lib/Transforms/Scalar/LoopUnroll.cpp

   1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file was developed by the LLVM research group and is distributed under
   6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This pass implements a simple loop unroller.  It works best when loops have
  11 // been canonicalized by the -indvars pass, allowing it to determine the trip
  12 // counts of loops easily.
  13 //
  14 // This pass is currently extremely limited.  It only currently only unrolls
  15 // single basic block loops that execute a constant number of times.
  16 //
  17 //===----------------------------------------------------------------------===//
  18
  19 #define DEBUG_TYPE "loop-unroll"
  20 #include "llvm/Transforms/Scalar.h"
  21 #include "llvm/Constants.h"
  22 #include "llvm/Function.h"
  23 #include "llvm/Instructions.h"
  24 #include "llvm/Analysis/LoopInfo.h"
  25 #include "llvm/Transforms/Utils/Cloning.h"
  26 #include "llvm/Transforms/Utils/Local.h"
  27 #include "llvm/Support/CommandLine.h"
  28 #include "llvm/Support/Debug.h"
  29 #include "llvm/ADT/Statistic.h"
  30 #include "llvm/ADT/STLExtras.h"
  31 #include "llvm/IntrinsicInst.h"
  32 #include <cstdio>
  33 #include <set>
  34 #include <algorithm>
  35 #include <iostream>
  36 using namespace llvm;
  37
  38 namespace {
  39   Statistic<> NumUnrolled("loop-unroll", "Number of loops completely unrolled");
  40
  41   cl::opt<unsigned>
  42   UnrollThreshold("unroll-threshold", cl::init(100), cl::Hidden,
  43                   cl::desc("The cut-off point for loop unrolling"));
  44
  45   class LoopUnroll : public FunctionPass {
  46     LoopInfo *LI;  // The current loop information
  47   public:
  48     virtual bool runOnFunction(Function &F);
  49     bool visitLoop(Loop *L);
  50
  51     /// This transformation requires natural loop information & requires that
  52     /// loop preheaders be inserted into the CFG...
  53     ///
  54     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  55       AU.addRequiredID(LoopSimplifyID);
  56       AU.addRequired<LoopInfo>();
  57       AU.addPreserved<LoopInfo>();
  58     }
  59   };
  60   RegisterOpt<LoopUnroll> X("loop-unroll", "Unroll loops");
  61 }
  62
  63 FunctionPass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
  64
  65 bool LoopUnroll::runOnFunction(Function &F) {
  66   bool Changed = false;
  67   LI = &getAnalysis<LoopInfo>();
  68
  69   // Transform all the top-level loops.  Copy the loop list so that the child
  70   // can update the loop tree if it needs to delete the loop.
  71   std::vector<Loop*> SubLoops(LI->begin(), LI->end());
  72   for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
  73     Changed |= visitLoop(SubLoops[i]);
  74
  75   return Changed;
  76 }
  77
  78 /// ApproximateLoopSize - Approximate the size of the loop after it has been
  79 /// unrolled.
  80 static unsigned ApproximateLoopSize(const Loop *L) {
  81   unsigned Size = 0;
  82   for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
  83     BasicBlock *BB = L->getBlocks()[i];
  84     Instruction *Term = BB->getTerminator();
  85     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
  86       if (isa<PHINode>(I) && BB == L->getHeader()) {
  87         // Ignore PHI nodes in the header.
  88       } else if (I->hasOneUse() && I->use_back() == Term) {
  89         // Ignore instructions only used by the loop terminator.
  90       } else if (DbgInfoIntrinsic *DbgI = dyn_cast<DbgInfoIntrinsic>(I)) {
  91         // Ignore debug instructions
  92       } else {
  93         ++Size;
  94       }
  95
  96       // TODO: Ignore expressions derived from PHI and constants if inval of phi
  97       // is a constant, or if operation is associative.  This will get induction
  98       // variables.
  99     }
 100   }
 101
 102   return Size;
 103 }
 104
 105 // RemapInstruction - Convert the instruction operands from referencing the
 106 // current values into those specified by ValueMap.
 107 //
 108 static inline void RemapInstruction(Instruction *I,
 109                                     std::map<const Value *, Value*> &ValueMap) {
 110   for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
 111     Value *Op = I->getOperand(op);
 112     std::map<const Value *, Value*>::iterator It = ValueMap.find(Op);
 113     if (It != ValueMap.end()) Op = It->second;
 114     I->setOperand(op, Op);
 115   }
 116 }
 117
 118 bool LoopUnroll::visitLoop(Loop *L) {
 119   bool Changed = false;
 120
 121   // Recurse through all subloops before we process this loop.  Copy the loop
 122   // list so that the child can update the loop tree if it needs to delete the
 123   // loop.
