1//===----- CGCUDANV.cpp - Interface to NVIDIA CUDA Runtime ----------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This provides a class for CUDA code generation targeting the NVIDIA CUDA
10// runtime library.
11//
12//===----------------------------------------------------------------------===//
13
14#include "CGCUDARuntime.h"
15#include "CodeGenFunction.h"
16#include "CodeGenModule.h"
17#include "clang/AST/Decl.h"
18#include "clang/Basic/Cuda.h"
19#include "clang/CodeGen/CodeGenABITypes.h"
20#include "clang/CodeGen/ConstantInitBuilder.h"
21#include "llvm/IR/BasicBlock.h"
22#include "llvm/IR/Constants.h"
23#include "llvm/IR/DerivedTypes.h"
24#include "llvm/Support/Format.h"
25
26using namespace clang;
27using namespace CodeGen;
28
29namespace {
30constexpr unsigned CudaFatMagic = 0x466243b1;
31constexpr unsigned HIPFatMagic = 0x48495046; // "HIPF"
32
33class CGNVCUDARuntime : public CGCUDARuntime {
34
35private:
36  llvm::IntegerType *IntTy, *SizeTy;
37  llvm::Type *VoidTy;
38  llvm::PointerType *CharPtrTy, *VoidPtrTy, *VoidPtrPtrTy;
39
40  /// Convenience reference to LLVM Context
41  llvm::LLVMContext &Context;
42  /// Convenience reference to the current module
43  llvm::Module &TheModule;
44  /// Keeps track of kernel launch stubs emitted in this module
45  struct KernelInfo {
46    llvm::Function *Kernel;
47    const Decl *D;
48  };
49  llvm::SmallVector<KernelInfo, 16> EmittedKernels;
50  struct VarInfo {
51    llvm::GlobalVariable *Var;
52    const VarDecl *D;
53    unsigned Flag;
54  };
55  llvm::SmallVector<VarInfo, 16> DeviceVars;
56  /// Keeps track of variable containing handle of GPU binary. Populated by
57  /// ModuleCtorFunction() and used to create corresponding cleanup calls in
58  /// ModuleDtorFunction()
59  llvm::GlobalVariable *GpuBinaryHandle = nullptr;
60  /// Whether we generate relocatable device code.
61  bool RelocatableDeviceCode;
62  /// Mangle context for device.
63  std::unique_ptr<MangleContext> DeviceMC;
64
65  llvm::FunctionCallee getSetupArgumentFn() const;
66  llvm::FunctionCallee getLaunchFn() const;
67
68  llvm::FunctionType *getRegisterGlobalsFnTy() const;
69  llvm::FunctionType *getCallbackFnTy() const;
70  llvm::FunctionType *getRegisterLinkedBinaryFnTy() const;
71  std::string addPrefixToName(StringRef FuncName) const;
72  std::string addUnderscoredPrefixToName(StringRef FuncName) const;
73
74  /// Creates a function to register all kernel stubs generated in this module.
75  llvm::Function *makeRegisterGlobalsFn();
76
77  /// Helper function that generates a constant string and returns a pointer to
78  /// the start of the string.  The result of this function can be used anywhere
79  /// where the C code specifies const char*.
80  llvm::Constant *makeConstantString(const std::string &Str,
81                                     const std::string &Name = "",
82                                     const std::string &SectionName = "",
83                                     unsigned Alignment = 0) {
84    llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
85                               llvm::ConstantInt::get(SizeTy, 0)};
86    auto ConstStr = CGM.GetAddrOfConstantCString(Str, Name.c_str());
87    llvm::GlobalVariable *GV =
88        cast<llvm::GlobalVariable>(ConstStr.getPointer());
89    if (!SectionName.empty()) {
90      GV->setSection(SectionName);
91      // Mark the address as used which make sure that this section isn't
92      // merged and we will really have it in the object file.
93      GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::None);
94    }
95    if (Alignment)
96      GV->setAlignment(llvm::Align(Alignment));
97
98    return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
99                                                ConstStr.getPointer(), Zeros);
100  }
101
102  /// Helper function that generates an empty dummy function returning void.
