1 //===----- CGCUDANV.cpp - Interface to NVIDIA CUDA Runtime ----------------===//
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 // This provides a class for CUDA code generation targeting the NVIDIA CUDA
10 // runtime library.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "CGCUDARuntime.h"
15 #include "CodeGenFunction.h"
16 #include "CodeGenModule.h"
17 #include "clang/AST/Decl.h"
18 #include "clang/Basic/Cuda.h"
19 #include "clang/CodeGen/CodeGenABITypes.h"
20 #include "clang/CodeGen/ConstantInitBuilder.h"
21 #include "llvm/IR/BasicBlock.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/Support/Format.h"
25 
26 using namespace clang;
27 using namespace CodeGen;
28 
29 namespace {
30 constexpr unsigned CudaFatMagic = 0x466243b1;
31 constexpr unsigned HIPFatMagic = 0x48495046; // "HIPF"
32 
33 class CGNVCUDARuntime : public CGCUDARuntime {
34 
35 private:
36   llvm::IntegerType *IntTy, *SizeTy;
37   llvm::Type *VoidTy;
38   llvm::PointerType *CharPtrTy, *VoidPtrTy, *VoidPtrPtrTy;
39 
40   /// Convenience reference to LLVM Context
41   llvm::LLVMContext &Context;
42   /// Convenience reference to the current module
43   llvm::Module &TheModule;
44   /// Keeps track of kernel launch stubs emitted in this module
45   struct KernelInfo {
46     llvm::Function *Kernel;
47     const Decl *D;
48   };
49   llvm::SmallVector<KernelInfo, 16> EmittedKernels;
50   struct VarInfo {
51     llvm::GlobalVariable *Var;
52     const VarDecl *D;
53     DeviceVarFlags Flags;
54   };
55   llvm::SmallVector<VarInfo, 16> DeviceVars;
56   /// Keeps track of variable containing handle of GPU binary. Populated by
57   /// ModuleCtorFunction() and used to create corresponding cleanup calls in
58   /// ModuleDtorFunction()
59   llvm::GlobalVariable *GpuBinaryHandle = nullptr;
60   /// Whether we generate relocatable device code.
61   bool RelocatableDeviceCode;
62   /// Mangle context for device.
63   std::unique_ptr<MangleContext> DeviceMC;
64 
65   llvm::FunctionCallee getSetupArgumentFn() const;
66   llvm::FunctionCallee getLaunchFn() const;
67 
68   llvm::FunctionType *getRegisterGlobalsFnTy() const;
69   llvm::FunctionType *getCallbackFnTy() const;
70   llvm::FunctionType *getRegisterLinkedBinaryFnTy() const;
71   std::string addPrefixToName(StringRef FuncName) const;
72   std::string addUnderscoredPrefixToName(StringRef FuncName) const;
73 
74   /// Creates a function to register all kernel stubs generated in this module.
75   llvm::Function *makeRegisterGlobalsFn();
76 
77   /// Helper function that generates a constant string and returns a pointer to
78   /// the start of the string.  The result of this function can be used anywhere
79   /// where the C code specifies const char*.
80   llvm::Constant *makeConstantString(const std::string &Str,
81                                      const std::string &Name = "",
82                                      const std::string &SectionName = "",
83                                      unsigned Alignment = 0) {
84     llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
85                                llvm::ConstantInt::get(SizeTy, 0)};
86     auto ConstStr = CGM.GetAddrOfConstantCString(Str, Name.c_str());
87     llvm::GlobalVariable *GV =
88         cast<llvm::GlobalVariable>(ConstStr.getPointer());
89     if (!SectionName.empty()) {
90       GV->setSection(SectionName);
91       // Mark the address as used which make sure that this section isn't
92       // merged and we will really have it in the object file.
93       GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::None);
94     }
95     if (Alignment)
96       GV->setAlignment(llvm::Align(Alignment));
97 
98     return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
99                                                 ConstStr.getPointer(), Zeros);
100   }
101 
102   /// Helper function that generates an empty dummy function returning void.
