lib/Transforms/Scalar/SimplifyHalfPowrLibCalls.cpp

   1 //===- SimplifyHalfPowrLibCalls.cpp - Optimize specific half_powr calls ---===//
   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 a simple pass that applies an experimental
  11 // transformation on calls to specific functions.
  12 //
  13 //===----------------------------------------------------------------------===//
  14
  15 #define DEBUG_TYPE "simplify-libcalls-halfpowr"
  16 #include "llvm/Transforms/Scalar.h"
  17 #include "llvm/Instructions.h"
  18 #include "llvm/Intrinsics.h"
  19 #include "llvm/Module.h"
  20 #include "llvm/Pass.h"
  21 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
  22 #include "llvm/Transforms/Utils/Cloning.h"
  23 #include "llvm/Target/TargetData.h"
  24 #include "llvm/ADT/STLExtras.h"
  25 #include "llvm/Support/Compiler.h"
  26 #include "llvm/Support/Debug.h"
  27 #include "llvm/Config/config.h"
  28 using namespace llvm;
  29
  30 namespace {
  31   /// This pass optimizes well half_powr function calls.
  32   ///
  33   class VISIBILITY_HIDDEN SimplifyHalfPowrLibCalls : public FunctionPass {
  34     const TargetData *TD;
  35   public:
  36     static char ID; // Pass identification
  37     SimplifyHalfPowrLibCalls() : FunctionPass(&ID) {}
  38
  39     bool runOnFunction(Function &F);
  40
  41     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  42     }
  43
  44     Instruction *
  45     InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs,
  46                     Instruction *InsertPt);
  47   };
  48   char SimplifyHalfPowrLibCalls::ID = 0;
  49 } // end anonymous namespace.
  50
  51 static RegisterPass<SimplifyHalfPowrLibCalls>
  52 X("simplify-libcalls-halfpowr", "Simplify half_powr library calls");
  53
  54 // Public interface to the Simplify HalfPowr LibCalls pass.
  55 FunctionPass *llvm::createSimplifyHalfPowrLibCallsPass() {
  56   return new SimplifyHalfPowrLibCalls();
  57 }
  58
  59 /// InlineHalfPowrs - Inline a sequence of adjacent half_powr calls, rearranging
  60 /// their control flow to better facilitate subsequent optimization.
  61 Instruction *
  62 SimplifyHalfPowrLibCalls::InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs,
  63                                         Instruction *InsertPt) {
  64   std::vector<BasicBlock *> Bodies;
  65   BasicBlock *NewBlock = 0;
  66
  67   for (unsigned i = 0, e = HalfPowrs.size(); i != e; ++i) {
  68     CallInst *Call = cast<CallInst>(HalfPowrs[i]);
  69     Function *Callee = Call->getCalledFunction();
  70
  71     // Minimally sanity-check the CFG of half_powr to ensure that it contains
  72     // the the kind of code we expect.  If we're running this pass, we have
  73     // reason to believe it will be what we expect.
  74     Function::iterator I = Callee->begin();
  75     BasicBlock *Prologue = I++;
  76     if (I == Callee->end()) break;
  77     BasicBlock *SubnormalHandling = I++;
  78     if (I == Callee->end()) break;
  79     BasicBlock *Body = I++;
  80     if (I != Callee->end()) break;
  81     if (SubnormalHandling->getSinglePredecessor() != Prologue)
  82       break;
  83     BranchInst *PBI = dyn_cast<BranchInst>(Prologue->getTerminator());
  84     if (!PBI || !PBI->isConditional())
  85       break;
  86     BranchInst *SNBI = dyn_cast<BranchInst>(SubnormalHandling->getTerminator());
  87     if (!SNBI || SNBI->isConditional())
  88       break;
  89     if (!isa<ReturnInst>(Body->getTerminator()))
  90       break;
  91
  92     Instruction *NextInst = next(BasicBlock::iterator(Call));
  93
  94     // Inline the call, taking care of what code ends up where.
  95     NewBlock = SplitBlock(NextInst->getParent(), NextInst, this);
  96
  97     bool B = InlineFunction(Call, 0, TD);
  98     assert(B && "half_powr didn't inline?"); B=B;
  99
 100     BasicBlock *NewBody = NewBlock->getSinglePredecessor();
 101     assert(NewBody);
 102     Bodies.push_back(NewBody);
 103   }
 104
 105   if (!NewBlock)
 106     return InsertPt;
 107
 108   // Put the code for all the bodies into one block, to facilitate
 109   // subsequent optimization.
 110   (void)SplitEdge(NewBlock->getSinglePredecessor(), NewBlock, this);
 111   for (unsigned i = 0, e = Bodies.size(); i != e; ++i) {
 112     BasicBlock *Body = Bodies[i];
 113     Instruction *FNP = Body->getFirstNonPHI();
 114     // Splice the insts from body into NewBlock.
 115     NewBlock->getInstList().splice(NewBlock->begin(), Body->getInstList(),
 116                                    FNP, Body->getTerminator());
 117   }
 118
 119   return NewBlock->begin();
 120 }
 121
 122 /// runOnFunction - Top level algorithm.
 123 ///
 124 bool SimplifyHalfPowrLibCalls::runOnFunction(Function &F) {
 125   TD = getAnalysisIfAvailable<TargetData>();
 126
 127   bool Changed = false;
 128   std::vector<Instruction *> HalfPowrs;
 129   for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
 130     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
 131       // Look for calls.
 132       bool IsHalfPowr = false;
 133       if (CallInst *CI = dyn_cast<CallInst>(I)) {
 134         // Look for direct calls and calls to non-external functions.
 135         Function *Callee = CI->getCalledFunction();
 136         if (Callee && Callee->hasExternalLinkage()) {
 137           // Look for calls with well-known names.
 138           const char *CalleeName = Callee->getNameStart();
 139           if (strcmp(CalleeName, "__half_powrf4") == 0)
 140             IsHalfPowr = true;
 141         }
 142       }
 143       if (IsHalfPowr)
 144         HalfPowrs.push_back(I);
 145       // We're looking for sequences of up to three such calls, which we'll
 146       // simplify as a group.
 147       if ((!IsHalfPowr && !HalfPowrs.empty()) || HalfPowrs.size() == 3) {
 148         I = InlineHalfPowrs(HalfPowrs, I);
 149         E = I->getParent()->end();
 150         HalfPowrs.clear();
 151         Changed = true;
 152       }
 153     }
 154     assert(HalfPowrs.empty() && "Block had no terminator!");
 155   }
 156
 157   return Changed;
 158 }