103  llvm::Function *makeDummyFunction(llvm::FunctionType *FnTy) {
104    assert(FnTy->getReturnType()->isVoidTy() &&
105           "Can only generate dummy functions returning void!");
106    llvm::Function *DummyFunc = llvm::Function::Create(
107        FnTy, llvm::GlobalValue::InternalLinkage, "dummy", &TheModule);
108
109    llvm::BasicBlock *DummyBlock =
110        llvm::BasicBlock::Create(Context, "", DummyFunc);
111    CGBuilderTy FuncBuilder(CGM, Context);
112    FuncBuilder.SetInsertPoint(DummyBlock);
113    FuncBuilder.CreateRetVoid();
114
115    return DummyFunc;
116  }
117
118  void emitDeviceStubBodyLegacy(CodeGenFunction &CGF, FunctionArgList &Args);
119  void emitDeviceStubBodyNew(CodeGenFunction &CGF, FunctionArgList &Args);
120  std::string getDeviceSideName(const Decl *ND);
121
122public:
123  CGNVCUDARuntime(CodeGenModule &CGM);
124
125  void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
126  void registerDeviceVar(const VarDecl *VD, llvm::GlobalVariable &Var,
127                         unsigned Flags) override {
128    DeviceVars.push_back({&Var, VD, Flags});
129  }
130
131  /// Creates module constructor function
132  llvm::Function *makeModuleCtorFunction() override;
133  /// Creates module destructor function
134  llvm::Function *makeModuleDtorFunction() override;
135  /// Construct and return the stub name of a kernel.
136  std::string getDeviceStubName(llvm::StringRef Name) const override;
137};
138
139}
140
141std::string CGNVCUDARuntime::addPrefixToName(StringRef FuncName) const {
142  if (CGM.getLangOpts().HIP)
143    return ((Twine("hip") + Twine(FuncName)).str());
144  return ((Twine("cuda") + Twine(FuncName)).str());
145}
146std::string
147CGNVCUDARuntime::addUnderscoredPrefixToName(StringRef FuncName) const {
148  if (CGM.getLangOpts().HIP)
149    return ((Twine("__hip") + Twine(FuncName)).str());
150  return ((Twine("__cuda") + Twine(FuncName)).str());
151}
152
153CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
154    : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
155      TheModule(CGM.getModule()),
156      RelocatableDeviceCode(CGM.getLangOpts().GPURelocatableDeviceCode),
157      DeviceMC(CGM.getContext().createMangleContext(
158          CGM.getContext().getAuxTargetInfo())) {
159  CodeGen::CodeGenTypes &Types = CGM.getTypes();
160  ASTContext &Ctx = CGM.getContext();
161
162  IntTy = CGM.IntTy;
163  SizeTy = CGM.SizeTy;
164  VoidTy = CGM.VoidTy;
165
166  CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
167  VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
168  VoidPtrPtrTy = VoidPtrTy->getPointerTo();
169}
170
171llvm::FunctionCallee CGNVCUDARuntime::getSetupArgumentFn() const {
172  // cudaError_t cudaSetupArgument(void *, size_t, size_t)
173  llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
174  return CGM.CreateRuntimeFunction(
175      llvm::FunctionType::get(IntTy, Params, false),
176      addPrefixToName("SetupArgument"));
177}
178
179llvm::FunctionCallee CGNVCUDARuntime::getLaunchFn() const {
180  if (CGM.getLangOpts().HIP) {
181    // hipError_t hipLaunchByPtr(char *);
182    return CGM.CreateRuntimeFunction(
183        llvm::FunctionType::get(IntTy, CharPtrTy, false), "hipLaunchByPtr");
184  } else {
185    // cudaError_t cudaLaunch(char *);
186    return CGM.CreateRuntimeFunction(
187        llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
188  }
189}
190
191llvm::FunctionType *CGNVCUDARuntime::getRegisterGlobalsFnTy() const {
192  return llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false);
193}
194
195llvm::FunctionType *CGNVCUDARuntime::getCallbackFnTy() const {
196  return llvm::FunctionType::get(VoidTy, VoidPtrTy, false);
197}
198
199llvm::FunctionType *CGNVCUDARuntime::getRegisterLinkedBinaryFnTy() const {
200  auto CallbackFnTy = getCallbackFnTy();
201  auto RegisterGlobalsFnTy = getRegisterGlobalsFnTy();
202  llvm::Type *Params[] = {RegisterGlobalsFnTy->getPointerTo(), VoidPtrTy,
203                          VoidPtrTy, CallbackFnTy->getPointerTo()};
204  return llvm::FunctionType::get(VoidTy, Params, false);
205}
206
207std::string CGNVCUDARuntime::getDeviceSideName(const Decl *D) {
208  auto *ND = cast<const NamedDecl>(D);
209  std::string DeviceSideName;
210  if (DeviceMC->shouldMangleDeclName(ND)) {
211    SmallString<256> Buffer;
212    llvm::raw_svector_ostream Out(Buffer);
213    DeviceMC->mangleName(ND, Out);
214    DeviceSideName = Out.str();
215  } else
216    DeviceSideName = ND->getIdentifier()->getName();
217  return DeviceSideName;
218}
219
220void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
221                                     FunctionArgList &Args) {
222  // Ensure either we have different ABIs between host and device compilations,
223  // says host compilation following MSVC ABI but device compilation follows
224  // Itanium C++ ABI or, if they follow the same ABI, kernel names after
225  // mangling should be the same after name stubbing. The later checking is
226  // very important as the device kernel name being mangled in host-compilation
227  // is used to resolve the device binaries to be executed. Inconsistent naming
228  // result in undefined behavior. Even though we cannot check that naming
229  // directly between host- and device-compilations, the host- and
230  // device-mangling in host compilation could help catching certain ones.
