1 //===- InlineAdvisor.cpp - analysis pass implementation -------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements inlining decision-making APIs. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Analysis/InlineAdvisor.h" 15 #include "llvm/ADT/Statistic.h" 16 #include "llvm/Analysis/InlineCost.h" 17 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 18 #include "llvm/Analysis/ProfileSummaryInfo.h" 19 #include "llvm/Analysis/TargetLibraryInfo.h" 20 #include "llvm/Analysis/TargetTransformInfo.h" 21 #include "llvm/IR/Instructions.h" 22 #include "llvm/Support/raw_ostream.h" 23 24 #include <sstream> 25 26 using namespace llvm; 27 #define DEBUG_TYPE "inline" 28 29 // This weirdly named statistic tracks the number of times that, when attempting 30 // to inline a function A into B, we analyze the callers of B in order to see 31 // if those would be more profitable and blocked inline steps. 32 STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed"); 33 34 /// Flag to add inline messages as callsite attributes 'inline-remark'. 35 static cl::opt<bool> 36 InlineRemarkAttribute("inline-remark-attribute", cl::init(false), 37 cl::Hidden, 38 cl::desc("Enable adding inline-remark attribute to" 39 " callsites processed by inliner but decided" 40 " to be not inlined")); 41 42 // An integer used to limit the cost of inline deferral. The default negative 43 // number tells shouldBeDeferred to only take the secondary cost into account. 44 static cl::opt<int> 45 InlineDeferralScale("inline-deferral-scale", 46 cl::desc("Scale to limit the cost of inline deferral"), 47 cl::init(-1), cl::Hidden); 48 49 /// Return true if inlining of CB can block the caller from being 50 /// inlined which is proved to be more beneficial. \p IC is the 51 /// estimated inline cost associated with callsite \p CB. 52 /// \p TotalSecondaryCost will be set to the estimated cost of inlining the 53 /// caller if \p CB is suppressed for inlining. 54 static bool 55 shouldBeDeferred(Function *Caller, InlineCost IC, int &TotalSecondaryCost, 56 function_ref<InlineCost(CallBase &CB)> GetInlineCost) { 57 // For now we only handle local or inline functions. 58 if (!Caller->hasLocalLinkage() && !Caller->hasLinkOnceODRLinkage()) 59 return false; 60 // If the cost of inlining CB is non-positive, it is not going to prevent the 61 // caller from being inlined into its callers and hence we don't need to 62 // defer. 63 if (IC.getCost() <= 0) 64 return false; 65 // Try to detect the case where the current inlining candidate caller (call 66 // it B) is a static or linkonce-ODR function and is an inlining candidate 67 // elsewhere, and the current candidate callee (call it C) is large enough 68 // that inlining it into B would make B too big to inline later. In these 69 // circumstances it may be best not to inline C into B, but to inline B into 70 // its callers. 71 // 72 // This only applies to static and linkonce-ODR functions because those are 73 // expected to be available for inlining in the translation units where they 74 // are used. Thus we will always have the opportunity to make local inlining 75 // decisions. Importantly the linkonce-ODR linkage covers inline functions 76 // and templates in C++. 77 // 78 // FIXME: All of this logic should be sunk into getInlineCost. It relies on 79 // the internal implementation of the inline cost metrics rather than 80 // treating them as truly abstract units etc. 81 TotalSecondaryCost = 0; 82 // The candidate cost to be imposed upon the current function. 83 int CandidateCost = IC.getCost() - 1; 84 // If the caller has local linkage and can be inlined to all its callers, we 85 // can apply a huge negative bonus to TotalSecondaryCost. 86 bool ApplyLastCallBonus = Caller->hasLocalLinkage() && !Caller->hasOneUse(); 87 // This bool tracks what happens if we DO inline C into B. 88 bool InliningPreventsSomeOuterInline = false; 89 unsigned NumCallerUsers = 0; 90 for (User *U : Caller->users()) { 91 CallBase *CS2 = dyn_cast<CallBase>(U); 92 93 // If this isn't a call to Caller (it could be some other sort 94 // of reference) skip it. Such references will prevent the caller 95 // from being removed. 96 if (!CS2 || CS2->getCalledFunction() != Caller) { 97 ApplyLastCallBonus = false; 98 continue; 99 } 100 101 InlineCost IC2 = GetInlineCost(*CS2); 102 ++NumCallerCallersAnalyzed; 103 if (!IC2) { 104 ApplyLastCallBonus = false; 105 continue; 106 } 107 if (IC2.isAlways()) 108 continue; 109 110 // See if inlining of the original callsite would erase the cost delta of 111 // this callsite. We subtract off the penalty for the call instruction, 112 // which we would be deleting. 