lib/Analysis/LoopVR.cpp

   1 //===- LoopVR.cpp - Value Range analysis driven by loop information -------===//
   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 // FIXME: What does this do?
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #define DEBUG_TYPE "loopvr"
  15 #include "llvm/Analysis/LoopVR.h"
  16 #include "llvm/Constants.h"
  17 #include "llvm/Instructions.h"
  18 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
  19 #include "llvm/Assembly/Writer.h"
  20 #include "llvm/Support/CFG.h"
  21 #include "llvm/Support/Debug.h"
  22 #include "llvm/Support/raw_ostream.h"
  23 using namespace llvm;
  24
  25 char LoopVR::ID = 0;
  26 static RegisterPass<LoopVR> X("loopvr", "Loop Value Ranges", false, true);
  27
  28 /// getRange - determine the range for a particular SCEV within a given Loop
  29 ConstantRange LoopVR::getRange(SCEVHandle S, Loop *L, ScalarEvolution &SE) {
  30   SCEVHandle T = SE.getBackedgeTakenCount(L);
  31   if (isa<SCEVCouldNotCompute>(T))
  32     return ConstantRange(cast<IntegerType>(S->getType())->getBitWidth(), true);
  33
  34   T = SE.getTruncateOrZeroExtend(T, S->getType());
  35   return getRange(S, T, SE);
  36 }
  37
  38 /// getRange - determine the range for a particular SCEV with a given trip count
  39 ConstantRange LoopVR::getRange(SCEVHandle S, SCEVHandle T, ScalarEvolution &SE){
  40
  41   if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S))
  42     return ConstantRange(C->getValue()->getValue());
  43
  44   ConstantRange FullSet(cast<IntegerType>(S->getType())->getBitWidth(), true);
  45
  46   // {x,+,y,+,...z}. We detect overflow by checking the size of the set after
  47   // summing the upper and lower.
  48   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
  49     ConstantRange X = getRange(Add->getOperand(0), T, SE);
  50     if (X.isFullSet()) return FullSet;
  51     for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i) {
  52       ConstantRange Y = getRange(Add->getOperand(i), T, SE);
  53       if (Y.isFullSet()) return FullSet;
  54
  55       APInt Spread_X = X.getSetSize(), Spread_Y = Y.getSetSize();
  56       APInt NewLower = X.getLower() + Y.getLower();
  57       APInt NewUpper = X.getUpper() + Y.getUpper() - 1;
  58       if (NewLower == NewUpper)
  59         return FullSet;
  60
  61       X = ConstantRange(NewLower, NewUpper);
  62       if (X.getSetSize().ult(Spread_X) || X.getSetSize().ult(Spread_Y))
  63         return FullSet; // we've wrapped, therefore, full set.
  64     }
  65     return X;
  66   }
  67
  68   // {x,*,y,*,...,z}. In order to detect overflow, we use k*bitwidth where
  69   // k is the number of terms being multiplied.
  70   if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
  71     ConstantRange X = getRange(Mul->getOperand(0), T, SE);
  72     if (X.isFullSet()) return FullSet;
  73
  74     const IntegerType *Ty = IntegerType::get(X.getBitWidth());
  75     const IntegerType *ExTy = IntegerType::get(X.getBitWidth() *
  76                                                Mul->getNumOperands());
  77     ConstantRange XExt = X.zeroExtend(ExTy->getBitWidth());
  78
  79     for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i) {
  80       ConstantRange Y = getRange(Mul->getOperand(i), T, SE);
  81       if (Y.isFullSet()) return FullSet;
  82
  83       ConstantRange YExt = Y.zeroExtend(ExTy->getBitWidth());
  84       XExt = ConstantRange(XExt.getLower() * YExt.getLower(),
  85                            ((XExt.getUpper()-1) * (YExt.getUpper()-1)) + 1);
  86     }
  87     return XExt.truncate(Ty->getBitWidth());
  88   }
  89
  90   // X smax Y smax ... Z is: range(smax(X_smin, Y_smin, ..., Z_smin),
  91   //                               smax(X_smax, Y_smax, ..., Z_smax))
  92   // It doesn't matter if one of the SCEVs has FullSet because we're taking
  93   // a maximum of the minimums across all of them.
