1 //===----- CGCUDANV.cpp - Interface to NVIDIA CUDA Runtime ----------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This provides a class for CUDA code generation targeting the NVIDIA CUDA
11 // runtime library.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "CGCUDARuntime.h"
16 #include "CodeGenFunction.h"
17 #include "CodeGenModule.h"
18 #include "clang/AST/Decl.h"
19 #include "clang/CodeGen/ConstantInitBuilder.h"
20 #include "llvm/IR/BasicBlock.h"
21 #include "llvm/IR/CallSite.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   llvm::SmallVector<llvm::Function *, 16> EmittedKernels;
46   llvm::SmallVector<std::pair<llvm::GlobalVariable *, unsigned>, 16> DeviceVars;
47   /// Keeps track of variable containing handle of GPU binary. Populated by
48   /// ModuleCtorFunction() and used to create corresponding cleanup calls in
49   /// ModuleDtorFunction()
50   llvm::GlobalVariable *GpuBinaryHandle = nullptr;
51   /// Whether we generate relocatable device code.
52   bool RelocatableDeviceCode;
53 
54   llvm::Constant *getSetupArgumentFn() const;
55   llvm::Constant *getLaunchFn() const;
56 
57   llvm::FunctionType *getRegisterGlobalsFnTy() const;
58   llvm::FunctionType *getCallbackFnTy() const;
59   llvm::FunctionType *getRegisterLinkedBinaryFnTy() const;
60   std::string addPrefixToName(StringRef FuncName) const;
61   std::string addUnderscoredPrefixToName(StringRef FuncName) const;
62 
63   /// Creates a function to register all kernel stubs generated in this module.
64   llvm::Function *makeRegisterGlobalsFn();
65 
66   /// Helper function that generates a constant string and returns a pointer to
67   /// the start of the string.  The result of this function can be used anywhere
68   /// where the C code specifies const char*.
69   llvm::Constant *makeConstantString(const std::string &Str,
70                                      const std::string &Name = "",
71                                      const std::string &SectionName = "",
72                                      unsigned Alignment = 0) {
73     llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
74                                llvm::ConstantInt::get(SizeTy, 0)};
75     auto ConstStr = CGM.GetAddrOfConstantCString(Str, Name.c_str());
76     llvm::GlobalVariable *GV =
77         cast<llvm::GlobalVariable>(ConstStr.getPointer());
78     if (!SectionName.empty()) {
79       GV->setSection(SectionName);
80       // Mark the address as used which make sure that this section isn't
81       // merged and we will really have it in the object file.
82       GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::None);
83     }
84     if (Alignment)
85       GV->setAlignment(Alignment);
86 
87     return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
88                                                 ConstStr.getPointer(), Zeros);
89   }
90 
91   /// Helper function that generates an empty dummy function returning void.
