CodeMetrics.cpp revision 831737d329a727f53a1fb0572f7b7a8127208881
1//===- CodeMetrics.cpp - Code cost measurements ---------------------------===// 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 code cost measurement utilities. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/Analysis/CodeMetrics.h" 15#include "llvm/DataLayout.h" 16#include "llvm/Function.h" 17#include "llvm/IntrinsicInst.h" 18#include "llvm/Support/CallSite.h" 19 20using namespace llvm; 21 22/// callIsSmall - If a call is likely to lower to a single target instruction, 23/// or is otherwise deemed small return true. 24/// TODO: Perhaps calls like memcpy, strcpy, etc? 25bool llvm::callIsSmall(ImmutableCallSite CS) { 26 if (isa<IntrinsicInst>(CS.getInstruction())) 27 return true; 28 29 const Function *F = CS.getCalledFunction(); 30 if (!F) return false; 31 32 if (F->hasLocalLinkage()) return false; 33 34 if (!F->hasName()) return false; 35 36 StringRef Name = F->getName(); 37 38 // These will all likely lower to a single selection DAG node. 39 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" || 40 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || 41 Name == "sin" || Name == "sinf" || Name == "sinl" || 42 Name == "cos" || Name == "cosf" || Name == "cosl" || 43 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" ) 44 return true; 45 46 // These are all likely to be optimized into something smaller. 47 if (Name == "pow" || Name == "powf" || Name == "powl" || 48 Name == "exp2" || Name == "exp2l" || Name == "exp2f" || 49 Name == "floor" || Name == "floorf" || Name == "ceil" || 50 Name == "round" || Name == "ffs" || Name == "ffsl" || 51 Name == "abs" || Name == "labs" || Name == "llabs") 52 return true; 53 54 return false; 55} 56 57bool llvm::isInstructionFree(const Instruction *I, const DataLayout *TD) { 58 if (isa<PHINode>(I)) 59 return true; 60 61 // If a GEP has all constant indices, it will probably be folded with 62 // a load/store. 63 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) 64 return GEP->hasAllConstantIndices(); 65 66 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 67 switch (II->getIntrinsicID()) { 68 default: 69 return false; 70 case Intrinsic::dbg_declare: 71 case Intrinsic::dbg_value: 72 case Intrinsic::invariant_start: 73 case Intrinsic::invariant_end: 74 case Intrinsic::lifetime_start: 75 case Intrinsic::lifetime_end: 76 case Intrinsic::objectsize: 77 case Intrinsic::ptr_annotation: 78 case Intrinsic::var_annotation: 79 // These intrinsics don't count as size. 80 return true; 81 } 82 } 83 84 if (const CastInst *CI = dyn_cast<CastInst>(I)) { 85 // Noop casts, including ptr <-> int, don't count. 86 if (CI->isLosslessCast()) 87 return true; 88 89 Value *Op = CI->getOperand(0); 90 // An inttoptr cast is free so long as the input is a legal integer type 91 // which doesn't contain values outside the range of a pointer. 92 if (isa<IntToPtrInst>(CI) && TD && 93 TD->isLegalInteger(Op->getType()->getScalarSizeInBits()) && 94 Op->getType()->getScalarSizeInBits() <= TD->getPointerSizeInBits()) 95 return true; 96 97 // A ptrtoint cast is free so long as the result is large enough to store 98 // the pointer, and a legal integer type. 99 if (isa<PtrToIntInst>(CI) && TD && 100 TD->isLegalInteger(Op->getType()->getScalarSizeInBits()) && 101 Op->getType()->getScalarSizeInBits() >= TD->getPointerSizeInBits()) 102 return true; 103 104 // trunc to a native type is free (assuming the target has compare and 105 // shift-right of the same width). 106 if (TD && isa<TruncInst>(CI) && 107 TD->isLegalInteger(TD->getTypeSizeInBits(CI->getType()))) 108 return true; 109 // Result of a cmp instruction is often extended (to be used by other 110 // cmp instructions, logical or return instructions). These are usually 111 // nop on most sane targets. 