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//===----------------------------------------------------------------------===//
12
13#include "NVPTXCtorDtorLowering.h"
14#include "MCTargetDesc/NVPTXBaseInfo.h"
15#include "NVPTX.h"
16#include "llvm/ADT/StringExtras.h"
17#include "llvm/IR/CallingConv.h"
18#include "llvm/IR/Constants.h"
19#include "llvm/IR/Function.h"
20#include "llvm/IR/GlobalVariable.h"
21#include "llvm/IR/IRBuilder.h"
22#include "llvm/IR/Module.h"
23#include "llvm/IR/Value.h"
24#include "llvm/Pass.h"
25#include "llvm/Support/CommandLine.h"
26#include "llvm/Support/MD5.h"
27#include "llvm/Support/NVVMAttributes.h"
28#include "llvm/Transforms/Utils/ModuleUtils.h"
29
30using namespace llvm;
31
32#define DEBUG_TYPE "nvptx-lower-ctor-dtor"
33
34static cl::opt<std::string>
35 GlobalStr("nvptx-lower-global-ctor-dtor-id",
36 cl::desc("Override unique ID of ctor/dtor globals."),
37 cl::init(Val: ""), cl::Hidden);
38
39static cl::opt<bool>
40 CreateKernels("nvptx-emit-init-fini-kernel",
41 cl::desc("Emit kernels to call ctor/dtor globals."),
42 cl::init(Val: true), cl::Hidden);
43
44namespace {
45
46static std::string getHash(StringRef Str) {
47 llvm::MD5 Hasher;
48 llvm::MD5::MD5Result Hash;
49 Hasher.update(Str);
50 Hasher.final(Result&: Hash);
51 return llvm::utohexstr(X: Hash.low(), /*LowerCase=*/true);
52}
53
54static void addKernelAttrs(Function *F) {
55 F->addFnAttr(Kind: NVVMAttr::MaxClusterRank, Val: "1");
56 F->addFnAttr(Kind: NVVMAttr::MaxNTID, Val: "1");
57 F->setCallingConv(CallingConv::PTX_Kernel);
58}
59
60static Function *createInitOrFiniKernelFunction(Module &M, bool IsCtor) {
61 StringRef InitOrFiniKernelName =
62 IsCtor ? "nvptx$device$init" : "nvptx$device$fini";
63 if (M.getFunction(Name: InitOrFiniKernelName))
64 return nullptr;
65
66 Function *InitOrFiniKernel = Function::createWithDefaultAttr(
67 Ty: FunctionType::get(Result: Type::getVoidTy(C&: M.getContext()), isVarArg: false),
68 Linkage: GlobalValue::WeakODRLinkage, AddrSpace: 0, N: InitOrFiniKernelName, M: &M);
69 addKernelAttrs(F: InitOrFiniKernel);
70
71 return InitOrFiniKernel;
72}
73
74// We create the IR required to call each callback in this section. This is
75// equivalent to the following code. Normally, the linker would provide us with
76// the definitions of the init and fini array sections. The 'nvlink' linker does
77// not do this so initializing these values is done by the runtime.