92   llvm::Function *makeDummyFunction(llvm::FunctionType *FnTy) {
93     assert(FnTy->getReturnType()->isVoidTy() &&
94            "Can only generate dummy functions returning void!");
95     llvm::Function *DummyFunc = llvm::Function::Create(
96         FnTy, llvm::GlobalValue::InternalLinkage, "dummy", &TheModule);
97 
98     llvm::BasicBlock *DummyBlock =
99         llvm::BasicBlock::Create(Context, "", DummyFunc);
100     CGBuilderTy FuncBuilder(CGM, Context);
101     FuncBuilder.SetInsertPoint(DummyBlock);
102     FuncBuilder.CreateRetVoid();
103 
104     return DummyFunc;
105   }
106 
107   void emitDeviceStubBody(CodeGenFunction &CGF, FunctionArgList &Args);
108 
109 public:
110   CGNVCUDARuntime(CodeGenModule &CGM);
111 
112   void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
113   void registerDeviceVar(llvm::GlobalVariable &Var, unsigned Flags) override {
114     DeviceVars.push_back(std::make_pair(&Var, Flags));
115   }
116 
117   /// Creates module constructor function
118   llvm::Function *makeModuleCtorFunction() override;
119   /// Creates module destructor function
120   llvm::Function *makeModuleDtorFunction() override;
121 };
122 
123 }
124 
125 std::string CGNVCUDARuntime::addPrefixToName(StringRef FuncName) const {
126   if (CGM.getLangOpts().HIP)
127     return ((Twine("hip") + Twine(FuncName)).str());
128   return ((Twine("cuda") + Twine(FuncName)).str());
129 }
130 std::string
131 CGNVCUDARuntime::addUnderscoredPrefixToName(StringRef FuncName) const {
132   if (CGM.getLangOpts().HIP)
133     return ((Twine("__hip") + Twine(FuncName)).str());
134   return ((Twine("__cuda") + Twine(FuncName)).str());
135 }
136 
137 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
138     : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
139       TheModule(CGM.getModule()),
140       RelocatableDeviceCode(CGM.getLangOpts().GPURelocatableDeviceCode) {
141   CodeGen::CodeGenTypes &Types = CGM.getTypes();
142   ASTContext &Ctx = CGM.getContext();
143 
144   IntTy = CGM.IntTy;
145   SizeTy = CGM.SizeTy;
146   VoidTy = CGM.VoidTy;
147 
148   CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
149   VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
150   VoidPtrPtrTy = VoidPtrTy->getPointerTo();
151 }
152 
153 llvm::Constant *CGNVCUDARuntime::getSetupArgumentFn() const {
154   // cudaError_t cudaSetupArgument(void *, size_t, size_t)
155   llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
156   return CGM.CreateRuntimeFunction(
157       llvm::FunctionType::get(IntTy, Params, false),
158       addPrefixToName("SetupArgument"));
159 }
160 
161 llvm::Constant *CGNVCUDARuntime::getLaunchFn() const {
162   if (CGM.getLangOpts().HIP) {
163     // hipError_t hipLaunchByPtr(char *);
164     return CGM.CreateRuntimeFunction(
165         llvm::FunctionType::get(IntTy, CharPtrTy, false), "hipLaunchByPtr");
166   } else {
167     // cudaError_t cudaLaunch(char *);
168     return CGM.CreateRuntimeFunction(
169         llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
170   }
171 }
172 
173 llvm::FunctionType *CGNVCUDARuntime::getRegisterGlobalsFnTy() const {
174   return llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false);
175 }
176 
177 llvm::FunctionType *CGNVCUDARuntime::getCallbackFnTy() const {
178   return llvm::FunctionType::get(VoidTy, VoidPtrTy, false);
179 }
180 
181 llvm::FunctionType *CGNVCUDARuntime::getRegisterLinkedBinaryFnTy() const {
182   auto CallbackFnTy = getCallbackFnTy();
183   auto RegisterGlobalsFnTy = getRegisterGlobalsFnTy();
184   llvm::Type *Params[] = {RegisterGlobalsFnTy->getPointerTo(), VoidPtrTy,
185                           VoidPtrTy, CallbackFnTy->getPointerTo()};
186   return llvm::FunctionType::get(VoidTy, Params, false);
187 }
188 
189 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
190                                      FunctionArgList &Args) {
191   EmittedKernels.push_back(CGF.CurFn);
192   emitDeviceStubBody(CGF, Args);
193 }
194 
195 void CGNVCUDARuntime::emitDeviceStubBody(CodeGenFunction &CGF,
196                                          FunctionArgList &Args) {
197   // Emit a call to cudaSetupArgument for each arg in Args.
