xref: /freebsd/contrib/llvm-project/clang/lib/Driver/ToolChains/HIPAMD.cpp (revision 5b56413d04e608379c9a306373554a8e4d321bc0)
1 //===--- HIPAMD.cpp - HIP 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 "HIPAMD.h"
10 #include "AMDGPU.h"
11 #include "CommonArgs.h"
12 #include "HIPUtility.h"
13 #include "clang/Basic/Cuda.h"
14 #include "clang/Basic/TargetID.h"
15 #include "clang/Driver/Compilation.h"
16 #include "clang/Driver/Driver.h"
17 #include "clang/Driver/DriverDiagnostic.h"
18 #include "clang/Driver/InputInfo.h"
19 #include "clang/Driver/Options.h"
20 #include "clang/Driver/SanitizerArgs.h"
21 #include "llvm/Support/Alignment.h"
22 #include "llvm/Support/FileSystem.h"
23 #include "llvm/Support/Path.h"
24 #include "llvm/TargetParser/TargetParser.h"
25 
26 using namespace clang::driver;
27 using namespace clang::driver::toolchains;
28 using namespace clang::driver::tools;
29 using namespace clang;
30 using namespace llvm::opt;
31 
32 #if defined(_WIN32) || defined(_WIN64)
33 #define NULL_FILE "nul"
34 #else
35 #define NULL_FILE "/dev/null"
36 #endif
37 
38 static bool shouldSkipSanitizeOption(const ToolChain &TC,
39                                      const llvm::opt::ArgList &DriverArgs,
40                                      StringRef TargetID,
41                                      const llvm::opt::Arg *A) {
42   // For actions without targetID, do nothing.
43   if (TargetID.empty())
44     return false;
45   Option O = A->getOption();
46   if (!O.matches(options::OPT_fsanitize_EQ))
47     return false;
48 
49   if (!DriverArgs.hasFlag(options::OPT_fgpu_sanitize,
50                           options::OPT_fno_gpu_sanitize, true))
51     return true;
52 
53   auto &Diags = TC.getDriver().getDiags();
54 
55   // For simplicity, we only allow -fsanitize=address
56   SanitizerMask K = parseSanitizerValue(A->getValue(), /*AllowGroups=*/false);
57   if (K != SanitizerKind::Address)
58     return true;
59 
60   llvm::StringMap<bool> FeatureMap;
61   auto OptionalGpuArch = parseTargetID(TC.getTriple(), TargetID, &FeatureMap);
62 
63   assert(OptionalGpuArch && "Invalid Target ID");
64   (void)OptionalGpuArch;
65   auto Loc = FeatureMap.find("xnack");
66   if (Loc == FeatureMap.end() || !Loc->second) {
67     Diags.Report(
68         clang::diag::warn_drv_unsupported_option_for_offload_arch_req_feature)
69         << A->getAsString(DriverArgs) << TargetID << "xnack+";
70     return true;
71   }
72   return false;
73 }
74 
75 void AMDGCN::Linker::constructLlvmLinkCommand(Compilation &C,
76                                          const JobAction &JA,
77                                          const InputInfoList &Inputs,
78                                          const InputInfo &Output,
79                                          const llvm::opt::ArgList &Args) const {
80   // Construct llvm-link command.
81   // The output from llvm-link is a bitcode file.
82   ArgStringList LlvmLinkArgs;
83 
84   assert(!Inputs.empty() && "Must have at least one input.");
85 
86   LlvmLinkArgs.append({"-o", Output.getFilename()});
87   for (auto Input : Inputs)
88     LlvmLinkArgs.push_back(Input.getFilename());
89 
90   // Look for archive of bundled bitcode in arguments, and add temporary files
91   // for the extracted archive of bitcode to inputs.
