1 //===--- HIP.cpp - HIP 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 //===----------------------------------------------------------------------===//
11 #include "CommonArgs.h"
12 #include "InputInfo.h"
13 #include "clang/Basic/Cuda.h"
14 #include "clang/Driver/Compilation.h"
15 #include "clang/Driver/Driver.h"
16 #include "clang/Driver/DriverDiagnostic.h"
17 #include "clang/Driver/Options.h"
18 #include "llvm/Support/FileSystem.h"
19 #include "llvm/Support/Path.h"
20 #include "llvm/Support/TargetParser.h"
22 using namespace clang::driver;
23 using namespace clang::driver::toolchains;
24 using namespace clang::driver::tools;
25 using namespace clang;
26 using namespace llvm::opt;
28 #if defined(_WIN32) || defined(_WIN64)
29 #define NULL_FILE "nul"
31 #define NULL_FILE "/dev/null"
36 static void addBCLib(const Driver &D, const ArgList &Args,
37 ArgStringList &CmdArgs, ArgStringList LibraryPaths,
40 for (std::string LibraryPath : LibraryPaths) {
41 SmallString<128> Path(LibraryPath);
42 llvm::sys::path::append(Path, BCName);
44 if (llvm::sys::fs::exists(FullName)) {
45 CmdArgs.push_back("-mlink-builtin-bitcode");
46 CmdArgs.push_back(Args.MakeArgString(FullName));
50 D.Diag(diag::err_drv_no_such_file) << BCName;
54 void AMDGCN::Linker::constructLldCommand(Compilation &C, const JobAction &JA,
55 const InputInfoList &Inputs,
56 const InputInfo &Output,
57 const llvm::opt::ArgList &Args) const {
58 // Construct lld command.
59 // The output from ld.lld is an HSA code object file.
60 ArgStringList LldArgs{"-flavor", "gnu", "--no-undefined", "-shared",
61 "-plugin-opt=-amdgpu-internalize-symbols"};
63 auto &TC = getToolChain();
64 auto &D = TC.getDriver();
65 assert(!Inputs.empty() && "Must have at least one input.");
66 addLTOOptions(TC, Args, LldArgs, Output, Inputs[0],
67 D.getLTOMode() == LTOK_Thin);
69 // Extract all the -m options
70 std::vector<llvm::StringRef> Features;
71 amdgpu::getAMDGPUTargetFeatures(D, Args, Features);
73 // Add features to mattr such as cumode
74 std::string MAttrString = "-plugin-opt=-mattr=";
75 for (auto OneFeature : unifyTargetFeatures(Features)) {
76 MAttrString.append(Args.MakeArgString(OneFeature));
77 if (OneFeature != Features.back())
78 MAttrString.append(",");
80 if (!Features.empty())
81 LldArgs.push_back(Args.MakeArgString(MAttrString));
83 for (const Arg *A : Args.filtered(options::OPT_mllvm)) {
85 Args.MakeArgString(Twine("-plugin-opt=") + A->getValue(0)));
88 if (C.getDriver().isSaveTempsEnabled())
89 LldArgs.push_back("-save-temps");
91 LldArgs.append({"-o", Output.getFilename()});
92 for (auto Input : Inputs)
93 LldArgs.push_back(Input.getFilename());
94 const char *Lld = Args.MakeArgString(getToolChain().GetProgramPath("lld"));
95 C.addCommand(std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
96 Lld, LldArgs, Inputs));
99 // Construct a clang-offload-bundler command to bundle code objects for
100 // different GPU's into a HIP fat binary.
101 void AMDGCN::constructHIPFatbinCommand(Compilation &C, const JobAction &JA,
102 StringRef OutputFileName, const InputInfoList &Inputs,
103 const llvm::opt::ArgList &Args, const Tool& T) {
104 // Construct clang-offload-bundler command to bundle object files for
105 // for different GPU archs.
106 ArgStringList BundlerArgs;
107 BundlerArgs.push_back(Args.MakeArgString("-type=o"));
109 // ToDo: Remove the dummy host binary entry which is required by
110 // clang-offload-bundler.
