[RISCV] Add support for Smepmp 1.0 (#78489)
[llvm-project.git] / llvm / lib / Target / NVPTX / NVPTXCtorDtorLowering.cpp
blobf77a1f0272c88b8016eac8d2bfc0832e8a7de33c
1 //===-- NVPTXCtorDtorLowering.cpp - Handle global ctors and dtors --------===//
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 /// \file
10 /// This pass creates a unified init and fini kernel with the required metadata
11 //===----------------------------------------------------------------------===//
13 #include "NVPTXCtorDtorLowering.h"
14 #include "MCTargetDesc/NVPTXBaseInfo.h"
15 #include "NVPTX.h"
16 #include "llvm/ADT/StringExtras.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/Function.h"
19 #include "llvm/IR/GlobalVariable.h"
20 #include "llvm/IR/IRBuilder.h"
21 #include "llvm/IR/Module.h"
22 #include "llvm/IR/Value.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Transforms/Utils/ModuleUtils.h"
27 using namespace llvm;
29 #define DEBUG_TYPE "nvptx-lower-ctor-dtor"
31 static cl::opt<std::string>
32 GlobalStr("nvptx-lower-global-ctor-dtor-id",
33 cl::desc("Override unique ID of ctor/dtor globals."),
34 cl::init(""), cl::Hidden);
36 static cl::opt<bool>
37 CreateKernels("nvptx-emit-init-fini-kernel",
38 cl::desc("Emit kernels to call ctor/dtor globals."),
39 cl::init(true), cl::Hidden);
41 namespace {
43 static std::string getHash(StringRef Str) {
44 llvm::MD5 Hasher;
45 llvm::MD5::MD5Result Hash;
46 Hasher.update(Str);
47 Hasher.final(Hash);
48 return llvm::utohexstr(Hash.low(), /*LowerCase=*/true);
51 static void addKernelMetadata(Module &M, GlobalValue *GV) {
52 llvm::LLVMContext &Ctx = M.getContext();
54 // Get "nvvm.annotations" metadata node.
55 llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations");
57 llvm::Metadata *KernelMDVals[] = {
58 llvm::ConstantAsMetadata::get(GV), llvm::MDString::get(Ctx, "kernel"),
59 llvm::ConstantAsMetadata::get(
60 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
62 // This kernel is only to be called single-threaded.
63 llvm::Metadata *ThreadXMDVals[] = {
64 llvm::ConstantAsMetadata::get(GV), llvm::MDString::get(Ctx, "maxntidx"),
65 llvm::ConstantAsMetadata::get(
66 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
67 llvm::Metadata *ThreadYMDVals[] = {
68 llvm::ConstantAsMetadata::get(GV), llvm::MDString::get(Ctx, "maxntidy"),
69 llvm::ConstantAsMetadata::get(
70 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
71 llvm::Metadata *ThreadZMDVals[] = {
72 llvm::ConstantAsMetadata::get(GV), llvm::MDString::get(Ctx, "maxntidz"),
73 llvm::ConstantAsMetadata::get(
74 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
76 llvm::Metadata *BlockMDVals[] = {
77 llvm::ConstantAsMetadata::get(GV),
78 llvm::MDString::get(Ctx, "maxclusterrank"),
79 llvm::ConstantAsMetadata::get(
80 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))};
82 // Append metadata to nvvm.annotations.
83 MD->addOperand(llvm::MDNode::get(Ctx, KernelMDVals));
84 MD->addOperand(llvm::MDNode::get(Ctx, ThreadXMDVals));
85 MD->addOperand(llvm::MDNode::get(Ctx, ThreadYMDVals));
86 MD->addOperand(llvm::MDNode::get(Ctx, ThreadZMDVals));
87 MD->addOperand(llvm::MDNode::get(Ctx, BlockMDVals));
90 static Function *createInitOrFiniKernelFunction(Module &M, bool IsCtor) {
91 StringRef InitOrFiniKernelName =
92 IsCtor ? "nvptx$device$init" : "nvptx$device$fini";
93 if (M.getFunction(InitOrFiniKernelName))
94 return nullptr;
96 Function *InitOrFiniKernel = Function::createWithDefaultAttr(
97 FunctionType::get(Type::getVoidTy(M.getContext()), false),
98 GlobalValue::WeakODRLinkage, 0, InitOrFiniKernelName, &M);
99 addKernelMetadata(M, InitOrFiniKernel);
101 return InitOrFiniKernel;
104 // We create the IR required to call each callback in this section. This is
105 // equivalent to the following code. Normally, the linker would provide us with
106 // the definitions of the init and fini array sections. The 'nvlink' linker does
107 // not do this so initializing these values is done by the runtime.