92   auto TargetID = Args.getLastArgValue(options::OPT_mcpu_EQ);
93   AddStaticDeviceLibsLinking(C, *this, JA, Inputs, Args, LlvmLinkArgs, "amdgcn",
94                              TargetID, /*IsBitCodeSDL=*/true);
95 
96   const char *LlvmLink =
97     Args.MakeArgString(getToolChain().GetProgramPath("llvm-link"));
98   C.addCommand(std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
99                                          LlvmLink, LlvmLinkArgs, Inputs,
100                                          Output));
101 }
102 
103 void AMDGCN::Linker::constructLldCommand(Compilation &C, const JobAction &JA,
104                                          const InputInfoList &Inputs,
105                                          const InputInfo &Output,
106                                          const llvm::opt::ArgList &Args) const {
107   // Construct lld command.
108   // The output from ld.lld is an HSA code object file.
109   ArgStringList LldArgs{"-flavor",
110                         "gnu",
111                         "-m",
112                         "elf64_amdgpu",
113                         "--no-undefined",
114                         "-shared",
115                         "-plugin-opt=-amdgpu-internalize-symbols"};
116   if (Args.hasArg(options::OPT_hipstdpar))
117     LldArgs.push_back("-plugin-opt=-amdgpu-enable-hipstdpar");
118 
119   auto &TC = getToolChain();
120   auto &D = TC.getDriver();
121   assert(!Inputs.empty() && "Must have at least one input.");
122   bool IsThinLTO = D.getLTOMode(/*IsOffload=*/true) == LTOK_Thin;
123   addLTOOptions(TC, Args, LldArgs, Output, Inputs[0], IsThinLTO);
124 
125   // Extract all the -m options
126   std::vector<llvm::StringRef> Features;
127   amdgpu::getAMDGPUTargetFeatures(D, TC.getTriple(), Args, Features);
128 
129   // Add features to mattr such as cumode
130   std::string MAttrString = "-plugin-opt=-mattr=";
131   for (auto OneFeature : unifyTargetFeatures(Features)) {
132     MAttrString.append(Args.MakeArgString(OneFeature));
133     if (OneFeature != Features.back())
134       MAttrString.append(",");
135   }
136   if (!Features.empty())
137     LldArgs.push_back(Args.MakeArgString(MAttrString));
138 
139   // ToDo: Remove this option after AMDGPU backend supports ISA-level linking.
140   // Since AMDGPU backend currently does not support ISA-level linking, all
141   // called functions need to be imported.
142   if (IsThinLTO)
143     LldArgs.push_back(Args.MakeArgString("-plugin-opt=-force-import-all"));
144 
145   if (C.getDriver().isSaveTempsEnabled())
146     LldArgs.push_back("-save-temps");
147 
148   addLinkerCompressDebugSectionsOption(TC, Args, LldArgs);
149 
150   // Given that host and device linking happen in separate processes, the device
151   // linker doesn't always have the visibility as to which device symbols are
152   // needed by a program, especially for the device symbol dependencies that are
153   // introduced through the host symbol resolution.
154   // For example: host_A() (A.obj) --> host_B(B.obj) --> device_kernel_B()
155   // (B.obj) In this case, the device linker doesn't know that A.obj actually
156   // depends on the kernel functions in B.obj.  When linking to static device
157   // library, the device linker may drop some of the device global symbols if
158   // they aren't referenced.  As a workaround, we are adding to the
159   // --whole-archive flag such that all global symbols would be linked in.
160   LldArgs.push_back("--whole-archive");
161 
162   for (auto *Arg : Args.filtered(options::OPT_Xoffload_linker)) {
163     StringRef ArgVal = Arg->getValue(1);
164     auto SplitArg = ArgVal.split("-mllvm=");
165     if (!SplitArg.second.empty()) {
166       LldArgs.push_back(
167           Args.MakeArgString(Twine("-plugin-opt=") + SplitArg.second));
168     } else {
169       LldArgs.push_back(Args.MakeArgString(ArgVal));
170     }
171     Arg->claim();
172   }
173 
174   LldArgs.append({"-o", Output.getFilename()});
175   for (auto Input : Inputs)
176     LldArgs.push_back(Input.getFilename());
177 
178   // Look for archive of bundled bitcode in arguments, and add temporary files
179   // for the extracted archive of bitcode to inputs.
