm::errs() << "SPACK package " << Prefix << " not found at " << Cand.Path
63 << '\n';
64 return {};
65 }
66
67 if (SubDirs.size() > 1 && Verbose) {
68 llvm::errs() << "Cannot use SPACK package " << Prefix << " at " << Cand.Path
69 << " due to multiple installations for the same version\n";
70 }
71 return {};
72 }
73
scanLibDevicePath(llvm::StringRef Path)74 void RocmInstallationDetector::scanLibDevicePath(llvm::StringRef Path) {
75 assert(!Path.empty());
76
77 const StringRef Suffix(".bc");
78 const StringRef Suffix2(".amdgcn.bc");
79
80 std::error_code EC;
81 for (llvm::vfs::directory_iterator LI = D.getVFS().dir_begin(Path, EC), LE;
82 !EC && LI != LE; LI = LI.increment(EC)) {
83 StringRef FilePath = LI->path();
84 StringRef FileName = llvm::sys::path::filename(FilePath);
85 if (!FileName.endswith(Suffix))
86 continue;
87
88 StringRef BaseName;
89 if (FileName.endswith(Suffix2))
90 BaseName = FileName.drop_back(Suffix2.size());
91 else if (FileName.endswith(Suffix))
92 BaseName = FileName.drop_back(Suffix.size());
93
94 if (BaseName == "ocml") {
95 OCML = FilePath;
96 } else if (BaseName == "ockl") {
97 OCKL = FilePath;
98 } else if (BaseName == "opencl") {
99 OpenCL = FilePath;
100 } else if (BaseName == "hip") {
101 HIP = FilePath;
102 } else if (BaseName == "asanrtl") {
103 AsanRTL = FilePath;
104 } else if (BaseName == "oclc_finite_only_off") {
105 FiniteOnly.Off = FilePath;
106 } else if (BaseName == "oclc_finite_only_on") {
107 FiniteOnly.On = FilePath;
108 } else if (BaseName == "oclc_daz_opt_on") {
109 DenormalsAreZero.On = FilePath;
110 } else if (BaseName == "oclc_daz_opt_off") {
111 DenormalsAreZero.Off = FilePath;
112 } else if (BaseName == "oclc_correctly_rounded_sqrt_on") {
113 CorrectlyRoundedSqrt.On = FilePath;
114 } else if (BaseName == "oclc_correctly_rounded_sqrt_off") {
115 CorrectlyRoundedSqrt.Off = FilePath;
116 } else if (BaseName == "oclc_unsafe_math_on") {
117 UnsafeMath.On = FilePath;
118 } else if (BaseName == "oclc_unsafe_math_off") {
119 UnsafeMath.Off = FilePath;
120 } else if (BaseName == "oclc_wavefrontsize64_on") {
121 WavefrontSize64.On = FilePath;
122 } else if (BaseName == "oclc_wavefrontsize64_off") {
123 WavefrontSize64.Off = FilePath;
124 } else {
125 // Process all bitcode filenames that look like
126 // ocl_isa_version_XXX.amdgcn.bc
127 const StringRef DeviceLibPrefix = "oclc_isa_version_";
128 if (!BaseName.startswith(DeviceLibPrefix))
129 continue;
130
131 StringRef IsaVersionNumber =
132 BaseName.drop_front(DeviceLibPrefix.size());
133
134 llvm::Twine GfxName = Twine("gfx") + IsaVersionNumber;
135 SmallString<8> Tmp;
136 LibDeviceMap.insert(
137 std::make_pair(GfxName.toStringRef(Tmp), FilePath.str()));
138 }
139 }
140 }
141
142 // Parse and extract version numbers from `.hipVersion`. Return `true` if
143 // the parsing fails.
parseHIPVersionFile(llvm::StringRef V)144 bool RocmInstallationDetector::parseHIPVersionFile(llvm::StringRef V) {
145 SmallVector<StringRef, 4> VersionParts;
146 V.split(VersionParts, '\n');
147 unsigned Major = ~0U;
148 unsigned Minor = ~0U;
149 for (auto Part : VersionParts) {
150 auto Splits = Part.rtrim().split('=');
151 if (Splits.first == "HIP_VERSION_MAJOR") {
152 if (Splits.second.getAsInteger(0, Major))
153 return true;
154 } else if (Splits.first == "HIP_VERSION_MINOR") {
155 if (Splits.second.getAsInteger(0, Minor))
156 return true;
157 } else if (Splits.first == "HIP_VERSION_PATCH")
158 VersionPatch = Splits.second.str();
159 }
160 if (Major == ~0U || Minor == ~0U)
161 return true;
162 VersionMajorMinor = llvm::VersionTuple(Major, Minor);
163 DetectedVersion =
164 (Twine(Major) + "." + Twine(Minor) + "." + VersionPatch).str();
165 return false;
166 }
167
168 /// \returns a list of candidate directories for ROCm installation, which is
169 /// cached and populated only once.
