LLVM API Documentation

LoopStrengthReduce.cpp
Go to the documentation of this file.
00001 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This transformation analyzes and transforms the induction variables (and
00011 // computations derived from them) into forms suitable for efficient execution
00012 // on the target.
00013 //
00014 // This pass performs a strength reduction on array references inside loops that
00015 // have as one or more of their components the loop induction variable, it
00016 // rewrites expressions to take advantage of scaled-index addressing modes
00017 // available on the target, and it performs a variety of other optimizations
00018 // related to loop induction variables.
00019 //
00020 // Terminology note: this code has a lot of handling for "post-increment" or
00021 // "post-inc" users. This is not talking about post-increment addressing modes;
00022 // it is instead talking about code like this:
00023 //
00024 //   %i = phi [ 0, %entry ], [ %i.next, %latch ]
00025 //   ...
00026 //   %i.next = add %i, 1
00027 //   %c = icmp eq %i.next, %n
00028 //
00029 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
00030 // it's useful to think about these as the same register, with some uses using
00031 // the value of the register before the add and some using // it after. In this
00032 // example, the icmp is a post-increment user, since it uses %i.next, which is
00033 // the value of the induction variable after the increment. The other common
00034 // case of post-increment users is users outside the loop.
00035 //
00036 // TODO: More sophistication in the way Formulae are generated and filtered.
00037 //
00038 // TODO: Handle multiple loops at a time.
00039 //
00040 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead
00041 //       of a GlobalValue?
00042 //
00043 // TODO: When truncation is free, truncate ICmp users' operands to make it a
00044 //       smaller encoding (on x86 at least).
00045 //
00046 // TODO: When a negated register is used by an add (such as in a list of
00047 //       multiple base registers, or as the increment expression in an addrec),
00048 //       we may not actually need both reg and (-1 * reg) in registers; the
00049 //       negation can be implemented by using a sub instead of an add. The
00050 //       lack of support for taking this into consideration when making
00051 //       register pressure decisions is partly worked around by the "Special"
00052 //       use kind.
00053 //
00054 //===----------------------------------------------------------------------===//
00055 
00056 #define DEBUG_TYPE "loop-reduce"
00057 #include "llvm/Transforms/Scalar.h"
00058 #include "llvm/ADT/DenseSet.h"
00059 #include "llvm/ADT/Hashing.h"
00060 #include "llvm/ADT/STLExtras.h"
00061 #include "llvm/ADT/SetVector.h"
00062 #include "llvm/ADT/SmallBitVector.h"
00063 #include "llvm/Analysis/IVUsers.h"
00064 #include "llvm/Analysis/LoopPass.h"
00065 #include "llvm/Analysis/ScalarEvolutionExpander.h"
00066 #include "llvm/Analysis/TargetTransformInfo.h"
00067 #include "llvm/IR/Constants.h"
00068 #include "llvm/IR/DerivedTypes.h"
00069 #include "llvm/IR/Dominators.h"
00070 #include "llvm/IR/Instructions.h"
00071 #include "llvm/IR/IntrinsicInst.h"
00072 #include "llvm/IR/ValueHandle.h"
00073 #include "llvm/Support/CommandLine.h"
00074 #include "llvm/Support/Debug.h"
00075 #include "llvm/Support/raw_ostream.h"
00076 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
00077 #include "llvm/Transforms/Utils/Local.h"
00078 #include <algorithm>
00079 using namespace llvm;
00080 
00081 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for
00082 /// bail out. This threshold is far beyond the number of users that LSR can
00083 /// conceivably solve, so it should not affect generated code, but catches the
00084 /// worst cases before LSR burns too much compile time and stack space.
00085 static const unsigned MaxIVUsers = 200;
00086 
00087 // Temporary flag to cleanup congruent phis after LSR phi expansion.
00088 // It's currently disabled until we can determine whether it's truly useful or
00089 // not. The flag should be removed after the v3.0 release.
00090 // This is now needed for ivchains.
00091 static cl::opt<bool> EnablePhiElim(
00092   "enable-lsr-phielim", cl::Hidden, cl::init(true),
00093   cl::desc("Enable LSR phi elimination"));
00094 
00095 #ifndef NDEBUG
00096 // Stress test IV chain generation.
00097 static cl::opt<bool> StressIVChain(
00098   "stress-ivchain", cl::Hidden, cl::init(false),
00099   cl::desc("Stress test LSR IV chains"));
00100 #else
00101 static bool StressIVChain = false;
00102 #endif
00103 
00104 namespace {
00105 
00106 /// RegSortData - This class holds data which is used to order reuse candidates.
00107 class RegSortData {
00108 public:
00109   /// UsedByIndices - This represents the set of LSRUse indices which reference
00110   /// a particular register.
00111   SmallBitVector UsedByIndices;
00112 
00113   RegSortData() {}
00114 
00115   void print(raw_ostream &OS) const;
00116   void dump() const;
00117 };
00118 
00119 }
00120 
00121 void RegSortData::print(raw_ostream &OS) const {
00122   OS << "[NumUses=" << UsedByIndices.count() << ']';
00123 }
00124 
00125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
00126 void RegSortData::dump() const {
00127   print(errs()); errs() << '\n';
00128 }
00129 #endif
00130 
00131 namespace {
00132 
00133 /// RegUseTracker - Map register candidates to information about how they are
00134 /// used.
00135 class RegUseTracker {
00136   typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
00137 
00138   RegUsesTy RegUsesMap;
00139   SmallVector<const SCEV *, 16> RegSequence;
00140 
00141 public:
00142   void CountRegister(const SCEV *Reg, size_t LUIdx);
00143   void DropRegister(const SCEV *Reg, size_t LUIdx);
00144   void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx);
00145 
00146   bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
00147 
00148   const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
00149 
00150   void clear();
00151 
00152   typedef SmallVectorImpl<const SCEV *>::iterator iterator;
00153   typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
00154   iterator begin() { return RegSequence.begin(); }
00155   iterator end()   { return RegSequence.end(); }
00156   const_iterator begin() const { return RegSequence.begin(); }
00157   const_iterator end() const   { return RegSequence.end(); }
00158 };
00159 
00160 }
00161 
00162 void
00163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
00164   std::pair<RegUsesTy::iterator, bool> Pair =
00165     RegUsesMap.insert(std::make_pair(Reg, RegSortData()));
00166   RegSortData &RSD = Pair.first->second;
00167   if (Pair.second)
00168     RegSequence.push_back(Reg);
00169   RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
00170   RSD.UsedByIndices.set(LUIdx);
00171 }
00172 
00173 void
00174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
00175   RegUsesTy::iterator It = RegUsesMap.find(Reg);
00176   assert(It != RegUsesMap.end());
00177   RegSortData &RSD = It->second;
00178   assert(RSD.UsedByIndices.size() > LUIdx);
00179   RSD.UsedByIndices.reset(LUIdx);
00180 }
00181 
00182 void
00183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) {
00184   assert(LUIdx <= LastLUIdx);
00185 
00186   // Update RegUses. The data structure is not optimized for this purpose;
00187   // we must iterate through it and update each of the bit vectors.
00188   for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
00189        I != E; ++I) {
00190     SmallBitVector &UsedByIndices = I->second.UsedByIndices;
00191     if (LUIdx < UsedByIndices.size())
00192       UsedByIndices[LUIdx] =
00193         LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0;
00194     UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx));
00195   }
00196 }
00197 
00198 bool
00199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
00200   RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
00201   if (I == RegUsesMap.end())
00202     return false;
00203   const SmallBitVector &UsedByIndices = I->second.UsedByIndices;
00204   int i = UsedByIndices.find_first();
00205   if (i == -1) return false;
00206   if ((size_t)i != LUIdx) return true;
00207   return UsedByIndices.find_next(i) != -1;
00208 }
00209 
00210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
00211   RegUsesTy::const_iterator I = RegUsesMap.find(Reg);
00212   assert(I != RegUsesMap.end() && "Unknown register!");
00213   return I->second.UsedByIndices;
00214 }
00215 
00216 void RegUseTracker::clear() {
00217   RegUsesMap.clear();
00218   RegSequence.clear();
00219 }
00220 
00221 namespace {
00222 
00223 /// Formula - This class holds information that describes a formula for
00224 /// computing satisfying a use. It may include broken-out immediates and scaled
00225 /// registers.
00226 struct Formula {
00227   /// Global base address used for complex addressing.
00228   GlobalValue *BaseGV;
00229 
00230   /// Base offset for complex addressing.
00231   int64_t BaseOffset;
00232 
00233   /// Whether any complex addressing has a base register.
00234   bool HasBaseReg;
00235 
00236   /// The scale of any complex addressing.
00237   int64_t Scale;
00238 
00239   /// BaseRegs - The list of "base" registers for this use. When this is
00240   /// non-empty,
00241   SmallVector<const SCEV *, 4> BaseRegs;
00242 
00243   /// ScaledReg - The 'scaled' register for this use. This should be non-null
00244   /// when Scale is not zero.
00245   const SCEV *ScaledReg;
00246 
00247   /// UnfoldedOffset - An additional constant offset which added near the
00248   /// use. This requires a temporary register, but the offset itself can
00249   /// live in an add immediate field rather than a register.
00250   int64_t UnfoldedOffset;
00251 
00252   Formula()
00253       : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0),
00254         UnfoldedOffset(0) {}
00255 
00256   void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE);
00257 
00258   unsigned getNumRegs() const;
00259   Type *getType() const;
00260 
00261   void DeleteBaseReg(const SCEV *&S);
00262 
00263   bool referencesReg(const SCEV *S) const;
00264   bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
00265                                   const RegUseTracker &RegUses) const;
00266 
00267   void print(raw_ostream &OS) const;
00268   void dump() const;
00269 };
00270 
00271 }
00272 
00273 /// DoInitialMatch - Recursion helper for InitialMatch.
00274 static void DoInitialMatch(const SCEV *S, Loop *L,
00275                            SmallVectorImpl<const SCEV *> &Good,
00276                            SmallVectorImpl<const SCEV *> &Bad,
00277                            ScalarEvolution &SE) {
00278   // Collect expressions which properly dominate the loop header.
00279   if (SE.properlyDominates(S, L->getHeader())) {
00280     Good.push_back(S);
00281     return;
00282   }
00283 
00284   // Look at add operands.
00285   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
00286     for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
00287          I != E; ++I)
00288       DoInitialMatch(*I, L, Good, Bad, SE);
00289     return;
00290   }
00291 
00292   // Look at addrec operands.
00293   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
00294     if (!AR->getStart()->isZero()) {
00295       DoInitialMatch(AR->getStart(), L, Good, Bad, SE);
00296       DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0),
00297                                       AR->getStepRecurrence(SE),
00298                                       // FIXME: AR->getNoWrapFlags()
00299                                       AR->getLoop(), SCEV::FlagAnyWrap),
00300                      L, Good, Bad, SE);
00301       return;
00302     }
00303 
00304   // Handle a multiplication by -1 (negation) if it didn't fold.
00305   if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
00306     if (Mul->getOperand(0)->isAllOnesValue()) {
00307       SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
00308       const SCEV *NewMul = SE.getMulExpr(Ops);
00309 
00310       SmallVector<const SCEV *, 4> MyGood;
00311       SmallVector<const SCEV *, 4> MyBad;
00312       DoInitialMatch(NewMul, L, MyGood, MyBad, SE);
00313       const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
00314         SE.getEffectiveSCEVType(NewMul->getType())));
00315       for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
00316            E = MyGood.end(); I != E; ++I)
00317         Good.push_back(SE.getMulExpr(NegOne, *I));
00318       for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
00319            E = MyBad.end(); I != E; ++I)
00320         Bad.push_back(SE.getMulExpr(NegOne, *I));
00321       return;
00322     }
00323 
00324   // Ok, we can't do anything interesting. Just stuff the whole thing into a
00325   // register and hope for the best.
00326   Bad.push_back(S);
00327 }
00328 
00329 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
00330 /// attempting to keep all loop-invariant and loop-computable values in a
00331 /// single base register.
00332 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) {
00333   SmallVector<const SCEV *, 4> Good;
00334   SmallVector<const SCEV *, 4> Bad;
00335   DoInitialMatch(S, L, Good, Bad, SE);
00336   if (!Good.empty()) {
00337     const SCEV *Sum = SE.getAddExpr(Good);
00338     if (!Sum->isZero())
00339       BaseRegs.push_back(Sum);
00340     HasBaseReg = true;
00341   }
00342   if (!Bad.empty()) {
00343     const SCEV *Sum = SE.getAddExpr(Bad);
00344     if (!Sum->isZero())
00345       BaseRegs.push_back(Sum);
00346     HasBaseReg = true;
00347   }
00348 }
00349 
00350 /// getNumRegs - Return the total number of register operands used by this
00351 /// formula. This does not include register uses implied by non-constant
00352 /// addrec strides.
00353 unsigned Formula::getNumRegs() const {
00354   return !!ScaledReg + BaseRegs.size();
00355 }
00356 
00357 /// getType - Return the type of this formula, if it has one, or null
00358 /// otherwise. This type is meaningless except for the bit size.
00359 Type *Formula::getType() const {
00360   return !BaseRegs.empty() ? BaseRegs.front()->getType() :
00361          ScaledReg ? ScaledReg->getType() :
00362          BaseGV ? BaseGV->getType() :
00363          0;
00364 }
00365 
00366 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list.
00367 void Formula::DeleteBaseReg(const SCEV *&S) {
00368   if (&S != &BaseRegs.back())
00369     std::swap(S, BaseRegs.back());
00370   BaseRegs.pop_back();
00371 }
00372 
00373 /// referencesReg - Test if this formula references the given register.
00374 bool Formula::referencesReg(const SCEV *S) const {
00375   return S == ScaledReg ||
00376          std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
00377 }
00378 
00379 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
00380 /// which are used by uses other than the use with the given index.
00381 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
00382                                          const RegUseTracker &RegUses) const {
00383   if (ScaledReg)
00384     if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
00385       return true;
00386   for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
00387        E = BaseRegs.end(); I != E; ++I)
00388     if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
00389       return true;
00390   return false;
00391 }
00392 
00393 void Formula::print(raw_ostream &OS) const {
00394   bool First = true;
00395   if (BaseGV) {
00396     if (!First) OS << " + "; else First = false;
00397     BaseGV->printAsOperand(OS, /*PrintType=*/false);
00398   }
00399   if (BaseOffset != 0) {
00400     if (!First) OS << " + "; else First = false;
00401     OS << BaseOffset;
00402   }
00403   for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
00404        E = BaseRegs.end(); I != E; ++I) {
00405     if (!First) OS << " + "; else First = false;
00406     OS << "reg(" << **I << ')';
00407   }
00408   if (HasBaseReg && BaseRegs.empty()) {
00409     if (!First) OS << " + "; else First = false;
00410     OS << "**error: HasBaseReg**";
00411   } else if (!HasBaseReg && !BaseRegs.empty()) {
00412     if (!First) OS << " + "; else First = false;
00413     OS << "**error: !HasBaseReg**";
00414   }
00415   if (Scale != 0) {
00416     if (!First) OS << " + "; else First = false;
00417     OS << Scale << "*reg(";
00418     if (ScaledReg)
00419       OS << *ScaledReg;
00420     else
00421       OS << "<unknown>";
00422     OS << ')';
00423   }
00424   if (UnfoldedOffset != 0) {
00425     if (!First) OS << " + ";
00426     OS << "imm(" << UnfoldedOffset << ')';
00427   }
00428 }
00429 
00430 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
00431 void Formula::dump() const {
00432   print(errs()); errs() << '\n';
00433 }
00434 #endif
00435 
00436 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
00437 /// without changing its value.
00438 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
00439   Type *WideTy =
00440     IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1);
00441   return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
00442 }
00443 
00444 /// isAddSExtable - Return true if the given add can be sign-extended
00445 /// without changing its value.
00446 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
00447   Type *WideTy =
00448     IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1);
00449   return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
00450 }
00451 
00452 /// isMulSExtable - Return true if the given mul can be sign-extended
00453 /// without changing its value.
00454 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) {
00455   Type *WideTy =
00456     IntegerType::get(SE.getContext(),
00457                      SE.getTypeSizeInBits(M->getType()) * M->getNumOperands());
00458   return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy));
00459 }
00460 
00461 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
00462 /// and if the remainder is known to be zero,  or null otherwise. If
00463 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
00464 /// to Y, ignoring that the multiplication may overflow, which is useful when
00465 /// the result will be used in a context where the most significant bits are
00466 /// ignored.
00467 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
00468                                 ScalarEvolution &SE,
00469                                 bool IgnoreSignificantBits = false) {
00470   // Handle the trivial case, which works for any SCEV type.
00471   if (LHS == RHS)
00472     return SE.getConstant(LHS->getType(), 1);
00473 
00474   // Handle a few RHS special cases.
00475   const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
00476   if (RC) {
00477     const APInt &RA = RC->getValue()->getValue();
00478     // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do
00479     // some folding.
00480     if (RA.isAllOnesValue())
00481       return SE.getMulExpr(LHS, RC);
00482     // Handle x /s 1 as x.
00483     if (RA == 1)
00484       return LHS;
00485   }
00486 
00487   // Check for a division of a constant by a constant.
00488   if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
00489     if (!RC)
00490       return 0;
00491     const APInt &LA = C->getValue()->getValue();
00492     const APInt &RA = RC->getValue()->getValue();
00493     if (LA.srem(RA) != 0)
00494       return 0;
00495     return SE.getConstant(LA.sdiv(RA));
00496   }
00497 
00498   // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
00499   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
00500     if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
00501       const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
00502                                       IgnoreSignificantBits);
00503       if (!Step) return 0;
00504       const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
00505                                        IgnoreSignificantBits);
00506       if (!Start) return 0;
00507       // FlagNW is independent of the start value, step direction, and is
00508       // preserved with smaller magnitude steps.
00509       // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
00510       return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap);
00511     }
00512     return 0;
00513   }
00514 
00515   // Distribute the sdiv over add operands, if the add doesn't overflow.
00516   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
00517     if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
00518       SmallVector<const SCEV *, 8> Ops;
00519       for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
00520            I != E; ++I) {
00521         const SCEV *Op = getExactSDiv(*I, RHS, SE,
00522                                       IgnoreSignificantBits);
00523         if (!Op) return 0;
00524         Ops.push_back(Op);
00525       }
00526       return SE.getAddExpr(Ops);
00527     }
00528     return 0;
00529   }
00530 
00531   // Check for a multiply operand that we can pull RHS out of.
00532   if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) {
00533     if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
00534       SmallVector<const SCEV *, 4> Ops;
00535       bool Found = false;
00536       for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
00537            I != E; ++I) {
00538         const SCEV *S = *I;
00539         if (!Found)
00540           if (const SCEV *Q = getExactSDiv(S, RHS, SE,
00541                                            IgnoreSignificantBits)) {
00542             S = Q;
00543             Found = true;
00544           }
00545         Ops.push_back(S);
00546       }
00547       return Found ? SE.getMulExpr(Ops) : 0;
00548     }
00549     return 0;
00550   }
00551 
00552   // Otherwise we don't know.
00553   return 0;
00554 }
00555 
00556 /// ExtractImmediate - If S involves the addition of a constant integer value,
00557 /// return that integer value, and mutate S to point to a new SCEV with that
00558 /// value excluded.
00559 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
00560   if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
00561     if (C->getValue()->getValue().getMinSignedBits() <= 64) {
00562       S = SE.getConstant(C->getType(), 0);
00563       return C->getValue()->getSExtValue();
00564     }
00565   } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
00566     SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
00567     int64_t Result = ExtractImmediate(NewOps.front(), SE);
00568     if (Result != 0)
00569       S = SE.getAddExpr(NewOps);
00570     return Result;
00571   } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
00572     SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
00573     int64_t Result = ExtractImmediate(NewOps.front(), SE);
00574     if (Result != 0)
00575       S = SE.getAddRecExpr(NewOps, AR->getLoop(),
00576                            // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
00577                            SCEV::FlagAnyWrap);
00578     return Result;
00579   }
00580   return 0;
00581 }
00582 
00583 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
00584 /// return that symbol, and mutate S to point to a new SCEV with that
00585 /// value excluded.
00586 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
00587   if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
00588     if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
00589       S = SE.getConstant(GV->getType(), 0);
00590       return GV;
00591     }
00592   } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
00593     SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
00594     GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
00595     if (Result)
00596       S = SE.getAddExpr(NewOps);
00597     return Result;
00598   } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
00599     SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
00600     GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
00601     if (Result)
00602       S = SE.getAddRecExpr(NewOps, AR->getLoop(),
00603                            // FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
00604                            SCEV::FlagAnyWrap);
00605     return Result;
00606   }
00607   return 0;
00608 }
00609 
00610 /// isAddressUse - Returns true if the specified instruction is using the
00611 /// specified value as an address.
00612 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
00613   bool isAddress = isa<LoadInst>(Inst);
00614   if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
00615     if (SI->getOperand(1) == OperandVal)
00616       isAddress = true;
00617   } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
00618     // Addressing modes can also be folded into prefetches and a variety
00619     // of intrinsics.
00620     switch (II->getIntrinsicID()) {
00621       default: break;
00622       case Intrinsic::prefetch:
00623       case Intrinsic::x86_sse_storeu_ps:
00624       case Intrinsic::x86_sse2_storeu_pd:
00625       case Intrinsic::x86_sse2_storeu_dq:
00626       case Intrinsic::x86_sse2_storel_dq:
00627         if (II->getArgOperand(0) == OperandVal)
00628           isAddress = true;
00629         break;
00630     }
00631   }
00632   return isAddress;
00633 }
00634 
00635 /// getAccessType - Return the type of the memory being accessed.
00636 static Type *getAccessType(const Instruction *Inst) {
00637   Type *AccessTy = Inst->getType();
00638   if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
00639     AccessTy = SI->getOperand(0)->getType();
00640   else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
00641     // Addressing modes can also be folded into prefetches and a variety
00642     // of intrinsics.