180   auto TargetID = Args.getLastArgValue(options::OPT_mcpu_EQ);
181   AddStaticDeviceLibsLinking(C, *this, JA, Inputs, Args, LldArgs, "amdgcn",
182                              TargetID, /*IsBitCodeSDL=*/true);
183 
184   LldArgs.push_back("--no-whole-archive");
185 
186   const char *Lld = Args.MakeArgString(getToolChain().GetProgramPath("lld"));
187   C.addCommand(std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
188                                          Lld, LldArgs, Inputs, Output));
189 }
190 
191 // For amdgcn the inputs of the linker job are device bitcode and output is
192 // either an object file or bitcode (-emit-llvm). It calls llvm-link, opt,
193 // llc, then lld steps.
194 void AMDGCN::Linker::ConstructJob(Compilation &C, const JobAction &JA,
195                                   const InputInfo &Output,
196                                   const InputInfoList &Inputs,
197                                   const ArgList &Args,
198                                   const char *LinkingOutput) const {
199   if (Inputs.size() > 0 &&
200       Inputs[0].getType() == types::TY_Image &&
201       JA.getType() == types::TY_Object)
202     return HIP::constructGenerateObjFileFromHIPFatBinary(C, Output, Inputs,
203                                                          Args, JA, *this);
204 
205   if (JA.getType() == types::TY_HIP_FATBIN)
206     return HIP::constructHIPFatbinCommand(C, JA, Output.getFilename(), Inputs,
207                                           Args, *this);
208 
209   if (JA.getType() == types::TY_LLVM_BC)
210     return constructLlvmLinkCommand(C, JA, Inputs, Output, Args);
211 
212   return constructLldCommand(C, JA, Inputs, Output, Args);
213 }
214 
215 HIPAMDToolChain::HIPAMDToolChain(const Driver &D, const llvm::Triple &Triple,
216                                  const ToolChain &HostTC, const ArgList &Args)
217     : ROCMToolChain(D, Triple, Args), HostTC(HostTC) {
218   // Lookup binaries into the driver directory, this is used to
219   // discover the clang-offload-bundler executable.
220   getProgramPaths().push_back(getDriver().Dir);
221 
222   // Diagnose unsupported sanitizer options only once.
223   if (!Args.hasFlag(options::OPT_fgpu_sanitize, options::OPT_fno_gpu_sanitize,
224                     true))
225     return;
226   for (auto *A : Args.filtered(options::OPT_fsanitize_EQ)) {
227     SanitizerMask K = parseSanitizerValue(A->getValue(), /*AllowGroups=*/false);
228     if (K != SanitizerKind::Address)
229       D.getDiags().Report(clang::diag::warn_drv_unsupported_option_for_target)
230           << A->getAsString(Args) << getTriple().str();
231   }
232 }
233 
234 void HIPAMDToolChain::addClangTargetOptions(
235     const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args,
236     Action::OffloadKind DeviceOffloadingKind) const {
237   HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind);
238 
239   assert(DeviceOffloadingKind == Action::OFK_HIP &&
240          "Only HIP offloading kinds are supported for GPUs.");
241 
242   CC1Args.push_back("-fcuda-is-device");
243 
244   if (!DriverArgs.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc,
245                           false))
246     CC1Args.append({"-mllvm", "-amdgpu-internalize-symbols"});
247   if (DriverArgs.hasArgNoClaim(options::OPT_hipstdpar))
248     CC1Args.append({"-mllvm", "-amdgpu-enable-hipstdpar"});
249 
250   StringRef MaxThreadsPerBlock =
251       DriverArgs.getLastArgValue(options::OPT_gpu_max_threads_per_block_EQ);
252   if (!MaxThreadsPerBlock.empty()) {
253     std::string ArgStr =
254         (Twine("--gpu-max-threads-per-block=") + MaxThreadsPerBlock).str();
255     CC1Args.push_back(DriverArgs.MakeArgStringRef(ArgStr));
256   }
257 
258   CC1Args.push_back("-fcuda-allow-variadic-functions");
259 
260   // Default to "hidden" visibility, as object level linking will not be
261   // supported for the foreseeable future.