170 const SmallVectorImpl<RocmInstallationDetector::Candidate> &
getInstallationPathCandidates()171 RocmInstallationDetector::getInstallationPathCandidates() {
172
173 // Return the cached candidate list if it has already been populated.
174 if (!ROCmSearchDirs.empty())
175 return ROCmSearchDirs;
176
177 auto DoPrintROCmSearchDirs = [&]() {
178 if (PrintROCmSearchDirs)
179 for (auto Cand : ROCmSearchDirs) {
180 llvm::errs() << "ROCm installation search path";
181 if (Cand.isSPACK())
182 llvm::errs() << " (Spack " << Cand.SPACKReleaseStr << ")";
183 llvm::errs() << ": " << Cand.Path << '\n';
184 }
185 };
186
187 // For candidate specified by --rocm-path we do not do strict check, i.e.,
188 // checking existence of HIP version file and device library files.
189 if (!RocmPathArg.empty()) {
190 ROCmSearchDirs.emplace_back(RocmPathArg.str());
191 DoPrintROCmSearchDirs();
192 return ROCmSearchDirs;
193 } else if (const char *RocmPathEnv = ::getenv("ROCM_PATH")) {
194 if (!StringRef(RocmPathEnv).empty()) {
195 ROCmSearchDirs.emplace_back(RocmPathEnv);
196 DoPrintROCmSearchDirs();
197 return ROCmSearchDirs;
198 }
199 }
200
201 // Try to find relative to the compiler binary.
202 const char *InstallDir = D.getInstalledDir();
203
204 // Check both a normal Unix prefix position of the clang binary, as well as
205 // the Windows-esque layout the ROCm packages use with the host architecture
206 // subdirectory of bin.
207 auto DeduceROCmPath = [](StringRef ClangPath) {
208 // Strip off directory (usually bin)
209 StringRef ParentDir = llvm::sys::path::parent_path(ClangPath);
210 StringRef ParentName = llvm::sys::path::filename(ParentDir);
211
212 // Some builds use bin/{host arch}, so go up again.
213 if (ParentName == "bin") {
214 ParentDir = llvm::sys::path::parent_path(ParentDir);
215 ParentName = llvm::sys::path::filename(ParentDir);
216 }
217
218 // Detect ROCm packages built with SPACK.
219 // clang is installed at
220 // <rocm_root>/llvm-amdgpu-<rocm_release_string>-<hash>/bin directory.
221 // We only consider the parent directory of llvm-amdgpu package as ROCm
222 // installation candidate for SPACK.
223 if (ParentName.startswith("llvm-amdgpu-")) {
224 auto SPACKPostfix =
225 ParentName.drop_front(strlen("llvm-amdgpu-")).split('-');
226 auto SPACKReleaseStr = SPACKPostfix.first;
227 if (!SPACKReleaseStr.empty()) {
228 ParentDir = llvm::sys::path::parent_path(ParentDir);
229 return Candidate(ParentDir.str(), /*StrictChecking=*/true,
230 SPACKReleaseStr);
231 }
232 }
233
234 // Some versions of the rocm llvm package install to /opt/rocm/llvm/bin
235 // Some versions of the aomp package install to /opt/rocm/aomp/bin
236 if (ParentName == "llvm" || ParentName.startswith("aomp"))
237 ParentDir = llvm::sys::path::parent_path(ParentDir);
238
239 return Candidate(ParentDir.str(), /*StrictChecking=*/true);
240 };
241
242 // Deduce ROCm path by the path used to invoke clang. Do not resolve symbolic
243 // link of clang itself.
244 ROCmSearchDirs.emplace_back(DeduceROCmPath(InstallDir));
245
246 // Deduce ROCm path by the real path of the invoked clang, resolving symbolic
247 // link of clang itself.
248 llvm::SmallString<256> RealClangPath;
249 llvm::sys::fs::real_path(D.getClangProgramPath(), RealClangPath);
250 auto ParentPath = llvm::sys::path::parent_path(RealClangPath);
251 if (ParentPath != InstallDir)
252 ROCmSearchDirs.emplace_back(DeduceROCmPath(ParentPath));
253
254 // Device library may be installed in clang resource directory.
255 ROCmSearchDirs.emplace_back(D.ResourceDir,
256 /*StrictChecking=*/true);
257
258 ROCmSearchDirs.emplace_back(D.SysRoot + "/opt/rocm",
259 /*StrictChecking=*/true);
260
261 // Find the latest /opt/rocm-{release} directory.
262 std::error_code EC;
263 std::string LatestROCm;
264 llvm::VersionTuple LatestVer;
265 // Get ROCm version from ROCm directory name.