109 // extern "C" void **__init_array_start = nullptr;
110 // extern "C" void **__init_array_end = nullptr;
111 // extern "C" void **__fini_array_start = nullptr;
112 // extern "C" void **__fini_array_end = nullptr;
114 // using InitCallback = void();
115 // using FiniCallback = void();
117 // void call_init_array_callbacks() {
118 // for (auto start = __init_array_start; start != __init_array_end; ++start)
119 // reinterpret_cast<InitCallback *>(*start)();
120 // }
122 // void call_init_array_callbacks() {
123 // size_t fini_array_size = __fini_array_end - __fini_array_start;
124 // for (size_t i = fini_array_size; i > 0; --i)
125 // reinterpret_cast<FiniCallback *>(__fini_array_start[i - 1])();
126 // }
127 static void createInitOrFiniCalls(Function &F, bool IsCtor) {
128 Module &M = *F.getParent();
129 LLVMContext &C = M.getContext();
131 IRBuilder<> IRB(BasicBlock::Create(C, "entry", &F));
132 auto *LoopBB = BasicBlock::Create(C, "while.entry", &F);
133 auto *ExitBB = BasicBlock::Create(C, "while.end", &F);
134 Type *PtrTy = IRB.getPtrTy(llvm::ADDRESS_SPACE_GLOBAL);
136 auto *Begin = M.getOrInsertGlobal(
137 IsCtor ? "__init_array_start" : "__fini_array_start",
138 PointerType::get(C, 0), [&]() {
139 auto *GV = new GlobalVariable(
140 M, PointerType::get(C, 0),
141 /*isConstant=*/false, GlobalValue::WeakAnyLinkage,
142 Constant::getNullValue(PointerType::get(C, 0)),
143 IsCtor ? "__init_array_start" : "__fini_array_start",
144 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal,
145 /*AddressSpace=*/llvm::ADDRESS_SPACE_GLOBAL);
146 GV->setVisibility(GlobalVariable::ProtectedVisibility);
147 return GV;
149 auto *End = M.getOrInsertGlobal(
150 IsCtor ? "__init_array_end" : "__fini_array_end", PointerType::get(C, 0),
151 [&]() {
152 auto *GV = new GlobalVariable(
153 M, PointerType::get(C, 0),
154 /*isConstant=*/false, GlobalValue::WeakAnyLinkage,
155 Constant::getNullValue(PointerType::get(C, 0)),
156 IsCtor ? "__init_array_end" : "__fini_array_end",
157 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal,
158 /*AddressSpace=*/llvm::ADDRESS_SPACE_GLOBAL);
159 GV->setVisibility(GlobalVariable::ProtectedVisibility);
160 return GV;
163 // The constructor type is suppoed to allow using the argument vectors, but
164 // for now we just call them with no arguments.
165 auto *CallBackTy = FunctionType::get(IRB.getVoidTy(), {});
167 // The destructor array must be called in reverse order. Get an expression to
168 // the end of the array and iterate backwards in that case.
169 Value *BeginVal = IRB.CreateLoad(Begin->getType(), Begin, "begin");
170 Value *EndVal = IRB.CreateLoad(Begin->getType(), End, "stop");
171 if (!IsCtor) {
172 auto *BeginInt = IRB.CreatePtrToInt(BeginVal, IntegerType::getInt64Ty(C));
173 auto *EndInt = IRB.CreatePtrToInt(EndVal, IntegerType::getInt64Ty(C));
174 auto *SubInst = IRB.CreateSub(EndInt, BeginInt);
175 auto *Offset = IRB.CreateAShr(
176 SubInst, ConstantInt::get(IntegerType::getInt64Ty(C), 3), "offset",
177 /*IsExact=*/true);
178 auto *ValuePtr = IRB.CreateGEP(PointerType::get(C, 0), BeginVal,
179 ArrayRef<Value *>({Offset}));
180 EndVal = BeginVal;
181 BeginVal = IRB.CreateInBoundsGEP(
182 PointerType::get(C, 0), ValuePtr,
183 ArrayRef<Value *>(ConstantInt::get(IntegerType::getInt64Ty(C), -1)),
184 "start");
186 IRB.CreateCondBr(
187 IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_NE : ICmpInst::ICMP_UGT, BeginVal,
188 EndVal),
189 LoopBB, ExitBB);
190 IRB.SetInsertPoint(LoopBB);
191 auto *CallBackPHI = IRB.CreatePHI(PtrTy, 2, "ptr");
192 auto *CallBack = IRB.CreateLoad(IRB.getPtrTy(F.getAddressSpace()),
193 CallBackPHI, "callback");
194 IRB.CreateCall(CallBackTy, CallBack);
195 auto *NewCallBack =
196 IRB.CreateConstGEP1_64(PtrTy, CallBackPHI, IsCtor ? 1 : -1, "next");
197 auto *EndCmp = IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_ULT,
198 NewCallBack, EndVal, "end");
199 CallBackPHI->addIncoming(BeginVal, &F.getEntryBlock());
200 CallBackPHI->addIncoming(NewCallBack, LoopBB);
201 IRB.CreateCondBr(EndCmp, ExitBB, LoopBB);
202 IRB.SetInsertPoint(ExitBB);
203 IRB.CreateRetVoid();
206 static bool createInitOrFiniGlobals(Module &M, GlobalVariable *GV,
207 bool IsCtor) {
208 ConstantArray *GA = dyn_cast<ConstantArray>(GV->getInitializer());
209 if (!GA || GA->getNumOperands() == 0)
210 return false;
212 // NVPTX has no way to emit variables at specific sections or support for
213 // the traditional constructor sections. Instead, we emit mangled global
214 // names so the runtime can build the list manually.