00643     switch (II->getIntrinsicID()) {
00644     default: break;
00645     case Intrinsic::x86_sse_storeu_ps:
00646     case Intrinsic::x86_sse2_storeu_pd:
00647     case Intrinsic::x86_sse2_storeu_dq:
00648     case Intrinsic::x86_sse2_storel_dq:
00649       AccessTy = II->getArgOperand(0)->getType();
00650       break;
00651     }
00652   }
00653 
00654   // All pointers have the same requirements, so canonicalize them to an
00655   // arbitrary pointer type to minimize variation.
00656   if (PointerType *PTy = dyn_cast<PointerType>(AccessTy))
00657     AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
00658                                 PTy->getAddressSpace());
00659 
00660   return AccessTy;
00661 }
00662 
00663 /// isExistingPhi - Return true if this AddRec is already a phi in its loop.
00664 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
00665   for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
00666        PHINode *PN = dyn_cast<PHINode>(I); ++I) {
00667     if (SE.isSCEVable(PN->getType()) &&
00668         (SE.getEffectiveSCEVType(PN->getType()) ==
00669          SE.getEffectiveSCEVType(AR->getType())) &&
00670         SE.getSCEV(PN) == AR)
00671       return true;
00672   }
00673   return false;
00674 }
00675 
00676 /// Check if expanding this expression is likely to incur significant cost. This
00677 /// is tricky because SCEV doesn't track which expressions are actually computed
00678 /// by the current IR.
00679 ///
00680 /// We currently allow expansion of IV increments that involve adds,
00681 /// multiplication by constants, and AddRecs from existing phis.
00682 ///
00683 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an
00684 /// obvious multiple of the UDivExpr.
00685 static bool isHighCostExpansion(const SCEV *S,
00686                                 SmallPtrSet<const SCEV*, 8> &Processed,
00687                                 ScalarEvolution &SE) {
00688   // Zero/One operand expressions
00689   switch (S->getSCEVType()) {
00690   case scUnknown:
00691   case scConstant:
00692     return false;
00693   case scTruncate:
00694     return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(),
00695                                Processed, SE);
00696   case scZeroExtend:
00697     return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(),
00698                                Processed, SE);
00699   case scSignExtend:
00700     return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(),
00701                                Processed, SE);
00702   }
00703 
00704   if (!Processed.insert(S))
00705     return false;
00706 
00707   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
00708     for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
00709          I != E; ++I) {
00710       if (isHighCostExpansion(*I, Processed, SE))
00711         return true;
00712     }
00713     return false;
00714   }
00715 
00716   if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
00717     if (Mul->getNumOperands() == 2) {
00718       // Multiplication by a constant is ok
00719       if (isa<SCEVConstant>(Mul->getOperand(0)))
00720         return isHighCostExpansion(Mul->getOperand(1), Processed, SE);
00721 
00722       // If we have the value of one operand, check if an existing
00723       // multiplication already generates this expression.
00724       if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) {
00725         Value *UVal = U->getValue();
00726         for (User *UR : UVal->users()) {
00727           // If U is a constant, it may be used by a ConstantExpr.
00728           Instruction *UI = dyn_cast<Instruction>(UR);
00729           if (UI && UI->getOpcode() == Instruction::Mul &&
00730               SE.isSCEVable(UI->getType())) {
00731             return SE.getSCEV(UI) == Mul;
00732           }
00733         }
00734       }
00735     }
00736   }
00737 
00738   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
00739     if (isExistingPhi(AR, SE))
00740       return false;
00741   }
00742 
00743   // Fow now, consider any other type of expression (div/mul/min/max) high cost.
00744   return true;
00745 }
00746 
00747 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
00748 /// specified set are trivially dead, delete them and see if this makes any of
00749 /// their operands subsequently dead.
00750 static bool
00751 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
00752   bool Changed = false;
00753 
00754   while (!DeadInsts.empty()) {
00755     Value *V = DeadInsts.pop_back_val();
00756     Instruction *I = dyn_cast_or_null<Instruction>(V);
00757 
00758     if (I == 0 || !isInstructionTriviallyDead(I))
00759       continue;
00760 
00761     for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
00762       if (Instruction *U = dyn_cast<Instruction>(*OI)) {
00763         *OI = 0;
00764         if (U->use_empty())
00765           DeadInsts.push_back(U);
00766       }
00767 
00768     I->eraseFromParent();
00769     Changed = true;
00770   }
00771 
00772   return Changed;
00773 }
00774 
00775 namespace {
00776 class LSRUse;
00777 }
00778 // Check if it is legal to fold 2 base registers.
00779 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
00780                              const Formula &F);
00781 // Get the cost of the scaling factor used in F for LU.
00782 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
00783                                      const LSRUse &LU, const Formula &F);
00784 
00785 namespace {
00786 
00787 /// Cost - This class is used to measure and compare candidate formulae.
00788 class Cost {
00789   /// TODO: Some of these could be merged. Also, a lexical ordering
00790   /// isn't always optimal.
00791   unsigned NumRegs;
00792   unsigned AddRecCost;
00793   unsigned NumIVMuls;
00794   unsigned NumBaseAdds;
00795   unsigned ImmCost;
00796   unsigned SetupCost;
00797   unsigned ScaleCost;
00798 
00799 public:
00800   Cost()
00801     : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
00802       SetupCost(0), ScaleCost(0) {}
00803 
00804   bool operator<(const Cost &Other) const;
00805 
00806   void Lose();
00807 
00808 #ifndef NDEBUG
00809   // Once any of the metrics loses, they must all remain losers.
00810   bool isValid() {
00811     return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds
00812              | ImmCost | SetupCost | ScaleCost) != ~0u)
00813       || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds
00814            & ImmCost & SetupCost & ScaleCost) == ~0u);
00815   }
00816 #endif
00817 
00818   bool isLoser() {
00819     assert(isValid() && "invalid cost");
00820     return NumRegs == ~0u;
00821   }
00822 
00823   void RateFormula(const TargetTransformInfo &TTI,
00824                    const Formula &F,
00825                    SmallPtrSet<const SCEV *, 16> &Regs,
00826                    const DenseSet<const SCEV *> &VisitedRegs,
00827                    const Loop *L,
00828                    const SmallVectorImpl<int64_t> &Offsets,
00829                    ScalarEvolution &SE, DominatorTree &DT,
00830                    const LSRUse &LU,
00831                    SmallPtrSet<const SCEV *, 16> *LoserRegs = 0);
00832 
00833   void print(raw_ostream &OS) const;
00834   void dump() const;
00835 
00836 private:
00837   void RateRegister(const SCEV *Reg,
00838                     SmallPtrSet<const SCEV *, 16> &Regs,
00839                     const Loop *L,
00840                     ScalarEvolution &SE, DominatorTree &DT);
00841   void RatePrimaryRegister(const SCEV *Reg,
00842                            SmallPtrSet<const SCEV *, 16> &Regs,
00843                            const Loop *L,
00844                            ScalarEvolution &SE, DominatorTree &DT,
00845                            SmallPtrSet<const SCEV *, 16> *LoserRegs);
00846 };
00847 
00848 }
00849 
00850 /// RateRegister - Tally up interesting quantities from the given register.
00851 void Cost::RateRegister(const SCEV *Reg,
00852                         SmallPtrSet<const SCEV *, 16> &Regs,
00853                         const Loop *L,
00854                         ScalarEvolution &SE, DominatorTree &DT) {
00855   if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
00856     // If this is an addrec for another loop, don't second-guess its addrec phi
00857     // nodes. LSR isn't currently smart enough to reason about more than one
00858     // loop at a time. LSR has already run on inner loops, will not run on outer
00859     // loops, and cannot be expected to change sibling loops.
00860     if (AR->getLoop() != L) {
00861       // If the AddRec exists, consider it's register free and leave it alone.
00862       if (isExistingPhi(AR, SE))
00863         return;
00864 
00865       // Otherwise, do not consider this formula at all.
00866       Lose();
00867       return;
00868     }
00869     AddRecCost += 1; /// TODO: This should be a function of the stride.
00870 
00871     // Add the step value register, if it needs one.
00872     // TODO: The non-affine case isn't precisely modeled here.
00873     if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) {
00874       if (!Regs.count(AR->getOperand(1))) {
00875         RateRegister(AR->getOperand(1), Regs, L, SE, DT);
00876         if (isLoser())
00877           return;
00878       }
00879     }
00880   }
00881   ++NumRegs;
00882 
00883   // Rough heuristic; favor registers which don't require extra setup
00884   // instructions in the preheader.
00885   if (!isa<SCEVUnknown>(Reg) &&
00886       !isa<SCEVConstant>(Reg) &&
00887       !(isa<SCEVAddRecExpr>(Reg) &&
00888         (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
00889          isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
00890     ++SetupCost;
00891 
00892     NumIVMuls += isa<SCEVMulExpr>(Reg) &&
00893                  SE.hasComputableLoopEvolution(Reg, L);
00894 }
00895 
00896 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
00897 /// before, rate it. Optional LoserRegs provides a way to declare any formula
00898 /// that refers to one of those regs an instant loser.
00899 void Cost::RatePrimaryRegister(const SCEV *Reg,
00900                                SmallPtrSet<const SCEV *, 16> &Regs,
00901                                const Loop *L,
00902                                ScalarEvolution &SE, DominatorTree &DT,
00903                                SmallPtrSet<const SCEV *, 16> *LoserRegs) {
00904   if (LoserRegs && LoserRegs->count(Reg)) {
00905     Lose();
00906     return;
00907   }
00908   if (Regs.insert(Reg)) {
00909     RateRegister(Reg, Regs, L, SE, DT);
00910     if (LoserRegs && isLoser())
00911       LoserRegs->insert(Reg);
00912   }
00913 }
00914 
00915 void Cost::RateFormula(const TargetTransformInfo &TTI,
00916                        const Formula &F,
00917                        SmallPtrSet<const SCEV *, 16> &Regs,
00918                        const DenseSet<const SCEV *> &VisitedRegs,
00919                        const Loop *L,
00920                        const SmallVectorImpl<int64_t> &Offsets,
00921                        ScalarEvolution &SE, DominatorTree &DT,
00922                        const LSRUse &LU,
00923                        SmallPtrSet<const SCEV *, 16> *LoserRegs) {
00924   // Tally up the registers.
00925   if (const SCEV *ScaledReg = F.ScaledReg) {
00926     if (VisitedRegs.count(ScaledReg)) {
00927       Lose();
00928       return;
00929     }
00930     RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs);
00931     if (isLoser())
00932       return;
00933   }
00934   for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
00935        E = F.BaseRegs.end(); I != E; ++I) {
00936     const SCEV *BaseReg = *I;
00937     if (VisitedRegs.count(BaseReg)) {
00938       Lose();
00939       return;
00940     }
00941     RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs);
00942     if (isLoser())
00943       return;
00944   }
00945 
00946   // Determine how many (unfolded) adds we'll need inside the loop.
00947   size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0);
00948   if (NumBaseParts > 1)
00949     // Do not count the base and a possible second register if the target
00950     // allows to fold 2 registers.
00951     NumBaseAdds += NumBaseParts - (1 + isLegal2RegAMUse(TTI, LU, F));
00952 
00953   // Accumulate non-free scaling amounts.
00954   ScaleCost += getScalingFactorCost(TTI, LU, F);
00955 
00956   // Tally up the non-zero immediates.
00957   for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
00958        E = Offsets.end(); I != E; ++I) {
00959     int64_t Offset = (uint64_t)*I + F.BaseOffset;
00960     if (F.BaseGV)
00961       ImmCost += 64; // Handle symbolic values conservatively.
00962                      // TODO: This should probably be the pointer size.
00963     else if (Offset != 0)
00964       ImmCost += APInt(64, Offset, true).getMinSignedBits();
00965   }
00966   assert(isValid() && "invalid cost");
00967 }
00968 
00969 /// Lose - Set this cost to a losing value.
00970 void Cost::Lose() {
00971   NumRegs = ~0u;
00972   AddRecCost = ~0u;
00973   NumIVMuls = ~0u;
00974   NumBaseAdds = ~0u;
00975   ImmCost = ~0u;
00976   SetupCost = ~0u;
00977   ScaleCost = ~0u;
00978 }
00979 
00980 /// operator< - Choose the lower cost.
00981 bool Cost::operator<(const Cost &Other) const {
00982   return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost,
00983                   ImmCost, SetupCost) <
00984          std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls,
00985                   Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost,
00986                   Other.SetupCost);
00987 }
00988 
00989 void Cost::print(raw_ostream &OS) const {
00990   OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
00991   if (AddRecCost != 0)
00992     OS << ", with addrec cost " << AddRecCost;
00993   if (NumIVMuls != 0)
00994     OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
00995   if (NumBaseAdds != 0)
00996     OS << ", plus " << NumBaseAdds << " base add"
00997        << (NumBaseAdds == 1 ? "" : "s");
00998   if (ScaleCost != 0)
00999     OS << ", plus " << ScaleCost << " scale cost";
01000   if (ImmCost != 0)
01001     OS << ", plus " << ImmCost << " imm cost";
01002   if (SetupCost != 0)
01003     OS << ", plus " << SetupCost << " setup cost";
01004 }
01005 
01006 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
01007 void Cost::dump() const {
01008   print(errs()); errs() << '\n';
01009 }
01010 #endif
01011 
01012 namespace {
01013 
01014 /// LSRFixup - An operand value in an instruction which is to be replaced
01015 /// with some equivalent, possibly strength-reduced, replacement.
01016 struct LSRFixup {
01017   /// UserInst - The instruction which will be updated.
01018   Instruction *UserInst;
01019 
01020   /// OperandValToReplace - The operand of the instruction which will
01021   /// be replaced. The operand may be used more than once; every instance
01022   /// will be replaced.
01023   Value *OperandValToReplace;
01024 
01025   /// PostIncLoops - If this user is to use the post-incremented value of an
01026   /// induction variable, this variable is non-null and holds the loop
01027   /// associated with the induction variable.
01028   PostIncLoopSet PostIncLoops;
01029 
01030   /// LUIdx - The index of the LSRUse describing the expression which
01031   /// this fixup needs, minus an offset (below).
01032   size_t LUIdx;
01033 
01034   /// Offset - A constant offset to be added to the LSRUse expression.
01035   /// This allows multiple fixups to share the same LSRUse with different
01036   /// offsets, for example in an unrolled loop.
01037   int64_t Offset;
01038 
01039   bool isUseFullyOutsideLoop(const Loop *L) const;
01040 
01041   LSRFixup();
01042 
01043   void print(raw_ostream &OS) const;
01044   void dump() const;
01045 };
01046 
01047 }
01048 
01049 LSRFixup::LSRFixup()
01050   : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {}
01051 
01052 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
01053 /// value outside of the given loop.
01054 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
01055   // PHI nodes use their value in their incoming blocks.
01056   if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
01057     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
01058       if (PN->getIncomingValue(i) == OperandValToReplace &&
01059           L->contains(PN->getIncomingBlock(i)))
01060         return false;
01061     return true;
01062   }
01063 
01064   return !L->contains(UserInst);
01065 }
01066 
01067 void LSRFixup::print(raw_ostream &OS) const {
01068   OS << "UserInst=";
01069   // Store is common and interesting enough to be worth special-casing.
01070   if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
01071     OS << "store ";
01072     Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false);
01073   } else if (UserInst->getType()->isVoidTy())
01074     OS << UserInst->getOpcodeName();
01075   else
01076     UserInst->printAsOperand(OS, /*PrintType=*/false);
01077 
01078   OS << ", OperandValToReplace=";
01079   OperandValToReplace->printAsOperand(OS, /*PrintType=*/false);
01080 
01081   for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
01082        E = PostIncLoops.end(); I != E; ++I) {
01083     OS << ", PostIncLoop=";
01084     (*I)->getHeader()->printAsOperand(OS, /*PrintType=*/false);
01085   }
01086 
01087   if (LUIdx != ~size_t(0))
01088     OS << ", LUIdx=" << LUIdx;
01089 
01090   if (Offset != 0)
01091     OS << ", Offset=" << Offset;
01092 }
01093 
01094 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
01095 void LSRFixup::dump() const {
01096   print(errs()); errs() << '\n';
01097 }
01098 #endif
01099 
01100 namespace {
01101 
01102 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
01103 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
01104 struct UniquifierDenseMapInfo {
01105   static SmallVector<const SCEV *, 4> getEmptyKey() {
01106     SmallVector<const SCEV *, 4>  V;
01107     V.push_back(reinterpret_cast<const SCEV *>(-1));
01108     return V;
01109   }
01110 
01111   static SmallVector<const SCEV *, 4> getTombstoneKey() {
01112     SmallVector<const SCEV *, 4> V;
01113     V.push_back(reinterpret_cast<const SCEV *>(-2));
01114     return V;
01115   }
01116 
01117   static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) {
01118     return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
01119   }
01120 
01121   static bool isEqual(const SmallVector<const SCEV *, 4> &LHS,
01122                       const SmallVector<const SCEV *, 4> &RHS) {
01123     return LHS == RHS;
01124   }
01125 };
01126 
01127 /// LSRUse - This class holds the state that LSR keeps for each use in
01128 /// IVUsers, as well as uses invented by LSR itself. It includes information
01129 /// about what kinds of things can be folded into the user, information about
01130 /// the user itself, and information about how the use may be satisfied.
01131 /// TODO: Represent multiple users of the same expression in common?
01132 class LSRUse {
01133   DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier;
01134 
01135 public:
01136   /// KindType - An enum for a kind of use, indicating what types of
01137   /// scaled and immediate operands it might support.
01138   enum KindType {
01139     Basic,   ///< A normal use, with no folding.
01140     Special, ///< A special case of basic, allowing -1 scales.
01141     Address, ///< An address use; folding according to TargetLowering
01142     ICmpZero ///< An equality icmp with both operands folded into one.
01143     // TODO: Add a generic icmp too?
01144   };
01145 
01146   typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair;
01147 
01148   KindType Kind;
01149   Type *AccessTy;
01150 
01151   SmallVector<int64_t, 8> Offsets;
01152   int64_t MinOffset;
01153   int64_t MaxOffset;
01154 
01155   /// AllFixupsOutsideLoop - This records whether all of the fixups using this
01156   /// LSRUse are outside of the loop, in which case some special-case heuristics
01157   /// may be used.
01158   bool AllFixupsOutsideLoop;
01159 
01160   /// RigidFormula is set to true to guarantee that this use will be associated
01161   /// with a single formula--the one that initially matched. Some SCEV
01162   /// expressions cannot be expanded. This allows LSR to consider the registers
01163   /// used by those expressions without the need to expand them later after
01164   /// changing the formula.
01165   bool RigidFormula;
01166 
01167   /// WidestFixupType - This records the widest use type for any fixup using
01168   /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
01169   /// max fixup widths to be equivalent, because the narrower one may be relying
01170   /// on the implicit truncation to truncate away bogus bits.
01171   Type *WidestFixupType;
01172 
01173   /// Formulae - A list of ways to build a value that can satisfy this user.
01174   /// After the list is populated, one of these is selected heuristically and
01175   /// used to formulate a replacement for OperandValToReplace in UserInst.
01176   SmallVector<Formula, 12> Formulae;
01177 
01178   /// Regs - The set of register candidates used by all formulae in this LSRUse.
01179   SmallPtrSet<const SCEV *, 4> Regs;
01180 
01181   LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T),
01182                                       MinOffset(INT64_MAX),
01183                                       MaxOffset(INT64_MIN),
01184                                       AllFixupsOutsideLoop(true),
01185                                       RigidFormula(false),
01186                                       WidestFixupType(0) {}
01187 
01188   bool HasFormulaWithSameRegs(const Formula &F) const;
01189   bool InsertFormula(const Formula &F);
01190   void DeleteFormula(Formula &F);
01191   void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses);
01192 
01193   void print(raw_ostream &OS) const;
01194   void dump() const;
01195 };
01196 
01197 }
01198 
01199 /// HasFormula - Test whether this use as a formula which has the same
01200 /// registers as the given formula.
01201 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
01202   SmallVector<const SCEV *, 4> Key = F.BaseRegs;
01203   if (F.ScaledReg) Key.push_back(F.ScaledReg);
01204   // Unstable sort by host order ok, because this is only used for uniquifying.
01205   std::sort(Key.begin(), Key.end());
01206   return Uniquifier.count(Key);
01207 }
01208 
01209 /// InsertFormula - If the given formula has not yet been inserted, add it to
01210 /// the list, and return true. Return false otherwise.
01211 bool LSRUse::InsertFormula(const Formula &F) {
01212   if (!Formulae.empty() && RigidFormula)
01213     return false;
01214 
01215   SmallVector<const SCEV *, 4> Key = F.BaseRegs;
01216   if (F.ScaledReg) Key.push_back(F.ScaledReg);
01217   // Unstable sort by host order ok, because this is only used for uniquifying.
01218   std::sort(Key.begin(), Key.end());
01219 
01220   if (!Uniquifier.insert(Key).second)
01221     return false;
01222 
01223   // Using a register to hold the value of 0 is not profitable.
01224   assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
01225          "Zero allocated in a scaled register!");
01226 #ifndef NDEBUG
01227   for (SmallVectorImpl<const SCEV *>::const_iterator I =
01228        F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
01229     assert(!(*I)->isZero() && "Zero allocated in a base register!");
01230 #endif
01231 
01232   // Add the formula to the list.
01233   Formulae.push_back(F);
01234 
01235   // Record registers now being used by this use.
01236   Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
01237 
01238   return true;
01239 }
01240 
01241 /// DeleteFormula - Remove the given formula from this use's list.
01242 void LSRUse::DeleteFormula(Formula &F) {
01243   if (&F != &Formulae.back())
01244     std::swap(F, Formulae.back());
01245   Formulae.pop_back();
01246 }
01247 
01248 /// RecomputeRegs - Recompute the Regs field, and update RegUses.
01249 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) {
01250   // Now that we've filtered out some formulae, recompute the Regs set.
01251   SmallPtrSet<const SCEV *, 4> OldRegs = Regs;
01252   Regs.clear();
01253   for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(),
01254        E = Formulae.end(); I != E; ++I) {
01255     const Formula &F = *I;
01256     if (F.ScaledReg) Regs.insert(F.ScaledReg);
01257     Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
01258   }
01259 
01260   // Update the RegTracker.
