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     if (Alignment)
81       GV->setAlignment(Alignment);
82 
83     return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
84                                                 ConstStr.getPointer(), Zeros);
85   }
86 
87   /// Helper function that generates an empty dummy function returning void.
88   llvm::Function *makeDummyFunction(llvm::FunctionType *FnTy) {
89     assert(FnTy->getReturnType()->isVoidTy() &&
90            "Can only generate dummy functions returning void!");
91     llvm::Function *DummyFunc = llvm::Function::Create(
92         FnTy, llvm::GlobalValue::InternalLinkage, "dummy", &TheModule);
93 
94     llvm::BasicBlock *DummyBlock =
95         llvm::BasicBlock::Create(Context, "", DummyFunc);
96     CGBuilderTy FuncBuilder(CGM, Context);
97     FuncBuilder.SetInsertPoint(DummyBlock);
98     FuncBuilder.CreateRetVoid();
99 
100     return DummyFunc;
101   }
102 
103   void emitDeviceStubBody(CodeGenFunction &CGF, FunctionArgList &Args);
104 
105 public:
106   CGNVCUDARuntime(CodeGenModule &CGM);
107 
108   void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
109   void registerDeviceVar(llvm::GlobalVariable &Var, unsigned Flags) override {
110     DeviceVars.push_back(std::make_pair(&Var, Flags));
111   }
112 
113   /// Creates module constructor function
114   llvm::Function *makeModuleCtorFunction() override;
115   /// Creates module destructor function
116   llvm::Function *makeModuleDtorFunction() override;
117 };
118 
119 }
120 
121 std::string CGNVCUDARuntime::addPrefixToName(StringRef FuncName) const {
122   if (CGM.getLangOpts().HIP)
123     return ((Twine("hip") + Twine(FuncName)).str());
124   return ((Twine("cuda") + Twine(FuncName)).str());
125 }
126 std::string
127 CGNVCUDARuntime::addUnderscoredPrefixToName(StringRef FuncName) const {
128   if (CGM.getLangOpts().HIP)
129     return ((Twine("__hip") + Twine(FuncName)).str());
130   return ((Twine("__cuda") + Twine(FuncName)).str());
131 }
132 
133 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
134     : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
135       TheModule(CGM.getModule()),
136       RelocatableDeviceCode(CGM.getLangOpts().CUDARelocatableDeviceCode) {
137   CodeGen::CodeGenTypes &Types = CGM.getTypes();
138   ASTContext &Ctx = CGM.getContext();
139 
140   IntTy = CGM.IntTy;
141   SizeTy = CGM.SizeTy;
142   VoidTy = CGM.VoidTy;
143 
144   CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
145   VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
146   VoidPtrPtrTy = VoidPtrTy->getPointerTo();
147 }
148 
149 llvm::Constant *CGNVCUDARuntime::getSetupArgumentFn() const {
150   // cudaError_t cudaSetupArgument(void *, size_t, size_t)
151   llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
152   return CGM.CreateRuntimeFunction(
153       llvm::FunctionType::get(IntTy, Params, false),
154       addPrefixToName("SetupArgument"));
155 }
156 
157 llvm::Constant *CGNVCUDARuntime::getLaunchFn() const {
158   if (CGM.getLangOpts().HIP) {
159     // hipError_t hipLaunchByPtr(char *);
160     return CGM.CreateRuntimeFunction(
161         llvm::FunctionType::get(IntTy, CharPtrTy, false), "hipLaunchByPtr");
162   } else {
163     // cudaError_t cudaLaunch(char *);
164     return CGM.CreateRuntimeFunction(
165         llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
166   }
167 }
168 
169 llvm::FunctionType *CGNVCUDARuntime::getRegisterGlobalsFnTy() const {
170   return llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false);
171 }
172 
173 llvm::FunctionType *CGNVCUDARuntime::getCallbackFnTy() const {
174   return llvm::FunctionType::get(VoidTy, VoidPtrTy, false);
175 }
176 
177 llvm::FunctionType *CGNVCUDARuntime::getRegisterLinkedBinaryFnTy() const {
178   auto CallbackFnTy = getCallbackFnTy();
179   auto RegisterGlobalsFnTy = getRegisterGlobalsFnTy();
180   llvm::Type *Params[] = {RegisterGlobalsFnTy->getPointerTo(), VoidPtrTy,
181                           VoidPtrTy, CallbackFnTy->getPointerTo()};
182   return llvm::FunctionType::get(VoidTy, Params, false);
183 }
184 
185 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
186                                      FunctionArgList &Args) {
187   EmittedKernels.push_back(CGF.CurFn);
188   emitDeviceStubBody(CGF, Args);
189 }
190 
191 void CGNVCUDARuntime::emitDeviceStubBody(CodeGenFunction &CGF,
192                                          FunctionArgList &Args) {
193   // Emit a call to cudaSetupArgument for each arg in Args.
