1 //===-- AMDGPUAlwaysInlinePass.cpp - Promote Allocas ----------------------===// 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 marks all internal functions as always_inline and creates 11 /// duplicates of all other functions and marks the duplicates as always_inline. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "AMDGPU.h" 16 #include "AMDGPUTargetMachine.h" 17 #include "Utils/AMDGPUBaseInfo.h" 18 #include "llvm/IR/Module.h" 19 #include "llvm/Pass.h" 20 #include "llvm/Support/CommandLine.h" 21 22 using namespace llvm; 23 24 namespace { 25 26 static cl::opt<bool> StressCalls( 27 "amdgpu-stress-function-calls", 28 cl::Hidden, 29 cl::desc("Force all functions to be noinline"), 30 cl::init(false)); 31 32 class AMDGPUAlwaysInline : public ModulePass { 33 bool GlobalOpt; 34 35 public: 36 static char ID; 37 38 AMDGPUAlwaysInline(bool GlobalOpt = false) : 39 ModulePass(ID), GlobalOpt(GlobalOpt) { } 40 bool runOnModule(Module &M) override; 41 42 void getAnalysisUsage(AnalysisUsage &AU) const override { 43 AU.setPreservesAll(); 44 } 45 }; 46 47 } // End anonymous namespace 48 49 INITIALIZE_PASS(AMDGPUAlwaysInline, "amdgpu-always-inline", 50 "AMDGPU Inline All Functions", false, false) 51 52 char AMDGPUAlwaysInline::ID = 0; 53 54 static void 55 recursivelyVisitUsers(GlobalValue &GV, 56 SmallPtrSetImpl<Function *> &FuncsToAlwaysInline) { 57 SmallVector<User *, 16> Stack; 58 59 SmallPtrSet<const Value *, 8> Visited; 60 61 for (User *U : GV.users()) 62 Stack.push_back(U); 63 64 while (!Stack.empty()) { 65 User *U = Stack.pop_back_val(); 66 if (!Visited.insert(U).second) 67 continue; 68 69 if (Instruction *I = dyn_cast<Instruction>(U)) { 70 Function *F = I->getParent()->getParent(); 71 if (!AMDGPU::isEntryFunctionCC(F->getCallingConv())) { 72 // FIXME: This is a horrible hack. We should always respect noinline, 73 // and just let us hit the error when we can't handle this. 74 // 75 // Unfortunately, clang adds noinline to all functions at -O0. We have 76 // to override this here. until that's fixed. 77 F->removeFnAttr(Attribute::NoInline); 78 79 FuncsToAlwaysInline.insert(F); 80 Stack.push_back(F); 81 } 82 83 // No need to look at further users, but we do need to inline any callers. 84 continue; 85 } 86 87 for (User *UU : U->users()) 88 Stack.push_back(UU); 89 } 90 } 91 92 static bool alwaysInlineImpl(Module &M, bool GlobalOpt) { 93 std::vector<GlobalAlias*> AliasesToRemove; 94 95 SmallPtrSet<Function *, 8> FuncsToAlwaysInline; 96 SmallPtrSet<Function *, 8> FuncsToNoInline; 97 98 for (GlobalAlias &A : M.aliases()) { 99 if (Function* F = dyn_cast<Function>(A.getAliasee())) { 100 A.replaceAllUsesWith(F); 101 AliasesToRemove.push_back(&A); 102 } 103 104 // FIXME: If the aliasee isn't a function, it's some kind of constant expr 105 // cast that won't be inlined through. 106 } 107 108 if (GlobalOpt) { 109 for (GlobalAlias* A : AliasesToRemove) { 110 A->eraseFromParent(); 111 } 112 } 113 114 // Always force inlining of any function that uses an LDS global address. This 115 // is something of a workaround because we don't have a way of supporting LDS 116 // objects defined in functions. LDS is always allocated by a kernel, and it 117 // is difficult to manage LDS usage if a function may be used by multiple 118 // kernels. 119 // 120 // OpenCL doesn't allow declaring LDS in non-kernels, so in practice this 121 // should only appear when IPO passes manages to move LDs defined in a kernel 122 // into a single user function. 123 124 for (GlobalVariable &GV : M.globals()) { 125 // TODO: Region address 126 unsigned AS = GV.getAddressSpace(); 127 if (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::REGION_ADDRESS) 128 continue; 129 130 recursivelyVisitUsers(GV, FuncsToAlwaysInline); 131 } 132 133 if (!AMDGPUTargetMachine::EnableFunctionCalls || StressCalls) { 134 auto IncompatAttr 135 = StressCalls ? Attribute::AlwaysInline : Attribute::NoInline; 136 137 for (Function &F : M) { 138 if (!F.isDeclaration() && !F.use_empty() && 139 !F.hasFnAttribute(IncompatAttr)) { 140 if (StressCalls) { 141 if (!FuncsToAlwaysInline.count(&F)) 142 FuncsToNoInline.insert(&F); 143 } else 144 FuncsToAlwaysInline.insert(&F); 145 } 146 } 147 } 148 149 for (Function *F : FuncsToAlwaysInline) 150 F->addFnAttr(Attribute::AlwaysInline); 151 152 for (Function *F : FuncsToNoInline) 153 F->addFnAttr(Attribute::NoInline); 154 155 return !FuncsToAlwaysInline.empty() || !FuncsToNoInline.empty(); 156 } 157 158 bool AMDGPUAlwaysInline::runOnModule(Module &M) { 159 return alwaysInlineImpl(M, GlobalOpt); 160 } 161 162 ModulePass *llvm::createAMDGPUAlwaysInlinePass(bool GlobalOpt) { 163 return new AMDGPUAlwaysInline(GlobalOpt); 164 } 165 166 PreservedAnalyses AMDGPUAlwaysInlinePass::run(Module &M, 167 ModuleAnalysisManager &AM) { 168 alwaysInlineImpl(M, GlobalOpt); 169 return PreservedAnalyses::all(); 170 } 171