01261   for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(),
01262        E = OldRegs.end(); I != E; ++I)
01263     if (!Regs.count(*I))
01264       RegUses.DropRegister(*I, LUIdx);
01265 }
01266 
01267 void LSRUse::print(raw_ostream &OS) const {
01268   OS << "LSR Use: Kind=";
01269   switch (Kind) {
01270   case Basic:    OS << "Basic"; break;
01271   case Special:  OS << "Special"; break;
01272   case ICmpZero: OS << "ICmpZero"; break;
01273   case Address:
01274     OS << "Address of ";
01275     if (AccessTy->isPointerTy())
01276       OS << "pointer"; // the full pointer type could be really verbose
01277     else
01278       OS << *AccessTy;
01279   }
01280 
01281   OS << ", Offsets={";
01282   for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
01283        E = Offsets.end(); I != E; ++I) {
01284     OS << *I;
01285     if (std::next(I) != E)
01286       OS << ',';
01287   }
01288   OS << '}';
01289 
01290   if (AllFixupsOutsideLoop)
01291     OS << ", all-fixups-outside-loop";
01292 
01293   if (WidestFixupType)
01294     OS << ", widest fixup type: " << *WidestFixupType;
01295 }
01296 
01297 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
01298 void LSRUse::dump() const {
01299   print(errs()); errs() << '\n';
01300 }
01301 #endif
01302 
01303 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
01304 /// be completely folded into the user instruction at isel time. This includes
01305 /// address-mode folding and special icmp tricks.
01306 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind,
01307                        Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset,
01308                        bool HasBaseReg, int64_t Scale) {
01309   switch (Kind) {
01310   case LSRUse::Address:
01311     return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
01312 
01313     // Otherwise, just guess that reg+reg addressing is legal.
01314     //return ;
01315 
01316   case LSRUse::ICmpZero:
01317     // There's not even a target hook for querying whether it would be legal to
01318     // fold a GV into an ICmp.
01319     if (BaseGV)
01320       return false;
01321 
01322     // ICmp only has two operands; don't allow more than two non-trivial parts.
01323     if (Scale != 0 && HasBaseReg && BaseOffset != 0)
01324       return false;
01325 
01326     // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
01327     // putting the scaled register in the other operand of the icmp.
01328     if (Scale != 0 && Scale != -1)
01329       return false;
01330 
01331     // If we have low-level target information, ask the target if it can fold an
01332     // integer immediate on an icmp.
01333     if (BaseOffset != 0) {
01334       // We have one of:
01335       // ICmpZero     BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset
01336       // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset
01337       // Offs is the ICmp immediate.
01338       if (Scale == 0)
01339         // The cast does the right thing with INT64_MIN.
01340         BaseOffset = -(uint64_t)BaseOffset;
01341       return TTI.isLegalICmpImmediate(BaseOffset);
01342     }
01343 
01344     // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
01345     return true;
01346 
01347   case LSRUse::Basic:
01348     // Only handle single-register values.
01349     return !BaseGV && Scale == 0 && BaseOffset == 0;
01350 
01351   case LSRUse::Special:
01352     // Special case Basic to handle -1 scales.
01353     return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0;
01354   }
01355 
01356   llvm_unreachable("Invalid LSRUse Kind!");
01357 }
01358 
01359 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
01360                        int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
01361                        GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg,
01362                        int64_t Scale) {
01363   // Check for overflow.
01364   if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) !=
01365       (MinOffset > 0))
01366     return false;
01367   MinOffset = (uint64_t)BaseOffset + MinOffset;
01368   if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) !=
01369       (MaxOffset > 0))
01370     return false;
01371   MaxOffset = (uint64_t)BaseOffset + MaxOffset;
01372 
01373   return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg,
01374                     Scale) &&
01375          isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale);
01376 }
01377 
01378 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset,
01379                        int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy,
01380                        const Formula &F) {
01381   return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV,
01382                     F.BaseOffset, F.HasBaseReg, F.Scale);
01383 }
01384 
01385 static bool isLegal2RegAMUse(const TargetTransformInfo &TTI, const LSRUse &LU,
01386                              const Formula &F) {
01387   // If F is used as an Addressing Mode, it may fold one Base plus one
01388   // scaled register. If the scaled register is nil, do as if another
01389   // element of the base regs is a 1-scaled register.
01390   // This is possible if BaseRegs has at least 2 registers.
01391 
01392   // If this is not an address calculation, this is not an addressing mode
01393   // use.
01394   if (LU.Kind !=  LSRUse::Address)
01395     return false;
01396 
01397   // F is already scaled.
01398   if (F.Scale != 0)
01399     return false;
01400 
01401   // We need to keep one register for the base and one to scale.
01402   if (F.BaseRegs.size() < 2)
01403     return false;
01404 
01405   return isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
01406                     F.BaseGV, F.BaseOffset, F.HasBaseReg, 1);
01407  }
01408 
01409 static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
01410                                      const LSRUse &LU, const Formula &F) {
01411   if (!F.Scale)
01412     return 0;
01413   assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind,
01414                     LU.AccessTy, F) && "Illegal formula in use.");
01415 
01416   switch (LU.Kind) {
01417   case LSRUse::Address: {
01418     // Check the scaling factor cost with both the min and max offsets.
01419     int ScaleCostMinOffset =
01420       TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
01421                                F.BaseOffset + LU.MinOffset,
01422                                F.HasBaseReg, F.Scale);
01423     int ScaleCostMaxOffset =
01424       TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV,
01425                                F.BaseOffset + LU.MaxOffset,
01426                                F.HasBaseReg, F.Scale);
01427 
01428     assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 &&
01429            "Legal addressing mode has an illegal cost!");
01430     return std::max(ScaleCostMinOffset, ScaleCostMaxOffset);
01431   }
01432   case LSRUse::ICmpZero:
01433     // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
01434     // Therefore, return 0 in case F.Scale == -1.
01435     return F.Scale != -1;
01436 
01437   case LSRUse::Basic:
01438   case LSRUse::Special:
01439     return 0;
01440   }
01441 
01442   llvm_unreachable("Invalid LSRUse Kind!");
01443 }
01444 
01445 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
01446                              LSRUse::KindType Kind, Type *AccessTy,
01447                              GlobalValue *BaseGV, int64_t BaseOffset,
01448                              bool HasBaseReg) {
01449   // Fast-path: zero is always foldable.
01450   if (BaseOffset == 0 && !BaseGV) return true;
01451 
01452   // Conservatively, create an address with an immediate and a
01453   // base and a scale.
01454   int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
01455 
01456   // Canonicalize a scale of 1 to a base register if the formula doesn't
01457   // already have a base register.
01458   if (!HasBaseReg && Scale == 1) {
01459     Scale = 0;
01460     HasBaseReg = true;
01461   }
01462 
01463   return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale);
01464 }
01465 
01466 static bool isAlwaysFoldable(const TargetTransformInfo &TTI,
01467                              ScalarEvolution &SE, int64_t MinOffset,
01468                              int64_t MaxOffset, LSRUse::KindType Kind,
01469                              Type *AccessTy, const SCEV *S, bool HasBaseReg) {
01470   // Fast-path: zero is always foldable.
01471   if (S->isZero()) return true;
01472 
01473   // Conservatively, create an address with an immediate and a
01474   // base and a scale.
01475   int64_t BaseOffset = ExtractImmediate(S, SE);
01476   GlobalValue *BaseGV = ExtractSymbol(S, SE);
01477 
01478   // If there's anything else involved, it's not foldable.
01479   if (!S->isZero()) return false;
01480 
01481   // Fast-path: zero is always foldable.
01482   if (BaseOffset == 0 && !BaseGV) return true;
01483 
01484   // Conservatively, create an address with an immediate and a
01485   // base and a scale.
01486   int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
01487 
01488   return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV,
01489                     BaseOffset, HasBaseReg, Scale);
01490 }
01491 
01492 namespace {
01493 
01494 /// IVInc - An individual increment in a Chain of IV increments.
01495 /// Relate an IV user to an expression that computes the IV it uses from the IV
01496 /// used by the previous link in the Chain.
01497 ///
01498 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the
01499 /// original IVOperand. The head of the chain's IVOperand is only valid during
01500 /// chain collection, before LSR replaces IV users. During chain generation,
01501 /// IncExpr can be used to find the new IVOperand that computes the same
01502 /// expression.
01503 struct IVInc {
01504   Instruction *UserInst;
01505   Value* IVOperand;
01506   const SCEV *IncExpr;
01507 
01508   IVInc(Instruction *U, Value *O, const SCEV *E):
01509     UserInst(U), IVOperand(O), IncExpr(E) {}
01510 };
01511 
01512 // IVChain - The list of IV increments in program order.
01513 // We typically add the head of a chain without finding subsequent links.
01514 struct IVChain {
01515   SmallVector<IVInc,1> Incs;
01516   const SCEV *ExprBase;
01517 
01518   IVChain() : ExprBase(0) {}
01519 
01520   IVChain(const IVInc &Head, const SCEV *Base)
01521     : Incs(1, Head), ExprBase(Base) {}
01522 
01523   typedef SmallVectorImpl<IVInc>::const_iterator const_iterator;
01524 
01525   // begin - return the first increment in the chain.
01526   const_iterator begin() const {
01527     assert(!Incs.empty());
01528     return std::next(Incs.begin());
01529   }
01530   const_iterator end() const {
01531     return Incs.end();
01532   }
01533 
01534   // hasIncs - Returns true if this chain contains any increments.
01535   bool hasIncs() const { return Incs.size() >= 2; }
01536 
01537   // add - Add an IVInc to the end of this chain.
01538   void add(const IVInc &X) { Incs.push_back(X); }
01539 
01540   // tailUserInst - Returns the last UserInst in the chain.
01541   Instruction *tailUserInst() const { return Incs.back().UserInst; }
01542 
01543   // isProfitableIncrement - Returns true if IncExpr can be profitably added to
01544   // this chain.
01545   bool isProfitableIncrement(const SCEV *OperExpr,
01546                              const SCEV *IncExpr,
01547                              ScalarEvolution&);
01548 };
01549 
01550 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses.
01551 /// Distinguish between FarUsers that definitely cross IV increments and
01552 /// NearUsers that may be used between IV increments.
01553 struct ChainUsers {
01554   SmallPtrSet<Instruction*, 4> FarUsers;
01555   SmallPtrSet<Instruction*, 4> NearUsers;
01556 };
01557 
01558 /// LSRInstance - This class holds state for the main loop strength reduction
01559 /// logic.
01560 class LSRInstance {
01561   IVUsers &IU;
01562   ScalarEvolution &SE;
01563   DominatorTree &DT;
01564   LoopInfo &LI;
01565   const TargetTransformInfo &TTI;
01566   Loop *const L;
01567   bool Changed;
01568 
01569   /// IVIncInsertPos - This is the insert position that the current loop's
01570   /// induction variable increment should be placed. In simple loops, this is
01571   /// the latch block's terminator. But in more complicated cases, this is a
01572   /// position which will dominate all the in-loop post-increment users.
01573   Instruction *IVIncInsertPos;
01574 
01575   /// Factors - Interesting factors between use strides.
01576   SmallSetVector<int64_t, 8> Factors;
01577 
01578   /// Types - Interesting use types, to facilitate truncation reuse.
01579   SmallSetVector<Type *, 4> Types;
01580 
01581   /// Fixups - The list of operands which are to be replaced.
01582   SmallVector<LSRFixup, 16> Fixups;
01583 
01584   /// Uses - The list of interesting uses.
01585   SmallVector<LSRUse, 16> Uses;
01586 
01587   /// RegUses - Track which uses use which register candidates.
01588   RegUseTracker RegUses;
01589 
01590   // Limit the number of chains to avoid quadratic behavior. We don't expect to
01591   // have more than a few IV increment chains in a loop. Missing a Chain falls
01592   // back to normal LSR behavior for those uses.
01593   static const unsigned MaxChains = 8;
01594 
01595   /// IVChainVec - IV users can form a chain of IV increments.
01596   SmallVector<IVChain, MaxChains> IVChainVec;
01597 
01598   /// IVIncSet - IV users that belong to profitable IVChains.
01599   SmallPtrSet<Use*, MaxChains> IVIncSet;
01600 
01601   void OptimizeShadowIV();
01602   bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
01603   ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
01604   void OptimizeLoopTermCond();
01605 
01606   void ChainInstruction(Instruction *UserInst, Instruction *IVOper,
01607                         SmallVectorImpl<ChainUsers> &ChainUsersVec);
01608   void FinalizeChain(IVChain &Chain);
01609   void CollectChains();
01610   void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
01611                        SmallVectorImpl<WeakVH> &DeadInsts);
01612 
01613   void CollectInterestingTypesAndFactors();
01614   void CollectFixupsAndInitialFormulae();
01615 
01616   LSRFixup &getNewFixup() {
01617     Fixups.push_back(LSRFixup());
01618     return Fixups.back();
01619   }
01620 
01621   // Support for sharing of LSRUses between LSRFixups.
01622   typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy;
01623   UseMapTy UseMap;
01624 
01625   bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
01626                           LSRUse::KindType Kind, Type *AccessTy);
01627 
01628   std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
01629                                     LSRUse::KindType Kind,
01630                                     Type *AccessTy);
01631 
01632   void DeleteUse(LSRUse &LU, size_t LUIdx);
01633 
01634   LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
01635 
01636   void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
01637   void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
01638   void CountRegisters(const Formula &F, size_t LUIdx);
01639   bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
01640 
01641   void CollectLoopInvariantFixupsAndFormulae();
01642 
01643   void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
01644                               unsigned Depth = 0);
01645   void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
01646   void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
01647   void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
01648   void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
01649   void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
01650   void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
01651   void GenerateCrossUseConstantOffsets();
01652   void GenerateAllReuseFormulae();
01653 
01654   void FilterOutUndesirableDedicatedRegisters();
01655 
01656   size_t EstimateSearchSpaceComplexity() const;
01657   void NarrowSearchSpaceByDetectingSupersets();
01658   void NarrowSearchSpaceByCollapsingUnrolledCode();
01659   void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
01660   void NarrowSearchSpaceByPickingWinnerRegs();
01661   void NarrowSearchSpaceUsingHeuristics();
01662 
01663   void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
01664                     Cost &SolutionCost,
01665                     SmallVectorImpl<const Formula *> &Workspace,
01666                     const Cost &CurCost,
01667                     const SmallPtrSet<const SCEV *, 16> &CurRegs,
01668                     DenseSet<const SCEV *> &VisitedRegs) const;
01669   void Solve(SmallVectorImpl<const Formula *> &Solution) const;
01670 
01671   BasicBlock::iterator
01672     HoistInsertPosition(BasicBlock::iterator IP,
01673                         const SmallVectorImpl<Instruction *> &Inputs) const;
01674   BasicBlock::iterator
01675     AdjustInsertPositionForExpand(BasicBlock::iterator IP,
01676                                   const LSRFixup &LF,
01677                                   const LSRUse &LU,
01678                                   SCEVExpander &Rewriter) const;
01679 
01680   Value *Expand(const LSRFixup &LF,
01681                 const Formula &F,
01682                 BasicBlock::iterator IP,
01683                 SCEVExpander &Rewriter,
01684                 SmallVectorImpl<WeakVH> &DeadInsts) const;
01685   void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
01686                      const Formula &F,
01687                      SCEVExpander &Rewriter,
01688                      SmallVectorImpl<WeakVH> &DeadInsts,
01689                      Pass *P) const;
01690   void Rewrite(const LSRFixup &LF,
01691                const Formula &F,
01692                SCEVExpander &Rewriter,
01693                SmallVectorImpl<WeakVH> &DeadInsts,
01694                Pass *P) const;
01695   void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
01696                          Pass *P);
01697 
01698 public:
01699   LSRInstance(Loop *L, Pass *P);
01700 
01701   bool getChanged() const { return Changed; }
01702 
01703   void print_factors_and_types(raw_ostream &OS) const;
01704   void print_fixups(raw_ostream &OS) const;
01705   void print_uses(raw_ostream &OS) const;
01706   void print(raw_ostream &OS) const;
01707   void dump() const;
01708 };
01709 
01710 }
01711 
01712 /// OptimizeShadowIV - If IV is used in a int-to-float cast
01713 /// inside the loop then try to eliminate the cast operation.
01714 void LSRInstance::OptimizeShadowIV() {
01715   const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
01716   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
01717     return;
01718 
01719   for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
01720        UI != E; /* empty */) {
01721     IVUsers::const_iterator CandidateUI = UI;
01722     ++UI;
01723     Instruction *ShadowUse = CandidateUI->getUser();
01724     Type *DestTy = 0;
01725     bool IsSigned = false;
01726 
01727     /* If shadow use is a int->float cast then insert a second IV
01728        to eliminate this cast.
01729 
01730          for (unsigned i = 0; i < n; ++i)
01731            foo((double)i);
01732 
01733        is transformed into
01734 
01735          double d = 0.0;
01736          for (unsigned i = 0; i < n; ++i, ++d)
01737            foo(d);
01738     */
01739     if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) {
01740       IsSigned = false;
01741       DestTy = UCast->getDestTy();
01742     }
01743     else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) {
01744       IsSigned = true;
01745       DestTy = SCast->getDestTy();
01746     }
01747     if (!DestTy) continue;
01748 
01749     // If target does not support DestTy natively then do not apply
01750     // this transformation.
01751     if (!TTI.isTypeLegal(DestTy)) continue;
01752 
01753     PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
01754     if (!PH) continue;
01755     if (PH->getNumIncomingValues() != 2) continue;
01756 
01757     Type *SrcTy = PH->getType();
01758     int Mantissa = DestTy->getFPMantissaWidth();
01759     if (Mantissa == -1) continue;
01760     if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
01761       continue;
01762 
01763     unsigned Entry, Latch;
01764     if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
01765       Entry = 0;
01766       Latch = 1;
01767     } else {
01768       Entry = 1;
01769       Latch = 0;
01770     }
01771 
01772     ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
01773     if (!Init) continue;
01774     Constant *NewInit = ConstantFP::get(DestTy, IsSigned ?
01775                                         (double)Init->getSExtValue() :
01776                                         (double)Init->getZExtValue());
01777 
01778     BinaryOperator *Incr =
01779       dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
01780     if (!Incr) continue;
01781     if (Incr->getOpcode() != Instruction::Add
01782         && Incr->getOpcode() != Instruction::Sub)
01783       continue;
01784 
01785     /* Initialize new IV, double d = 0.0 in above example. */
01786     ConstantInt *C = 0;
01787     if (Incr->getOperand(0) == PH)
01788       C = dyn_cast<ConstantInt>(Incr->getOperand(1));
01789     else if (Incr->getOperand(1) == PH)
01790       C = dyn_cast<ConstantInt>(Incr->getOperand(0));
01791     else
01792       continue;
01793 
01794     if (!C) continue;
01795 
01796     // Ignore negative constants, as the code below doesn't handle them
01797     // correctly. TODO: Remove this restriction.
01798     if (!C->getValue().isStrictlyPositive()) continue;
01799 
01800     /* Add new PHINode. */
01801     PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH);
01802 
01803     /* create new increment. '++d' in above example. */
01804     Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
01805     BinaryOperator *NewIncr =
01806       BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
01807                                Instruction::FAdd : Instruction::FSub,
01808                              NewPH, CFP, "IV.S.next.", Incr);
01809 
01810     NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
01811     NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
01812 
01813     /* Remove cast operation */
01814     ShadowUse->replaceAllUsesWith(NewPH);
01815     ShadowUse->eraseFromParent();
01816     Changed = true;
01817     break;
01818   }
01819 }
01820 
01821 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
01822 /// set the IV user and stride information and return true, otherwise return
01823 /// false.
01824 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) {
01825   for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
01826     if (UI->getUser() == Cond) {
01827       // NOTE: we could handle setcc instructions with multiple uses here, but
01828       // InstCombine does it as well for simple uses, it's not clear that it
01829       // occurs enough in real life to handle.
01830       CondUse = UI;
01831       return true;
01832     }
01833   return false;
01834 }
01835 
01836 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
01837 /// a max computation.
01838 ///
01839 /// This is a narrow solution to a specific, but acute, problem. For loops
01840 /// like this:
01841 ///
01842 ///   i = 0;
01843 ///   do {
01844 ///     p[i] = 0.0;
01845 ///   } while (++i < n);
01846 ///
01847 /// the trip count isn't just 'n', because 'n' might not be positive. And
01848 /// unfortunately this can come up even for loops where the user didn't use
01849 /// a C do-while loop. For example, seemingly well-behaved top-test loops
01850 /// will commonly be lowered like this:
01851 //
01852 ///   if (n > 0) {
01853 ///     i = 0;
01854 ///     do {
01855 ///       p[i] = 0.0;
01856 ///     } while (++i < n);
01857 ///   }
01858 ///
01859 /// and then it's possible for subsequent optimization to obscure the if
01860 /// test in such a way that indvars can't find it.
01861 ///
01862 /// When indvars can't find the if test in loops like this, it creates a
01863 /// max expression, which allows it to give the loop a canonical
01864 /// induction variable:
01865 ///
01866 ///   i = 0;
01867 ///   max = n < 1 ? 1 : n;
01868 ///   do {
01869 ///     p[i] = 0.0;
01870 ///   } while (++i != max);
01871 ///
01872 /// Canonical induction variables are necessary because the loop passes
01873 /// are designed around them. The most obvious example of this is the
01874 /// LoopInfo analysis, which doesn't remember trip count values. It
01875 /// expects to be able to rediscover the trip count each time it is
01876 /// needed, and it does this using a simple analysis that only succeeds if
01877 /// the loop has a canonical induction variable.
01878 ///
01879 /// However, when it comes time to generate code, the maximum operation
01880 /// can be quite costly, especially if it's inside of an outer loop.
01881 ///
01882 /// This function solves this problem by detecting this type of loop and
01883 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
01884 /// the instructions for the maximum computation.
01885 ///
01886 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
01887   // Check that the loop matches the pattern we're looking for.
01888   if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
01889       Cond->getPredicate() != CmpInst::ICMP_NE)
01890     return Cond;
01891 
01892   SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
01893   if (!Sel || !Sel->hasOneUse()) return Cond;
01894 
01895   const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
01896   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
01897     return Cond;
01898   const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1);
01899 
01900   // Add one to the backedge-taken count to get the trip count.
