xref: /freebsd/contrib/llvm-project/clang/lib/Driver/ToolChains/Cuda.h (revision 47ef2a131091508e049ab10cad7f91a3c1342cd9)
1 //===--- Cuda.h - Cuda 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 #ifndef LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_CUDA_H
10 #define LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_CUDA_H
11 
12 #include "clang/Basic/Cuda.h"
13 #include "clang/Driver/Action.h"
14 #include "clang/Driver/Multilib.h"
15 #include "clang/Driver/Tool.h"
16 #include "clang/Driver/ToolChain.h"
17 #include "llvm/Support/Compiler.h"
18 #include "llvm/Support/VersionTuple.h"
19 #include <bitset>
20 #include <set>
21 #include <vector>
22 
23 namespace clang {
24 namespace driver {
25 
26 /// A class to find a viable CUDA installation
27 class CudaInstallationDetector {
28 private:
29   const Driver &D;
30   bool IsValid = false;
31   CudaVersion Version = CudaVersion::UNKNOWN;
32   std::string InstallPath;
33   std::string BinPath;
34   std::string LibDevicePath;
35   std::string IncludePath;
36   llvm::StringMap<std::string> LibDeviceMap;
37 
38   // CUDA architectures for which we have raised an error in
39   // CheckCudaVersionSupportsArch.
40   mutable std::bitset<(int)OffloadArch::LAST> ArchsWithBadVersion;
41 
42 public:
43   CudaInstallationDetector(const Driver &D, const llvm::Triple &HostTriple,
44                            const llvm::opt::ArgList &Args);
45 
46   void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs,
47                           llvm::opt::ArgStringList &CC1Args) const;
48 
49   /// Emit an error if Version does not support the given Arch.
50   ///
51   /// If either Version or Arch is unknown, does not emit an error.  Emits at
52   /// most one error per Arch.
53   void CheckCudaVersionSupportsArch(OffloadArch Arch) const;
54 
55   /// Check whether we detected a valid Cuda install.
56   bool isValid() const { return IsValid; }
57   /// Print information about the detected CUDA installation.
58   void print(raw_ostream &OS) const;
59 
60   /// Get the detected Cuda install's version.
61   CudaVersion version() const {
62     return Version == CudaVersion::NEW ? CudaVersion::PARTIALLY_SUPPORTED
63                                        : Version;
64   }
65   /// Get the detected Cuda installation path.
66   StringRef getInstallPath() const { return InstallPath; }
67   /// Get the detected path to Cuda's bin directory.
68   StringRef getBinPath() const { return BinPath; }
69   /// Get the detected Cuda Include path.
70   StringRef getIncludePath() const { return IncludePath; }
71   /// Get the detected Cuda device library path.
72   StringRef getLibDevicePath() const { return LibDevicePath; }
73   /// Get libdevice file for given architecture
74   std::string getLibDeviceFile(StringRef Gpu) const {
75     return LibDeviceMap.lookup(Gpu);
76   }
77   void WarnIfUnsupportedVersion();
78 };
79 
80 namespace tools {
81 namespace NVPTX {
82 
83 // Run ptxas, the NVPTX assembler.
84 class LLVM_LIBRARY_VISIBILITY Assembler final : public Tool {
85 public:
86   Assembler(const ToolChain &TC) : Tool("NVPTX::Assembler", "ptxas", TC) {}
87 
88   bool hasIntegratedCPP() const override { return false; }
89 
90   void ConstructJob(Compilation &C, const JobAction &JA,
91                     const InputInfo &Output, const InputInfoList &Inputs,
92                     const llvm::opt::ArgList &TCArgs,
93                     const char *LinkingOutput) const override;
94 };
95 
96 // Runs fatbinary, which combines GPU object files ("cubin" files) and/or PTX
97 // assembly into a single output file.
98 class LLVM_LIBRARY_VISIBILITY FatBinary : public Tool {
99 public:
100   FatBinary(const ToolChain &TC) : Tool("NVPTX::Linker", "fatbinary", TC) {}
101 
102   bool hasIntegratedCPP() const override { return false; }
103 
104   void ConstructJob(Compilation &C, const JobAction &JA,
105                     const InputInfo &Output, const InputInfoList &Inputs,
106                     const llvm::opt::ArgList &TCArgs,
107                     const char *LinkingOutput) const override;
108 };
109 
110 // Runs nvlink, which links GPU object files ("cubin" files) into a single file.