266 auto GetROCmVersion = [](StringRef DirName) {
267 llvm::VersionTuple V;
268 std::string VerStr = DirName.drop_front(strlen("rocm-")).str();
269 // The ROCm directory name follows the format of
270 // rocm-{major}.{minor}.{subMinor}[-{build}]
271 std::replace(VerStr.begin(), VerStr.end(), '-', '.');
272 V.tryParse(VerStr);
273 return V;
274 };
275 for (llvm::vfs::directory_iterator
276 File = D.getVFS().dir_begin(D.SysRoot + "/opt", EC),
277 FileEnd;
278 File != FileEnd && !EC; File.increment(EC)) {
279 llvm::StringRef FileName = llvm::sys::path::filename(File->path());
280 if (!FileName.startswith("rocm-"))
281 continue;
282 if (LatestROCm.empty()) {
283 LatestROCm = FileName.str();
284 LatestVer = GetROCmVersion(LatestROCm);
285 continue;
286 }
287 auto Ver = GetROCmVersion(FileName);
288 if (LatestVer < Ver) {
289 LatestROCm = FileName.str();
290 LatestVer = Ver;
291 }
292 }
293 if (!LatestROCm.empty())
294 ROCmSearchDirs.emplace_back(D.SysRoot + "/opt/" + LatestROCm,
295 /*StrictChecking=*/true);
296
297 DoPrintROCmSearchDirs();
298 return ROCmSearchDirs;
299 }
300
RocmInstallationDetector(const Driver & D,const llvm::Triple & HostTriple,const llvm::opt::ArgList & Args,bool DetectHIPRuntime,bool DetectDeviceLib)301 RocmInstallationDetector::RocmInstallationDetector(
302 const Driver &D, const llvm::Triple &HostTriple,
303 const llvm::opt::ArgList &Args, bool DetectHIPRuntime, bool DetectDeviceLib)
304 : D(D) {
305 Verbose = Args.hasArg(options::OPT_v);
306 RocmPathArg = Args.getLastArgValue(clang::driver::options::OPT_rocm_path_EQ);
307 PrintROCmSearchDirs =
308 Args.hasArg(clang::driver::options::OPT_print_rocm_search_dirs);
309 RocmDeviceLibPathArg =
310 Args.getAllArgValues(clang::driver::options::OPT_rocm_device_lib_path_EQ);
311 HIPPathArg = Args.getLastArgValue(clang::driver::options::OPT_hip_path_EQ);
312 if (auto *A = Args.getLastArg(clang::driver::options::OPT_hip_version_EQ)) {
313 HIPVersionArg = A->getValue();
314 unsigned Major = 0;
315 unsigned Minor = 0;
316 SmallVector<StringRef, 3> Parts;
317 HIPVersionArg.split(Parts, '.');
318 if (Parts.size())
319 Parts[0].getAsInteger(0, Major);
320 if (Parts.size() > 1)
321 Parts[1].getAsInteger(0, Minor);
322 if (Parts.size() > 2)
323 VersionPatch = Parts[2].str();
324 if (VersionPatch.empty())
325 VersionPatch = "0";
326 if (Major == 0 || Minor == 0)
327 D.Diag(diag::err_drv_invalid_value)
328 << A->getAsString(Args) << HIPVersionArg;
329
330 VersionMajorMinor = llvm::VersionTuple(Major, Minor);
331 DetectedVersion =
332 (Twine(Major) + "." + Twine(Minor) + "." + VersionPatch).str();
333 } else {
334 VersionPatch = DefaultVersionPatch;
335 VersionMajorMinor =
336 llvm::VersionTuple(DefaultVersionMajor, DefaultVersionMinor);
337 DetectedVersion = (Twine(DefaultVersionMajor) + "." +
338 Twine(DefaultVersionMinor) + "." + VersionPatch)
339 .str();
340 }
341
342 if (DetectHIPRuntime)
343 detectHIPRuntime();
344 if (DetectDeviceLib)
345 detectDeviceLibrary();
346 }
347
detectDeviceLibrary()348 void RocmInstallationDetector::detectDeviceLibrary() {
349 assert(LibDevicePath.empty());
350
351 if (!RocmDeviceLibPathArg.empty())
352 LibDevicePath = RocmDeviceLibPathArg[RocmDeviceLibPathArg.size() - 1];
353 else if (const char *LibPathEnv = ::getenv("HIP_DEVICE_LIB_PATH"))
354 LibDevicePath = LibPathEnv;
355
356 auto &FS = D.getVFS();
357 if (!LibDevicePath.empty()) {
358 // Maintain compatability with HIP flag/envvar pointing directly at the
359 // bitcode library directory. This points directly at the library path instead
360 // of the rocm root installation.
361 if (!FS.exists(LibDevicePath))
362 return;
363
364 scanLibDevicePath(LibDevicePath);
365 HasDeviceLibrary = allGenericLibsValid() && !LibDeviceMap.empty();
366 return;
367 }
368
369 // The install path situation in old versions of ROCm is a real mess, and
370 // use a different install layout. Multiple copies of the device libraries
371 // exist for each frontend project, and differ depending on which build
372 // system produced the packages. Standalone OpenCL builds also have a
373 // different directory structure from the ROCm OpenCL package.
374 auto &ROCmDirs = getInstallationPathCandidates();
375 for (const auto &Candidate : ROCmDirs) {
376 auto CandidatePath = Candidate.Path;
377
378 // Check device library exists at the given path.