01901   const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount);
01902   if (IterationCount != SE.getSCEV(Sel)) return Cond;
01903 
01904   // Check for a max calculation that matches the pattern. There's no check
01905   // for ICMP_ULE here because the comparison would be with zero, which
01906   // isn't interesting.
01907   CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE;
01908   const SCEVNAryExpr *Max = 0;
01909   if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) {
01910     Pred = ICmpInst::ICMP_SLE;
01911     Max = S;
01912   } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) {
01913     Pred = ICmpInst::ICMP_SLT;
01914     Max = S;
01915   } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) {
01916     Pred = ICmpInst::ICMP_ULT;
01917     Max = U;
01918   } else {
01919     // No match; bail.
01920     return Cond;
01921   }
01922 
01923   // To handle a max with more than two operands, this optimization would
01924   // require additional checking and setup.
01925   if (Max->getNumOperands() != 2)
01926     return Cond;
01927 
01928   const SCEV *MaxLHS = Max->getOperand(0);
01929   const SCEV *MaxRHS = Max->getOperand(1);
01930 
01931   // ScalarEvolution canonicalizes constants to the left. For < and >, look
01932   // for a comparison with 1. For <= and >=, a comparison with zero.
01933   if (!MaxLHS ||
01934       (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One)))
01935     return Cond;
01936 
01937   // Check the relevant induction variable for conformance to
01938   // the pattern.
01939   const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
01940   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
01941   if (!AR || !AR->isAffine() ||
01942       AR->getStart() != One ||
01943       AR->getStepRecurrence(SE) != One)
01944     return Cond;
01945 
01946   assert(AR->getLoop() == L &&
01947          "Loop condition operand is an addrec in a different loop!");
01948 
01949   // Check the right operand of the select, and remember it, as it will
01950   // be used in the new comparison instruction.
01951   Value *NewRHS = 0;
01952   if (ICmpInst::isTrueWhenEqual(Pred)) {
01953     // Look for n+1, and grab n.
01954     if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1)))
01955       if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
01956          if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
01957            NewRHS = BO->getOperand(0);
01958     if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2)))
01959       if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1)))
01960         if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS)
01961           NewRHS = BO->getOperand(0);
01962     if (!NewRHS)
01963       return Cond;
01964   } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
01965     NewRHS = Sel->getOperand(1);
01966   else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
01967     NewRHS = Sel->getOperand(2);
01968   else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS))
01969     NewRHS = SU->getValue();
01970   else
01971     // Max doesn't match expected pattern.
01972     return Cond;
01973 
01974   // Determine the new comparison opcode. It may be signed or unsigned,
01975   // and the original comparison may be either equality or inequality.
01976   if (Cond->getPredicate() == CmpInst::ICMP_EQ)
01977     Pred = CmpInst::getInversePredicate(Pred);
01978 
01979   // Ok, everything looks ok to change the condition into an SLT or SGE and
01980   // delete the max calculation.
01981   ICmpInst *NewCond =
01982     new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
01983 
01984   // Delete the max calculation instructions.
01985   Cond->replaceAllUsesWith(NewCond);
01986   CondUse->setUser(NewCond);
01987   Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
01988   Cond->eraseFromParent();
01989   Sel->eraseFromParent();
01990   if (Cmp->use_empty())
01991     Cmp->eraseFromParent();
01992   return NewCond;
01993 }
01994 
01995 /// OptimizeLoopTermCond - Change loop terminating condition to use the
01996 /// postinc iv when possible.
01997 void
01998 LSRInstance::OptimizeLoopTermCond() {
01999   SmallPtrSet<Instruction *, 4> PostIncs;
02000 
02001   BasicBlock *LatchBlock = L->getLoopLatch();
02002   SmallVector<BasicBlock*, 8> ExitingBlocks;
02003   L->getExitingBlocks(ExitingBlocks);
02004 
02005   for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
02006     BasicBlock *ExitingBlock = ExitingBlocks[i];
02007 
02008     // Get the terminating condition for the loop if possible.  If we
02009     // can, we want to change it to use a post-incremented version of its
02010     // induction variable, to allow coalescing the live ranges for the IV into
02011     // one register value.
02012 
02013     BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
02014     if (!TermBr)
02015       continue;
02016     // FIXME: Overly conservative, termination condition could be an 'or' etc..
02017     if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
02018       continue;
02019 
02020     // Search IVUsesByStride to find Cond's IVUse if there is one.
02021     IVStrideUse *CondUse = 0;
02022     ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
02023     if (!FindIVUserForCond(Cond, CondUse))
02024       continue;
02025 
02026     // If the trip count is computed in terms of a max (due to ScalarEvolution
02027     // being unable to find a sufficient guard, for example), change the loop
02028     // comparison to use SLT or ULT instead of NE.
02029     // One consequence of doing this now is that it disrupts the count-down
02030     // optimization. That's not always a bad thing though, because in such
02031     // cases it may still be worthwhile to avoid a max.
02032     Cond = OptimizeMax(Cond, CondUse);
02033 
02034     // If this exiting block dominates the latch block, it may also use
02035     // the post-inc value if it won't be shared with other uses.
02036     // Check for dominance.
02037     if (!DT.dominates(ExitingBlock, LatchBlock))
02038       continue;
02039 
02040     // Conservatively avoid trying to use the post-inc value in non-latch
02041     // exits if there may be pre-inc users in intervening blocks.
02042     if (LatchBlock != ExitingBlock)
02043       for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
02044         // Test if the use is reachable from the exiting block. This dominator
02045         // query is a conservative approximation of reachability.
02046         if (&*UI != CondUse &&
02047             !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
02048           // Conservatively assume there may be reuse if the quotient of their
02049           // strides could be a legal scale.
02050           const SCEV *A = IU.getStride(*CondUse, L);
02051           const SCEV *B = IU.getStride(*UI, L);
02052           if (!A || !B) continue;
02053           if (SE.getTypeSizeInBits(A->getType()) !=
02054               SE.getTypeSizeInBits(B->getType())) {
02055             if (SE.getTypeSizeInBits(A->getType()) >
02056                 SE.getTypeSizeInBits(B->getType()))
02057               B = SE.getSignExtendExpr(B, A->getType());
02058             else
02059               A = SE.getSignExtendExpr(A, B->getType());
02060           }
02061           if (const SCEVConstant *D =
02062                 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
02063             const ConstantInt *C = D->getValue();
02064             // Stride of one or negative one can have reuse with non-addresses.
02065             if (C->isOne() || C->isAllOnesValue())
02066               goto decline_post_inc;
02067             // Avoid weird situations.
02068             if (C->getValue().getMinSignedBits() >= 64 ||
02069                 C->getValue().isMinSignedValue())
02070               goto decline_post_inc;
02071             // Check for possible scaled-address reuse.
02072             Type *AccessTy = getAccessType(UI->getUser());
02073             int64_t Scale = C->getSExtValue();
02074             if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
02075                                           /*BaseOffset=*/ 0,
02076                                           /*HasBaseReg=*/ false, Scale))
02077               goto decline_post_inc;
02078             Scale = -Scale;
02079             if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0,
02080                                           /*BaseOffset=*/ 0,
02081                                           /*HasBaseReg=*/ false, Scale))
02082               goto decline_post_inc;
02083           }
02084         }
02085 
02086     DEBUG(dbgs() << "  Change loop exiting icmp to use postinc iv: "
02087                  << *Cond << '\n');
02088 
02089     // It's possible for the setcc instruction to be anywhere in the loop, and
02090     // possible for it to have multiple users.  If it is not immediately before
02091     // the exiting block branch, move it.
02092     if (&*++BasicBlock::iterator(Cond) != TermBr) {
02093       if (Cond->hasOneUse()) {
02094         Cond->moveBefore(TermBr);
02095       } else {
02096         // Clone the terminating condition and insert into the loopend.
02097         ICmpInst *OldCond = Cond;
02098         Cond = cast<ICmpInst>(Cond->clone());
02099         Cond->setName(L->getHeader()->getName() + ".termcond");
02100         ExitingBlock->getInstList().insert(TermBr, Cond);
02101 
02102         // Clone the IVUse, as the old use still exists!
02103         CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace());
02104         TermBr->replaceUsesOfWith(OldCond, Cond);
02105       }
02106     }
02107 
02108     // If we get to here, we know that we can transform the setcc instruction to
02109     // use the post-incremented version of the IV, allowing us to coalesce the
02110     // live ranges for the IV correctly.
02111     CondUse->transformToPostInc(L);
02112     Changed = true;
02113 
02114     PostIncs.insert(Cond);
02115   decline_post_inc:;
02116   }
02117 
02118   // Determine an insertion point for the loop induction variable increment. It
02119   // must dominate all the post-inc comparisons we just set up, and it must
02120   // dominate the loop latch edge.
02121   IVIncInsertPos = L->getLoopLatch()->getTerminator();
02122   for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
02123        E = PostIncs.end(); I != E; ++I) {
02124     BasicBlock *BB =
02125       DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
02126                                     (*I)->getParent());
02127     if (BB == (*I)->getParent())
02128       IVIncInsertPos = *I;
02129     else if (BB != IVIncInsertPos->getParent())
02130       IVIncInsertPos = BB->getTerminator();
02131   }
02132 }
02133 
02134 /// reconcileNewOffset - Determine if the given use can accommodate a fixup
02135 /// at the given offset and other details. If so, update the use and
02136 /// return true.
02137 bool
02138 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
02139                                 LSRUse::KindType Kind, Type *AccessTy) {
02140   int64_t NewMinOffset = LU.MinOffset;
02141   int64_t NewMaxOffset = LU.MaxOffset;
02142   Type *NewAccessTy = AccessTy;
02143 
02144   // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
02145   // something conservative, however this can pessimize in the case that one of
02146   // the uses will have all its uses outside the loop, for example.
02147   if (LU.Kind != Kind)
02148     return false;
02149   // Conservatively assume HasBaseReg is true for now.
02150   if (NewOffset < LU.MinOffset) {
02151     if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
02152                           LU.MaxOffset - NewOffset, HasBaseReg))
02153       return false;
02154     NewMinOffset = NewOffset;
02155   } else if (NewOffset > LU.MaxOffset) {
02156     if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
02157                           NewOffset - LU.MinOffset, HasBaseReg))
02158       return false;
02159     NewMaxOffset = NewOffset;
02160   }
02161   // Check for a mismatched access type, and fall back conservatively as needed.
02162   // TODO: Be less conservative when the type is similar and can use the same
02163   // addressing modes.
02164   if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
02165     NewAccessTy = Type::getVoidTy(AccessTy->getContext());
02166 
02167   // Update the use.
02168   LU.MinOffset = NewMinOffset;
02169   LU.MaxOffset = NewMaxOffset;
02170   LU.AccessTy = NewAccessTy;
02171   if (NewOffset != LU.Offsets.back())
02172     LU.Offsets.push_back(NewOffset);
02173   return true;
02174 }
02175 
02176 /// getUse - Return an LSRUse index and an offset value for a fixup which
02177 /// needs the given expression, with the given kind and optional access type.
02178 /// Either reuse an existing use or create a new one, as needed.
02179 std::pair<size_t, int64_t>
02180 LSRInstance::getUse(const SCEV *&Expr,
02181                     LSRUse::KindType Kind, Type *AccessTy) {
02182   const SCEV *Copy = Expr;
02183   int64_t Offset = ExtractImmediate(Expr, SE);
02184 
02185   // Basic uses can't accept any offset, for example.
02186   if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0,
02187                         Offset, /*HasBaseReg=*/ true)) {
02188     Expr = Copy;
02189     Offset = 0;
02190   }
02191 
02192   std::pair<UseMapTy::iterator, bool> P =
02193     UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0));
02194   if (!P.second) {
02195     // A use already existed with this base.
02196     size_t LUIdx = P.first->second;
02197     LSRUse &LU = Uses[LUIdx];
02198     if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
02199       // Reuse this use.
02200       return std::make_pair(LUIdx, Offset);
02201   }
02202 
02203   // Create a new use.
02204   size_t LUIdx = Uses.size();
02205   P.first->second = LUIdx;
02206   Uses.push_back(LSRUse(Kind, AccessTy));
02207   LSRUse &LU = Uses[LUIdx];
02208 
02209   // We don't need to track redundant offsets, but we don't need to go out
02210   // of our way here to avoid them.
02211   if (LU.Offsets.empty() || Offset != LU.Offsets.back())
02212     LU.Offsets.push_back(Offset);
02213 
02214   LU.MinOffset = Offset;
02215   LU.MaxOffset = Offset;
02216   return std::make_pair(LUIdx, Offset);
02217 }
02218 
02219 /// DeleteUse - Delete the given use from the Uses list.
02220 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) {
02221   if (&LU != &Uses.back())
02222     std::swap(LU, Uses.back());
02223   Uses.pop_back();
02224 
02225   // Update RegUses.
02226   RegUses.SwapAndDropUse(LUIdx, Uses.size());
02227 }
02228 
02229 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has
02230 /// a formula that has the same registers as the given formula.
02231 LSRUse *
02232 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
02233                                        const LSRUse &OrigLU) {
02234   // Search all uses for the formula. This could be more clever.
02235   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
02236     LSRUse &LU = Uses[LUIdx];
02237     // Check whether this use is close enough to OrigLU, to see whether it's
02238     // worthwhile looking through its formulae.
02239     // Ignore ICmpZero uses because they may contain formulae generated by
02240     // GenerateICmpZeroScales, in which case adding fixup offsets may
02241     // be invalid.
02242     if (&LU != &OrigLU &&
02243         LU.Kind != LSRUse::ICmpZero &&
02244         LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy &&
02245         LU.WidestFixupType == OrigLU.WidestFixupType &&
02246         LU.HasFormulaWithSameRegs(OrigF)) {
02247       // Scan through this use's formulae.
02248       for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
02249            E = LU.Formulae.end(); I != E; ++I) {
02250         const Formula &F = *I;
02251         // Check to see if this formula has the same registers and symbols
02252         // as OrigF.
02253         if (F.BaseRegs == OrigF.BaseRegs &&
02254             F.ScaledReg == OrigF.ScaledReg &&
02255             F.BaseGV == OrigF.BaseGV &&
02256             F.Scale == OrigF.Scale &&
02257             F.UnfoldedOffset == OrigF.UnfoldedOffset) {
02258           if (F.BaseOffset == 0)
02259             return &LU;
02260           // This is the formula where all the registers and symbols matched;
02261           // there aren't going to be any others. Since we declined it, we
02262           // can skip the rest of the formulae and proceed to the next LSRUse.
02263           break;
02264         }
02265       }
02266     }
02267   }
02268 
02269   // Nothing looked good.
02270   return 0;
02271 }
02272 
02273 void LSRInstance::CollectInterestingTypesAndFactors() {
02274   SmallSetVector<const SCEV *, 4> Strides;
02275 
02276   // Collect interesting types and strides.
02277   SmallVector<const SCEV *, 4> Worklist;
02278   for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
02279     const SCEV *Expr = IU.getExpr(*UI);
02280 
02281     // Collect interesting types.
02282     Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
02283 
02284     // Add strides for mentioned loops.
02285     Worklist.push_back(Expr);
02286     do {
02287       const SCEV *S = Worklist.pop_back_val();
02288       if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
02289         if (AR->getLoop() == L)
02290           Strides.insert(AR->getStepRecurrence(SE));
02291         Worklist.push_back(AR->getStart());
02292       } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
02293         Worklist.append(Add->op_begin(), Add->op_end());
02294       }
02295     } while (!Worklist.empty());
02296   }
02297 
02298   // Compute interesting factors from the set of interesting strides.
02299   for (SmallSetVector<const SCEV *, 4>::const_iterator
02300        I = Strides.begin(), E = Strides.end(); I != E; ++I)
02301     for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
02302          std::next(I); NewStrideIter != E; ++NewStrideIter) {
02303       const SCEV *OldStride = *I;
02304       const SCEV *NewStride = *NewStrideIter;
02305 
02306       if (SE.getTypeSizeInBits(OldStride->getType()) !=
02307           SE.getTypeSizeInBits(NewStride->getType())) {
02308         if (SE.getTypeSizeInBits(OldStride->getType()) >
02309             SE.getTypeSizeInBits(NewStride->getType()))
02310           NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
02311         else
02312           OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
02313       }
02314       if (const SCEVConstant *Factor =
02315             dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
02316                                                         SE, true))) {
02317         if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
02318           Factors.insert(Factor->getValue()->getValue().getSExtValue());
02319       } else if (const SCEVConstant *Factor =
02320                    dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
02321                                                                NewStride,
02322                                                                SE, true))) {
02323         if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
02324           Factors.insert(Factor->getValue()->getValue().getSExtValue());
02325       }
02326     }
02327 
02328   // If all uses use the same type, don't bother looking for truncation-based
02329   // reuse.
02330   if (Types.size() == 1)
02331     Types.clear();
02332 
02333   DEBUG(print_factors_and_types(dbgs()));
02334 }
02335 
02336 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed
02337 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped
02338 /// Instructions to IVStrideUses, we could partially skip this.
02339 static User::op_iterator
02340 findIVOperand(User::op_iterator OI, User::op_iterator OE,
02341               Loop *L, ScalarEvolution &SE) {
02342   for(; OI != OE; ++OI) {
02343     if (Instruction *Oper = dyn_cast<Instruction>(*OI)) {
02344       if (!SE.isSCEVable(Oper->getType()))
02345         continue;
02346 
02347       if (const SCEVAddRecExpr *AR =
02348           dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) {
02349         if (AR->getLoop() == L)
02350           break;
02351       }
02352     }
02353   }
02354   return OI;
02355 }
02356 
02357 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst
02358 /// operands, so wrap it in a convenient helper.
02359 static Value *getWideOperand(Value *Oper) {
02360   if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper))
02361     return Trunc->getOperand(0);
02362   return Oper;
02363 }
02364 
02365 /// isCompatibleIVType - Return true if we allow an IV chain to include both
02366 /// types.
02367 static bool isCompatibleIVType(Value *LVal, Value *RVal) {
02368   Type *LType = LVal->getType();
02369   Type *RType = RVal->getType();
02370   return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy());
02371 }
02372 
02373 /// getExprBase - Return an approximation of this SCEV expression's "base", or
02374 /// NULL for any constant. Returning the expression itself is
02375 /// conservative. Returning a deeper subexpression is more precise and valid as
02376 /// long as it isn't less complex than another subexpression. For expressions
02377 /// involving multiple unscaled values, we need to return the pointer-type
02378 /// SCEVUnknown. This avoids forming chains across objects, such as:
02379 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a.
02380 ///
02381 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost
02382 /// SCEVUnknown, we simply return the rightmost SCEV operand.
02383 static const SCEV *getExprBase(const SCEV *S) {
02384   switch (S->getSCEVType()) {
02385   default: // uncluding scUnknown.
02386     return S;
02387   case scConstant:
02388     return 0;
02389   case scTruncate:
02390     return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand());
02391   case scZeroExtend:
02392     return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand());
02393   case scSignExtend:
02394     return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand());
02395   case scAddExpr: {
02396     // Skip over scaled operands (scMulExpr) to follow add operands as long as
02397     // there's nothing more complex.
02398     // FIXME: not sure if we want to recognize negation.
02399     const SCEVAddExpr *Add = cast<SCEVAddExpr>(S);
02400     for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()),
02401            E(Add->op_begin()); I != E; ++I) {
02402       const SCEV *SubExpr = *I;
02403       if (SubExpr->getSCEVType() == scAddExpr)
02404         return getExprBase(SubExpr);
02405 
02406       if (SubExpr->getSCEVType() != scMulExpr)
02407         return SubExpr;
02408     }
02409     return S; // all operands are scaled, be conservative.
02410   }
02411   case scAddRecExpr:
02412     return getExprBase(cast<SCEVAddRecExpr>(S)->getStart());
02413   }
02414 }
02415 
02416 /// Return true if the chain increment is profitable to expand into a loop
02417 /// invariant value, which may require its own register. A profitable chain
02418 /// increment will be an offset relative to the same base. We allow such offsets
02419 /// to potentially be used as chain increment as long as it's not obviously
02420 /// expensive to expand using real instructions.
02421 bool IVChain::isProfitableIncrement(const SCEV *OperExpr,
02422                                     const SCEV *IncExpr,
02423                                     ScalarEvolution &SE) {
02424   // Aggressively form chains when -stress-ivchain.
02425   if (StressIVChain)
02426     return true;
02427 
02428   // Do not replace a constant offset from IV head with a nonconstant IV
02429   // increment.
02430   if (!isa<SCEVConstant>(IncExpr)) {
02431     const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand));
02432     if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr)))
02433       return 0;
02434   }
02435 
02436   SmallPtrSet<const SCEV*, 8> Processed;
02437   return !isHighCostExpansion(IncExpr, Processed, SE);
02438 }
02439 
02440 /// Return true if the number of registers needed for the chain is estimated to
02441 /// be less than the number required for the individual IV users. First prohibit
02442 /// any IV users that keep the IV live across increments (the Users set should
02443 /// be empty). Next count the number and type of increments in the chain.
02444 ///
02445 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't
02446 /// effectively use postinc addressing modes. Only consider it profitable it the
02447 /// increments can be computed in fewer registers when chained.
02448 ///
02449 /// TODO: Consider IVInc free if it's already used in another chains.
02450 static bool
02451 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
02452                   ScalarEvolution &SE, const TargetTransformInfo &TTI) {
02453   if (StressIVChain)
02454     return true;
02455 
02456   if (!Chain.hasIncs())
02457     return false;
02458 
02459   if (!Users.empty()) {
02460     DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n";
02461           for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(),
02462                  E = Users.end(); I != E; ++I) {
02463             dbgs() << "  " << **I << "\n";
02464           });
02465     return false;
02466   }
02467   assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
02468 
02469   // The chain itself may require a register, so intialize cost to 1.
02470   int cost = 1;
02471 
02472   // A complete chain likely eliminates the need for keeping the original IV in
02473   // a register. LSR does not currently know how to form a complete chain unless
02474   // the header phi already exists.