231  assert((CGF.CGM.getContext().getAuxTargetInfo() &&
232          (CGF.CGM.getContext().getAuxTargetInfo()->getCXXABI() !=
233           CGF.CGM.getContext().getTargetInfo().getCXXABI())) ||
234         getDeviceStubName(getDeviceSideName(CGF.CurFuncDecl)) ==
235             CGF.CurFn->getName());
236
237  EmittedKernels.push_back({CGF.CurFn, CGF.CurFuncDecl});
238  if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
239                         CudaFeature::CUDA_USES_NEW_LAUNCH) ||
240      CGF.getLangOpts().HIPUseNewLaunchAPI)
241    emitDeviceStubBodyNew(CGF, Args);
242  else
243    emitDeviceStubBodyLegacy(CGF, Args);
244}
245
246// CUDA 9.0+ uses new way to launch kernels. Parameters are packed in a local
247// array and kernels are launched using cudaLaunchKernel().
248void CGNVCUDARuntime::emitDeviceStubBodyNew(CodeGenFunction &CGF,
249                                            FunctionArgList &Args) {
250  // Build the shadow stack entry at the very start of the function.
251
252  // Calculate amount of space we will need for all arguments.  If we have no
253  // args, allocate a single pointer so we still have a valid pointer to the
254  // argument array that we can pass to runtime, even if it will be unused.
255  Address KernelArgs = CGF.CreateTempAlloca(
256      VoidPtrTy, CharUnits::fromQuantity(16), "kernel_args",
257      llvm::ConstantInt::get(SizeTy, std::max<size_t>(1, Args.size())));
258  // Store pointers to the arguments in a locally allocated launch_args.
259  for (unsigned i = 0; i < Args.size(); ++i) {
260    llvm::Value* VarPtr = CGF.GetAddrOfLocalVar(Args[i]).getPointer();
261    llvm::Value *VoidVarPtr = CGF.Builder.CreatePointerCast(VarPtr, VoidPtrTy);
262    CGF.Builder.CreateDefaultAlignedStore(
263        VoidVarPtr, CGF.Builder.CreateConstGEP1_32(KernelArgs.getPointer(), i));
264  }
265
266  llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
267
268  // Lookup cudaLaunchKernel/hipLaunchKernel function.
269  // cudaError_t cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
270  //                              void **args, size_t sharedMem,
271  //                              cudaStream_t stream);
272  // hipError_t hipLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
273  //                            void **args, size_t sharedMem,
274  //                            hipStream_t stream);
275  TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
276  DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
277  auto LaunchKernelName = addPrefixToName("LaunchKernel");
278  IdentifierInfo &cudaLaunchKernelII =
279      CGM.getContext().Idents.get(LaunchKernelName);
280  FunctionDecl *cudaLaunchKernelFD = nullptr;
281  for (const auto &Result : DC->lookup(&cudaLaunchKernelII)) {
282    if (FunctionDecl *FD = dyn_cast<FunctionDecl>(Result))
283      cudaLaunchKernelFD = FD;
284  }
285
286  if (cudaLaunchKernelFD == nullptr) {
287    CGM.Error(CGF.CurFuncDecl->getLocation(),
288              "Can't find declaration for " + LaunchKernelName);
289    return;
290  }
291  // Create temporary dim3 grid_dim, block_dim.
