X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FScalarEvolution.cpp;h=d1274d78a30575aad9e8d2acbffc5c1262e6c66b;hb=12af22e8cc217827cf4f118b0f5e4ebbda9925ae;hp=4afbadeaf52e13b55ea34791e30765c5596306e0;hpb=bd7cba0d8114059dfa3550cbe85e4c50ca77eae2;p=oota-llvm.git diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp index 4afbadeaf52..d1274d78a30 100644 --- a/lib/Analysis/ScalarEvolution.cpp +++ b/lib/Analysis/ScalarEvolution.cpp @@ -1,4 +1,4 @@ -//===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===// +//===- ScalarEvolution.cpp - Scalar Evolution Analysis --------------------===// // // The LLVM Compiler Infrastructure // @@ -58,16 +58,17 @@ // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "scalar-evolution" #include "llvm/Analysis/ScalarEvolution.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/Statistic.h" +#include "llvm/Analysis/AssumptionTracker.h" #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" #include "llvm/Analysis/ValueTracking.h" +#include "llvm/IR/ConstantRange.h" #include "llvm/IR/Constants.h" #include "llvm/IR/DataLayout.h" #include "llvm/IR/DerivedTypes.h" @@ -80,7 +81,6 @@ #include "llvm/IR/LLVMContext.h" #include "llvm/IR/Operator.h" #include "llvm/Support/CommandLine.h" -#include "llvm/Support/ConstantRange.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/MathExtras.h" @@ -89,6 +89,8 @@ #include using namespace llvm; +#define DEBUG_TYPE "scalar-evolution" + STATISTIC(NumArrayLenItCounts, "Number of trip counts computed with array length"); STATISTIC(NumTripCountsComputed, @@ -112,6 +114,7 @@ VerifySCEV("verify-scev", INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution", "Scalar Evolution Analysis", false, true) +INITIALIZE_PASS_DEPENDENCY(AssumptionTracker) INITIALIZE_PASS_DEPENDENCY(LoopInfo) INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) @@ -182,7 +185,7 @@ void SCEV::print(raw_ostream &OS) const { case scUMaxExpr: case scSMaxExpr: { const SCEVNAryExpr *NAry = cast(this); - const char *OpStr = 0; + const char *OpStr = nullptr; switch (NAry->getSCEVType()) { case scAddExpr: OpStr = " + "; break; case scMulExpr: OpStr = " * "; break; @@ -312,7 +315,7 @@ const SCEV *ScalarEvolution::getConstant(ConstantInt *V) { FoldingSetNodeID ID; ID.AddInteger(scConstant); ID.AddPointer(V); - void *IP = 0; + void *IP = nullptr; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V); UniqueSCEVs.InsertNode(S, IP); @@ -365,7 +368,7 @@ void SCEVUnknown::deleted() { SE->UniqueSCEVs.RemoveNode(this); // Release the value. - setValPtr(0); + setValPtr(nullptr); } void SCEVUnknown::allUsesReplacedWith(Value *New) { @@ -829,7 +832,7 @@ const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op, ID.AddInteger(scTruncate); ID.AddPointer(Op); ID.AddPointer(Ty); - void *IP = 0; + void *IP = nullptr; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; // Fold if the operand is constant. @@ -919,7 +922,7 @@ const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op, ID.AddInteger(scZeroExtend); ID.AddPointer(Op); ID.AddPointer(Ty); - void *IP = 0; + void *IP = nullptr; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; // zext(trunc(x)) --> zext(x) or x or trunc(x) @@ -1072,7 +1075,7 @@ static const SCEV *getOverflowLimitForStep(const SCEV *Step, return SE->getConstant(APInt::getSignedMaxValue(BitWidth) - SE->getSignedRange(Step).getSignedMin()); } - return 0; + return nullptr; } // The recurrence AR has been shown to have no signed wrap. Typically, if we can @@ -1091,19 +1094,18 @@ static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR, // Check for a simple looking step prior to loop entry. const SCEVAddExpr *SA = dyn_cast(Start); if (!SA) - return 0; + return nullptr; // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV // subtraction is expensive. For this purpose, perform a quick and dirty // difference, by checking for Step in the operand list. SmallVector DiffOps; - for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end(); - I != E; ++I) { - if (*I != Step) - DiffOps.push_back(*I); - } + for (const SCEV *Op : SA->operands()) + if (Op != Step) + DiffOps.push_back(Op); + if (DiffOps.size() == SA->getNumOperands()) - return 0; + return nullptr; // This is a postinc AR. Check for overflow on the preinc recurrence using the // same three conditions that getSignExtendedExpr checks. @@ -1139,7 +1141,7 @@ static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR, SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) { return PreStart; } - return 0; + return nullptr; } // Get the normalized sign-extended expression for this AddRec's Start. @@ -1181,7 +1183,7 @@ const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op, ID.AddInteger(scSignExtend); ID.AddPointer(Op); ID.AddPointer(Ty); - void *IP = 0; + void *IP = nullptr; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; // If the input value is provably positive, build a zext instead. @@ -1201,6 +1203,23 @@ const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op, return getTruncateOrSignExtend(X, Ty); } + // sext(C1 + (C2 * x)) --> C1 + sext(C2 * x) if C1 < C2 + if (auto SA = dyn_cast(Op)) { + if (SA->getNumOperands() == 2) { + auto SC1 = dyn_cast(SA->getOperand(0)); + auto SMul = dyn_cast(SA->getOperand(1)); + if (SMul && SC1) { + if (auto SC2 = dyn_cast(SMul->getOperand(0))) { + const APInt &C1 = SC1->getValue()->getValue(); + const APInt &C2 = SC2->getValue()->getValue(); + if (C1.isStrictlyPositive() && C2.isStrictlyPositive() && + C2.ugt(C1) && C2.isPowerOf2()) + return getAddExpr(getSignExtendExpr(SC1, Ty), + getSignExtendExpr(SMul, Ty)); + } + } + } + } // If the input value is a chrec scev, and we can prove that the value // did not overflow the old, smaller, value, we can sign extend all of the // operands (often constants). This allows analysis of something like @@ -1292,6 +1311,22 @@ const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op, L, AR->getNoWrapFlags()); } } + // If Start and Step are constants, check if we can apply this + // transformation: + // sext{C1,+,C2} --> C1 + sext{0,+,C2} if C1 < C2 + auto SC1 = dyn_cast(Start); + auto SC2 = dyn_cast(Step); + if (SC1 && SC2) { + const APInt &C1 = SC1->getValue()->getValue(); + const APInt &C2 = SC2->getValue()->getValue(); + if (C1.isStrictlyPositive() && C2.isStrictlyPositive() && C2.ugt(C1) && + C2.isPowerOf2()) { + Start = getSignExtendExpr(Start, Ty); + const SCEV *NewAR = getAddRecExpr(getConstant(AR->getType(), 0), Step, + L, AR->getNoWrapFlags()); + return getAddExpr(Start, getSignExtendExpr(NewAR, Ty)); + } + } } // The cast wasn't folded; create an explicit cast node. @@ -1340,9 +1375,8 @@ const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op, // Force the cast to be folded into the operands of an addrec. if (const SCEVAddRecExpr *AR = dyn_cast(Op)) { SmallVector Ops; - for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end(); - I != E; ++I) - Ops.push_back(getAnyExtendExpr(*I, Ty)); + for (const SCEV *Op : AR->operands()) + Ops.push_back(getAnyExtendExpr(Op, Ty)); return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW); } @@ -1360,7 +1394,7 @@ const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op, /// what it does, given a sequence of operands that would form an add /// expression like this: /// -/// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r) +/// m + n + 13 + (A * (o + p + (B * (q + m + 29)))) + r + (-1 * r) /// /// where A and B are constants, update the map with these values: /// @@ -1811,7 +1845,7 @@ const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl &Ops, ID.AddInteger(scAddExpr); for (unsigned i = 0, e = Ops.size(); i != e; ++i) ID.AddPointer(Ops[i]); - void *IP = 0; + void *IP = nullptr; SCEVAddExpr *S = static_cast(UniqueSCEVs.FindNodeOrInsertPos(ID, IP)); if (!S) { @@ -2029,71 +2063,66 @@ const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl &Ops, // Okay, if there weren't any loop invariants to be folded, check to see if // there are multiple AddRec's with the same loop induction variable being // multiplied together. If so, we can fold them. + + // {A1,+,A2,+,...,+,An} * {B1,+,B2,+,...,+,Bn} + // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [ + // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z + // ]]],+,...up to x=2n}. + // Note that the arguments to choose() are always integers with values + // known at compile time, never SCEV objects. + // + // The implementation