103   llvm::Function *makeDummyFunction(llvm::FunctionType *FnTy) {
104     assert(FnTy->getReturnType()->isVoidTy() &&
105            "Can only generate dummy functions returning void!");
106     llvm::Function *DummyFunc = llvm::Function::Create(
107         FnTy, llvm::GlobalValue::InternalLinkage, "dummy", &TheModule);
108 
109     llvm::BasicBlock *DummyBlock =
110         llvm::BasicBlock::Create(Context, "", DummyFunc);
111     CGBuilderTy FuncBuilder(CGM, Context);
112     FuncBuilder.SetInsertPoint(DummyBlock);
113     FuncBuilder.CreateRetVoid();
114 
115     return DummyFunc;
116   }
117 
118   void emitDeviceStubBodyLegacy(CodeGenFunction &CGF, FunctionArgList &Args);
119   void emitDeviceStubBodyNew(CodeGenFunction &CGF, FunctionArgList &Args);
120   std::string getDeviceSideName(const NamedDecl *ND) override;
121 
122 public:
123   CGNVCUDARuntime(CodeGenModule &CGM);
124 
125   void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
126   void registerDeviceVar(const VarDecl *VD, llvm::GlobalVariable &Var,
127                          bool Extern, bool Constant) override {
128     DeviceVars.push_back({&Var,
129                           VD,
130                           {DeviceVarFlags::Variable, Extern, Constant,
131                            /*Normalized*/ false, /*Type*/ 0}});
132   }
133   void registerDeviceSurf(const VarDecl *VD, llvm::GlobalVariable &Var,
134                           bool Extern, int Type) override {
135     DeviceVars.push_back({&Var,
136                           VD,
137                           {DeviceVarFlags::Surface, Extern, /*Constant*/ false,
138                            /*Normalized*/ false, Type}});
139   }
140   void registerDeviceTex(const VarDecl *VD, llvm::GlobalVariable &Var,
141                          bool Extern, int Type, bool Normalized) override {
142     DeviceVars.push_back({&Var,
143                           VD,
144                           {DeviceVarFlags::Texture, Extern, /*Constant*/ false,
145                            Normalized, Type}});
146   }
147 
148   /// Creates module constructor function
149   llvm::Function *makeModuleCtorFunction() override;
150   /// Creates module destructor function
151   llvm::Function *makeModuleDtorFunction() override;
152 };
153 
154 }
155 
156 std::string CGNVCUDARuntime::addPrefixToName(StringRef FuncName) const {
157   if (CGM.getLangOpts().HIP)
158     return ((Twine("hip") + Twine(FuncName)).str());
159   return ((Twine("cuda") + Twine(FuncName)).str());
160 }
161 std::string
162 CGNVCUDARuntime::addUnderscoredPrefixToName(StringRef FuncName) const {
163   if (CGM.getLangOpts().HIP)
164     return ((Twine("__hip") + Twine(FuncName)).str());
165   return ((Twine("__cuda") + Twine(FuncName)).str());
166 }
167 
168 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
169     : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
170       TheModule(CGM.getModule()),
171       RelocatableDeviceCode(CGM.getLangOpts().GPURelocatableDeviceCode),
172       DeviceMC(CGM.getContext().createMangleContext(
173           CGM.getContext().getAuxTargetInfo())) {
174   CodeGen::CodeGenTypes &Types = CGM.getTypes();
175   ASTContext &Ctx = CGM.getContext();
176 
177   IntTy = CGM.IntTy;
178   SizeTy = CGM.SizeTy;
179   VoidTy = CGM.VoidTy;
180 
181   CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
182   VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
183   VoidPtrPtrTy = VoidPtrTy->getPointerTo();
184 }
185 
186 llvm::FunctionCallee CGNVCUDARuntime::getSetupArgumentFn() const {
187   // cudaError_t cudaSetupArgument(void *, size_t, size_t)
188   llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
189   return CGM.CreateRuntimeFunction(
190       llvm::FunctionType::get(IntTy, Params, false),
191       addPrefixToName("SetupArgument"));
192 }
193 
194 llvm::FunctionCallee CGNVCUDARuntime::getLaunchFn() const {
195   if (CGM.getLangOpts().HIP) {
196     // hipError_t hipLaunchByPtr(char *);
197     return CGM.CreateRuntimeFunction(
198         llvm::FunctionType::get(IntTy, CharPtrTy, false), "hipLaunchByPtr");
199   } else {
200     // cudaError_t cudaLaunch(char *);
201     return CGM.CreateRuntimeFunction(
202         llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
203   }
204 }
205 
206 llvm::FunctionType *CGNVCUDARuntime::getRegisterGlobalsFnTy() const {
207   return llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false);
208 }
209 
210 llvm::FunctionType *CGNVCUDARuntime::getCallbackFnTy() const {
211   return llvm::FunctionType::get(VoidTy, VoidPtrTy, false);
212 }
213 
214 llvm::FunctionType *CGNVCUDARuntime::getRegisterLinkedBinaryFnTy() const {
215   auto CallbackFnTy = getCallbackFnTy();
216   auto RegisterGlobalsFnTy = getRegisterGlobalsFnTy();
217   llvm::Type *Params[] = {RegisterGlobalsFnTy->getPointerTo(), VoidPtrTy,
218                           VoidPtrTy, CallbackFnTy->getPointerTo()};
219   return llvm::FunctionType::get(VoidTy, Params, false);
220 }
221 
222 std::string CGNVCUDARuntime::getDeviceSideName(const NamedDecl *ND) {
223   GlobalDecl GD;
224   // D could be either a kernel or a variable.