02475   if (isa<PHINode>(Chain.tailUserInst())
02476       && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) {
02477     --cost;
02478   }
02479   const SCEV *LastIncExpr = 0;
02480   unsigned NumConstIncrements = 0;
02481   unsigned NumVarIncrements = 0;
02482   unsigned NumReusedIncrements = 0;
02483   for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
02484        I != E; ++I) {
02485 
02486     if (I->IncExpr->isZero())
02487       continue;
02488 
02489     // Incrementing by zero or some constant is neutral. We assume constants can
02490     // be folded into an addressing mode or an add's immediate operand.
02491     if (isa<SCEVConstant>(I->IncExpr)) {
02492       ++NumConstIncrements;
02493       continue;
02494     }
02495 
02496     if (I->IncExpr == LastIncExpr)
02497       ++NumReusedIncrements;
02498     else
02499       ++NumVarIncrements;
02500 
02501     LastIncExpr = I->IncExpr;
02502   }
02503   // An IV chain with a single increment is handled by LSR's postinc
02504   // uses. However, a chain with multiple increments requires keeping the IV's
02505   // value live longer than it needs to be if chained.
02506   if (NumConstIncrements > 1)
02507     --cost;
02508 
02509   // Materializing increment expressions in the preheader that didn't exist in
02510   // the original code may cost a register. For example, sign-extended array
02511   // indices can produce ridiculous increments like this:
02512   // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64)))
02513   cost += NumVarIncrements;
02514 
02515   // Reusing variable increments likely saves a register to hold the multiple of
02516   // the stride.
02517   cost -= NumReusedIncrements;
02518 
02519   DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost
02520                << "\n");
02521 
02522   return cost < 0;
02523 }
02524 
02525 /// ChainInstruction - Add this IV user to an existing chain or make it the head
02526 /// of a new chain.
02527 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper,
02528                                    SmallVectorImpl<ChainUsers> &ChainUsersVec) {
02529   // When IVs are used as types of varying widths, they are generally converted
02530   // to a wider type with some uses remaining narrow under a (free) trunc.
02531   Value *const NextIV = getWideOperand(IVOper);
02532   const SCEV *const OperExpr = SE.getSCEV(NextIV);
02533   const SCEV *const OperExprBase = getExprBase(OperExpr);
02534 
02535   // Visit all existing chains. Check if its IVOper can be computed as a
02536   // profitable loop invariant increment from the last link in the Chain.
02537   unsigned ChainIdx = 0, NChains = IVChainVec.size();
02538   const SCEV *LastIncExpr = 0;
02539   for (; ChainIdx < NChains; ++ChainIdx) {
02540     IVChain &Chain = IVChainVec[ChainIdx];
02541 
02542     // Prune the solution space aggressively by checking that both IV operands
02543     // are expressions that operate on the same unscaled SCEVUnknown. This
02544     // "base" will be canceled by the subsequent getMinusSCEV call. Checking
02545     // first avoids creating extra SCEV expressions.
02546     if (!StressIVChain && Chain.ExprBase != OperExprBase)
02547       continue;
02548 
02549     Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand);
02550     if (!isCompatibleIVType(PrevIV, NextIV))
02551       continue;
02552 
02553     // A phi node terminates a chain.
02554     if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst()))
02555       continue;
02556 
02557     // The increment must be loop-invariant so it can be kept in a register.
02558     const SCEV *PrevExpr = SE.getSCEV(PrevIV);
02559     const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr);
02560     if (!SE.isLoopInvariant(IncExpr, L))
02561       continue;
02562 
02563     if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) {
02564       LastIncExpr = IncExpr;
02565       break;
02566     }
02567   }
02568   // If we haven't found a chain, create a new one, unless we hit the max. Don't
02569   // bother for phi nodes, because they must be last in the chain.
02570   if (ChainIdx == NChains) {
02571     if (isa<PHINode>(UserInst))
02572       return;
02573     if (NChains >= MaxChains && !StressIVChain) {
02574       DEBUG(dbgs() << "IV Chain Limit\n");
02575       return;
02576     }
02577     LastIncExpr = OperExpr;
02578     // IVUsers may have skipped over sign/zero extensions. We don't currently
02579     // attempt to form chains involving extensions unless they can be hoisted
02580     // into this loop's AddRec.
02581     if (!isa<SCEVAddRecExpr>(LastIncExpr))
02582       return;
02583     ++NChains;
02584     IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr),
02585                                  OperExprBase));
02586     ChainUsersVec.resize(NChains);
02587     DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst
02588                  << ") IV=" << *LastIncExpr << "\n");
02589   } else {
02590     DEBUG(dbgs() << "IV Chain#" << ChainIdx << "  Inc: (" << *UserInst
02591                  << ") IV+" << *LastIncExpr << "\n");
02592     // Add this IV user to the end of the chain.
02593     IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr));
02594   }
02595   IVChain &Chain = IVChainVec[ChainIdx];
02596 
02597   SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers;
02598   // This chain's NearUsers become FarUsers.
02599   if (!LastIncExpr->isZero()) {
02600     ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(),
02601                                             NearUsers.end());
02602     NearUsers.clear();
02603   }
02604 
02605   // All other uses of IVOperand become near uses of the chain.
02606   // We currently ignore intermediate values within SCEV expressions, assuming
02607   // they will eventually be used be the current chain, or can be computed
02608   // from one of the chain increments. To be more precise we could
02609   // transitively follow its user and only add leaf IV users to the set.
02610   for (User *U : IVOper->users()) {
02611     Instruction *OtherUse = dyn_cast<Instruction>(U);
02612     if (!OtherUse)
02613       continue;
02614     // Uses in the chain will no longer be uses if the chain is formed.
02615     // Include the head of the chain in this iteration (not Chain.begin()).
02616     IVChain::const_iterator IncIter = Chain.Incs.begin();
02617     IVChain::const_iterator IncEnd = Chain.Incs.end();
02618     for( ; IncIter != IncEnd; ++IncIter) {
02619       if (IncIter->UserInst == OtherUse)
02620         break;
02621     }
02622     if (IncIter != IncEnd)
02623       continue;
02624 
02625     if (SE.isSCEVable(OtherUse->getType())
02626         && !isa<SCEVUnknown>(SE.getSCEV(OtherUse))
02627         && IU.isIVUserOrOperand(OtherUse)) {
02628       continue;
02629     }
02630     NearUsers.insert(OtherUse);
02631   }
02632 
02633   // Since this user is part of the chain, it's no longer considered a use
02634   // of the chain.
02635   ChainUsersVec[ChainIdx].FarUsers.erase(UserInst);
02636 }
02637 
02638 /// CollectChains - Populate the vector of Chains.
02639 ///
02640 /// This decreases ILP at the architecture level. Targets with ample registers,
02641 /// multiple memory ports, and no register renaming probably don't want
02642 /// this. However, such targets should probably disable LSR altogether.
02643 ///
02644 /// The job of LSR is to make a reasonable choice of induction variables across
02645 /// the loop. Subsequent passes can easily "unchain" computation exposing more
02646 /// ILP *within the loop* if the target wants it.
02647 ///
02648 /// Finding the best IV chain is potentially a scheduling problem. Since LSR
02649 /// will not reorder memory operations, it will recognize this as a chain, but
02650 /// will generate redundant IV increments. Ideally this would be corrected later
02651 /// by a smart scheduler:
02652 ///        = A[i]
02653 ///        = A[i+x]
02654 /// A[i]   =
02655 /// A[i+x] =
02656 ///
02657 /// TODO: Walk the entire domtree within this loop, not just the path to the
02658 /// loop latch. This will discover chains on side paths, but requires
02659 /// maintaining multiple copies of the Chains state.
02660 void LSRInstance::CollectChains() {
02661   DEBUG(dbgs() << "Collecting IV Chains.\n");
02662   SmallVector<ChainUsers, 8> ChainUsersVec;
02663 
02664   SmallVector<BasicBlock *,8> LatchPath;
02665   BasicBlock *LoopHeader = L->getHeader();
02666   for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch());
02667        Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) {
02668     LatchPath.push_back(Rung->getBlock());
02669   }
02670   LatchPath.push_back(LoopHeader);
02671 
02672   // Walk the instruction stream from the loop header to the loop latch.
02673   for (SmallVectorImpl<BasicBlock *>::reverse_iterator
02674          BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend();
02675        BBIter != BBEnd; ++BBIter) {
02676     for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end();
02677          I != E; ++I) {
02678       // Skip instructions that weren't seen by IVUsers analysis.
02679       if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I))
02680         continue;
02681 
02682       // Ignore users that are part of a SCEV expression. This way we only
02683       // consider leaf IV Users. This effectively rediscovers a portion of
02684       // IVUsers analysis but in program order this time.
02685       if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I)))
02686         continue;
02687 
02688       // Remove this instruction from any NearUsers set it may be in.
02689       for (unsigned ChainIdx = 0, NChains = IVChainVec.size();
02690            ChainIdx < NChains; ++ChainIdx) {
02691         ChainUsersVec[ChainIdx].NearUsers.erase(I);
02692       }
02693       // Search for operands that can be chained.
02694       SmallPtrSet<Instruction*, 4> UniqueOperands;
02695       User::op_iterator IVOpEnd = I->op_end();
02696       User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE);
02697       while (IVOpIter != IVOpEnd) {
02698         Instruction *IVOpInst = cast<Instruction>(*IVOpIter);
02699         if (UniqueOperands.insert(IVOpInst))
02700           ChainInstruction(I, IVOpInst, ChainUsersVec);
02701         IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
02702       }
02703     } // Continue walking down the instructions.
02704   } // Continue walking down the domtree.
02705   // Visit phi backedges to determine if the chain can generate the IV postinc.
02706   for (BasicBlock::iterator I = L->getHeader()->begin();
02707        PHINode *PN = dyn_cast<PHINode>(I); ++I) {
02708     if (!SE.isSCEVable(PN->getType()))
02709       continue;
02710 
02711     Instruction *IncV =
02712       dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch()));
02713     if (IncV)
02714       ChainInstruction(PN, IncV, ChainUsersVec);
02715   }
02716   // Remove any unprofitable chains.
02717   unsigned ChainIdx = 0;
02718   for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
02719        UsersIdx < NChains; ++UsersIdx) {
02720     if (!isProfitableChain(IVChainVec[UsersIdx],
02721                            ChainUsersVec[UsersIdx].FarUsers, SE, TTI))
02722       continue;
02723     // Preserve the chain at UsesIdx.
02724     if (ChainIdx != UsersIdx)
02725       IVChainVec[ChainIdx] = IVChainVec[UsersIdx];
02726     FinalizeChain(IVChainVec[ChainIdx]);
02727     ++ChainIdx;
02728   }
02729   IVChainVec.resize(ChainIdx);
02730 }
02731 
02732 void LSRInstance::FinalizeChain(IVChain &Chain) {
02733   assert(!Chain.Incs.empty() && "empty IV chains are not allowed");
02734   DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n");
02735 
02736   for (IVChain::const_iterator I = Chain.begin(), E = Chain.end();
02737        I != E; ++I) {
02738     DEBUG(dbgs() << "        Inc: " << *I->UserInst << "\n");
02739     User::op_iterator UseI =
02740       std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand);
02741     assert(UseI != I->UserInst->op_end() && "cannot find IV operand");
02742     IVIncSet.insert(UseI);
02743   }
02744 }
02745 
02746 /// Return true if the IVInc can be folded into an addressing mode.
02747 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
02748                              Value *Operand, const TargetTransformInfo &TTI) {
02749   const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
02750   if (!IncConst || !isAddressUse(UserInst, Operand))
02751     return false;
02752 
02753   if (IncConst->getValue()->getValue().getMinSignedBits() > 64)
02754     return false;
02755 
02756   int64_t IncOffset = IncConst->getValue()->getSExtValue();
02757   if (!isAlwaysFoldable(TTI, LSRUse::Address,
02758                         getAccessType(UserInst), /*BaseGV=*/ 0,
02759                         IncOffset, /*HaseBaseReg=*/ false))
02760     return false;
02761 
02762   return true;
02763 }
02764 
02765 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to
02766 /// materialize the IV user's operand from the previous IV user's operand.
02767 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter,
02768                                   SmallVectorImpl<WeakVH> &DeadInsts) {
02769   // Find the new IVOperand for the head of the chain. It may have been replaced
02770   // by LSR.
02771   const IVInc &Head = Chain.Incs[0];
02772   User::op_iterator IVOpEnd = Head.UserInst->op_end();
02773   // findIVOperand returns IVOpEnd if it can no longer find a valid IV user.
02774   User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(),
02775                                              IVOpEnd, L, SE);
02776   Value *IVSrc = 0;
02777   while (IVOpIter != IVOpEnd) {
02778     IVSrc = getWideOperand(*IVOpIter);
02779 
02780     // If this operand computes the expression that the chain needs, we may use
02781     // it. (Check this after setting IVSrc which is used below.)
02782     //
02783     // Note that if Head.IncExpr is wider than IVSrc, then this phi is too
02784     // narrow for the chain, so we can no longer use it. We do allow using a
02785     // wider phi, assuming the LSR checked for free truncation. In that case we
02786     // should already have a truncate on this operand such that
02787     // getSCEV(IVSrc) == IncExpr.
02788     if (SE.getSCEV(*IVOpIter) == Head.IncExpr
02789         || SE.getSCEV(IVSrc) == Head.IncExpr) {
02790       break;
02791     }
02792     IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE);
02793   }
02794   if (IVOpIter == IVOpEnd) {
02795     // Gracefully give up on this chain.
02796     DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n");
02797     return;
02798   }
02799 
02800   DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n");
02801   Type *IVTy = IVSrc->getType();
02802   Type *IntTy = SE.getEffectiveSCEVType(IVTy);
02803   const SCEV *LeftOverExpr = 0;
02804   for (IVChain::const_iterator IncI = Chain.begin(),
02805          IncE = Chain.end(); IncI != IncE; ++IncI) {
02806 
02807     Instruction *InsertPt = IncI->UserInst;
02808     if (isa<PHINode>(InsertPt))
02809       InsertPt = L->getLoopLatch()->getTerminator();
02810 
02811     // IVOper will replace the current IV User's operand. IVSrc is the IV
02812     // value currently held in a register.
02813     Value *IVOper = IVSrc;
02814     if (!IncI->IncExpr->isZero()) {
02815       // IncExpr was the result of subtraction of two narrow values, so must
02816       // be signed.
02817       const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy);
02818       LeftOverExpr = LeftOverExpr ?
02819         SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr;
02820     }
02821     if (LeftOverExpr && !LeftOverExpr->isZero()) {
02822       // Expand the IV increment.
02823       Rewriter.clearPostInc();
02824       Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt);
02825       const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc),
02826                                              SE.getUnknown(IncV));
02827       IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt);
02828 
02829       // If an IV increment can't be folded, use it as the next IV value.
02830       if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
02831                             TTI)) {
02832         assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
02833         IVSrc = IVOper;
02834         LeftOverExpr = 0;
02835       }
02836     }
02837     Type *OperTy = IncI->IVOperand->getType();
02838     if (IVTy != OperTy) {
02839       assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) &&
02840              "cannot extend a chained IV");
02841       IRBuilder<> Builder(InsertPt);
02842       IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain");
02843     }
02844     IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper);
02845     DeadInsts.push_back(IncI->IVOperand);
02846   }
02847   // If LSR created a new, wider phi, we may also replace its postinc. We only
02848   // do this if we also found a wide value for the head of the chain.
02849   if (isa<PHINode>(Chain.tailUserInst())) {
02850     for (BasicBlock::iterator I = L->getHeader()->begin();
02851          PHINode *Phi = dyn_cast<PHINode>(I); ++I) {
02852       if (!isCompatibleIVType(Phi, IVSrc))
02853         continue;
02854       Instruction *PostIncV = dyn_cast<Instruction>(
02855         Phi->getIncomingValueForBlock(L->getLoopLatch()));
02856       if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc)))
02857         continue;
02858       Value *IVOper = IVSrc;
02859       Type *PostIncTy = PostIncV->getType();
02860       if (IVTy != PostIncTy) {
02861         assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types");
02862         IRBuilder<> Builder(L->getLoopLatch()->getTerminator());
02863         Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc());
02864         IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain");
02865       }
02866       Phi->replaceUsesOfWith(PostIncV, IVOper);
02867       DeadInsts.push_back(PostIncV);
02868     }
02869   }
02870 }
02871 
02872 void LSRInstance::CollectFixupsAndInitialFormulae() {
02873   for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
02874     Instruction *UserInst = UI->getUser();
02875     // Skip IV users that are part of profitable IV Chains.
02876     User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(),
02877                                        UI->getOperandValToReplace());
02878     assert(UseI != UserInst->op_end() && "cannot find IV operand");
02879     if (IVIncSet.count(UseI))
02880       continue;
02881 
02882     // Record the uses.
02883     LSRFixup &LF = getNewFixup();
02884     LF.UserInst = UserInst;
02885     LF.OperandValToReplace = UI->getOperandValToReplace();
02886     LF.PostIncLoops = UI->getPostIncLoops();
02887 
02888     LSRUse::KindType Kind = LSRUse::Basic;
02889     Type *AccessTy = 0;
02890     if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
02891       Kind = LSRUse::Address;
02892       AccessTy = getAccessType(LF.UserInst);
02893     }
02894 
02895     const SCEV *S = IU.getExpr(*UI);
02896 
02897     // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
02898     // (N - i == 0), and this allows (N - i) to be the expression that we work
02899     // with rather than just N or i, so we can consider the register
02900     // requirements for both N and i at the same time. Limiting this code to
02901     // equality icmps is not a problem because all interesting loops use
02902     // equality icmps, thanks to IndVarSimplify.
02903     if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
02904       if (CI->isEquality()) {
02905         // Swap the operands if needed to put the OperandValToReplace on the
02906         // left, for consistency.
02907         Value *NV = CI->getOperand(1);
02908         if (NV == LF.OperandValToReplace) {
02909           CI->setOperand(1, CI->getOperand(0));
02910           CI->setOperand(0, NV);
02911           NV = CI->getOperand(1);
02912           Changed = true;
02913         }
02914 
02915         // x == y  -->  x - y == 0
02916         const SCEV *N = SE.getSCEV(NV);
02917         if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
02918           // S is normalized, so normalize N before folding it into S
02919           // to keep the result normalized.
02920           N = TransformForPostIncUse(Normalize, N, CI, 0,
02921                                      LF.PostIncLoops, SE, DT);
02922           Kind = LSRUse::ICmpZero;
02923           S = SE.getMinusSCEV(N, S);
02924         }
02925 
02926         // -1 and the negations of all interesting strides (except the negation
02927         // of -1) are now also interesting.
02928         for (size_t i = 0, e = Factors.size(); i != e; ++i)
02929           if (Factors[i] != -1)
02930             Factors.insert(-(uint64_t)Factors[i]);
02931         Factors.insert(-1);
02932       }
02933 
02934     // Set up the initial formula for this use.
02935     std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
02936     LF.LUIdx = P.first;
02937     LF.Offset = P.second;
02938     LSRUse &LU = Uses[LF.LUIdx];
02939     LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
02940     if (!LU.WidestFixupType ||
02941         SE.getTypeSizeInBits(LU.WidestFixupType) <
02942         SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
02943       LU.WidestFixupType = LF.OperandValToReplace->getType();
02944 
02945     // If this is the first use of this LSRUse, give it a formula.
02946     if (LU.Formulae.empty()) {
02947       InsertInitialFormula(S, LU, LF.LUIdx);
02948       CountRegisters(LU.Formulae.back(), LF.LUIdx);
02949     }
02950   }
02951 
02952   DEBUG(print_fixups(dbgs()));
02953 }
02954 
02955 /// InsertInitialFormula - Insert a formula for the given expression into
02956 /// the given use, separating out loop-variant portions from loop-invariant
02957 /// and loop-computable portions.
02958 void
02959 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
02960   // Mark uses whose expressions cannot be expanded.
02961   if (!isSafeToExpand(S, SE))
02962     LU.RigidFormula = true;
02963 
02964   Formula F;
02965   F.InitialMatch(S, L, SE);
02966   bool Inserted = InsertFormula(LU, LUIdx, F);
02967   assert(Inserted && "Initial formula already exists!"); (void)Inserted;
02968 }
02969 
02970 /// InsertSupplementalFormula - Insert a simple single-register formula for
02971 /// the given expression into the given use.
02972 void
02973 LSRInstance::InsertSupplementalFormula(const SCEV *S,
02974                                        LSRUse &LU, size_t LUIdx) {
02975   Formula F;
02976   F.BaseRegs.push_back(S);
02977   F.HasBaseReg = true;
02978   bool Inserted = InsertFormula(LU, LUIdx, F);
02979   assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
02980 }
02981 
02982 /// CountRegisters - Note which registers are used by the given formula,
02983 /// updating RegUses.
02984 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
02985   if (F.ScaledReg)
02986     RegUses.CountRegister(F.ScaledReg, LUIdx);
02987   for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
02988        E = F.BaseRegs.end(); I != E; ++I)
02989     RegUses.CountRegister(*I, LUIdx);
02990 }
02991 
02992 /// InsertFormula - If the given formula has not yet been inserted, add it to
02993 /// the list, and return true. Return false otherwise.
02994 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
02995   if (!LU.InsertFormula(F))
02996     return false;
02997 
02998   CountRegisters(F, LUIdx);
02999   return true;
03000 }
03001 
03002 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
03003 /// loop-invariant values which we're tracking. These other uses will pin these
03004 /// values in registers, making them less profitable for elimination.
03005 /// TODO: This currently misses non-constant addrec step registers.
03006 /// TODO: Should this give more weight to users inside the loop?
03007 void
03008 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
03009   SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
03010   SmallPtrSet<const SCEV *, 8> Inserted;
03011 
03012   while (!Worklist.empty()) {
03013     const SCEV *S = Worklist.pop_back_val();
03014 
03015     if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
03016       Worklist.append(N->op_begin(), N->op_end());
03017     else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
03018       Worklist.push_back(C->getOperand());
03019     else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
03020       Worklist.push_back(D->getLHS());
03021       Worklist.push_back(D->getRHS());
03022     } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) {
03023       if (!Inserted.insert(US)) continue;
03024       const Value *V = US->getValue();
03025       if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
03026         // Look for instructions defined outside the loop.