  94   if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) {
  95     ConstantRange X = getRange(SMax->getOperand(0), T, SE);
  96     if (X.isFullSet()) return FullSet;
  97
  98     APInt smin = X.getSignedMin(), smax = X.getSignedMax();
  99     for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i) {
 100       ConstantRange Y = getRange(SMax->getOperand(i), T, SE);
 101       smin = APIntOps::smax(smin, Y.getSignedMin());
 102       smax = APIntOps::smax(smax, Y.getSignedMax());
 103     }
 104     if (smax + 1 == smin) return FullSet;
 105     return ConstantRange(smin, smax + 1);
 106   }
 107
 108   // X umax Y umax ... Z is: range(umax(X_umin, Y_umin, ..., Z_umin),
 109   //                               umax(X_umax, Y_umax, ..., Z_umax))
 110   // It doesn't matter if one of the SCEVs has FullSet because we're taking
 111   // a maximum of the minimums across all of them.
 112   if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) {
 113     ConstantRange X = getRange(UMax->getOperand(0), T, SE);
 114     if (X.isFullSet()) return FullSet;
 115
 116     APInt umin = X.getUnsignedMin(), umax = X.getUnsignedMax();
 117     for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i) {
 118       ConstantRange Y = getRange(UMax->getOperand(i), T, SE);
 119       umin = APIntOps::umax(umin, Y.getUnsignedMin());
 120       umax = APIntOps::umax(umax, Y.getUnsignedMax());
 121     }
 122     if (umax + 1 == umin) return FullSet;
 123     return ConstantRange(umin, umax + 1);
 124   }
 125
 126   // L udiv R. Luckily, there's only ever 2 sides to a udiv.
 127   if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) {
 128     ConstantRange L = getRange(UDiv->getLHS(), T, SE);
 129     ConstantRange R = getRange(UDiv->getRHS(), T, SE);
 130     if (L.isFullSet() && R.isFullSet()) return FullSet;
 131
 132     if (R.getUnsignedMax() == 0) {
 133       // RHS must be single-element zero. Return an empty set.
 134       return ConstantRange(R.getBitWidth(), false);
 135     }
 136
 137     APInt Lower = L.getUnsignedMin().udiv(R.getUnsignedMax());
 138
 139     APInt Upper;
 140
 141     if (R.getUnsignedMin() == 0) {
 142       // Just because it contains zero, doesn't mean it will also contain one.
 143       // Use maximalIntersectWith to get the right behaviour.
 144       ConstantRange NotZero(APInt(L.getBitWidth(), 1),
 145                             APInt::getNullValue(L.getBitWidth()));
 146       R = R.maximalIntersectWith(NotZero);
 147     }
 148
 149     // But, the maximal intersection might still include zero. If it does, then
 150     // we know it also included one.
 151     if (R.contains(APInt::getNullValue(L.getBitWidth())))
 152       Upper = L.getUnsignedMax();
 153     else
 154       Upper = L.getUnsignedMax().udiv(R.getUnsignedMin());
 155
 156     return ConstantRange(Lower, Upper);
 157   }
 158
 159   // ConstantRange already implements the cast operators.
 160
 161   if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) {
 162     T = SE.getTruncateOrZeroExtend(T, ZExt->getOperand()->getType());
 163     ConstantRange X = getRange(ZExt->getOperand(), T, SE);
 164     return X.zeroExtend(cast<IntegerType>(ZExt->getType())->getBitWidth());
 165   }
 166
 167   if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) {
 168     T = SE.getTruncateOrZeroExtend(T, SExt->getOperand()->getType());
 169     ConstantRange X = getRange(SExt->getOperand(), T, SE);
 170     return X.signExtend(cast<IntegerType>(SExt->getType())->getBitWidth());
 171   }
 172
 173   if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) {
 174     T = SE.getTruncateOrZeroExtend(T, Trunc->getOperand()->getType());
 175     ConstantRange X = getRange(Trunc->getOperand(), T, SE);
 176     if (X.isFullSet()) return FullSet;
 177     return X.truncate(cast<IntegerType>(Trunc->getType())->getBitWidth());
 178   }
 179
 180   if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) {
