xref: /netbsd-src/external/apache2/llvm/dist/clang/lib/CodeGen/CGCUDANV.cpp (revision e038c9c4676b0f19b1b7dd08a940c6ed64a6d5ae) 
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 "CGCXXABI.h"
16 #include "CodeGenFunction.h"
17 #include "CodeGenModule.h"
18 #include "clang/AST/Decl.h"
19 #include "clang/Basic/Cuda.h"
20 #include "clang/CodeGen/CodeGenABITypes.h"
21 #include "clang/CodeGen/ConstantInitBuilder.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/DerivedTypes.h"
25 #include "llvm/IR/ReplaceConstant.h"
26 #include "llvm/Support/Format.h"
27 
28 using namespace clang;
29 using namespace CodeGen;
30 
31 namespace {
32 constexpr unsigned CudaFatMagic = 0x466243b1;
33 constexpr unsigned HIPFatMagic = 0x48495046; // "HIPF"
34 
35 class CGNVCUDARuntime : public CGCUDARuntime {
36 
37 private:
38   llvm::IntegerType *IntTy, *SizeTy;
39   llvm::Type *VoidTy;
40   llvm::PointerType *CharPtrTy, *VoidPtrTy, *VoidPtrPtrTy;
41 
42   /// Convenience reference to LLVM Context
43   llvm::LLVMContext &Context;
44   /// Convenience reference to the current module
45   llvm::Module &TheModule;
46   /// Keeps track of kernel launch stubs and handles emitted in this module
47   struct KernelInfo {
48     llvm::Function *Kernel; // stub function to help launch kernel
49     const Decl *D;
50   };
51   llvm::SmallVector<KernelInfo, 16> EmittedKernels;
52   // Map a device stub function to a symbol for identifying kernel in host code.
53   // For CUDA, the symbol for identifying the kernel is the same as the device
54   // stub function. For HIP, they are different.
55   llvm::DenseMap<llvm::Function *, llvm::GlobalValue *> KernelHandles;
56   // Map a kernel handle to the kernel stub.
57   llvm::DenseMap<llvm::GlobalValue *, llvm::Function *> KernelStubs;
58   struct VarInfo {
59     llvm::GlobalVariable *Var;
60     const VarDecl *D;
61     DeviceVarFlags Flags;
62   };
63   llvm::SmallVector<VarInfo, 16> DeviceVars;
64   /// Keeps track of variable containing handle of GPU binary. Populated by
65   /// ModuleCtorFunction() and used to create corresponding cleanup calls in
66   /// ModuleDtorFunction()
67   llvm::GlobalVariable *GpuBinaryHandle = nullptr;
68   /// Whether we generate relocatable device code.
69   bool RelocatableDeviceCode;
70   /// Mangle context for device.
71   std::unique_ptr<MangleContext> DeviceMC;
72 
73   llvm::FunctionCallee getSetupArgumentFn() const;
74   llvm::FunctionCallee getLaunchFn() const;
75 
76   llvm::FunctionType *getRegisterGlobalsFnTy() const;
77   llvm::FunctionType *getCallbackFnTy() const;
78   llvm::FunctionType *getRegisterLinkedBinaryFnTy() const;
79   std::string addPrefixToName(StringRef FuncName) const;
80   std::string addUnderscoredPrefixToName(StringRef FuncName) const;
81 
82   /// Creates a function to register all kernel stubs generated in this module.
83   llvm::Function *makeRegisterGlobalsFn();
84 
85   /// Helper function that generates a constant string and returns a pointer to
86   /// the start of the string.  The result of this function can be used anywhere
87   /// where the C code specifies const char*.
makeConstantString(const std::string & Str,const std::string & Name="",const std::string & SectionName="",unsigned Alignment=0)88   llvm::Constant *makeConstantString(const std::string &Str,
89                                      const std::string &Name = "",
90                                      const std::string &SectionName = "",
91                                      unsigned Alignment = 0) {
92     llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
93                                llvm::ConstantInt::get(SizeTy, 0)};
94     auto ConstStr = CGM.GetAddrOfConstantCString(Str, Name.c_str());
95     llvm::GlobalVariable *GV =
96         cast<llvm::GlobalVariable>(ConstStr.getPointer());
97     if (!SectionName.empty()) {
98       GV->setSection(SectionName);
99       // Mark the address as used which make sure that this section isn't
100       // merged and we will really have it in the object file.
101       GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::None);
102     }
103     if (Alignment)
104       GV->setAlignment(llvm::Align(Alignment));
105 
106     return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
107                                                 ConstStr.getPointer(), Zeros);
108   }
109 
110   /// Helper function that generates an empty dummy function returning void.
makeDummyFunction(llvm::FunctionType * FnTy)111   llvm::Function *makeDummyFunction(llvm::FunctionType *FnTy) {
112     assert(FnTy->getReturnType()->isVoidTy() &&
113            "Can only generate dummy functions returning void!");
114     llvm::Function *DummyFunc = llvm::Function::Create(
115         FnTy, llvm::GlobalValue::InternalLinkage, "dummy", &TheModule);
116 
117     llvm::BasicBlock *DummyBlock =
118         llvm::BasicBlock::Create(Context, "", DummyFunc);
119     CGBuilderTy FuncBuilder(CGM, Context);
120     FuncBuilder.SetInsertPoint(DummyBlock);
121     FuncBuilder.CreateRetVoid();
122 
123     return DummyFunc;
124   }
125 
126   void emitDeviceStubBodyLegacy(CodeGenFunction &CGF, FunctionArgList &Args);
127   void emitDeviceStubBodyNew(CodeGenFunction &CGF, FunctionArgList &Args);
128   std::string getDeviceSideName(const NamedDecl *ND) override;
129 
registerDeviceVar(const VarDecl * VD,llvm::GlobalVariable & Var,bool Extern,bool Constant)130   void registerDeviceVar(const VarDecl *VD, llvm::GlobalVariable &Var,
131                          bool Extern, bool Constant) {
132     DeviceVars.push_back({&Var,
133                           VD,
134                           {DeviceVarFlags::Variable, Extern, Constant,
135                            VD->hasAttr<HIPManagedAttr>(),
136                            /*Normalized*/ false, 0}});
137   }
registerDeviceSurf(const VarDecl * VD,llvm::GlobalVariable & Var,bool Extern,int Type)138   void registerDeviceSurf(const VarDecl *VD, llvm::GlobalVariable &Var,
139                           bool Extern, int Type) {
140     DeviceVars.push_back({&Var,
141                           VD,
142                           {DeviceVarFlags::Surface, Extern, /*Constant*/ false,
143                            /*Managed*/ false,
144                            /*Normalized*/ false, Type}});
145   }
registerDeviceTex(const VarDecl * VD,llvm::GlobalVariable & Var,bool Extern,int Type,bool Normalized)146   void registerDeviceTex(const VarDecl *VD, llvm::GlobalVariable &Var,
147                          bool Extern, int Type, bool Normalized) {
148     DeviceVars.push_back({&Var,
149                           VD,
150                           {DeviceVarFlags::Texture, Extern, /*Constant*/ false,
151                            /*Managed*/ false, Normalized, Type}});
152   }
153 
154   /// Creates module constructor function
155   llvm::Function *makeModuleCtorFunction();
156   /// Creates module destructor function
157   llvm::Function *makeModuleDtorFunction();
158   /// Transform managed variables for device compilation.