 124   std::vector<Loop*> SubLoops(L->begin(), L->end());
 125   for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
 126     Changed |= visitLoop(SubLoops[i]);
 127
 128   // We only handle single basic block loops right now.
 129   if (L->getBlocks().size() != 1)
 130     return Changed;
 131
 132   BasicBlock *BB = L->getHeader();
 133   BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
 134   if (BI == 0) return Changed;  // Must end in a conditional branch
 135
 136   ConstantInt *TripCountC = dyn_cast_or_null<ConstantInt>(L->getTripCount());
 137   if (!TripCountC) return Changed;  // Must have constant trip count!
 138
 139   uint64_t TripCountFull = TripCountC->getRawValue();
 140   if (TripCountFull != TripCountC->getRawValue() || TripCountFull == 0)
 141     return Changed; // More than 2^32 iterations???
 142
 143   unsigned LoopSize = ApproximateLoopSize(L);
 144   DEBUG(std::cerr << "Loop Unroll: F[" << BB->getParent()->getName()
 145         << "] Loop %" << BB->getName() << " Loop Size = " << LoopSize
 146         << " Trip Count = " << TripCountFull << " - ");
 147   uint64_t Size = (uint64_t)LoopSize*TripCountFull;
 148   if (Size > UnrollThreshold) {
 149     DEBUG(std::cerr << "TOO LARGE: " << Size << ">" << UnrollThreshold << "\n");
 150     return Changed;
 151   }
 152   DEBUG(std::cerr << "UNROLLING!\n");
 153
 154   unsigned TripCount = (unsigned)TripCountFull;
 155
 156   BasicBlock *LoopExit = BI->getSuccessor(L->contains(BI->getSuccessor(0)));
 157
 158   // Create a new basic block to temporarily hold all of the cloned code.
 159   BasicBlock *NewBlock = new BasicBlock();
 160
 161   // For the first iteration of the loop, we should use the precloned values for
 162   // PHI nodes.  Insert associations now.
 163   std::map<const Value*, Value*> LastValueMap;
 164   std::vector<PHINode*> OrigPHINode;
 165   for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++I) {
 166     PHINode *PN = cast<PHINode>(I);
 167     OrigPHINode.push_back(PN);
 168     if (Instruction *I =dyn_cast<Instruction>(PN->getIncomingValueForBlock(BB)))
 169       if (I->getParent() == BB)
 170         LastValueMap[I] = I;
 171   }
 172
 173   // Remove the exit branch from the loop
 174   BB->getInstList().erase(BI);
 175
 176   assert(TripCount != 0 && "Trip count of 0 is impossible!");
 177   for (unsigned It = 1; It != TripCount; ++It) {
 178     char SuffixBuffer[100];
 179     sprintf(SuffixBuffer, ".%d", It);
 180     std::map<const Value*, Value*> ValueMap;
 181     BasicBlock *New = CloneBasicBlock(BB, ValueMap, SuffixBuffer);
 182
 183     // Loop over all of the PHI nodes in the block, changing them to use the
 184     // incoming values from the previous block.
 185     for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
 186       PHINode *NewPHI = cast<PHINode>(ValueMap[OrigPHINode[i]]);
 187       Value *InVal = NewPHI->getIncomingValueForBlock(BB);
 188       if (Instruction *InValI = dyn_cast<Instruction>(InVal))
 189         if (InValI->getParent() == BB)
 190           InVal = LastValueMap[InValI];
 191       ValueMap[OrigPHINode[i]] = InVal;
 192       New->getInstList().erase(NewPHI);
 193     }
 194
 195     for (BasicBlock::iterator I = New->begin(), E = New->end(); I != E; ++I)
 196       RemapInstruction(I, ValueMap);
 197
 198     // Now that all of the instructions are remapped, splice them into the end
 199     // of the NewBlock.
 200     NewBlock->getInstList().splice(NewBlock->end(), New->getInstList());
 201     delete New;
 202
 203     // LastValue map now contains values from this iteration.
 204     std::swap(LastValueMap, ValueMap);
 205   }
 206
 207   // If there was more than one iteration, replace any uses of values computed
 208   // in the loop with values computed during the last iteration of the loop.
 209   if (TripCount != 1) {
 210     std::set<User*> Users;
 211     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
 212       Users.insert(I->use_begin(), I->use_end());
 213
 214     // We don't want to reprocess entries with PHI nodes in them.  For this
 215     // reason, we look at each operand of each user exactly once, performing the
 216     // substitution exactly once.