198   llvm::Constant *cudaSetupArgFn = getSetupArgumentFn();
199   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
200   CharUnits Offset = CharUnits::Zero();
201   for (const VarDecl *A : Args) {
202     CharUnits TyWidth, TyAlign;
203     std::tie(TyWidth, TyAlign) =
204         CGM.getContext().getTypeInfoInChars(A->getType());
205     Offset = Offset.alignTo(TyAlign);
206     llvm::Value *Args[] = {
207         CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
208                                       VoidPtrTy),
209         llvm::ConstantInt::get(SizeTy, TyWidth.getQuantity()),
210         llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
211     };
212     llvm::CallSite CS = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
213     llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
214     llvm::Value *CSZero = CGF.Builder.CreateICmpEQ(CS.getInstruction(), Zero);
215     llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
216     CGF.Builder.CreateCondBr(CSZero, NextBlock, EndBlock);
217     CGF.EmitBlock(NextBlock);
218     Offset += TyWidth;
219   }
220 
221   // Emit the call to cudaLaunch
222   llvm::Constant *cudaLaunchFn = getLaunchFn();
223   llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
224   CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
225   CGF.EmitBranch(EndBlock);
226 
227   CGF.EmitBlock(EndBlock);
228 }
229 
230 /// Creates a function that sets up state on the host side for CUDA objects that
231 /// have a presence on both the host and device sides. Specifically, registers
232 /// the host side of kernel functions and device global variables with the CUDA
233 /// runtime.
234 /// \code
235 /// void __cuda_register_globals(void** GpuBinaryHandle) {
236 ///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
237 ///    ...
238 ///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
239 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
240 ///    ...
241 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
242 /// }
243 /// \endcode
244 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
245   // No need to register anything
246   if (EmittedKernels.empty() && DeviceVars.empty())
247     return nullptr;
248 
249   llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
250       getRegisterGlobalsFnTy(), llvm::GlobalValue::InternalLinkage,
251       addUnderscoredPrefixToName("_register_globals"), &TheModule);
252   llvm::BasicBlock *EntryBB =
253       llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
254   CGBuilderTy Builder(CGM, Context);
255   Builder.SetInsertPoint(EntryBB);
256 
257   // void __cudaRegisterFunction(void **, const char *, char *, const char *,
258   //                             int, uint3*, uint3*, dim3*, dim3*, int*)
259   llvm::Type *RegisterFuncParams[] = {
260       VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
261       VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
262   llvm::Constant *RegisterFunc = CGM.CreateRuntimeFunction(
263       llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
264       addUnderscoredPrefixToName("RegisterFunction"));
265 
266   // Extract GpuBinaryHandle passed as the first argument passed to
267   // __cuda_register_globals() and generate __cudaRegisterFunction() call for
268   // each emitted kernel.
269   llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
270   for (llvm::Function *Kernel : EmittedKernels) {
271     llvm::Constant *KernelName = makeConstantString(Kernel->getName());
272     llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
273     llvm::Value *Args[] = {
274         &GpuBinaryHandlePtr, Builder.CreateBitCast(Kernel, VoidPtrTy),
275         KernelName, KernelName, llvm::ConstantInt::get(IntTy, -1), NullPtr,
276         NullPtr, NullPtr, NullPtr,
277         llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
278     Builder.CreateCall(RegisterFunc, Args);
279   }
280 
281   // void __cudaRegisterVar(void **, char *, char *, const char *,
282   //                        int, int, int, int)
283   llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
284                                      CharPtrTy,    IntTy,     IntTy,
285                                      IntTy,        IntTy};
286   llvm::Constant *RegisterVar = CGM.CreateRuntimeFunction(
287       llvm::FunctionType::get(IntTy, RegisterVarParams, false),
288       addUnderscoredPrefixToName("RegisterVar"));
289   for (auto &Pair : DeviceVars) {