292  ParmVarDecl *GridDimParam = cudaLaunchKernelFD->getParamDecl(1);
293  QualType Dim3Ty = GridDimParam->getType();
294  Address GridDim =
295      CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "grid_dim");
296  Address BlockDim =
297      CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "block_dim");
298  Address ShmemSize =
299      CGF.CreateTempAlloca(SizeTy, CGM.getSizeAlign(), "shmem_size");
300  Address Stream =
301      CGF.CreateTempAlloca(VoidPtrTy, CGM.getPointerAlign(), "stream");
302  llvm::FunctionCallee cudaPopConfigFn = CGM.CreateRuntimeFunction(
303      llvm::FunctionType::get(IntTy,
304                              {/*gridDim=*/GridDim.getType(),
305                               /*blockDim=*/BlockDim.getType(),
306                               /*ShmemSize=*/ShmemSize.getType(),
307                               /*Stream=*/Stream.getType()},
308                              /*isVarArg=*/false),
309      addUnderscoredPrefixToName("PopCallConfiguration"));
310
311  CGF.EmitRuntimeCallOrInvoke(cudaPopConfigFn,
312                              {GridDim.getPointer(), BlockDim.getPointer(),
313                               ShmemSize.getPointer(), Stream.getPointer()});
314
315  // Emit the call to cudaLaunch
316  llvm::Value *Kernel = CGF.Builder.CreatePointerCast(CGF.CurFn, VoidPtrTy);
317  CallArgList LaunchKernelArgs;
318  LaunchKernelArgs.add(RValue::get(Kernel),
319                       cudaLaunchKernelFD->getParamDecl(0)->getType());
320  LaunchKernelArgs.add(RValue::getAggregate(GridDim), Dim3Ty);
321  LaunchKernelArgs.add(RValue::getAggregate(BlockDim), Dim3Ty);
322  LaunchKernelArgs.add(RValue::get(KernelArgs.getPointer()),
323                       cudaLaunchKernelFD->getParamDecl(3)->getType());
324  LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(ShmemSize)),
325                       cudaLaunchKernelFD->getParamDecl(4)->getType());
326  LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(Stream)),
327                       cudaLaunchKernelFD->getParamDecl(5)->getType());
328
329  QualType QT = cudaLaunchKernelFD->getType();
330  QualType CQT = QT.getCanonicalType();
331  llvm::Type *Ty = CGM.getTypes().ConvertType(CQT);
332  llvm::FunctionType *FTy = dyn_cast<llvm::FunctionType>(Ty);
333
334  const CGFunctionInfo &FI =
335      CGM.getTypes().arrangeFunctionDeclaration(cudaLaunchKernelFD);
336  llvm::FunctionCallee cudaLaunchKernelFn =
337      CGM.CreateRuntimeFunction(FTy, LaunchKernelName);
338  CGF.EmitCall(FI, CGCallee::forDirect(cudaLaunchKernelFn), ReturnValueSlot(),
339               LaunchKernelArgs);
340  CGF.EmitBranch(EndBlock);
341
342  CGF.EmitBlock(EndBlock);
343}
344
345void CGNVCUDARuntime::emitDeviceStubBodyLegacy(CodeGenFunction &CGF,
346                                               FunctionArgList &Args) {
347  // Emit a call to cudaSetupArgument for each arg in Args.
348  llvm::FunctionCallee cudaSetupArgFn = getSetupArgumentFn();
349  llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
350  CharUnits Offset = CharUnits::Zero();
351  for (const VarDecl *A : Args) {
352    CharUnits TyWidth, TyAlign;
353    std::tie(TyWidth, TyAlign) =
354        CGM.getContext().getTypeInfoInChars(A->getType());
355    Offset = Offset.alignTo(TyAlign);
356    llvm::Value *Args[] = {
357        CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
358                                      VoidPtrTy),
359        llvm::ConstantInt::get(SizeTy, TyWidth.getQuantity()),
360        llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
361    };
362    llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
363    llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
364    llvm::Value *CBZero = CGF.Builder.CreateICmpEQ(CB, Zero);
365    llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
366    CGF.Builder.CreateCondBr(CBZero, NextBlock, EndBlock);
367    CGF.EmitBlock(NextBlock);
368    Offset += TyWidth;
369  }
370
371  // Emit the call to cudaLaunch
372  llvm::FunctionCallee cudaLaunchFn = getLaunchFn();
373  llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
374  CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
375  CGF.EmitBranch(EndBlock);
376
377  CGF.EmitBlock(EndBlock);
378}
379
380/// Creates a function that sets up state on the host side for CUDA objects that
381/// have a presence on both the host and device sides. Specifically, registers
382/// the host side of kernel functions and device global variables with the CUDA
383/// runtime.
384/// \code
385/// void __cuda_register_globals(void** GpuBinaryHandle) {
386///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
387///    ...
388///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
389///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
390///    ...