262   if (!DriverArgs.hasArg(options::OPT_fvisibility_EQ,
263                          options::OPT_fvisibility_ms_compat)) {
264     CC1Args.append({"-fvisibility=hidden"});
265     CC1Args.push_back("-fapply-global-visibility-to-externs");
266   }
267 
268   for (auto BCFile : getDeviceLibs(DriverArgs)) {
269     CC1Args.push_back(BCFile.ShouldInternalize ? "-mlink-builtin-bitcode"
270                                                : "-mlink-bitcode-file");
271     CC1Args.push_back(DriverArgs.MakeArgString(BCFile.Path));
272   }
273 }
274 
275 llvm::opt::DerivedArgList *
276 HIPAMDToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
277                                StringRef BoundArch,
278                                Action::OffloadKind DeviceOffloadKind) const {
279   DerivedArgList *DAL =
280       HostTC.TranslateArgs(Args, BoundArch, DeviceOffloadKind);
281   if (!DAL)
282     DAL = new DerivedArgList(Args.getBaseArgs());
283 
284   const OptTable &Opts = getDriver().getOpts();
285 
286   for (Arg *A : Args) {
287     if (!shouldSkipSanitizeOption(*this, Args, BoundArch, A))
288       DAL->append(A);
289   }
290 
291   if (!BoundArch.empty()) {
292     DAL->eraseArg(options::OPT_mcpu_EQ);
293     DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_mcpu_EQ), BoundArch);
294     checkTargetID(*DAL);
295   }
296 
297   return DAL;
298 }
299 
300 Tool *HIPAMDToolChain::buildLinker() const {
301   assert(getTriple().getArch() == llvm::Triple::amdgcn);
302   return new tools::AMDGCN::Linker(*this);
303 }
304 
305 void HIPAMDToolChain::addClangWarningOptions(ArgStringList &CC1Args) const {
306   HostTC.addClangWarningOptions(CC1Args);
307 }
308 
309 ToolChain::CXXStdlibType
310 HIPAMDToolChain::GetCXXStdlibType(const ArgList &Args) const {
311   return HostTC.GetCXXStdlibType(Args);
312 }
313 
314 void HIPAMDToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
315                                                 ArgStringList &CC1Args) const {
316   HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args);
317 }
318 
319 void HIPAMDToolChain::AddClangCXXStdlibIncludeArgs(
320     const ArgList &Args, ArgStringList &CC1Args) const {
321   HostTC.AddClangCXXStdlibIncludeArgs(Args, CC1Args);
322 }
323 
324 void HIPAMDToolChain::AddIAMCUIncludeArgs(const ArgList &Args,
325                                           ArgStringList &CC1Args) const {
326   HostTC.AddIAMCUIncludeArgs(Args, CC1Args);
327 }
328 
329 void HIPAMDToolChain::AddHIPIncludeArgs(const ArgList &DriverArgs,
330                                         ArgStringList &CC1Args) const {
331   RocmInstallation->AddHIPIncludeArgs(DriverArgs, CC1Args);
332 }
333 
334 SanitizerMask HIPAMDToolChain::getSupportedSanitizers() const {
335   // The HIPAMDToolChain only supports sanitizers in the sense that it allows
336   // sanitizer arguments on the command line if they are supported by the host
337   // toolchain. The HIPAMDToolChain will actually ignore any command line
338   // arguments for any of these "supported" sanitizers. That means that no
339   // sanitization of device code is actually supported at this time.
340   //
341   // This behavior is necessary because the host and device toolchains
342   // invocations often share the command line, so the device toolchain must
343   // tolerate flags meant only for the host toolchain.
