//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Nate Begeman and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
#include "llvm/Transforms/Scalar.h"
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/Type.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Support/CFG.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Target/TargetData.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/Compiler.h"
#include <set>
using namespace llvm;
+STATISTIC(NumReduced , "Number of GEPs strength reduced");
+STATISTIC(NumInserted, "Number of PHIs inserted");
+STATISTIC(NumVariable, "Number of PHIs with variable strides");
+STATISTIC(NumEliminated , "Number of strides eliminated");
+
namespace {
- Statistic NumReduced ("loop-reduce", "Number of GEPs strength reduced");
- Statistic NumInserted("loop-reduce", "Number of PHIs inserted");
- Statistic NumVariable("loop-reduce","Number of PHIs with variable strides");
+
+ struct BasedUser;
/// IVStrideUse - Keep track of one use of a strided induction variable, where
/// the stride is stored externally. The Offset member keeps track of the
- /// offset from the IV, User is the actual user of the operand, and 'Operand'
- /// is the operand # of the User that is the use.
- struct IVStrideUse {
+ /// offset from the IV, User is the actual user of the operand, and
+ /// 'OperandValToReplace' is the operand of the User that is the use.
+ struct VISIBILITY_HIDDEN IVStrideUse {
SCEVHandle Offset;
Instruction *User;
Value *OperandValToReplace;
/// have an operand that is based on the trip count multiplied by some stride.
/// The stride for all of these users is common and kept external to this
/// structure.
- struct IVUsersOfOneStride {
+ struct VISIBILITY_HIDDEN IVUsersOfOneStride {
/// Users - Keep track of all of the users of this stride as well as the
/// initial value and the operand that uses the IV.
std::vector<IVStrideUse> Users;
/// IVInfo - This structure keeps track of one IV expression inserted during
/// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
/// well as the PHI node and increment value created for rewrite.
- struct IVExpr {
+ struct VISIBILITY_HIDDEN IVExpr {
SCEVHandle Stride;
SCEVHandle Base;
PHINode *PHI;
Value *IncV;
- IVExpr()
- : Stride(SCEVUnknown::getIntegerSCEV(0, Type::UIntTy)),
- Base (SCEVUnknown::getIntegerSCEV(0, Type::UIntTy)) {}
IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
Value *incv)
: Stride(stride), Base(base), PHI(phi), IncV(incv) {}
/// IVsOfOneStride - This structure keeps track of all IV expression inserted
/// during StrengthReduceStridedIVUsers for a particular stride of the IV.
- struct IVsOfOneStride {
+ struct VISIBILITY_HIDDEN IVsOfOneStride {
std::vector<IVExpr> IVs;
void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
}
};
- class VISIBILITY_HIDDEN LoopStrengthReduce : public FunctionPass {
+ class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
LoopInfo *LI;
- ETForest *EF;
+ DominatorTree *DT;
ScalarEvolution *SE;
const TargetData *TD;
const Type *UIntPtrTy;
/// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
/// We use this to iterate over the IVUsesByStride collection without being
/// dependent on random ordering of pointers in the process.
- std::vector<SCEVHandle> StrideOrder;
+ SmallVector<SCEVHandle, 16> StrideOrder;
/// CastedValues - As we need to cast values to uintptr_t, this keeps track
/// of the casted version of each value. This is accessed by
/// getCastedVersionOf.
- std::map<Value*, Value*> CastedPointers;
+ DenseMap<Value*, Value*> CastedPointers;
/// DeadInsts - Keep track of instructions we may have made dead, so that
/// we can remove them after we are done working.
- std::set<Instruction*> DeadInsts;
+ SetVector<Instruction*> DeadInsts;
/// TLI - Keep a pointer of a TargetLowering to consult for determining
/// transformation profitability.
const TargetLowering *TLI;
public:
- LoopStrengthReduce(const TargetLowering *tli = NULL)
- : TLI(tli) {
+ static char ID; // Pass ID, replacement for typeid
+ explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
+ LoopPass((intptr_t)&ID), TLI(tli) {
}
- virtual bool runOnFunction(Function &) {
- LI = &getAnalysis<LoopInfo>();
- EF = &getAnalysis<ETForest>();
- SE = &getAnalysis<ScalarEvolution>();
- TD = &getAnalysis<TargetData>();
- UIntPtrTy = TD->getIntPtrType();
- Changed = false;
-
- for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
- runOnLoop(*I);
-
- return Changed;
- }
+ bool runOnLoop(Loop *L, LPPassManager &LPM);
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
// We split critical edges, so we change the CFG. However, we do update
// many analyses if they are around.
AU.addPreservedID(LoopSimplifyID);
AU.addPreserved<LoopInfo>();
- AU.addPreserved<DominatorSet>();
- AU.addPreserved<ETForest>();
- AU.addPreserved<ImmediateDominators>();
AU.addPreserved<DominanceFrontier>();
AU.addPreserved<DominatorTree>();
AU.addRequiredID(LoopSimplifyID);
AU.addRequired<LoopInfo>();
- AU.addRequired<ETForest>();
+ AU.addRequired<DominatorTree>();
AU.addRequired<TargetData>();
AU.addRequired<ScalarEvolution>();
}
/// getCastedVersionOf - Return the specified value casted to uintptr_t.
///
- Value *getCastedVersionOf(Value *V);
+ Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
private:
- void runOnLoop(Loop *L);
bool AddUsersIfInteresting(Instruction *I, Loop *L,
- std::set<Instruction*> &Processed);
- SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
-
+ SmallPtrSet<Instruction*,16> &Processed);
+ SCEVHandle GetExpressionSCEV(Instruction *E);
+ ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
+ IVStrideUse* &CondUse,
+ const SCEVHandle* &CondStride);
void OptimizeIndvars(Loop *L);
-
- unsigned CheckForIVReuse(const SCEVHandle&, IVExpr&, const Type*);
-
+ bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
+ const SCEVHandle *&CondStride);
+ bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
+ unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
+ IVExpr&, const Type*,
+ const std::vector<BasedUser>& UsersToProcess);
+ bool ValidStride(bool, int64_t,
+ const std::vector<BasedUser>& UsersToProcess);
+ SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
+ IVUsersOfOneStride &Uses,
+ Loop *L,
+ bool &AllUsesAreAddresses,
+ std::vector<BasedUser> &UsersToProcess);
void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
IVUsersOfOneStride &Uses,
Loop *L, bool isOnlyStride);
- void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
+ void DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts);
};
- RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
}
-FunctionPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
+char LoopStrengthReduce::ID = 0;
+static RegisterPass<LoopStrengthReduce>
+X("loop-reduce", "Loop Strength Reduction");
+
+LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
return new LoopStrengthReduce(TLI);
}
/// getCastedVersionOf - Return the specified value casted to uintptr_t. This
/// assumes that the Value* V is of integer or pointer type only.