391///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
392/// }
393/// \endcode
394llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
395  // No need to register anything
396  if (EmittedKernels.empty() && DeviceVars.empty())
397    return nullptr;
398
399  llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
400      getRegisterGlobalsFnTy(), llvm::GlobalValue::InternalLinkage,
401      addUnderscoredPrefixToName("_register_globals"), &TheModule);
402  llvm::BasicBlock *EntryBB =
403      llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
404  CGBuilderTy Builder(CGM, Context);
405  Builder.SetInsertPoint(EntryBB);
406
407  // void __cudaRegisterFunction(void **, const char *, char *, const char *,
408  //                             int, uint3*, uint3*, dim3*, dim3*, int*)
409  llvm::Type *RegisterFuncParams[] = {
410      VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
411      VoidPtrTy,    VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo()};
412  llvm::FunctionCallee RegisterFunc = CGM.CreateRuntimeFunction(
413      llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
414      addUnderscoredPrefixToName("RegisterFunction"));
415
416  // Extract GpuBinaryHandle passed as the first argument passed to
417  // __cuda_register_globals() and generate __cudaRegisterFunction() call for
418  // each emitted kernel.
419  llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
420  for (auto &&I : EmittedKernels) {
421    llvm::Constant *KernelName = makeConstantString(getDeviceSideName(I.D));
422    llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
423    llvm::Value *Args[] = {
424        &GpuBinaryHandlePtr,
425        Builder.CreateBitCast(I.Kernel, VoidPtrTy),
426        KernelName,
427        KernelName,
428        llvm::ConstantInt::get(IntTy, -1),
429        NullPtr,
430        NullPtr,
431        NullPtr,
432        NullPtr,
433        llvm::ConstantPointerNull::get(IntTy->getPointerTo())};
434    Builder.CreateCall(RegisterFunc, Args);
435  }
436
437  // void __cudaRegisterVar(void **, char *, char *, const char *,
438  //                        int, int, int, int)
439  llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
440                                     CharPtrTy,    IntTy,     IntTy,
441                                     IntTy,        IntTy};
442  llvm::FunctionCallee RegisterVar = CGM.CreateRuntimeFunction(
443      llvm::FunctionType::get(IntTy, RegisterVarParams, false),
444      addUnderscoredPrefixToName("RegisterVar"));
445  for (auto &&Info : DeviceVars) {
446    llvm::GlobalVariable *Var = Info.Var;
447    unsigned Flags = Info.Flag;
448    llvm::Constant *VarName = makeConstantString(getDeviceSideName(Info.D));
449    uint64_t VarSize =
450        CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
451    llvm::Value *Args[] = {
452        &GpuBinaryHandlePtr,
453        Builder.CreateBitCast(Var, VoidPtrTy),
454        VarName,
455        VarName,
456        llvm::ConstantInt::get(IntTy, (Flags & ExternDeviceVar) ? 1 : 0),
457        llvm::ConstantInt::get(IntTy, VarSize),
458        llvm::ConstantInt::get(IntTy, (Flags & ConstantDeviceVar) ? 1 : 0),
459        llvm::ConstantInt::get(IntTy, 0)};
460    Builder.CreateCall(RegisterVar, Args);
461  }
462
463  Builder.CreateRetVoid();
464  return RegisterKernelsFunc;
465}
466
467/// Creates a global constructor function for the module:
468///
469/// For CUDA:
470/// \code
471/// void __cuda_module_ctor(void*) {
472///     Handle = __cudaRegisterFatBinary(GpuBinaryBlob);
473///     __cuda_register_globals(Handle);
474/// }
475/// \endcode
476///
477/// For HIP:
478/// \code
479/// void __hip_module_ctor(void*) {
480///     if (__hip_gpubin_handle == 0) {
481///         __hip_gpubin_handle  = __hipRegisterFatBinary(GpuBinaryBlob);
482///         __hip_register_globals(__hip_gpubin_handle);
483///     }
484/// }
485/// \endcode
486llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
487  bool IsHIP = CGM.getLangOpts().HIP;
488  bool IsCUDA = CGM.getLangOpts().CUDA;
489  // No need to generate ctors/dtors if there is no GPU binary.
490  StringRef CudaGpuBinaryFileName = CGM.getCodeGenOpts().CudaGpuBinaryFileName;
491  if (CudaGpuBinaryFileName.empty() && !IsHIP)
492    return nullptr;
493  if ((IsHIP || (IsCUDA && !RelocatableDeviceCode)) && EmittedKernels.empty() &&
494      DeviceVars.empty())
495    return nullptr;
496
497  // void __{cuda|hip}_register_globals(void* handle);
498  llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
499  // We always need a function to pass in as callback. Create a dummy
500  // implementation if we don't need to register anything.