159   void transformManagedVars();
160 
161 public:
162   CGNVCUDARuntime(CodeGenModule &CGM);
163 
164   llvm::GlobalValue *getKernelHandle(llvm::Function *F, GlobalDecl GD) override;
getKernelStub(llvm::GlobalValue * Handle)165   llvm::Function *getKernelStub(llvm::GlobalValue *Handle) override {
166     auto Loc = KernelStubs.find(Handle);
167     assert(Loc != KernelStubs.end());
168     return Loc->second;
169   }
170   void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
171   void handleVarRegistration(const VarDecl *VD,
172                              llvm::GlobalVariable &Var) override;
173   void
174   internalizeDeviceSideVar(const VarDecl *D,
175                            llvm::GlobalValue::LinkageTypes &Linkage) override;
176 
177   llvm::Function *finalizeModule() override;
178 };
179 
180 }
181 
addPrefixToName(StringRef FuncName) const182 std::string CGNVCUDARuntime::addPrefixToName(StringRef FuncName) const {
183   if (CGM.getLangOpts().HIP)
184     return ((Twine("hip") + Twine(FuncName)).str());
185   return ((Twine("cuda") + Twine(FuncName)).str());
186 }
187 std::string
addUnderscoredPrefixToName(StringRef FuncName) const188 CGNVCUDARuntime::addUnderscoredPrefixToName(StringRef FuncName) const {
189   if (CGM.getLangOpts().HIP)
190     return ((Twine("__hip") + Twine(FuncName)).str());
191   return ((Twine("__cuda") + Twine(FuncName)).str());
192 }
193 
CGNVCUDARuntime(CodeGenModule & CGM)194 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
195     : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
196       TheModule(CGM.getModule()),
197       RelocatableDeviceCode(CGM.getLangOpts().GPURelocatableDeviceCode),
198       DeviceMC(CGM.getContext().createMangleContext(
199           CGM.getContext().getAuxTargetInfo())) {
200   CodeGen::CodeGenTypes &Types = CGM.getTypes();
201   ASTContext &Ctx = CGM.getContext();
202 
203   IntTy = CGM.IntTy;
204   SizeTy = CGM.SizeTy;
205   VoidTy = CGM.VoidTy;
206 
207   CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
208   VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
209   VoidPtrPtrTy = VoidPtrTy->getPointerTo();
210   if (CGM.getContext().getAuxTargetInfo()) {
211     // If the host and device have different C++ ABIs, mark it as the device
212     // mangle context so that the mangling needs to retrieve the additonal
213     // device lambda mangling number instead of the regular host one.
214     DeviceMC->setDeviceMangleContext(
215         CGM.getContext().getTargetInfo().getCXXABI().isMicrosoft() &&
216         CGM.getContext().getAuxTargetInfo()->getCXXABI().isItaniumFamily());
217   }
218 }
219 
getSetupArgumentFn() const220 llvm::FunctionCallee CGNVCUDARuntime::getSetupArgumentFn() const {
221   // cudaError_t cudaSetupArgument(void *, size_t, size_t)
222   llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
223   return CGM.CreateRuntimeFunction(
224       llvm::FunctionType::get(IntTy, Params, false),
225       addPrefixToName("SetupArgument"));
226 }
227 
getLaunchFn() const228 llvm::FunctionCallee CGNVCUDARuntime::getLaunchFn() const {
229   if (CGM.getLangOpts().HIP) {
230     // hipError_t hipLaunchByPtr(char *);
231     return CGM.CreateRuntimeFunction(
232         llvm::FunctionType::get(IntTy, CharPtrTy, false), "hipLaunchByPtr");
233   } else {
234     // cudaError_t cudaLaunch(char *);
235     return CGM.CreateRuntimeFunction(
236         llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
237   }
238 }
239 
getRegisterGlobalsFnTy() const240 llvm::FunctionType *CGNVCUDARuntime::getRegisterGlobalsFnTy() const {
241   return llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false);
242 }
243 
getCallbackFnTy() const244 llvm::FunctionType *CGNVCUDARuntime::getCallbackFnTy() const {
245   return llvm::FunctionType::get(VoidTy, VoidPtrTy, false);
246 }
247 
getRegisterLinkedBinaryFnTy() const248 llvm::FunctionType *CGNVCUDARuntime::getRegisterLinkedBinaryFnTy() const {
249   auto CallbackFnTy = getCallbackFnTy();
250   auto RegisterGlobalsFnTy = getRegisterGlobalsFnTy();
251   llvm::Type *Params[] = {RegisterGlobalsFnTy->getPointerTo(), VoidPtrTy,
252                           VoidPtrTy, CallbackFnTy->getPointerTo()};
253   return llvm::FunctionType::get(VoidTy, Params, false);
254 }
255 
getDeviceSideName(const NamedDecl * ND)256 std::string CGNVCUDARuntime::getDeviceSideName(const NamedDecl *ND) {
257   GlobalDecl GD;
258   // D could be either a kernel or a variable.
259   if (auto *FD = dyn_cast<FunctionDecl>(ND))
260     GD = GlobalDecl(FD, KernelReferenceKind::Kernel);
261   else
262     GD = GlobalDecl(ND);
263   std::string DeviceSideName;
264   MangleContext *MC;
265   if (CGM.getLangOpts().CUDAIsDevice)
266     MC = &CGM.getCXXABI().getMangleContext();
267   else
268     MC = DeviceMC.get();
269   if (MC->shouldMangleDeclName(ND)) {
270     SmallString<256> Buffer;
271     llvm::raw_svector_ostream Out(Buffer);
272     MC->mangleName(GD, Out);
273     DeviceSideName = std::string(Out.str());
274   } else
275     DeviceSideName = std::string(ND->getIdentifier()->getName());
276 
277   // Make unique name for device side static file-scope variable for HIP.