225   if (auto *FD = dyn_cast<FunctionDecl>(ND))
226     GD = GlobalDecl(FD, KernelReferenceKind::Kernel);
227   else
228     GD = GlobalDecl(ND);
229   std::string DeviceSideName;
230   if (DeviceMC->shouldMangleDeclName(ND)) {
231     SmallString<256> Buffer;
232     llvm::raw_svector_ostream Out(Buffer);
233     DeviceMC->mangleName(GD, Out);
234     DeviceSideName = std::string(Out.str());
235   } else
236     DeviceSideName = std::string(ND->getIdentifier()->getName());
237   return DeviceSideName;
238 }
239 
240 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
241                                      FunctionArgList &Args) {
242   EmittedKernels.push_back({CGF.CurFn, CGF.CurFuncDecl});
243   if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
244                          CudaFeature::CUDA_USES_NEW_LAUNCH) ||
245       CGF.getLangOpts().HIPUseNewLaunchAPI)
246     emitDeviceStubBodyNew(CGF, Args);
247   else
248     emitDeviceStubBodyLegacy(CGF, Args);
249 }
250 
251 // CUDA 9.0+ uses new way to launch kernels. Parameters are packed in a local
252 // array and kernels are launched using cudaLaunchKernel().
253 void CGNVCUDARuntime::emitDeviceStubBodyNew(CodeGenFunction &CGF,
254                                             FunctionArgList &Args) {
255   // Build the shadow stack entry at the very start of the function.
256 
257   // Calculate amount of space we will need for all arguments.  If we have no
258   // args, allocate a single pointer so we still have a valid pointer to the
259   // argument array that we can pass to runtime, even if it will be unused.
260   Address KernelArgs = CGF.CreateTempAlloca(
261       VoidPtrTy, CharUnits::fromQuantity(16), "kernel_args",
262       llvm::ConstantInt::get(SizeTy, std::max<size_t>(1, Args.size())));
263   // Store pointers to the arguments in a locally allocated launch_args.
264   for (unsigned i = 0; i < Args.size(); ++i) {
265     llvm::Value* VarPtr = CGF.GetAddrOfLocalVar(Args[i]).getPointer();
266     llvm::Value *VoidVarPtr = CGF.Builder.CreatePointerCast(VarPtr, VoidPtrTy);
267     CGF.Builder.CreateDefaultAlignedStore(
268         VoidVarPtr, CGF.Builder.CreateConstGEP1_32(KernelArgs.getPointer(), i));
269   }
270 
271   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
272 
273   // Lookup cudaLaunchKernel/hipLaunchKernel function.
274   // cudaError_t cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
275   //                              void **args, size_t sharedMem,
276   //                              cudaStream_t stream);
277   // hipError_t hipLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
278   //                            void **args, size_t sharedMem,
279   //                            hipStream_t stream);
280   TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
281   DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
282   auto LaunchKernelName = addPrefixToName("LaunchKernel");
283   IdentifierInfo &cudaLaunchKernelII =
284       CGM.getContext().Idents.get(LaunchKernelName);
285   FunctionDecl *cudaLaunchKernelFD = nullptr;
286   for (const auto &Result : DC->lookup(&cudaLaunchKernelII)) {
287     if (FunctionDecl *FD = dyn_cast<FunctionDecl>(Result))
288       cudaLaunchKernelFD = FD;
289   }
290 
291   if (cudaLaunchKernelFD == nullptr) {
292     CGM.Error(CGF.CurFuncDecl->getLocation(),
293               "Can't find declaration for " + LaunchKernelName);
294     return;
295   }
296   // Create temporary dim3 grid_dim, block_dim.
