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       AU.addRequired<TargetData>();
  43     }
  44
  45     Instruction *
  46     InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs,
  47                     Instruction *InsertPt);
  48   };
  49   char SimplifyHalfPowrLibCalls::ID = 0;
  50 } // end anonymous namespace.
  51
  52 static RegisterPass<SimplifyHalfPowrLibCalls>
  53 X("simplify-libcalls-halfpowr", "Simplify half_powr library calls");
  54
  55 // Public interface to the Simplify HalfPowr LibCalls pass.
  56 FunctionPass *llvm::createSimplifyHalfPowrLibCallsPass() {
  57   return new SimplifyHalfPowrLibCalls();
  58 }
  59
  60 /// InlineHalfPowrs - Inline a sequence of adjacent half_powr calls, rearranging
  61 /// their control flow to better facilitate subsequent optimization.
  62 Instruction *
  63 SimplifyHalfPowrLibCalls::InlineHalfPowrs(const std::vector<Instruction *> &HalfPowrs,
  64                                         Instruction *InsertPt) {
  65   std::vector<BasicBlock *> Bodies;
  66   BasicBlock *NewBlock = 0;
  67
  68   for (unsigned i = 0, e = HalfPowrs.size(); i != e; ++i) {
  69     CallInst *Call = cast<CallInst>(HalfPowrs[i]);
  70     Function *Callee = Call->getCalledFunction();
  71
  72     // Minimally sanity-check the CFG of half_powr to ensure that it contains
  73     // the the kind of code we expect.  If we're running this pass, we have
  74     // reason to believe it will be what we expect.
  75     Function::iterator I = Callee->begin();
  76     BasicBlock *Prologue = I++;
  77     if (I == Callee->end()) break;
  78     BasicBlock *SubnormalHandling = I++;
  79     if (I == Callee->end()) break;
  80     BasicBlock *Body = I++;
  81     if (I != Callee->end()) break;
  82     if (SubnormalHandling->getSinglePredecessor() != Prologue)
  83       break;
  84     BranchInst *PBI = dyn_cast<BranchInst>(Prologue->getTerminator());
  85     if (!PBI || !PBI->isConditional())
  86       break;
  87     BranchInst *SNBI = dyn_cast<BranchInst>(SubnormalHandling->getTerminator());
  88     if (!SNBI || SNBI->isConditional())
  89       break;
  90     if (!isa<ReturnInst>(Body->getTerminator()))
  91       break;
  92
  93     Instruction *NextInst = next(BasicBlock::iterator(Call));
  94
  95     // Inline the call, taking care of what code ends up where.
  96     NewBlock = SplitBlock(NextInst->getParent(), NextInst, this);
  97
  98     bool B = InlineFunction(Call, 0, TD);
  99     assert(B && "half_powr didn't inline?"); B=B;
 100
 101     BasicBlock *NewBody = NewBlock->getSinglePredecessor();
 102     assert(NewBody);
 103     Bodies.push_back(NewBody);
 104   }
 105
 106   if (!NewBlock)
 107     return InsertPt;
 108
 109   // Put the code for all the bodies into one block, to facilitate
 110   // subsequent optimization.
 111   (void)SplitEdge(NewBlock->getSinglePredecessor(), NewBlock, this);
 112   for (unsigned i = 0, e = Bodies.size(); i != e; ++i) {
 113     BasicBlock *Body = Bodies[i];
 114     Instruction *FNP = Body->getFirstNonPHI();
 115     // Splice the insts from body into NewBlock.
 116     NewBlock->getInstList().splice(NewBlock->begin(), Body->getInstList(),
 117                                    FNP, Body->getTerminator());
 118   }
 119
 120   return NewBlock->begin();
 121 }
 122
 123 /// runOnFunction - Top level algorithm.
 124 ///
 125 bool SimplifyHalfPowrLibCalls::runOnFunction(Function &F) {
 126   TD = &getAnalysis<TargetData>();
 127
 128   bool Changed = false;
 129   std::vector<Instruction *> HalfPowrs;
 130   for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
 131     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
 132       // Look for calls.
 133       bool IsHalfPowr = false;
 134       if (CallInst *CI = dyn_cast<CallInst>(I)) {
 135         // Look for direct calls and calls to non-external functions.
 136         Function *Callee = CI->getCalledFunction();
 137         if (Callee && Callee->hasExternalLinkage()) {
 138           // Look for calls with well-known names.
 139           const char *CalleeName = Callee->getNameStart();
 140           if (strcmp(CalleeName, "__half_powrf4") == 0)
 141             IsHalfPowr = true;
 142         }
 143       }
 144       if (IsHalfPowr)
 145         HalfPowrs.push_back(I);
 146       // We're looking for sequences of up to three such calls, which we'll
 147       // simplify as a group.
 148       if ((!IsHalfPowr && !HalfPowrs.empty()) || HalfPowrs.size() == 3) {
 149         I = InlineHalfPowrs(HalfPowrs, I);
 150         E = I->getParent()->end();
 151         HalfPowrs.clear();
 152         Changed = true;
 153       }
 154     }
 155     assert(HalfPowrs.empty() && "Block had no terminator!");
 156   }
 157
 158   return Changed;
 159 }