344   return HostTC.getSupportedSanitizers();
345 }
346 
347 VersionTuple HIPAMDToolChain::computeMSVCVersion(const Driver *D,
348                                                  const ArgList &Args) const {
349   return HostTC.computeMSVCVersion(D, Args);
350 }
351 
352 llvm::SmallVector<ToolChain::BitCodeLibraryInfo, 12>
353 HIPAMDToolChain::getDeviceLibs(const llvm::opt::ArgList &DriverArgs) const {
354   llvm::SmallVector<BitCodeLibraryInfo, 12> BCLibs;
355   if (DriverArgs.hasArg(options::OPT_nogpulib))
356     return {};
357   ArgStringList LibraryPaths;
358 
359   // Find in --hip-device-lib-path and HIP_LIBRARY_PATH.
360   for (StringRef Path : RocmInstallation->getRocmDeviceLibPathArg())
361     LibraryPaths.push_back(DriverArgs.MakeArgString(Path));
362 
363   addDirectoryList(DriverArgs, LibraryPaths, "", "HIP_DEVICE_LIB_PATH");
364 
365   // Maintain compatability with --hip-device-lib.
366   auto BCLibArgs = DriverArgs.getAllArgValues(options::OPT_hip_device_lib_EQ);
367   if (!BCLibArgs.empty()) {
368     llvm::for_each(BCLibArgs, [&](StringRef BCName) {
369       StringRef FullName;
370       for (StringRef LibraryPath : LibraryPaths) {
371         SmallString<128> Path(LibraryPath);
372         llvm::sys::path::append(Path, BCName);
373         FullName = Path;
374         if (llvm::sys::fs::exists(FullName)) {
375           BCLibs.push_back(FullName);
376           return;
377         }
378       }
379       getDriver().Diag(diag::err_drv_no_such_file) << BCName;
380     });
381   } else {
382     if (!RocmInstallation->hasDeviceLibrary()) {
383       getDriver().Diag(diag::err_drv_no_rocm_device_lib) << 0;
384       return {};
385     }
386     StringRef GpuArch = getGPUArch(DriverArgs);
387     assert(!GpuArch.empty() && "Must have an explicit GPU arch.");
388 
389     // If --hip-device-lib is not set, add the default bitcode libraries.
390     if (DriverArgs.hasFlag(options::OPT_fgpu_sanitize,
391                            options::OPT_fno_gpu_sanitize, true) &&
392         getSanitizerArgs(DriverArgs).needsAsanRt()) {
393       auto AsanRTL = RocmInstallation->getAsanRTLPath();
394       if (AsanRTL.empty()) {
395         unsigned DiagID = getDriver().getDiags().getCustomDiagID(
396             DiagnosticsEngine::Error,
397             "AMDGPU address sanitizer runtime library (asanrtl) is not found. "
398             "Please install ROCm device library which supports address "
399             "sanitizer");
400         getDriver().Diag(DiagID);
401         return {};
402       } else
403         BCLibs.emplace_back(AsanRTL, /*ShouldInternalize=*/false);
404     }
405 
406     // Add the HIP specific bitcode library.
407     BCLibs.push_back(RocmInstallation->getHIPPath());
408 
409     // Add common device libraries like ocml etc.
410     for (StringRef N : getCommonDeviceLibNames(DriverArgs, GpuArch.str()))
411       BCLibs.emplace_back(N);
412 
413     // Add instrument lib.
414     auto InstLib =
415         DriverArgs.getLastArgValue(options::OPT_gpu_instrument_lib_EQ);
416     if (InstLib.empty())
417       return BCLibs;
418     if (llvm::sys::fs::exists(InstLib))
419       BCLibs.push_back(InstLib);
420     else
421       getDriver().Diag(diag::err_drv_no_such_file) << InstLib;
422   }
423 
424   return BCLibs;
425 }
426 
427 void HIPAMDToolChain::checkTargetID(
428     const llvm::opt::ArgList &DriverArgs) const {
429   auto PTID = getParsedTargetID(DriverArgs);
430   if (PTID.OptionalTargetID && !PTID.OptionalGPUArch) {
431     getDriver().Diag(clang::diag::err_drv_bad_target_id)
432         << *PTID.OptionalTargetID;
433   }
434 }
435