1 //===----- CGCUDANV.cpp - Interface to NVIDIA CUDA Runtime ----------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This provides a class for CUDA code generation targeting the NVIDIA CUDA
10 // runtime library.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "CGCUDARuntime.h"
15 #include "CodeGenFunction.h"
16 #include "CodeGenModule.h"
17 #include "clang/AST/Decl.h"
18 #include "clang/Basic/Cuda.h"
19 #include "clang/CodeGen/CodeGenABITypes.h"
20 #include "clang/CodeGen/ConstantInitBuilder.h"
21 #include "llvm/IR/BasicBlock.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/DerivedTypes.h"
24 #include "llvm/Support/Format.h"
25
26 using namespace clang;
27 using namespace CodeGen;
28
29 namespace {
30 constexpr unsigned CudaFatMagic = 0x466243b1;
31 constexpr unsigned HIPFatMagic = 0x48495046; // "HIPF"
32
33 class CGNVCUDARuntime : public CGCUDARuntime {
34
35 private:
36 llvm::IntegerType *IntTy, *SizeTy;
37 llvm::Type *VoidTy;
38 llvm::PointerType *CharPtrTy, *VoidPtrTy, *VoidPtrPtrTy;
39
40 /// Convenience reference to LLVM Context
41 llvm::LLVMContext &Context;
42 /// Convenience reference to the current module
43 llvm::Module &TheModule;
44 /// Keeps track of kernel launch stubs emitted in this module
45 struct KernelInfo {
46 llvm::Function *Kernel;
47 const Decl *D;
48 };
49 llvm::SmallVector<KernelInfo, 16> EmittedKernels;
50 struct VarInfo {
51 llvm::GlobalVariable *Var;
52 const VarDecl *D;
53 DeviceVarFlags Flags;
54 };
55 llvm::SmallVector<VarInfo, 16> DeviceVars;
56 /// Keeps track of variable containing handle of GPU binary. Populated by
57 /// ModuleCtorFunction() and used to create corresponding cleanup calls in
58 /// ModuleDtorFunction()
59 llvm::GlobalVariable *GpuBinaryHandle = nullptr;
60 /// Whether we generate relocatable device code.
61 bool RelocatableDeviceCode;
62 /// Mangle context for device.
63 std::unique_ptr<MangleContext> DeviceMC;
64
65 llvm::FunctionCallee getSetupArgumentFn() const;
66 llvm::FunctionCallee getLaunchFn() const;
67
68 llvm::FunctionType *getRegisterGlobalsFnTy() const;
69 llvm::FunctionType *getCallbackFnTy() const;
70 llvm::FunctionType *getRegisterLinkedBinaryFnTy() const;
71 std::string addPrefixToName(StringRef FuncName) const;
72 std::string addUnderscoredPrefixToName(StringRef FuncName) const;
73
74 /// Creates a function to register all kernel stubs generated in this module.
75 llvm::Function *makeRegisterGlobalsFn();
76
77 /// Helper function that generates a constant string and returns a pointer to
78 /// the start of the string. The result of this function can be used anywhere
79 /// where the C code specifies const char*.
makeConstantString(const std::string & Str,const std::string & Name="",const std::string & SectionName="",unsigned Alignment=0)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.
makeDummyFunction(llvm::FunctionType * FnTy)103 llvm::Function *makeDummyFunction(llvm::FunctionType *FnTy) {
104 assert(FnTy->getReturnType()->isVoidTy() &&
105 "Can only generate dummy functions returning void!");
106 llvm::Function *DummyFunc = llvm::Function::Create(
107 FnTy, llvm::GlobalValue::InternalLinkage, "dummy", &TheModule);
108
109 llvm::BasicBlock *DummyBlock =
110 llvm::BasicBlock::Create(Context, "", DummyFunc);
111 CGBuilderTy FuncBuilder(CGM, Context);
112 FuncBuilder.SetInsertPoint(DummyBlock);
113 FuncBuilder.CreateRetVoid();
114
115 return DummyFunc;
116 }
117
118 void emitDeviceStubBodyLegacy(CodeGenFunction &CGF, FunctionArgList &Args);
119 void emitDeviceStubBodyNew(CodeGenFunction &CGF, FunctionArgList &Args);
120 std::string getDeviceSideName(const NamedDecl *ND) override;
121
122 public:
123 CGNVCUDARuntime(CodeGenModule &CGM);
124
125 void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
registerDeviceVar(const VarDecl * VD,llvm::GlobalVariable & Var,bool Extern,bool Constant)126 void registerDeviceVar(const VarDecl *VD, llvm::GlobalVariable &Var,
127 bool Extern, bool Constant) override {
128 DeviceVars.push_back({&Var,
129 VD,
130 {DeviceVarFlags::Variable, Extern, Constant,