avoids pointless extra computations when the two + // addrec's are of different length (mathematically, it's equivalent to + // an infinite stream of zeros on the right). + bool OpsModified = false; for (unsigned OtherIdx = Idx+1; - OtherIdx < Ops.size() && isa(Ops[OtherIdx]); + OtherIdx != Ops.size() && isa(Ops[OtherIdx]); ++OtherIdx) { - if (AddRecLoop != cast(Ops[OtherIdx])->getLoop()) + const SCEVAddRecExpr *OtherAddRec = + dyn_cast(Ops[OtherIdx]); + if (!OtherAddRec || OtherAddRec->getLoop() != AddRecLoop) continue; - // {A1,+,A2,+,...,+,An} * {B1,+,B2,+,...,+,Bn} - // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [ - // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z - // ]]],+,...up to x=2n}. - // Note that the arguments to choose() are always integers with values - // known at compile time, never SCEV objects. - // - // The implementation avoids pointless extra computations when the two - // addrec's are of different length (mathematically, it's equivalent to - // an infinite stream of zeros on the right). - bool OpsModified = false; - for (; OtherIdx != Ops.size() && isa(Ops[OtherIdx]); - ++OtherIdx) { - const SCEVAddRecExpr *OtherAddRec = - dyn_cast(Ops[OtherIdx]); - if (!OtherAddRec || OtherAddRec->getLoop() != AddRecLoop) - continue; - - bool Overflow = false; - Type *Ty = AddRec->getType(); - bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64; - SmallVector AddRecOps; - for (int x = 0, xe = AddRec->getNumOperands() + - OtherAddRec->getNumOperands() - 1; x != xe && !Overflow; ++x) { - const SCEV *Term = getConstant(Ty, 0); - for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) { - uint64_t Coeff1 = Choose(x, 2*x - y, Overflow); - for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1), - ze = std::min(x+1, (int)OtherAddRec->getNumOperands()); - z < ze && !Overflow; ++z) { - uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow); - uint64_t Coeff; - if (LargerThan64Bits) - Coeff = umul_ov(Coeff1, Coeff2, Overflow); - else - Coeff = Coeff1*Coeff2; - const SCEV *CoeffTerm = getConstant(Ty, Coeff); - const SCEV *Term1 = AddRec->getOperand(y-z); - const SCEV *Term2 = OtherAddRec->getOperand(z); - Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2)); - } + bool Overflow = false; + Type *Ty = AddRec->getType(); + bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64; + SmallVector AddRecOps; + for (int x = 0, xe = AddRec->getNumOperands() + + OtherAddRec->getNumOperands() - 1; x != xe && !Overflow; ++x) { + const SCEV *Term = getConstant(Ty, 0); + for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) { + uint64_t Coeff1 = Choose(x, 2*x - y, Overflow); + for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1), + ze = std::min(x+1, (int)OtherAddRec->getNumOperands()); + z < ze && !Overflow; ++z) { + uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow); + uint64_t Coeff; + if (LargerThan64Bits) + Coeff = umul_ov(Coeff1, Coeff2, Overflow); + else + Coeff = Coeff1*Coeff2; + const SCEV *CoeffTerm = getConstant(Ty, Coeff); + const SCEV *Term1 = AddRec->getOperand(y-z); + const SCEV *Term2 = OtherAddRec->getOperand(z); + Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2)); } - AddRecOps.push_back(Term); - } - if (!Overflow) { - const SCEV *NewAddRec = getAddRecExpr(AddRecOps, AddRec->getLoop(), - SCEV::FlagAnyWrap); - if (Ops.size() == 2) return NewAddRec; - Ops[Idx] = NewAddRec; - Ops.erase(Ops.begin() + OtherIdx); --OtherIdx; - OpsModified = true; - AddRec = dyn_cast(NewAddRec); - if (!AddRec) - break; } + AddRecOps.push_back(Term); + } + if (!Overflow) { + const SCEV *NewAddRec = getAddRecExpr(AddRecOps, AddRec->getLoop(), + SCEV::FlagAnyWrap); + if (Ops.size() == 2) return NewAddRec; + Ops[Idx] = NewAddRec; + Ops.erase(Ops.begin() + OtherIdx); --OtherIdx; + OpsModified = true; + AddRec = dyn_cast(NewAddRec); + if (!AddRec) + break; } - if (OpsModified) - return getMulExpr(Ops); } + if (OpsModified) + return getMulExpr(Ops); // Otherwise couldn't fold anything into this recurrence. Move onto the // next one. @@ -2105,7 +2134,7 @@ const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl &Ops, ID.AddInteger(scMulExpr); for (unsigned i = 0, e = Ops.size(); i != e; ++i) ID.AddPointer(Ops[i]); - void *IP = 0; + void *IP = nullptr; SCEVMulExpr *S = static_cast(UniqueSCEVs.FindNodeOrInsertPos(ID, IP)); if (!S) { @@ -2230,7 +2259,7 @@ const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS, ID.AddInteger(scUDivExpr); ID.AddPointer(LHS); ID.AddPointer(RHS); - void *IP = 0; + void *IP = nullptr; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator), LHS, RHS); @@ -2425,7 +2454,7 @@ ScalarEvolution::getAddRecExpr(SmallVectorImpl &Operands, for (unsigned i = 0, e = Operands.size(); i != e; ++i) ID.AddPointer(Operands[i]); ID.AddPointer(L); - void *IP = 0; + void *IP = nullptr; SCEVAddRecExpr *S = static_cast(UniqueSCEVs.FindNodeOrInsertPos(ID, IP)); if (!S) { @@ -2533,7 +2562,7 @@ ScalarEvolution::getSMaxExpr(SmallVectorImpl &Ops) { ID.AddInteger(scSMaxExpr); for (unsigned i = 0, e = Ops.size(); i != e; ++i) ID.AddPointer(Ops[i]); - void *IP = 0; + void *IP = nullptr; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; const SCEV **O = SCEVAllocator.Allocate(Ops.size()); std::uninitialized_copy(Ops.begin(), Ops.end(), O); @@ -2637,7 +2666,7 @@ ScalarEvolution::getUMaxExpr(SmallVectorImpl &Ops) { ID.AddInteger(scUMaxExpr); for (unsigned i = 0, e = Ops.size(); i != e; ++i) ID.AddPointer(Ops[i]); - void *IP = 0; + void *IP = nullptr; if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; const SCEV **O = SCEVAllocator.Allocate(Ops.size()); std::uninitialized_copy(Ops.begin(), Ops.end(), O); @@ -2704,7 +2733,7 @@ const SCEV *ScalarEvolution::getUnknown(Value *V) { FoldingSetNodeID ID; ID.AddInteger(scUnknown); ID.AddPointer(V); - void *IP = 0; + void *IP = nullptr; if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) { assert(cast(S)->getValue() == V && "Stale SCEVUnknown in uniquing map!"); @@ -3010,7 +3039,7 @@ const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) { return getPointerBase(Cast->getOperand()); } else if (const SCEVNAryExpr *NAry = dyn_cast(V)) { - const SCEV *PtrOp = 0; + const SCEV *PtrOp = nullptr; for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end(); I != E; ++I) { if ((*I)->getType()->isPointerTy()) { @@ -3033,9 +3062,8 @@ static void PushDefUseChildren(Instruction *I, SmallVectorImpl &Worklist) { // Push the def-use children onto the Worklist stack. - for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); - UI != UE; ++UI) - Worklist.push_back(cast(*UI)); + for (User *U : I->users()) + Worklist.push_back(cast(U)); } /// ForgetSymbolicValue - This looks up computed SCEV values for all @@ -3091,20 +3119,20 @@ const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) { // The loop may have multiple entrances or multiple exits; we can analyze // this phi as an addrec if it has a unique entry value and a unique // backedge value. - Value *BEValueV = 0, *StartValueV = 0; + Value *BEValueV = nullptr, *StartValueV = nullptr; for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { Value *V = PN->getIncomingValue(i); if (L->contains(PN->getIncomingBlock(i))) { if (!BEValueV) { BEValueV = V; } else if (BEValueV != V) { - BEValueV = 0; + BEValueV = nullptr; break; } } else if (!StartValueV) { StartValueV = V; } else if (StartValueV != V) { - StartValueV = 0; + StartValueV = nullptr; break; } } @@ -3232,7 +3260,7 @@ const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) { // PHI's incoming blocks are in a different loop, in which case doing so // risks breaking LCSSA form. Instcombine would normally zap these, but // it doesn't have DominatorTree information, so it may miss cases. - if (Value *V = SimplifyInstruction(PN, DL, TLI, DT)) + if (Value *V = SimplifyInstruction(PN, DL, TLI, DT, AT)) if (LI->replacementPreservesLCSSAForm(PN, V)) return getSCEV(V); @@ -3364,7 +3392,7 @@ ScalarEvolution::GetMinTrailingZeros(const SCEV *S) { // For a SCEVUnknown, ask ValueTracking. unsigned BitWidth = getTypeSizeInBits(U->getType()); APInt Zeros(BitWidth, 0), Ones(BitWidth, 0); - ComputeMaskedBits(U->getValue(), Zeros, Ones); + computeKnownBits(U->getValue(), Zeros, Ones, DL, 0, AT, nullptr, DT); return Zeros.countTrailingOnes(); } @@ -3503,7 +3531,7 @@ ScalarEvolution::getUnsignedRange(const SCEV *S) { if (const SCEVUnknown *U = dyn_cast(S)) { // For a SCEVUnknown, ask