194   llvm::Constant *cudaSetupArgFn = getSetupArgumentFn();
195   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
196   CharUnits Offset = CharUnits::Zero();
197   for (const VarDecl *A : Args) {
198     CharUnits TyWidth, TyAlign;
199     std::tie(TyWidth, TyAlign) =
200         CGM.getContext().getTypeInfoInChars(A->getType());
201     Offset = Offset.alignTo(TyAlign);
202     llvm::Value *Args[] = {
203         CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
204                                       VoidPtrTy),
205         llvm::ConstantInt::get(SizeTy, TyWidth.getQuantity()),
206         llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
207     };
208     llvm::CallSite CS = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
209     llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
210     llvm::Value *CSZero = CGF.Builder.CreateICmpEQ(CS.getInstruction(), Zero);
211     llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
212     CGF.Builder.CreateCondBr(CSZero, NextBlock, EndBlock);
213     CGF.EmitBlock(NextBlock);
214     Offset += TyWidth;
215   }
216 
217   // Emit the call to cudaLaunch
218   llvm::Constant *cudaLaunchFn = getLaunchFn();
219   llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
220   CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
221   CGF.EmitBranch(EndBlock);
222 
223   CGF.EmitBlock(EndBlock);
224 }
225 
226 /// Creates a function that sets up state on the host side for CUDA objects that
227 /// have a presence on both the host and device sides. Specifically, registers
228 /// the host side of kernel functions and device global variables with the CUDA
229 /// runtime.
230 /// \code
231 /// void __cuda_register_globals(void** GpuBinaryHandle) {
232 ///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
233 ///    ...
234 ///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
235 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
236 ///    ...
237 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
238 /// }
239 /// \endcode
240 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
241   // No need to register anything
242   if (EmittedKernels.empty() && DeviceVars.empty())
243     return nullptr;
244 
245   llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
246       getRegisterGlobalsFnTy(), llvm::GlobalValue::InternalLinkage,
247       addUnderscoredPrefixToName("_register_globals"), &TheModule);
248   llvm::BasicBlock *EntryBB =
249       llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
250   CGBuilderTy Builder(CGM, Context);
251   Builder.SetInsertPoint(EntryBB);
252 
253   // void __cudaRegisterFunction(void **, const char *, char *, const char *,
254   //                             int, uint3*, uint3*, dim3*, dim3*, int*)
255   llvm::Type *RegisterFuncParams[] = {
256       VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
257       VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
258   llvm::Constant *RegisterFunc = CGM.CreateRuntimeFunction(
259       llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
260       addUnderscoredPrefixToName("RegisterFunction"));
261 
262   // Extract GpuBinaryHandle passed as the first argument passed to
263   // __cuda_register_globals() and generate __cudaRegisterFunction() call for
264   // each emitted kernel.
265   llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
266   for (llvm::Function *Kernel : EmittedKernels) {
267     llvm::Constant *KernelName = makeConstantString(Kernel->getName());
268     llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
269     llvm::Value *Args[] = {
270         &GpuBinaryHandlePtr, Builder.CreateBitCast(Kernel, VoidPtrTy),
271         KernelName, KernelName, llvm::ConstantInt::get(IntTy, -1), NullPtr,
272         NullPtr, NullPtr, NullPtr,
273         llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
274     Builder.CreateCall(RegisterFunc, Args);
275   }
276 
277   // void __cudaRegisterVar(void **, char *, char *, const char *,
278   //                        int, int, int, int)
279   llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
280                                      CharPtrTy,    IntTy,     IntTy,
281                                      IntTy,        IntTy};
282   llvm::Constant *RegisterVar = CGM.CreateRuntimeFunction(
283       llvm::FunctionType::get(IntTy, RegisterVarParams, false),
284       addUnderscoredPrefixToName("RegisterVar"));
285   for (auto &Pair : DeviceVars) {
286     llvm::GlobalVariable *Var = Pair.first;
287     unsigned Flags = Pair.second;
288     llvm::Constant *VarName = makeConstantString(Var->getName());
289     uint64_t VarSize =
290         CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
291     llvm::Value *Args[] = {
292         &GpuBinaryHandlePtr,
293         Builder.CreateBitCast(Var, VoidPtrTy),
294         VarName,
295         VarName,
296         llvm::ConstantInt::get(IntTy, (Flags & ExternDeviceVar) ? 1 : 0),
297         llvm::ConstantInt::get(IntTy, VarSize),
298         llvm::ConstantInt::get(IntTy, (Flags & ConstantDeviceVar) ? 1 : 0),
299         llvm::ConstantInt::get(IntTy, 0)};
300     Builder.CreateCall(RegisterVar, Args);
301   }
302 
303   Builder.CreateRetVoid();
304   return RegisterKernelsFunc;
305 }
306 
307 /// Creates a global constructor function for the module:
308 /// \code
309 /// void __cuda_module_ctor(void*) {
310 ///     Handle = __cudaRegisterFatBinary(GpuBinaryBlob);
311 ///     __cuda_register_globals(Handle);
312 /// }
313 /// \endcode
314 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
315   bool IsHIP = CGM.getLangOpts().HIP;
316   // No need to generate ctors/dtors if there is no GPU binary.
