1 //===-- AMDGPUCtorDtorLowering.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 "AMDGPUCtorDtorLowering.h"
14 #include "AMDGPU.h"
15 #include "llvm/IR/Constants.h"
16 #include "llvm/IR/Function.h"
17 #include "llvm/IR/GlobalVariable.h"
18 #include "llvm/IR/IRBuilder.h"
19 #include "llvm/IR/Module.h"
20 #include "llvm/IR/Value.h"
21 #include "llvm/Pass.h"
22 #include "llvm/Transforms/Utils/ModuleUtils.h"
23
24 using namespace llvm;
25
26 #define DEBUG_TYPE "amdgpu-lower-ctor-dtor"
27
28 namespace {
29
createInitOrFiniKernelFunction(Module & M,bool IsCtor)30 static Function *createInitOrFiniKernelFunction(Module &M, bool IsCtor) {
31 StringRef InitOrFiniKernelName = "amdgcn.device.init";
32 if (!IsCtor)
33 InitOrFiniKernelName = "amdgcn.device.fini";
34 if (M.getFunction(InitOrFiniKernelName))
35 return nullptr;
36
37 Function *InitOrFiniKernel = Function::createWithDefaultAttr(
38 FunctionType::get(Type::getVoidTy(M.getContext()), false),
39 GlobalValue::WeakODRLinkage, 0, InitOrFiniKernelName, &M);
40 InitOrFiniKernel->setCallingConv(CallingConv::AMDGPU_KERNEL);
41 InitOrFiniKernel->addFnAttr("amdgpu-flat-work-group-size", "1,1");
42 if (IsCtor)
43 InitOrFiniKernel->addFnAttr("device-init");
44 else
45 InitOrFiniKernel->addFnAttr("device-fini");
46 return InitOrFiniKernel;
47 }
48
49 // The linker will provide the associated symbols to allow us to traverse the
50 // global constructors / destructors in priority order. We create the IR
51 // required to call each callback in this section. This is equivalent to the
52 // following code.
53 //
54 // extern "C" void * __init_array_start[];
55 // extern "C" void * __init_array_end[];
56 // extern "C" void * __fini_array_start[];
57 // extern "C" void * __fini_array_end[];
58 //
59 // using InitCallback = void();
60 // using FiniCallback = void(void);
61 //
62 // void call_init_array_callbacks() {
63 // for (auto start = __init_array_start; start != __init_array_end; ++start)
64 // reinterpret_cast<InitCallback *>(*start)();
65 // }
66 //
67 // void call_fini_array_callbacks() {
68 // size_t fini_array_size = __fini_array_end - __fini_array_start;
69 // for (size_t i = fini_array_size; i > 0; --i)
70 // reinterpret_cast<FiniCallback *>(__fini_array_start[i - 1])();
71 // }
createInitOrFiniCalls(Function & F,bool IsCtor)72 static void createInitOrFiniCalls(Function &F, bool IsCtor) {
73 Module &M = *F.getParent();
74 LLVMContext &C = M.getContext();
75
76 IRBuilder<> IRB(BasicBlock::Create(C, "entry", &F));
77 auto *LoopBB = BasicBlock::Create(C, "while.entry", &F);
78 auto *ExitBB = BasicBlock::Create(C, "while.end", &F);
79 Type *PtrTy = IRB.getPtrTy(AMDGPUAS::GLOBAL_ADDRESS);
80
81 auto *Begin = M.getOrInsertGlobal(
82 IsCtor ? "__init_array_start" : "__fini_array_start",
83 ArrayType::get(PtrTy, 0), [&]() {
84 return new GlobalVariable(
85 M, ArrayType::get(PtrTy, 0),
86 /*isConstant=*/true, GlobalValue::ExternalLinkage,
87 /*Initializer=*/nullptr,
88 IsCtor ? "__init_array_start" : "__fini_array_start",
89 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal,
90 /*AddressSpace=*/1);
91 });
92 auto *End = M.getOrInsertGlobal(
93 IsCtor ? "__init_array_end" : "__fini_array_end",
94 ArrayType::get(PtrTy, 0), [&]() {
95 return new GlobalVariable(
96 M, ArrayType::get(PtrTy, 0),
97 /*isConstant=*/true, GlobalValue::ExternalLinkage,
98 /*Initializer=*/nullptr,
99 IsCtor ? "__init_array_end" : "__fini_array_end",
100 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal,
101 /*AddressSpace=*/1);
102 });
103
104 // The constructor type is suppoed to allow using the argument vectors, but
105 // for now we just call them with no arguments.
106 auto *CallBackTy = FunctionType::get(IRB.getVoidTy(), {});
107
108 Value *Start = Begin;
109 Value *Stop = End;
110 // The destructor array must be called in reverse order. Get a constant
111 // expression to the end of the array and iterate backwards instead.