///
-Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
+Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
+ Value *V) {
if (V->getType() == UIntPtrTy) return V;
if (Constant *CB = dyn_cast<Constant>(V))
- if (CB->getType()->isInteger())
- return ConstantExpr::getIntegerCast(CB, UIntPtrTy,
- CB->getType()->isSigned());
- else
- return ConstantExpr::getPtrToInt(CB, UIntPtrTy);
+ return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
Value *&New = CastedPointers[V];
if (New) return New;
- New = SCEVExpander::InsertCastOfTo(V, UIntPtrTy);
+ New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
DeadInsts.insert(cast<Instruction>(New));
return New;
}
/// specified set are trivially dead, delete them and see if this makes any of
/// their operands subsequently dead.
void LoopStrengthReduce::
-DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
+DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts) {
while (!Insts.empty()) {
- Instruction *I = *Insts.begin();
- Insts.erase(Insts.begin());
+ Instruction *I = Insts.back();
+ Insts.pop_back();
+
+ if (PHINode *PN = dyn_cast<PHINode>(I)) {
+ // If all incoming values to the Phi are the same, we can replace the Phi
+ // with that value.
+ if (Value *PNV = PN->hasConstantValue()) {
+ if (Instruction *U = dyn_cast<Instruction>(PNV))
+ Insts.insert(U);
+ SE->deleteValueFromRecords(PN);
+ PN->replaceAllUsesWith(PNV);
+ PN->eraseFromParent();
+ Changed = true;
+ continue;
+ }
+ }
+
if (isInstructionTriviallyDead(I)) {
- for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
- if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
+ for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
+ if (Instruction *U = dyn_cast<Instruction>(*i))
Insts.insert(U);
- SE->deleteInstructionFromRecords(I);
+ SE->deleteValueFromRecords(I);
I->eraseFromParent();
Changed = true;
}
/// GetExpressionSCEV - Compute and return the SCEV for the specified
/// instruction.
-SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
+SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
+ // Pointer to pointer bitcast instructions return the same value as their
+ // operand.
+ if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
+ if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
+ return SE->getSCEV(BCI);
+ SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
+ SE->setSCEV(BCI, R);
+ return R;
+ }
+
// Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
// If this is a GEP that SE doesn't know about, compute it now and insert it.
// If this is not a GEP, or if we have already done this computation, just let
return SE->getSCEV(Exp);
// Analyze all of the subscripts of this getelementptr instruction, looking
- // for uses that are determined by the trip count of L. First, skip all
- // operands the are not dependent on the IV.
+ // for uses that are determined by the trip count of the loop. First, skip
+ // all operands the are not dependent on the IV.
// Build up the base expression. Insert an LLVM cast of the pointer to
// uintptr_t first.
- SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
+ SCEVHandle GEPVal = SE->getUnknown(
+ getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
gep_type_iterator GTI = gep_type_begin(GEP);
- for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
+ for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
+ i != e; ++i, ++GTI) {
// If this is a use of a recurrence that we can analyze, and it comes before
// Op does in the GEP operand list, we will handle this when we process this
// operand.
if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
const StructLayout *SL = TD->getStructLayout(STy);
- unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
- uint64_t Offset = SL->MemberOffsets[Idx];
- GEPVal = SCEVAddExpr::get(GEPVal,
- SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
+ unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
+ uint64_t Offset = SL->getElementOffset(Idx);
+ GEPVal = SE->getAddExpr(GEPVal,
+ SE->getIntegerSCEV(Offset, UIntPtrTy));
} else {
- Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
+ unsigned GEPOpiBits =
+ (*i)->getType()->getPrimitiveSizeInBits();
+ unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
+ Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
+ Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
+ Instruction::BitCast));
+ Value *OpVal = getCastedVersionOf(opcode, *i);
SCEVHandle Idx = SE->getSCEV(OpVal);
- uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
+ uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
if (TypeSize != 1)
- Idx = SCEVMulExpr::get(Idx,
- SCEVConstant::get(ConstantInt::get(UIntPtrTy,
- TypeSize)));
- GEPVal = SCEVAddExpr::get(GEPVal, Idx);
+ Idx = SE->getMulExpr(Idx,
+ SE->getConstant(ConstantInt::get(UIntPtrTy,
+ TypeSize)));
+ GEPVal = SE->getAddExpr(GEPVal, Idx);
}
}
/// is. The stride must be a loop invariant expression, but the start may be
/// a mix of loop invariant and loop variant expressions.
static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
- SCEVHandle &Start, SCEVHandle &Stride) {
+ SCEVHandle &Start, SCEVHandle &Stride,
+ ScalarEvolution *SE) {
SCEVHandle TheAddRec = Start; // Initialize to zero.
// If the outer level is an AddExpr, the operands are all start values except
if (SCEVAddRecExpr *AddRec =
dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
if (AddRec->getLoop() == L)
- TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
+ TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
else
return false; // Nested IV of some sort?
} else {
- Start = SCEVAddExpr::get(Start, AE->getOperand(i));
+ Start = SE->getAddExpr(Start, AE->getOperand(i));
}
} else if (isa<SCEVAddRecExpr>(SH)) {
// FIXME: Generalize to non-affine IV's.
if (!AddRec->isAffine()) return false;
- Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
+ Start = SE->getAddExpr(Start, AddRec->getOperand(0));
if (!isa<SCEVConstant>(AddRec->getOperand(1)))
DOUT << "[" << L->getHeader()->getName()
<< "] Variable stride: " << *AddRec << "\n";
Stride = AddRec->getOperand(1);
- // Check that all constant strides are the unsigned type, we don't want to
- // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
- // merged.
- assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
- "Constants should be canonicalized to unsigned!");
-
return true;
}
/// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
/// should use the post-inc value).
static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
- Loop *L, ETForest *EF, Pass *P) {
+ Loop *L, DominatorTree *DT, Pass *P,
+ SetVector<Instruction*> &DeadInsts){
// If the user is in the loop, use the preinc value.
if (L->contains(User->getParent())) return false;
// Ok, the user is outside of the loop. If it is dominated by the latch
// block, use the post-inc value.