111 std::string BundlerTargetArg = "-targets=host-x86_64-unknown-linux";
112 std::string BundlerInputArg = "-inputs=" NULL_FILE;
114 for (const auto &II : Inputs) {
115 const auto* A = II.getAction();
116 BundlerTargetArg = BundlerTargetArg + ",hip-amdgcn-amd-amdhsa-" +
117 StringRef(A->getOffloadingArch()).str();
118 BundlerInputArg = BundlerInputArg + "," + II.getFilename();
120 BundlerArgs.push_back(Args.MakeArgString(BundlerTargetArg));
121 BundlerArgs.push_back(Args.MakeArgString(BundlerInputArg));
123 auto BundlerOutputArg = Args.MakeArgString(
124 std::string("-outputs=").append(std::string(OutputFileName)));
125 BundlerArgs.push_back(BundlerOutputArg);
127 const char *Bundler = Args.MakeArgString(
128 T.getToolChain().GetProgramPath("clang-offload-bundler"));
129 C.addCommand(std::make_unique<Command>(JA, T, ResponseFileSupport::None(),
130 Bundler, BundlerArgs, Inputs));
133 /// Add Generated HIP Object File which has device images embedded into the
134 /// host to the argument list for linking. Using MC directives, embed the
135 /// device code and also define symbols required by the code generation so that
136 /// the image can be retrieved at runtime.
137 void AMDGCN::Linker::constructGenerateObjFileFromHIPFatBinary(
138 Compilation &C, const InputInfo &Output,
139 const InputInfoList &Inputs, const ArgList &Args,
140 const JobAction &JA) const {
141 const ToolChain &TC = getToolChain();
143 std::string(llvm::sys::path::stem(Output.getFilename()));
145 // Create Temp Object File Generator,
146 // Offload Bundled file and Bundled Object file.
147 // Keep them if save-temps is enabled.
148 const char *McinFile;
149 const char *BundleFile;
150 if (C.getDriver().isSaveTempsEnabled()) {
151 McinFile = C.getArgs().MakeArgString(Name + ".mcin");
152 BundleFile = C.getArgs().MakeArgString(Name + ".hipfb");
154 auto TmpNameMcin = C.getDriver().GetTemporaryPath(Name, "mcin");
155 McinFile = C.addTempFile(C.getArgs().MakeArgString(TmpNameMcin));
156 auto TmpNameFb = C.getDriver().GetTemporaryPath(Name, "hipfb");
157 BundleFile = C.addTempFile(C.getArgs().MakeArgString(TmpNameFb));
159 constructHIPFatbinCommand(C, JA, BundleFile, Inputs, Args, *this);
161 // Create a buffer to write the contents of the temp obj generator.
162 std::string ObjBuffer;
163 llvm::raw_string_ostream ObjStream(ObjBuffer);
165 // Add MC directives to embed target binaries. We ensure that each
166 // section and image is 16-byte aligned. This is not mandatory, but
167 // increases the likelihood of data to be aligned with a cache block
168 // in several main host machines.
169 ObjStream << "# HIP Object Generator\n";
170 ObjStream << "# *** Automatically generated by Clang ***\n";
171 ObjStream << " .type __hip_fatbin,@object\n";
172 ObjStream << " .section .hip_fatbin,\"aMS\",@progbits,1\n";
173 ObjStream << " .data\n";
174 ObjStream << " .globl __hip_fatbin\n";
175 ObjStream << " .p2align 3\n";
176 ObjStream << "__hip_fatbin:\n";
177 ObjStream << " .incbin \"" << BundleFile << "\"\n";
180 // Dump the contents of the temp object file gen if the user requested that.
181 // We support this option to enable testing of behavior with -###.
182 if (C.getArgs().hasArg(options::OPT_fhip_dump_offload_linker_script))
183 llvm::errs() << ObjBuffer;
185 // Open script file and write the contents.
187 llvm::raw_fd_ostream Objf(McinFile, EC, llvm::sys::fs::OF_None);
190 C.getDriver().Diag(clang::diag::err_unable_to_make_temp) << EC.message();
196 ArgStringList McArgs{"-o", Output.getFilename(),
197 McinFile, "--filetype=obj"};
198 const char *Mc = Args.MakeArgString(TC.GetProgramPath("llvm-mc"));
199 C.addCommand(std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
200 Mc, McArgs, Inputs));
203 // For amdgcn the inputs of the linker job are device bitcode and output is
204 // object file. It calls llvm-link, opt, llc, then lld steps.