 181     const SCEVConstant *Trip = dyn_cast<SCEVConstant>(T);
 182     if (!Trip) return FullSet;
 183
 184     if (AddRec->isAffine()) {
 185       SCEVHandle StartHandle = AddRec->getStart();
 186       SCEVHandle StepHandle = AddRec->getOperand(1);
 187
 188       const SCEVConstant *Step = dyn_cast<SCEVConstant>(StepHandle);
 189       if (!Step) return FullSet;
 190
 191       uint32_t ExWidth = 2 * Trip->getValue()->getBitWidth();
 192       APInt TripExt = Trip->getValue()->getValue(); TripExt.zext(ExWidth);
 193       APInt StepExt = Step->getValue()->getValue(); StepExt.zext(ExWidth);
 194       if ((TripExt * StepExt).ugt(APInt::getLowBitsSet(ExWidth, ExWidth >> 1)))
 195         return FullSet;
 196
 197       SCEVHandle EndHandle = SE.getAddExpr(StartHandle,
 198                                            SE.getMulExpr(T, StepHandle));
 199       const SCEVConstant *Start = dyn_cast<SCEVConstant>(StartHandle);
 200       const SCEVConstant *End = dyn_cast<SCEVConstant>(EndHandle);
 201       if (!Start || !End) return FullSet;
 202
 203       const APInt &StartInt = Start->getValue()->getValue();
 204       const APInt &EndInt = End->getValue()->getValue();
 205       const APInt &StepInt = Step->getValue()->getValue();
 206
 207       if (StepInt.isNegative()) {
 208         if (EndInt == StartInt + 1) return FullSet;
 209         return ConstantRange(EndInt, StartInt + 1);
 210       } else {
 211         if (StartInt == EndInt + 1) return FullSet;
 212         return ConstantRange(StartInt, EndInt + 1);
 213       }
 214     }
 215   }
 216
 217   // TODO: non-affine addrec, udiv, SCEVUnknown (narrowed from elsewhere)?
 218
 219   return FullSet;
 220 }
 221
 222 bool LoopVR::runOnFunction(Function &F) { Map.clear(); return false; }
 223
 224 void LoopVR::print(std::ostream &os, const Module *) const {
 225   raw_os_ostream OS(os);
 226   for (std::map<Value *, ConstantRange *>::const_iterator I = Map.begin(),
 227        E = Map.end(); I != E; ++I) {
 228     OS << *I->first << ": " << *I->second << '\n';
 229   }
 230 }
 231
 232 void LoopVR::releaseMemory() {
 233   for (std::map<Value *, ConstantRange *>::iterator I = Map.begin(),
 234        E = Map.end(); I != E; ++I) {
 235     delete I->second;
 236   }
 237
 238   Map.clear();
 239 }
 240
 241 ConstantRange LoopVR::compute(Value *V) {
 242   if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
 243     return ConstantRange(CI->getValue());
 244
 245   Instruction *I = dyn_cast<Instruction>(V);
 246   if (!I)
 247     return ConstantRange(cast<IntegerType>(V->getType())->getBitWidth(), false);
 248
 249   LoopInfo &LI = getAnalysis<LoopInfo>();
 250
 251   Loop *L = LI.getLoopFor(I->getParent());
 252   if (!L || L->isLoopInvariant(I))
 253     return ConstantRange(cast<IntegerType>(V->getType())->getBitWidth(), false);
 254
 255   ScalarEvolution &SE = getAnalysis<ScalarEvolution>();
 256
 257   SCEVHandle S = SE.getSCEV(I);
 258   if (isa<SCEVUnknown>(S) || isa<SCEVCouldNotCompute>(S))
 259     return ConstantRange(cast<IntegerType>(V->getType())->getBitWidth(), false);
 260
 261   return ConstantRange(getRange(S, L, SE));
 262 }
 263
 264 ConstantRange LoopVR::get(Value *V) {
 265   std::map<Value *, ConstantRange *>::iterator I = Map.find(V);
 266   if (I == Map.end()) {
 267     ConstantRange *CR = new ConstantRange(compute(V));
 268     Map[V] = CR;
 269     return *CR;
 270   }
 271
 272   return *I->second;
 273 }
 274
 275 void LoopVR::remove(Value *V) {
 276   std::map<Value *, ConstantRange *>::iterator I = Map.find(V);
 277   if (I != Map.end()) {
 278     delete I->second;
 279     Map.erase(I);
 280   }
 281 }
 282
 283 void LoopVR::narrow(Value *V, const ConstantRange &CR) {
 284   if (CR.isFullSet()) return;
 285
 286   std::map<Value *, ConstantRange *>::iterator I = Map.find(V);
 287   if (I == Map.end())
 288     Map[V] = new ConstantRange(CR);
 289   else
 290     Map[V] = new ConstantRange(Map[V]->maximalIntersectWith(CR));
 291 }