131 /*Normalized*/ false, /*Type*/ 0}});
132 }
registerDeviceSurf(const VarDecl * VD,llvm::GlobalVariable & Var,bool Extern,int Type)133 void registerDeviceSurf(const VarDecl *VD, llvm::GlobalVariable &Var,
134 bool Extern, int Type) override {
135 DeviceVars.push_back({&Var,
136 VD,
137 {DeviceVarFlags::Surface, Extern, /*Constant*/ false,
138 /*Normalized*/ false, Type}});
139 }
registerDeviceTex(const VarDecl * VD,llvm::GlobalVariable & Var,bool Extern,int Type,bool Normalized)140 void registerDeviceTex(const VarDecl *VD, llvm::GlobalVariable &Var,
141 bool Extern, int Type, bool Normalized) override {
142 DeviceVars.push_back({&Var,
143 VD,
144 {DeviceVarFlags::Texture, Extern, /*Constant*/ false,
145 Normalized, Type}});
146 }
147
148 /// Creates module constructor function
149 llvm::Function *makeModuleCtorFunction() override;
150 /// Creates module destructor function
151 llvm::Function *makeModuleDtorFunction() override;
152 };
153
154 }
155
addPrefixToName(StringRef FuncName) const156 std::string CGNVCUDARuntime::addPrefixToName(StringRef FuncName) const {
157 if (CGM.getLangOpts().HIP)
158 return ((Twine("hip") + Twine(FuncName)).str());
159 return ((Twine("cuda") + Twine(FuncName)).str());
160 }
161 std::string
addUnderscoredPrefixToName(StringRef FuncName) const162 CGNVCUDARuntime::addUnderscoredPrefixToName(StringRef FuncName) const {
163 if (CGM.getLangOpts().HIP)
164 return ((Twine("__hip") + Twine(FuncName)).str());
165 return ((Twine("__cuda") + Twine(FuncName)).str());
166 }
167
CGNVCUDARuntime(CodeGenModule & CGM)168 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
169 : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
170 TheModule(CGM.getModule()),
171 RelocatableDeviceCode(CGM.getLangOpts().GPURelocatableDeviceCode),
172 DeviceMC(CGM.getContext().createMangleContext(
173 CGM.getContext().getAuxTargetInfo())) {
174 CodeGen::CodeGenTypes &Types = CGM.getTypes();
175 ASTContext &Ctx = CGM.getContext();
176
177 IntTy = CGM.IntTy;
178 SizeTy = CGM.SizeTy;
179 VoidTy = CGM.VoidTy;
180
181 unsigned DefaultAS = CGM.getTargetCodeGenInfo().getDefaultAS();
182 CharPtrTy = llvm::PointerType::get(Types.ConvertType(Ctx.CharTy), DefaultAS);
183 VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
184 VoidPtrPtrTy = VoidPtrTy->getPointerTo(DefaultAS);
185 }
186
getSetupArgumentFn() const187 llvm::FunctionCallee CGNVCUDARuntime::getSetupArgumentFn() const {
188 // cudaError_t cudaSetupArgument(void *, size_t, size_t)
189 llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
190 return CGM.CreateRuntimeFunction(
191 llvm::FunctionType::get(IntTy, Params, false),
192 addPrefixToName("SetupArgument"));
193 }
194
getLaunchFn() const195 llvm::FunctionCallee CGNVCUDARuntime::getLaunchFn() const {
196 if (CGM.getLangOpts().HIP) {
197 // hipError_t hipLaunchByPtr(char *);
198 return CGM.CreateRuntimeFunction(
199 llvm::FunctionType::get(IntTy, CharPtrTy, false), "hipLaunchByPtr");
200 } else {
201 // cudaError_t cudaLaunch(char *);
202 return CGM.CreateRuntimeFunction(
203 llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
204 }
205 }
206
getRegisterGlobalsFnTy() const207 llvm::FunctionType *CGNVCUDARuntime::getRegisterGlobalsFnTy() const {
208 return llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false);
209 }
210
getCallbackFnTy() const211 llvm::FunctionType *CGNVCUDARuntime::getCallbackFnTy() const {
212 return llvm::FunctionType::get(VoidTy, VoidPtrTy, false);
213 }
214
getRegisterLinkedBinaryFnTy() const215 llvm::FunctionType *CGNVCUDARuntime::getRegisterLinkedBinaryFnTy() const {
216 auto CallbackFnTy = getCallbackFnTy();
217 auto RegisterGlobalsFnTy = getRegisterGlobalsFnTy();
218 llvm::Type *Params[] = {RegisterGlobalsFnTy->getPointerTo(), VoidPtrTy,
219 VoidPtrTy, CallbackFnTy->getPointerTo()};
220 return llvm::FunctionType::get(VoidTy, Params, false);
221 }
222
getDeviceSideName(const NamedDecl * ND)223 std::string CGNVCUDARuntime::getDeviceSideName(const NamedDecl *ND) {
224 GlobalDecl GD;
225 // D could be either a kernel or a variable.