03027         if (L->contains(Inst)) continue;
03028       } else if (isa<UndefValue>(V))
03029         // Undef doesn't have a live range, so it doesn't matter.
03030         continue;
03031       for (const Use &U : V->uses()) {
03032         const Instruction *UserInst = dyn_cast<Instruction>(U.getUser());
03033         // Ignore non-instructions.
03034         if (!UserInst)
03035           continue;
03036         // Ignore instructions in other functions (as can happen with
03037         // Constants).
03038         if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
03039           continue;
03040         // Ignore instructions not dominated by the loop.
03041         const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
03042           UserInst->getParent() :
03043           cast<PHINode>(UserInst)->getIncomingBlock(
03044             PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
03045         if (!DT.dominates(L->getHeader(), UseBB))
03046           continue;
03047         // Ignore uses which are part of other SCEV expressions, to avoid
03048         // analyzing them multiple times.
03049         if (SE.isSCEVable(UserInst->getType())) {
03050           const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
03051           // If the user is a no-op, look through to its uses.
03052           if (!isa<SCEVUnknown>(UserS))
03053             continue;
03054           if (UserS == US) {
03055             Worklist.push_back(
03056               SE.getUnknown(const_cast<Instruction *>(UserInst)));
03057             continue;
03058           }
03059         }
03060         // Ignore icmp instructions which are already being analyzed.
03061         if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
03062           unsigned OtherIdx = !U.getOperandNo();
03063           Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
03064           if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L))
03065             continue;
03066         }
03067 
03068         LSRFixup &LF = getNewFixup();
03069         LF.UserInst = const_cast<Instruction *>(UserInst);
03070         LF.OperandValToReplace = U;
03071         std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
03072         LF.LUIdx = P.first;
03073         LF.Offset = P.second;
03074         LSRUse &LU = Uses[LF.LUIdx];
03075         LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
03076         if (!LU.WidestFixupType ||
03077             SE.getTypeSizeInBits(LU.WidestFixupType) <
03078             SE.getTypeSizeInBits(LF.OperandValToReplace->getType()))
03079           LU.WidestFixupType = LF.OperandValToReplace->getType();
03080         InsertSupplementalFormula(US, LU, LF.LUIdx);
03081         CountRegisters(LU.Formulae.back(), Uses.size() - 1);
03082         break;
03083       }
03084     }
03085   }
03086 }
03087 
03088 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
03089 /// separate registers. If C is non-null, multiply each subexpression by C.
03090 ///
03091 /// Return remainder expression after factoring the subexpressions captured by
03092 /// Ops. If Ops is complete, return NULL.
03093 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C,
03094                                    SmallVectorImpl<const SCEV *> &Ops,
03095                                    const Loop *L,
03096                                    ScalarEvolution &SE,
03097                                    unsigned Depth = 0) {
03098   // Arbitrarily cap recursion to protect compile time.
03099   if (Depth >= 3)
03100     return S;
03101 
03102   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
03103     // Break out add operands.
03104     for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
03105          I != E; ++I) {
03106       const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1);
03107       if (Remainder)
03108         Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
03109     }
03110     return 0;
03111   } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
03112     // Split a non-zero base out of an addrec.
03113     if (AR->getStart()->isZero())
03114       return S;
03115 
03116     const SCEV *Remainder = CollectSubexprs(AR->getStart(),
03117                                             C, Ops, L, SE, Depth+1);
03118     // Split the non-zero AddRec unless it is part of a nested recurrence that
03119     // does not pertain to this loop.
03120     if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) {
03121       Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder);
03122       Remainder = 0;
03123     }
03124     if (Remainder != AR->getStart()) {
03125       if (!Remainder)
03126         Remainder = SE.getConstant(AR->getType(), 0);
03127       return SE.getAddRecExpr(Remainder,
03128                               AR->getStepRecurrence(SE),
03129                               AR->getLoop(),
03130                               //FIXME: AR->getNoWrapFlags(SCEV::FlagNW)
03131                               SCEV::FlagAnyWrap);
03132     }
03133   } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
03134     // Break (C * (a + b + c)) into C*a + C*b + C*c.
03135     if (Mul->getNumOperands() != 2)
03136       return S;
03137     if (const SCEVConstant *Op0 =
03138         dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
03139       C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0;
03140       const SCEV *Remainder =
03141         CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1);
03142       if (Remainder)
03143         Ops.push_back(SE.getMulExpr(C, Remainder));
03144       return 0;
03145     }
03146   }
03147   return S;
03148 }
03149 
03150 /// GenerateReassociations - Split out subexpressions from adds and the bases of
03151 /// addrecs.
03152 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
03153                                          Formula Base,
03154                                          unsigned Depth) {
03155   // Arbitrarily cap recursion to protect compile time.
03156   if (Depth >= 3) return;
03157 
03158   for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
03159     const SCEV *BaseReg = Base.BaseRegs[i];
03160 
03161     SmallVector<const SCEV *, 8> AddOps;
03162     const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE);
03163     if (Remainder)
03164       AddOps.push_back(Remainder);
03165 
03166     if (AddOps.size() == 1) continue;
03167 
03168     for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
03169          JE = AddOps.end(); J != JE; ++J) {
03170 
03171       // Loop-variant "unknown" values are uninteresting; we won't be able to
03172       // do anything meaningful with them.
03173       if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L))
03174         continue;
03175 
03176       // Don't pull a constant into a register if the constant could be folded
03177       // into an immediate field.
03178       if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
03179                            LU.AccessTy, *J, Base.getNumRegs() > 1))
03180         continue;
03181 
03182       // Collect all operands except *J.
03183       SmallVector<const SCEV *, 8> InnerAddOps(
03184           ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J);
03185       InnerAddOps.append(std::next(J),
03186                          ((const SmallVector<const SCEV *, 8> &)AddOps).end());
03187 
03188       // Don't leave just a constant behind in a register if the constant could
03189       // be folded into an immediate field.
03190       if (InnerAddOps.size() == 1 &&
03191           isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind,
03192                            LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1))
03193         continue;
03194 
03195       const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
03196       if (InnerSum->isZero())
03197         continue;
03198       Formula F = Base;
03199 
03200       // Add the remaining pieces of the add back into the new formula.
03201       const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
03202       if (InnerSumSC &&
03203           SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
03204           TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
03205                                   InnerSumSC->getValue()->getZExtValue())) {
03206         F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
03207                            InnerSumSC->getValue()->getZExtValue();
03208         F.BaseRegs.erase(F.BaseRegs.begin() + i);
03209       } else
03210         F.BaseRegs[i] = InnerSum;
03211 
03212       // Add J as its own register, or an unfolded immediate.
03213       const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
03214       if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
03215           TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
03216                                   SC->getValue()->getZExtValue()))
03217         F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
03218                            SC->getValue()->getZExtValue();
03219       else
03220         F.BaseRegs.push_back(*J);
03221 
03222       if (InsertFormula(LU, LUIdx, F))
03223         // If that formula hadn't been seen before, recurse to find more like
03224         // it.
03225         GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
03226     }
03227   }
03228 }
03229 
03230 /// GenerateCombinations - Generate a formula consisting of all of the
03231 /// loop-dominating registers added into a single register.
03232 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
03233                                        Formula Base) {
03234   // This method is only interesting on a plurality of registers.
03235   if (Base.BaseRegs.size() <= 1) return;
03236 
03237   Formula F = Base;
03238   F.BaseRegs.clear();
03239   SmallVector<const SCEV *, 4> Ops;
03240   for (SmallVectorImpl<const SCEV *>::const_iterator
03241        I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
03242     const SCEV *BaseReg = *I;
03243     if (SE.properlyDominates(BaseReg, L->getHeader()) &&
03244         !SE.hasComputableLoopEvolution(BaseReg, L))
03245       Ops.push_back(BaseReg);
03246     else
03247       F.BaseRegs.push_back(BaseReg);
03248   }
03249   if (Ops.size() > 1) {
03250     const SCEV *Sum = SE.getAddExpr(Ops);
03251     // TODO: If Sum is zero, it probably means ScalarEvolution missed an
03252     // opportunity to fold something. For now, just ignore such cases
03253     // rather than proceed with zero in a register.
03254     if (!Sum->isZero()) {
03255       F.BaseRegs.push_back(Sum);
03256       (void)InsertFormula(LU, LUIdx, F);
03257     }
03258   }
03259 }
03260 
03261 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
03262 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
03263                                           Formula Base) {
03264   // We can't add a symbolic offset if the address already contains one.
03265   if (Base.BaseGV) return;
03266 
03267   for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
03268     const SCEV *G = Base.BaseRegs[i];
03269     GlobalValue *GV = ExtractSymbol(G, SE);
03270     if (G->isZero() || !GV)
03271       continue;
03272     Formula F = Base;
03273     F.BaseGV = GV;
03274     if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
03275       continue;
03276     F.BaseRegs[i] = G;
03277     (void)InsertFormula(LU, LUIdx, F);
03278   }
03279 }
03280 
03281 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
03282 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
03283                                           Formula Base) {
03284   // TODO: For now, just add the min and max offset, because it usually isn't
03285   // worthwhile looking at everything inbetween.
03286   SmallVector<int64_t, 2> Worklist;
03287   Worklist.push_back(LU.MinOffset);
03288   if (LU.MaxOffset != LU.MinOffset)
03289     Worklist.push_back(LU.MaxOffset);
03290 
03291   for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
03292     const SCEV *G = Base.BaseRegs[i];
03293 
03294     for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
03295          E = Worklist.end(); I != E; ++I) {
03296       Formula F = Base;
03297       F.BaseOffset = (uint64_t)Base.BaseOffset - *I;
03298       if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind,
03299                      LU.AccessTy, F)) {
03300         // Add the offset to the base register.
03301         const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
03302         // If it cancelled out, drop the base register, otherwise update it.
03303         if (NewG->isZero()) {
03304           std::swap(F.BaseRegs[i], F.BaseRegs.back());
03305           F.BaseRegs.pop_back();
03306         } else
03307           F.BaseRegs[i] = NewG;
03308 
03309         (void)InsertFormula(LU, LUIdx, F);
03310       }
03311     }
03312 
03313     int64_t Imm = ExtractImmediate(G, SE);
03314     if (G->isZero() || Imm == 0)
03315       continue;
03316     Formula F = Base;
03317     F.BaseOffset = (uint64_t)F.BaseOffset + Imm;
03318     if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F))
03319       continue;
03320     F.BaseRegs[i] = G;
03321     (void)InsertFormula(LU, LUIdx, F);
03322   }
03323 }
03324 
03325 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
03326 /// the comparison. For example, x == y -> x*c == y*c.
03327 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
03328                                          Formula Base) {
03329   if (LU.Kind != LSRUse::ICmpZero) return;
03330 
03331   // Determine the integer type for the base formula.
03332   Type *IntTy = Base.getType();
03333   if (!IntTy) return;
03334   if (SE.getTypeSizeInBits(IntTy) > 64) return;
03335 
03336   // Don't do this if there is more than one offset.
03337   if (LU.MinOffset != LU.MaxOffset) return;
03338 
03339   assert(!Base.BaseGV && "ICmpZero use is not legal!");
03340 
03341   // Check each interesting stride.
03342   for (SmallSetVector<int64_t, 8>::const_iterator
03343        I = Factors.begin(), E = Factors.end(); I != E; ++I) {
03344     int64_t Factor = *I;
03345 
03346     // Check that the multiplication doesn't overflow.
03347     if (Base.BaseOffset == INT64_MIN && Factor == -1)
03348       continue;
03349     int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor;
03350     if (NewBaseOffset / Factor != Base.BaseOffset)
03351       continue;
03352     // If the offset will be truncated at this use, check that it is in bounds.
03353     if (!IntTy->isPointerTy() &&
03354         !ConstantInt::isValueValidForType(IntTy, NewBaseOffset))
03355       continue;
03356 
03357     // Check that multiplying with the use offset doesn't overflow.
03358     int64_t Offset = LU.MinOffset;
03359     if (Offset == INT64_MIN && Factor == -1)
03360       continue;
03361     Offset = (uint64_t)Offset * Factor;
03362     if (Offset / Factor != LU.MinOffset)
03363       continue;
03364     // If the offset will be truncated at this use, check that it is in bounds.
03365     if (!IntTy->isPointerTy() &&
03366         !ConstantInt::isValueValidForType(IntTy, Offset))
03367       continue;
03368 
03369     Formula F = Base;
03370     F.BaseOffset = NewBaseOffset;
03371 
03372     // Check that this scale is legal.
03373     if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F))
03374       continue;
03375 
03376     // Compensate for the use having MinOffset built into it.
03377     F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset;
03378 
03379     const SCEV *FactorS = SE.getConstant(IntTy, Factor);
03380 
03381     // Check that multiplying with each base register doesn't overflow.
03382     for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
03383       F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
03384       if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
03385         goto next;
03386     }
03387 
03388     // Check that multiplying with the scaled register doesn't overflow.
03389     if (F.ScaledReg) {
03390       F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
03391       if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
03392         continue;
03393     }
03394 
03395     // Check that multiplying with the unfolded offset doesn't overflow.
03396     if (F.UnfoldedOffset != 0) {
03397       if (F.UnfoldedOffset == INT64_MIN && Factor == -1)
03398         continue;
03399       F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor;
03400       if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset)
03401         continue;
03402       // If the offset will be truncated, check that it is in bounds.
03403       if (!IntTy->isPointerTy() &&
03404           !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset))
03405         continue;
03406     }
03407 
03408     // If we make it here and it's legal, add it.
03409     (void)InsertFormula(LU, LUIdx, F);
03410   next:;
03411   }
03412 }
03413 
03414 /// GenerateScales - Generate stride factor reuse formulae by making use of
03415 /// scaled-offset address modes, for example.
03416 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) {
03417   // Determine the integer type for the base formula.
03418   Type *IntTy = Base.getType();
03419   if (!IntTy) return;
03420 
03421   // If this Formula already has a scaled register, we can't add another one.
03422   if (Base.Scale != 0) return;
03423 
03424   // Check each interesting stride.
03425   for (SmallSetVector<int64_t, 8>::const_iterator
03426        I = Factors.begin(), E = Factors.end(); I != E; ++I) {
03427     int64_t Factor = *I;
03428 
03429     Base.Scale = Factor;
03430     Base.HasBaseReg = Base.BaseRegs.size() > 1;
03431     // Check whether this scale is going to be legal.
03432     if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
03433                     Base)) {
03434       // As a special-case, handle special out-of-loop Basic users specially.
03435       // TODO: Reconsider this special case.
03436       if (LU.Kind == LSRUse::Basic &&
03437           isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special,
03438                      LU.AccessTy, Base) &&
03439           LU.AllFixupsOutsideLoop)
03440         LU.Kind = LSRUse::Special;
03441       else
03442         continue;
03443     }
03444     // For an ICmpZero, negating a solitary base register won't lead to
03445     // new solutions.
03446     if (LU.Kind == LSRUse::ICmpZero &&
03447         !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV)
03448       continue;
03449     // For each addrec base reg, apply the scale, if possible.
03450     for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
03451       if (const SCEVAddRecExpr *AR =
03452             dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
03453         const SCEV *FactorS = SE.getConstant(IntTy, Factor);
03454         if (FactorS->isZero())
03455           continue;
03456         // Divide out the factor, ignoring high bits, since we'll be
03457         // scaling the value back up in the end.
03458         if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
03459           // TODO: This could be optimized to avoid all the copying.
03460           Formula F = Base;
03461           F.ScaledReg = Quotient;
03462           F.DeleteBaseReg(F.BaseRegs[i]);
03463           (void)InsertFormula(LU, LUIdx, F);
03464         }
03465       }
03466   }
03467 }
03468 
03469 /// GenerateTruncates - Generate reuse formulae from different IV types.
03470 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
03471   // Don't bother truncating symbolic values.
03472   if (Base.BaseGV) return;
03473 
03474   // Determine the integer type for the base formula.
03475   Type *DstTy = Base.getType();
03476   if (!DstTy) return;
03477   DstTy = SE.getEffectiveSCEVType(DstTy);
03478 
03479   for (SmallSetVector<Type *, 4>::const_iterator
03480        I = Types.begin(), E = Types.end(); I != E; ++I) {
03481     Type *SrcTy = *I;
03482     if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) {
03483       Formula F = Base;
03484 
03485       if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
03486       for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
03487            JE = F.BaseRegs.end(); J != JE; ++J)
03488         *J = SE.getAnyExtendExpr(*J, SrcTy);
03489 
03490       // TODO: This assumes we've done basic processing on all uses and
03491       // have an idea what the register usage is.
03492       if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
03493         continue;
03494 
03495       (void)InsertFormula(LU, LUIdx, F);
03496     }
03497   }
03498 }
03499 
03500 namespace {
03501 
03502 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
03503 /// defer modifications so that the search phase doesn't have to worry about
03504 /// the data structures moving underneath it.
03505 struct WorkItem {
03506   size_t LUIdx;
03507   int64_t Imm;
03508   const SCEV *OrigReg;
03509 
03510   WorkItem(size_t LI, int64_t I, const SCEV *R)
03511     : LUIdx(LI), Imm(I), OrigReg(R) {}
03512 
03513   void print(raw_ostream &OS) const;
03514   void dump() const;
03515 };
03516 
03517 }
03518 
03519 void WorkItem::print(raw_ostream &OS) const {
03520   OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
03521      << " , add offset " << Imm;
03522 }
03523 
03524 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
03525 void WorkItem::dump() const {
03526   print(errs()); errs() << '\n';
03527 }
03528 #endif
03529 
03530 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
03531 /// distance apart and try to form reuse opportunities between them.
03532 void LSRInstance::GenerateCrossUseConstantOffsets() {
03533   // Group the registers by their value without any added constant offset.
03534   typedef std::map<int64_t, const SCEV *> ImmMapTy;
03535   typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
03536   RegMapTy Map;
03537   DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
03538   SmallVector<const SCEV *, 8> Sequence;
03539   for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
03540        I != E; ++I) {
03541     const SCEV *Reg = *I;
03542     int64_t Imm = ExtractImmediate(Reg, SE);
03543     std::pair<RegMapTy::iterator, bool> Pair =
03544       Map.insert(std::make_pair(Reg, ImmMapTy()));
03545     if (Pair.second)
03546       Sequence.push_back(Reg);
03547     Pair.first->second.insert(std::make_pair(Imm, *I));
03548     UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
03549   }
03550 
03551   // Now examine each set of registers with the same base value. Build up
03552   // a list of work to do and do the work in a separate step so that we're
03553   // not adding formulae and register counts while we're searching.
03554   SmallVector<WorkItem, 32> WorkItems;
03555   SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
03556   for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
03557        E = Sequence.end(); I != E; ++I) {
03558     const SCEV *Reg = *I;
03559     const ImmMapTy &Imms = Map.find(Reg)->second;
03560 
03561     // It's not worthwhile looking for reuse if there's only one offset.
03562     if (Imms.size() == 1)
03563       continue;
03564 
03565     DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
03566           for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
03567                J != JE; ++J)
03568             dbgs() << ' ' << J->first;
03569           dbgs() << '\n');
03570 
03571     // Examine each offset.
03572     for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
03573          J != JE; ++J) {
03574       const SCEV *OrigReg = J->second;
03575 
03576       int64_t JImm = J->first;
03577       const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
03578 
03579       if (!isa<SCEVConstant>(OrigReg) &&
03580           UsedByIndicesMap[Reg].count() == 1) {
03581         DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
03582         continue;
03583       }
03584 
03585       // Conservatively examine offsets between this orig reg a few selected
03586       // other orig regs.
03587       ImmMapTy::const_iterator OtherImms[] = {
03588         Imms.begin(), std::prev(Imms.end()),
03589         Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) /
03590                          2)
03591       };
03592       for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
03593         ImmMapTy::const_iterator M = OtherImms[i];
03594         if (M == J || M == JE) continue;
03595 
03596         // Compute the difference between the two.
03597         int64_t Imm = (uint64_t)JImm - M->first;
03598         for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
03599              LUIdx = UsedByIndices.find_next(LUIdx))
03600           // Make a memo of this use, offset, and register tuple.
03601           if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
03602             WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
03603       }
03604     }
03605   }
03606 
03607   Map.clear();
03608   Sequence.clear();
03609   UsedByIndicesMap.clear();
03610   UniqueItems.clear();
03611 
03612   // Now iterate through the worklist and add new formulae.
03613   for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
03614        E = WorkItems.end(); I != E; ++I) {
03615     const WorkItem &WI = *I;
03616     size_t LUIdx = WI.LUIdx;
03617     LSRUse &LU = Uses[LUIdx];
03618     int64_t Imm = WI.Imm;
03619     const SCEV *OrigReg = WI.OrigReg;
03620 
03621     Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
03622     const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
03623     unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
03624 
03625     // TODO: Use a more targeted data structure.
03626     for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
03627       const Formula &F = LU.Formulae[L];
03628       // Use the immediate in the scaled register.
03629       if (F.ScaledReg == OrigReg) {
03630         int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale;
03631         // Don't create 50 + reg(-50).
03632         if (F.referencesReg(SE.getSCEV(
03633                    ConstantInt::get(IntTy, -(uint64_t)Offset))))
03634           continue;
03635         Formula NewF = F;
03636         NewF.BaseOffset = Offset;
03637         if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
03638                         NewF))
03639           continue;
03640         NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
03641 
03642         // If the new scale is a constant in a register, and adding the constant
03643         // value to the immediate would produce a value closer to zero than the
03644         // immediate itself, then the formula isn't worthwhile.
03645         if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
03646           if (C->getValue()->isNegative() !=
03647                 (NewF.BaseOffset < 0) &&
03648               (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale))
03649                 .ule(abs64(NewF.BaseOffset)))
03650             continue;
03651 
03652         // OK, looks good.
03653         (void)InsertFormula(LU, LUIdx, NewF);
03654       } else {
03655         // Use the immediate in a base register.