501  if (RelocatableDeviceCode && !RegisterGlobalsFunc)
502    RegisterGlobalsFunc = makeDummyFunction(getRegisterGlobalsFnTy());
503
504  // void ** __{cuda|hip}RegisterFatBinary(void *);
505  llvm::FunctionCallee RegisterFatbinFunc = CGM.CreateRuntimeFunction(
506      llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
507      addUnderscoredPrefixToName("RegisterFatBinary"));
508  // struct { int magic, int version, void * gpu_binary, void * dont_care };
509  llvm::StructType *FatbinWrapperTy =
510      llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy);
511
512  // Register GPU binary with the CUDA runtime, store returned handle in a
513  // global variable and save a reference in GpuBinaryHandle to be cleaned up
514  // in destructor on exit. Then associate all known kernels with the GPU binary
515  // handle so CUDA runtime can figure out what to call on the GPU side.
516  std::unique_ptr<llvm::MemoryBuffer> CudaGpuBinary = nullptr;
517  if (!CudaGpuBinaryFileName.empty()) {
518    llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CudaGpuBinaryOrErr =
519        llvm::MemoryBuffer::getFileOrSTDIN(CudaGpuBinaryFileName);
520    if (std::error_code EC = CudaGpuBinaryOrErr.getError()) {
521      CGM.getDiags().Report(diag::err_cannot_open_file)
522          << CudaGpuBinaryFileName << EC.message();
523      return nullptr;
524    }
525    CudaGpuBinary = std::move(CudaGpuBinaryOrErr.get());
526  }
527
528  llvm::Function *ModuleCtorFunc = llvm::Function::Create(
529      llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
530      llvm::GlobalValue::InternalLinkage,
531      addUnderscoredPrefixToName("_module_ctor"), &TheModule);
532  llvm::BasicBlock *CtorEntryBB =
533      llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
534  CGBuilderTy CtorBuilder(CGM, Context);
535
536  CtorBuilder.SetInsertPoint(CtorEntryBB);
537
538  const char *FatbinConstantName;
539  const char *FatbinSectionName;
540  const char *ModuleIDSectionName;
541  StringRef ModuleIDPrefix;
542  llvm::Constant *FatBinStr;
543  unsigned FatMagic;
544  if (IsHIP) {
545    FatbinConstantName = ".hip_fatbin";
546    FatbinSectionName = ".hipFatBinSegment";
547
548    ModuleIDSectionName = "__hip_module_id";
549    ModuleIDPrefix = "__hip_";
550
551    if (CudaGpuBinary) {
552      // If fatbin is available from early finalization, create a string
553      // literal containing the fat binary loaded from the given file.
554      FatBinStr = makeConstantString(CudaGpuBinary->getBuffer(), "",
555                                     FatbinConstantName, 8);
556    } else {
557      // If fatbin is not available, create an external symbol
558      // __hip_fatbin in section .hip_fatbin. The external symbol is supposed
559      // to contain the fat binary but will be populated somewhere else,
560      // e.g. by lld through link script.
561      FatBinStr = new llvm::GlobalVariable(
562        CGM.getModule(), CGM.Int8Ty,
563        /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
564        "__hip_fatbin", nullptr,
565        llvm::GlobalVariable::NotThreadLocal);
566      cast<llvm::GlobalVariable>(FatBinStr)->setSection(FatbinConstantName);
567    }
568
569    FatMagic = HIPFatMagic;
570  } else {
571    if (RelocatableDeviceCode)
572      FatbinConstantName = CGM.getTriple().isMacOSX()
573                               ? "__NV_CUDA,__nv_relfatbin"
574                               : "__nv_relfatbin";
575    else
576      FatbinConstantName =
577          CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
578    // NVIDIA's cuobjdump looks for fatbins in this section.
579    FatbinSectionName =
580        CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
581
582    ModuleIDSectionName = CGM.getTriple().isMacOSX()
583                              ? "__NV_CUDA,__nv_module_id"
584                              : "__nv_module_id";
585    ModuleIDPrefix = "__nv_";
586
587    // For CUDA, create a string literal containing the fat binary loaded from
588    // the given file.
589    FatBinStr = makeConstantString(CudaGpuBinary->getBuffer(), "",
590                                   FatbinConstantName, 8);
591    FatMagic = CudaFatMagic;
592  }
593
594  // Create initialized wrapper structure that points to the loaded GPU binary
595  ConstantInitBuilder Builder(CGM);
596  auto Values = Builder.beginStruct(FatbinWrapperTy);
597  // Fatbin wrapper magic.
598  Values.addInt(IntTy, FatMagic);
599  // Fatbin version.
600  Values.addInt(IntTy, 1);
601  // Data.
602  Values.add(FatBinStr);
603  // Unused in fatbin v1.