317   StringRef CudaGpuBinaryFileName = CGM.getCodeGenOpts().CudaGpuBinaryFileName;
318   if (CudaGpuBinaryFileName.empty() && !IsHIP)
319     return nullptr;
320 
321   // void __{cuda|hip}_register_globals(void* handle);
322   llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
323   // We always need a function to pass in as callback. Create a dummy
324   // implementation if we don't need to register anything.
325   if (RelocatableDeviceCode && !RegisterGlobalsFunc)
326     RegisterGlobalsFunc = makeDummyFunction(getRegisterGlobalsFnTy());
327 
328   // void ** __{cuda|hip}RegisterFatBinary(void *);
329   llvm::Constant *RegisterFatbinFunc = CGM.CreateRuntimeFunction(
330       llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
331       addUnderscoredPrefixToName("RegisterFatBinary"));
332   // struct { int magic, int version, void * gpu_binary, void * dont_care };
333   llvm::StructType *FatbinWrapperTy =
334       llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy);
335 
336   // Register GPU binary with the CUDA runtime, store returned handle in a
337   // global variable and save a reference in GpuBinaryHandle to be cleaned up
338   // in destructor on exit. Then associate all known kernels with the GPU binary
339   // handle so CUDA runtime can figure out what to call on the GPU side.
340   std::unique_ptr<llvm::MemoryBuffer> CudaGpuBinary;
341   if (!IsHIP) {
342     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CudaGpuBinaryOrErr =
343         llvm::MemoryBuffer::getFileOrSTDIN(CudaGpuBinaryFileName);
344     if (std::error_code EC = CudaGpuBinaryOrErr.getError()) {
345       CGM.getDiags().Report(diag::err_cannot_open_file)
346           << CudaGpuBinaryFileName << EC.message();
347       return nullptr;
348     }
349     CudaGpuBinary = std::move(CudaGpuBinaryOrErr.get());
350   }
351 
352   llvm::Function *ModuleCtorFunc = llvm::Function::Create(
353       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
354       llvm::GlobalValue::InternalLinkage,
355       addUnderscoredPrefixToName("_module_ctor"), &TheModule);
356   llvm::BasicBlock *CtorEntryBB =
357       llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
358   CGBuilderTy CtorBuilder(CGM, Context);
359 
360   CtorBuilder.SetInsertPoint(CtorEntryBB);
361 
362   const char *FatbinConstantName;
363   const char *FatbinSectionName;
364   const char *ModuleIDSectionName;
365   StringRef ModuleIDPrefix;
366   llvm::Constant *FatBinStr;
367   unsigned FatMagic;
368   if (IsHIP) {
369     FatbinConstantName = ".hip_fatbin";
370     FatbinSectionName = ".hipFatBinSegment";
371 
372     ModuleIDSectionName = "__hip_module_id";
373     ModuleIDPrefix = "__hip_";
374 
375     // For HIP, create an external symbol __hip_fatbin in section .hip_fatbin.
376     // The external symbol is supposed to contain the fat binary but will be
377     // populated somewhere else, e.g. by lld through link script.
378     FatBinStr = new llvm::GlobalVariable(
379         CGM.getModule(), CGM.Int8Ty,
380         /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
381         "__hip_fatbin", nullptr,
382         llvm::GlobalVariable::NotThreadLocal);
383     cast<llvm::GlobalVariable>(FatBinStr)->setSection(FatbinConstantName);
384 
385     FatMagic = HIPFatMagic;
386   } else {
387     if (RelocatableDeviceCode)
388       // TODO: Figure out how this is called on mac OS!
389       FatbinConstantName = "__nv_relfatbin";
390     else
391       FatbinConstantName =
392           CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
393     // NVIDIA's cuobjdump looks for fatbins in this section.
394     FatbinSectionName =
395         CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
396 
397     // TODO: Figure out how this is called on mac OS!
398     ModuleIDSectionName = "__nv_module_id";
399     ModuleIDPrefix = "__nv_";
400 
401     // For CUDA, create a string literal containing the fat binary loaded from
402     // the given file.