290     llvm::GlobalVariable *Var = Pair.first;
291     unsigned Flags = Pair.second;
292     llvm::Constant *VarName = makeConstantString(Var->getName());
293     uint64_t VarSize =
294         CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
295     llvm::Value *Args[] = {
296         &GpuBinaryHandlePtr,
297         Builder.CreateBitCast(Var, VoidPtrTy),
298         VarName,
299         VarName,
300         llvm::ConstantInt::get(IntTy, (Flags & ExternDeviceVar) ? 1 : 0),
301         llvm::ConstantInt::get(IntTy, VarSize),
302         llvm::ConstantInt::get(IntTy, (Flags & ConstantDeviceVar) ? 1 : 0),
303         llvm::ConstantInt::get(IntTy, 0)};
304     Builder.CreateCall(RegisterVar, Args);
305   }
306 
307   Builder.CreateRetVoid();
308   return RegisterKernelsFunc;
309 }
310 
311 /// Creates a global constructor function for the module:
312 ///
313 /// For CUDA:
314 /// \code
315 /// void __cuda_module_ctor(void*) {
316 ///     Handle = __cudaRegisterFatBinary(GpuBinaryBlob);
317 ///     __cuda_register_globals(Handle);
318 /// }
319 /// \endcode
320 ///
321 /// For HIP:
322 /// \code
323 /// void __hip_module_ctor(void*) {
324 ///     if (__hip_gpubin_handle == 0) {
325 ///         __hip_gpubin_handle  = __hipRegisterFatBinary(GpuBinaryBlob);
326 ///         __hip_register_globals(__hip_gpubin_handle);
327 ///     }
328 /// }
329 /// \endcode
330 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
331   bool IsHIP = CGM.getLangOpts().HIP;
332   // No need to generate ctors/dtors if there is no GPU binary.
333   StringRef CudaGpuBinaryFileName = CGM.getCodeGenOpts().CudaGpuBinaryFileName;
334   if (CudaGpuBinaryFileName.empty() && !IsHIP)
335     return nullptr;
336 
337   // void __{cuda|hip}_register_globals(void* handle);
338   llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
339   // We always need a function to pass in as callback. Create a dummy
340   // implementation if we don't need to register anything.
341   if (RelocatableDeviceCode && !RegisterGlobalsFunc)
342     RegisterGlobalsFunc = makeDummyFunction(getRegisterGlobalsFnTy());
343 
344   // void ** __{cuda|hip}RegisterFatBinary(void *);
345   llvm::Constant *RegisterFatbinFunc = CGM.CreateRuntimeFunction(
346       llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
347       addUnderscoredPrefixToName("RegisterFatBinary"));
348   // struct { int magic, int version, void * gpu_binary, void * dont_care };
349   llvm::StructType *FatbinWrapperTy =
350       llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy);
351 
352   // Register GPU binary with the CUDA runtime, store returned handle in a
353   // global variable and save a reference in GpuBinaryHandle to be cleaned up
354   // in destructor on exit. Then associate all known kernels with the GPU binary
355   // handle so CUDA runtime can figure out what to call on the GPU side.
356   std::unique_ptr<llvm::MemoryBuffer> CudaGpuBinary = nullptr;
357   if (!CudaGpuBinaryFileName.empty()) {
358     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CudaGpuBinaryOrErr =
359         llvm::MemoryBuffer::getFileOrSTDIN(CudaGpuBinaryFileName);
360     if (std::error_code EC = CudaGpuBinaryOrErr.getError()) {
361       CGM.getDiags().Report(diag::err_cannot_open_file)
362           << CudaGpuBinaryFileName << EC.message();
363       return nullptr;
364     }
365     CudaGpuBinary = std::move(CudaGpuBinaryOrErr.get());
366   }
367 
368   llvm::Function *ModuleCtorFunc = llvm::Function::Create(
369       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
370       llvm::GlobalValue::InternalLinkage,
371       addUnderscoredPrefixToName("_module_ctor"), &TheModule);
372   llvm::BasicBlock *CtorEntryBB =
373       llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
374   CGBuilderTy CtorBuilder(CGM, Context);
375 
376   CtorBuilder.SetInsertPoint(CtorEntryBB);
377 
378   const char *FatbinConstantName;
379   const char *FatbinSectionName;
380   const char *ModuleIDSectionName;
381   StringRef ModuleIDPrefix;
382   llvm::Constant *FatBinStr;
383   unsigned FatMagic;
384   if (IsHIP) {
385     FatbinConstantName = ".hip_fatbin";
386     FatbinSectionName = ".hipFatBinSegment";
387 
388     ModuleIDSectionName = "__hip_module_id";
389     ModuleIDPrefix = "__hip_";
390 
391     if (CudaGpuBinary) {
392       // If fatbin is available from early finalization, create a string
393       // literal containing the fat binary loaded from the given file.
