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/Debug.h"
  26 using namespace llvm;
  27
  28 namespace {
  29   /// This pass optimizes well half_powr function calls.
  30   ///
  31   class SimplifyHalfPowrLibCalls : public FunctionPass {
  32     const TargetData *TD;
  33   public:
  34     static char ID; // Pass identification
  35     SimplifyHalfPowrLibCalls() : FunctionPass(ID) {
  36       initializeSimplifyHalfPowrLibCallsPass(*PassRegistry::getPassRegistry());
  37     }
  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 INITIALIZE_PASS(SimplifyHalfPowrLibCalls, "simplify-libcalls-halfpowr",
  52                 "Simplify half_powr library calls", false, false)
  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::
  63 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 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 = llvm::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     InlineFunctionInfo IFI(0, TD);
  99     bool B = InlineFunction(Call, IFI);
 100     assert(B && "half_powr didn't inline?");
 101     (void)B;
 102
 103     BasicBlock *NewBody = NewBlock->getSinglePredecessor();
 104     assert(NewBody);
 105     Bodies.push_back(NewBody);
 106   }
 107
 108   if (!NewBlock)
 109     return InsertPt;
 110
 111   // Put the code for all the bodies into one block, to facilitate
 112   // subsequent optimization.
 113   (void)SplitEdge(NewBlock->getSinglePredecessor(), NewBlock, this);
 114   for (unsigned i = 0, e = Bodies.size(); i != e; ++i) {
 115     BasicBlock *Body = Bodies[i];
 116     Instruction *FNP = Body->getFirstNonPHI();
 117     // Splice the insts from body into NewBlock.
 118     NewBlock->getInstList().splice(NewBlock->begin(), Body->getInstList(),
 119                                    FNP, Body->getTerminator());
 120   }
 121
 122   return NewBlock->begin();
 123 }
 124
 125 /// runOnFunction - Top level algorithm.
 126 ///
 127 bool SimplifyHalfPowrLibCalls::runOnFunction(Function &F) {
 128   TD = getAnalysisIfAvailable<TargetData>();
 129
 130   bool Changed = false;
 131   std::vector<Instruction *> HalfPowrs;
 132   for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
 133     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
 134       // Look for calls.
 135       bool IsHalfPowr = false;
 136       if (CallInst *CI = dyn_cast<CallInst>(I)) {
 137         // Look for direct calls and calls to non-external functions.
 138         Function *Callee = CI->getCalledFunction();
 139         if (Callee && Callee->hasExternalLinkage()) {
 140           // Look for calls with well-known names.
 141           if (Callee->getName() == "__half_powrf4")
 142             IsHalfPowr = true;
 143         }
 144       }
 145       if (IsHalfPowr)
 146         HalfPowrs.push_back(I);
 147       // We're looking for sequences of up to three such calls, which we'll
 148       // simplify as a group.
 149       if ((!IsHalfPowr && !HalfPowrs.empty()) || HalfPowrs.size() == 3) {
 150         I = InlineHalfPowrs(HalfPowrs, I);
 151         E = I->getParent()->end();
 152         HalfPowrs.clear();
 153         Changed = true;
 154       }
 155     }
 156     assert(HalfPowrs.empty() && "Block had no terminator!");
 157   }
 158
 159   return Changed;
 160 }