403     FatBinStr = makeConstantString(CudaGpuBinary->getBuffer(), "",
404                                    FatbinConstantName, 8);
405     FatMagic = CudaFatMagic;
406   }
407 
408   // Create initialized wrapper structure that points to the loaded GPU binary
409   ConstantInitBuilder Builder(CGM);
410   auto Values = Builder.beginStruct(FatbinWrapperTy);
411   // Fatbin wrapper magic.
412   Values.addInt(IntTy, FatMagic);
413   // Fatbin version.
414   Values.addInt(IntTy, 1);
415   // Data.
416   Values.add(FatBinStr);
417   // Unused in fatbin v1.
418   Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
419   llvm::GlobalVariable *FatbinWrapper = Values.finishAndCreateGlobal(
420       addUnderscoredPrefixToName("_fatbin_wrapper"), CGM.getPointerAlign(),
421       /*constant*/ true);
422   FatbinWrapper->setSection(FatbinSectionName);
423 
424   // Register binary with CUDA/HIP runtime. This is substantially different in
425   // default mode vs. separate compilation!
426   if (!RelocatableDeviceCode) {
427     // GpuBinaryHandle = __{cuda|hip}RegisterFatBinary(&FatbinWrapper);
428     llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
429         RegisterFatbinFunc,
430         CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
431     GpuBinaryHandle = new llvm::GlobalVariable(
432         TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
433         llvm::ConstantPointerNull::get(VoidPtrPtrTy),
434         addUnderscoredPrefixToName("_gpubin_handle"));
435 
436     CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
437                                    CGM.getPointerAlign());
438 
439     // Call __{cuda|hip}_register_globals(GpuBinaryHandle);
440     if (RegisterGlobalsFunc)
441       CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
442   } else {
443     // Generate a unique module ID.
444     SmallString<64> ModuleID;
445     llvm::raw_svector_ostream OS(ModuleID);
446     OS << ModuleIDPrefix << llvm::format("%x", FatbinWrapper->getGUID());
447     llvm::Constant *ModuleIDConstant =
448         makeConstantString(ModuleID.str(), "", ModuleIDSectionName, 32);
449 
450     // Create an alias for the FatbinWrapper that nvcc or hip backend will
451     // look for.
452     llvm::GlobalAlias::create(llvm::GlobalValue::ExternalLinkage,
453                               Twine("__fatbinwrap") + ModuleID, FatbinWrapper);
454 
455     // void __{cuda|hip}RegisterLinkedBinary%ModuleID%(void (*)(void *), void *,
456     // void *, void (*)(void **))
457     SmallString<128> RegisterLinkedBinaryName(
458         addUnderscoredPrefixToName("RegisterLinkedBinary"));
459     RegisterLinkedBinaryName += ModuleID;
460     llvm::Constant *RegisterLinkedBinaryFunc = CGM.CreateRuntimeFunction(
461         getRegisterLinkedBinaryFnTy(), RegisterLinkedBinaryName);
462 
463     assert(RegisterGlobalsFunc && "Expecting at least dummy function!");
464     llvm::Value *Args[] = {RegisterGlobalsFunc,
465                            CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy),
466                            ModuleIDConstant,
467                            makeDummyFunction(getCallbackFnTy())};
468     CtorBuilder.CreateCall(RegisterLinkedBinaryFunc, Args);
469   }
470 
471   CtorBuilder.CreateRetVoid();
472   return ModuleCtorFunc;
473 }
474 
475 /// Creates a global destructor function that unregisters the GPU code blob
476 /// registered by constructor.
477 /// \code
478 /// void __cuda_module_dtor(void*) {
479 ///     __cudaUnregisterFatBinary(Handle);
480 /// }
481 /// \endcode
482 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
483   // No need for destructor if we don't have a handle to unregister.
484   if (!GpuBinaryHandle)
485     return nullptr;
486 
487   // void __cudaUnregisterFatBinary(void ** handle);
488   llvm::Constant *UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
489       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
490       addUnderscoredPrefixToName("UnregisterFatBinary"));
491 
492   llvm::Function *ModuleDtorFunc = llvm::Function::Create(
493       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
494       llvm::GlobalValue::InternalLinkage,
495       addUnderscoredPrefixToName("_module_dtor"), &TheModule);
496 
497   llvm::BasicBlock *DtorEntryBB =
498       llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
499   CGBuilderTy DtorBuilder(CGM, Context);
500   DtorBuilder.SetInsertPoint(DtorEntryBB);
501 
502   auto HandleValue =
503       DtorBuilder.CreateAlignedLoad(GpuBinaryHandle, CGM.getPointerAlign());
504   DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
505 
506   DtorBuilder.CreateRetVoid();
507   return ModuleDtorFunc;
508 }
509 
510 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
511   return new CGNVCUDARuntime(CGM);
512 }
513