394       FatBinStr = makeConstantString(CudaGpuBinary->getBuffer(), "",
395                                      FatbinConstantName, 8);
396     } else {
397       // If fatbin is not available, create an external symbol
398       // __hip_fatbin in section .hip_fatbin. The external symbol is supposed
399       // to contain the fat binary but will be populated somewhere else,
400       // e.g. by lld through link script.
401       FatBinStr = new llvm::GlobalVariable(
402         CGM.getModule(), CGM.Int8Ty,
403         /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
404         "__hip_fatbin", nullptr,
405         llvm::GlobalVariable::NotThreadLocal);
406       cast<llvm::GlobalVariable>(FatBinStr)->setSection(FatbinConstantName);
407     }
408 
409     FatMagic = HIPFatMagic;
410   } else {
411     if (RelocatableDeviceCode)
412       FatbinConstantName = CGM.getTriple().isMacOSX()
413                                ? "__NV_CUDA,__nv_relfatbin"
414                                : "__nv_relfatbin";
415     else
416       FatbinConstantName =
417           CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
418     // NVIDIA's cuobjdump looks for fatbins in this section.
419     FatbinSectionName =
420         CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
421 
422     ModuleIDSectionName = CGM.getTriple().isMacOSX()
423                               ? "__NV_CUDA,__nv_module_id"
424                               : "__nv_module_id";
425     ModuleIDPrefix = "__nv_";
426 
427     // For CUDA, create a string literal containing the fat binary loaded from
428     // the given file.
429     FatBinStr = makeConstantString(CudaGpuBinary->getBuffer(), "",
430                                    FatbinConstantName, 8);
431     FatMagic = CudaFatMagic;
432   }
433 
434   // Create initialized wrapper structure that points to the loaded GPU binary
435   ConstantInitBuilder Builder(CGM);
436   auto Values = Builder.beginStruct(FatbinWrapperTy);
437   // Fatbin wrapper magic.
438   Values.addInt(IntTy, FatMagic);
439   // Fatbin version.
440   Values.addInt(IntTy, 1);
441   // Data.
442   Values.add(FatBinStr);
443   // Unused in fatbin v1.
444   Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
445   llvm::GlobalVariable *FatbinWrapper = Values.finishAndCreateGlobal(
446       addUnderscoredPrefixToName("_fatbin_wrapper"), CGM.getPointerAlign(),
447       /*constant*/ true);
448   FatbinWrapper->setSection(FatbinSectionName);
449 
450   // There is only one HIP fat binary per linked module, however there are
451   // multiple constructor functions. Make sure the fat binary is registered
452   // only once. The constructor functions are executed by the dynamic loader
453   // before the program gains control. The dynamic loader cannot execute the
454   // constructor functions concurrently since doing that would not guarantee
455   // thread safety of the loaded program. Therefore we can assume sequential
456   // execution of constructor functions here.
457   if (IsHIP) {
458     auto Linkage = CudaGpuBinary ? llvm::GlobalValue::InternalLinkage :
459         llvm::GlobalValue::LinkOnceAnyLinkage;
460     llvm::BasicBlock *IfBlock =
461         llvm::BasicBlock::Create(Context, "if", ModuleCtorFunc);
462     llvm::BasicBlock *ExitBlock =
463         llvm::BasicBlock::Create(Context, "exit", ModuleCtorFunc);
464     // The name, size, and initialization pattern of this variable is part
465     // of HIP ABI.