78//
79// extern "C" void **__init_array_start = nullptr;
80// extern "C" void **__init_array_end = nullptr;
81// extern "C" void **__fini_array_start = nullptr;
82// extern "C" void **__fini_array_end = nullptr;
83//
84// using InitCallback = void();
85// using FiniCallback = void();
86//
87// void call_init_array_callbacks() {
88// for (auto start = __init_array_start; start != __init_array_end; ++start)
89// reinterpret_cast<InitCallback *>(*start)();
90// }
91//
92// void call_fini_array_callbacks() {
93// size_t fini_array_size = __fini_array_end - __fini_array_start;
94// for (size_t i = fini_array_size; i > 0; --i)
95// reinterpret_cast<FiniCallback *>(__fini_array_start[i - 1])();
96// }
97static void createInitOrFiniCalls(Function &F, bool IsCtor) {
98 Module &M = *F.getParent();
99 LLVMContext &C = M.getContext();
100
101 IRBuilder<> IRB(BasicBlock::Create(Context&: C, Name: "entry", Parent: &F));
102 auto *LoopBB = BasicBlock::Create(Context&: C, Name: "while.entry", Parent: &F);
103 auto *ExitBB = BasicBlock::Create(Context&: C, Name: "while.end", Parent: &F);
104 Type *PtrTy = IRB.getPtrTy(AddrSpace: llvm::ADDRESS_SPACE_GLOBAL);
105
106 auto *Begin = M.getOrInsertGlobal(
107 Name: IsCtor ? "__init_array_start" : "__fini_array_start",
108 Ty: PointerType::get(C, AddressSpace: 0), CreateGlobalCallback: [&]() {
109 auto *GV = new GlobalVariable(
110 M, PointerType::get(C, AddressSpace: 0),
111 /*isConstant=*/false, GlobalValue::WeakAnyLinkage,
112 Constant::getNullValue(Ty: PointerType::get(C, AddressSpace: 0)),
113 IsCtor ? "__init_array_start" : "__fini_array_start",
114 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal,
115 /*AddressSpace=*/llvm::ADDRESS_SPACE_GLOBAL);
116 GV->setVisibility(GlobalVariable::ProtectedVisibility);
117 return GV;
118 });
119 auto *End = M.getOrInsertGlobal(
120 Name: IsCtor ? "__init_array_end" : "__fini_array_end", Ty: PointerType::get(C, AddressSpace: 0),
121 CreateGlobalCallback: [&]() {
122 auto *GV = new GlobalVariable(
123 M, PointerType::get(C, AddressSpace: 0),
124 /*isConstant=*/false, GlobalValue::WeakAnyLinkage,
125 Constant::getNullValue(Ty: PointerType::get(C, AddressSpace: 0)),
126 IsCtor ? "__init_array_end" : "__fini_array_end",
127 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal,
128 /*AddressSpace=*/llvm::ADDRESS_SPACE_GLOBAL);
129 GV->setVisibility(GlobalVariable::ProtectedVisibility);
130 return GV;
131 });
132
133 // The constructor type is suppoed to allow using the argument vectors, but
134 // for now we just call them with no arguments.
135 auto *CallBackTy = FunctionType::get(Result: IRB.getVoidTy(), isVarArg: {});
136
137 // The destructor array must be called in reverse order. Get an expression to
138 // the end of the array and iterate backwards in that case.
139 Value *BeginVal = IRB.CreateLoad(Ty: Begin->getType(), Ptr: Begin, Name: "begin");
140 Value *EndVal = IRB.CreateLoad(Ty: Begin->getType(), Ptr: End, Name: "stop");
141 if (!IsCtor) {
142 Value *OldBeginVal = BeginVal;
143 BeginVal =
144 IRB.CreateInBoundsGEP(Ty: PointerType::get(C, AddressSpace: 0), Ptr: EndVal,
145 IdxList: ArrayRef<Value *>(ConstantInt::getAllOnesValue(
146 Ty: IntegerType::getInt64Ty(C))),
147 Name: "start");
148 EndVal = OldBeginVal;
149 }
150 IRB.CreateCondBr(
151 Cond: IRB.CreateCmp(Pred: IsCtor ? ICmpInst::ICMP_NE : ICmpInst::ICMP_UGE, LHS: BeginVal,
152 RHS: EndVal),
153 True: LoopBB, False: ExitBB);
154 IRB.SetInsertPoint(LoopBB);
155 auto *CallBackPHI = IRB.CreatePHI(Ty: PtrTy, NumReservedValues: 2, Name: "ptr");
156 auto *CallBack = IRB.CreateLoad(Ty: IRB.getPtrTy(AddrSpace: F.getAddressSpace()),
157 Ptr: CallBackPHI, Name: "callback");
158 IRB.CreateCall(FTy: CallBackTy, Callee: CallBack);
159 auto *NewCallBack =
160 IRB.CreateConstGEP1_64(Ty: PtrTy, Ptr: CallBackPHI, Idx0: IsCtor ? 1 : -1, Name: "next");
161 auto *EndCmp = IRB.CreateCmp(Pred: IsCtor ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_ULT,
162 LHS: NewCallBack, RHS: EndVal, Name: "end");
163 CallBackPHI->addIncoming(V: BeginVal, BB: &F.getEntryBlock());
164 CallBackPHI->addIncoming(V: NewCallBack, BB: LoopBB);
165 IRB.CreateCondBr(Cond: EndCmp, True: ExitBB, False: LoopBB);
166 IRB.SetInsertPoint(ExitBB);
167 IRB.CreateRetVoid();
168}
169
170static bool createInitOrFiniGlobals(Module &M, GlobalVariable *GV,
171 bool IsCtor) {
172 ConstantArray *GA = dyn_cast<ConstantArray>(Val: GV->getInitializer());
173 if (!GA || GA->getNumOperands() == 0)
174 return false;
175
176 // NVPTX has no way to emit variables at specific sections or support for
177 // the traditional constructor sections. Instead, we emit mangled global
178 // names so the runtime can build the list manually.