ValueTracking. APInt Zeros(BitWidth, 0), Ones(BitWidth, 0); - ComputeMaskedBits(U->getValue(), Zeros, Ones, DL); + computeKnownBits(U->getValue(), Zeros, Ones, DL, 0, AT, nullptr, DT); if (Ones == ~Zeros + 1) return setUnsignedRange(U, ConservativeResult); return setUnsignedRange(U, @@ -3655,7 +3683,7 @@ ScalarEvolution::getSignedRange(const SCEV *S) { // For a SCEVUnknown, ask ValueTracking. if (!U->getValue()->getType()->isIntegerTy() && !DL) return setSignedRange(U, ConservativeResult); - unsigned NS = ComputeNumSignBits(U->getValue(), DL); + unsigned NS = ComputeNumSignBits(U->getValue(), DL, 0, AT, nullptr, DT); if (NS <= 1) return setSignedRange(U, ConservativeResult); return setSignedRange(U, ConservativeResult.intersectWith( @@ -3756,13 +3784,14 @@ const SCEV *ScalarEvolution::createSCEV(Value *V) { // Instcombine's ShrinkDemandedConstant may strip bits out of // constants, obscuring what would otherwise be a low-bits mask. - // Use ComputeMaskedBits to compute what ShrinkDemandedConstant + // Use computeKnownBits to compute what ShrinkDemandedConstant // knew about to reconstruct a low-bits mask value. unsigned LZ = A.countLeadingZeros(); unsigned TZ = A.countTrailingZeros(); unsigned BitWidth = A.getBitWidth(); APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); - ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, DL); + computeKnownBits(U->getOperand(0), KnownZero, KnownOne, DL, + 0, AT, nullptr, DT); APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ - TZ).shl(TZ); @@ -4317,9 +4346,9 @@ ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const { if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute(); assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info"); - const SCEV *BECount = 0; + const SCEV *BECount = nullptr; for (const ExitNotTakenInfo *ENT = &ExitNotTaken; - ENT != 0; ENT = ENT->getNextExit()) { + ENT != nullptr; ENT = ENT->getNextExit()) { assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV"); @@ -4337,7 +4366,7 @@ const SCEV * ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock, ScalarEvolution *SE) const { for (const ExitNotTakenInfo *ENT = &ExitNotTaken; - ENT != 0; ENT = ENT->getNextExit()) { + ENT != nullptr; ENT = ENT->getNextExit()) { if (ENT->ExitingBlock == ExitingBlock) return ENT->ExactNotTaken; @@ -4360,7 +4389,7 @@ bool ScalarEvolution::BackedgeTakenInfo::hasOperand(const SCEV *S, return false; for (const ExitNotTakenInfo *ENT = &ExitNotTaken; - ENT != 0; ENT = ENT->getNextExit()) { + ENT != nullptr; ENT = ENT->getNextExit()) { if (ENT->ExactNotTaken != SE->getCouldNotCompute() && SE->hasOperand(ENT->ExactNotTaken, S)) { @@ -4399,8 +4428,8 @@ ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo( /// clear - Invalidate this result and free the ExitNotTakenInfo array. void ScalarEvolution::BackedgeTakenInfo::clear() { - ExitNotTaken.ExitingBlock = 0; - ExitNotTaken.ExactNotTaken = 0; + ExitNotTaken.ExitingBlock = nullptr; + ExitNotTaken.ExactNotTaken = nullptr; delete[] ExitNotTaken.getNextExit(); } @@ -4411,38 +4440,63 @@ ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) { SmallVector ExitingBlocks; L->getExitingBlocks(ExitingBlocks); - // Examine all exits and pick the most conservative values. - const SCEV *MaxBECount = getCouldNotCompute(); + SmallVector, 4> ExitCounts; bool CouldComputeBECount = true; BasicBlock *Latch = L->getLoopLatch(); // may be NULL. - const SCEV *LatchMaxCount = 0; - SmallVector, 4> ExitCounts; + const SCEV *MustExitMaxBECount = nullptr; + const SCEV *MayExitMaxBECount = nullptr; + + // Compute the ExitLimit for each loop exit. Use this to populate ExitCounts + // and compute maxBECount. for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { - ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]); + BasicBlock *ExitBB = ExitingBlocks[i]; + ExitLimit EL = ComputeExitLimit(L, ExitBB); + + // 1. For each exit that can be computed, add an entry to ExitCounts. + // CouldComputeBECount is true only if all exits can be computed. if (EL.Exact == getCouldNotCompute()) // We couldn't compute an exact value for this exit, so // we won't be able to compute an exact value for the loop. CouldComputeBECount = false; else - ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact)); - - if (MaxBECount == getCouldNotCompute()) - MaxBECount = EL.Max; - else if (EL.Max != getCouldNotCompute()) { - // We cannot take the "min" MaxBECount, because non-unit stride loops may - // skip some loop tests. Taking the max over the exits is sufficiently - // conservative. TODO: We could do better taking into consideration - // non-latch exits that dominate the latch. - if (EL.MustExit && ExitingBlocks[i] == Latch) - LatchMaxCount = EL.Max; - else - MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max); + ExitCounts.push_back(std::make_pair(ExitBB, EL.Exact)); + + // 2. Derive the loop's MaxBECount from each exit's max number of + // non-exiting iterations. Partition the loop exits into two kinds: + // LoopMustExits and LoopMayExits. + // + // A LoopMustExit meets two requirements: + // + // (a) Its ExitLimit.MustExit flag must be set which indicates that the exit + // test condition cannot be skipped (the tested variable has unit stride or + // the test is less-than or greater-than, rather than a strict inequality). + // + // (b) It must dominate the loop latch, hence must be tested on every loop + // iteration. + // + // If any computable LoopMustExit is found, then MaxBECount is the minimum + // EL.Max of computable LoopMustExits. Otherwise, MaxBECount is + // conservatively the maximum EL.Max, where CouldNotCompute is considered + // greater than any computable EL.Max. + if (EL.MustExit && EL.Max != getCouldNotCompute() && Latch && + DT->dominates(ExitBB, Latch)) { + if (!MustExitMaxBECount) + MustExitMaxBECount = EL.Max; + else { + MustExitMaxBECount = + getUMinFromMismatchedTypes(MustExitMaxBECount, EL.Max); + } + } else if (MayExitMaxBECount != getCouldNotCompute()) { + if (!MayExitMaxBECount || EL.Max == getCouldNotCompute()) + MayExitMaxBECount = EL.Max; + else { + MayExitMaxBECount = + getUMaxFromMismatchedTypes(MayExitMaxBECount, EL.Max); + } } } - // Be more precise in the easy case of a loop latch that must exit. - if (LatchMaxCount) { - MaxBECount = getUMinFromMismatchedTypes(MaxBECount, LatchMaxCount); - } + const SCEV *MaxBECount = MustExitMaxBECount ? MustExitMaxBECount : + (MayExitMaxBECount ? MayExitMaxBECount : getCouldNotCompute()); return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount); } @@ -4455,7 +4509,7 @@ ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) { // exit at this block and remember the exit block and whether all other targets // lead to the loop header. bool MustExecuteLoopHeader = true; - BasicBlock *Exit = 0; + BasicBlock *Exit = nullptr; for (succ_iterator SI = succ_begin(ExitingBlock), SE = succ_end(ExitingBlock); SI != SE; ++SI) if (!L->contains(*SI)) { @@ -4801,7 +4855,7 @@ ScalarEvolution::ComputeLoadConstantCompareExitLimit( return getCouldNotCompute(); // Okay, we allow one non-constant index into the GEP instruction. - Value *VarIdx = 0; + Value *VarIdx = nullptr; std::vector Indexes; unsigned VarIdxNum = 0; for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i) @@ -4811,7 +4865,7 @@ ScalarEvolution::ComputeLoadConstantCompareExitLimit( if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's. VarIdx = GEP->getOperand(i); VarIdxNum = i-2; - Indexes.push_back(0); + Indexes.push_back(nullptr); } // Loop-invariant loads may be a byproduct of loop optimization. Skip them. @@ -4842,7 +4896,7 @@ ScalarEvolution::ComputeLoadConstantCompareExitLimit( Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(), Indexes); - if (Result == 0) break; // Cannot compute! + if (!Result) break; // Cannot compute! // Evaluate the condition for this iteration. Result = ConstantExpr::getICmp(predicate, Result, RHS); @@ -4903,14 +4957,14 @@ getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L, // Otherwise, we can evaluate this instruction if all of its operands are // constant or derived from a PHI node themselves. - PHINode *PHI = 0; + PHINode *PHI = nullptr; for (Instruction::op_iterator OpI = UseInst->op_begin(), OpE = UseInst->op_end(); OpI != OpE; ++OpI) { if (isa(*OpI)) continue; Instruction *OpInst = dyn_cast(*OpI); - if (!OpInst || !canConstantEvolve(OpInst, L)) return 0; + if (!OpInst || !canConstantEvolve(OpInst, L)) return