466     GpuBinaryHandle = new llvm::GlobalVariable(
467         TheModule, VoidPtrPtrTy, /*isConstant=*/false,
468         Linkage,
469         /*Initializer=*/llvm::ConstantPointerNull::get(VoidPtrPtrTy),
470         "__hip_gpubin_handle");
471     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getQuantity());
472     // Prevent the weak symbol in different shared libraries being merged.
473     if (Linkage != llvm::GlobalValue::InternalLinkage)
474       GpuBinaryHandle->setVisibility(llvm::GlobalValue::HiddenVisibility);
475     Address GpuBinaryAddr(
476         GpuBinaryHandle,
477         CharUnits::fromQuantity(GpuBinaryHandle->getAlignment()));
478     {
479       auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
480       llvm::Constant *Zero =
481           llvm::Constant::getNullValue(HandleValue->getType());
482       llvm::Value *EQZero = CtorBuilder.CreateICmpEQ(HandleValue, Zero);
483       CtorBuilder.CreateCondBr(EQZero, IfBlock, ExitBlock);
484     }
485     {
486       CtorBuilder.SetInsertPoint(IfBlock);
487       // GpuBinaryHandle = __hipRegisterFatBinary(&FatbinWrapper);
488       llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
489           RegisterFatbinFunc,
490           CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
491       CtorBuilder.CreateStore(RegisterFatbinCall, GpuBinaryAddr);
492       CtorBuilder.CreateBr(ExitBlock);
493     }
494     {
495       CtorBuilder.SetInsertPoint(ExitBlock);
496       // Call __hip_register_globals(GpuBinaryHandle);
497       if (RegisterGlobalsFunc) {
498         auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
499         CtorBuilder.CreateCall(RegisterGlobalsFunc, HandleValue);
500       }
501     }
502   } else if (!RelocatableDeviceCode) {
503     // Register binary with CUDA runtime. This is substantially different in
504     // default mode vs. separate compilation!
505     // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
506     llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
507         RegisterFatbinFunc,
508         CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
509     GpuBinaryHandle = new llvm::GlobalVariable(
510         TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
511         llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
512     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getQuantity());
513     CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
514                                    CGM.getPointerAlign());
515 
516     // Call __cuda_register_globals(GpuBinaryHandle);
517     if (RegisterGlobalsFunc)
518       CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
519   } else {
520     // Generate a unique module ID.
521     SmallString<64> ModuleID;
522     llvm::raw_svector_ostream OS(ModuleID);
523     OS << ModuleIDPrefix << llvm::format("%" PRIx64, FatbinWrapper->getGUID());
524     llvm::Constant *ModuleIDConstant =
525         makeConstantString(ModuleID.str(), "", ModuleIDSectionName, 32);
526 
527     // Create an alias for the FatbinWrapper that nvcc will look for.
528     llvm::GlobalAlias::create(llvm::GlobalValue::ExternalLinkage,
529                               Twine("__fatbinwrap") + ModuleID, FatbinWrapper);
530 
531     // void __cudaRegisterLinkedBinary%ModuleID%(void (*)(void *), void *,
532     // void *, void (*)(void **))
533     SmallString<128> RegisterLinkedBinaryName("__cudaRegisterLinkedBinary");
534     RegisterLinkedBinaryName += ModuleID;
535     llvm::Constant *RegisterLinkedBinaryFunc = CGM.CreateRuntimeFunction(
536         getRegisterLinkedBinaryFnTy(), RegisterLinkedBinaryName);
537 
538     assert(RegisterGlobalsFunc && "Expecting at least dummy function!");
539     llvm::Value *Args[] = {RegisterGlobalsFunc,
540                            CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy),
541                            ModuleIDConstant,
542                            makeDummyFunction(getCallbackFnTy())};
543     CtorBuilder.CreateCall(RegisterLinkedBinaryFunc, Args);
544   }
545 
546   // Create destructor and register it with atexit() the way NVCC does it. Doing
547   // it during regular destructor phase worked in CUDA before 9.2 but results in
548   // double-free in 9.2.