226 if (auto *FD = dyn_cast<FunctionDecl>(ND))
227 GD = GlobalDecl(FD, KernelReferenceKind::Kernel);
228 else
229 GD = GlobalDecl(ND);
230 std::string DeviceSideName;
231 if (DeviceMC->shouldMangleDeclName(ND)) {
232 SmallString<256> Buffer;
233 llvm::raw_svector_ostream Out(Buffer);
234 DeviceMC->mangleName(GD, Out);
235 DeviceSideName = std::string(Out.str());
236 } else
237 DeviceSideName = std::string(ND->getIdentifier()->getName());
238 return DeviceSideName;
239 }
240
emitDeviceStub(CodeGenFunction & CGF,FunctionArgList & Args)241 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
242 FunctionArgList &Args) {
243 EmittedKernels.push_back({CGF.CurFn, CGF.CurFuncDecl});
244 if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
245 CudaFeature::CUDA_USES_NEW_LAUNCH) ||
246 (CGF.getLangOpts().HIP && CGF.getLangOpts().HIPUseNewLaunchAPI))
247 emitDeviceStubBodyNew(CGF, Args);
248 else
249 emitDeviceStubBodyLegacy(CGF, Args);
250 }
251
252 // CUDA 9.0+ uses new way to launch kernels. Parameters are packed in a local
253 // array and kernels are launched using cudaLaunchKernel().
emitDeviceStubBodyNew(CodeGenFunction & CGF,FunctionArgList & Args)254 void CGNVCUDARuntime::emitDeviceStubBodyNew(CodeGenFunction &CGF,
255 FunctionArgList &Args) {
256 // Build the shadow stack entry at the very start of the function.
257
258 // Calculate amount of space we will need for all arguments. If we have no
259 // args, allocate a single pointer so we still have a valid pointer to the
260 // argument array that we can pass to runtime, even if it will be unused.
261 Address KernelArgs = CGF.CreateTempAlloca(
262 VoidPtrTy, CharUnits::fromQuantity(16), "kernel_args",
263 llvm::ConstantInt::get(SizeTy, std::max<size_t>(1, Args.size())));
264 // Store pointers to the arguments in a locally allocated launch_args.
265 for (unsigned i = 0; i < Args.size(); ++i) {
266 llvm::Value* VarPtr = CGF.GetAddrOfLocalVar(Args[i]).getPointer();
267 llvm::Value *VoidVarPtr = CGF.Builder.CreatePointerCast(VarPtr, VoidPtrTy);
268 CGF.Builder.CreateDefaultAlignedStore(
269 VoidVarPtr, CGF.Builder.CreateConstGEP1_32(KernelArgs.getPointer(), i));
270 }
271
272 llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
273
274 // Lookup cudaLaunchKernel/hipLaunchKernel function.
275 // cudaError_t cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
276 // void **args, size_t sharedMem,
277 // cudaStream_t stream);
278 // hipError_t hipLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
279 // void **args, size_t sharedMem,
280 // hipStream_t stream);
281 TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
282 DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
283 auto LaunchKernelName = addPrefixToName("LaunchKernel");
284 IdentifierInfo &cudaLaunchKernelII =
285 CGM.getContext().Idents.get(LaunchKernelName);
286 FunctionDecl *cudaLaunchKernelFD = nullptr;
287 for (const auto &Result : DC->lookup(&cudaLaunchKernelII)) {
288 if (FunctionDecl *FD = dyn_cast<FunctionDecl>(Result))
289 cudaLaunchKernelFD = FD;
290 }
291
292 if (cudaLaunchKernelFD == nullptr) {
293 CGM.Error(CGF.CurFuncDecl->getLocation(),
294 "Can't find declaration for " + LaunchKernelName);
295 return;
296 }
297 // Create temporary dim3 grid_dim, block_dim.