604  Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
605  llvm::GlobalVariable *FatbinWrapper = Values.finishAndCreateGlobal(
606      addUnderscoredPrefixToName("_fatbin_wrapper"), CGM.getPointerAlign(),
607      /*constant*/ true);
608  FatbinWrapper->setSection(FatbinSectionName);
609
610  // There is only one HIP fat binary per linked module, however there are
611  // multiple constructor functions. Make sure the fat binary is registered
612  // only once. The constructor functions are executed by the dynamic loader
613  // before the program gains control. The dynamic loader cannot execute the
614  // constructor functions concurrently since doing that would not guarantee
615  // thread safety of the loaded program. Therefore we can assume sequential
616  // execution of constructor functions here.
617  if (IsHIP) {
618    auto Linkage = CudaGpuBinary ? llvm::GlobalValue::InternalLinkage :
619        llvm::GlobalValue::LinkOnceAnyLinkage;
620    llvm::BasicBlock *IfBlock =
621        llvm::BasicBlock::Create(Context, "if", ModuleCtorFunc);
622    llvm::BasicBlock *ExitBlock =
623        llvm::BasicBlock::Create(Context, "exit", ModuleCtorFunc);
624    // The name, size, and initialization pattern of this variable is part
625    // of HIP ABI.
626    GpuBinaryHandle = new llvm::GlobalVariable(
627        TheModule, VoidPtrPtrTy, /*isConstant=*/false,
628        Linkage,
629        /*Initializer=*/llvm::ConstantPointerNull::get(VoidPtrPtrTy),
630        "__hip_gpubin_handle");
631    GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
632    // Prevent the weak symbol in different shared libraries being merged.
633    if (Linkage != llvm::GlobalValue::InternalLinkage)
634      GpuBinaryHandle->setVisibility(llvm::GlobalValue::HiddenVisibility);
635    Address GpuBinaryAddr(
636        GpuBinaryHandle,
637        CharUnits::fromQuantity(GpuBinaryHandle->getAlignment()));
638    {
639      auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
640      llvm::Constant *Zero =
641          llvm::Constant::getNullValue(HandleValue->getType());
642      llvm::Value *EQZero = CtorBuilder.CreateICmpEQ(HandleValue, Zero);
643      CtorBuilder.CreateCondBr(EQZero, IfBlock, ExitBlock);
644    }
645    {
646      CtorBuilder.SetInsertPoint(IfBlock);
647      // GpuBinaryHandle = __hipRegisterFatBinary(&FatbinWrapper);
648      llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
649          RegisterFatbinFunc,
650          CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
651      CtorBuilder.CreateStore(RegisterFatbinCall, GpuBinaryAddr);
652      CtorBuilder.CreateBr(ExitBlock);
653    }
654    {
655      CtorBuilder.SetInsertPoint(ExitBlock);
656      // Call __hip_register_globals(GpuBinaryHandle);
657      if (RegisterGlobalsFunc) {
658        auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
659        CtorBuilder.CreateCall(RegisterGlobalsFunc, HandleValue);
660      }
661    }
662  } else if (!RelocatableDeviceCode) {
663    // Register binary with CUDA runtime. This is substantially different in
664    // default mode vs. separate compilation!
665    // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
666    llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
667        RegisterFatbinFunc,
668        CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
669    GpuBinaryHandle = new llvm::GlobalVariable(
670        TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
671        llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
672    GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
673    CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
674                                   CGM.getPointerAlign());
675
676    // Call __cuda_register_globals(GpuBinaryHandle);
677    if (RegisterGlobalsFunc)
678      CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
679
680    // Call __cudaRegisterFatBinaryEnd(Handle) if this CUDA version needs it.
681    if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
682                           CudaFeature::CUDA_USES_FATBIN_REGISTER_END)) {
683      // void __cudaRegisterFatBinaryEnd(void **);
684      llvm::FunctionCallee RegisterFatbinEndFunc = CGM.CreateRuntimeFunction(
685          llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
686          "__cudaRegisterFatBinaryEnd");
687      CtorBuilder.CreateCall(RegisterFatbinEndFunc, RegisterFatbinCall);
688    }
689  } else {
690    // Generate a unique module ID.
691    SmallString<64> ModuleID;
692    llvm::raw_svector_ostream OS(ModuleID);
693    OS << ModuleIDPrefix << llvm::format("%" PRIx64, FatbinWrapper->getGUID());
694    llvm::Constant *ModuleIDConstant =
695        makeConstantString(ModuleID.str(), "", ModuleIDSectionName, 32);
696
697    // Create an alias for the FatbinWrapper that nvcc will look for.
