path: root/lib/Transforms/Scalar/LoopUnroll.cpp

//===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass implements a simple loop unroller.  It works best when loops have
// been canonicalized by the -indvars pass, allowing it to determine the trip
// counts of loops easily.
//
// This pass will multi-block loops only if they contain no non-unrolled 
// subloops.  The process of unrolling can produce extraneous basic blocks 
// linked with unconditional branches.  This will be corrected in the future.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "loop-unroll"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Transforms/Utils/Cloning.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/IntrinsicInst.h"
#include <cstdio>
#include <algorithm>
using namespace llvm;

STATISTIC(NumCompletelyUnrolled, "Number of loops completely unrolled");
STATISTIC(NumUnrolled,    "Number of loops unrolled (completely or otherwise)");

namespace {
  cl::opt<unsigned>
  UnrollThreshold
    ("unroll-threshold", cl::init(100), cl::Hidden,
     cl::desc("The cut-off point for automatic loop unrolling"));

  cl::opt<unsigned>
  UnrollCount
    ("unroll-count", cl::init(0), cl::Hidden,
     cl::desc("Use this unroll count for all loops, for testing purposes"));

  class VISIBILITY_HIDDEN LoopUnroll : public LoopPass {
    LoopInfo *LI;  // The current loop information
  public:
    static char ID; // Pass ID, replacement for typeid
    LoopUnroll() : LoopPass((intptr_t)&ID) {}

    /// A magic value for use with the Threshold parameter to indicate
    /// that the loop unroll should be performed regardless of how much
    /// code expansion would result.
    static const unsigned NoThreshold = UINT_MAX;

    bool runOnLoop(Loop *L, LPPassManager &LPM);
    bool unrollLoop(Loop *L, unsigned Count, unsigned Threshold);
    BasicBlock *FoldBlockIntoPredecessor(BasicBlock *BB);

    /// This transformation requires natural loop information & requires that
    /// loop preheaders be inserted into the CFG...
    ///
    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
      AU.addRequiredID(LoopSimplifyID);
      AU.addRequiredID(LCSSAID);
      AU.addRequired<LoopInfo>();
      AU.addPreservedID(LCSSAID);
      AU.addPreserved<LoopInfo>();
    }
  };
  char LoopUnroll::ID = 0;
  RegisterPass<LoopUnroll> X("loop-unroll", "Unroll loops");
}

LoopPass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }

/// ApproximateLoopSize - Approximate the size of the loop.
static unsigned ApproximateLoopSize(const Loop *L) {
  unsigned Size = 0;
  for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) {
    BasicBlock *BB = L->getBlocks()[i];
    Instruction *Term = BB->getTerminator();
    for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
      if (isa<PHINode>(I) && BB == L->getHeader()) {
        // Ignore PHI nodes in the header.
      } else if (I->hasOneUse() && I->use_back() == Term) {
        // Ignore instructions only used by the loop terminator.
      } else if (isa<DbgInfoIntrinsic>(I)) {
        // Ignore debug instructions
      } else {
        ++Size;
      }

      // TODO: Ignore expressions derived from PHI and constants if inval of phi
      // is a constant, or if operation is associative.  This will get induction
      // variables.
    }
  }

  return Size;
}

// RemapInstruction - Convert the instruction operands from referencing the
// current values into those specified by ValueMap.
//
static inline void RemapInstruction(Instruction *I,
                                    DenseMap<const Value *, Value*> &ValueMap) {
  for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
    Value *Op = I->getOperand(op);
    DenseMap<const Value *, Value*>::iterator It = ValueMap.find(Op);
    if (It != ValueMap.end()) Op = It->second;
    I->setOperand(op, Op);
  }
}

// FoldBlockIntoPredecessor - Folds a basic block into its predecessor if it
// only has one predecessor, and that predecessor only has one successor.
// Returns the new combined block.
BasicBlock *LoopUnroll::FoldBlockIntoPredecessor(BasicBlock *BB) {
  // Merge basic blocks into their predecessor if there is only one distinct
  // pred, and if there is only one distinct successor of the predecessor, and
  // if there are no PHI nodes.
  //
  BasicBlock *OnlyPred = BB->getSinglePredecessor();
  if (!OnlyPred) return 0;

  if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
    return 0;

  DOUT << "Merging: " << *BB << "into: " << *OnlyPred;

  // Resolve any PHI nodes at the start of the block.  They are all
  // guaranteed to have exactly one entry if they exist, unless there are
  // multiple duplicate (but guaranteed to be equal) entries for the
  // incoming edges.  This occurs when there are multiple edges from
  // OnlyPred to OnlySucc.
  //
  while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
    PN->replaceAllUsesWith(PN->getIncomingValue(0));
    BB->getInstList().pop_front();  // Delete the phi node...
  }

  // Delete the unconditional branch from the predecessor...
  OnlyPred->getInstList().pop_back();

  // Move all definitions in the successor to the predecessor...
  OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());

  // Make all PHI nodes that referred to BB now refer to Pred as their
  // source...
  BB->replaceAllUsesWith(OnlyPred);

  std::string OldName = BB->getName();

  // Erase basic block from the function...
  LI->removeBlock(BB);
  BB->eraseFromParent();

  // Inherit predecessor's name if it exists...
  if (!OldName.empty() && !OnlyPred->hasName())
    OnlyPred->setName(OldName);

  return OnlyPred;
}

bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
  LI = &getAnalysis<LoopInfo>();

  // Unroll the loop.
  if (!unrollLoop(L, UnrollCount, UnrollThreshold))
    return false;

  // Update the loop information for this loop.
  // If we completely unrolled the loop, remove it from the parent.
  if (L->getNumBackEdges() == 0)
    LPM.deleteLoopFromQueue(L);

  return true;
}

/// Unroll the given loop by UnrollCount, or by a heuristically-determined
/// value if Count is zero. If Threshold is not NoThreshold, it is a value
/// to limit code size expansion. If the loop size would expand beyond the
/// threshold value, unrolling is suppressed. The return value is true if
/// any transformations are performed.
///
bool LoopUnroll::unrollLoop(Loop *L, unsigned Count, unsigned Threshold) {
  assert(L->isLCSSAForm());

  BasicBlock *Header = L->getHeader();
  BasicBlock *LatchBlock = L->getLoopLatch();
  BranchInst *BI = dyn_cast<BranchInst>(LatchBlock->getTerminator());

  DOUT << "Loop Unroll: F[" << Header->getParent()->getName()
       << "] Loop %" << Header->getName() << "\n";

  if (!BI || BI->isUnconditional()) {
    // The loop-rorate pass can be helpful to avoid this in many cases.
    DOUT << "  Can't unroll; loop not terminated by a conditional branch.\n";
    return false;
  }

  // Determine the trip count and/or trip multiple. A TripCount value of zero
  // is used to mean an unknown trip count. The TripMultiple value is the
  // greatest known integer multiple of the trip count.
  unsigned TripCount = 0;
  unsigned TripMultiple = 1;
  if (Value *TripCountValue = L->getTripCount()) {
    if (ConstantInt *TripCountC = dyn_cast<ConstantInt>(TripCountValue)) {
      // Guard against huge trip counts. This also guards against assertions in
      // APInt from the use of getZE