298 ParmVarDecl *GridDimParam = cudaLaunchKernelFD->getParamDecl(1);
299 QualType Dim3Ty = GridDimParam->getType();
300 Address GridDim =
301 CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "grid_dim");
302 Address BlockDim =
303 CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "block_dim");
304 Address ShmemSize =
305 CGF.CreateTempAlloca(SizeTy, CGM.getSizeAlign(), "shmem_size");
306 Address Stream =
307 CGF.CreateTempAlloca(VoidPtrTy, CGM.getPointerAlign(), "stream");
308 llvm::FunctionCallee cudaPopConfigFn = CGM.CreateRuntimeFunction(
309 llvm::FunctionType::get(IntTy,
310 {/*gridDim=*/GridDim.getType(),
311 /*blockDim=*/BlockDim.getType(),
312 /*ShmemSize=*/ShmemSize.getType(),
313 /*Stream=*/Stream.getType()},
314 /*isVarArg=*/false),
315 addUnderscoredPrefixToName("PopCallConfiguration"));
316
317 CGF.EmitRuntimeCallOrInvoke(cudaPopConfigFn,
318 {GridDim.getPointer(), BlockDim.getPointer(),
319 ShmemSize.getPointer(), Stream.getPointer()});
320
321 // Emit the call to cudaLaunch
322 llvm::Value *Kernel = CGF.Builder.CreatePointerCast(CGF.CurFn, VoidPtrTy);
323 CallArgList LaunchKernelArgs;
324 LaunchKernelArgs.add(RValue::get(Kernel),
325 cudaLaunchKernelFD->getParamDecl(0)->getType());
326 LaunchKernelArgs.add(RValue::getAggregate(GridDim), Dim3Ty);
327 LaunchKernelArgs.add(RValue::getAggregate(BlockDim), Dim3Ty);
328 LaunchKernelArgs.add(RValue::get(KernelArgs.getPointer()),
329 cudaLaunchKernelFD->getParamDecl(3)->getType());
330 LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(ShmemSize)),
331 cudaLaunchKernelFD->getParamDecl(4)->getType());
332 LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(Stream)),
333 cudaLaunchKernelFD->getParamDecl(5)->getType());
334
335 QualType QT = cudaLaunchKernelFD->getType();
336 QualType CQT = QT.getCanonicalType();
337 llvm::Type *Ty = CGM.getTypes().ConvertType(CQT);
338 llvm::FunctionType *FTy = dyn_cast<llvm::FunctionType>(Ty);
339
340 const CGFunctionInfo &FI =
341 CGM.getTypes().arrangeFunctionDeclaration(cudaLaunchKernelFD);
342 llvm::FunctionCallee cudaLaunchKernelFn =
343 CGM.CreateRuntimeFunction(FTy, LaunchKernelName);
344 CGF.EmitCall(FI, CGCallee::forDirect(cudaLaunchKernelFn), ReturnValueSlot(),
345 LaunchKernelArgs);
346 CGF.EmitBranch(EndBlock);
347
348 CGF.EmitBlock(EndBlock);
349 }
350
emitDeviceStubBodyLegacy(CodeGenFunction & CGF,FunctionArgList & Args)351 void CGNVCUDARuntime::emitDeviceStubBodyLegacy(CodeGenFunction &CGF,
352 FunctionArgList &Args) {
353 // Emit a call to cudaSetupArgument for each arg in Args.
354 llvm::FunctionCallee cudaSetupArgFn = getSetupArgumentFn();
355 llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
356 CharUnits Offset = CharUnits::Zero();
357 for (const VarDecl *A : Args) {
358 CharUnits TyWidth, TyAlign;
359 std::tie(TyWidth, TyAlign) =
360 CGM.getContext().getTypeInfoInChars(A->getType());
361 Offset = Offset.alignTo(TyAlign);
362 llvm::Value *Args[] = {
363 CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
364 VoidPtrTy),
365 llvm::ConstantInt::get(SizeTy, TyWidth.getQuantity()),
366 llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
367 };
368 llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
369 llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
370 llvm::Value *CBZero = CGF.Builder.CreateICmpEQ(CB, Zero);
371 llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
372 CGF.Builder.CreateCondBr(CBZero, NextBlock, EndBlock);
373 CGF.EmitBlock(NextBlock);
374 Offset += TyWidth;
375 }
376
377 // Emit the call to cudaLaunch
378 llvm::FunctionCallee cudaLaunchFn = getLaunchFn();
379 llvm::Value *Arg = CGF.Builder.CreatePointerCast(CGF.CurFn, CharPtrTy);
380 CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
381 CGF.EmitBranch(EndBlock);
382
383 CGF.EmitBlock(EndBlock);
384 }
385
386 /// Creates a function that sets up state on the host side for CUDA objects that
387 /// have a presence on both the host and device sides. Specifically, registers
388 /// the host side of kernel functions and device global variables with the CUDA
389 /// runtime.
390 /// \code
391 /// void __cuda_register_globals(void** GpuBinaryHandle) {
392 /// __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
393 /// ...
394 /// __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
395 /// __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
396 /// ...
397 /// __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
398 /// }
399 /// \endcode
makeRegisterGlobalsFn()400 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
401 // No need to register anything
402 if (EmittedKernels.empty() && DeviceVars.empty())
403 return nullptr;
404
405 llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
406 getRegisterGlobalsFnTy(), llvm::GlobalValue::InternalLinkage,
407 addUnderscoredPrefixToName("_register_globals"), &TheModule);
408 llvm::BasicBlock *EntryBB =
409 llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
410 CGBuilderTy Builder(CGM, Context);
411 Builder.SetInsertPoint(EntryBB);
412
413 unsigned DefaultAS = CGM.getTargetCodeGenInfo().getDefaultAS();
414 // void __cudaRegisterFunction(void **, const char *, char *, const char *,
415 // int, uint3*, uint3*, dim3*, dim3*, int*)
416 llvm::Type *RegisterFuncParams[] = {
417 VoidPtrPtrTy, CharPtrTy, CharPtrTy, CharPtrTy, IntTy,
418 VoidPtrTy, VoidPtrTy, VoidPtrTy, VoidPtrTy, IntTy->getPointerTo(DefaultAS)};
419 llvm::FunctionCallee RegisterFunc = CGM.CreateRuntimeFunction(
420 llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
421 addUnderscoredPrefixToName("RegisterFunction"));
422
423 // Extract GpuBinaryHandle passed as the first argument passed to
424 // __cuda_register_globals() and generate __cudaRegisterFunction() call for
425 // each emitted kernel.