278   if (CGM.getContext().shouldExternalizeStaticVar(ND) &&
279       CGM.getLangOpts().GPURelocatableDeviceCode &&
280       !CGM.getLangOpts().CUID.empty()) {
281     SmallString<256> Buffer;
282     llvm::raw_svector_ostream Out(Buffer);
283     Out << DeviceSideName;
284     CGM.printPostfixForExternalizedStaticVar(Out);
285     DeviceSideName = std::string(Out.str());
286   }
287   return DeviceSideName;
288 }
289 
emitDeviceStub(CodeGenFunction & CGF,FunctionArgList & Args)290 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
291                                      FunctionArgList &Args) {
292   EmittedKernels.push_back({CGF.CurFn, CGF.CurFuncDecl});
293   if (auto *GV = dyn_cast<llvm::GlobalVariable>(KernelHandles[CGF.CurFn])) {
294     GV->setLinkage(CGF.CurFn->getLinkage());
295     GV->setInitializer(CGF.CurFn);
296   }
297   if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
298                          CudaFeature::CUDA_USES_NEW_LAUNCH) ||
299       (CGF.getLangOpts().HIP && CGF.getLangOpts().HIPUseNewLaunchAPI))
300     emitDeviceStubBodyNew(CGF, Args);
301   else
302     emitDeviceStubBodyLegacy(CGF, Args);
303 }
304 
305 // CUDA 9.0+ uses new way to launch kernels. Parameters are packed in a local
306 // array and kernels are launched using cudaLaunchKernel().
emitDeviceStubBodyNew(CodeGenFunction & CGF,FunctionArgList & Args)307 void CGNVCUDARuntime::emitDeviceStubBodyNew(CodeGenFunction &CGF,
308                                             FunctionArgList &Args) {
309   // Build the shadow stack entry at the very start of the function.
310 
311   // Calculate amount of space we will need for all arguments.  If we have no
312   // args, allocate a single pointer so we still have a valid pointer to the
313   // argument array that we can pass to runtime, even if it will be unused.
314   Address KernelArgs = CGF.CreateTempAlloca(
315       VoidPtrTy, CharUnits::fromQuantity(16), "kernel_args",
316       llvm::ConstantInt::get(SizeTy, std::max<size_t>(1, Args.size())));
317   // Store pointers to the arguments in a locally allocated launch_args.
318   for (unsigned i = 0; i < Args.size(); ++i) {
319     llvm::Value* VarPtr = CGF.GetAddrOfLocalVar(Args[i]).getPointer();
320     llvm::Value *VoidVarPtr = CGF.Builder.CreatePointerCast(VarPtr, VoidPtrTy);
321     CGF.Builder.CreateDefaultAlignedStore(
322         VoidVarPtr, CGF.Builder.CreateConstGEP1_32(KernelArgs.getPointer(), i));
323   }
324 
325   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
326 
327   // Lookup cudaLaunchKernel/hipLaunchKernel function.
328   // cudaError_t cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
329   //                              void **args, size_t sharedMem,
330   //                              cudaStream_t stream);
331   // hipError_t hipLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
332   //                            void **args, size_t sharedMem,
333   //                            hipStream_t stream);
334   TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
335   DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
336   auto LaunchKernelName = addPrefixToName("LaunchKernel");
337   IdentifierInfo &cudaLaunchKernelII =
338       CGM.getContext().Idents.get(LaunchKernelName);
339   FunctionDecl *cudaLaunchKernelFD = nullptr;
340   for (auto *Result : DC->lookup(&cudaLaunchKernelII)) {
341     if (FunctionDecl *FD = dyn_cast<FunctionDecl>(Result))
342       cudaLaunchKernelFD = FD;
343   }
344 
345   if (cudaLaunchKernelFD == nullptr) {
346     CGM.Error(CGF.CurFuncDecl->getLocation(),
347               "Can't find declaration for " + LaunchKernelName);
348     return;
349   }
350   // Create temporary dim3 grid_dim, block_dim.
351   ParmVarDecl *GridDimParam = cudaLaunchKernelFD->getParamDecl(1);
352   QualType Dim3Ty = GridDimParam->getType();
353   Address GridDim =
354       CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "grid_dim");
355   Address BlockDim =
356       CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "block_dim");
357   Address ShmemSize =
358       CGF.CreateTempAlloca(SizeTy, CGM.getSizeAlign(), "shmem_size");
359   Address Stream =
360       CGF.CreateTempAlloca(VoidPtrTy, CGM.getPointerAlign(), "stream");
361   llvm::FunctionCallee cudaPopConfigFn = CGM.CreateRuntimeFunction(
362       llvm::FunctionType::get(IntTy,
363                               {/*gridDim=*/GridDim.getType(),
364                                /*blockDim=*/BlockDim.getType(),
365                                /*ShmemSize=*/ShmemSize.getType(),
366                                /*Stream=*/Stream.getType()},
367                               /*isVarArg=*/false),
368       addUnderscoredPrefixToName("PopCallConfiguration"));
369 
370   CGF.EmitRuntimeCallOrInvoke(cudaPopConfigFn,
371                               {GridDim.getPointer(), BlockDim.getPointer(),
372                                ShmemSize.getPointer(), Stream.getPointer()});
373 
374   // Emit the call to cudaLaunch
375   llvm::Value *Kernel =
376       CGF.Builder.CreatePointerCast(KernelHandles[CGF.CurFn], VoidPtrTy);
377   CallArgList LaunchKernelArgs;
378   LaunchKernelArgs.add(RValue::get(Kernel),
379                        cudaLaunchKernelFD->getParamDecl(0)->getType());
380   LaunchKernelArgs.add(RValue::getAggregate(GridDim), Dim3Ty);
381   LaunchKernelArgs.add(RValue::getAggregate(BlockDim), Dim3Ty);
382   LaunchKernelArgs.add(RValue::get(KernelArgs.getPointer()),
383                        cudaLaunchKernelFD->getParamDecl(3)->getType());
384   LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(ShmemSize)),
385                        cudaLaunchKernelFD->getParamDecl(4)->getType());
386   LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(Stream)),
387                        cudaLaunchKernelFD->getParamDecl(5)->getType());
388 
389   QualType QT = cudaLaunchKernelFD->getType();
390   QualType CQT = QT.getCanonicalType();
391   llvm::Type *Ty = CGM.getTypes().ConvertType(CQT);
392   llvm::FunctionType *FTy = dyn_cast<llvm::FunctionType>(Ty);
393 
394   const CGFunctionInfo &FI =
395       CGM.getTypes().arrangeFunctionDeclaration(cudaLaunchKernelFD);
396   llvm::FunctionCallee cudaLaunchKernelFn =
397       CGM.CreateRuntimeFunction(FTy, LaunchKernelName);
398   CGF.EmitCall(FI, CGCallee::forDirect(cudaLaunchKernelFn), ReturnValueSlot(),
399                LaunchKernelArgs);
400   CGF.EmitBranch(EndBlock);
401 
402   CGF.EmitBlock(EndBlock);
403 }
404 
emitDeviceStubBodyLegacy(CodeGenFunction & CGF,FunctionArgList & Args)405 void CGNVCUDARuntime::emitDeviceStubBodyLegacy(CodeGenFunction &CGF,
406                                                FunctionArgList &Args) {
407   // Emit a call to cudaSetupArgument for each arg in Args.
