lib/Transforms/Scalar/LoopUnrollPass.cpp

   1 //===-- LoopUnroll.cpp - Loop unroller pass -------------------------------===//
   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 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
  15 #define DEBUG_TYPE "loop-unroll"
  16 #include "llvm/IntrinsicInst.h"
  17 #include "llvm/Transforms/Scalar.h"
  18 #include "llvm/Analysis/LoopPass.h"
  19 #include "llvm/Analysis/CodeMetrics.h"
  20 #include "llvm/Analysis/ScalarEvolution.h"
  21 #include "llvm/Support/CommandLine.h"
  22 #include "llvm/Support/Debug.h"
  23 #include "llvm/Support/raw_ostream.h"
  24 #include "llvm/Transforms/Utils/UnrollLoop.h"
  25 #include <climits>
  26
  27 using namespace llvm;
  28
  29 static cl::opt<unsigned>
  30 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
  31   cl::desc("The cut-off point for automatic loop unrolling"));
  32
  33 static cl::opt<unsigned>
  34 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
  35   cl::desc("Use this unroll count for all loops, for testing purposes"));
  36
  37 static cl::opt<bool>
  38 UnrollAllowPartial("unroll-allow-partial", cl::init(false), cl::Hidden,
  39   cl::desc("Allows loops to be partially unrolled until "
  40            "-unroll-threshold loop size is reached."));
  41
  42 namespace {
  43   class LoopUnroll : public LoopPass {
  44   public:
  45     static char ID; // Pass ID, replacement for typeid
  46     LoopUnroll() : LoopPass(ID) {
  47       initializeLoopUnrollPass(*PassRegistry::getPassRegistry());
  48     }
  49
  50     /// A magic value for use with the Threshold parameter to indicate
  51     /// that the loop unroll should be performed regardless of how much
  52     /// code expansion would result.
  53     static const unsigned NoThreshold = UINT_MAX;
  54
  55     // Threshold to use when optsize is specified (and there is no
  56     // explicit -unroll-threshold).
  57     static const unsigned OptSizeUnrollThreshold = 50;
  58
  59     unsigned CurrentThreshold;
  60
  61     bool runOnLoop(Loop *L, LPPassManager &LPM);
  62
  63     /// This transformation requires natural loop information & requires that
  64     /// loop preheaders be inserted into the CFG...
  65     ///
  66     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  67       AU.addRequired<LoopInfo>();
  68       AU.addPreserved<LoopInfo>();
  69       AU.addRequiredID(LoopSimplifyID);
  70       AU.addPreservedID(LoopSimplifyID);
  71       AU.addRequiredID(LCSSAID);
  72       AU.addPreservedID(LCSSAID);
  73       AU.addPreserved<ScalarEvolution>();
  74       // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
  75       // If loop unroll does not preserve dom info then LCSSA pass on next
  76       // loop will receive invalid dom info.
  77       // For now, recreate dom info, if loop is unrolled.
  78       AU.addPreserved<DominatorTree>();
  79     }
  80   };
  81 }
  82
  83 char LoopUnroll::ID = 0;
  84 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
  85 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
  86 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
  87 INITIALIZE_PASS_DEPENDENCY(LCSSA)
  88 INITIALIZE_PASS_END(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
  89
  90 Pass *llvm::createLoopUnrollPass() { return new LoopUnroll(); }
  91
  92 /// ApproximateLoopSize - Approximate the size of the loop.
  93 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls) {
  94   CodeMetrics Metrics;
  95   for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
  96        I != E; ++I)
  97     Metrics.analyzeBasicBlock(*I);
  98   NumCalls = Metrics.NumInlineCandidates;
  99
 100   unsigned LoopSize = Metrics.NumInsts;
 101
 102   // Don't allow an estimate of size zero.  This would allows unrolling of loops
 103   // with huge iteration counts, which is a compile time problem even if it's
 104   // not a problem for code quality.
 105   if (LoopSize == 0) LoopSize = 1;
 106
 107   return LoopSize;
 108 }
 109
 110 bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
 111   LoopInfo *LI = &getAnalysis<LoopInfo>();
 112
 113   BasicBlock *Header = L->getHeader();
 114   DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
 115         << "] Loop %" << Header->getName() << "\n");
 116   (void)Header;
 117
 118   // Determine the current unrolling threshold.  While this is normally set
 119   // from UnrollThreshold, it is overridden to a smaller value if the current
 120   // function is marked as optimize-for-size, and the unroll threshold was
 121   // not user specified.
 122   CurrentThreshold = UnrollThreshold;
 123   if (Header->getParent()->hasFnAttr(Attribute::OptimizeForSize) &&
 124       UnrollThreshold.getNumOccurrences() == 0)
 125     CurrentThreshold = OptSizeUnrollThreshold;
 126
 127   // Find trip count
 128   unsigned TripCount = L->getSmallConstantTripCount();
 129   unsigned Count = UnrollCount;
 130
 131   // Automatically select an unroll count.
 132   if (Count == 0) {
 133     // Conservative heuristic: if we know the trip count, see if we can
 134     // completely unroll (subject to the threshold, checked below); otherwise
 135     // try to find greatest modulo of the trip count which is still under
 136     // threshold value.
 137     if (TripCount == 0)
 138       return false;
 139     Count = TripCount;
 140   }
 141
 142   // Enforce the threshold.
 143   if (CurrentThreshold != NoThreshold) {
 144     unsigned NumInlineCandidates;
 145     unsigned LoopSize = ApproximateLoopSize(L, NumInlineCandidates);
 146     DEBUG(dbgs() << "  Loop Size = " << LoopSize << "\n");
 147     if (NumInlineCandidates != 0) {
 148       DEBUG(dbgs() << "  Not unrolling loop with inlinable calls.\n");
 149       return false;
 150     }
 151     uint64_t Size = (uint64_t)LoopSize*Count;
 152     if (TripCount != 1 && Size > CurrentThreshold) {
 153       DEBUG(dbgs() << "  Too large to fully unroll with count: " << Count
 154             << " because size: " << Size << ">" << CurrentThreshold << "\n");
 155       if (!UnrollAllowPartial) {
 156         DEBUG(dbgs() << "  will not try to unroll partially because "
 157               << "-unroll-allow-partial not given\n");
 158         return false;
 159       }
 160       // Reduce unroll count to be modulo of TripCount for partial unrolling
 161       Count = CurrentThreshold / LoopSize;
 162       while (Count != 0 && TripCount%Count != 0) {
 163         Count--;
 164       }
 165       if (Count < 2) {
 166         DEBUG(dbgs() << "  could not unroll partially\n");
 167         return false;
 168       }
 169       DEBUG(dbgs() << "  partially unrolling with count: " << Count << "\n");
 170     }
 171   }
 172
 173   // Unroll the loop.
 174   Function *F = L->getHeader()->getParent();
 175   if (!UnrollLoop(L, Count, LI, &LPM))
 176     return false;
 177
 178   // FIXME: Reconstruct dom info, because it is not preserved properly.
 179   if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>())
 180     DT->runOnFunction(*F);
 181   return true;
 182 }