03656         for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
03657           const SCEV *BaseReg = F.BaseRegs[N];
03658           if (BaseReg != OrigReg)
03659             continue;
03660           Formula NewF = F;
03661           NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm;
03662           if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset,
03663                           LU.Kind, LU.AccessTy, NewF)) {
03664             if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
03665               continue;
03666             NewF = F;
03667             NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
03668           }
03669           NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
03670 
03671           // If the new formula has a constant in a register, and adding the
03672           // constant value to the immediate would produce a value closer to
03673           // zero than the immediate itself, then the formula isn't worthwhile.
03674           for (SmallVectorImpl<const SCEV *>::const_iterator
03675                J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
03676                J != JE; ++J)
03677             if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
03678               if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt(
03679                    abs64(NewF.BaseOffset)) &&
03680                   (C->getValue()->getValue() +
03681                    NewF.BaseOffset).countTrailingZeros() >=
03682                    countTrailingZeros<uint64_t>(NewF.BaseOffset))
03683                 goto skip_formula;
03684 
03685           // Ok, looks good.
03686           (void)InsertFormula(LU, LUIdx, NewF);
03687           break;
03688         skip_formula:;
03689         }
03690       }
03691     }
03692   }
03693 }
03694 
03695 /// GenerateAllReuseFormulae - Generate formulae for each use.
03696 void
03697 LSRInstance::GenerateAllReuseFormulae() {
03698   // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
03699   // queries are more precise.
03700   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
03701     LSRUse &LU = Uses[LUIdx];
03702     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
03703       GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
03704     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
03705       GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
03706   }
03707   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
03708     LSRUse &LU = Uses[LUIdx];
03709     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
03710       GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
03711     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
03712       GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
03713     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
03714       GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
03715     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
03716       GenerateScales(LU, LUIdx, LU.Formulae[i]);
03717   }
03718   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
03719     LSRUse &LU = Uses[LUIdx];
03720     for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
03721       GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
03722   }
03723 
03724   GenerateCrossUseConstantOffsets();
03725 
03726   DEBUG(dbgs() << "\n"
03727                   "After generating reuse formulae:\n";
03728         print_uses(dbgs()));
03729 }
03730 
03731 /// If there are multiple formulae with the same set of registers used
03732 /// by other uses, pick the best one and delete the others.
03733 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
03734   DenseSet<const SCEV *> VisitedRegs;
03735   SmallPtrSet<const SCEV *, 16> Regs;
03736   SmallPtrSet<const SCEV *, 16> LoserRegs;
03737 #ifndef NDEBUG
03738   bool ChangedFormulae = false;
03739 #endif
03740 
03741   // Collect the best formula for each unique set of shared registers. This
03742   // is reset for each use.
03743   typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo>
03744     BestFormulaeTy;
03745   BestFormulaeTy BestFormulae;
03746 
03747   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
03748     LSRUse &LU = Uses[LUIdx];
03749     DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n');
03750 
03751     bool Any = false;
03752     for (size_t FIdx = 0, NumForms = LU.Formulae.size();
03753          FIdx != NumForms; ++FIdx) {
03754       Formula &F = LU.Formulae[FIdx];
03755 
03756       // Some formulas are instant losers. For example, they may depend on
03757       // nonexistent AddRecs from other loops. These need to be filtered
03758       // immediately, otherwise heuristics could choose them over others leading
03759       // to an unsatisfactory solution. Passing LoserRegs into RateFormula here
03760       // avoids the need to recompute this information across formulae using the
03761       // same bad AddRec. Passing LoserRegs is also essential unless we remove
03762       // the corresponding bad register from the Regs set.
03763       Cost CostF;
03764       Regs.clear();
03765       CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU,
03766                         &LoserRegs);
03767       if (CostF.isLoser()) {
03768         // During initial formula generation, undesirable formulae are generated
03769         // by uses within other loops that have some non-trivial address mode or
03770         // use the postinc form of the IV. LSR needs to provide these formulae
03771         // as the basis of rediscovering the desired formula that uses an AddRec
03772         // corresponding to the existing phi. Once all formulae have been
03773         // generated, these initial losers may be pruned.
03774         DEBUG(dbgs() << "  Filtering loser "; F.print(dbgs());
03775               dbgs() << "\n");
03776       }
03777       else {
03778         SmallVector<const SCEV *, 4> Key;
03779         for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
03780                JE = F.BaseRegs.end(); J != JE; ++J) {
03781           const SCEV *Reg = *J;
03782           if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
03783             Key.push_back(Reg);
03784         }
03785         if (F.ScaledReg &&
03786             RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
03787           Key.push_back(F.ScaledReg);
03788         // Unstable sort by host order ok, because this is only used for
03789         // uniquifying.
03790         std::sort(Key.begin(), Key.end());
03791 
03792         std::pair<BestFormulaeTy::const_iterator, bool> P =
03793           BestFormulae.insert(std::make_pair(Key, FIdx));
03794         if (P.second)
03795           continue;
03796 
03797         Formula &Best = LU.Formulae[P.first->second];
03798 
03799         Cost CostBest;
03800         Regs.clear();
03801         CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE,
03802                              DT, LU);
03803         if (CostF < CostBest)
03804           std::swap(F, Best);
03805         DEBUG(dbgs() << "  Filtering out formula "; F.print(dbgs());
03806               dbgs() << "\n"
03807                         "    in favor of formula "; Best.print(dbgs());
03808               dbgs() << '\n');
03809       }
03810 #ifndef NDEBUG
03811       ChangedFormulae = true;
03812 #endif
03813       LU.DeleteFormula(F);
03814       --FIdx;
03815       --NumForms;
03816       Any = true;
03817     }
03818 
03819     // Now that we've filtered out some formulae, recompute the Regs set.
03820     if (Any)
03821       LU.RecomputeRegs(LUIdx, RegUses);
03822 
03823     // Reset this to prepare for the next use.
03824     BestFormulae.clear();
03825   }
03826 
03827   DEBUG(if (ChangedFormulae) {
03828           dbgs() << "\n"
03829                     "After filtering out undesirable candidates:\n";
03830           print_uses(dbgs());
03831         });
03832 }
03833 
03834 // This is a rough guess that seems to work fairly well.
03835 static const size_t ComplexityLimit = UINT16_MAX;
03836 
03837 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of
03838 /// solutions the solver might have to consider. It almost never considers
03839 /// this many solutions because it prune the search space, but the pruning
03840 /// isn't always sufficient.
03841 size_t LSRInstance::EstimateSearchSpaceComplexity() const {
03842   size_t Power = 1;
03843   for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
03844        E = Uses.end(); I != E; ++I) {
03845     size_t FSize = I->Formulae.size();
03846     if (FSize >= ComplexityLimit) {
03847       Power = ComplexityLimit;
03848       break;
03849     }
03850     Power *= FSize;
03851     if (Power >= ComplexityLimit)
03852       break;
03853   }
03854   return Power;
03855 }
03856 
03857 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset
03858 /// of the registers of another formula, it won't help reduce register
03859 /// pressure (though it may not necessarily hurt register pressure); remove
03860 /// it to simplify the system.
03861 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() {
03862   if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
03863     DEBUG(dbgs() << "The search space is too complex.\n");
03864 
03865     DEBUG(dbgs() << "Narrowing the search space by eliminating formulae "
03866                     "which use a superset of registers used by other "
03867                     "formulae.\n");
03868 
03869     for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
03870       LSRUse &LU = Uses[LUIdx];
03871       bool Any = false;
03872       for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
03873         Formula &F = LU.Formulae[i];
03874         // Look for a formula with a constant or GV in a register. If the use
03875         // also has a formula with that same value in an immediate field,
03876         // delete the one that uses a register.
03877         for (SmallVectorImpl<const SCEV *>::const_iterator
03878              I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) {
03879           if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) {
03880             Formula NewF = F;
03881             NewF.BaseOffset += C->getValue()->getSExtValue();
03882             NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
03883                                 (I - F.BaseRegs.begin()));
03884             if (LU.HasFormulaWithSameRegs(NewF)) {
03885               DEBUG(dbgs() << "  Deleting "; F.print(dbgs()); dbgs() << '\n');
03886               LU.DeleteFormula(F);
03887               --i;
03888               --e;
03889               Any = true;
03890               break;
03891             }
03892           } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) {
03893             if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue()))
03894               if (!F.BaseGV) {
03895                 Formula NewF = F;
03896                 NewF.BaseGV = GV;
03897                 NewF.BaseRegs.erase(NewF.BaseRegs.begin() +
03898                                     (I - F.BaseRegs.begin()));
03899                 if (LU.HasFormulaWithSameRegs(NewF)) {
03900                   DEBUG(dbgs() << "  Deleting "; F.print(dbgs());
03901                         dbgs() << '\n');
03902                   LU.DeleteFormula(F);
03903                   --i;
03904                   --e;
03905                   Any = true;
03906                   break;
03907                 }
03908               }
03909           }
03910         }
03911       }
03912       if (Any)
03913         LU.RecomputeRegs(LUIdx, RegUses);
03914     }
03915 
03916     DEBUG(dbgs() << "After pre-selection:\n";
03917           print_uses(dbgs()));
03918   }
03919 }
03920 
03921 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers
03922 /// for expressions like A, A+1, A+2, etc., allocate a single register for
03923 /// them.
03924 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
03925   if (EstimateSearchSpaceComplexity() < ComplexityLimit)
03926     return;
03927 
03928   DEBUG(dbgs() << "The search space is too complex.\n"
03929                   "Narrowing the search space by assuming that uses separated "
03930                   "by a constant offset will use the same registers.\n");
03931 
03932   // This is especially useful for unrolled loops.
03933 
03934   for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
03935     LSRUse &LU = Uses[LUIdx];
03936     for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
03937          E = LU.Formulae.end(); I != E; ++I) {
03938       const Formula &F = *I;
03939       if (F.BaseOffset == 0 || F.Scale != 0)
03940         continue;
03941 
03942       LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU);
03943       if (!LUThatHas)
03944         continue;
03945 
03946       if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false,
03947                               LU.Kind, LU.AccessTy))
03948         continue;
03949 
03950       DEBUG(dbgs() << "  Deleting use "; LU.print(dbgs()); dbgs() << '\n');
03951 
03952       LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
03953 
03954       // Update the relocs to reference the new use.
03955       for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
03956            E = Fixups.end(); I != E; ++I) {
03957         LSRFixup &Fixup = *I;
03958         if (Fixup.LUIdx == LUIdx) {
03959           Fixup.LUIdx = LUThatHas - &Uses.front();
03960           Fixup.Offset += F.BaseOffset;
03961           // Add the new offset to LUThatHas' offset list.
03962           if (LUThatHas->Offsets.back() != Fixup.Offset) {
03963             LUThatHas->Offsets.push_back(Fixup.Offset);
03964             if (Fixup.Offset > LUThatHas->MaxOffset)
03965               LUThatHas->MaxOffset = Fixup.Offset;
03966             if (Fixup.Offset < LUThatHas->MinOffset)
03967               LUThatHas->MinOffset = Fixup.Offset;
03968           }
03969           DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n');
03970         }
03971         if (Fixup.LUIdx == NumUses-1)
03972           Fixup.LUIdx = LUIdx;
03973       }
03974 
03975       // Delete formulae from the new use which are no longer legal.
03976       bool Any = false;
03977       for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
03978         Formula &F = LUThatHas->Formulae[i];
03979         if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset,
03980                         LUThatHas->Kind, LUThatHas->AccessTy, F)) {
03981           DEBUG(dbgs() << "  Deleting "; F.print(dbgs());
03982                 dbgs() << '\n');
03983           LUThatHas->DeleteFormula(F);
03984           --i;
03985           --e;
03986           Any = true;
03987         }
03988       }
03989 
03990       if (Any)
03991         LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
03992 
03993       // Delete the old use.
03994       DeleteUse(LU, LUIdx);
03995       --LUIdx;
03996       --NumUses;
03997       break;
03998     }
03999   }
04000 
04001   DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs()));
04002 }
04003 
04004 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call
04005 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that
04006 /// we've done more filtering, as it may be able to find more formulae to
04007 /// eliminate.
04008 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){
04009   if (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
04010     DEBUG(dbgs() << "The search space is too complex.\n");
04011 
04012     DEBUG(dbgs() << "Narrowing the search space by re-filtering out "
04013                     "undesirable dedicated registers.\n");
04014 
04015     FilterOutUndesirableDedicatedRegisters();
04016 
04017     DEBUG(dbgs() << "After pre-selection:\n";
04018           print_uses(dbgs()));
04019   }
04020 }
04021 
04022 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely
04023 /// to be profitable, and then in any use which has any reference to that
04024 /// register, delete all formulae which do not reference that register.
04025 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() {
04026   // With all other options exhausted, loop until the system is simple
04027   // enough to handle.
04028   SmallPtrSet<const SCEV *, 4> Taken;
04029   while (EstimateSearchSpaceComplexity() >= ComplexityLimit) {
04030     // Ok, we have too many of formulae on our hands to conveniently handle.
04031     // Use a rough heuristic to thin out the list.
04032     DEBUG(dbgs() << "The search space is too complex.\n");
04033 
04034     // Pick the register which is used by the most LSRUses, which is likely
04035     // to be a good reuse register candidate.
04036     const SCEV *Best = 0;
04037     unsigned BestNum = 0;
04038     for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
04039          I != E; ++I) {
04040       const SCEV *Reg = *I;
04041       if (Taken.count(Reg))
04042         continue;
04043       if (!Best)
04044         Best = Reg;
04045       else {
04046         unsigned Count = RegUses.getUsedByIndices(Reg).count();
04047         if (Count > BestNum) {
04048           Best = Reg;
04049           BestNum = Count;
04050         }
04051       }
04052     }
04053 
04054     DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
04055                  << " will yield profitable reuse.\n");
04056     Taken.insert(Best);
04057 
04058     // In any use with formulae which references this register, delete formulae
04059     // which don't reference it.
04060     for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
04061       LSRUse &LU = Uses[LUIdx];
04062       if (!LU.Regs.count(Best)) continue;
04063 
04064       bool Any = false;
04065       for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
04066         Formula &F = LU.Formulae[i];
04067         if (!F.referencesReg(Best)) {
04068           DEBUG(dbgs() << "  Deleting "; F.print(dbgs()); dbgs() << '\n');
04069           LU.DeleteFormula(F);
04070           --e;
04071           --i;
04072           Any = true;
04073           assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?");
04074           continue;
04075         }
04076       }
04077 
04078       if (Any)
04079         LU.RecomputeRegs(LUIdx, RegUses);
04080     }
04081 
04082     DEBUG(dbgs() << "After pre-selection:\n";
04083           print_uses(dbgs()));
04084   }
04085 }
04086 
04087 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
04088 /// formulae to choose from, use some rough heuristics to prune down the number
04089 /// of formulae. This keeps the main solver from taking an extraordinary amount
04090 /// of time in some worst-case scenarios.
04091 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
04092   NarrowSearchSpaceByDetectingSupersets();
04093   NarrowSearchSpaceByCollapsingUnrolledCode();
04094   NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters();
04095   NarrowSearchSpaceByPickingWinnerRegs();
04096 }
04097 
04098 /// SolveRecurse - This is the recursive solver.
04099 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
04100                                Cost &SolutionCost,
04101                                SmallVectorImpl<const Formula *> &Workspace,
04102                                const Cost &CurCost,
04103                                const SmallPtrSet<const SCEV *, 16> &CurRegs,
04104                                DenseSet<const SCEV *> &VisitedRegs) const {
04105   // Some ideas:
04106   //  - prune more:
04107   //    - use more aggressive filtering
04108   //    - sort the formula so that the most profitable solutions are found first
04109   //    - sort the uses too
04110   //  - search faster:
04111   //    - don't compute a cost, and then compare. compare while computing a cost
04112   //      and bail early.
04113   //    - track register sets with SmallBitVector
04114 
04115   const LSRUse &LU = Uses[Workspace.size()];
04116 
04117   // If this use references any register that's already a part of the
04118   // in-progress solution, consider it a requirement that a formula must
04119   // reference that register in order to be considered. This prunes out
04120   // unprofitable searching.
04121   SmallSetVector<const SCEV *, 4> ReqRegs;
04122   for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
04123        E = CurRegs.end(); I != E; ++I)
04124     if (LU.Regs.count(*I))
04125       ReqRegs.insert(*I);
04126 
04127   SmallPtrSet<const SCEV *, 16> NewRegs;
04128   Cost NewCost;
04129   for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
04130        E = LU.Formulae.end(); I != E; ++I) {
04131     const Formula &F = *I;
04132 
04133     // Ignore formulae which do not use any of the required registers.
04134     bool SatisfiedReqReg = true;
04135     for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
04136          JE = ReqRegs.end(); J != JE; ++J) {
04137       const SCEV *Reg = *J;
04138       if ((!F.ScaledReg || F.ScaledReg != Reg) &&
04139           std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
04140           F.BaseRegs.end()) {
04141         SatisfiedReqReg = false;
04142         break;
04143       }
04144     }
04145     if (!SatisfiedReqReg) {
04146       // If none of the formulae satisfied the required registers, then we could
04147       // clear ReqRegs and try again. Currently, we simply give up in this case.
04148       continue;
04149     }
04150 
04151     // Evaluate the cost of the current formula. If it's already worse than
04152     // the current best, prune the search at that point.
04153     NewCost = CurCost;
04154     NewRegs = CurRegs;
04155     NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT,
04156                         LU);
04157     if (NewCost < SolutionCost) {
04158       Workspace.push_back(&F);
04159       if (Workspace.size() != Uses.size()) {
04160         SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
04161                      NewRegs, VisitedRegs);
04162         if (F.getNumRegs() == 1 && Workspace.size() == 1)
04163           VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
04164       } else {
04165         DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
04166               dbgs() << ".\n Regs:";
04167               for (SmallPtrSet<const SCEV *, 16>::const_iterator
04168                    I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
04169                 dbgs() << ' ' << **I;
04170               dbgs() << '\n');
04171 
04172         SolutionCost = NewCost;
04173         Solution = Workspace;
04174       }
04175       Workspace.pop_back();
04176     }
04177   }
04178 }
04179 
04180 /// Solve - Choose one formula from each use. Return the results in the given
04181 /// Solution vector.
04182 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
04183   SmallVector<const Formula *, 8> Workspace;
04184   Cost SolutionCost;
04185   SolutionCost.Lose();
04186   Cost CurCost;
04187   SmallPtrSet<const SCEV *, 16> CurRegs;
04188   DenseSet<const SCEV *> VisitedRegs;
04189   Workspace.reserve(Uses.size());
04190 
04191   // SolveRecurse does all the work.
04192   SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
04193                CurRegs, VisitedRegs);
04194   if (Solution.empty()) {
04195     DEBUG(dbgs() << "\nNo Satisfactory Solution\n");
04196     return;
04197   }
04198 
04199   // Ok, we've now made all our decisions.
04200   DEBUG(dbgs() << "\n"
04201                   "The chosen solution requires "; SolutionCost.print(dbgs());
04202         dbgs() << ":\n";
04203         for (size_t i = 0, e = Uses.size(); i != e; ++i) {
04204           dbgs() << "  ";
04205           Uses[i].print(dbgs());
04206           dbgs() << "\n"
04207                     "    ";
04208           Solution[i]->print(dbgs());
04209           dbgs() << '\n';
04210         });
04211 
04212   assert(Solution.size() == Uses.size() && "Malformed solution!");
04213 }
04214 
04215 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
04216 /// the dominator tree far as we can go while still being dominated by the
04217 /// input positions. This helps canonicalize the insert position, which
04218 /// encourages sharing.
04219 BasicBlock::iterator
04220 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
04221                                  const SmallVectorImpl<Instruction *> &Inputs)
04222                                                                          const {
04223   for (;;) {
04224     const Loop *IPLoop = LI.getLoopFor(IP->getParent());
04225     unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
04226 
04227     BasicBlock *IDom;
04228     for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
04229       if (!Rung) return IP;
04230       Rung = Rung->getIDom();
04231       if (!Rung) return IP;
04232       IDom = Rung->getBlock();
04233 
04234       // Don't climb into a loop though.
04235       const Loop *IDomLoop = LI.getLoopFor(IDom);
04236       unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
04237       if (IDomDepth <= IPLoopDepth &&
04238           (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
04239         break;
04240     }
04241 
04242     bool AllDominate = true;
04243     Instruction *BetterPos = 0;
04244     Instruction *Tentative = IDom->getTerminator();
04245     for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
04246          E = Inputs.end(); I != E; ++I) {
04247       Instruction *Inst = *I;
04248       if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
04249         AllDominate = false;
04250         break;
04251       }
04252       // Attempt to find an insert position in the middle of the block,
04253       // instead of at the end, so that it can be used for other expansions.
04254       if (IDom == Inst->getParent() &&
04255           (!BetterPos || !DT.dominates(Inst, BetterPos)))
04256         BetterPos = std::next(BasicBlock::iterator(Inst));
04257     }
04258     if (!AllDominate)
04259       break;
04260     if (BetterPos)
04261       IP = BetterPos;
04262     else
04263       IP = Tentative;
04264   }
04265 
04266   return IP;
04267 }
04268 
04269 /// AdjustInsertPositionForExpand - Determine an input position which will be
04270 /// dominated by the operands and which will dominate the result.
04271 BasicBlock::iterator
04272 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP,
04273                                            const LSRFixup &LF,
04274                                            const LSRUse &LU,
04275                                            SCEVExpander &Rewriter) const {
04276   // Collect some instructions which must be dominated by the
04277   // expanding replacement. These must be dominated by any operands that
04278   // will be required in the expansion.
04279   SmallVector<Instruction *, 4> Inputs;
04280   if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
04281     Inputs.push_back(I);
04282   if (LU.Kind == LSRUse::ICmpZero)
04283     if (Instruction *I =
04284           dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
04285       Inputs.push_back(I);
04286   if (LF.PostIncLoops.count(L)) {
04287     if (LF.isUseFullyOutsideLoop(L))
04288       Inputs.push_back(L->getLoopLatch()->getTerminator());
04289     else
04290       Inputs.push_back(IVIncInsertPos);
04291   }
04292   // The expansion must also be dominated by the increment positions of any
04293   // loops it for which it is using post-inc mode.