379 auto CheckDeviceLib = [&](StringRef Path) {
380 bool CheckLibDevice = (!NoBuiltinLibs || Candidate.StrictChecking);
381 if (CheckLibDevice && !FS.exists(Path))
382 return false;
383
384 scanLibDevicePath(Path);
385
386 if (!NoBuiltinLibs) {
387 // Check that the required non-target libraries are all available.
388 if (!allGenericLibsValid())
389 return false;
390
391 // Check that we have found at least one libdevice that we can link in
392 // if -nobuiltinlib hasn't been specified.
393 if (LibDeviceMap.empty())
394 return false;
395 }
396 return true;
397 };
398
399 // The possible structures are:
400 // - ${ROCM_ROOT}/amdgcn/bitcode/*
401 // - ${ROCM_ROOT}/lib/*
402 // - ${ROCM_ROOT}/lib/bitcode/*
403 // so try to detect these layouts.
404 static constexpr std::array<const char *, 2> SubDirsList[] = {
405 {"amdgcn", "bitcode"},
406 {"lib", ""},
407 {"lib", "bitcode"},
408 };
409
410 // Make a path by appending sub-directories to InstallPath.
411 auto MakePath = [&](const llvm::ArrayRef<const char *> &SubDirs) {
412 // Device library built by SPACK is installed to
413 // <rocm_root>/rocm-device-libs-<rocm_release_string>-<hash> directory.
414 auto SPACKPath = findSPACKPackage(Candidate, "rocm-device-libs");
415 auto Path = SPACKPath.empty() ? CandidatePath : SPACKPath;
416 for (auto SubDir : SubDirs)
417 llvm::sys::path::append(Path, SubDir);
418 return Path;
419 };
420
421 for (auto SubDirs : SubDirsList) {
422 LibDevicePath = MakePath(SubDirs);
423 HasDeviceLibrary = CheckDeviceLib(LibDevicePath);
424 if (HasDeviceLibrary)
425 return;
426 }
427 }
428 }
429
detectHIPRuntime()430 void RocmInstallationDetector::detectHIPRuntime() {
431 SmallVector<Candidate, 4> HIPSearchDirs;
432 if (!HIPPathArg.empty())
433 HIPSearchDirs.emplace_back(HIPPathArg.str(), /*StrictChecking=*/true);
434 else
435 HIPSearchDirs.append(getInstallationPathCandidates());
436 auto &FS = D.getVFS();
437
438 for (const auto &Candidate : HIPSearchDirs) {
439 InstallPath = Candidate.Path;
440 if (InstallPath.empty() || !FS.exists(InstallPath))
441 continue;
442 // HIP runtime built by SPACK is installed to
443 // <rocm_root>/hip-<rocm_release_string>-<hash> directory.
444 auto SPACKPath = findSPACKPackage(Candidate, "hip");
445 InstallPath = SPACKPath.empty() ? InstallPath : SPACKPath;
446
447 BinPath = InstallPath;
448 llvm::sys::path::append(BinPath, "bin");
449 IncludePath = InstallPath;
450 llvm::sys::path::append(IncludePath, "include");
451 LibPath = InstallPath;
452 llvm::sys::path::append(LibPath, "lib");
453
454 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> VersionFile =
455 FS.getBufferForFile(BinPath + "/.hipVersion");
456 if (!VersionFile && Candidate.StrictChecking)
457 continue;
458
459 if (HIPVersionArg.empty() && VersionFile)
460 if (parseHIPVersionFile((*VersionFile)->getBuffer()))
461 continue;
462
463 HasHIPRuntime = true;
464 return;
465 }
466 HasHIPRuntime = false;
467 }
468
print(raw_ostream & OS) const469 void RocmInstallationDetector::print(raw_ostream &OS) const {
470 if (hasHIPRuntime())
471 OS << "Found HIP installation: " << InstallPath << ", version "
472 << DetectedVersion << '\n';
473 }
474
AddHIPIncludeArgs(const ArgList & DriverArgs,ArgStringList & CC1Args) const475 void RocmInstallationDetector::AddHIPIncludeArgs(const ArgList &DriverArgs,
476 ArgStringList &CC1Args) const {
477 bool UsesRuntimeWrapper = VersionMajorMinor > llvm::VersionTuple(3, 5);
478
479 if (!DriverArgs.hasArg(options::OPT_nobuiltininc)) {
480 // HIP header includes standard library wrapper headers under clang
481 // cuda_wrappers directory. Since these wrapper headers include_next
482 // standard C++ headers, whereas libc++ headers include_next other clang
483 // headers. The include paths have to follow this order:
484 // - wrapper include path
485 // - standard C++ include path
486 // - other clang include path
487 // Since standard C++ and other clang include paths are added in other
488 // places after this function, here we only need to make sure wrapper
489 // include path is added.
490 //
491 // ROCm 3.5 does not fully support the wrapper headers. Therefore it needs
492 // a workaround.