426 llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
427 for (auto &&I : EmittedKernels) {
428 llvm::Constant *KernelName =
429 makeConstantString(getDeviceSideName(cast<NamedDecl>(I.D)));
430 llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
431 llvm::Value *Args[] = {
432 &GpuBinaryHandlePtr,
433 Builder.CreateBitCast(I.Kernel, VoidPtrTy),
434 KernelName,
435 KernelName,
436 llvm::ConstantInt::get(IntTy, -1),
437 NullPtr,
438 NullPtr,
439 NullPtr,
440 NullPtr,
441 llvm::ConstantPointerNull::get(IntTy->getPointerTo(DefaultAS))};
442 Builder.CreateCall(RegisterFunc, Args);
443 }
444
445 llvm::Type *VarSizeTy = IntTy;
446 // For HIP or CUDA 9.0+, device variable size is type of `size_t`.
447 if (CGM.getLangOpts().HIP ||
448 ToCudaVersion(CGM.getTarget().getSDKVersion()) >= CudaVersion::CUDA_90)
449 VarSizeTy = SizeTy;
450
451 // void __cudaRegisterVar(void **, char *, char *, const char *,
452 // int, int, int, int)
453 llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
454 CharPtrTy, IntTy, VarSizeTy,
455 IntTy, IntTy};
456 llvm::FunctionCallee RegisterVar = CGM.CreateRuntimeFunction(
457 llvm::FunctionType::get(VoidTy, RegisterVarParams, false),
458 addUnderscoredPrefixToName("RegisterVar"));
459 // void __cudaRegisterSurface(void **, const struct surfaceReference *,
460 // const void **, const char *, int, int);
461 llvm::FunctionCallee RegisterSurf = CGM.CreateRuntimeFunction(
462 llvm::FunctionType::get(
463 VoidTy, {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy},
464 false),
465 addUnderscoredPrefixToName("RegisterSurface"));
466 // void __cudaRegisterTexture(void **, const struct textureReference *,
467 // const void **, const char *, int, int, int)
468 llvm::FunctionCallee RegisterTex = CGM.CreateRuntimeFunction(
469 llvm::FunctionType::get(
470 VoidTy,
471 {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy, IntTy},
472 false),
473 addUnderscoredPrefixToName("RegisterTexture"));
474 for (auto &&Info : DeviceVars) {
475 llvm::GlobalVariable *Var = Info.Var;
476 llvm::Constant *VarName = makeConstantString(getDeviceSideName(Info.D));
477 switch (Info.Flags.getKind()) {
478 case DeviceVarFlags::Variable: {
479 uint64_t VarSize =
480 CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
481 llvm::Value *Args[] = {
482 &GpuBinaryHandlePtr,
483 Builder.CreateBitCast(Var, VoidPtrTy),
484 VarName,
485 VarName,
486 llvm::ConstantInt::get(IntTy, Info.Flags.isExtern()),
487 llvm::ConstantInt::get(VarSizeTy, VarSize),
488 llvm::ConstantInt::get(IntTy, Info.Flags.isConstant()),
489 llvm::ConstantInt::get(IntTy, 0)};
490 Builder.CreateCall(RegisterVar, Args);
491 break;
492 }
493 case DeviceVarFlags::Surface:
494 Builder.CreateCall(
495 RegisterSurf,
496 {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
497 VarName, llvm::ConstantInt::get(IntTy, Info.Flags.getSurfTexType()),
498 llvm::ConstantInt::get(IntTy, Info.Flags.isExtern())});
499 break;
500 case DeviceVarFlags::Texture:
501 Builder.CreateCall(
502 RegisterTex,
503 {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
504 VarName, llvm::ConstantInt::get(IntTy, Info.Flags.getSurfTexType()),
505 llvm::ConstantInt::get(IntTy, Info.Flags.isNormalized()),
506 llvm::ConstantInt::get(IntTy, Info.Flags.isExtern())});
507 break;
508 }
509 }
510
511 Builder.CreateRetVoid();
512 return RegisterKernelsFunc;
513 }
514
515 /// Creates a global constructor function for the module:
516 ///
517 /// For CUDA:
518 /// \code
519 /// void __cuda_module_ctor(void*) {
520 /// Handle = __cudaRegisterFatBinary(GpuBinaryBlob);
521 /// __cuda_register_globals(Handle);
522 /// }
523 /// \endcode
524 ///
525 /// For HIP:
526 /// \code
527 /// void __hip_module_ctor(void*) {
528 /// if (__hip_gpubin_handle == 0) {
529 /// __hip_gpubin_handle = __hipRegisterFatBinary(GpuBinaryBlob);
530 /// __hip_register_globals(__hip_gpubin_handle);
531 /// }
532 /// }
533 /// \endcode
makeModuleCtorFunction()534 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
535 bool IsHIP = CGM.getLangOpts().HIP;
536 bool IsCUDA = CGM.getLangOpts().CUDA;
537 // No need to generate ctors/dtors if there is no GPU binary.