215 for (Value *V : GA->operands()) {
216 auto *CS = cast<ConstantStruct>(V);
217 auto *F = cast<Constant>(CS->getOperand(1));
218 uint64_t Priority = cast<ConstantInt>(CS->getOperand(0))->getSExtValue();
219 std::string PriorityStr = "." + std::to_string(Priority);
220 // We append a semi-unique hash and the priority to the global name.
221 std::string GlobalID =
222 !GlobalStr.empty() ? GlobalStr : getHash(M.getSourceFileName());
223 std::string NameStr =
224 ((IsCtor ? "__init_array_object_" : "__fini_array_object_") +
225 F->getName() + "_" + GlobalID + "_" + std::to_string(Priority))
226 .str();
227 // PTX does not support exported names with '.' in them.
228 llvm::transform(NameStr, NameStr.begin(),
229 [](char c) { return c == '.' ? '_' : c; });
231 auto *GV = new GlobalVariable(M, F->getType(), /*IsConstant=*/true,
232 GlobalValue::ExternalLinkage, F, NameStr,
233 nullptr, GlobalValue::NotThreadLocal,
234 /*AddressSpace=*/4);
235 // This isn't respected by Nvidia, simply put here for clarity.
236 GV->setSection(IsCtor ? ".init_array" + PriorityStr
237 : ".fini_array" + PriorityStr);
238 GV->setVisibility(GlobalVariable::ProtectedVisibility);
239 appendToUsed(M, {GV});
242 return true;
245 static bool createInitOrFiniKernel(Module &M, StringRef GlobalName,
246 bool IsCtor) {
247 GlobalVariable *GV = M.getGlobalVariable(GlobalName);
248 if (!GV || !GV->hasInitializer())
249 return false;
251 if (!createInitOrFiniGlobals(M, GV, IsCtor))
252 return false;
254 if (!CreateKernels)
255 return true;
257 Function *InitOrFiniKernel = createInitOrFiniKernelFunction(M, IsCtor);
258 if (!InitOrFiniKernel)
259 return false;
261 createInitOrFiniCalls(*InitOrFiniKernel, IsCtor);
263 GV->eraseFromParent();
264 return true;
267 static bool lowerCtorsAndDtors(Module &M) {
268 bool Modified = false;
269 Modified |= createInitOrFiniKernel(M, "llvm.global_ctors", /*IsCtor =*/true);
270 Modified |= createInitOrFiniKernel(M, "llvm.global_dtors", /*IsCtor =*/false);
271 return Modified;
274 class NVPTXCtorDtorLoweringLegacy final : public ModulePass {
275 public:
276 static char ID;
277 NVPTXCtorDtorLoweringLegacy() : ModulePass(ID) {}
278 bool runOnModule(Module &M) override { return lowerCtorsAndDtors(M); }
281 } // End anonymous namespace
283 PreservedAnalyses NVPTXCtorDtorLoweringPass::run(Module &M,
284 ModuleAnalysisManager &AM) {
285 return lowerCtorsAndDtors(M) ? PreservedAnalyses::none()
286 : PreservedAnalyses::all();
289 char NVPTXCtorDtorLoweringLegacy::ID = 0;
290 char &llvm::NVPTXCtorDtorLoweringLegacyPassID = NVPTXCtorDtorLoweringLegacy::ID;
291 INITIALIZE_PASS(NVPTXCtorDtorLoweringLegacy, DEBUG_TYPE,
292 "Lower ctors and dtors for NVPTX", false, false)
294 ModulePass *llvm::createNVPTXCtorDtorLoweringLegacyPass() {
295 return new NVPTXCtorDtorLoweringLegacy();