408   llvm::FunctionCallee cudaSetupArgFn = getSetupArgumentFn();
409   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
410   CharUnits Offset = CharUnits::Zero();
411   for (const VarDecl *A : Args) {
412     auto TInfo = CGM.getContext().getTypeInfoInChars(A->getType());
413     Offset = Offset.alignTo(TInfo.Align);
414     llvm::Value *Args[] = {
415         CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
416                                       VoidPtrTy),
417         llvm::ConstantInt::get(SizeTy, TInfo.Width.getQuantity()),
418         llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
419     };
420     llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
421     llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
422     llvm::Value *CBZero = CGF.Builder.CreateICmpEQ(CB, Zero);
423     llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
424     CGF.Builder.CreateCondBr(CBZero, NextBlock, EndBlock);
425     CGF.EmitBlock(NextBlock);
426     Offset += TInfo.Width;
427   }
428 
429   // Emit the call to cudaLaunch
430   llvm::FunctionCallee cudaLaunchFn = getLaunchFn();
431   llvm::Value *Arg =
432       CGF.Builder.CreatePointerCast(KernelHandles[CGF.CurFn], CharPtrTy);
433   CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
434   CGF.EmitBranch(EndBlock);
435 
436   CGF.EmitBlock(EndBlock);
437 }
438 
439 // Replace the original variable Var with the address loaded from variable
440 // ManagedVar populated by HIP runtime.
replaceManagedVar(llvm::GlobalVariable * Var,llvm::GlobalVariable * ManagedVar)441 static void replaceManagedVar(llvm::GlobalVariable *Var,
442                               llvm::GlobalVariable *ManagedVar) {
443   SmallVector<SmallVector<llvm::User *, 8>, 8> WorkList;
444   for (auto &&VarUse : Var->uses()) {
445     WorkList.push_back({VarUse.getUser()});
446   }
447   while (!WorkList.empty()) {
448     auto &&WorkItem = WorkList.pop_back_val();
449     auto *U = WorkItem.back();
450     if (isa<llvm::ConstantExpr>(U)) {
451       for (auto &&UU : U->uses()) {
452         WorkItem.push_back(UU.getUser());
453         WorkList.push_back(WorkItem);
454         WorkItem.pop_back();
455       }
456       continue;
457     }
458     if (auto *I = dyn_cast<llvm::Instruction>(U)) {
459       llvm::Value *OldV = Var;
460       llvm::Instruction *NewV =
461           new llvm::LoadInst(Var->getType(), ManagedVar, "ld.managed", false,
462                              llvm::Align(Var->getAlignment()), I);
463       WorkItem.pop_back();
464       // Replace constant expressions directly or indirectly using the managed
465       // variable with instructions.
466       for (auto &&Op : WorkItem) {
467         auto *CE = cast<llvm::ConstantExpr>(Op);
468         auto *NewInst = llvm::createReplacementInstr(CE, I);
469         NewInst->replaceUsesOfWith(OldV, NewV);
470         OldV = CE;
471         NewV = NewInst;
472       }
473       I->replaceUsesOfWith(OldV, NewV);
474     } else {
475       llvm_unreachable("Invalid use of managed variable");
476     }
477   }
478 }
479 
480 /// Creates a function that sets up state on the host side for CUDA objects that
481 /// have a presence on both the host and device sides. Specifically, registers
482 /// the host side of kernel functions and device global variables with the CUDA
483 /// runtime.
484 /// \code
485 /// void __cuda_register_globals(void** GpuBinaryHandle) {
486 ///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
487 ///    ...
488 ///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
489 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
490 ///    ...
491 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
492 /// }
493 /// \endcode
makeRegisterGlobalsFn()494 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
495   // No need to register anything
496   if (EmittedKernels.empty() && DeviceVars.empty())
497     return nullptr;
498 
499   llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
500       getRegisterGlobalsFnTy(), llvm::GlobalValue::InternalLinkage,
501       addUnderscoredPrefixToName("_register_globals"), &TheModule);
502   llvm::BasicBlock *EntryBB =
503       llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
504   CGBuilderTy Builder(CGM, Context);
505   Builder.SetInsertPoint(EntryBB);
506 
507   // void __cudaRegisterFunction(void **, const char *, char *, const char *,
508   //                             int, uint3*, uint3*, dim3*, dim3*, int*)
509   llvm::Type *RegisterFuncParams[] = {
510       VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
511       VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
512   llvm::FunctionCallee RegisterFunc = CGM.CreateRuntimeFunction(
513       llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
514       addUnderscoredPrefixToName("RegisterFunction"));
515 
516   // Extract GpuBinaryHandle passed as the first argument passed to
517   // __cuda_register_globals() and generate __cudaRegisterFunction() call for
518   // each emitted kernel.
519   llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
520   for (auto &&I : EmittedKernels) {
521     llvm::Constant *KernelName =
522         makeConstantString(getDeviceSideName(cast<NamedDecl>(I.D)));
523     llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
524     llvm::Value *Args[] = {
525         &GpuBinaryHandlePtr,
526         Builder.CreateBitCast(KernelHandles[I.Kernel], VoidPtrTy),
527         KernelName,
528         KernelName,
529         llvm::ConstantInt::get(IntTy, -1),
530         NullPtr,
531         NullPtr,
532         NullPtr,
533         NullPtr,
534         llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
535     Builder.CreateCall(RegisterFunc, Args);
536   }
537 
538   llvm::Type *VarSizeTy = IntTy;
539   // For HIP or CUDA 9.0+, device variable size is type of `size_t`.