nullptr; PHINode *P = dyn_cast(OpInst); if (!P) @@ -4924,8 +4978,10 @@ getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L, P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap); PHIMap[OpInst] = P; } - if (P == 0) return 0; // Not evolving from PHI - if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs. + if (!P) + return nullptr; // Not evolving from PHI + if (PHI && PHI != P) + return nullptr; // Evolving from multiple different PHIs. PHI = P; } // This is a expression evolving from a constant PHI! @@ -4939,7 +4995,7 @@ getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L, /// constraints, return null. static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) { Instruction *I = dyn_cast(V); - if (I == 0 || !canConstantEvolve(I, L)) return 0; + if (!I || !canConstantEvolve(I, L)) return nullptr; if (PHINode *PN = dyn_cast(I)) { return PN; @@ -4961,18 +5017,18 @@ static Constant *EvaluateExpression(Value *V, const Loop *L, // Convenient constant check, but redundant for recursive calls. if (Constant *C = dyn_cast(V)) return C; Instruction *I = dyn_cast(V); - if (!I) return 0; + if (!I) return nullptr; if (Constant *C = Vals.lookup(I)) return C; // An instruction inside the loop depends on a value outside the loop that we // weren't given a mapping for, or a value such as a call inside the loop. - if (!canConstantEvolve(I, L)) return 0; + if (!canConstantEvolve(I, L)) return nullptr; // An unmapped PHI can be due to a branch or another loop inside this loop, // or due to this not being the initial iteration through a loop where we // couldn't compute the evolution of this particular PHI last time. - if (isa(I)) return 0; + if (isa(I)) return nullptr; std::vector Operands(I->getNumOperands()); @@ -4980,12 +5036,12 @@ static Constant *EvaluateExpression(Value *V, const Loop *L, Instruction *Operand = dyn_cast(I->getOperand(i)); if (!Operand) { Operands[i] = dyn_cast(I->getOperand(i)); - if (!Operands[i]) return 0; + if (!Operands[i]) return nullptr; continue; } Constant *C = EvaluateExpression(Operand, L, Vals, DL, TLI); Vals[Operand] = C; - if (!C) return 0; + if (!C) return nullptr; Operands[i] = C; } @@ -5014,7 +5070,7 @@ ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN, return I->second; if (BEs.ugt(MaxBruteForceIterations)) - return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it. + return ConstantEvolutionLoopExitValue[PN] = nullptr; // Not going to evaluate it. Constant *&RetVal = ConstantEvolutionLoopExitValue[PN]; @@ -5026,22 +5082,22 @@ ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN, // entry must be a constant (coming in from outside of the loop), and the // second must be derived from the same PHI. bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1)); - PHINode *PHI = 0; + PHINode *PHI = nullptr; for (BasicBlock::iterator I = Header->begin(); (PHI = dyn_cast(I)); ++I) { Constant *StartCST = dyn_cast(PHI->getIncomingValue(!SecondIsBackedge)); - if (StartCST == 0) continue; + if (!StartCST) continue; CurrentIterVals[PHI] = StartCST; } if (!CurrentIterVals.count(PN)) - return RetVal = 0; + return RetVal = nullptr; Value *BEValue = PN->getIncomingValue(SecondIsBackedge); // Execute the loop symbolically to determine the exit value. if (BEs.getActiveBits() >= 32) - return RetVal = 0; // More than 2^32-1 iterations?? Not doing it! + return RetVal = nullptr; // More than 2^32-1 iterations?? Not doing it! unsigned NumIterations = BEs.getZExtValue(); // must be in range unsigned IterationNum = 0; @@ -5054,8 +5110,8 @@ ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN, DenseMap NextIterVals; Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, DL, TLI); - if (NextPHI == 0) - return 0; // Couldn't evaluate! + if (!NextPHI) + return nullptr; // Couldn't evaluate! NextIterVals[PN] = NextPHI; bool StoppedEvolving = NextPHI == CurrentIterVals[PN]; @@ -5102,7 +5158,7 @@ const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) { PHINode *PN = getConstantEvolvingPHI(Cond, L); - if (PN == 0) return getCouldNotCompute(); + if (!PN) return getCouldNotCompute(); // If the loop is canonicalized, the PHI will have exactly two entries. // That's the only form we support here. @@ -5115,12 +5171,12 @@ const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L, // One entry must be a constant (coming in from outside of the loop), and the // second must be derived from the same PHI. bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1)); - PHINode *PHI = 0; + PHINode *PHI = nullptr; for (BasicBlock::iterator I = Header->begin(); (PHI = dyn_cast(I)); ++I) { Constant *StartCST = dyn_cast(PHI->getIncomingValue(!SecondIsBackedge)); - if (StartCST == 0) continue; + if (!StartCST) continue; CurrentIterVals[PHI] = StartCST; } if (!CurrentIterVals.count(PN)) @@ -5190,7 +5246,7 @@ const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) { if (Values[u].first == L) return Values[u].second ? Values[u].second : V; } - Values.push_back(std::make_pair(L, static_cast(0))); + Values.push_back(std::make_pair(L, static_cast(nullptr))); // Otherwise compute it. const SCEV *C = computeSCEVAtScope(V, L); SmallVector, 2> &Values2 = ValuesAtScopes[V]; @@ -5244,7 +5300,7 @@ static Constant *BuildConstantFromSCEV(const SCEV *V) { } for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) { Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i)); - if (!C2) return 0; + if (!C2) return nullptr; // First pointer! if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) { @@ -5259,7 +5315,7 @@ static Constant *BuildConstantFromSCEV(const SCEV *V) { // Don't bother trying to sum two pointers. We probably can't // statically compute a load that results from it anyway. if (C2->getType()->isPointerTy()) - return 0; + return nullptr; if (PointerType *PTy = dyn_cast(C->getType())) { if (PTy->getElementType()->isStructTy()) @@ -5277,10 +5333,10 @@ static Constant *BuildConstantFromSCEV(const SCEV *V) { const SCEVMulExpr *SM = cast(V); if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) { // Don't bother with pointers at all. - if (C->getType()->isPointerTy()) return 0; + if (C->getType()->isPointerTy()) return nullptr; for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) { Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i)); - if (!C2 || C2->getType()->isPointerTy()) return 0; + if (!C2 || C2->getType()->isPointerTy()) return nullptr; C = ConstantExpr::getMul(C, C2); } return C; @@ -5299,7 +5355,7 @@ static Constant *BuildConstantFromSCEV(const SCEV *V) { case scUMaxExpr: break; // TODO: smax, umax. } - return 0; + return nullptr; } const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) { @@ -5366,7 +5422,7 @@ const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) { // Check to see if getSCEVAtScope actually made an improvement. if (MadeImprovement) { - Constant *C = 0; + Constant *C = nullptr; if (const CmpInst *CI = dyn_cast(I)) C = ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0], Operands[1], DL, @@ -5698,7 +5754,7 @@ ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L, bool IsSubExpr) { // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step. // We have not yet seen any such cases. const SCEVConstant *StepC = dyn_cast(Step); - if (StepC == 0 || StepC->getValue()->equalsInt(0)) + if (!StepC || StepC->getValue()->equalsInt(0)) return getCouldNotCompute(); // For positive steps (counting up until unsigned overflow): @@ -5745,6 +5801,16 @@ ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L, bool IsSubExpr) { getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step); return ExitLimit(Exact, Exact, /*MustExit=*/false); } + + // If Step is a power of two that evenly divides Start we know that the loop + // will always terminate. Start may not be a constant so we just have the + // number of trailing zeros available. This is safe even in presence of + // overflow as the recurrence will overflow to exactly 0. + const APInt &StepV = StepC->getValue()->getValue(); + if (StepV.isPowerOf2() && + GetMinTrailingZeros(getNegativeSCEV(Start)) >= StepV.countTrailingZeros()) + return getUDivExactExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step); + // Then, try to solve the above equation provided that Start is constant. if (const SCEVConstant *StartC = dyn_cast(Start)) return SolveLinEquationWithOverflow(StepC->getValue()->getValue(), @@ -6127,18 +6193,30 @@ bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred, // If