- if (EF->dominates(LatchBlock, User->getParent()))
+ if (DT->dominates(LatchBlock, User->getParent()))
return true;
// There is one case we have to be careful of: PHI nodes. These little guys
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
if (PN->getIncomingValue(i) == IV) {
++NumUses;
- if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
+ if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
return false;
}
// post-incremented value.
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
if (PN->getIncomingValue(i) == IV) {
- SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P,
- true);
+ SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
// Splitting the critical edge can reduce the number of entries in this
// PHI.
e = PN->getNumIncomingValues();
if (--NumUses == 0) break;
}
+
+ // PHI node might have become a constant value after SplitCriticalEdge.
+ DeadInsts.insert(User);
return true;
}
/// reducible SCEV, recursively add its users to the IVUsesByStride set and
/// return true. Otherwise, return false.
bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
- std::set<Instruction*> &Processed) {
+ SmallPtrSet<Instruction*,16> &Processed) {
if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
- return false; // Void and FP expressions cannot be reduced.
- if (!Processed.insert(I).second)
+ return false; // Void and FP expressions cannot be reduced.
+ if (!Processed.insert(I))
return true; // Instruction already handled.
// Get the symbolic expression for this instruction.
- SCEVHandle ISE = GetExpressionSCEV(I, L);
+ SCEVHandle ISE = GetExpressionSCEV(I);
if (isa<SCEVCouldNotCompute>(ISE)) return false;
// Get the start and stride for this expression.
- SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
+ SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
SCEVHandle Stride = Start;
- if (!getSCEVStartAndStride(ISE, L, Start, Stride))
+ if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
return false; // Non-reducible symbolic expression, bail out.
-
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
- Instruction *User = cast<Instruction>(*UI);
+
+ std::vector<Instruction *> IUsers;
+ // Collect all I uses now because IVUseShouldUsePostIncValue may
+ // invalidate use_iterator.
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
+ IUsers.push_back(cast<Instruction>(*UI));
+
+ for (unsigned iused_index = 0, iused_size = IUsers.size();
+ iused_index != iused_size; ++iused_index) {
+
+ Instruction *User = IUsers[iused_index];
// Do not infinitely recurse on PHI nodes.
if (isa<PHINode>(User) && Processed.count(User))
// Okay, we found a user that we cannot reduce. Analyze the instruction
// and decide what to do with it. If we are a use inside of the loop, use
// the value before incrementation, otherwise use it after incrementation.
- if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
+ if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
// The value used will be incremented by the stride more than we are
// expecting, so subtract this off.
- SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
+ SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
StrideUses.addUser(NewStart, User, I);
StrideUses.Users.back().isUseOfPostIncrementedValue = true;
DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
/// BasedUser - For a particular base value, keep information about how we've
/// partitioned the expression so far.
struct BasedUser {
+ /// SE - The current ScalarEvolution object.
+ ScalarEvolution *SE;
+
/// Base - The Base value for the PHI node that needs to be inserted for
/// this use. As the use is processed, information gets moved from this
/// field to the Imm field (below). BasedUser values are sorted by this
// the loop.
bool isUseOfPostIncrementedValue;
- BasedUser(IVStrideUse &IVSU)
- : Base(IVSU.Offset), Inst(IVSU.User),
+ BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
+ : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
OperandValToReplace(IVSU.OperandValToReplace),
- Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
+ Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
// Once we rewrite the code to insert the new IVs we want, update the
// operands of Inst to use the new expression 'NewBase', with 'Imm' added
// to it.
void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
- SCEVExpander &Rewriter, Loop *L,
- Pass *P);
+ Instruction *InsertPt,
+ SCEVExpander &Rewriter, Loop *L, Pass *P,
+ SetVector<Instruction*> &DeadInsts);
Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
SCEVExpander &Rewriter,
}
// If there is no immediate value, skip the next part.
- if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
- if (SC->getValue()->isNullValue())
- return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
- OperandValToReplace->getType());
+ if (Imm->isZero())
+ return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
+
+ // If we are inserting the base and imm values in the same block, make sure to
+ // adjust the IP position if insertion reused a result.
+ if (IP == BaseInsertPt)
+ IP = Rewriter.getInsertionPoint();
// Always emit the immediate (if non-zero) into the same block as the user.
- SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
- return Rewriter.expandCodeFor(NewValSCEV, IP,
- OperandValToReplace->getType());
+ SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
+ return Rewriter.expandCodeFor(NewValSCEV, IP);
+
}
// Once we rewrite the code to insert the new IVs we want, update the
// operands of Inst to use the new expression 'NewBase', with 'Imm' added
-// to it.
+// to it. NewBasePt is the last instruction which contributes to the
+// value of NewBase in the case that it's a diffferent instruction from
+// the PHI that NewBase is computed from, or null otherwise.
+//
void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
- SCEVExpander &Rewriter,
- Loop *L, Pass *P) {
+ Instruction *NewBasePt,
+ SCEVExpander &Rewriter, Loop *L, Pass *P,
+ SetVector<Instruction*> &DeadInsts) {
if (!isa<PHINode>(Inst)) {
- Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
+ // By default, insert code at the user instruction.
+ BasicBlock::iterator InsertPt = Inst;
+
+ // However, if the Operand is itself an instruction, the (potentially
+ // complex) inserted code may be shared by many users. Because of this, we
+ // want to emit code for the computation of the operand right before its old
+ // computation. This is usually safe, because we obviously used to use the
+ // computation when it was computed in its current block. However, in some
+ // cases (e.g. use of a post-incremented induction variable) the NewBase
+ // value will be pinned to live somewhere after the original computation.
+ // In this case, we have to back off.
+ if (!isUseOfPostIncrementedValue) {
+ if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
+ InsertPt = NewBasePt;
+ ++InsertPt;
+ } else if (Instruction *OpInst
+ = dyn_cast<Instruction>(OperandValToReplace)) {
+ InsertPt = OpInst;
+ while (isa<PHINode>(InsertPt)) ++InsertPt;
+ }
+ }
+ Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
+ // Adjust the type back to match the Inst. Note that we can't use InsertPt
+ // here because the SCEVExpander may have inserted the instructions after
+ // that point, in its efforts to avoid inserting redundant expressions.