540   if (CGM.getLangOpts().HIP ||
541       ToCudaVersion(CGM.getTarget().getSDKVersion()) >= CudaVersion::CUDA_90)
542     VarSizeTy = SizeTy;
543 
544   // void __cudaRegisterVar(void **, char *, char *, const char *,
545   //                        int, int, int, int)
546   llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
547                                      CharPtrTy,    IntTy,     VarSizeTy,
548                                      IntTy,        IntTy};
549   llvm::FunctionCallee RegisterVar = CGM.CreateRuntimeFunction(
550       llvm::FunctionType::get(VoidTy, RegisterVarParams, false),
551       addUnderscoredPrefixToName("RegisterVar"));
552   // void __hipRegisterManagedVar(void **, char *, char *, const char *,
553   //                              size_t, unsigned)
554   llvm::Type *RegisterManagedVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
555                                             CharPtrTy,    VarSizeTy, IntTy};
556   llvm::FunctionCallee RegisterManagedVar = CGM.CreateRuntimeFunction(
557       llvm::FunctionType::get(VoidTy, RegisterManagedVarParams, false),
558       addUnderscoredPrefixToName("RegisterManagedVar"));
559   // void __cudaRegisterSurface(void **, const struct surfaceReference *,
560   //                            const void **, const char *, int, int);
561   llvm::FunctionCallee RegisterSurf = CGM.CreateRuntimeFunction(
562       llvm::FunctionType::get(
563           VoidTy, {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy},
564           false),
565       addUnderscoredPrefixToName("RegisterSurface"));
566   // void __cudaRegisterTexture(void **, const struct textureReference *,
567   //                            const void **, const char *, int, int, int)
568   llvm::FunctionCallee RegisterTex = CGM.CreateRuntimeFunction(
569       llvm::FunctionType::get(
570           VoidTy,
571           {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy, IntTy},
572           false),
573       addUnderscoredPrefixToName("RegisterTexture"));
574   for (auto &&Info : DeviceVars) {
575     llvm::GlobalVariable *Var = Info.Var;
576     assert((!Var->isDeclaration() || Info.Flags.isManaged()) &&
577            "External variables should not show up here, except HIP managed "
578            "variables");
579     llvm::Constant *VarName = makeConstantString(getDeviceSideName(Info.D));
580     switch (Info.Flags.getKind()) {
581     case DeviceVarFlags::Variable: {
582       uint64_t VarSize =
583           CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
584       if (Info.Flags.isManaged()) {
585         auto ManagedVar = new llvm::GlobalVariable(
586             CGM.getModule(), Var->getType(),
587             /*isConstant=*/false, Var->getLinkage(),
588             /*Init=*/Var->isDeclaration()
589                 ? nullptr
590                 : llvm::ConstantPointerNull::get(Var->getType()),
591             /*Name=*/"", /*InsertBefore=*/nullptr,
592             llvm::GlobalVariable::NotThreadLocal);
593         ManagedVar->setDSOLocal(Var->isDSOLocal());
594         ManagedVar->setVisibility(Var->getVisibility());
595         ManagedVar->setExternallyInitialized(true);
596         ManagedVar->takeName(Var);
597         Var->setName(Twine(ManagedVar->getName() + ".managed"));
598         replaceManagedVar(Var, ManagedVar);
599         llvm::Value *Args[] = {
600             &GpuBinaryHandlePtr,
601             Builder.CreateBitCast(ManagedVar, VoidPtrTy),
602             Builder.CreateBitCast(Var, VoidPtrTy),
603             VarName,
604             llvm::ConstantInt::get(VarSizeTy, VarSize),
605             llvm::ConstantInt::get(IntTy, Var->getAlignment())};
606         if (!Var->isDeclaration())
607           Builder.CreateCall(RegisterManagedVar, Args);
608       } else {
609         llvm::Value *Args[] = {
610             &GpuBinaryHandlePtr,
611             Builder.CreateBitCast(Var, VoidPtrTy),
612             VarName,
613             VarName,
614             llvm::ConstantInt::get(IntTy, Info.Flags.isExtern()),
615             llvm::ConstantInt::get(VarSizeTy, VarSize),
616             llvm::ConstantInt::get(IntTy, Info.Flags.isConstant()),
617             llvm::ConstantInt::get(IntTy, 0)};
618         Builder.CreateCall(RegisterVar, Args);
619       }
620       break;
621     }
622     case DeviceVarFlags::Surface:
623       Builder.CreateCall(
624           RegisterSurf,
625           {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
626            VarName, llvm::ConstantInt::get(IntTy, Info.Flags.getSurfTexType()),
627            llvm::ConstantInt::get(IntTy, Info.Flags.isExtern())});
628       break;
629     case DeviceVarFlags::Texture:
630       Builder.CreateCall(
631           RegisterTex,
632           {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
633            VarName, llvm::ConstantInt::get(IntTy, Info.Flags.getSurfTexType()),
634            llvm::ConstantInt::get(IntTy, Info.Flags.isNormalized()),
635            llvm::ConstantInt::get(IntTy, Info.Flags.isExtern())});
636       break;
637     }
638   }
639 
640   Builder.CreateRetVoid();
641   return RegisterKernelsFunc;
642 }
643 
644 /// Creates a global constructor function for the module:
645 ///
646 /// For CUDA:
647 /// \code
648 /// void __cuda_module_ctor(void*) {
649 ///     Handle = __cudaRegisterFatBinary(GpuBinaryBlob);
650 ///     __cuda_register_globals(Handle);
651 /// }
652 /// \endcode
653 ///
654 /// For HIP:
655 /// \code
656 /// void __hip_module_ctor(void*) {
657 ///     if (__hip_gpubin_handle == 0) {
658 ///         __hip_gpubin_handle  = __hipRegisterFatBinary(GpuBinaryBlob);
659 ///         __hip_register_globals(__hip_gpubin_handle);
660 ///     }
661 /// }
662 /// \endcode
makeModuleCtorFunction()663 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
664   bool IsHIP = CGM.getLangOpts().HIP;
665   bool IsCUDA = CGM.getLangOpts().CUDA;
666   // No need to generate ctors/dtors if there is no GPU binary.
667   StringRef CudaGpuBinaryFileName = CGM.getCodeGenOpts().CudaGpuBinaryFileName;
668   if (CudaGpuBinaryFileName.empty() && !IsHIP)
669     return nullptr;
670   if ((IsHIP || (IsCUDA && !RelocatableDeviceCode)) && EmittedKernels.empty() &&
671       DeviceVars.empty())
672     return nullptr;
673 
674   // void __{cuda|hip}_register_globals(void* handle);
675   llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
676   // We always need a function to pass in as callback. Create a dummy
677   // implementation if we don't need to register anything.
678   if (RelocatableDeviceCode && !RegisterGlobalsFunc)
679     RegisterGlobalsFunc = makeDummyFunction(getRegisterGlobalsFnTy());
680 
681   // void ** __{cuda|hip}RegisterFatBinary(void *);
682   llvm::FunctionCallee RegisterFatbinFunc = CGM.CreateRuntimeFunction(
683       llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
684       addUnderscoredPrefixToName("RegisterFatBinary"));
685   // struct { int magic, int version, void * gpu_binary, void * dont_care };
686   llvm::StructType *FatbinWrapperTy =
687       llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy);
688 
689   // Register GPU binary with the CUDA runtime, store returned handle in a
690   // global variable and save a reference in GpuBinaryHandle to be cleaned up
691   // in destructor on exit. Then associate all known kernels with the GPU binary
692   // handle so CUDA runtime can figure out what to call on the GPU side.
693   std::unique_ptr<llvm::MemoryBuffer> CudaGpuBinary = nullptr;
694   if (!CudaGpuBinaryFileName.empty()) {
695     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CudaGpuBinaryOrErr =
696         llvm::MemoryBuffer::getFileOrSTDIN(CudaGpuBinaryFileName);
697     if (std::error_code EC = CudaGpuBinaryOrErr.getError()) {
698       CGM.getDiags().Report(diag::err_cannot_open_file)
699           << CudaGpuBinaryFileName << EC.message();
700       return nullptr;
701     }
702     CudaGpuBinary = std::move(CudaGpuBinaryOrErr.get());
703   }
704 
705   llvm::Function *ModuleCtorFunc = llvm::Function::Create(
706       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
707       llvm::GlobalValue::InternalLinkage,
708       addUnderscoredPrefixToName("_module_ctor"), &TheModule);
709   llvm::BasicBlock *CtorEntryBB =
710       llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
711   CGBuilderTy CtorBuilder(CGM, Context);
712 
713   CtorBuilder.SetInsertPoint(CtorEntryBB);
714 
715   const char *FatbinConstantName;
716   const char *FatbinSectionName;
717   const char *ModuleIDSectionName;
718   StringRef ModuleIDPrefix;
719   llvm::Constant *FatBinStr;
720   unsigned FatMagic;
721   if (IsHIP) {
722     FatbinConstantName = ".hip_fatbin";
723     FatbinSectionName = ".hipFatBinSegment";
724 
725     ModuleIDSectionName = "__hip_module_id";
726     ModuleIDPrefix = "__hip_";
727 
728     if (CudaGpuBinary) {
729       // If fatbin is available from early finalization, create a string
730       // literal containing the fat binary loaded from the given file.
