lib/Transforms/Scalar/PartiallyInlineLibCalls.cpp

   1 //===--- PartiallyInlineLibCalls.cpp - Partially inline libcalls ----------===//
   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 pass tries to partially inline the fast path of well-known library
  11 // functions, such as using square-root instructions for cases where sqrt()
  12 // does not need to set errno.
  13 //
  14 //===----------------------------------------------------------------------===//
  15
  16 #include "llvm/Analysis/TargetTransformInfo.h"
  17 #include "llvm/IR/IRBuilder.h"
  18 #include "llvm/IR/Intrinsics.h"
  19 #include "llvm/Pass.h"
  20 #include "llvm/Support/CommandLine.h"
  21 #include "llvm/Target/TargetLibraryInfo.h"
  22 #include "llvm/Transforms/Scalar.h"
  23 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
  24
  25 using namespace llvm;
  26
  27 #define DEBUG_TYPE "partially-inline-libcalls"
  28
  29 namespace {
  30   class PartiallyInlineLibCalls : public FunctionPass {
  31   public:
  32     static char ID;
  33
  34     PartiallyInlineLibCalls() :
  35       FunctionPass(ID) {
  36       initializePartiallyInlineLibCallsPass(*PassRegistry::getPassRegistry());
  37     }
  38
  39     void getAnalysisUsage(AnalysisUsage &AU) const override;
  40     bool runOnFunction(Function &F) override;
  41
  42   private:
  43     /// Optimize calls to sqrt.
  44     bool optimizeSQRT(CallInst *Call, Function *CalledFunc,
  45                       BasicBlock &CurrBB, Function::iterator &BB);
  46   };
  47
  48   char PartiallyInlineLibCalls::ID = 0;
  49 }
  50
  51 INITIALIZE_PASS(PartiallyInlineLibCalls, "partially-inline-libcalls",
  52                 "Partially inline calls to library functions", false, false)
  53
  54 void PartiallyInlineLibCalls::getAnalysisUsage(AnalysisUsage &AU) const {
  55   AU.addRequired<TargetLibraryInfo>();
  56   AU.addRequired<TargetTransformInfo>();
  57   FunctionPass::getAnalysisUsage(AU);
  58 }
  59
  60 bool PartiallyInlineLibCalls::runOnFunction(Function &F) {
  61   bool Changed = false;
  62   Function::iterator CurrBB;
  63   TargetLibraryInfo *TLI = &getAnalysis<TargetLibraryInfo>();
  64   const TargetTransformInfo *TTI = &getAnalysis<TargetTransformInfo>();
  65   for (Function::iterator BB = F.begin(), BE = F.end(); BB != BE;) {
  66     CurrBB = BB++;
  67
  68     for (BasicBlock::iterator II = CurrBB->begin(), IE = CurrBB->end();
  69          II != IE; ++II) {
  70       CallInst *Call = dyn_cast<CallInst>(&*II);
  71       Function *CalledFunc;
  72
  73       if (!Call || !(CalledFunc = Call->getCalledFunction()))
  74         continue;
  75
  76       // Skip if function either has local linkage or is not a known library
  77       // function.
  78       LibFunc::Func LibFunc;
  79       if (CalledFunc->hasLocalLinkage() || !CalledFunc->hasName() ||
  80           !TLI->getLibFunc(CalledFunc->getName(), LibFunc))
  81         continue;
  82
  83       switch (LibFunc) {
  84       case LibFunc::sqrtf:
  85       case LibFunc::sqrt:
  86         if (TTI->haveFastSqrt(Call->getType()) &&
  87             optimizeSQRT(Call, CalledFunc, *CurrBB, BB))
  88           break;
  89         continue;
  90       default:
  91         continue;
  92       }
  93
  94       Changed = true;
  95       break;
  96     }
  97   }
  98
  99   return Changed;
 100 }
 101
 102 bool PartiallyInlineLibCalls::optimizeSQRT(CallInst *Call,
 103                                            Function *CalledFunc,
 104                                            BasicBlock &CurrBB,
 105                                            Function::iterator &BB) {
 106   // There is no need to change the IR, since backend will emit sqrt
 107   // instruction if the call has already been marked read-only.
 108   if (Call->onlyReadsMemory())
 109     return false;
 110
 111   // The call must have the expected result type.
 112   if (!Call->getType()->isFloatingPointTy())
 113     return false;
 114
 115   // Do the following transformation:
 116   //
 117   // (before)
 118   // dst = sqrt(src)
 119   //
 120   // (after)
 121   // v0 = sqrt_noreadmem(src) # native sqrt instruction.
 122   // if (v0 is a NaN)
 123   //   v1 = sqrt(src)         # library call.
 124   // dst = phi(v0, v1)
 125   //
 126
 127   // Move all instructions following Call to newly created block JoinBB.
 128   // Create phi and replace all uses.
 129   BasicBlock *JoinBB = llvm::SplitBlock(&CurrBB, Call->getNextNode(), this);
 130   IRBuilder<> Builder(JoinBB, JoinBB->begin());
 131   PHINode *Phi = Builder.CreatePHI(Call->getType(), 2);
 132   Call->replaceAllUsesWith(Phi);
 133
 134   // Create basic block LibCallBB and insert a call to library function sqrt.
 135   BasicBlock *LibCallBB = BasicBlock::Create(CurrBB.getContext(), "call.sqrt",
 136                                              CurrBB.getParent(), JoinBB);
 137   Builder.SetInsertPoint(LibCallBB);
 138   Instruction *LibCall = Call->clone();
 139   Builder.Insert(LibCall);
 140   Builder.CreateBr(JoinBB);
 141
 142   // Add attribute "readnone" so that backend can use a native sqrt instruction
 143   // for this call. Insert a FP compare instruction and a conditional branch
 144   // at the end of CurrBB.
 145   Call->addAttribute(AttributeSet::FunctionIndex, Attribute::ReadNone);
 146   CurrBB.getTerminator()->eraseFromParent();
 147   Builder.SetInsertPoint(&CurrBB);
 148   Value *FCmp = Builder.CreateFCmpOEQ(Call, Call);
 149   Builder.CreateCondBr(FCmp, JoinBB, LibCallBB);
 150
 151   // Add phi operands.
 152   Phi->addIncoming(Call, &CurrBB);
 153   Phi->addIncoming(LibCall, LibCallBB);
 154
 155   BB = JoinBB;
 156   return true;
 157 }
 158
 159 FunctionPass *llvm::createPartiallyInlineLibCallsPass() {
 160   return new PartiallyInlineLibCalls();
 161 }