+ if (isa<PointerType>(OperandValToReplace->getType())) {
+ NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
+ NewVal,
+ OperandValToReplace->getType());
+ }
// Replace the use of the operand Value with the new Phi we just created.
Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
- DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
+ DOUT << " CHANGED: IMM =" << *Imm;
+ DOUT << " \tNEWBASE =" << *NewBase;
+ DOUT << " \tInst = " << *Inst;
return;
}
// have multiple entries for the same predecessor. We use a map to make sure
// that a PHI node only has a single Value* for each predecessor (which also
// prevents us from inserting duplicate code in some blocks).
- std::map<BasicBlock*, Value*> InsertedCode;
+ DenseMap<BasicBlock*, Value*> InsertedCode;
PHINode *PN = cast<PHINode>(Inst);
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
if (PN->getIncomingValue(i) == OperandValToReplace) {
(PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
// First step, split the critical edge.
- SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
+ SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
// Next step: move the basic block. In particular, if the PHI node
// is outside of the loop, and PredTI is in the loop, we want to
// Insert the code into the end of the predecessor block.
Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
+
+ // Adjust the type back to match the PHI. Note that we can't use
+ // InsertPt here because the SCEVExpander may have inserted its
+ // instructions after that point, in its efforts to avoid inserting
+ // redundant expressions.
+ if (isa<PointerType>(PN->getType())) {
+ Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
+ Code,
+ PN->getType());
+ }
}
// Replace the use of the operand Value with the new Phi we just created.
Rewriter.clear();
}
}
+
+ // PHI node might have become a constant value after SplitCriticalEdge.
+ DeadInsts.insert(Inst);
+
DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
}
/// isTargetConstant - Return true if the following can be referenced by the
/// immediate field of a target instruction.
-static bool isTargetConstant(const SCEVHandle &V, const TargetLowering *TLI) {
+static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
+ const TargetLowering *TLI) {
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
- int64_t V = SC->getValue()->getSExtValue();
- if (TLI)
- return TLI->isLegalAddressImmediate(V);
- else
+ int64_t VC = SC->getValue()->getSExtValue();
+ if (TLI) {
+ TargetLowering::AddrMode AM;
+ AM.BaseOffs = VC;
+ return TLI->isLegalAddressingMode(AM, UseTy);
+ } else {
// Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
- return (V > -(1 << 16) && V < (1 << 16)-1);
+ return (VC > -(1 << 16) && VC < (1 << 16)-1);
+ }
}
if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
- if (CE->getOpcode() == Instruction::PtrToInt) {
+ if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
Constant *Op0 = CE->getOperand(0);
- if (isa<GlobalValue>(Op0) && TLI &&
- TLI->isLegalAddressImmediate(cast<GlobalValue>(Op0)))
- return true;
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
+ TargetLowering::AddrMode AM;
+ AM.BaseGV = GV;
+ return TLI->isLegalAddressingMode(AM, UseTy);
+ }
}
return false;
}
/// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
/// loop varying to the Imm operand.
static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
- Loop *L) {
+ Loop *L, ScalarEvolution *SE) {
if (Val->isLoopInvariant(L)) return; // Nothing to do.
if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
if (!SAE->getOperand(i)->isLoopInvariant(L)) {
// If this is a loop-variant expression, it must stay in the immediate
// field of the expression.
- Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
+ Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
} else {
NewOps.push_back(SAE->getOperand(i));
}
if (NewOps.empty())
- Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
+ Val = SE->getIntegerSCEV(0, Val->getType());
else
- Val = SCEVAddExpr::get(NewOps);
+ Val = SE->getAddExpr(NewOps);
} else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
// Try to pull immediates out of the start value of nested addrec's.
SCEVHandle Start = SARE->getStart();
- MoveLoopVariantsToImediateField(Start, Imm, L);
+ MoveLoopVariantsToImediateField(Start, Imm, L, SE);
std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
Ops[0] = Start;
- Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
+ Val = SE->getAddRecExpr(Ops, SARE->getLoop());
} else {
// Otherwise, all of Val is variant, move the whole thing over.
- Imm = SCEVAddExpr::get(Imm, Val);
- Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
+ Imm = SE->getAddExpr(Imm, Val);
+ Val = SE->getIntegerSCEV(0, Val->getType());
}
}
/// that can fit into the immediate field of instructions in the target.
/// Accumulate these immediate values into the Imm value.
static void MoveImmediateValues(const TargetLowering *TLI,
+ Instruction *User,
SCEVHandle &Val, SCEVHandle &Imm,
- bool isAddress, Loop *L) {
+ bool isAddress, Loop *L,
+ ScalarEvolution *SE) {
+ const Type *UseTy = User->getType();
+ if (StoreInst *SI = dyn_cast<StoreInst>(User))
+ UseTy = SI->getOperand(0)->getType();
+
if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
std::vector<SCEVHandle> NewOps;
NewOps.reserve(SAE->getNumOperands());
for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
SCEVHandle NewOp = SAE->getOperand(i);
- MoveImmediateValues(TLI, NewOp, Imm, isAddress, L);
+ MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
if (!NewOp->isLoopInvariant(L)) {
// If this is a loop-variant expression, it must stay in the immediate
// field of the expression.
- Imm = SCEVAddExpr::get(Imm, NewOp);
+ Imm = SE->getAddExpr(Imm, NewOp);
} else {
NewOps.push_back(NewOp);
}
}
if (NewOps.empty())
- Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
+ Val = SE->getIntegerSCEV(0, Val->getType());
else
- Val = SCEVAddExpr::get(NewOps);
+ Val = SE->getAddExpr(NewOps);
return;
} else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
// Try to pull immediates out of the start value of nested addrec's.
SCEVHandle Start = SARE->getStart();
- MoveImmediateValues(TLI, Start, Imm, isAddress, L);
+ MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
if (Start != SARE->getStart()) {
std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
Ops[0] = Start;
- Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
+ Val = SE->getAddRecExpr(Ops, SARE->getLoop());
}
return;
} else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
// Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
- if (isAddress && isTargetConstant(SME->getOperand(0), TLI) &&
+ if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
- SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
+ SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
SCEVHandle NewOp = SME->getOperand(1);
- MoveImmediateValues(TLI, NewOp, SubImm, isAddress, L);
+ MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
// If we extracted something out of the subexpressions, see if we can
// simplify this!
if (NewOp != SME->getOperand(1)) {
// Scale SubImm up by "8". If the result is a target constant, we are
// good.
- SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
- if (isTargetConstant(SubImm, TLI)) {
+ SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
+ if (isTargetConstant(SubImm, UseTy, TLI)) {
// Accumulate the immediate.
- Imm = SCEVAddExpr::get(Imm, SubImm);
+ Imm = SE->getAddExpr(Imm, SubImm);
// Update what is left of 'Val'.
- Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
+ Val = SE->getMulExpr(SME->getOperand(0), NewOp);
return;
}
}
// Loop-variant expressions must stay in the immediate field of the
// expression.
- if ((isAddress && isTargetConstant(Val, TLI)) ||
+ if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
!Val->isLoopInvariant(L)) {
- Imm = SCEVAddExpr::get(Imm, Val);
- Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
+ Imm = SE->getAddExpr(Imm, Val);
+ Val = SE->getIntegerSCEV(0, Val->getType());
return;
}
/// added together. This is used to reassociate common addition subexprs
/// together for maximal sharing when rewriting bases.
static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
- SCEVHandle Expr) {
+ SCEVHandle Expr,
+ ScalarEvolution *SE) {
if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
- SeparateSubExprs(SubExprs, AE->getOperand(j));
+ SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
} else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
- SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
+ SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
if (SARE->getOperand(0) == Zero) {
SubExprs.push_back(Expr);
} else {
// Compute the addrec with zero as its base.
std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
Ops[0] = Zero; // Start with zero base.
- SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
+ SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
- SeparateSubExprs(SubExprs, SARE->getOperand(0));
+ SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
}
- } else if (!isa<SCEVConstant>(Expr) ||
- !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
+ } else if (!Expr->isZero()) {
// Do not add zero.
SubExprs.push_back(Expr);
}
/// removed, accumulated, and returned. This looks for things like (a+b+c) and
/// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
static SCEVHandle
-RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
+RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
+ ScalarEvolution *SE) {
unsigned NumUses = Uses.size();
// Only one use? Use its base, regardless of what it is!
- SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
+ SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
SCEVHandle Result = Zero;
if (NumUses == 1) {
std::swap(Result, Uses[0].Base);
if (Uses[i].Base == Zero) return Zero;
// Split the expression into subexprs.
- SeparateSubExprs(SubExprs, Uses[i].Base);
+ SeparateSubExprs(SubExprs, Uses[i].Base, SE);
// Add one to SubExpressionUseCounts for each subexpr present.
for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
if (++SubExpressionUseCounts[SubExprs[j]] == 1)
SubExpressionUseCounts.find(UniqueSubExprs[i]);
assert(I != SubExpressionUseCounts.end() && "Entry not found?");
if (I->second == NumUses) { // Found CSE!
- Result = SCEVAddExpr::get(Result, I->first);
+ Result = SE->getAddExpr(Result, I->first);
} else {
// Remove non-cse's from SubExpressionUseCounts.
SubExpressionUseCounts.erase(I);
// Otherwise, remove all of the CSE's we found from each of the base values.
for (unsigned i = 0; i != NumUses; ++i) {
// Split the expression into subexprs.
- SeparateSubExprs(SubExprs, Uses[i].Base);
+ SeparateSubExprs(SubExprs, Uses[i].Base, SE);
// Remove any common subexpressions.
for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
if (SubExprs.empty())
Uses[i].Base = Zero;
else
- Uses[i].Base = SCEVAddExpr::get(SubExprs);
+ Uses[i].Base = SE->getAddExpr(SubExprs);
SubExprs.clear();
}
return Result;
}
-/// isZero - returns true if the scalar evolution expression is zero.
+/// ValidStride - Check whether the given Scale is valid for all loads and
+/// stores in UsersToProcess.
///
-static bool isZero(SCEVHandle &V) {
- if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
- return SC->getValue()->getZExtValue() == 0;
- return false;
+bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
+ int64_t Scale,
+ const std::vector<BasedUser>& UsersToProcess) {
+ if (!TLI)
+ return true;
+
+ for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
+ // If this is a load or other access, pass the type of the access in.
+ const Type *AccessTy = Type::VoidTy;
+ if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
+ AccessTy = SI->getOperand(0)->getType();
+ else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
+ AccessTy = LI->getType();
+ else if (isa<PHINode>(UsersToProcess[i].Inst))
+ continue;
+
+ TargetLowering::AddrMode AM;
+ if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
+ AM.BaseOffs = SC->getValue()->getSExtValue();
+ AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
+ AM.Scale = Scale;
+
+ // If load[imm+r*scale] is illegal, bail out.
+ if (!TLI->isLegalAddressingMode(AM, AccessTy))
+ return false;
+ }
+ return true;
}
+/// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
+/// a nop.
+bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
+ const Type *Ty2) {
+ if (Ty1 == Ty2)
+ return false;
+ if (TLI && TLI->isTruncateFree(Ty1, Ty2))
+ return false;
+ return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
+ !(isa<PointerType>(Ty2) &&
+ Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
+ !(isa<PointerType>(Ty1) &&
+ Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
+}
/// CheckForIVReuse - Returns the multiple if the stride is the multiple
/// of a previous stride and it is a legal value for the target addressing
-/// mode scale component. This allows the users of this stride to be rewritten
-/// as prev iv * factor. It returns 0 if no reuse is possible.
-unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride,
- IVExpr &IV, const Type *Ty) {
- if (!TLI) return 0;
-
+/// mode scale component and optional base reg. This allows the users of
+/// this stride to be rewritten as prev iv * factor. It returns 0 if no
+/// reuse is possible.
+unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
+ bool AllUsesAreAddresses,
+ const SCEVHandle &Stride,
+ IVExpr &IV, const Type *Ty,
+ const std::vector<BasedUser>& UsersToProcess) {
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
int64_t SInt = SC->getValue()->getSExtValue();
- if (SInt == 1) return 0;
-
- for (TargetLowering::legal_am_scale_iterator
- I = TLI->legal_am_scale_begin(), E = TLI->legal_am_scale_end();
- I != E; ++I) {
- unsigned Scale = *I;
- if (unsigned(abs(SInt)) < Scale || (SInt % Scale) != 0)
+ for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
+ ++NewStride) {
+ std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
+ IVsByStride.find(StrideOrder[NewStride]);
+ if (SI == IVsByStride.end())
continue;
- std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
- IVsByStride.find(SCEVUnknown::getIntegerSCEV(SInt/Scale, Type::UIntTy));
- if (SI == IVsByStride.end())
+ int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
+ if (SI->first != Stride &&
+ (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
continue;
- for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
- IE = SI->second.IVs.end(); II != IE; ++II)
- // FIXME: Only handle base == 0 for now.