112 if (!IsCtor) {
113 Type *Int64Ty = IntegerType::getInt64Ty(C);
114 auto *EndPtr = IRB.CreatePtrToInt(End, Int64Ty);
115 auto *BeginPtr = IRB.CreatePtrToInt(Begin, Int64Ty);
116 auto *ByteSize = IRB.CreateSub(EndPtr, BeginPtr);
117 auto *Size = IRB.CreateAShr(ByteSize, ConstantInt::get(Int64Ty, 3));
118 auto *Offset = IRB.CreateSub(Size, ConstantInt::get(Int64Ty, 1));
119 Start = IRB.CreateInBoundsGEP(
120 ArrayType::get(IRB.getPtrTy(), 0), Begin,
121 ArrayRef<Value *>({ConstantInt::get(Int64Ty, 0), Offset}));
122 Stop = Begin;
123 }
124
125 IRB.CreateCondBr(
126 IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_NE : ICmpInst::ICMP_UGE, Start,
127 Stop),
128 LoopBB, ExitBB);
129 IRB.SetInsertPoint(LoopBB);
130 auto *CallBackPHI = IRB.CreatePHI(PtrTy, 2, "ptr");
131 auto *CallBack = IRB.CreateLoad(IRB.getPtrTy(F.getAddressSpace()),
132 CallBackPHI, "callback");
133 IRB.CreateCall(CallBackTy, CallBack);
134 auto *NewCallBack =
135 IRB.CreateConstGEP1_64(PtrTy, CallBackPHI, IsCtor ? 1 : -1, "next");
136 auto *EndCmp = IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_ULT,
137 NewCallBack, Stop, "end");
138 CallBackPHI->addIncoming(Start, &F.getEntryBlock());
139 CallBackPHI->addIncoming(NewCallBack, LoopBB);
140 IRB.CreateCondBr(EndCmp, ExitBB, LoopBB);
141 IRB.SetInsertPoint(ExitBB);
142 IRB.CreateRetVoid();
143 }
144
createInitOrFiniKernel(Module & M,StringRef GlobalName,bool IsCtor)145 static bool createInitOrFiniKernel(Module &M, StringRef GlobalName,
146 bool IsCtor) {
147 GlobalVariable *GV = M.getGlobalVariable(GlobalName);
148 if (!GV || !GV->hasInitializer())
149 return false;
150 ConstantArray *GA = dyn_cast<ConstantArray>(GV->getInitializer());
151 if (!GA || GA->getNumOperands() == 0)
152 return false;
153
154 Function *InitOrFiniKernel = createInitOrFiniKernelFunction(M, IsCtor);
155 if (!InitOrFiniKernel)
156 return false;
157
158 createInitOrFiniCalls(*InitOrFiniKernel, IsCtor);
159
160 appendToUsed(M, {InitOrFiniKernel});
161 return true;
162 }
163
lowerCtorsAndDtors(Module & M)164 static bool lowerCtorsAndDtors(Module &M) {
165 bool Modified = false;
166 Modified |= createInitOrFiniKernel(M, "llvm.global_ctors", /*IsCtor =*/true);
167 Modified |= createInitOrFiniKernel(M, "llvm.global_dtors", /*IsCtor =*/false);
168 return Modified;
169 }
170
171 class AMDGPUCtorDtorLoweringLegacy final : public ModulePass {
172 public:
173 static char ID;
AMDGPUCtorDtorLoweringLegacy()174 AMDGPUCtorDtorLoweringLegacy() : ModulePass(ID) {}
runOnModule(Module & M)175 bool runOnModule(Module &M) override { return lowerCtorsAndDtors(M); }
176 };
177
178 } // End anonymous namespace
179
run(Module & M,ModuleAnalysisManager & AM)180 PreservedAnalyses AMDGPUCtorDtorLoweringPass::run(Module &M,
181 ModuleAnalysisManager &AM) {
182 return lowerCtorsAndDtors(M) ? PreservedAnalyses::none()
183 : PreservedAnalyses::all();
184 }
185
186 char AMDGPUCtorDtorLoweringLegacy::ID = 0;
187 char &llvm::AMDGPUCtorDtorLoweringLegacyPassID =
188 AMDGPUCtorDtorLoweringLegacy::ID;
189 INITIALIZE_PASS(AMDGPUCtorDtorLoweringLegacy, DEBUG_TYPE,
190 "Lower ctors and dtors for AMDGPU", false, false)
191
createAMDGPUCtorDtorLoweringLegacyPass()192 ModulePass *llvm::createAMDGPUCtorDtorLoweringLegacyPass() {
193 return new AMDGPUCtorDtorLoweringLegacy();
194 }
195