297   ParmVarDecl *GridDimParam = cudaLaunchKernelFD->getParamDecl(1);
298   QualType Dim3Ty = GridDimParam->getType();
299   Address GridDim =
300       CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "grid_dim");
301   Address BlockDim =
302       CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "block_dim");
303   Address ShmemSize =
304       CGF.CreateTempAlloca(SizeTy, CGM.getSizeAlign(), "shmem_size");
305   Address Stream =
306       CGF.CreateTempAlloca(VoidPtrTy, CGM.getPointerAlign(), "stream");
307   llvm::FunctionCallee cudaPopConfigFn = CGM.CreateRuntimeFunction(
308       llvm::FunctionType::get(IntTy,
309                               {/*gridDim=*/GridDim.getType(),
310                                /*blockDim=*/BlockDim.getType(),
311                                /*ShmemSize=*/ShmemSize.getType(),
312                                /*Stream=*/Stream.getType()},
313                               /*isVarArg=*/false),
314       addUnderscoredPrefixToName("PopCallConfiguration"));
315 
316   CGF.EmitRuntimeCallOrInvoke(cudaPopConfigFn,
317                               {GridDim.getPointer(), BlockDim.getPointer(),
318                                ShmemSize.getPointer(), Stream.getPointer()});
319 
320   // Emit the call to cudaLaunch
321   llvm::Value *Kernel = CGF.Builder.CreatePointerCast(CGF.CurFn, VoidPtrTy);
322   CallArgList LaunchKernelArgs;
323   LaunchKernelArgs.add(RValue::get(Kernel),
324                        cudaLaunchKernelFD->getParamDecl(0)->getType());
325   LaunchKernelArgs.add(RValue::getAggregate(GridDim), Dim3Ty);
326   LaunchKernelArgs.add(RValue::getAggregate(BlockDim), Dim3Ty);
327   LaunchKernelArgs.add(RValue::get(KernelArgs.getPointer()),
328                        cudaLaunchKernelFD->getParamDecl(3)->getType());
329   LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(ShmemSize)),
330                        cudaLaunchKernelFD->getParamDecl(4)->getType());
331   LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(Stream)),
332                        cudaLaunchKernelFD->getParamDecl(5)->getType());
333 
334   QualType QT = cudaLaunchKernelFD->getType();
335   QualType CQT = QT.getCanonicalType();
336   llvm::Type *Ty = CGM.getTypes().ConvertType(CQT);
337   llvm::FunctionType *FTy = dyn_cast<llvm::FunctionType>(Ty);
338 
339   const CGFunctionInfo &FI =
340       CGM.getTypes().arrangeFunctionDeclaration(cudaLaunchKernelFD);
341   llvm::FunctionCallee cudaLaunchKernelFn =
342       CGM.CreateRuntimeFunction(FTy, LaunchKernelName);
343   CGF.EmitCall(FI, CGCallee::forDirect(cudaLaunchKernelFn), ReturnValueSlot(),
344                LaunchKernelArgs);
345   CGF.EmitBranch(EndBlock);
346 
347   CGF.EmitBlock(EndBlock);
348 }
349 
350 void CGNVCUDARuntime::emitDeviceStubBodyLegacy(CodeGenFunction &CGF,
351                                                FunctionArgList &Args) {
352   // Emit a call to cudaSetupArgument for each arg in Args.
353   llvm::FunctionCallee cudaSetupArgFn = getSetupArgumentFn();
354   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
355   CharUnits Offset = CharUnits::Zero();
356   for (const VarDecl *A : Args) {
357     CharUnits TyWidth, TyAlign;
358     std::tie(TyWidth, TyAlign) =
359         CGM.getContext().getTypeInfoInChars(A->getType());
360     Offset = Offset.alignTo(TyAlign);
361     llvm::Value *Args[] = {
362         CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
363                                       VoidPtrTy),
364         llvm::ConstantInt::get(SizeTy, TyWidth.getQuantity()),
365         llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
366     };
367     llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
368     llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
369     llvm::Value *CBZero = CGF.Builder.CreateICmpEQ(CB, Zero);
370     llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
371     CGF.Builder.CreateCondBr(CBZero, NextBlock, EndBlock);
372     CGF.EmitBlock(NextBlock);
373     Offset += TyWidth;
374   }
375 
376   // Emit the call to cudaLaunch
377   llvm::FunctionCallee cudaLaunchFn = getLaunchFn();
378   llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
379   CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
380   CGF.EmitBranch(EndBlock);
381 
382   CGF.EmitBlock(EndBlock);
383 }
384 
385 /// Creates a function that sets up state on the host side for CUDA objects that
386 /// have a presence on both the host and device sides. Specifically, registers
387 /// the host side of kernel functions and device global variables with the CUDA
388 /// runtime.
389 /// \code
390 /// void __cuda_register_globals(void** GpuBinaryHandle) {
391 ///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
392 ///    ...
393 ///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
394 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
395 ///    ...
396 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
397 /// }
398 /// \endcode
399 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
400   // No need to register anything
401   if (EmittedKernels.empty() && DeviceVars.empty())
402     return nullptr;
403 
404   llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
405       getRegisterGlobalsFnTy(), llvm::GlobalValue::InternalLinkage,
406       addUnderscoredPrefixToName("_register_globals"), &TheModule);
407   llvm::BasicBlock *EntryBB =
408       llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
409   CGBuilderTy Builder(CGM, Context);
410   Builder.SetInsertPoint(EntryBB);
411 
412   // void __cudaRegisterFunction(void **, const char *, char *, const char *,
413   //                             int, uint3*, uint3*, dim3*, dim3*, int*)
414   llvm::Type *RegisterFuncParams[] = {
415       VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
416       VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
417   llvm::FunctionCallee RegisterFunc = CGM.CreateRuntimeFunction(
418       llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
419       addUnderscoredPrefixToName("RegisterFunction"));
420 
421   // Extract GpuBinaryHandle passed as the first argument passed to
422   // __cuda_register_globals() and generate __cudaRegisterFunction() call for
423   // each emitted kernel.