 217     for (std::set<User*>::iterator UI = Users.begin(), E = Users.end(); UI != E;
 218          ++UI) {
 219       Instruction *I = cast<Instruction>(*UI);
 220       if (I->getParent() != BB && I->getParent() != NewBlock)
 221         RemapInstruction(I, LastValueMap);
 222     }
 223   }
 224
 225   // Now that we cloned the block as many times as we needed, stitch the new
 226   // code into the original block and delete the temporary block.
 227   BB->getInstList().splice(BB->end(), NewBlock->getInstList());
 228   delete NewBlock;
 229
 230   // Now loop over the PHI nodes in the original block, setting them to their
 231   // incoming values.
 232   BasicBlock *Preheader = L->getLoopPreheader();
 233   for (unsigned i = 0, e = OrigPHINode.size(); i != e; ++i) {
 234     PHINode *PN = OrigPHINode[i];
 235     PN->replaceAllUsesWith(PN->getIncomingValueForBlock(Preheader));
 236     BB->getInstList().erase(PN);
 237   }
 238
 239   // Finally, add an unconditional branch to the block to continue into the exit
 240   // block.
 241   new BranchInst(LoopExit, BB);
 242
 243   // At this point, the code is well formed.  We now do a quick sweep over the
 244   // inserted code, doing constant propagation and dead code elimination as we
 245   // go.
 246   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
 247     Instruction *Inst = I++;
 248
 249     if (isInstructionTriviallyDead(Inst))
 250       BB->getInstList().erase(Inst);
 251     else if (Constant *C = ConstantFoldInstruction(Inst)) {
 252       Inst->replaceAllUsesWith(C);
 253       BB->getInstList().erase(Inst);
 254     }
 255   }
 256
 257   // Update the loop information for this loop.
 258   Loop *Parent = L->getParentLoop();
 259
 260   // Move all of the basic blocks in the loop into the parent loop.
 261   LI->changeLoopFor(BB, Parent);
 262
 263   // Remove the loop from the parent.
 264   if (Parent)
 265     delete Parent->removeChildLoop(std::find(Parent->begin(), Parent->end(),L));
 266   else
 267     delete LI->removeLoop(std::find(LI->begin(), LI->end(), L));
 268
 269   // Remove single-entry Phis from the exit block.
 270   for (BasicBlock::iterator ExitInstr = LoopExit->begin();
 271        PHINode* PN = dyn_cast<PHINode>(ExitInstr); ++ExitInstr) {
 272     assert(PN->getNumIncomingValues() == 1
 273            && "Block should only have one pred, so Phi's must be single entry");
 274     PN->replaceAllUsesWith(PN->getOperand(0));
 275     PN->eraseFromParent();
 276   }
 277
 278   // FIXME: Should update dominator analyses
 279
 280   // Now that everything is up-to-date that will be, we fold the loop block into
 281   // the preheader and exit block, updating our analyses as we go.
 282   LoopExit->getInstList().splice(LoopExit->begin(), BB->getInstList(),
 283                                  BB->getInstList().begin(),
 284                                  prior(BB->getInstList().end()));
 285   LoopExit->getInstList().splice(LoopExit->begin(), Preheader->getInstList(),
 286                                  Preheader->getInstList().begin(),
 287                                  prior(Preheader->getInstList().end()));
 288
 289   // Make all other blocks in the program branch to LoopExit now instead of
 290   // Preheader.
 291   Preheader->replaceAllUsesWith(LoopExit);
 292
 293   Function *F = LoopExit->getParent();
 294   if (Parent) {
 295     // Otherwise, if this is a sub-loop, and the preheader was the loop header
 296     // of the parent loop, move the exit block to be the new parent loop header.
 297     if (Parent->getHeader() == Preheader) {
 298       assert(Parent->contains(LoopExit) &&
 299              "Exit block isn't contained in parent?");
 300       Parent->moveToHeader(LoopExit);
 301     }
 302   } else {
 303     // If the preheader was the entry block of this function, move the exit
 304     // block to be the new entry of the function.
 305     if (Preheader == &F->front())
 306       F->getBasicBlockList().splice(F->begin(),
 307                                     F->getBasicBlockList(), LoopExit);
 308   }
 309
 310   // Remove BB and LoopExit from our analyses.
 311   LI->removeBlock(Preheader);
 312   LI->removeBlock(BB);
 313
 314   // Actually delete the blocks now.
 315   F->getBasicBlockList().erase(Preheader);
 316   F->getBasicBlockList().erase(BB);
 317
 318   ++NumUnrolled;
 319   return true;
 320 }