538 StringRef CudaGpuBinaryFileName = CGM.getCodeGenOpts().CudaGpuBinaryFileName;
539 if (CudaGpuBinaryFileName.empty() && !IsHIP)
540 return nullptr;
541 if ((IsHIP || (IsCUDA && !RelocatableDeviceCode)) && EmittedKernels.empty() &&
542 DeviceVars.empty())
543 return nullptr;
544
545 // void __{cuda|hip}_register_globals(void* handle);
546 llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
547 // We always need a function to pass in as callback. Create a dummy
548 // implementation if we don't need to register anything.
549 if (RelocatableDeviceCode && !RegisterGlobalsFunc)
550 RegisterGlobalsFunc = makeDummyFunction(getRegisterGlobalsFnTy());
551
552 // void ** __{cuda|hip}RegisterFatBinary(void *);
553 llvm::FunctionCallee RegisterFatbinFunc = CGM.CreateRuntimeFunction(
554 llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
555 addUnderscoredPrefixToName("RegisterFatBinary"));
556 // struct { int magic, int version, void * gpu_binary, void * dont_care };
557 llvm::StructType *FatbinWrapperTy =
558 llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy);
559
560 // Register GPU binary with the CUDA runtime, store returned handle in a
561 // global variable and save a reference in GpuBinaryHandle to be cleaned up
562 // in destructor on exit. Then associate all known kernels with the GPU binary
563 // handle so CUDA runtime can figure out what to call on the GPU side.
564 std::unique_ptr<llvm::MemoryBuffer> CudaGpuBinary = nullptr;
565 if (!CudaGpuBinaryFileName.empty()) {
566 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> CudaGpuBinaryOrErr =
567 llvm::MemoryBuffer::getFileOrSTDIN(CudaGpuBinaryFileName);
568 if (std::error_code EC = CudaGpuBinaryOrErr.getError()) {
569 CGM.getDiags().Report(diag::err_cannot_open_file)
570 << CudaGpuBinaryFileName << EC.message();
571 return nullptr;
572 }
573 CudaGpuBinary = std::move(CudaGpuBinaryOrErr.get());
574 }
575
576 llvm::Function *ModuleCtorFunc = llvm::Function::Create(
577 llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
578 llvm::GlobalValue::InternalLinkage,
579 addUnderscoredPrefixToName("_module_ctor"), &TheModule);
580 llvm::BasicBlock *CtorEntryBB =
581 llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
582 CGBuilderTy CtorBuilder(CGM, Context);
583
584 CtorBuilder.SetInsertPoint(CtorEntryBB);
585
586 const char *FatbinConstantName;
587 const char *FatbinSectionName;
588 const char *ModuleIDSectionName;
589 StringRef ModuleIDPrefix;
590 llvm::Constant *FatBinStr;
591 unsigned FatMagic;
592 if (IsHIP) {
593 FatbinConstantName = ".hip_fatbin";
594 FatbinSectionName = ".hipFatBinSegment";
595
596 ModuleIDSectionName = "__hip_module_id";
597 ModuleIDPrefix = "__hip_";
598
599 if (CudaGpuBinary) {
600 // If fatbin is available from early finalization, create a string
601 // literal containing the fat binary loaded from the given file.
602 FatBinStr = makeConstantString(std::string(CudaGpuBinary->getBuffer()),
603 "", FatbinConstantName, 8);
604 } else {
605 // If fatbin is not available, create an external symbol
606 // __hip_fatbin in section .hip_fatbin. The external symbol is supposed
607 // to contain the fat binary but will be populated somewhere else,
608 // e.g. by lld through link script.
609 FatBinStr = new llvm::GlobalVariable(
610 CGM.getModule(), CGM.Int8Ty,
611 /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
612 "__hip_fatbin", nullptr,
613 llvm::GlobalVariable::NotThreadLocal);
614 cast<llvm::GlobalVariable>(FatBinStr)->setSection(FatbinConstantName);
615 }
616
617 FatMagic = HIPFatMagic;
618 } else {
619 if (RelocatableDeviceCode)
620 FatbinConstantName = CGM.getTriple().isMacOSX()
621 ? "__NV_CUDA,__nv_relfatbin"
622 : "__nv_relfatbin";
623 else
624 FatbinConstantName =
625 CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
626 // NVIDIA's cuobjdump looks for fatbins in this section.
627 FatbinSectionName =
628 CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
629
630 ModuleIDSectionName = CGM.getTriple().isMacOSX()
631 ? "__NV_CUDA,__nv_module_id"
632 : "__nv_module_id";
633 ModuleIDPrefix = "__nv_";
634
635 // For CUDA, create a string literal containing the fat binary loaded from
636 // the given file.
637 FatBinStr = makeConstantString(std::string(CudaGpuBinary->getBuffer()), "",
638 FatbinConstantName, 8);
639 FatMagic = CudaFatMagic;
640 }
641
642 // Create initialized wrapper structure that points to the loaded GPU binary
643 ConstantInitBuilder Builder(CGM);
644 auto Values = Builder.beginStruct(FatbinWrapperTy);
645 // Fatbin wrapper magic.