113 if (IC2.getCostDelta() <= CandidateCost) { 114 InliningPreventsSomeOuterInline = true; 115 TotalSecondaryCost += IC2.getCost(); 116 NumCallerUsers++; 117 } 118 } 119 120 if (!InliningPreventsSomeOuterInline) 121 return false; 122 123 // If all outer calls to Caller would get inlined, the cost for the last 124 // one is set very low by getInlineCost, in anticipation that Caller will 125 // be removed entirely. We did not account for this above unless there 126 // is only one caller of Caller. 127 if (ApplyLastCallBonus) 128 TotalSecondaryCost -= InlineConstants::LastCallToStaticBonus; 129 130 // If InlineDeferralScale is negative, then ignore the cost of primary 131 // inlining -- IC.getCost() multiplied by the number of callers to Caller. 132 if (InlineDeferralScale < 0) 133 return TotalSecondaryCost < IC.getCost(); 134 135 int TotalCost = TotalSecondaryCost + IC.getCost() * NumCallerUsers; 136 int Allowance = IC.getCost() * InlineDeferralScale; 137 return TotalCost < Allowance; 138 } 139 140 namespace llvm { 141 static std::basic_ostream<char> &operator<<(std::basic_ostream<char> &R, 142 const ore::NV &Arg) { 143 return R << Arg.Val; 144 } 145 146 template <class RemarkT> 147 RemarkT &operator<<(RemarkT &&R, const InlineCost &IC) { 148 using namespace ore; 149 if (IC.isAlways()) { 150 R << "(cost=always)"; 151 } else if (IC.isNever()) { 152 R << "(cost=never)"; 153 } else { 154 R << "(cost=" << ore::NV("Cost", IC.getCost()) 155 << ", threshold=" << ore::NV("Threshold", IC.getThreshold()) << ")"; 156 } 157 if (const char *Reason = IC.getReason()) 158 R << ": " << ore::NV("Reason", Reason); 159 return R; 160 } 161 } // namespace llvm 162 163 std::string llvm::inlineCostStr(const InlineCost &IC) { 164 std::stringstream Remark; 165 Remark << IC; 166 return Remark.str(); 167 } 168 169 void llvm::setInlineRemark(CallBase &CB, StringRef Message) { 170 if (!InlineRemarkAttribute) 171 return; 172 173 Attribute Attr = Attribute::get(CB.getContext(), "inline-remark", Message); 174 CB.addAttribute(AttributeList::FunctionIndex, Attr); 175 } 176 177 /// Return the cost only if the inliner should attempt to inline at the given 178 /// CallSite. If we return the cost, we will emit an optimisation remark later 179 /// using that cost, so we won't do so from this function. Return None if 180 /// inlining should not be attempted. 181 Optional<InlineCost> 182 llvm::shouldInline(CallBase &CB, 183 function_ref<InlineCost(CallBase &CB)> GetInlineCost, 184 OptimizationRemarkEmitter &ORE) { 185 using namespace ore; 186 187 InlineCost IC = GetInlineCost(CB); 188 Instruction *Call = &CB; 189 Function *Callee = CB.getCalledFunction(); 190 Function *Caller = CB.getCaller(); 191 192 if (IC.isAlways()) { 193 LLVM_DEBUG(dbgs() << " Inlining " << inlineCostStr(IC) 194 << ", Call: " << CB << "\n"); 195 return IC; 196 } 197 198 if (!IC) { 199 LLVM_DEBUG(dbgs() << " NOT Inlining " << inlineCostStr(IC) 200 << ", Call: " << CB << "\n"); 201 if (IC.isNever()) { 202 ORE.emit([&]() { 203 return OptimizationRemarkMissed(DEBUG_TYPE, "NeverInline", Call) 204 << NV("Callee", Callee) << " not inlined into " 205 << NV("Caller", Caller) << " because it should never be inlined " 206 << IC; 207 }); 208 } else { 209 ORE.emit([&]() { 210 return OptimizationRemarkMissed(DEBUG_TYPE, "TooCostly", Call) 211 << NV("Callee", Callee) << " not inlined into " 212 << NV("Caller", Caller) << " because too costly to inline " 213 << IC; 214 }); 215 } 216 setInlineRemark(CB, inlineCostStr(IC)); 217 return None; 218 } 219 220 int TotalSecondaryCost = 0; 221 if (shouldBeDeferred(Caller, IC, TotalSecondaryCost, GetInlineCost)) { 222 LLVM_DEBUG(dbgs() << " NOT Inlining: " << CB 223 << " Cost = " << IC.getCost() 224 << ", outer Cost = " << TotalSecondaryCost << '\n'); 225 ORE.emit([&]() { 226 return OptimizationRemarkMissed(DEBUG_TYPE, "IncreaseCostInOtherContexts", 227 Call) 228 << "Not inlining. Cost of inlining " << NV("Callee", Callee) 229 << " increases the cost of inlining " << NV("Caller", Caller) 230 << " in other contexts"; 231 }); 232 setInlineRemark(CB, "deferred"); 233 // IC does not bool() to false, so get an InlineCost that will. 234 // This will not be inspected to make an error message. 235 return None; 236 } 237 238 LLVM_DEBUG(dbgs() << " Inlining " << inlineCostStr(IC) << ", Call: " << CB 239 << '\n'); 240 return IC; 241 } 242 243 void llvm::emitInlinedInto(OptimizationRemarkEmitter &ORE, DebugLoc DLoc, 244 const BasicBlock *Block, const Function &Callee, 245 const Function &Caller, const InlineCost &IC) { 246 ORE.emit([&]() { 247 bool AlwaysInline = IC.isAlways(); 248 StringRef RemarkName = AlwaysInline ? "AlwaysInline" : "Inlined"; 249 return OptimizationRemark(DEBUG_TYPE, RemarkName, DLoc, Block) 250 << ore::NV("Callee", &Callee) << " inlined into " 251 << ore::NV("Caller", &Caller) << " with " << IC; 252 }); 253 } 254