731       const unsigned HIPCodeObjectAlign = 4096;
732       FatBinStr =
733           makeConstantString(std::string(CudaGpuBinary->getBuffer()), "",
734                              FatbinConstantName, HIPCodeObjectAlign);
735     } else {
736       // If fatbin is not available, create an external symbol
737       // __hip_fatbin in section .hip_fatbin. The external symbol is supposed
738       // to contain the fat binary but will be populated somewhere else,
739       // e.g. by lld through link script.
740       FatBinStr = new llvm::GlobalVariable(
741         CGM.getModule(), CGM.Int8Ty,
742         /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
743         "__hip_fatbin", nullptr,
744         llvm::GlobalVariable::NotThreadLocal);
745       cast<llvm::GlobalVariable>(FatBinStr)->setSection(FatbinConstantName);
746     }
747 
748     FatMagic = HIPFatMagic;
749   } else {
750     if (RelocatableDeviceCode)
751       FatbinConstantName = CGM.getTriple().isMacOSX()
752                                ? "__NV_CUDA,__nv_relfatbin"
753                                : "__nv_relfatbin";
754     else
755       FatbinConstantName =
756           CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
757     // NVIDIA's cuobjdump looks for fatbins in this section.
758     FatbinSectionName =
759         CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
760 
761     ModuleIDSectionName = CGM.getTriple().isMacOSX()
762                               ? "__NV_CUDA,__nv_module_id"
763                               : "__nv_module_id";
764     ModuleIDPrefix = "__nv_";
765 
766     // For CUDA, create a string literal containing the fat binary loaded from
767     // the given file.
768     FatBinStr = makeConstantString(std::string(CudaGpuBinary->getBuffer()), "",
769                                    FatbinConstantName, 8);
770     FatMagic = CudaFatMagic;
771   }
772 
773   // Create initialized wrapper structure that points to the loaded GPU binary
774   ConstantInitBuilder Builder(CGM);
775   auto Values = Builder.beginStruct(FatbinWrapperTy);
776   // Fatbin wrapper magic.
777   Values.addInt(IntTy, FatMagic);
778   // Fatbin version.
779   Values.addInt(IntTy, 1);
780   // Data.
781   Values.add(FatBinStr);
782   // Unused in fatbin v1.
783   Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
784   llvm::GlobalVariable *FatbinWrapper = Values.finishAndCreateGlobal(
785       addUnderscoredPrefixToName("_fatbin_wrapper"), CGM.getPointerAlign(),
786       /*constant*/ true);
787   FatbinWrapper->setSection(FatbinSectionName);
788 
789   // There is only one HIP fat binary per linked module, however there are
790   // multiple constructor functions. Make sure the fat binary is registered
791   // only once. The constructor functions are executed by the dynamic loader
792   // before the program gains control. The dynamic loader cannot execute the
793   // constructor functions concurrently since doing that would not guarantee
794   // thread safety of the loaded program. Therefore we can assume sequential
795   // execution of constructor functions here.
796   if (IsHIP) {
797     auto Linkage = CudaGpuBinary ? llvm::GlobalValue::InternalLinkage :
798         llvm::GlobalValue::LinkOnceAnyLinkage;
799     llvm::BasicBlock *IfBlock =
800         llvm::BasicBlock::Create(Context, "if", ModuleCtorFunc);
801     llvm::BasicBlock *ExitBlock =
802         llvm::BasicBlock::Create(Context, "exit", ModuleCtorFunc);
803     // The name, size, and initialization pattern of this variable is part
804     // of HIP ABI.
805     GpuBinaryHandle = new llvm::GlobalVariable(
806         TheModule, VoidPtrPtrTy, /*isConstant=*/false,
807         Linkage,
808         /*Initializer=*/llvm::ConstantPointerNull::get(VoidPtrPtrTy),
809         "__hip_gpubin_handle");
810     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
811     // Prevent the weak symbol in different shared libraries being merged.
812     if (Linkage != llvm::GlobalValue::InternalLinkage)
813       GpuBinaryHandle->setVisibility(llvm::GlobalValue::HiddenVisibility);
814     Address GpuBinaryAddr(
815         GpuBinaryHandle,
816         CharUnits::fromQuantity(GpuBinaryHandle->getAlignment()));
817     {
818       auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
819       llvm::Constant *Zero =
820           llvm::Constant::getNullValue(HandleValue->getType());
821       llvm::Value *EQZero = CtorBuilder.CreateICmpEQ(HandleValue, Zero);
822       CtorBuilder.CreateCondBr(EQZero, IfBlock, ExitBlock);
823     }
824     {
825       CtorBuilder.SetInsertPoint(IfBlock);
826       // GpuBinaryHandle = __hipRegisterFatBinary(&FatbinWrapper);
827       llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
828           RegisterFatbinFunc,
829           CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
830       CtorBuilder.CreateStore(RegisterFatbinCall, GpuBinaryAddr);
831       CtorBuilder.CreateBr(ExitBlock);
832     }
833     {
834       CtorBuilder.SetInsertPoint(ExitBlock);
835       // Call __hip_register_globals(GpuBinaryHandle);
836       if (RegisterGlobalsFunc) {
837         auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
838         CtorBuilder.CreateCall(RegisterGlobalsFunc, HandleValue);
839       }
840     }
841   } else if (!RelocatableDeviceCode) {
842     // Register binary with CUDA runtime. This is substantially different in
843     // default mode vs. separate compilation!
844     // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
845     llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
846         RegisterFatbinFunc,
847         CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
848     GpuBinaryHandle = new llvm::GlobalVariable(
849         TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
850         llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
851     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
852     CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
853                                    CGM.getPointerAlign());
854 
855     // Call __cuda_register_globals(GpuBinaryHandle);
856     if (RegisterGlobalsFunc)
857       CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
858 
859     // Call __cudaRegisterFatBinaryEnd(Handle) if this CUDA version needs it.
860     if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
861                            CudaFeature::CUDA_USES_FATBIN_REGISTER_END)) {
862       // void __cudaRegisterFatBinaryEnd(void **);
863       llvm::FunctionCallee RegisterFatbinEndFunc = CGM.CreateRuntimeFunction(
864           llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
865           "__cudaRegisterFatBinaryEnd");
866       CtorBuilder.CreateCall(RegisterFatbinEndFunc, RegisterFatbinCall);
867     }
868   } else {
869     // Generate a unique module ID.