646 Values.addInt(IntTy, FatMagic);
647 // Fatbin version.
648 Values.addInt(IntTy, 1);
649 // Data.
650 Values.add(FatBinStr);
651 // Unused in fatbin v1.
652 Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
653 llvm::GlobalVariable *FatbinWrapper = Values.finishAndCreateGlobal(
654 addUnderscoredPrefixToName("_fatbin_wrapper"), CGM.getPointerAlign(),
655 /*constant*/ true);
656 FatbinWrapper->setSection(FatbinSectionName);
657
658 // There is only one HIP fat binary per linked module, however there are
659 // multiple constructor functions. Make sure the fat binary is registered
660 // only once. The constructor functions are executed by the dynamic loader
661 // before the program gains control. The dynamic loader cannot execute the
662 // constructor functions concurrently since doing that would not guarantee
663 // thread safety of the loaded program. Therefore we can assume sequential
664 // execution of constructor functions here.
665 if (IsHIP) {
666 auto Linkage = CudaGpuBinary ? llvm::GlobalValue::InternalLinkage :
667 llvm::GlobalValue::LinkOnceAnyLinkage;
668 llvm::BasicBlock *IfBlock =
669 llvm::BasicBlock::Create(Context, "if", ModuleCtorFunc);
670 llvm::BasicBlock *ExitBlock =
671 llvm::BasicBlock::Create(Context, "exit", ModuleCtorFunc);
672 // The name, size, and initialization pattern of this variable is part
673 // of HIP ABI.
674 GpuBinaryHandle = new llvm::GlobalVariable(
675 TheModule, VoidPtrPtrTy, /*isConstant=*/false,
676 Linkage,
677 /*Initializer=*/llvm::ConstantPointerNull::get(VoidPtrPtrTy),
678 "__hip_gpubin_handle");
679 GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
680 // Prevent the weak symbol in different shared libraries being merged.
681 if (Linkage != llvm::GlobalValue::InternalLinkage)
682 GpuBinaryHandle->setVisibility(llvm::GlobalValue::HiddenVisibility);
683 Address GpuBinaryAddr(
684 GpuBinaryHandle,
685 CharUnits::fromQuantity(GpuBinaryHandle->getAlignment()));
686 {
687 auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
688 llvm::Constant *Zero =
689 llvm::Constant::getNullValue(HandleValue->getType());
690 llvm::Value *EQZero = CtorBuilder.CreateICmpEQ(HandleValue, Zero);
691 CtorBuilder.CreateCondBr(EQZero, IfBlock, ExitBlock);
692 }
693 {
694 CtorBuilder.SetInsertPoint(IfBlock);
695 // GpuBinaryHandle = __hipRegisterFatBinary(&FatbinWrapper);
696 llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
697 RegisterFatbinFunc,
698 CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
699 CtorBuilder.CreateStore(RegisterFatbinCall, GpuBinaryAddr);
700 CtorBuilder.CreateBr(ExitBlock);
701 }
702 {
703 CtorBuilder.SetInsertPoint(ExitBlock);
704 // Call __hip_register_globals(GpuBinaryHandle);
705 if (RegisterGlobalsFunc) {
706 auto HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
707 CtorBuilder.CreateCall(RegisterGlobalsFunc, HandleValue);
708 }
709 }
710 } else if (!RelocatableDeviceCode) {
711 // Register binary with CUDA runtime. This is substantially different in
712 // default mode vs. separate compilation!
713 // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
714 llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
715 RegisterFatbinFunc,
716 CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
717 GpuBinaryHandle = new llvm::GlobalVariable(
718 TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
719 llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
720 GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
721 CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
722 CGM.getPointerAlign());
723
724 // Call __cuda_register_globals(GpuBinaryHandle);
725 if (RegisterGlobalsFunc)
726 CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
727
728 // Call __cudaRegisterFatBinaryEnd(Handle) if this CUDA version needs it.
729 if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
730 CudaFeature::CUDA_USES_FATBIN_REGISTER_END)) {
731 // void __cudaRegisterFatBinaryEnd(void **);
732 llvm::FunctionCallee RegisterFatbinEndFunc = CGM.CreateRuntimeFunction(
733 llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
734 "__cudaRegisterFatBinaryEnd");
735 CtorBuilder.CreateCall(RegisterFatbinEndFunc, RegisterFatbinCall);
736 }
737 } else {
738 // Generate a unique module ID.
739 SmallString<64> ModuleID;
740 llvm::raw_svector_ostream OS(ModuleID);
741 OS << ModuleIDPrefix << llvm::format("%" PRIx64, FatbinWrapper->getGUID());
742 llvm::Constant *ModuleIDConstant = makeConstantString(
743 std::string(ModuleID.str()), "", ModuleIDSectionName, 32);
744
745 // Create an alias for the FatbinWrapper that nvcc will look for.