04294   for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
04295        E = LF.PostIncLoops.end(); I != E; ++I) {
04296     const Loop *PIL = *I;
04297     if (PIL == L) continue;
04298 
04299     // Be dominated by the loop exit.
04300     SmallVector<BasicBlock *, 4> ExitingBlocks;
04301     PIL->getExitingBlocks(ExitingBlocks);
04302     if (!ExitingBlocks.empty()) {
04303       BasicBlock *BB = ExitingBlocks[0];
04304       for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
04305         BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
04306       Inputs.push_back(BB->getTerminator());
04307     }
04308   }
04309 
04310   assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP)
04311          && !isa<DbgInfoIntrinsic>(LowestIP) &&
04312          "Insertion point must be a normal instruction");
04313 
04314   // Then, climb up the immediate dominator tree as far as we can go while
04315   // still being dominated by the input positions.
04316   BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs);
04317 
04318   // Don't insert instructions before PHI nodes.
04319   while (isa<PHINode>(IP)) ++IP;
04320 
04321   // Ignore landingpad instructions.
04322   while (isa<LandingPadInst>(IP)) ++IP;
04323 
04324   // Ignore debug intrinsics.
04325   while (isa<DbgInfoIntrinsic>(IP)) ++IP;
04326 
04327   // Set IP below instructions recently inserted by SCEVExpander. This keeps the
04328   // IP consistent across expansions and allows the previously inserted
04329   // instructions to be reused by subsequent expansion.
04330   while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP;
04331 
04332   return IP;
04333 }
04334 
04335 /// Expand - Emit instructions for the leading candidate expression for this
04336 /// LSRUse (this is called "expanding").
04337 Value *LSRInstance::Expand(const LSRFixup &LF,
04338                            const Formula &F,
04339                            BasicBlock::iterator IP,
04340                            SCEVExpander &Rewriter,
04341                            SmallVectorImpl<WeakVH> &DeadInsts) const {
04342   const LSRUse &LU = Uses[LF.LUIdx];
04343   if (LU.RigidFormula)
04344     return LF.OperandValToReplace;
04345 
04346   // Determine an input position which will be dominated by the operands and
04347   // which will dominate the result.
04348   IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter);
04349 
04350   // Inform the Rewriter if we have a post-increment use, so that it can
04351   // perform an advantageous expansion.
04352   Rewriter.setPostInc(LF.PostIncLoops);
04353 
04354   // This is the type that the user actually needs.
04355   Type *OpTy = LF.OperandValToReplace->getType();
04356   // This will be the type that we'll initially expand to.
04357   Type *Ty = F.getType();
04358   if (!Ty)
04359     // No type known; just expand directly to the ultimate type.
04360     Ty = OpTy;
04361   else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
04362     // Expand directly to the ultimate type if it's the right size.
04363     Ty = OpTy;
04364   // This is the type to do integer arithmetic in.
04365   Type *IntTy = SE.getEffectiveSCEVType(Ty);
04366 
04367   // Build up a list of operands to add together to form the full base.
04368   SmallVector<const SCEV *, 8> Ops;
04369 
04370   // Expand the BaseRegs portion.
04371   for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
04372        E = F.BaseRegs.end(); I != E; ++I) {
04373     const SCEV *Reg = *I;
04374     assert(!Reg->isZero() && "Zero allocated in a base register!");
04375 
04376     // If we're expanding for a post-inc user, make the post-inc adjustment.
04377     PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
04378     Reg = TransformForPostIncUse(Denormalize, Reg,
04379                                  LF.UserInst, LF.OperandValToReplace,
04380                                  Loops, SE, DT);
04381 
04382     Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
04383   }
04384 
04385   // Expand the ScaledReg portion.
04386   Value *ICmpScaledV = 0;
04387   if (F.Scale != 0) {
04388     const SCEV *ScaledS = F.ScaledReg;
04389 
04390     // If we're expanding for a post-inc user, make the post-inc adjustment.
04391     PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
04392     ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
04393                                      LF.UserInst, LF.OperandValToReplace,
04394                                      Loops, SE, DT);
04395 
04396     if (LU.Kind == LSRUse::ICmpZero) {
04397       // An interesting way of "folding" with an icmp is to use a negated
04398       // scale, which we'll implement by inserting it into the other operand
04399       // of the icmp.
04400       assert(F.Scale == -1 &&
04401              "The only scale supported by ICmpZero uses is -1!");
04402       ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
04403     } else {
04404       // Otherwise just expand the scaled register and an explicit scale,
04405       // which is expected to be matched as part of the address.
04406 
04407       // Flush the operand list to suppress SCEVExpander hoisting address modes.
04408       if (!Ops.empty() && LU.Kind == LSRUse::Address) {
04409         Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
04410         Ops.clear();
04411         Ops.push_back(SE.getUnknown(FullV));
04412       }
04413       ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
04414       ScaledS = SE.getMulExpr(ScaledS,
04415                               SE.getConstant(ScaledS->getType(), F.Scale));
04416       Ops.push_back(ScaledS);
04417     }
04418   }
04419 
04420   // Expand the GV portion.
04421   if (F.BaseGV) {
04422     // Flush the operand list to suppress SCEVExpander hoisting.
04423     if (!Ops.empty()) {
04424       Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
04425       Ops.clear();
04426       Ops.push_back(SE.getUnknown(FullV));
04427     }
04428     Ops.push_back(SE.getUnknown(F.BaseGV));
04429   }
04430 
04431   // Flush the operand list to suppress SCEVExpander hoisting of both folded and
04432   // unfolded offsets. LSR assumes they both live next to their uses.
04433   if (!Ops.empty()) {
04434     Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
04435     Ops.clear();
04436     Ops.push_back(SE.getUnknown(FullV));
04437   }
04438 
04439   // Expand the immediate portion.
04440   int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset;
04441   if (Offset != 0) {
04442     if (LU.Kind == LSRUse::ICmpZero) {
04443       // The other interesting way of "folding" with an ICmpZero is to use a
04444       // negated immediate.
04445       if (!ICmpScaledV)
04446         ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset);
04447       else {
04448         Ops.push_back(SE.getUnknown(ICmpScaledV));
04449         ICmpScaledV = ConstantInt::get(IntTy, Offset);
04450       }
04451     } else {
04452       // Just add the immediate values. These again are expected to be matched
04453       // as part of the address.
04454       Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
04455     }
04456   }
04457 
04458   // Expand the unfolded offset portion.
04459   int64_t UnfoldedOffset = F.UnfoldedOffset;
04460   if (UnfoldedOffset != 0) {
04461     // Just add the immediate values.
04462     Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy,
04463                                                        UnfoldedOffset)));
04464   }
04465 
04466   // Emit instructions summing all the operands.
04467   const SCEV *FullS = Ops.empty() ?
04468                       SE.getConstant(IntTy, 0) :
04469                       SE.getAddExpr(Ops);
04470   Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
04471 
04472   // We're done expanding now, so reset the rewriter.
04473   Rewriter.clearPostInc();
04474 
04475   // An ICmpZero Formula represents an ICmp which we're handling as a
04476   // comparison against zero. Now that we've expanded an expression for that
04477   // form, update the ICmp's other operand.
04478   if (LU.Kind == LSRUse::ICmpZero) {
04479     ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
04480     DeadInsts.push_back(CI->getOperand(1));
04481     assert(!F.BaseGV && "ICmp does not support folding a global value and "
04482                            "a scale at the same time!");
04483     if (F.Scale == -1) {
04484       if (ICmpScaledV->getType() != OpTy) {
04485         Instruction *Cast =
04486           CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
04487                                                    OpTy, false),
04488                            ICmpScaledV, OpTy, "tmp", CI);
04489         ICmpScaledV = Cast;
04490       }
04491       CI->setOperand(1, ICmpScaledV);
04492     } else {
04493       assert(F.Scale == 0 &&
04494              "ICmp does not support folding a global value and "
04495              "a scale at the same time!");
04496       Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
04497                                            -(uint64_t)Offset);
04498       if (C->getType() != OpTy)
04499         C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
04500                                                           OpTy, false),
04501                                   C, OpTy);
04502 
04503       CI->setOperand(1, C);
04504     }
04505   }
04506 
04507   return FullV;
04508 }
04509 
04510 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
04511 /// of their operands effectively happens in their predecessor blocks, so the
04512 /// expression may need to be expanded in multiple places.
04513 void LSRInstance::RewriteForPHI(PHINode *PN,
04514                                 const LSRFixup &LF,
04515                                 const Formula &F,
04516                                 SCEVExpander &Rewriter,
04517                                 SmallVectorImpl<WeakVH> &DeadInsts,
04518                                 Pass *P) const {
04519   DenseMap<BasicBlock *, Value *> Inserted;
04520   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
04521     if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
04522       BasicBlock *BB = PN->getIncomingBlock(i);
04523 
04524       // If this is a critical edge, split the edge so that we do not insert
04525       // the code on all predecessor/successor paths.  We do this unless this
04526       // is the canonical backedge for this loop, which complicates post-inc
04527       // users.
04528       if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
04529           !isa<IndirectBrInst>(BB->getTerminator())) {
04530         BasicBlock *Parent = PN->getParent();
04531         Loop *PNLoop = LI.getLoopFor(Parent);
04532         if (!PNLoop || Parent != PNLoop->getHeader()) {
04533           // Split the critical edge.
04534           BasicBlock *NewBB = 0;
04535           if (!Parent->isLandingPad()) {
04536             NewBB = SplitCriticalEdge(BB, Parent, P,
04537                                       /*MergeIdenticalEdges=*/true,
04538                                       /*DontDeleteUselessPhis=*/true);
04539           } else {
04540             SmallVector<BasicBlock*, 2> NewBBs;
04541             SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs);
04542             NewBB = NewBBs[0];
04543           }
04544           // If NewBB==NULL, then SplitCriticalEdge refused to split because all
04545           // phi predecessors are identical. The simple thing to do is skip
04546           // splitting in this case rather than complicate the API.
04547           if (NewBB) {
04548             // If PN is outside of the loop and BB is in the loop, we want to
04549             // move the block to be immediately before the PHI block, not
04550             // immediately after BB.
04551             if (L->contains(BB) && !L->contains(PN))
04552               NewBB->moveBefore(PN->getParent());
04553 
04554             // Splitting the edge can reduce the number of PHI entries we have.
04555             e = PN->getNumIncomingValues();
04556             BB = NewBB;
04557             i = PN->getBasicBlockIndex(BB);
04558           }
04559         }
04560       }
04561 
04562       std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
04563         Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
04564       if (!Pair.second)
04565         PN->setIncomingValue(i, Pair.first->second);
04566       else {
04567         Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
04568 
04569         // If this is reuse-by-noop-cast, insert the noop cast.
04570         Type *OpTy = LF.OperandValToReplace->getType();
04571         if (FullV->getType() != OpTy)
04572           FullV =
04573             CastInst::Create(CastInst::getCastOpcode(FullV, false,
04574                                                      OpTy, false),
04575                              FullV, LF.OperandValToReplace->getType(),
04576                              "tmp", BB->getTerminator());
04577 
04578         PN->setIncomingValue(i, FullV);
04579         Pair.first->second = FullV;
04580       }
04581     }
04582 }
04583 
04584 /// Rewrite - Emit instructions for the leading candidate expression for this
04585 /// LSRUse (this is called "expanding"), and update the UserInst to reference
04586 /// the newly expanded value.
04587 void LSRInstance::Rewrite(const LSRFixup &LF,
04588                           const Formula &F,
04589                           SCEVExpander &Rewriter,
04590                           SmallVectorImpl<WeakVH> &DeadInsts,
04591                           Pass *P) const {
04592   // First, find an insertion point that dominates UserInst. For PHI nodes,
04593   // find the nearest block which dominates all the relevant uses.
04594   if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
04595     RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
04596   } else {
04597     Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
04598 
04599     // If this is reuse-by-noop-cast, insert the noop cast.
04600     Type *OpTy = LF.OperandValToReplace->getType();
04601     if (FullV->getType() != OpTy) {
04602       Instruction *Cast =
04603         CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
04604                          FullV, OpTy, "tmp", LF.UserInst);
04605       FullV = Cast;
04606     }
04607 
04608     // Update the user. ICmpZero is handled specially here (for now) because
04609     // Expand may have updated one of the operands of the icmp already, and
04610     // its new value may happen to be equal to LF.OperandValToReplace, in
04611     // which case doing replaceUsesOfWith leads to replacing both operands
04612     // with the same value. TODO: Reorganize this.
04613     if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
04614       LF.UserInst->setOperand(0, FullV);
04615     else
04616       LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
04617   }
04618 
04619   DeadInsts.push_back(LF.OperandValToReplace);
04620 }
04621 
04622 /// ImplementSolution - Rewrite all the fixup locations with new values,
04623 /// following the chosen solution.
04624 void
04625 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
04626                                Pass *P) {
04627   // Keep track of instructions we may have made dead, so that
04628   // we can remove them after we are done working.
04629   SmallVector<WeakVH, 16> DeadInsts;
04630 
04631   SCEVExpander Rewriter(SE, "lsr");
04632 #ifndef NDEBUG
04633   Rewriter.setDebugType(DEBUG_TYPE);
04634 #endif
04635   Rewriter.disableCanonicalMode();
04636   Rewriter.enableLSRMode();
04637   Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
04638 
04639   // Mark phi nodes that terminate chains so the expander tries to reuse them.
04640   for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
04641          ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
04642     if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst()))
04643       Rewriter.setChainedPhi(PN);
04644   }
04645 
04646   // Expand the new value definitions and update the users.
04647   for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
04648        E = Fixups.end(); I != E; ++I) {
04649     const LSRFixup &Fixup = *I;
04650 
04651     Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P);
04652 
04653     Changed = true;
04654   }
04655 
04656   for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(),
04657          ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) {
04658     GenerateIVChain(*ChainI, Rewriter, DeadInsts);
04659     Changed = true;
04660   }
04661   // Clean up after ourselves. This must be done before deleting any
04662   // instructions.
04663   Rewriter.clear();
04664 
04665   Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
04666 }
04667 
04668 LSRInstance::LSRInstance(Loop *L, Pass *P)
04669     : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()),
04670       DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()),
04671       LI(P->getAnalysis<LoopInfo>()),
04672       TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false),
04673       IVIncInsertPos(0) {
04674   // If LoopSimplify form is not available, stay out of trouble.
04675   if (!L->isLoopSimplifyForm())
04676     return;
04677 
04678   // If there's no interesting work to be done, bail early.
04679   if (IU.empty()) return;
04680 
04681   // If there's too much analysis to be done, bail early. We won't be able to
04682   // model the problem anyway.
04683   unsigned NumUsers = 0;
04684   for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
04685     if (++NumUsers > MaxIVUsers) {
04686       DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L
04687             << "\n");
04688       return;
04689     }
04690   }
04691 
04692 #ifndef NDEBUG
04693   // All dominating loops must have preheaders, or SCEVExpander may not be able
04694   // to materialize an AddRecExpr whose Start is an outer AddRecExpr.
04695   //
04696   // IVUsers analysis should only create users that are dominated by simple loop
04697   // headers. Since this loop should dominate all of its users, its user list
04698   // should be empty if this loop itself is not within a simple loop nest.
04699   for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader());
04700        Rung; Rung = Rung->getIDom()) {
04701     BasicBlock *BB = Rung->getBlock();
04702     const Loop *DomLoop = LI.getLoopFor(BB);
04703     if (DomLoop && DomLoop->getHeader() == BB) {
04704       assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest");
04705     }
04706   }
04707 #endif // DEBUG
04708 
04709   DEBUG(dbgs() << "\nLSR on loop ";
04710         L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false);
04711         dbgs() << ":\n");
04712 
04713   // First, perform some low-level loop optimizations.
04714   OptimizeShadowIV();
04715   OptimizeLoopTermCond();
04716 
04717   // If loop preparation eliminates all interesting IV users, bail.
04718   if (IU.empty()) return;
04719 
04720   // Skip nested loops until we can model them better with formulae.
04721   if (!L->empty()) {
04722     DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n");
04723     return;
04724   }
04725 
04726   // Start collecting data and preparing for the solver.
04727   CollectChains();
04728   CollectInterestingTypesAndFactors();
04729   CollectFixupsAndInitialFormulae();
04730   CollectLoopInvariantFixupsAndFormulae();
04731 
04732   assert(!Uses.empty() && "IVUsers reported at least one use");
04733   DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
04734         print_uses(dbgs()));
04735 
04736   // Now use the reuse data to generate a bunch of interesting ways
04737   // to formulate the values needed for the uses.
04738   GenerateAllReuseFormulae();
04739 
04740   FilterOutUndesirableDedicatedRegisters();
04741   NarrowSearchSpaceUsingHeuristics();
04742 
04743   SmallVector<const Formula *, 8> Solution;
04744   Solve(Solution);
04745 
04746   // Release memory that is no longer needed.
04747   Factors.clear();
04748   Types.clear();
04749   RegUses.clear();
04750 
04751   if (Solution.empty())
04752     return;
04753 
04754 #ifndef NDEBUG
04755   // Formulae should be legal.
04756   for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end();
04757        I != E; ++I) {
04758     const LSRUse &LU = *I;
04759     for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
04760                                                   JE = LU.Formulae.end();
04761          J != JE; ++J)
04762       assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy,
04763                         *J) && "Illegal formula generated!");
04764   };
04765 #endif
04766 
04767   // Now that we've decided what we want, make it so.
04768   ImplementSolution(Solution, P);
04769 }
04770 
04771 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
04772   if (Factors.empty() && Types.empty()) return;
04773 
04774   OS << "LSR has identified the following interesting factors and types: ";
04775   bool First = true;
04776 
04777   for (SmallSetVector<int64_t, 8>::const_iterator
04778        I = Factors.begin(), E = Factors.end(); I != E; ++I) {
04779     if (!First) OS << ", ";
04780     First = false;
04781     OS << '*' << *I;
04782   }
04783 
04784   for (SmallSetVector<Type *, 4>::const_iterator
04785        I = Types.begin(), E = Types.end(); I != E; ++I) {
04786     if (!First) OS << ", ";
04787     First = false;
04788     OS << '(' << **I << ')';
04789   }
04790   OS << '\n';
04791 }
04792 
04793 void LSRInstance::print_fixups(raw_ostream &OS) const {
04794   OS << "LSR is examining the following fixup sites:\n";
04795   for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
04796        E = Fixups.end(); I != E; ++I) {
04797     dbgs() << "  ";
04798     I->print(OS);
04799     OS << '\n';
04800   }
04801 }
04802 
04803 void LSRInstance::print_uses(raw_ostream &OS) const {
04804   OS << "LSR is examining the following uses:\n";
04805   for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
04806        E = Uses.end(); I != E; ++I) {
04807     const LSRUse &LU = *I;
04808     dbgs() << "  ";
04809     LU.print(OS);
04810     OS << '\n';
04811     for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
04812          JE = LU.Formulae.end(); J != JE; ++J) {
04813       OS << "    ";
04814       J->print(OS);
04815       OS << '\n';
04816     }
04817   }
04818 }
04819 
04820 void LSRInstance::print(raw_ostream &OS) const {
04821   print_factors_and_types(OS);
04822   print_fixups(OS);
04823   print_uses(OS);
04824 }
04825 
04826 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
04827 void LSRInstance::dump() const {
04828   print(errs()); errs() << '\n';
04829 }
04830 #endif
04831 
04832 namespace {
04833 
04834 class LoopStrengthReduce : public LoopPass {
04835 public:
04836   static char ID; // Pass ID, replacement for typeid
04837   LoopStrengthReduce();
04838 
04839 private:
04840   bool runOnLoop(Loop *L, LPPassManager &LPM) override;
04841   void getAnalysisUsage(AnalysisUsage &AU) const override;
04842 };
04843 
04844 }
04845 
04846 char LoopStrengthReduce::ID = 0;
04847 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce",
04848                 "Loop Strength Reduction", false, false)
04849 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
04850 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
04851 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
04852 INITIALIZE_PASS_DEPENDENCY(IVUsers)
04853 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
04854 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
04855 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
04856                 "Loop Strength Reduction", false, false)
04857 
04858 
04859 Pass *llvm::createLoopStrengthReducePass() {
04860   return new LoopStrengthReduce();
04861 }
04862 
04863 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) {
04864   initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
04865 }
04866 
04867 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
04868   // We split critical edges, so we change the CFG.  However, we do update
04869   // many analyses if they are around.
04870   AU.addPreservedID(LoopSimplifyID);
04871 
04872   AU.addRequired<LoopInfo>();
04873   AU.addPreserved<LoopInfo>();
04874   AU.addRequiredID(LoopSimplifyID);
04875   AU.addRequired<DominatorTreeWrapperPass>();
04876   AU.addPreserved<DominatorTreeWrapperPass>();
04877   AU.addRequired<ScalarEvolution>();
04878   AU.addPreserved<ScalarEvolution>();
04879   // Requiring LoopSimplify a second time here prevents IVUsers from running
04880   // twice, since LoopSimplify was invalidated by running ScalarEvolution.
04881   AU.addRequiredID(LoopSimplifyID);
04882   AU.addRequired<IVUsers>();
04883   AU.addPreserved<IVUsers>();
04884   AU.addRequired<TargetTransformInfo>();
04885 }
04886 
04887 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
04888   if (skipOptnoneFunction(L))
04889     return false;
04890 
04891   bool Changed = false;
04892 
04893   // Run the main LSR transformation.
04894   Changed |= LSRInstance(L, this).getChanged();
04895 
04896   // Remove any extra phis created by processing inner loops.
04897   Changed |= DeleteDeadPHIs(L->getHeader());
04898   if (EnablePhiElim && L->isLoopSimplifyForm()) {
04899     SmallVector<WeakVH, 16> DeadInsts;
04900     SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr");
04901 #ifndef NDEBUG
04902     Rewriter.setDebugType(DEBUG_TYPE);
04903 #endif
04904     unsigned numFolded = Rewriter.replaceCongruentIVs(
04905         L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts,
04906         &getAnalysis<TargetTransformInfo>());
04907     if (numFolded) {
04908       Changed = true;
04909       DeleteTriviallyDeadInstructions(DeadInsts);
04910       DeleteDeadPHIs(L->getHeader());
04911     }
04912   }
04913   return Changed;
04914 }