870     SmallString<64> ModuleID;
871     llvm::raw_svector_ostream OS(ModuleID);
872     OS << ModuleIDPrefix << llvm::format("%" PRIx64, FatbinWrapper->getGUID());
873     llvm::Constant *ModuleIDConstant = makeConstantString(
874         std::string(ModuleID.str()), "", ModuleIDSectionName, 32);
875 
876     // Create an alias for the FatbinWrapper that nvcc will look for.
877     llvm::GlobalAlias::create(llvm::GlobalValue::ExternalLinkage,
878                               Twine("__fatbinwrap") + ModuleID, FatbinWrapper);
879 
880     // void __cudaRegisterLinkedBinary%ModuleID%(void (*)(void *), void *,
881     // void *, void (*)(void **))
882     SmallString<128> RegisterLinkedBinaryName("__cudaRegisterLinkedBinary");
883     RegisterLinkedBinaryName += ModuleID;
884     llvm::FunctionCallee RegisterLinkedBinaryFunc = CGM.CreateRuntimeFunction(
885         getRegisterLinkedBinaryFnTy(), RegisterLinkedBinaryName);
886 
887     assert(RegisterGlobalsFunc && "Expecting at least dummy function!");
888     llvm::Value *Args[] = {RegisterGlobalsFunc,
889                            CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy),
890                            ModuleIDConstant,
891                            makeDummyFunction(getCallbackFnTy())};
892     CtorBuilder.CreateCall(RegisterLinkedBinaryFunc, Args);
893   }
894 
895   // Create destructor and register it with atexit() the way NVCC does it. Doing
896   // it during regular destructor phase worked in CUDA before 9.2 but results in
897   // double-free in 9.2.
898   if (llvm::Function *CleanupFn = makeModuleDtorFunction()) {
899     // extern "C" int atexit(void (*f)(void));
900     llvm::FunctionType *AtExitTy =
901         llvm::FunctionType::get(IntTy, CleanupFn->getType(), false);
902     llvm::FunctionCallee AtExitFunc =
903         CGM.CreateRuntimeFunction(AtExitTy, "atexit", llvm::AttributeList(),
904                                   /*Local=*/true);
905     CtorBuilder.CreateCall(AtExitFunc, CleanupFn);
906   }
907 
908   CtorBuilder.CreateRetVoid();
909   return ModuleCtorFunc;
910 }
911 
912 /// Creates a global destructor function that unregisters the GPU code blob
913 /// registered by constructor.
914 ///
915 /// For CUDA:
916 /// \code
917 /// void __cuda_module_dtor(void*) {
918 ///     __cudaUnregisterFatBinary(Handle);
919 /// }
920 /// \endcode
921 ///
922 /// For HIP:
923 /// \code
924 /// void __hip_module_dtor(void*) {
925 ///     if (__hip_gpubin_handle) {
926 ///         __hipUnregisterFatBinary(__hip_gpubin_handle);
927 ///         __hip_gpubin_handle = 0;
928 ///     }
929 /// }
930 /// \endcode
makeModuleDtorFunction()931 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
932   // No need for destructor if we don't have a handle to unregister.
933   if (!GpuBinaryHandle)
934     return nullptr;
935 
936   // void __cudaUnregisterFatBinary(void ** handle);
937   llvm::FunctionCallee UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
938       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
939       addUnderscoredPrefixToName("UnregisterFatBinary"));
940 
941   llvm::Function *ModuleDtorFunc = llvm::Function::Create(
942       llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
943       llvm::GlobalValue::InternalLinkage,
944       addUnderscoredPrefixToName("_module_dtor"), &TheModule);
945 
946   llvm::BasicBlock *DtorEntryBB =
947       llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
948   CGBuilderTy DtorBuilder(CGM, Context);
949   DtorBuilder.SetInsertPoint(DtorEntryBB);
950 
951   Address GpuBinaryAddr(GpuBinaryHandle, CharUnits::fromQuantity(
952                                              GpuBinaryHandle->getAlignment()));
953   auto HandleValue = DtorBuilder.CreateLoad(GpuBinaryAddr);
954   // There is only one HIP fat binary per linked module, however there are
955   // multiple destructor functions. Make sure the fat binary is unregistered
956   // only once.
957   if (CGM.getLangOpts().HIP) {
958     llvm::BasicBlock *IfBlock =
959         llvm::BasicBlock::Create(Context, "if", ModuleDtorFunc);
960     llvm::BasicBlock *ExitBlock =
961         llvm::BasicBlock::Create(Context, "exit", ModuleDtorFunc);
962     llvm::Constant *Zero = llvm::Constant::getNullValue(HandleValue->getType());
963     llvm::Value *NEZero = DtorBuilder.CreateICmpNE(HandleValue, Zero);
964     DtorBuilder.CreateCondBr(NEZero, IfBlock, ExitBlock);
965 
966     DtorBuilder.SetInsertPoint(IfBlock);
967     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
968     DtorBuilder.CreateStore(Zero, GpuBinaryAddr);
969     DtorBuilder.CreateBr(ExitBlock);
970 
971     DtorBuilder.SetInsertPoint(ExitBlock);
972   } else {
973     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
974   }
975   DtorBuilder.CreateRetVoid();
976   return ModuleDtorFunc;
977 }
978 
CreateNVCUDARuntime(CodeGenModule & CGM)979 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
980   return new CGNVCUDARuntime(CGM);
981 }
982 
internalizeDeviceSideVar(const VarDecl * D,llvm::GlobalValue::LinkageTypes & Linkage)983 void CGNVCUDARuntime::internalizeDeviceSideVar(
984     const VarDecl *D, llvm::GlobalValue::LinkageTypes &Linkage) {
985   // For -fno-gpu-rdc, host-side shadows of external declarations of device-side
986   // global variables become internal definitions. These have to be internal in
987   // order to prevent name conflicts with global host variables with the same
988   // name in a different TUs.
989   //
990   // For -fgpu-rdc, the shadow variables should not be internalized because
991   // they may be accessed by different TU.
992   if (CGM.getLangOpts().GPURelocatableDeviceCode)
993     return;
994 
995   // __shared__ variables are odd. Shadows do get created, but
996   // they are not registered with the CUDA runtime, so they
997   // can't really be used to access their device-side
998   // counterparts. It's not clear yet whether it's nvcc's bug or
999   // a feature, but we've got to do the same for compatibility.
1000   if (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>() ||
1001       D->hasAttr<CUDASharedAttr>() ||
1002       D->getType()->isCUDADeviceBuiltinSurfaceType() ||
1003       D->getType()->isCUDADeviceBuiltinTextureType()) {
1004     Linkage = llvm::GlobalValue::InternalLinkage;
1005   }
1006 }
1007 
handleVarRegistration(const VarDecl * D,llvm::GlobalVariable & GV)1008 void CGNVCUDARuntime::handleVarRegistration(const VarDecl *D,
1009                                             llvm::GlobalVariable &GV) {
1010   if (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>()) {
1011     // Shadow variables and their properties must be registered with CUDA
1012     // runtime. Skip Extern global variables, which will be registered in
1013     // the TU where they are defined.