179 for (Value *V : GA->operands()) {
180 auto *CS = cast<ConstantStruct>(Val: V);
181 auto *F = cast<Constant>(Val: CS->getOperand(i_nocapture: 1));
182 uint64_t Priority = cast<ConstantInt>(Val: CS->getOperand(i_nocapture: 0))->getSExtValue();
183 std::string PriorityStr = "." + std::to_string(val: Priority);
184 // We append a semi-unique hash and the priority to the global name.
185 std::string GlobalID =
186 !GlobalStr.empty() ? GlobalStr : getHash(Str: M.getSourceFileName());
187 std::string NameStr =
188 ((IsCtor ? "__init_array_object_" : "__fini_array_object_") +
189 F->getName() + "_" + GlobalID + "_" + std::to_string(val: Priority))
190 .str();
191 // PTX does not support exported names with '.' in them.
192 llvm::transform(Range&: NameStr, d_first: NameStr.begin(),
193 F: [](char c) { return c == '.' ? '_' : c; });
194
195 auto *GV = new GlobalVariable(M, F->getType(), /*IsConstant=*/true,
196 GlobalValue::ExternalLinkage, F, NameStr,
197 nullptr, GlobalValue::NotThreadLocal,
198 /*AddressSpace=*/4);
199 // This isn't respected by Nvidia, simply put here for clarity.
200 GV->setSection(IsCtor ? ".init_array" + PriorityStr
201 : ".fini_array" + PriorityStr);
202 GV->setVisibility(GlobalVariable::ProtectedVisibility);
203 appendToUsed(M, Values: {GV});
204 }
205
206 return true;
207}
208
209static bool createInitOrFiniKernel(Module &M, StringRef GlobalName,
210 bool IsCtor) {
211 GlobalVariable *GV = M.getGlobalVariable(Name: GlobalName);
212 if (!GV || !GV->hasInitializer())
213 return false;
214
215 if (!createInitOrFiniGlobals(M, GV, IsCtor))
216 return false;
217
218 if (!CreateKernels)
219 return true;
220
221 Function *InitOrFiniKernel = createInitOrFiniKernelFunction(M, IsCtor);
222 if (!InitOrFiniKernel)
223 return false;
224
225 createInitOrFiniCalls(F&: *InitOrFiniKernel, IsCtor);
226
227 GV->eraseFromParent();
228 return true;
229}
230
231static bool lowerCtorsAndDtors(Module &M) {
232 bool Modified = false;
233 Modified |= createInitOrFiniKernel(M, GlobalName: "llvm.global_ctors", /*IsCtor =*/true);
234 Modified |= createInitOrFiniKernel(M, GlobalName: "llvm.global_dtors", /*IsCtor =*/false);
235 return Modified;
236}
237
238class NVPTXCtorDtorLoweringLegacy final : public ModulePass {
239public:
240 static char ID;
241 NVPTXCtorDtorLoweringLegacy() : ModulePass(ID) {}
242 bool runOnModule(Module &M) override { return lowerCtorsAndDtors(M); }
243};
244
245} // End anonymous namespace
246
247PreservedAnalyses NVPTXCtorDtorLoweringPass::run(Module &M,
248 ModuleAnalysisManager &AM) {
249 return lowerCtorsAndDtors(M) ? PreservedAnalyses::none()
250 : PreservedAnalyses::all();
251}
252
253char NVPTXCtorDtorLoweringLegacy::ID = 0;
254INITIALIZE_PASS(NVPTXCtorDtorLoweringLegacy, DEBUG_TYPE,
255 "Lower ctors and dtors for NVPTX", false, false)
256
257ModulePass *llvm::createNVPTXCtorDtorLoweringLegacyPass() {
258 return new NVPTXCtorDtorLoweringLegacy();
259}
260