424   llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
425   for (auto &&I : EmittedKernels) {
426     llvm::Constant *KernelName =
427         makeConstantString(getDeviceSideName(cast<NamedDecl>(I.D)));
428     llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
429     llvm::Value *Args[] = {
430         &GpuBinaryHandlePtr,
431         Builder.CreateBitCast(I.Kernel, VoidPtrTy),
432         KernelName,
433         KernelName,
434         llvm::ConstantInt::get(IntTy, -1),
435         NullPtr,
436         NullPtr,
437         NullPtr,
438         NullPtr,
439         llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
440     Builder.CreateCall(RegisterFunc, Args);
441   }
442 
443   // void __cudaRegisterVar(void **, char *, char *, const char *,
444   //                        int, int, int, int)
445   llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
446                                      CharPtrTy,    IntTy,     IntTy,
447                                      IntTy,        IntTy};
448   llvm::FunctionCallee RegisterVar = CGM.CreateRuntimeFunction(
449       llvm::FunctionType::get(IntTy, RegisterVarParams, false),
450       addUnderscoredPrefixToName("RegisterVar"));
451   // void __cudaRegisterSurface(void **, const struct surfaceReference *,
452   //                            const void **, const char *, int, int);
453   llvm::FunctionCallee RegisterSurf = CGM.CreateRuntimeFunction(
454       llvm::FunctionType::get(
455           VoidTy, {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy},
456           false),
457       addUnderscoredPrefixToName("RegisterSurface"));
458   // void __cudaRegisterTexture(void **, const struct textureReference *,
459   //                            const void **, const char *, int, int, int)
460   llvm::FunctionCallee RegisterTex = CGM.CreateRuntimeFunction(
461       llvm::FunctionType::get(
462           VoidTy,
463           {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy, IntTy},
464           false),
465       addUnderscoredPrefixToName("RegisterTexture"));
466   for (auto &&Info : DeviceVars) {
467     llvm::GlobalVariable *Var = Info.Var;
468     llvm::Constant *VarName = makeConstantString(getDeviceSideName(Info.D));
469     switch (Info.Flags.Kind) {
470     case DeviceVarFlags::Variable: {
471       uint64_t VarSize =
472           CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
473       llvm::Value *Args[] = {&GpuBinaryHandlePtr,
474                              Builder.CreateBitCast(Var, VoidPtrTy),
475                              VarName,
476                              VarName,
477                              llvm::ConstantInt::get(IntTy, Info.Flags.Extern),
478                              llvm::ConstantInt::get(IntTy, VarSize),
479                              llvm::ConstantInt::get(IntTy, Info.Flags.Constant),
480                              llvm::ConstantInt::get(IntTy, 0)};
481       Builder.CreateCall(RegisterVar, Args);
482       break;
483     }
484     case DeviceVarFlags::Surface:
485       Builder.CreateCall(
486           RegisterSurf,
487           {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
488            VarName, llvm::ConstantInt::get(IntTy, Info.Flags.SurfTexType),
489            llvm::ConstantInt::get(IntTy, Info.Flags.Extern)});
490       break;
491     case DeviceVarFlags::Texture:
492       Builder.CreateCall(
493           RegisterTex,
494           {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
495            VarName, llvm::ConstantInt::get(IntTy, Info.Flags.SurfTexType),
496            llvm::ConstantInt::get(IntTy, Info.Flags.Normalized),
497            llvm::ConstantInt::get(IntTy, Info.Flags.Extern)});
498       break;
499     }
500   }
501 
502   Builder.CreateRetVoid();
503   return RegisterKernelsFunc;
504 }
505 
506 /// Creates a global constructor function for the module:
507 ///
508 /// For CUDA:
509 /// \code
510 /// void __cuda_module_ctor(void*) {
511 ///     Handle = __cudaRegisterFatBinary(GpuBinaryBlob);
512 ///     __cuda_register_globals(Handle);
513 /// }
514 /// \endcode
515 ///
516 /// For HIP:
517 /// \code
518 /// void __hip_module_ctor(void*) {
519 ///     if (__hip_gpubin_handle == 0) {
520 ///         __hip_gpubin_handle  = __hipRegisterFatBinary(GpuBinaryBlob);
521 ///         __hip_register_globals(__hip_gpubin_handle);
522 ///     }
523 /// }
524 /// \endcode
525 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
526   bool IsHIP = CGM.getLangOpts().HIP;
527   bool IsCUDA = CGM.getLangOpts().CUDA;
528   // No need to generate ctors/dtors if there is no GPU binary.
529   StringRef CudaGpuBinaryFileName = CGM.getCodeGenOpts().CudaGpuBinaryFileName;
530   if (CudaGpuBinaryFileName.empty() && !IsHIP)
531     return nullptr;
532   if ((IsHIP || (IsCUDA && !RelocatableDeviceCode)) && EmittedKernels.empty() &&
533       DeviceVars.empty())
534     return nullptr;
535 
536   // void __{cuda|hip}_register_globals(void* handle);
537   llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
538   // We always need a function to pass in as callback. Create a dummy
539   // implementation if we don't need to register anything.