549   if (llvm::Function *CleanupFn = makeModuleDtorFunction()) {
550     // extern "C" int atexit(void (*f)(void));
551     llvm::FunctionType *AtExitTy =
552         llvm::FunctionType::get(IntTy, CleanupFn->getType(), false);
553     llvm::Constant *AtExitFunc =
554         CGM.CreateRuntimeFunction(AtExitTy, "atexit", llvm::AttributeList(),
555                                   /*Local=*/true);
556     CtorBuilder.CreateCall(AtExitFunc, CleanupFn);
557   }
558 
559   CtorBuilder.CreateRetVoid();
560   return ModuleCtorFunc;
561 }
562 
563 /// Creates a global destructor function that unregisters the GPU code blob
564 /// registered by constructor.
565 ///
566 /// For CUDA:
567 /// \code
568 /// void __cuda_module_dtor(void*) {
569 ///     __cudaUnregisterFatBinary(Handle);
570 /// }
571 /// \endcode
572 ///
573 /// For HIP:
574 /// \code
575 /// void __hip_module_dtor(void*) {
576 ///     if (__hip_gpubin_handle) {
577 ///         __hipUnregisterFatBinary(__hip_gpubin_handle);
578 ///         __hip_gpubin_handle = 0;
579 ///     }
580 /// }
581 /// \endcode
582 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
583   // No need for destructor if we don't have a handle to unregister.
584   if (!GpuBinaryHandle)
585     return nullptr;
586 
587   // void __cudaUnregisterFatBinary(void ** handle);
588   llvm::Constant *UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
589       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
590       addUnderscoredPrefixToName("UnregisterFatBinary"));
591 
592   llvm::Function *ModuleDtorFunc = llvm::Function::Create(
593       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
594       llvm::GlobalValue::InternalLinkage,
595       addUnderscoredPrefixToName("_module_dtor"), &TheModule);
596 
597   llvm::BasicBlock *DtorEntryBB =
598       llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
599   CGBuilderTy DtorBuilder(CGM, Context);
600   DtorBuilder.SetInsertPoint(DtorEntryBB);
601 
602   Address GpuBinaryAddr(GpuBinaryHandle, CharUnits::fromQuantity(
603                                              GpuBinaryHandle->getAlignment()));
604   auto HandleValue = DtorBuilder.CreateLoad(GpuBinaryAddr);
605   // There is only one HIP fat binary per linked module, however there are
606   // multiple destructor functions. Make sure the fat binary is unregistered
607   // only once.
608   if (CGM.getLangOpts().HIP) {
609     llvm::BasicBlock *IfBlock =
610         llvm::BasicBlock::Create(Context, "if", ModuleDtorFunc);
611     llvm::BasicBlock *ExitBlock =
612         llvm::BasicBlock::Create(Context, "exit", ModuleDtorFunc);
613     llvm::Constant *Zero = llvm::Constant::getNullValue(HandleValue->getType());
614     llvm::Value *NEZero = DtorBuilder.CreateICmpNE(HandleValue, Zero);
615     DtorBuilder.CreateCondBr(NEZero, IfBlock, ExitBlock);
616 
617     DtorBuilder.SetInsertPoint(IfBlock);
618     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
619     DtorBuilder.CreateStore(Zero, GpuBinaryAddr);
620     DtorBuilder.CreateBr(ExitBlock);
621 
622     DtorBuilder.SetInsertPoint(ExitBlock);
623   } else {
624     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
625   }
626   DtorBuilder.CreateRetVoid();
627   return ModuleDtorFunc;
628 }
629 
630 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
631   return new CGNVCUDARuntime(CGM);
632 }
633