205 void AMDGCN::Linker::ConstructJob(Compilation &C, const JobAction &JA,
206 const InputInfo &Output,
207 const InputInfoList &Inputs,
209 const char *LinkingOutput) const {
210 if (Inputs.size() > 0 &&
211 Inputs[0].getType() == types::TY_Image &&
212 JA.getType() == types::TY_Object)
213 return constructGenerateObjFileFromHIPFatBinary(C, Output, Inputs, Args, JA);
215 if (JA.getType() == types::TY_HIP_FATBIN)
216 return constructHIPFatbinCommand(C, JA, Output.getFilename(), Inputs, Args, *this);
218 return constructLldCommand(C, JA, Inputs, Output, Args);
221 HIPToolChain::HIPToolChain(const Driver &D, const llvm::Triple &Triple,
222 const ToolChain &HostTC, const ArgList &Args)
223 : ROCMToolChain(D, Triple, Args), HostTC(HostTC) {
224 // Lookup binaries into the driver directory, this is used to
225 // discover the clang-offload-bundler executable.
226 getProgramPaths().push_back(getDriver().Dir);
229 void HIPToolChain::addClangTargetOptions(
230 const llvm::opt::ArgList &DriverArgs,
231 llvm::opt::ArgStringList &CC1Args,
232 Action::OffloadKind DeviceOffloadingKind) const {
233 HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind);
235 StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_mcpu_EQ);
236 assert(!GpuArch.empty() && "Must have an explicit GPU arch.");
238 assert(DeviceOffloadingKind == Action::OFK_HIP &&
239 "Only HIP offloading kinds are supported for GPUs.");
240 auto Kind = llvm::AMDGPU::parseArchAMDGCN(GpuArch);
241 const StringRef CanonArch = llvm::AMDGPU::getArchNameAMDGCN(Kind);
243 CC1Args.push_back("-fcuda-is-device");
245 if (DriverArgs.hasFlag(options::OPT_fcuda_approx_transcendentals,
246 options::OPT_fno_cuda_approx_transcendentals, false))
247 CC1Args.push_back("-fcuda-approx-transcendentals");
249 if (DriverArgs.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc,
251 CC1Args.push_back("-fgpu-rdc");
253 CC1Args.append({"-mllvm", "-amdgpu-internalize-symbols"});
255 StringRef MaxThreadsPerBlock =
256 DriverArgs.getLastArgValue(options::OPT_gpu_max_threads_per_block_EQ);
257 if (!MaxThreadsPerBlock.empty()) {
259 std::string("--gpu-max-threads-per-block=") + MaxThreadsPerBlock.str();
260 CC1Args.push_back(DriverArgs.MakeArgStringRef(ArgStr));
263 if (DriverArgs.hasFlag(options::OPT_fgpu_allow_device_init,
264 options::OPT_fno_gpu_allow_device_init, false))
265 CC1Args.push_back("-fgpu-allow-device-init");
267 CC1Args.push_back("-fcuda-allow-variadic-functions");
269 // Default to "hidden" visibility, as object level linking will not be
270 // supported for the foreseeable future.
271 if (!DriverArgs.hasArg(options::OPT_fvisibility_EQ,
272 options::OPT_fvisibility_ms_compat)) {
273 CC1Args.append({"-fvisibility", "hidden"});
274 CC1Args.push_back("-fapply-global-visibility-to-externs");
277 if (DriverArgs.hasArg(options::OPT_nogpulib))
279 ArgStringList LibraryPaths;
281 // Find in --hip-device-lib-path and HIP_LIBRARY_PATH.
282 for (auto Path : RocmInstallation.getRocmDeviceLibPathArg())
283 LibraryPaths.push_back(DriverArgs.MakeArgString(Path));
285 addDirectoryList(DriverArgs, LibraryPaths, "", "HIP_DEVICE_LIB_PATH");
287 // Maintain compatability with --hip-device-lib.
288 auto BCLibs = DriverArgs.getAllArgValues(options::OPT_hip_device_lib_EQ);
289 if (!BCLibs.empty()) {
290 for (auto Lib : BCLibs)
291 addBCLib(getDriver(), DriverArgs, CC1Args, LibraryPaths, Lib);
293 if (!RocmInstallation.hasDeviceLibrary()) {
294 getDriver().Diag(diag::err_drv_no_rocm_device_lib) << 0;
298 std::string LibDeviceFile = RocmInstallation.getLibDeviceFile(CanonArch);
299 if (LibDeviceFile.empty()) {
300 getDriver().Diag(diag::err_drv_no_rocm_device_lib) << 1 << GpuArch;
304 // If --hip-device-lib is not set, add the default bitcode libraries.