540   if (RelocatableDeviceCode && !RegisterGlobalsFunc)
541     RegisterGlobalsFunc = makeDummyFunction(getRegisterGlobalsFnTy());
542 
543   // void ** __{cuda|hip}RegisterFatBinary(void *);
544   llvm::FunctionCallee RegisterFatbinFunc = CGM.CreateRuntimeFunction(
545       llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
546       addUnderscoredPrefixToName("RegisterFatBinary"));
547   // struct { int magic, int version, void * gpu_binary, void * dont_care };
548   llvm::StructType *FatbinWrapperTy =
549       llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy);
550 
551   // Register GPU binary with the CUDA runtime, store returned handle in a
552   // global variable and save a reference in GpuBinaryHandle to be cleaned up
553   // in destructor on exit. Then associate all known kernels with the GPU binary
554   // handle so CUDA runtime can figure out what to call on the GPU side.
555   std::unique_ptr<llvm::MemoryBuffer> CudaGpuBinary = nullptr;
556   if (!CudaGpuBinaryFileName.empty()) {
557     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CudaGpuBinaryOrErr =
558         llvm::MemoryBuffer::getFileOrSTDIN(CudaGpuBinaryFileName);
559     if (std::error_code EC = CudaGpuBinaryOrErr.getError()) {
560       CGM.getDiags().Report(diag::err_cannot_open_file)
561           << CudaGpuBinaryFileName << EC.message();
562       return nullptr;
563     }
564     CudaGpuBinary = std::move(CudaGpuBinaryOrErr.get());
565   }
566 
567   llvm::Function *ModuleCtorFunc = llvm::Function::Create(
568       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
569       llvm::GlobalValue::InternalLinkage,
570       addUnderscoredPrefixToName("_module_ctor"), &TheModule);
571   llvm::BasicBlock *CtorEntryBB =
572       llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
573   CGBuilderTy CtorBuilder(CGM, Context);
574 
575   CtorBuilder.SetInsertPoint(CtorEntryBB);
576 
577   const char *FatbinConstantName;
578   const char *FatbinSectionName;
579   const char *ModuleIDSectionName;
580   StringRef ModuleIDPrefix;
581   llvm::Constant *FatBinStr;
582   unsigned FatMagic;
583   if (IsHIP) {
584     FatbinConstantName = ".hip_fatbin";
585     FatbinSectionName = ".hipFatBinSegment";
586 
587     ModuleIDSectionName = "__hip_module_id";
588     ModuleIDPrefix = "__hip_";
589 
590     if (CudaGpuBinary) {
591       // If fatbin is available from early finalization, create a string
592       // literal containing the fat binary loaded from the given file.
593       FatBinStr = makeConstantString(std::string(CudaGpuBinary->getBuffer()),
594                                      "", FatbinConstantName, 8);
595     } else {
596       // If fatbin is not available, create an external symbol
597       // __hip_fatbin in section .hip_fatbin. The external symbol is supposed
598       // to contain the fat binary but will be populated somewhere else,
599       // e.g. by lld through link script.
600       FatBinStr = new llvm::GlobalVariable(
601         CGM.getModule(), CGM.Int8Ty,
602         /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
603         "__hip_fatbin", nullptr,
604         llvm::GlobalVariable::NotThreadLocal);
605       cast<llvm::GlobalVariable>(FatBinStr)->setSection(FatbinConstantName);
606     }
607 
608     FatMagic = HIPFatMagic;
609   } else {
610     if (RelocatableDeviceCode)
611       FatbinConstantName = CGM.getTriple().isMacOSX()
612                                ? "__NV_CUDA,__nv_relfatbin"
613                                : "__nv_relfatbin";
614     else
615       FatbinConstantName =
616           CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
617     // NVIDIA's cuobjdump looks for fatbins in this section.
618     FatbinSectionName =
619         CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
620 
621     ModuleIDSectionName = CGM.getTriple().isMacOSX()
622                               ? "__NV_CUDA,__nv_module_id"
623                               : "__nv_module_id";
624     ModuleIDPrefix = "__nv_";
625 
626     // For CUDA, create a string literal containing the fat binary loaded from
627     // the given file.
628     FatBinStr = makeConstantString(std::string(CudaGpuBinary->getBuffer()), "",
629                                    FatbinConstantName, 8);
630     FatMagic = CudaFatMagic;
631   }
632 
633   // Create initialized wrapper structure that points to the loaded GPU binary
634   ConstantInitBuilder Builder(CGM);
635   auto Values = Builder.beginStruct(FatbinWrapperTy);
636   // Fatbin wrapper magic.