112 if (isa<CmpInst>(CI->getOperand(0))) 113 return true; 114 } 115 116 return false; 117} 118 119/// analyzeBasicBlock - Fill in the current structure with information gleaned 120/// from the specified block. 121void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB, 122 const DataLayout *TD) { 123 ++NumBlocks; 124 unsigned NumInstsBeforeThisBB = NumInsts; 125 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end(); 126 II != E; ++II) { 127 if (isInstructionFree(II, TD)) 128 continue; 129 130 // Special handling for calls. 131 if (isa<CallInst>(II) || isa<InvokeInst>(II)) { 132 ImmutableCallSite CS(cast<Instruction>(II)); 133 134 if (const Function *F = CS.getCalledFunction()) { 135 // If a function is both internal and has a single use, then it is 136 // extremely likely to get inlined in the future (it was probably 137 // exposed by an interleaved devirtualization pass). 138 if (!CS.isNoInline() && F->hasInternalLinkage() && F->hasOneUse()) 139 ++NumInlineCandidates; 140 141 // If this call is to function itself, then the function is recursive. 142 // Inlining it into other functions is a bad idea, because this is 143 // basically just a form of loop peeling, and our metrics aren't useful 144 // for that case. 145 if (F == BB->getParent()) 146 isRecursive = true; 147 } 148 149 if (!callIsSmall(CS)) { 150 // Each argument to a call takes on average one instruction to set up. 151 NumInsts += CS.arg_size(); 152 153 // We don't want inline asm to count as a call - that would prevent loop 154 // unrolling. The argument setup cost is still real, though. 155 if (!isa<InlineAsm>(CS.getCalledValue())) 156 ++NumCalls; 157 } 158 } 159 160 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) { 161 if (!AI->isStaticAlloca()) 162 this->usesDynamicAlloca = true; 163 } 164 165 if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy()) 166 ++NumVectorInsts; 167 168 if (const CallInst *CI = dyn_cast<CallInst>(II)) 169 if (CI->hasFnAttr(Attribute::NoDuplicate)) 170 notDuplicatable = true; 171 172 if (const InvokeInst *InvI = dyn_cast<InvokeInst>(II)) 173 if (InvI->hasFnAttr(Attribute::NoDuplicate)) 174 notDuplicatable = true; 175 176 ++NumInsts; 177 } 178 179 if (isa<ReturnInst>(BB->getTerminator())) 180 ++NumRets; 181 182 // We never want to inline functions that contain an indirectbr. This is 183 // incorrect because all the blockaddress's (in static global initializers 184 // for example) would be referring to the original function, and this indirect 185 // jump would jump from the inlined copy of the function into the original 186 // function which is extremely undefined behavior. 187 // FIXME: This logic isn't really right; we can safely inline functions 188 // with indirectbr's as long as no other function or global references the 189 // blockaddress of a block within the current function. And as a QOI issue, 190 // if someone is using a blockaddress without an indirectbr, and that 191 // reference somehow ends up in another function or global, we probably 192 // don't want to inline this function. 193 notDuplicatable |= isa<IndirectBrInst>(BB->getTerminator()); 194 195 // Remember NumInsts for this BB. 196 NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB; 197} 198 199void CodeMetrics::analyzeFunction(Function *F, const DataLayout *TD) { 200 // If this function contains a call that "returns twice" (e.g., setjmp or 201 // _setjmp) and it isn't marked with "returns twice" itself, never inline it. 202 // This is a hack because we depend on the user marking their local variables 203 // as volatile if they are live across a setjmp call, and they probably 204 // won't do this in callers. 205 exposesReturnsTwice = F->callsFunctionThatReturnsTwice() && 206 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, 207 Attribute::ReturnsTwice); 208 209 // Look at the size of the callee. 210 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 211 analyzeBasicBlock(&*BB, TD); 212} 213