493 SmallString<128> P(D.ResourceDir);
494 if (UsesRuntimeWrapper)
495 llvm::sys::path::append(P, "include", "cuda_wrappers");
496 CC1Args.push_back("-internal-isystem");
497 CC1Args.push_back(DriverArgs.MakeArgString(P));
498 }
499
500 if (DriverArgs.hasArg(options::OPT_nogpuinc))
501 return;
502
503 if (!hasHIPRuntime()) {
504 D.Diag(diag::err_drv_no_hip_runtime);
505 return;
506 }
507
508 CC1Args.push_back("-internal-isystem");
509 CC1Args.push_back(DriverArgs.MakeArgString(getIncludePath()));
510 if (UsesRuntimeWrapper)
511 CC1Args.append({"-include", "__clang_hip_runtime_wrapper.h"});
512 }
513
ConstructJob(Compilation & C,const JobAction & JA,const InputInfo & Output,const InputInfoList & Inputs,const ArgList & Args,const char * LinkingOutput) const514 void amdgpu::Linker::ConstructJob(Compilation &C, const JobAction &JA,
515 const InputInfo &Output,
516 const InputInfoList &Inputs,
517 const ArgList &Args,
518 const char *LinkingOutput) const {
519
520 std::string Linker = getToolChain().GetProgramPath(getShortName());
521 ArgStringList CmdArgs;
522 addLinkerCompressDebugSectionsOption(getToolChain(), Args, CmdArgs);
523 AddLinkerInputs(getToolChain(), Inputs, Args, CmdArgs, JA);
524 CmdArgs.push_back("-shared");
525 CmdArgs.push_back("-o");
526 CmdArgs.push_back(Output.getFilename());
527 C.addCommand(std::make_unique<Command>(
528 JA, *this, ResponseFileSupport::AtFileCurCP(), Args.MakeArgString(Linker),
529 CmdArgs, Inputs, Output));
530 }
531
getAMDGPUTargetFeatures(const Driver & D,const llvm::Triple & Triple,const llvm::opt::ArgList & Args,std::vector<StringRef> & Features)532 void amdgpu::getAMDGPUTargetFeatures(const Driver &D,
533 const llvm::Triple &Triple,
534 const llvm::opt::ArgList &Args,
535 std::vector<StringRef> &Features) {
536 // Add target ID features to -target-feature options. No diagnostics should
537 // be emitted here since invalid target ID is diagnosed at other places.
538 StringRef TargetID = Args.getLastArgValue(options::OPT_mcpu_EQ);
539 if (!TargetID.empty()) {
540 llvm::StringMap<bool> FeatureMap;
541 auto OptionalGpuArch = parseTargetID(Triple, TargetID, &FeatureMap);
542 if (OptionalGpuArch) {
543 StringRef GpuArch = OptionalGpuArch.getValue();
544 // Iterate through all possible target ID features for the given GPU.
545 // If it is mapped to true, add +feature.
546 // If it is mapped to false, add -feature.
547 // If it is not in the map (default), do not add it
548 for (auto &&Feature : getAllPossibleTargetIDFeatures(Triple, GpuArch)) {
549 auto Pos = FeatureMap.find(Feature);
550 if (Pos == FeatureMap.end())
551 continue;
552 Features.push_back(Args.MakeArgStringRef(
553 (Twine(Pos->second ? "+" : "-") + Feature).str()));
554 }
555 }
556 }
557
558 if (Args.hasFlag(options::OPT_mwavefrontsize64,
559 options::OPT_mno_wavefrontsize64, false))
560 Features.push_back("+wavefrontsize64");
561
562 handleTargetFeaturesGroup(
563 Args, Features, options::OPT_m_amdgpu_Features_Group);
564 }
565
566 /// AMDGPU Toolchain
AMDGPUToolChain(const Driver & D,const llvm::Triple & Triple,const ArgList & Args)567 AMDGPUToolChain::AMDGPUToolChain(const Driver &D, const llvm::Triple &Triple,
568 const ArgList &Args)
569 : Generic_ELF(D, Triple, Args),
570 OptionsDefault(
571 {{options::OPT_O, "3"}, {options::OPT_cl_std_EQ, "CL1.2"}}) {
572 // Check code object version options. Emit warnings for legacy options
573 // and errors for the last invalid code object version options.
574 // It is done here to avoid repeated warning or error messages for
575 // each tool invocation.