LHS or RHS is an addrec, check to see if the condition is true in // every iteration of the loop. - if (const SCEVAddRecExpr *AR = dyn_cast(LHS)) - if (isLoopEntryGuardedByCond( - AR->getLoop(), Pred, AR->getStart(), RHS) && - isLoopBackedgeGuardedByCond( - AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS)) - return true; - if (const SCEVAddRecExpr *AR = dyn_cast(RHS)) - if (isLoopEntryGuardedByCond( - AR->getLoop(), Pred, LHS, AR->getStart()) && - isLoopBackedgeGuardedByCond( - AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this))) - return true; + // If LHS and RHS are both addrec, both conditions must be true in + // every iteration of the loop. + const SCEVAddRecExpr *LAR = dyn_cast(LHS); + const SCEVAddRecExpr *RAR = dyn_cast(RHS); + bool LeftGuarded = false; + bool RightGuarded = false; + if (LAR) { + const Loop *L = LAR->getLoop(); + if (isLoopEntryGuardedByCond(L, Pred, LAR->getStart(), RHS) && + isLoopBackedgeGuardedByCond(L, Pred, LAR->getPostIncExpr(*this), RHS)) { + if (!RAR) return true; + LeftGuarded = true; + } + } + if (RAR) { + const Loop *L = RAR->getLoop(); + if (isLoopEntryGuardedByCond(L, Pred, LHS, RAR->getStart()) && + isLoopBackedgeGuardedByCond(L, Pred, LHS, RAR->getPostIncExpr(*this))) { + if (!LAR) return true; + RightGuarded = true; + } + } + if (LeftGuarded && RightGuarded) + return true; // Otherwise see what can be done with known constant ranges. return isKnownPredicateWithRanges(Pred, LHS, RHS); @@ -6156,7 +6234,6 @@ ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred, default: llvm_unreachable("Unexpected ICmpInst::Predicate value!"); case ICmpInst::ICMP_SGT: - Pred = ICmpInst::ICMP_SLT; std::swap(LHS, RHS); case ICmpInst::ICMP_SLT: { ConstantRange LHSRange = getSignedRange(LHS); @@ -6168,7 +6245,6 @@ ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred, break; } case ICmpInst::ICMP_SGE: - Pred = ICmpInst::ICMP_SLE; std::swap(LHS, RHS); case ICmpInst::ICMP_SLE: { ConstantRange LHSRange = getSignedRange(LHS); @@ -6180,7 +6256,6 @@ ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred, break; } case ICmpInst::ICMP_UGT: - Pred = ICmpInst::ICMP_ULT; std::swap(LHS, RHS); case ICmpInst::ICMP_ULT: { ConstantRange LHSRange = getUnsignedRange(LHS); @@ -6192,7 +6267,6 @@ ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred, break; } case ICmpInst::ICMP_UGE: - Pred = ICmpInst::ICMP_ULE; std::swap(LHS, RHS); case ICmpInst::ICMP_ULE: { ConstantRange LHSRange = getUnsignedRange(LHS); @@ -6239,13 +6313,22 @@ ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L, BranchInst *LoopContinuePredicate = dyn_cast(Latch->getTerminator()); - if (!LoopContinuePredicate || - LoopContinuePredicate->isUnconditional()) - return false; + if (LoopContinuePredicate && LoopContinuePredicate->isConditional() && + isImpliedCond(Pred, LHS, RHS, + LoopContinuePredicate->getCondition(), + LoopContinuePredicate->getSuccessor(0) != L->getHeader())) + return true; + + // Check conditions due to any @llvm.assume intrinsics. + for (auto &CI : AT->assumptions(F)) { + if (!DT->dominates(CI, Latch->getTerminator())) + continue; - return isImpliedCond(Pred, LHS, RHS, - LoopContinuePredicate->getCondition(), - LoopContinuePredicate->getSuccessor(0) != L->getHeader()); + if (isImpliedCond(Pred, LHS, RHS, CI->getArgOperand(0), false)) + return true; + } + + return false; } /// isLoopEntryGuardedByCond - Test whether entry to the loop is protected @@ -6279,6 +6362,15 @@ ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L, return true; } + // Check conditions due to any @llvm.assume intrinsics. + for (auto &CI : AT->assumptions(F)) { + if (!DT->dominates(CI, L->getHeader())) + continue; + + if (isImpliedCond(Pred, LHS, RHS, CI->getArgOperand(0), false)) + return true; + } + return false; } @@ -6598,7 +6690,7 @@ ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS, IsSigned ? APIntOps::smin(getSignedRange(RHS).getSignedMax(), Limit) : APIntOps::umin(getUnsignedRange(RHS).getUnsignedMax(), Limit); - const SCEV *MaxBECount = getCouldNotCompute(); + const SCEV *MaxBECount; if (isa(BECount)) MaxBECount = BECount; else @@ -6809,6 +6901,105 @@ const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range, return SE.getCouldNotCompute(); } +namespace { +struct FindUndefs { + bool Found; + FindUndefs() : Found(false) {} + + bool follow(const SCEV *S) { + if (const SCEVUnknown *C = dyn_cast(S)) { + if (isa(C->getValue())) + Found = true; + } else if (const SCEVConstant *C = dyn_cast(S)) { + if (isa(C->getValue())) + Found = true; + } + + // Keep looking if we haven't found it yet. + return !Found; + } + bool isDone() const { + // Stop recursion if we have found an undef. + return Found; + } +}; +} + +// Return true when S contains at least an undef value. +static inline bool +containsUndefs(const SCEV *S) { + FindUndefs F; + SCEVTraversal ST(F); + ST.visitAll(S); + + return F.Found; +} + +namespace { +// Collect all steps of SCEV expressions. +struct SCEVCollectStrides { + ScalarEvolution &SE; + SmallVectorImpl &Strides; + + SCEVCollectStrides(ScalarEvolution &SE, SmallVectorImpl &S) + : SE(SE), Strides(S) {} + + bool follow(const SCEV *S) { + if (const SCEVAddRecExpr *AR = dyn_cast(S)) + Strides.push_back(AR->getStepRecurrence(SE)); + return true; + } + bool isDone() const { return false; } +}; + +// Collect all SCEVUnknown and SCEVMulExpr expressions. +struct SCEVCollectTerms { + SmallVectorImpl &Terms; + + SCEVCollectTerms(SmallVectorImpl &T) + : Terms(T) {} + + bool follow(const SCEV *S) { + if (isa(S) || isa(S)) { + if (!containsUndefs(S)) + Terms.push_back(S); + + // Stop recursion: once we collected a term, do not walk its operands. + return false; + } + + // Keep looking. + return true; + } + bool isDone() const { return false; } +}; +} + +/// Find parametric terms in this SCEVAddRecExpr. +void SCEVAddRecExpr::collectParametricTerms( + ScalarEvolution &SE, SmallVectorImpl &Terms) const { + SmallVector Strides; + SCEVCollectStrides StrideCollector(SE, Strides); + visitAll(this, StrideCollector); + + DEBUG({ + dbgs() << "Strides:\n"; + for (const SCEV *S : Strides) + dbgs() << *S << "\n"; + }); + + for (const SCEV *S : Strides) { + SCEVCollectTerms TermCollector(Terms); + visitAll(S, TermCollector); + } + + DEBUG({ + dbgs() << "Terms:\n"; + for (const SCEV *T : Terms) + dbgs() << *T << "\n"; + }); +} + static const APInt srem(const SCEVConstant *C1, const SCEVConstant *C2) { APInt A = C1->getValue()->getValue(); APInt B = C2->getValue()->getValue(); @@ -6838,352 +7029,488 @@ static const APInt sdiv(const SCEVConstant *C1, const SCEVConstant *C2) { } namespace { -struct SCEVGCD : public SCEVVisitor { -public: - // Pattern match Step into Start. When Step is a multiply expression, find - // the largest subexpression of Step that appears in Start. When Start is an - // add expression, try to match Step in the subexpressions of Start, non - // matching subexpressions are returned under Remainder. - static const SCEV *findGCD(ScalarEvolution &SE, const SCEV *Start, - const SCEV *Step, const SCEV **Remainder) { - assert(Remainder && "Remainder should not be NULL"); - SCEVGCD R(SE, Step, SE.getConstant(Step->getType(), 0)); - const SCEV *Res = R.visit(Start); - *Remainder = R.Remainder; - return Res; - } +struct FindSCEVSize { + int Size; + FindSCEVSize() : Size(0) {} - SCEVGCD(ScalarEvolution &S, const SCEV *G, const SCEV *R) - : SE(S), GCD(G), Remainder(R) { - Zero = SE.getConstant(GCD->getType(), 0); - One = SE.getConstant(GCD->getType(), 1); + bool follow(const SCEV *S) { + ++Size; + // Keep looking at all operands of S. + return true; + } + bool isDone() const { + return false; } +}; +} - const SCEV *visitConstant(const SCEVConstant *Constant) { - if (GCD == Constant || Constant == Zero) - return GCD; +// Returns the size of the SCEV S. +static inline int sizeOfSCEV(const SCEV *S) { + FindSCEVSize F; + SCEVTraversal ST(F); + ST.visitAll(S); + return F.Size; +} - if (const SCEVConstant *CGCD = dyn_cast(GCD)) { - const SCEV *Res = SE.getConstant(gcd(Constant, CGCD)); - if (Res != One) - return Res; +namespace { - Remainder = SE.getConstant(srem(Constant, CGCD)); - Constant = cast(SE.getMinusSCEV(Constant, Remainder)); - Res = SE.getConstant(gcd(Constant, CGCD)); - return Res; +struct SCEVDivision : public SCEVVisitor { +public: + // Computes the Quotient and Remainder of the division of Numerator by + // Denominator. + static