637   Values.addInt(IntTy, FatMagic);
638   // Fatbin version.
639   Values.addInt(IntTy, 1);
640   // Data.
641   Values.add(FatBinStr);
642   // Unused in fatbin v1.
643   Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
644   llvm::GlobalVariable *FatbinWrapper = Values.finishAndCreateGlobal(
645       addUnderscoredPrefixToName("_fatbin_wrapper"), CGM.getPointerAlign(),
646       /*constant*/ true);
647   FatbinWrapper->setSection(FatbinSectionName);
648 
649   // There is only one HIP fat binary per linked module, however there are
650   // multiple constructor functions. Make sure the fat binary is registered
651   // only once. The constructor functions are executed by the dynamic loader
652   // before the program gains control. The dynamic loader cannot execute the
653   // constructor functions concurrently since doing that would not guarantee
654   // thread safety of the loaded program. Therefore we can assume sequential
655   // execution of constructor functions here.
656   if (IsHIP) {
657     auto Linkage = CudaGpuBinary ? llvm::GlobalValue::InternalLinkage :
658         llvm::GlobalValue::LinkOnceAnyLinkage;
659     llvm::BasicBlock *IfBlock =
660         llvm::BasicBlock::Create(Context, "if", ModuleCtorFunc);
661     llvm::BasicBlock *ExitBlock =
662         llvm::BasicBlock::Create(Context, "exit", ModuleCtorFunc);
663     // The name, size, and initialization pattern of this variable is part
664     // of HIP ABI.
665     GpuBinaryHandle = new llvm::GlobalVariable(
666         TheModule, VoidPtrPtrTy, /*isConstant=*/false,
667         Linkage,
668         /*Initializer=*/llvm::ConstantPointerNull::get(VoidPtrPtrTy),
669         "__hip_gpubin_handle");
670     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
671     // Prevent the weak symbol in different shared libraries being merged.
672     if (Linkage != llvm::GlobalValue::InternalLinkage)
673       GpuBinaryHandle->setVisibility(llvm::GlobalValue::HiddenVisibility);
674     Address GpuBinaryAddr(
675         GpuBinaryHandle,
676         CharUnits::fromQuantity(GpuBinaryHandle->getAlignment()));
677     {
678       auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
679       llvm::Constant *Zero =
680           llvm::Constant::getNullValue(HandleValue->getType());
681       llvm::Value *EQZero = CtorBuilder.CreateICmpEQ(HandleValue, Zero);
682       CtorBuilder.CreateCondBr(EQZero, IfBlock, ExitBlock);
683     }
684     {
685       CtorBuilder.SetInsertPoint(IfBlock);
686       // GpuBinaryHandle = __hipRegisterFatBinary(&FatbinWrapper);
687       llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
688           RegisterFatbinFunc,
689           CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
690       CtorBuilder.CreateStore(RegisterFatbinCall, GpuBinaryAddr);
691       CtorBuilder.CreateBr(ExitBlock);
692     }
693     {
694       CtorBuilder.SetInsertPoint(ExitBlock);
695       // Call __hip_register_globals(GpuBinaryHandle);
696       if (RegisterGlobalsFunc) {
697         auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
698         CtorBuilder.CreateCall(RegisterGlobalsFunc, HandleValue);
699       }
700     }
701   } else if (!RelocatableDeviceCode) {
702     // Register binary with CUDA runtime. This is substantially different in
703     // default mode vs. separate compilation!
704     // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
705     llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
706         RegisterFatbinFunc,
707         CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
708     GpuBinaryHandle = new llvm::GlobalVariable(
709         TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
710         llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
711     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
712     CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
713                                    CGM.getPointerAlign());
714 
715     // Call __cuda_register_globals(GpuBinaryHandle);
716     if (RegisterGlobalsFunc)
717       CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
718 
719     // Call __cudaRegisterFatBinaryEnd(Handle) if this CUDA version needs it.
720     if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
721                            CudaFeature::CUDA_USES_FATBIN_REGISTER_END)) {
722       // void __cudaRegisterFatBinaryEnd(void **);
723       llvm::FunctionCallee RegisterFatbinEndFunc = CGM.CreateRuntimeFunction(
724           llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
725           "__cudaRegisterFatBinaryEnd");
726       CtorBuilder.CreateCall(RegisterFatbinEndFunc, RegisterFatbinCall);
727     }
728   } else {
729     // Generate a unique module ID.
730     SmallString<64> ModuleID;
731     llvm::raw_svector_ostream OS(ModuleID);
732     OS << ModuleIDPrefix << llvm::format("%" PRIx64, FatbinWrapper->getGUID());
733     llvm::Constant *ModuleIDConstant = makeConstantString(
734         std::string(ModuleID.str()), "", ModuleIDSectionName, 32);
735 
736     // Create an alias for the FatbinWrapper that nvcc will look for.