576 checkAMDGPUCodeObjectVersion(D, Args);
577 }
578
buildLinker() const579 Tool *AMDGPUToolChain::buildLinker() const {
580 return new tools::amdgpu::Linker(*this);
581 }
582
583 DerivedArgList *
TranslateArgs(const DerivedArgList & Args,StringRef BoundArch,Action::OffloadKind DeviceOffloadKind) const584 AMDGPUToolChain::TranslateArgs(const DerivedArgList &Args, StringRef BoundArch,
585 Action::OffloadKind DeviceOffloadKind) const {
586
587 DerivedArgList *DAL =
588 Generic_ELF::TranslateArgs(Args, BoundArch, DeviceOffloadKind);
589
590 const OptTable &Opts = getDriver().getOpts();
591
592 if (!DAL)
593 DAL = new DerivedArgList(Args.getBaseArgs());
594
595 for (Arg *A : Args) {
596 if (!shouldSkipArgument(A))
597 DAL->append(A);
598 }
599
600 checkTargetID(*DAL);
601
602 if (!Args.getLastArgValue(options::OPT_x).equals("cl"))
603 return DAL;
604
605 // Phase 1 (.cl -> .bc)
606 if (Args.hasArg(options::OPT_c) && Args.hasArg(options::OPT_emit_llvm)) {
607 DAL->AddFlagArg(nullptr, Opts.getOption(getTriple().isArch64Bit()
608 ? options::OPT_m64
609 : options::OPT_m32));
610
611 // Have to check OPT_O4, OPT_O0 & OPT_Ofast separately
612 // as they defined that way in Options.td
613 if (!Args.hasArg(options::OPT_O, options::OPT_O0, options::OPT_O4,
614 options::OPT_Ofast))
615 DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_O),
616 getOptionDefault(options::OPT_O));
617 }
618
619 return DAL;
620 }
621
getDefaultDenormsAreZeroForTarget(llvm::AMDGPU::GPUKind Kind)622 bool AMDGPUToolChain::getDefaultDenormsAreZeroForTarget(
623 llvm::AMDGPU::GPUKind Kind) {
624
625 // Assume nothing without a specific target.
626 if (Kind == llvm::AMDGPU::GK_NONE)
627 return false;
628
629 const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind);
630
631 // Default to enabling f32 denormals by default on subtargets where fma is
632 // fast with denormals
633 const bool BothDenormAndFMAFast =
634 (ArchAttr & llvm::AMDGPU::FEATURE_FAST_FMA_F32) &&
635 (ArchAttr & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32);
636 return !BothDenormAndFMAFast;
637 }
638
getDefaultDenormalModeForType(const llvm::opt::ArgList & DriverArgs,const JobAction & JA,const llvm::fltSemantics * FPType) const639 llvm::DenormalMode AMDGPUToolChain::getDefaultDenormalModeForType(
640 const llvm::opt::ArgList &DriverArgs, const JobAction &JA,
641 const llvm::fltSemantics *FPType) const {
642 // Denormals should always be enabled for f16 and f64.
643 if (!FPType || FPType != &llvm::APFloat::IEEEsingle())
644 return llvm::DenormalMode::getIEEE();
645
646 if (JA.getOffloadingDeviceKind() == Action::OFK_HIP ||
647 JA.getOffloadingDeviceKind() == Action::OFK_Cuda) {
648 auto Arch = getProcessorFromTargetID(getTriple(), JA.getOffloadingArch());
649 auto Kind = llvm::AMDGPU::parseArchAMDGCN(Arch);
650 if (FPType && FPType == &llvm::APFloat::IEEEsingle() &&
651 DriverArgs.hasFlag(options::OPT_fgpu_flush_denormals_to_zero,
652 options::OPT_fno_gpu_flush_denormals_to_zero,
653 getDefaultDenormsAreZeroForTarget(Kind)))
654 return llvm::DenormalMode::getPreserveSign();
655
656 return llvm::DenormalMode::getIEEE();
657 }
658
659 const StringRef GpuArch = getGPUArch(DriverArgs);
660 auto Kind = llvm::AMDGPU::parseArchAMDGCN(GpuArch);
661
662 // TODO: There are way too many flags that change this. Do we need to check
663 // them all?
664 bool DAZ = DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) ||
665 getDefaultDenormsAreZeroForTarget(Kind);
666
667 // Outputs are flushed to zero (FTZ), preserving sign. Denormal inputs are
668 // also implicit treated as zero (DAZ).
669 return DAZ ? llvm::DenormalMode::getPreserveSign() :
670 llvm::DenormalMode::getIEEE();
671 }
672
isWave64(const llvm::opt::ArgList & DriverArgs,llvm::AMDGPU::GPUKind Kind)673 bool AMDGPUToolChain::isWave64(const llvm::opt::ArgList &DriverArgs,
674 llvm::AMDGPU::GPUKind Kind) {
675 const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind);
676 bool HasWave32 = (ArchAttr & llvm::AMDGPU::FEATURE_WAVE32);
677
678 return !HasWave32 || DriverArgs.hasFlag(
679 options::OPT_mwavefrontsize64, options::OPT_mno_wavefrontsize64, false);
680 }
681
682
683 /// ROCM Toolchain
ROCMToolChain(const Driver & D,const llvm::Triple & Triple,const ArgList & Args)684 ROCMToolChain::ROCMToolChain(const Driver &D, const llvm::Triple &Triple,
685 const ArgList &Args)
686 : AMDGPUToolChain(D, Triple, Args) {
687 RocmInstallation.detectDeviceLibrary();
688 }
689
addClangTargetOptions(const llvm::opt::ArgList & DriverArgs,llvm::opt::ArgStringList & CC1Args,Action::OffloadKind DeviceOffloadingKind) const690 void AMDGPUToolChain::addClangTargetOptions(
691 const llvm::opt::ArgList &DriverArgs,
692 llvm::opt::ArgStringList &CC1Args,
693 Action::OffloadKind DeviceOffloadingKind) const {
694 // Default to "hidden" visibility, as object level linking will not be
695 // supported for the foreseeable future.