111 class LLVM_LIBRARY_VISIBILITY Linker final : public Tool {
112 public:
113   Linker(const ToolChain &TC) : Tool("NVPTX::Linker", "nvlink", TC) {}
114 
115   bool hasIntegratedCPP() const override { return false; }
116 
117   void ConstructJob(Compilation &C, const JobAction &JA,
118                     const InputInfo &Output, const InputInfoList &Inputs,
119                     const llvm::opt::ArgList &TCArgs,
120                     const char *LinkingOutput) const override;
121 };
122 
123 void getNVPTXTargetFeatures(const Driver &D, const llvm::Triple &Triple,
124                             const llvm::opt::ArgList &Args,
125                             std::vector<StringRef> &Features);
126 
127 } // end namespace NVPTX
128 } // end namespace tools
129 
130 namespace toolchains {
131 
132 class LLVM_LIBRARY_VISIBILITY NVPTXToolChain : public ToolChain {
133 public:
134   NVPTXToolChain(const Driver &D, const llvm::Triple &Triple,
135                  const llvm::Triple &HostTriple, const llvm::opt::ArgList &Args,
136                  bool Freestanding);
137 
138   NVPTXToolChain(const Driver &D, const llvm::Triple &Triple,
139                  const llvm::opt::ArgList &Args);
140 
141   llvm::opt::DerivedArgList *
142   TranslateArgs(const llvm::opt::DerivedArgList &Args, StringRef BoundArch,
143                 Action::OffloadKind DeviceOffloadKind) const override;
144 
145   void
146   addClangTargetOptions(const llvm::opt::ArgList &DriverArgs,
147                         llvm::opt::ArgStringList &CC1Args,
148                         Action::OffloadKind DeviceOffloadKind) const override;
149 
150   // Never try to use the integrated assembler with CUDA; always fork out to
151   // ptxas.
152   bool useIntegratedAs() const override { return false; }
153   bool isCrossCompiling() const override { return true; }
154   bool isPICDefault() const override { return false; }
155   bool isPIEDefault(const llvm::opt::ArgList &Args) const override {
156     return false;
157   }
158   bool HasNativeLLVMSupport() const override { return true; }
159   bool isPICDefaultForced() const override { return false; }
160   bool SupportsProfiling() const override { return false; }
161 
162   bool IsMathErrnoDefault() const override { return false; }
163 
164   bool supportsDebugInfoOption(const llvm::opt::Arg *A) const override;
165   void adjustDebugInfoKind(llvm::codegenoptions::DebugInfoKind &DebugInfoKind,
166                            const llvm::opt::ArgList &Args) const override;
167 
168   // NVPTX supports only DWARF2.
169   unsigned GetDefaultDwarfVersion() const override { return 2; }
170   unsigned getMaxDwarfVersion() const override { return 2; }
171 
172   /// Uses nvptx-arch tool to get arch of the system GPU. Will return error
173   /// if unable to find one.
174   virtual Expected<SmallVector<std::string>>
175   getSystemGPUArchs(const llvm::opt::ArgList &Args) const override;
176 
177   CudaInstallationDetector CudaInstallation;
178 
179 protected:
180   Tool *buildAssembler() const override; // ptxas.
181   Tool *buildLinker() const override;    // nvlink.
182 
183 private:
184   bool Freestanding = false;
185 };
186 
187 class LLVM_LIBRARY_VISIBILITY CudaToolChain : public NVPTXToolChain {
188 public:
189   CudaToolChain(const Driver &D, const llvm::Triple &Triple,
190                 const ToolChain &HostTC, const llvm::opt::ArgList &Args);
191 
192   const llvm::Triple *getAuxTriple() const override {
193     return &HostTC.getTriple();
194   }
195 
196   bool HasNativeLLVMSupport() const override { return false; }
197 
198   std::string getInputFilename(const InputInfo &Input) const override;
199 
200   llvm::opt::DerivedArgList *
201   TranslateArgs(const llvm::opt::DerivedArgList &Args, StringRef BoundArch,
202                 Action::OffloadKind DeviceOffloadKind) const override;
203   void
204   addClangTargetOptions(const llvm::opt::ArgList &DriverArgs,
205                         llvm::opt::ArgStringList &CC1Args,
206                         Action::OffloadKind DeviceOffloadKind) const override;
207 
208   llvm::DenormalMode getDefaultDenormalModeForType(
209       const llvm::opt::ArgList &DriverArgs, const JobAction &JA,
210       const llvm::fltSemantics *FPType = nullptr) const override;
211 
212   void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs,
213                           llvm::opt::ArgStringList &CC1Args) const override;
214 
215   void addClangWarningOptions(llvm::opt::ArgStringList &CC1Args) const override;
216   CXXStdlibType GetCXXStdlibType(const llvm::opt::ArgList &Args) const override;
217   void
218   AddClangSystemIncludeArgs(const llvm::opt::ArgList &DriverArgs,
219                             llvm::opt::ArgStringList &CC1Args) const override;
220   void AddClangCXXStdlibIncludeArgs(
221       const llvm::opt::ArgList &Args,
222       llvm::opt::ArgStringList &CC1Args) const override;
223   void AddIAMCUIncludeArgs(const llvm::opt::ArgList &DriverArgs,
224                            llvm::opt::ArgStringList &CC1Args) const override;
225 
226   SanitizerMask getSupportedSanitizers() const override;
227 
228   VersionTuple
229   computeMSVCVersion(const Driver *D,
230                      const llvm::opt::ArgList &Args) const override;
231 
232   const ToolChain &HostTC;
233 
234 protected:
235   Tool *buildAssembler() const override; // ptxas
236   Tool *buildLinker() const override;    // fatbinary (ok, not really a linker)
237 };
238 
239 } // end namespace toolchains
240 } // end namespace driver
241 } // end namespace clang
242 
243 #endif // LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_CUDA_H
244