void divide(ScalarEvolution &SE, const SCEV *Numerator, + const SCEV *Denominator, const SCEV **Quotient, + const SCEV **Remainder) { + assert(Numerator && Denominator && "Uninitialized SCEV"); + + SCEVDivision D(SE, Numerator, Denominator); + + // Check for the trivial case here to avoid having to check for it in the + // rest of the code. + if (Numerator == Denominator) { + *Quotient = D.One; + *Remainder = D.Zero; + return; } - // When GCD is not a constant, it could be that the GCD is an Add, Mul, - // AddRec, etc., in which case we want to find out how many times the - // Constant divides the GCD: we then return that as the new GCD. - const SCEV *Rem = Zero; - const SCEV *Res = findGCD(SE, GCD, Constant, &Rem); + if (Numerator->isZero()) { + *Quotient = D.Zero; + *Remainder = D.Zero; + return; + } - if (Res == One || Rem != Zero) { - Remainder = Constant; - return One; + // Split the Denominator when it is a product. + if (const SCEVMulExpr *T = dyn_cast(Denominator)) { + const SCEV *Q, *R; + *Quotient = Numerator; + for (const SCEV *Op : T->operands()) { + divide(SE, *Quotient, Op, &Q, &R); + *Quotient = Q; + + // Bail out when the Numerator is not divisible by one of the terms of + // the Denominator. + if (!R->isZero()) { + *Quotient = D.Zero; + *Remainder = Numerator; + return; + } + } + *Remainder = D.Zero; + return; } - assert(isa(Res) && "Res should be a constant"); - Remainder = SE.getConstant(srem(Constant, cast(Res))); - return Res; + D.visit(Numerator); + *Quotient = D.Quotient; + *Remainder = D.Remainder; + } + + SCEVDivision(ScalarEvolution &S, const SCEV *Numerator, const SCEV *Denominator) + : SE(S), Denominator(Denominator) { + Zero = SE.getConstant(Denominator->getType(), 0); + One = SE.getConstant(Denominator->getType(), 1); + + // By default, we don't know how to divide Expr by Denominator. + // Providing the default here simplifies the rest of the code. + Quotient = Zero; + Remainder = Numerator; + } + + // Except in the trivial case described above, we do not know how to divide + // Expr by Denominator for the following functions with empty implementation. + void visitTruncateExpr(const SCEVTruncateExpr *Numerator) {} + void visitZeroExtendExpr(const SCEVZeroExtendExpr *Numerator) {} + void visitSignExtendExpr(const SCEVSignExtendExpr *Numerator) {} + void visitUDivExpr(const SCEVUDivExpr *Numerator) {} + void visitSMaxExpr(const SCEVSMaxExpr *Numerator) {} + void visitUMaxExpr(const SCEVUMaxExpr *Numerator) {} + void visitUnknown(const SCEVUnknown *Numerator) {} + void visitCouldNotCompute(const SCEVCouldNotCompute *Numerator) {} + + void visitConstant(const SCEVConstant *Numerator) { + if (const SCEVConstant *D = dyn_cast(Denominator)) { + Quotient = SE.getConstant(sdiv(Numerator, D)); + Remainder = SE.getConstant(srem(Numerator, D)); + return; + } } - const SCEV *visitTruncateExpr(const SCEVTruncateExpr *Expr) { - if (GCD != Expr) - Remainder = Expr; - return GCD; + void visitAddRecExpr(const SCEVAddRecExpr *Numerator) { + const SCEV *StartQ, *StartR, *StepQ, *StepR; + assert(Numerator->isAffine() && "Numerator should be affine"); + divide(SE, Numerator->getStart(), Denominator, &StartQ, &StartR); + divide(SE, Numerator->getStepRecurrence(SE), Denominator, &StepQ, &StepR); + Quotient = SE.getAddRecExpr(StartQ, StepQ, Numerator->getLoop(), + Numerator->getNoWrapFlags()); + Remainder = SE.getAddRecExpr(StartR, StepR, Numerator->getLoop(), + Numerator->getNoWrapFlags()); } - const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) { - if (GCD != Expr) - Remainder = Expr; - return GCD; - } + void visitAddExpr(const SCEVAddExpr *Numerator) { + SmallVector Qs, Rs; + Type *Ty = Denominator->getType(); - const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) { - if (GCD != Expr) - Remainder = Expr; - return GCD; - } + for (const SCEV *Op : Numerator->operands()) { + const SCEV *Q, *R; + divide(SE, Op, Denominator, &Q, &R); - const SCEV *visitAddExpr(const SCEVAddExpr *Expr) { - if (GCD == Expr) - return GCD; + // Bail out if types do not match. + if (Ty != Q->getType() || Ty != R->getType()) { + Quotient = Zero; + Remainder = Numerator; + return; + } - for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) { - const SCEV *Rem = Zero; - const SCEV *Res = findGCD(SE, Expr->getOperand(e - 1 - i), GCD, &Rem); + Qs.push_back(Q); + Rs.push_back(R); + } - // FIXME: There may be ambiguous situations: for instance, - // GCD(-4 + (3 * %m), 2 * %m) where 2 divides -4 and %m divides (3 * %m). - // The order in which the AddExpr is traversed computes a different GCD - // and Remainder. - if (Res != One) - GCD = Res; - if (Rem != Zero) - Remainder = SE.getAddExpr(Remainder, Rem); + if (Qs.size() == 1) { + Quotient = Qs[0]; + Remainder = Rs[0]; + return; } - return GCD; + Quotient = SE.getAddExpr(Qs); + Remainder = SE.getAddExpr(Rs); } - const SCEV *visitMulExpr(const SCEVMulExpr *Expr) { - if (GCD == Expr) - return GCD; + void visitMulExpr(const SCEVMulExpr *Numerator) { + SmallVector Qs; + Type *Ty = Denominator->getType(); - for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) { - if (Expr->getOperand(i) == GCD) - return GCD; - } + bool FoundDenominatorTerm = false; + for (const SCEV *Op : Numerator->operands()) { + // Bail out if types do not match. + if (Ty != Op->getType()) { + Quotient = Zero; + Remainder = Numerator; + return; + } + + if (FoundDenominatorTerm) { + Qs.push_back(Op); + continue; + } - // If we have not returned yet, it means that GCD is not part of Expr. - const SCEV *PartialGCD = One; - for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) { - const SCEV *Rem = Zero; - const SCEV *Res = findGCD(SE, Expr->getOperand(i), GCD, &Rem); - if (Rem != Zero) - // GCD does not divide Expr->getOperand(i). + // Check whether Denominator divides one of the product operands. + const SCEV *Q, *R; + divide(SE, Op, Denominator, &Q, &R); + if (!R->isZero()) { + Qs.push_back(Op); continue; + } - if (Res == GCD) - return GCD; - PartialGCD = SE.getMulExpr(PartialGCD, Res); - if (PartialGCD == GCD) - return GCD; - } - - if (PartialGCD != One) - return PartialGCD; - - Remainder = Expr; - const SCEVMulExpr *Mul = dyn_cast(GCD); - if (!Mul) - return PartialGCD; - - // When the GCD is a multiply expression, try to decompose it: - // this occurs when Step does not divide the Start expression - // as in: {(-4 + (3 * %m)),+,(2 * %m)} - for (int i = 0, e = Mul->getNumOperands(); i < e; ++i) { - const SCEV *Rem = Zero; - const SCEV *Res = findGCD(SE, Expr, Mul->getOperand(i), &Rem); - if (Rem == Zero) { - Remainder = Rem; - return Res; + // Bail out if types do not match. + if (Ty != Q->getType()) { + Quotient = Zero; + Remainder = Numerator; + return; } + + FoundDenominatorTerm = true; + Qs.push_back(Q); } - return PartialGCD; - } + if (FoundDenominatorTerm) { + Remainder = Zero; + if (Qs.size() == 1) + Quotient = Qs[0]; + else + Quotient = SE.getMulExpr(Qs); + return; + } - const SCEV *visitUDivExpr(const SCEVUDivExpr *Expr) { - if (GCD != Expr) - Remainder = Expr; - return GCD; - } + if (!isa(Denominator)) { + Quotient = Zero; + Remainder = Numerator; + return; + } - const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) { - if (GCD == Expr) - return GCD; + // The Remainder is obtained by replacing Denominator by 0 in Numerator. + ValueToValueMap RewriteMap; + RewriteMap[cast(Denominator)->getValue()] = + cast(Zero)->getValue(); + Remainder = SCEVParameterRewriter::rewrite(Numerator, SE, RewriteMap, true); + + if (Remainder->isZero()) { + // The Quotient is obtained by replacing Denominator by 1 in Numerator. + RewriteMap[cast(Denominator)->getValue()] = + cast(One)->getValue(); + Quotient = + SCEVParameterRewriter::rewrite(Numerator, SE, RewriteMap, true); + return; + } - if (!Expr->isAffine()) { - Remainder = Expr; - return GCD; + // Quotient is (Numerator - Remainder) divided by Denominator. + const SCEV *Q, *R; + const SCEV *Diff = SE.getMinusSCEV(Numerator, Remainder); + if (sizeOfSCEV(Diff) > sizeOfSCEV(Numerator)) { + // This SCEV does not seem to simplify: fail the division here. + Quotient = Zero; + Remainder = Numerator; + return; } + divide(SE, Diff, Denominator, &Q, &R); + assert(R == Zero && + "(Numerator - Remainder) should evenly divide Denominator"); + Quotient = Q; + } + +private: + ScalarEvolution &SE; + const SCEV *Denominator, *Quotient, *Remainder, *Zero, *One; +}; +} + +static bool findArrayDimensionsRec(ScalarEvolution &SE, + SmallVectorImpl &Terms, + SmallVectorImpl &Sizes) { + int Last = Terms.size() - 1; + const SCEV *Step = Terms[Last]; - const SCEV *Rem = Zero; - const SCEV *Res = findGCD(SE, Expr->getOperand(0), GCD, &Rem); - if (Rem != Zero) - Remainder = SE.getAddExpr(Remainder, Rem); + // End of recursion. + if (Last == 0) { + if (const SCEVMulExpr *M = dyn_cast(Step)) { + SmallVector Qs; + for (const SCEV *Op : M->operands()) + if (!isa(Op)) + Qs.push_back(Op); - Rem = Zero; - Res = findGCD(SE, Expr->getOperand(1), Res, &Rem); - if (Rem != Zero) { - Remainder = Expr; - return GCD; + Step = SE.getMulExpr(Qs); } - return Res; + Sizes.push_back(Step); + return true; } - const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) { - if (GCD != Expr) - Remainder = Expr; - return GCD; - } + for (const SCEV *&Term : Terms) { + // Normalize the terms before the next call to findArrayDimensionsRec. + const SCEV *Q, *R; + SCEVDivision::divide(SE, Term, Step, &Q, &R); - const SCEV *visitUMaxExpr(const SCEVUMaxExpr *Expr) { - if (GCD != Expr) - Remainder = Expr; - return GCD; - } + // Bail out when GCD does not evenly divide one of the terms. + if (!R->isZero()) + return false; - const SCEV *visitUnknown(const SCEVUnknown *Expr) { - if (GCD != Expr) - Remainder = Expr; - return GCD; + Term = Q; } - const SCEV *visitCouldNotCompute(const SCEVCouldNotCompute *Expr) { - return One; - } + // Remove all SCEVConstants. + Terms.erase(std::remove_if(Terms.begin(), Terms.end(), [](const SCEV *E) { + return isa(E); + }), + Terms.end()); -private: - ScalarEvolution &SE; - const SCEV *GCD, *Remainder, *Zero, *One; -}; + if (Terms.size() > 0) + if (!findArrayDimensionsRec(SE, Terms, Sizes)) + return false; -struct SCEVDivision : public SCEVVisitor { -public: - // Remove from Start all multiples of Step. - static const SCEV *divide(ScalarEvolution &SE, const SCEV *Start, - const SCEV *Step) { - SCEVDivision D(SE, Step); - const SCEV *Rem = D.Zero; - (void)Rem; - // The division is guaranteed to succeed: Step should divide Start with no - // remainder. - assert(Step == SCEVGCD::findGCD(SE, Start, Step, &Rem) && Rem == D.Zero && - "Step should divide Start with no remainder."); - return D.visit(Start); - } + Sizes.push_back(Step); + return true; +} + +namespace { +struct FindParameter { + bool FoundParameter; + FindParameter() : FoundParameter(false) {} - SCEVDivision(ScalarEvolution &S, const SCEV *G) : SE(S), GCD(G) { - Zero = SE.getConstant(GCD->getType(), 0); - One = SE.getConstant(GCD->getType(), 1); + bool follow(const SCEV *S) { + if (isa(S)) { + FoundParameter = true; + // Stop recursion: we found a parameter. + return false; + } + // Keep looking. + return true; } + bool isDone() const { + // Stop recursion if we have found a parameter. + return FoundParameter; + } +}; +} - const SCEV *visitConstant(const SCEVConstant *Constant) { - if (GCD == Constant) - return One; +// Returns true when S contains at least a SCEVUnknown parameter. +static inline bool +containsParameters(const SCEV *S) { + FindParameter F; + SCEVTraversal ST(F); + ST.visitAll(S); - if (const SCEVConstant *CGCD = dyn_cast(GCD)) - return SE.getConstant(sdiv(Constant, CGCD)); - return Constant; - } + return F.FoundParameter; +} - const SCEV *visitTruncateExpr(const SCEVTruncateExpr *Expr) { - if (GCD == Expr) - return One; - return Expr; - } +// Returns true when one of the SCEVs of Terms contains a SCEVUnknown parameter. +static inline bool +containsParameters(SmallVectorImpl &Terms) { + for (const SCEV *T : Terms) + if (containsParameters(T)) + return true; + return false; +} - const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) { - if (GCD == Expr) - return One; - return Expr; - } +// Return the number of product terms in S. +static inline int numberOfTerms(const SCEV *S) { + if (const SCEVMulExpr *Expr = dyn_cast(S)) + return Expr->getNumOperands(); + return 1; +} - const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) { - if (GCD == Expr) - return One; - return Expr; - } +static const SCEV *removeConstantFactors(ScalarEvolution &SE, const SCEV *T) { + if (isa(T)) + return nullptr; - const SCEV *visitAddExpr(const SCEVAddExpr *Expr) { - if (GCD == Expr) - return One; + if (isa(T)) + return T; - SmallVector Operands; - for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) - Operands.push_back(divide(SE, Expr->getOperand(i), GCD)); + if (const SCEVMulExpr *M = dyn_cast(T)) { + SmallVector Factors; + for (const SCEV *Op : M->operands()) + if (!isa(Op)) + Factors.push_back(Op); - if (Operands.size() == 1) - return Operands[0]; - return SE.getAddExpr(Operands); + return SE.getMulExpr(Factors); } - const SCEV *visitMulExpr(const SCEVMulExpr *Expr) { - if (GCD == Expr) - return One; + return T; +} - bool FoundGCDTerm = false; - for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) - if (Expr->getOperand(i) == GCD) - FoundGCDTerm = true; +/// Return the size of an element read or written by Inst. +const SCEV *ScalarEvolution::getElementSize(Instruction *Inst) { + Type *Ty; + if (StoreInst *Store = dyn_cast(Inst)) + Ty = Store->getValueOperand()->getType(); + else if (LoadInst *Load = dyn_cast(Inst)) + Ty = Load->getType(); + else + return nullptr; - SmallVector Operands; - if (FoundGCDTerm) { - FoundGCDTerm = false; - for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) { - if (FoundGCDTerm) - Operands.push_back(Expr->getOperand(i)); - else if (Expr->getOperand(i) == GCD) - FoundGCDTerm = true; - else - Operands.push_back(Expr->getOperand(i)); - } - } else { - FoundGCDTerm = false; - const SCEV *PartialGCD = One; - for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) { - if (PartialGCD == GCD) { - Operands.push_back(Expr->getOperand(i)); - continue; - } + Type *ETy = getEffectiveSCEVType(PointerType::getUnqual(Ty)); + return getSizeOfExpr(ETy, Ty); +} - const SCEV *Rem = Zero; - const SCEV *Res = SCEVGCD::findGCD(SE, Expr->getOperand(i), GCD, &Rem); - if (Rem == Zero) { - PartialGCD = SE.getMulExpr(PartialGCD, Res); - Operands.push_back(divide(SE, Expr->getOperand(i), GCD)); - } else { - Operands.push_back(Expr->getOperand(i)); - } - } - } +/// Second step of delinearization: compute the array dimensions Sizes from the +/// set of Terms extracted from the memory access function of this SCEVAddRec. +void ScalarEvolution::findArrayDimensions(SmallVectorImpl &Terms, + SmallVectorImpl &Sizes, + const SCEV *ElementSize) const { - if (Operands.size() == 1) - return Operands[0]; - return SE.getMulExpr(Operands); - } + if (Terms.size() < 1 || !ElementSize) + return; - const SCEV *visitUDivExpr(const SCEVUDivExpr *Expr) { - if (GCD == Expr) - return One; - return Expr; - } + // Early return when Terms do not contain parameters: we do not delinearize + // non parametric SCEVs. + if (!containsParameters(Terms)) + return; - const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) { - if (GCD == Expr) - return One; + DEBUG({ + dbgs() << "Terms:\n"; + for (const SCEV *T : Terms) + dbgs() << *T << "\n"; + }); - assert(Expr->isAffine() && "Expr should be affine"); + // Remove duplicates. + std::sort(Terms.begin(), Terms.end()); + Terms.erase(std::unique(Terms.begin(), Terms.end()), Terms.end()); - const SCEV *Start = divide(SE, Expr->getStart(), GCD); - const SCEV *Step = divide(SE, Expr->getStepRecurrence(SE), GCD); + // Put larger terms first. + std::sort(Terms.begin(), Terms.end(), [](const SCEV *LHS, const SCEV *RHS) { + return numberOfTerms(LHS) > numberOfTerms(RHS); + }); - return SE.getAddRecExpr(Start, Step, Expr->getLoop(), - Expr->getNoWrapFlags()); - } + ScalarEvolution &SE = *const_cast(this); - const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) { - if (GCD == Expr) - return One; - return Expr; + // Divide all terms by the element size. + for (const SCEV *&Term : Terms) { + const SCEV *Q, *R; + SCEVDivision::divide(SE, Term, ElementSize, &Q, &R); + Term = Q; } - const SCEV *visitUMaxExpr(const SCEVUMaxExpr *Expr) { - if (GCD == Expr) - return One; - return Expr; - } + SmallVector NewTerms; + + // Remove constant factors. + for (const SCEV *T : Terms) + if (const SCEV *NewT = removeConstantFactors(SE, T)) + NewTerms.push_back(NewT); + + DEBUG({ + dbgs() << "Terms after sorting:\n"; + for (const SCEV *T : NewTerms) + dbgs() << *T << "\n"; + }); - const SCEV *visitUnknown(const SCEVUnknown *Expr) { - if (GCD == Expr) - return One; - return Expr; + if (NewTerms.empty() || + !findArrayDimensionsRec(SE, NewTerms, Sizes)) { + Sizes.clear(); + return; } - const SCEV *visitCouldNotCompute(const SCEVCouldNotCompute *Expr) { - return Expr; + // The last element to be pushed into Sizes is the size of an element. + Sizes.push_back(ElementSize); + + DEBUG({ + dbgs() << "Sizes:\n"; + for (const SCEV *S : Sizes) + dbgs() << *S << "\n"; + }); +} + +/// Third step of delinearization: compute the access functions for the +/// Subscripts based on the dimensions in Sizes. +void SCEVAddRecExpr::computeAccessFunctions( + ScalarEvolution &SE, SmallVectorImpl &Subscripts, + SmallVectorImpl &Sizes) const { + + // Early exit in case this SCEV is not an affine multivariate function. + if (Sizes.empty() || !this->isAffine()) + return; + + const SCEV *Res = this; + int Last = Sizes.size() - 1; + for (int i = Last; i >= 0; i--) { + const SCEV *Q, *R; + SCEVDivision::divide(SE, Res, Sizes[i], &Q, &R); + + DEBUG({ + dbgs() << "Res: " << *Res << "\n"; + dbgs() << "Sizes[i]: " << *Sizes[i] << "\n"; + dbgs() << "Res divided by Sizes[i]:\n"; + dbgs() << "Quotient: " << *Q << "\n"; + dbgs() << "Remainder: " << *R << "\n"; + }); + + Res = Q; + + // Do not record the last subscript corresponding to the size of elements in + // the array. + if (i == Last) { + + // Bail out if the remainder is too complex. + if (isa(R)) { + Subscripts.clear(); + Sizes.clear(); + return; + } + + continue; + } + + // Record the access function for the current subscript. + Subscripts.push_back(R); } -private: - ScalarEvolution &SE; - const SCEV *GCD, *Zero, *One; -}; + // Also push in last position the remainder of the last division: it will be + // the access function of the innermost dimension. + Subscripts.push_back(Res); + + std::reverse(Subscripts.begin(), Subscripts.end()); + + DEBUG({ + dbgs() << "Subscripts:\n"; + for (const SCEV *S : Subscripts) + dbgs() << *S << "\n"; + }); } /// Splits the SCEV into two vectors of SCEVs representing the subscripts and @@ -7235,84 +7562,40 @@ private: /// asking for the SCEV of the memory access with respect to all enclosing /// loops, calling SCEV->delinearize on that and printing the results. -const SCEV * -SCEVAddRecExpr::delinearize(ScalarEvolution &SE, - SmallVectorImpl &Subscripts, - SmallVectorImpl &Sizes) const { - // Early exit in case this SCEV is not an affine multivariate function. - if (!this->isAffine()) - return this; - - const SCEV *Start = this->getStart(); - const SCEV *Step = this->getStepRecurrence(SE); - - // Build the SCEV representation of the canonical induction variable in the - // loop of this SCEV. - const SCEV *Zero = SE.getConstant(this->getType(), 0); - const SCEV *One = SE.getConstant(this->getType(), 1); - const SCEV *IV = - SE.getAddRecExpr(Zero, One, this->getLoop(), this->getNoWrapFlags()); - - DEBUG(dbgs() << "(delinearize: " << *this << "\n"); - - // When the stride of this SCEV is 1, do not compute the GCD: the size of this - // subscript is 1, and this same SCEV for the access function. - const SCEV *Remainder = Zero; - const SCEV *GCD = One; - - // Find the GCD and Remainder of the Start and Step coefficients of this SCEV. - if (Step != One && !Step->isAllOnesValue()) - GCD = SCEVGCD::findGCD(SE, Start, Step, &Remainder); - - DEBUG(dbgs() << "GCD: " << *GCD << "\n"); - DEBUG(dbgs() << "Remainder: " << *Remainder << "\n"); - - const SCEV *Quotient = Start; - if (GCD != One && !GCD->isAllOnesValue()) - // As findGCD computed Remainder, GCD divides "Start - Remainder." The - // Quotient is then this SCEV without Remainder, scaled down by the GCD. The - // Quotient is what will be used in the next subscript delinearization. - Quotient = SCEVDivision::divide(SE, SE.getMinusSCEV(Start, Remainder), GCD); - - DEBUG(dbgs() << "Quotient: " << *Quotient << "\n"); - - const SCEV *Rem = Quotient; - if (const SCEVAddRecExpr *AR = dyn_cast(Quotient)) - // Recursively call delinearize on the Quotient until there are no more - // multiples that can be recognized. - Rem = AR->delinearize(SE, Subscripts, Sizes); - - // Scale up the canonical induction variable IV by whatever remains from the - // Step after division by the GCD: the GCD is the size of all the sub-array. - if (Step != One && !Step->isAllOnesValue() && GCD != One && - !GCD->isAllOnesValue() && Step != GCD) { - Step = SCEVDivision::divide(SE, Step, GCD); - IV = SE.getMulExpr(IV, Step); - } - // The access function in the current subscript is computed as the canonical - // induction variable IV (potentially scaled up by the step) and offset by - // Rem, the offset of delinearization in the sub-array. - const SCEV *Index = SE.getAddExpr(IV, Rem); - - // Record the access function and the size of the current subscript. - Subscripts.push_back(Index); - Sizes.push_back(GCD); +void SCEVAddRecExpr::delinearize(ScalarEvolution &SE, + SmallVectorImpl &Subscripts, + SmallVectorImpl &Sizes, + const SCEV *ElementSize) const { + // First step: collect parametric terms. + SmallVector Terms; + collectParametricTerms(SE, Terms); -#ifndef NDEBUG - int Size = Sizes.size(); - DEBUG(dbgs() << "succeeded to delinearize " << *this << "\n"); - DEBUG(dbgs() << "ArrayDecl[UnknownSize]"); - for (int i = 0; i < Size - 1; i++) - DEBUG(dbgs() << "[" << *Sizes[i] << "]"); - DEBUG(dbgs() << " with elements of " << *Sizes[Size - 1] << " bytes.\n"); - - DEBUG(dbgs() << "ArrayRef"); - for (int i = 0; i < Size; i++) - DEBUG(dbgs() << "[" << *Subscripts[i] << "]"); - DEBUG(dbgs() << "\n)\n"); -#endif + if (Terms.empty()) + return; + + // Second step: find subscript sizes. + SE.findArrayDimensions(Terms, Sizes, ElementSize); + + if (Sizes.empty()) + return; + + // Third step: compute the access functions for each subscript. + computeAccessFunctions(SE, Subscripts, Sizes); + + if (Subscripts.empty()) + return; + + DEBUG({ + dbgs() << "succeeded to delinearize " << *this << "\n"; + dbgs() << "ArrayDecl[UnknownSize]"; + for (const SCEV *S : Sizes) + dbgs() << "[" << *S << "]"; - return Remainder; + dbgs() << "\nArrayRef"; + for (const SCEV *S : Subscripts) + dbgs() << "[" << *S << "]"; + dbgs() << "\n"; + }); } //===----------------------------------------------------------------------===// @@ -7334,11 +7617,8 @@ void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) { // so that future queries will recompute the expressions using the new // value. Value *Old = getValPtr(); - SmallVector Worklist; + SmallVector Worklist(Old->user_begin(), Old->user_end()); SmallPtrSet Visited; - for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end(); - UI != UE; ++UI) - Worklist.push_back(*UI); while (!Worklist.empty()) { User *U = Worklist.pop_back_val(); // Deleting the Old value will cause this to dangle. Postpone @@ -7350,9 +7630,7 @@ void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) { if (PHINode *PN = dyn_cast(U)) SE->ConstantEvolutionLoopExitValue.erase(PN); SE->ValueExprMap.erase(U); - for (Value::use_iterator UI = U->use_begin(), UE = U->use_end(); - UI != UE; ++UI) - Worklist.push_back(*UI); + Worklist.insert(Worklist.end(), U->user_begin(), U->user_end()); } // Delete the Old value. if (PHINode *PN = dyn_cast(Old)) @@ -7369,15 +7647,17 @@ ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se) //===----------------------------------------------------------------------===// ScalarEvolution::ScalarEvolution() - : FunctionPass(ID), ValuesAtScopes(64), LoopDispositions(64), BlockDispositions(64), FirstUnknown(0) { + : FunctionPass(ID), ValuesAtScopes(64), LoopDispositions(64), + BlockDispositions(64), FirstUnknown(nullptr) { initializeScalarEvolutionPass(*PassRegistry::getPassRegistry()); } bool ScalarEvolution::runOnFunction(Function &F) { this->F = &F; + AT = &getAnalysis(); LI = &getAnalysis(); DataLayoutPass *DLP = getAnalysisIfAvailable(); - DL = DLP ? &DLP->getDataLayout() : 0; + DL = DLP ? &DLP->getDataLayout() : nullptr; TLI = &getAnalysis(); DT = &getAnalysis().getDomTree(); return false; @@ -7388,7 +7668,7 @@ void ScalarEvolution::releaseMemory() { // destructors, so that they release their references to their values. for (SCEVUnknown *U = FirstUnknown; U; U = U->Next) U->~SCEVUnknown(); - FirstUnknown = 0; + FirstUnknown = nullptr; ValueExprMap.clear(); @@ -7415,6 +7695,7 @@ void ScalarEvolution::releaseMemory() { void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesAll(); + AU.addRequired(); AU.addRequiredTransitive(); AU.addRequiredTransitive(); AU.addRequired();