- // Only reuse previous IV if it would not require a type conversion.
- if (isZero(II->Base) &&
- II->Base->getType()->canLosslesslyBitCastTo(Ty)) {
- IV = *II;
- return Scale;
- }
+ int64_t Scale = SInt / SSInt;
+ // Check that this stride is valid for all the types used for loads and
+ // stores; if it can be used for some and not others, we might as well use
+ // the original stride everywhere, since we have to create the IV for it
+ // anyway. If the scale is 1, then we don't need to worry about folding
+ // multiplications.
+ if (Scale == 1 ||
+ (AllUsesAreAddresses &&
+ ValidStride(HasBaseReg, Scale, UsersToProcess)))
+ for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
+ IE = SI->second.IVs.end(); II != IE; ++II)
+ // FIXME: Only handle base == 0 for now.
+ // Only reuse previous IV if it would not require a type conversion.
+ if (II->Base->isZero() &&
+ !RequiresTypeConversion(II->Base->getType(), Ty)) {
+ IV = *II;
+ return Scale;
+ }
}
}
-
return 0;
}
return Val.isUseOfPostIncrementedValue;
}
-/// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
-/// stride of IV. All of the users may have different starting values, and this
-/// may not be the only stride (we know it is if isOnlyStride is true).
-void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
- IVUsersOfOneStride &Uses,
- Loop *L,
- bool isOnlyStride) {
- // Transform our list of users and offsets to a bit more complex table. In
- // this new vector, each 'BasedUser' contains 'Base' the base of the
- // strided accessas well as the old information from Uses. We progressively
- // move information from the Base field to the Imm field, until we eventually
- // have the full access expression to rewrite the use.
- std::vector<BasedUser> UsersToProcess;
+/// isNonConstantNegative - Return true if the specified scev is negated, but
+/// not a constant.
+static bool isNonConstantNegative(const SCEVHandle &Expr) {
+ SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
+ if (!Mul) return false;
+
+ // If there is a constant factor, it will be first.
+ SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
+ if (!SC) return false;
+
+ // Return true if the value is negative, this matches things like (-42 * V).
+ return SC->getValue()->getValue().isNegative();
+}
+
+/// isAddress - Returns true if the specified instruction is using the
+/// specified value as an address.
+static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
+ bool isAddress = isa<LoadInst>(Inst);
+ if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
+ if (SI->getOperand(1) == OperandVal)
+ isAddress = true;
+ } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
+ // Addressing modes can also be folded into prefetches and a variety
+ // of intrinsics.
+ switch (II->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::prefetch:
+ case Intrinsic::x86_sse2_loadu_dq:
+ case Intrinsic::x86_sse2_loadu_pd:
+ case Intrinsic::x86_sse_loadu_ps:
+ case Intrinsic::x86_sse_storeu_ps:
+ case Intrinsic::x86_sse2_storeu_pd:
+ case Intrinsic::x86_sse2_storeu_dq:
+ case Intrinsic::x86_sse2_storel_dq:
+ if (II->getOperand(1) == OperandVal)
+ isAddress = true;
+ break;
+ }
+ }
+ return isAddress;
+}
+
+// CollectIVUsers - Transform our list of users and offsets to a bit more
+// complex table. In this new vector, each 'BasedUser' contains 'Base', the base
+// of the strided accesses, as well as the old information from Uses. We
+// progressively move information from the Base field to the Imm field, until
+// we eventually have the full access expression to rewrite the use.
+SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
+ IVUsersOfOneStride &Uses,
+ Loop *L,
+ bool &AllUsesAreAddresses,
+ std::vector<BasedUser> &UsersToProcess) {
UsersToProcess.reserve(Uses.Users.size());
for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
- UsersToProcess.push_back(Uses.Users[i]);
+ UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
// Move any loop invariant operands from the offset field to the immediate
// field of the use, so that we don't try to use something before it is
// computed.
MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
- UsersToProcess.back().Imm, L);
+ UsersToProcess.back().Imm, L, SE);
assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
"Base value is not loop invariant!");
}
// "A+B"), emit it to the preheader, then remove the expression from the
// UsersToProcess base values.
SCEVHandle CommonExprs =
- RemoveCommonExpressionsFromUseBases(UsersToProcess);
-
- // Check if it is possible to reuse a IV with stride that is factor of this
- // stride. And the multiple is a number that can be encoded in the scale
- // field of the target addressing mode.
- PHINode *NewPHI = NULL;
- Value *IncV = NULL;
- IVExpr ReuseIV;
- unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
- CommonExprs->getType());
- if (RewriteFactor != 0) {
- DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
- << " and BASE " << *ReuseIV.Base << " :\n";
- NewPHI = ReuseIV.PHI;
- IncV = ReuseIV.IncV;
- }
+ RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
// Next, figure out what we can represent in the immediate fields of
// instructions. If we can represent anything there, move it to the imm
// fields of the BasedUsers. We do this so that it increases the commonality
// of the remaining uses.
+ unsigned NumPHI = 0;
for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
// If the user is not in the current loop, this means it is using the exit
// value of the IV. Do not put anything in the base, make sure it's all in
// the immediate field to allow as much factoring as possible.
if (!L->contains(UsersToProcess[i].Inst->getParent())) {
- UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
- UsersToProcess[i].Base);
+ UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
+ UsersToProcess[i].Base);
UsersToProcess[i].Base =
- SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
+ SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
} else {
// Addressing modes can be folded into loads and stores. Be careful that
// the store is through the expression, not of the expression though.
- bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
- if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
- if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
- isAddress = true;
+ bool isPHI = false;
+ bool isAddress = isAddressUse(UsersToProcess[i].Inst,
+ UsersToProcess[i].OperandValToReplace);
+ if (isa<PHINode>(UsersToProcess[i].Inst)) {
+ isPHI = true;
+ ++NumPHI;
+ }
+
+ // If this use isn't an address, then not all uses are addresses.