1014     //
1015     // Don't register a C++17 inline variable. The local symbol can be
1016     // discarded and referencing a discarded local symbol from outside the
1017     // comdat (__cuda_register_globals) is disallowed by the ELF spec.
1018     //
1019     // HIP managed variables need to be always recorded in device and host
1020     // compilations for transformation.
1021     //
1022     // HIP managed variables and variables in CUDADeviceVarODRUsedByHost are
1023     // added to llvm.compiler-used, therefore they are safe to be registered.
1024     if ((!D->hasExternalStorage() && !D->isInline()) ||
1025         CGM.getContext().CUDADeviceVarODRUsedByHost.contains(D) ||
1026         D->hasAttr<HIPManagedAttr>()) {
1027       registerDeviceVar(D, GV, !D->hasDefinition(),
1028                         D->hasAttr<CUDAConstantAttr>());
1029     }
1030   } else if (D->getType()->isCUDADeviceBuiltinSurfaceType() ||
1031              D->getType()->isCUDADeviceBuiltinTextureType()) {
1032     // Builtin surfaces and textures and their template arguments are
1033     // also registered with CUDA runtime.
1034     const auto *TD = cast<ClassTemplateSpecializationDecl>(
1035         D->getType()->castAs<RecordType>()->getDecl());
1036     const TemplateArgumentList &Args = TD->getTemplateArgs();
1037     if (TD->hasAttr<CUDADeviceBuiltinSurfaceTypeAttr>()) {
1038       assert(Args.size() == 2 &&
1039              "Unexpected number of template arguments of CUDA device "
1040              "builtin surface type.");
1041       auto SurfType = Args[1].getAsIntegral();
1042       if (!D->hasExternalStorage())
1043         registerDeviceSurf(D, GV, !D->hasDefinition(), SurfType.getSExtValue());
1044     } else {
1045       assert(Args.size() == 3 &&
1046              "Unexpected number of template arguments of CUDA device "
1047              "builtin texture type.");
1048       auto TexType = Args[1].getAsIntegral();
1049       auto Normalized = Args[2].getAsIntegral();
1050       if (!D->hasExternalStorage())
1051         registerDeviceTex(D, GV, !D->hasDefinition(), TexType.getSExtValue(),
1052                           Normalized.getZExtValue());
1053     }
1054   }
1055 }
1056 
1057 // Transform managed variables to pointers to managed variables in device code.
1058 // Each use of the original managed variable is replaced by a load from the
1059 // transformed managed variable. The transformed managed variable contains
1060 // the address of managed memory which will be allocated by the runtime.
transformManagedVars()1061 void CGNVCUDARuntime::transformManagedVars() {
1062   for (auto &&Info : DeviceVars) {
1063     llvm::GlobalVariable *Var = Info.Var;
1064     if (Info.Flags.getKind() == DeviceVarFlags::Variable &&
1065         Info.Flags.isManaged()) {
1066       auto ManagedVar = new llvm::GlobalVariable(
1067           CGM.getModule(), Var->getType(),
1068           /*isConstant=*/false, Var->getLinkage(),
1069           /*Init=*/Var->isDeclaration()
1070               ? nullptr
1071               : llvm::ConstantPointerNull::get(Var->getType()),
1072           /*Name=*/"", /*InsertBefore=*/nullptr,
1073           llvm::GlobalVariable::NotThreadLocal,
1074           CGM.getContext().getTargetAddressSpace(LangAS::cuda_device));
1075       ManagedVar->setDSOLocal(Var->isDSOLocal());
1076       ManagedVar->setVisibility(Var->getVisibility());
1077       ManagedVar->setExternallyInitialized(true);
1078       replaceManagedVar(Var, ManagedVar);
1079       ManagedVar->takeName(Var);
1080       Var->setName(Twine(ManagedVar->getName()) + ".managed");
1081       // Keep managed variables even if they are not used in device code since
1082       // they need to be allocated by the runtime.
1083       if (!Var->isDeclaration()) {
1084         assert(!ManagedVar->isDeclaration());
1085         CGM.addCompilerUsedGlobal(Var);
1086         CGM.addCompilerUsedGlobal(ManagedVar);
1087       }
1088     }
1089   }
1090 }
1091 
1092 // Returns module constructor to be added.
finalizeModule()1093 llvm::Function *CGNVCUDARuntime::finalizeModule() {
1094   if (CGM.getLangOpts().CUDAIsDevice) {
1095     transformManagedVars();
1096 
1097     // Mark ODR-used device variables as compiler used to prevent it from being
1098     // eliminated by optimization. This is necessary for device variables
1099     // ODR-used by host functions. Sema correctly marks them as ODR-used no
1100     // matter whether they are ODR-used by device or host functions.
1101     //
1102     // We do not need to do this if the variable has used attribute since it
1103     // has already been added.
1104     //
1105     // Static device variables have been externalized at this point, therefore
1106     // variables with LLVM private or internal linkage need not be added.
1107     for (auto &&Info : DeviceVars) {
1108       auto Kind = Info.Flags.getKind();
1109       if (!Info.Var->isDeclaration() &&
1110           !llvm::GlobalValue::isLocalLinkage(Info.Var->getLinkage()) &&
1111           (Kind == DeviceVarFlags::Variable ||
1112            Kind == DeviceVarFlags::Surface ||
1113            Kind == DeviceVarFlags::Texture) &&
1114           Info.D->isUsed() && !Info.D->hasAttr<UsedAttr>()) {
1115         CGM.addCompilerUsedGlobal(Info.Var);
1116       }
1117     }
1118     return nullptr;
1119   }
1120   return makeModuleCtorFunction();
1121 }
1122 
getKernelHandle(llvm::Function * F,GlobalDecl GD)1123 llvm::GlobalValue *CGNVCUDARuntime::getKernelHandle(llvm::Function *F,
1124                                                     GlobalDecl GD) {
1125   auto Loc = KernelHandles.find(F);
1126   if (Loc != KernelHandles.end())
1127     return Loc->second;
1128 
1129   if (!CGM.getLangOpts().HIP) {
1130     KernelHandles[F] = F;
1131     KernelStubs[F] = F;
1132     return F;
1133   }
1134 
1135   auto *Var = new llvm::GlobalVariable(
1136       TheModule, F->getType(), /*isConstant=*/true, F->getLinkage(),
1137       /*Initializer=*/nullptr,
1138       CGM.getMangledName(
1139           GD.getWithKernelReferenceKind(KernelReferenceKind::Kernel)));
1140   Var->setAlignment(CGM.getPointerAlign().getAsAlign());
1141   Var->setDSOLocal(F->isDSOLocal());
1142   Var->setVisibility(F->getVisibility());
1143   KernelHandles[F] = Var;
1144   KernelStubs[Var] = F;
1145   return Var;
1146 }
1147