696 if (!DriverArgs.hasArg(options::OPT_fvisibility_EQ,
697 options::OPT_fvisibility_ms_compat)) {
698 CC1Args.push_back("-fvisibility");
699 CC1Args.push_back("hidden");
700 CC1Args.push_back("-fapply-global-visibility-to-externs");
701 }
702 }
703
704 StringRef
getGPUArch(const llvm::opt::ArgList & DriverArgs) const705 AMDGPUToolChain::getGPUArch(const llvm::opt::ArgList &DriverArgs) const {
706 return getProcessorFromTargetID(
707 getTriple(), DriverArgs.getLastArgValue(options::OPT_mcpu_EQ));
708 }
709
checkTargetID(const llvm::opt::ArgList & DriverArgs) const710 void AMDGPUToolChain::checkTargetID(
711 const llvm::opt::ArgList &DriverArgs) const {
712 StringRef TargetID = DriverArgs.getLastArgValue(options::OPT_mcpu_EQ);
713 if (TargetID.empty())
714 return;
715
716 llvm::StringMap<bool> FeatureMap;
717 auto OptionalGpuArch = parseTargetID(getTriple(), TargetID, &FeatureMap);
718 if (!OptionalGpuArch) {
719 getDriver().Diag(clang::diag::err_drv_bad_target_id) << TargetID;
720 }
721 }
722
723 llvm::Error
detectSystemGPUs(const ArgList & Args,SmallVector<std::string,1> & GPUArchs) const724 AMDGPUToolChain::detectSystemGPUs(const ArgList &Args,
725 SmallVector<std::string, 1> &GPUArchs) const {
726 std::string Program;
727 if (Arg *A = Args.getLastArg(options::OPT_amdgpu_arch_tool_EQ))
728 Program = A->getValue();
729 else
730 Program = GetProgramPath(AMDGPU_ARCH_PROGRAM_NAME);
731 llvm::SmallString<64> OutputFile;
732 llvm::sys::fs::createTemporaryFile("print-system-gpus", "" /* No Suffix */,
733 OutputFile);
734 llvm::FileRemover OutputRemover(OutputFile.c_str());
735 llvm::Optional<llvm::StringRef> Redirects[] = {
736 {""},
737 StringRef(OutputFile),
738 {""},
739 };
740
741 std::string ErrorMessage;
742 if (int Result = llvm::sys::ExecuteAndWait(
743 Program.c_str(), {}, {}, Redirects, /* SecondsToWait */ 0,
744 /*MemoryLimit*/ 0, &ErrorMessage)) {
745 if (Result > 0) {
746 ErrorMessage = "Exited with error code " + std::to_string(Result);
747 } else if (Result == -1) {
748 ErrorMessage = "Execute failed: " + ErrorMessage;
749 } else {
750 ErrorMessage = "Crashed: " + ErrorMessage;
751 }
752
753 return llvm::createStringError(std::error_code(),
754 Program + ": " + ErrorMessage);
755 }
756
757 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> OutputBuf =
758 llvm::MemoryBuffer::getFile(OutputFile.c_str());
759 if (!OutputBuf) {
760 return llvm::createStringError(OutputBuf.getError(),
761 "Failed to read stdout of " + Program +
762 ": " + OutputBuf.getError().message());
763 }
764
765 for (llvm::line_iterator LineIt(**OutputBuf); !LineIt.is_at_end(); ++LineIt) {
766 GPUArchs.push_back(LineIt->str());
767 }
768 return llvm::Error::success();
769 }
770
getSystemGPUArch(const ArgList & Args,std::string & GPUArch) const771 llvm::Error AMDGPUToolChain::getSystemGPUArch(const ArgList &Args,
772 std::string &GPUArch) const {
773 // detect the AMDGPU installed in system
774 SmallVector<std::string, 1> GPUArchs;
775 auto Err = detectSystemGPUs(Args, GPUArchs);
776 if (Err) {
777 return Err;
778 }
779 if (GPUArchs.empty()) {
780 return llvm::createStringError(std::error_code(),
781 "No AMD GPU detected in the system");
782 }
783 GPUArch = GPUArchs[0];
784 if (GPUArchs.size() > 1) {
785 bool AllSame = std::all_of(
786 GPUArchs.begin(), GPUArchs.end(),
787 [&](const StringRef &GPUArch) { return GPUArch == GPUArchs.front(); });
788 if (!AllSame)
789 return llvm::createStringError(
790 std::error_code(), "Multiple AMD GPUs found with different archs");
791 }
792 return llvm::Error::success();
793 }
794
addClangTargetOptions(const llvm::opt::ArgList & DriverArgs,llvm::opt::ArgStringList & CC1Args,Action::OffloadKind DeviceOffloadingKind) const795 void ROCMToolChain::addClangTargetOptions(
796 const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args,
797 Action::OffloadKind DeviceOffloadingKind) const {
798 AMDGPUToolChain::addClangTargetOptions(DriverArgs, CC1Args,
799 DeviceOffloadingKind);
800
801 // For the OpenCL case where there is no offload target, accept -nostdlib to
802 // disable bitcode linking.