737     llvm::GlobalAlias::create(llvm::GlobalValue::ExternalLinkage,
738                               Twine("__fatbinwrap") + ModuleID, FatbinWrapper);
739 
740     // void __cudaRegisterLinkedBinary%ModuleID%(void (*)(void *), void *,
741     // void *, void (*)(void **))
742     SmallString<128> RegisterLinkedBinaryName("__cudaRegisterLinkedBinary");
743     RegisterLinkedBinaryName += ModuleID;
744     llvm::FunctionCallee RegisterLinkedBinaryFunc = CGM.CreateRuntimeFunction(
745         getRegisterLinkedBinaryFnTy(), RegisterLinkedBinaryName);
746 
747     assert(RegisterGlobalsFunc && "Expecting at least dummy function!");
748     llvm::Value *Args[] = {RegisterGlobalsFunc,
749                            CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy),
750                            ModuleIDConstant,
751                            makeDummyFunction(getCallbackFnTy())};
752     CtorBuilder.CreateCall(RegisterLinkedBinaryFunc, Args);
753   }
754 
755   // Create destructor and register it with atexit() the way NVCC does it. Doing
756   // it during regular destructor phase worked in CUDA before 9.2 but results in
757   // double-free in 9.2.
758   if (llvm::Function *CleanupFn = makeModuleDtorFunction()) {
759     // extern "C" int atexit(void (*f)(void));
760     llvm::FunctionType *AtExitTy =
761         llvm::FunctionType::get(IntTy, CleanupFn->getType(), false);
762     llvm::FunctionCallee AtExitFunc =
763         CGM.CreateRuntimeFunction(AtExitTy, "atexit", llvm::AttributeList(),
764                                   /*Local=*/true);
765     CtorBuilder.CreateCall(AtExitFunc, CleanupFn);
766   }
767 
768   CtorBuilder.CreateRetVoid();
769   return ModuleCtorFunc;
770 }
771 
772 /// Creates a global destructor function that unregisters the GPU code blob
773 /// registered by constructor.
774 ///
775 /// For CUDA:
776 /// \code
777 /// void __cuda_module_dtor(void*) {
778 ///     __cudaUnregisterFatBinary(Handle);
779 /// }
780 /// \endcode
781 ///
782 /// For HIP:
783 /// \code
784 /// void __hip_module_dtor(void*) {
785 ///     if (__hip_gpubin_handle) {
786 ///         __hipUnregisterFatBinary(__hip_gpubin_handle);
787 ///         __hip_gpubin_handle = 0;
788 ///     }
789 /// }
790 /// \endcode
791 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
792   // No need for destructor if we don't have a handle to unregister.
793   if (!GpuBinaryHandle)
794     return nullptr;
795 
796   // void __cudaUnregisterFatBinary(void ** handle);
797   llvm::FunctionCallee UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
798       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
799       addUnderscoredPrefixToName("UnregisterFatBinary"));
800 
801   llvm::Function *ModuleDtorFunc = llvm::Function::Create(
802       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
803       llvm::GlobalValue::InternalLinkage,
804       addUnderscoredPrefixToName("_module_dtor"), &TheModule);
805 
806   llvm::BasicBlock *DtorEntryBB =
807       llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
808   CGBuilderTy DtorBuilder(CGM, Context);
809   DtorBuilder.SetInsertPoint(DtorEntryBB);
810 
811   Address GpuBinaryAddr(GpuBinaryHandle, CharUnits::fromQuantity(
812                                              GpuBinaryHandle->getAlignment()));
813   auto HandleValue = DtorBuilder.CreateLoad(GpuBinaryAddr);
814   // There is only one HIP fat binary per linked module, however there are
815   // multiple destructor functions. Make sure the fat binary is unregistered
816   // only once.
817   if (CGM.getLangOpts().HIP) {
818     llvm::BasicBlock *IfBlock =
819         llvm::BasicBlock::Create(Context, "if", ModuleDtorFunc);
820     llvm::BasicBlock *ExitBlock =
821         llvm::BasicBlock::Create(Context, "exit", ModuleDtorFunc);
822     llvm::Constant *Zero = llvm::Constant::getNullValue(HandleValue->getType());
823     llvm::Value *NEZero = DtorBuilder.CreateICmpNE(HandleValue, Zero);
824     DtorBuilder.CreateCondBr(NEZero, IfBlock, ExitBlock);
825 
826     DtorBuilder.SetInsertPoint(IfBlock);
827     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
828     DtorBuilder.CreateStore(Zero, GpuBinaryAddr);
829     DtorBuilder.CreateBr(ExitBlock);
830 
831     DtorBuilder.SetInsertPoint(ExitBlock);
832   } else {
833     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
834   }
835   DtorBuilder.CreateRetVoid();
836   return ModuleDtorFunc;
837 }
838 
839 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
840   return new CGNVCUDARuntime(CGM);
841 }
842