746 llvm::GlobalAlias::create(llvm::GlobalValue::ExternalLinkage,
747 Twine("__fatbinwrap") + ModuleID, FatbinWrapper);
748
749 // void __cudaRegisterLinkedBinary%ModuleID%(void (*)(void *), void *,
750 // void *, void (*)(void **))
751 SmallString<128> RegisterLinkedBinaryName("__cudaRegisterLinkedBinary");
752 RegisterLinkedBinaryName += ModuleID;
753 llvm::FunctionCallee RegisterLinkedBinaryFunc = CGM.CreateRuntimeFunction(
754 getRegisterLinkedBinaryFnTy(), RegisterLinkedBinaryName);
755
756 assert(RegisterGlobalsFunc && "Expecting at least dummy function!");
757 llvm::Value *Args[] = {RegisterGlobalsFunc,
758 CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy),
759 ModuleIDConstant,
760 makeDummyFunction(getCallbackFnTy())};
761 CtorBuilder.CreateCall(RegisterLinkedBinaryFunc, Args);
762 }
763
764 // Create destructor and register it with atexit() the way NVCC does it. Doing
765 // it during regular destructor phase worked in CUDA before 9.2 but results in
766 // double-free in 9.2.
767 if (llvm::Function *CleanupFn = makeModuleDtorFunction()) {
768 // extern "C" int atexit(void (*f)(void));
769 llvm::FunctionType *AtExitTy =
770 llvm::FunctionType::get(IntTy, CleanupFn->getType(), false);
771 llvm::FunctionCallee AtExitFunc =
772 CGM.CreateRuntimeFunction(AtExitTy, "atexit", llvm::AttributeList(),
773 /*Local=*/true);
774 CtorBuilder.CreateCall(AtExitFunc, CleanupFn);
775 }
776
777 CtorBuilder.CreateRetVoid();
778 return ModuleCtorFunc;
779 }
780
781 /// Creates a global destructor function that unregisters the GPU code blob
782 /// registered by constructor.
783 ///
784 /// For CUDA:
785 /// \code
786 /// void __cuda_module_dtor(void*) {
787 /// __cudaUnregisterFatBinary(Handle);
788 /// }
789 /// \endcode
790 ///
791 /// For HIP:
792 /// \code
793 /// void __hip_module_dtor(void*) {
794 /// if (__hip_gpubin_handle) {
795 /// __hipUnregisterFatBinary(__hip_gpubin_handle);
796 /// __hip_gpubin_handle = 0;
797 /// }
798 /// }
799 /// \endcode
makeModuleDtorFunction()800 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
801 // No need for destructor if we don't have a handle to unregister.
802 if (!GpuBinaryHandle)
803 return nullptr;
804
805 // void __cudaUnregisterFatBinary(void ** handle);
806 llvm::FunctionCallee UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
807 llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
808 addUnderscoredPrefixToName("UnregisterFatBinary"));
809
810 llvm::Function *ModuleDtorFunc = llvm::Function::Create(
811 llvm::FunctionType::get(VoidTy, VoidPtrTy, false),
812 llvm::GlobalValue::InternalLinkage,
813 addUnderscoredPrefixToName("_module_dtor"), &TheModule);
814
815 llvm::BasicBlock *DtorEntryBB =
816 llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
817 CGBuilderTy DtorBuilder(CGM, Context);
818 DtorBuilder.SetInsertPoint(DtorEntryBB);
819
820 Address GpuBinaryAddr(GpuBinaryHandle, CharUnits::fromQuantity(
821 GpuBinaryHandle->getAlignment()));
822 auto HandleValue = DtorBuilder.CreateLoad(GpuBinaryAddr);
823 // There is only one HIP fat binary per linked module, however there are
824 // multiple destructor functions. Make sure the fat binary is unregistered
825 // only once.
826 if (CGM.getLangOpts().HIP) {
827 llvm::BasicBlock *IfBlock =
828 llvm::BasicBlock::Create(Context, "if", ModuleDtorFunc);
829 llvm::BasicBlock *ExitBlock =
830 llvm::BasicBlock::Create(Context, "exit", ModuleDtorFunc);
831 llvm::Constant *Zero = llvm::Constant::getNullValue(HandleValue->getType());
832 llvm::Value *NEZero = DtorBuilder.CreateICmpNE(HandleValue, Zero);
833 DtorBuilder.CreateCondBr(NEZero, IfBlock, ExitBlock);
834
835 DtorBuilder.SetInsertPoint(IfBlock);
836 DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
837 DtorBuilder.CreateStore(Zero, GpuBinaryAddr);
838 DtorBuilder.CreateBr(ExitBlock);
839
840 DtorBuilder.SetInsertPoint(ExitBlock);
841 } else {
842 DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
843 }
844 DtorBuilder.CreateRetVoid();
845 return ModuleDtorFunc;
846 }
847
CreateNVCUDARuntime(CodeGenModule & CGM)848 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
849 return new CGNVCUDARuntime(CGM);
850 }
851