803 if (DeviceOffloadingKind == Action::OFK_None &&
804 DriverArgs.hasArg(options::OPT_nostdlib))
805 return;
806
807 if (DriverArgs.hasArg(options::OPT_nogpulib))
808 return;
809
810 if (!RocmInstallation.hasDeviceLibrary()) {
811 getDriver().Diag(diag::err_drv_no_rocm_device_lib) << 0;
812 return;
813 }
814
815 // Get the device name and canonicalize it
816 const StringRef GpuArch = getGPUArch(DriverArgs);
817 auto Kind = llvm::AMDGPU::parseArchAMDGCN(GpuArch);
818 const StringRef CanonArch = llvm::AMDGPU::getArchNameAMDGCN(Kind);
819 std::string LibDeviceFile = RocmInstallation.getLibDeviceFile(CanonArch);
820 if (LibDeviceFile.empty()) {
821 getDriver().Diag(diag::err_drv_no_rocm_device_lib) << 1 << GpuArch;
822 return;
823 }
824
825 bool Wave64 = isWave64(DriverArgs, Kind);
826
827 // TODO: There are way too many flags that change this. Do we need to check
828 // them all?
829 bool DAZ = DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) ||
830 getDefaultDenormsAreZeroForTarget(Kind);
831 bool FiniteOnly = DriverArgs.hasArg(options::OPT_cl_finite_math_only);
832
833 bool UnsafeMathOpt =
834 DriverArgs.hasArg(options::OPT_cl_unsafe_math_optimizations);
835 bool FastRelaxedMath = DriverArgs.hasArg(options::OPT_cl_fast_relaxed_math);
836 bool CorrectSqrt =
837 DriverArgs.hasArg(options::OPT_cl_fp32_correctly_rounded_divide_sqrt);
838
839 // Add the OpenCL specific bitcode library.
840 llvm::SmallVector<std::string, 12> BCLibs;
841 BCLibs.push_back(RocmInstallation.getOpenCLPath().str());
842
843 // Add the generic set of libraries.
844 BCLibs.append(RocmInstallation.getCommonBitcodeLibs(
845 DriverArgs, LibDeviceFile, Wave64, DAZ, FiniteOnly, UnsafeMathOpt,
846 FastRelaxedMath, CorrectSqrt));
847
848 llvm::for_each(BCLibs, [&](StringRef BCFile) {
849 CC1Args.push_back("-mlink-builtin-bitcode");
850 CC1Args.push_back(DriverArgs.MakeArgString(BCFile));
851 });
852 }
853
854 llvm::SmallVector<std::string, 12>
getCommonBitcodeLibs(const llvm::opt::ArgList & DriverArgs,StringRef LibDeviceFile,bool Wave64,bool DAZ,bool FiniteOnly,bool UnsafeMathOpt,bool FastRelaxedMath,bool CorrectSqrt) const855 RocmInstallationDetector::getCommonBitcodeLibs(
856 const llvm::opt::ArgList &DriverArgs, StringRef LibDeviceFile, bool Wave64,
857 bool DAZ, bool FiniteOnly, bool UnsafeMathOpt, bool FastRelaxedMath,
858 bool CorrectSqrt) const {
859
860 llvm::SmallVector<std::string, 12> BCLibs;
861
862 auto AddBCLib = [&](StringRef BCFile) { BCLibs.push_back(BCFile.str()); };
863
864 AddBCLib(getOCMLPath());
865 AddBCLib(getOCKLPath());
866 AddBCLib(getDenormalsAreZeroPath(DAZ));
867 AddBCLib(getUnsafeMathPath(UnsafeMathOpt || FastRelaxedMath));
868 AddBCLib(getFiniteOnlyPath(FiniteOnly || FastRelaxedMath));
869 AddBCLib(getCorrectlyRoundedSqrtPath(CorrectSqrt));
870 AddBCLib(getWavefrontSize64Path(Wave64));
871 AddBCLib(LibDeviceFile);
872
873 return BCLibs;
874 }
875
shouldSkipArgument(const llvm::opt::Arg * A) const876 bool AMDGPUToolChain::shouldSkipArgument(const llvm::opt::Arg *A) const {
877 Option O = A->getOption();
878 if (O.matches(options::OPT_fPIE) || O.matches(options::OPT_fpie))
879 return true;
880 return false;
881 }
882