305 // TODO: There are way too many flags that change this. Do we need to check
307 bool DAZ = DriverArgs.hasFlag(options::OPT_fcuda_flush_denormals_to_zero,
308 options::OPT_fno_cuda_flush_denormals_to_zero,
309 getDefaultDenormsAreZeroForTarget(Kind));
310 // TODO: Check standard C++ flags?
311 bool FiniteOnly = false;
312 bool UnsafeMathOpt = false;
313 bool FastRelaxedMath = false;
314 bool CorrectSqrt = true;
315 bool Wave64 = isWave64(DriverArgs, Kind);
317 // Add the HIP specific bitcode library.
318 CC1Args.push_back("-mlink-builtin-bitcode");
319 CC1Args.push_back(DriverArgs.MakeArgString(RocmInstallation.getHIPPath()));
321 // Add the generic set of libraries.
322 RocmInstallation.addCommonBitcodeLibCC1Args(
323 DriverArgs, CC1Args, LibDeviceFile, Wave64, DAZ, FiniteOnly,
324 UnsafeMathOpt, FastRelaxedMath, CorrectSqrt);
328 llvm::opt::DerivedArgList *
329 HIPToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
331 Action::OffloadKind DeviceOffloadKind) const {
332 DerivedArgList *DAL =
333 HostTC.TranslateArgs(Args, BoundArch, DeviceOffloadKind);
335 DAL = new DerivedArgList(Args.getBaseArgs());
337 const OptTable &Opts = getDriver().getOpts();
339 for (Arg *A : Args) {
343 if (!BoundArch.empty()) {
344 DAL->eraseArg(options::OPT_mcpu_EQ);
345 DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_mcpu_EQ), BoundArch);
351 Tool *HIPToolChain::buildLinker() const {
352 assert(getTriple().getArch() == llvm::Triple::amdgcn);
353 return new tools::AMDGCN::Linker(*this);
356 void HIPToolChain::addClangWarningOptions(ArgStringList &CC1Args) const {
357 HostTC.addClangWarningOptions(CC1Args);
360 ToolChain::CXXStdlibType
361 HIPToolChain::GetCXXStdlibType(const ArgList &Args) const {
362 return HostTC.GetCXXStdlibType(Args);
365 void HIPToolChain::AddClangSystemIncludeArgs(const ArgList &DriverArgs,
366 ArgStringList &CC1Args) const {
367 HostTC.AddClangSystemIncludeArgs(DriverArgs, CC1Args);
370 void HIPToolChain::AddClangCXXStdlibIncludeArgs(const ArgList &Args,
371 ArgStringList &CC1Args) const {
372 HostTC.AddClangCXXStdlibIncludeArgs(Args, CC1Args);
375 void HIPToolChain::AddIAMCUIncludeArgs(const ArgList &Args,
376 ArgStringList &CC1Args) const {
377 HostTC.AddIAMCUIncludeArgs(Args, CC1Args);
380 void HIPToolChain::AddHIPIncludeArgs(const ArgList &DriverArgs,
381 ArgStringList &CC1Args) const {
382 RocmInstallation.AddHIPIncludeArgs(DriverArgs, CC1Args);
385 SanitizerMask HIPToolChain::getSupportedSanitizers() const {
386 // The HIPToolChain only supports sanitizers in the sense that it allows
387 // sanitizer arguments on the command line if they are supported by the host
388 // toolchain. The HIPToolChain will actually ignore any command line
389 // arguments for any of these "supported" sanitizers. That means that no
390 // sanitization of device code is actually supported at this time.
392 // This behavior is necessary because the host and device toolchains
393 // invocations often share the command line, so the device toolchain must
394 // tolerate flags meant only for the host toolchain.
395 return HostTC.getSupportedSanitizers();
398 VersionTuple HIPToolChain::computeMSVCVersion(const Driver *D,
399 const ArgList &Args) const {
400 return HostTC.computeMSVCVersion(D, Args);