698    llvm::GlobalAlias::create(llvm::GlobalValue::ExternalLinkage,
699                              Twine("__fatbinwrap") + ModuleID, FatbinWrapper);
700
701    // void __cudaRegisterLinkedBinary%ModuleID%(void (*)(void *), void *,
702    // void *, void (*)(void **))
703    SmallString<128> RegisterLinkedBinaryName("__cudaRegisterLinkedBinary");
704    RegisterLinkedBinaryName += ModuleID;
705    llvm::FunctionCallee RegisterLinkedBinaryFunc = CGM.CreateRuntimeFunction(
706        getRegisterLinkedBinaryFnTy(), RegisterLinkedBinaryName);
707
708    assert(RegisterGlobalsFunc && "Expecting at least dummy function!");
709    llvm::Value *Args[] = {RegisterGlobalsFunc,
710                           CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy),
711                           ModuleIDConstant,
712                           makeDummyFunction(getCallbackFnTy())};
713    CtorBuilder.CreateCall(RegisterLinkedBinaryFunc, Args);
714  }
715
716  // Create destructor and register it with atexit() the way NVCC does it. Doing
717  // it during regular destructor phase worked in CUDA before 9.2 but results in
718  // double-free in 9.2.
719  if (llvm::Function *CleanupFn = makeModuleDtorFunction()) {
720    // extern "C" int atexit(void (*f)(void));
721    llvm::FunctionType *AtExitTy =
722        llvm::FunctionType::get(IntTy, CleanupFn->getType(), false);
723    llvm::FunctionCallee AtExitFunc =
724        CGM.CreateRuntimeFunction(AtExitTy, "atexit", llvm::AttributeList(),
725                                  /*Local=*/true);
726    CtorBuilder.CreateCall(AtExitFunc, CleanupFn);
727  }
728
729  CtorBuilder.CreateRetVoid();
730  return ModuleCtorFunc;
731}
732
733/// Creates a global destructor function that unregisters the GPU code blob
734/// registered by constructor.
735///
736/// For CUDA:
737/// \code
738/// void __cuda_module_dtor(void*) {
739///     __cudaUnregisterFatBinary(Handle);
740/// }
741/// \endcode
742///
743/// For HIP:
744/// \code
745/// void __hip_module_dtor(void*) {
746///     if (__hip_gpubin_handle) {
747///         __hipUnregisterFatBinary(__hip_gpubin_handle);
748///         __hip_gpubin_handle = 0;
749///     }
750/// }
751/// \endcode
752llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
753  // No need for destructor if we don't have a handle to unregister.
754  if (!GpuBinaryHandle)
755    return nullptr;
756
757  // void __cudaUnregisterFatBinary(void ** handle);
758  llvm::FunctionCallee UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
759      llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
760      addUnderscoredPrefixToName("UnregisterFatBinary"));
761
762  llvm::Function *ModuleDtorFunc = llvm::Function::Create(
763      llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
764      llvm::GlobalValue::InternalLinkage,
765      addUnderscoredPrefixToName("_module_dtor"), &TheModule);
766
767  llvm::BasicBlock *DtorEntryBB =
768      llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
769  CGBuilderTy DtorBuilder(CGM, Context);
770  DtorBuilder.SetInsertPoint(DtorEntryBB);
771
772  Address GpuBinaryAddr(GpuBinaryHandle, CharUnits::fromQuantity(
773                                             GpuBinaryHandle->getAlignment()));
774  auto HandleValue = DtorBuilder.CreateLoad(GpuBinaryAddr);
775  // There is only one HIP fat binary per linked module, however there are
776  // multiple destructor functions. Make sure the fat binary is unregistered
777  // only once.
778  if (CGM.getLangOpts().HIP) {
779    llvm::BasicBlock *IfBlock =
780        llvm::BasicBlock::Create(Context, "if", ModuleDtorFunc);
781    llvm::BasicBlock *ExitBlock =
782        llvm::BasicBlock::Create(Context, "exit", ModuleDtorFunc);
783    llvm::Constant *Zero = llvm::Constant::getNullValue(HandleValue->getType());
784    llvm::Value *NEZero = DtorBuilder.CreateICmpNE(HandleValue, Zero);
785    DtorBuilder.CreateCondBr(NEZero, IfBlock, ExitBlock);
786
787    DtorBuilder.SetInsertPoint(IfBlock);
788    DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
789    DtorBuilder.CreateStore(Zero, GpuBinaryAddr);
790    DtorBuilder.CreateBr(ExitBlock);
791
792    DtorBuilder.SetInsertPoint(ExitBlock);
793  } else {
794    DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
795  }
796  DtorBuilder.CreateRetVoid();
797  return ModuleDtorFunc;
798}
799
800std::string CGNVCUDARuntime::getDeviceStubName(llvm::StringRef Name) const {
801  if (!CGM.getLangOpts().HIP)
802    return Name;
803  return (Name + ".stub").str();
804}
805
806CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
807  return new CGNVCUDARuntime(CGM);
808}
809