+ if (!isAddress && !isPHI)
+ AllUsesAreAddresses = false;
- MoveImmediateValues(TLI, UsersToProcess[i].Base, UsersToProcess[i].Imm,
- isAddress, L);
+ MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
+ UsersToProcess[i].Imm, isAddress, L, SE);
}
}
+ // If one of the use if a PHI node and all other uses are addresses, still
+ // allow iv reuse. Essentially we are trading one constant multiplication
+ // for one fewer iv.
+ if (NumPHI > 1)
+ AllUsesAreAddresses = false;
+
+ return CommonExprs;
+}
+
+/// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
+/// stride of IV. All of the users may have different starting values, and this
+/// may not be the only stride (we know it is if isOnlyStride is true).
+void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
+ IVUsersOfOneStride &Uses,
+ Loop *L,
+ bool isOnlyStride) {
+ // If all the users are moved to another stride, then there is nothing to do.
+ if (Uses.Users.empty())
+ return;
+
+ // Keep track if every use in UsersToProcess is an address. If they all are,
+ // we may be able to rewrite the entire collection of them in terms of a
+ // smaller-stride IV.
+ bool AllUsesAreAddresses = true;
+
+ // Transform our list of users and offsets to a bit more complex table. In
+ // this new vector, each 'BasedUser' contains 'Base' the base of the
+ // strided accessas well as the old information from Uses. We progressively
+ // move information from the Base field to the Imm field, until we eventually
+ // have the full access expression to rewrite the use.
+ std::vector<BasedUser> UsersToProcess;
+ SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
+ UsersToProcess);
+
+ // If we managed to find some expressions in common, we'll need to carry
+ // their value in a register and add it in for each use. This will take up
+ // a register operand, which potentially restricts what stride values are
+ // valid.
+ bool HaveCommonExprs = !CommonExprs->isZero();
+
+ // If all uses are addresses, check if it is possible to reuse an IV with a
+ // stride that is a factor of this stride. And that the multiple is a number
+ // that can be encoded in the scale field of the target addressing mode. And
+ // that we will have a valid instruction after this substition, including the
+ // immediate field, if any.
+ PHINode *NewPHI = NULL;
+ Value *IncV = NULL;
+ IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
+ SE->getIntegerSCEV(0, Type::Int32Ty),
+ 0, 0);
+ unsigned RewriteFactor = 0;
+ RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
+ Stride, ReuseIV, CommonExprs->getType(),
+ UsersToProcess);
+ if (RewriteFactor != 0) {
+ DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
+ << " and BASE " << *ReuseIV.Base << " :\n";
+ NewPHI = ReuseIV.PHI;
+ IncV = ReuseIV.IncV;
+ }
+
+ const Type *ReplacedTy = CommonExprs->getType();
+
// Now that we know what we need to do, insert the PHI node itself.
//
- DOUT << "INSERTING IV of STRIDE " << *Stride << " and BASE "
- << *CommonExprs << " :\n";
+ DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
+ << *Stride << " and BASE " << *CommonExprs << ": ";
SCEVExpander Rewriter(*SE, *LI);
SCEVExpander PreheaderRewriter(*SE, *LI);
BasicBlock *LatchBlock = L->getLoopLatch();
- const Type *ReplacedTy = CommonExprs->getType();
// Emit the initial base value into the loop preheader.
Value *CommonBaseV
- = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
- ReplacedTy);
+ = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
if (RewriteFactor == 0) {
// Create a new Phi for this base, and stick it in the loop header.
- NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
+ NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
++NumInserted;
// Add common base to the new Phi node.
NewPHI->addIncoming(CommonBaseV, Preheader);
+ // If the stride is negative, insert a sub instead of an add for the
+ // increment.
+ bool isNegative = isNonConstantNegative(Stride);
+ SCEVHandle IncAmount = Stride;
+ if (isNegative)
+ IncAmount = SE->getNegativeSCEV(Stride);
+
// Insert the stride into the preheader.
- Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
- ReplacedTy);
+ Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
if (!isa<ConstantInt>(StrideV)) ++NumVariable;
// Emit the increment of the base value before the terminator of the loop
// latch block, and add it to the Phi node.
- SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
- SCEVUnknown::get(StrideV));
+ SCEVHandle IncExp = SE->getUnknown(StrideV);
+ if (isNegative)
+ IncExp = SE->getNegativeSCEV(IncExp);
+ IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
- IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
- ReplacedTy);
+ IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
IncV->setName(NewPHI->getName()+".inc");
NewPHI->addIncoming(IncV, LatchBlock);
// Remember this in case a later stride is multiple of this.
IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
+
+ DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
} else {
Constant *C = dyn_cast<Constant>(CommonBaseV);
if (!C ||
(!C->isNullValue() &&
- !isTargetConstant(SCEVUnknown::get(CommonBaseV), TLI)))
+ !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
// We want the common base emitted into the preheader! This is just
// using cast as a copy so BitCast (no-op cast) is appropriate
CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
"commonbase", PreInsertPt);
}
+ DOUT << "\n";
// We want to emit code for users inside the loop first. To do this, we
// rearrange BasedUser so that the entries at the end have
// Get a base value.
SCEVHandle Base = UsersToProcess[i].Base;
- // Compact everything with this base to be consequetive with this one.
+ // Compact everything with this base to be consequtive with this one.
for (unsigned j = i+1; j != e; ++j) {
if (UsersToProcess[j].Base == Base) {
std::swap(UsersToProcess[i+1], UsersToProcess[j]);
while (!UsersToProcess.empty()) {
SCEVHandle Base = UsersToProcess.back().Base;
- DOUT << " INSERTING code for BASE = " << *Base << ":\n";
-
// Emit the code for Base into the preheader.
- Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
- ReplacedTy);
-
+ Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
+
+ DOUT << " INSERTING code for BASE = " << *Base << ":";
+ if (BaseV->hasName())
+ DOUT << " Result value name = %" << BaseV->getNameStr();
+ DOUT << "\n";
+
// If BaseV is a constant other than 0, make sure that it gets inserted into
// the preheader, instead of being forward substituted into the uses. We do
// this by forcing a BitCast (noop cast) to be inserted into the preheader
// in this case.
if (Constant *C = dyn_cast<Constant>(BaseV)) {
- if (!C->isNullValue() && !isTargetConstant(Base, TLI)) {
+ if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
// We want this constant emitted into the preheader! This is just
// using cast as a copy so BitCast (no-op cast) is appropriate
BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
- PreInsertPt);
+ PreInsertPt);
}
}
if (L->contains(User.Inst->getParent()))
User.Inst->moveBefore(LatchBlock->getTerminator());
}
- if (RewriteOp->getType() != ReplacedTy)
- RewriteOp = SCEVExpander::InsertCastOfTo(RewriteOp, ReplacedTy);
+ if (RewriteOp->getType() != ReplacedTy) {
+ Instruction::CastOps opcode = Instruction::Trunc;
+ if (ReplacedTy->getPrimitiveSizeInBits() ==
+ RewriteOp->getType()->getPrimitiveSizeInBits())
+ opcode = Instruction::BitCast;
+ RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
+ }
- SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
+ SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
+
+ // If we had to insert new instrutions for RewriteOp, we have to
+ // consider that they may not have been able to end up immediately
+ // next to RewriteOp, because non-PHI instructions may never precede
+ // PHI instructions in a block. In this case, remember where the last
+ // instruction was inserted so that if we're replacing a different
+ // PHI node, we can use the later point to expand the final
+ // RewriteExpr.
+ Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
+ if (RewriteOp == NewPHI) NewBasePt = 0;
// Clear the SCEVExpander's expression map so that we are guaranteed
// to have the code emitted where we expect it.
// If we are reusing the iv, then it must be multiplied by a constant
// factor take advantage of addressing mode scale component.
if (RewriteFactor != 0) {
- RewriteExpr =
- SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
- RewriteExpr->getType()),
- RewriteExpr);
+ RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
+ RewriteExpr->getType()),
+ RewriteExpr);
// The common base is emitted in the loop preheader. But since we
// are reusing an IV, it has not been used to initialize the PHI node.
// Add it to the expression used to rewrite the uses.
if (!isa<ConstantInt>(CommonBaseV) ||
- !cast<ConstantInt>(CommonBaseV)->isNullValue())
- RewriteExpr = SCEVAddExpr::get(RewriteExpr,
- SCEVUnknown::get(CommonBaseV));
+ !cast<ConstantInt>(CommonBaseV)->isZero())
+ RewriteExpr = SE->getAddExpr(RewriteExpr,
+ SE->getUnknown(CommonBaseV));
}
// Now that we know what we need to do, insert code before User for the
// immediate and any loop-variant expressions.
- if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
+ if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
// Add BaseV to the PHI value if needed.
- RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
+ RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
- User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
+ User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
+ Rewriter, L, this,
+ DeadInsts);
// Mark old value we replaced as possibly dead, so that it is elminated
// if we just replaced the last use of that value.
// different starting values, into different PHIs.
}
+/// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
+/// set the IV user and stride information and return true, otherwise return
+/// false.
+bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
+ const SCEVHandle *&CondStride) {
+ for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
+ ++Stride) {
+ std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
+ IVUsesByStride.find(StrideOrder[Stride]);
+ assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
+
+ for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
+ E = SI->second.Users.end(); UI != E; ++UI)
+ if (UI->User == Cond) {
+ // NOTE: we could handle setcc instructions with multiple uses here, but
+ // InstCombine does it as well for simple uses, it's not clear that it
+ // occurs enough in real life to handle.
+ CondUse = &*UI;
+ CondStride = &SI->first;
+ return true;
+ }
+ }
+ return false;
+}
+
+namespace {
+ // Constant strides come first which in turns are sorted by their absolute
+ // values. If absolute values are the same, then positive strides comes first.
+ // e.g.
+ // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
+ struct StrideCompare {
+ bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
+ SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
+ SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
+ if (LHSC && RHSC) {
+ int64_t LV = LHSC->getValue()->getSExtValue();
+ int64_t RV = RHSC->getValue()->getSExtValue();
+ uint64_t ALV = (LV < 0) ? -LV : LV;
+ uint64_t ARV = (RV < 0) ? -RV : RV;
+ if (ALV == ARV)
+ return LV > RV;
+ else
+ return ALV < ARV;
+ }
+ return (LHSC && !RHSC);
+ }
+ };
+}
+
+/// ChangeCompareStride - If a loop termination compare instruction is the
+/// only use of its stride, and the compaison is against a constant value,
+/// try eliminate the stride by moving the compare instruction to another
+/// stride and change its constant operand accordingly. e.g.
+///
+/// loop:
+/// ...
+/// v1 = v1 + 3
+/// v2 = v2 + 1
+/// if (v2 < 10) goto loop
+/// =>
+/// loop:
+/// ...
+/// v1 = v1 + 3
+/// if (v1 < 30) goto loop
+ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
+ IVStrideUse* &CondUse,
+ const SCEVHandle* &CondStride) {
+ if (StrideOrder.size() < 2 ||
+ IVUsesByStride[*CondStride].Users.size() != 1)
+ return Cond;
+ const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
+ if (!SC) return Cond;
+ ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
+ if (!C) return Cond;
+
+ ICmpInst::Predicate Predicate = Cond->getPredicate();
+ int64_t CmpSSInt = SC->getValue()->getSExtValue();
+ int64_t CmpVal = C->getValue().getSExtValue();
+ unsigned BitWidth = C->getValue().getBitWidth();
+ uint64_t SignBit = 1ULL << (BitWidth-1);
+ const Type *CmpTy = C->getType();
+ const Type *NewCmpTy = NULL;
+ unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
+ unsigned NewTyBits = 0;
+ int64_t NewCmpVal = CmpVal;
+ SCEVHandle *NewStride = NULL;
+ Value *NewIncV = NULL;
+ int64_t Scale = 1;
+
+ // Check stride constant and the comparision constant signs to detect
+ // overflow.
+ if (ICmpInst::isSignedPredicate(Predicate) &&
+ (CmpVal & SignBit) != (CmpSSInt & SignBit))
+ return Cond;
+
+ // Look for a suitable stride / iv as replacement.
+ std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
+ for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
+ std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
+ IVUsesByStride.find(StrideOrder[i]);
+ if (!isa<SCEVConstant>(SI->first))
+ continue;
+ int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
+ if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
+ continue;
+
+ Scale = SSInt / CmpSSInt;
+ NewCmpVal = CmpVal * Scale;
+ APInt Mul = APInt(BitWidth, NewCmpVal);
+ // Check for overflow.
+ if (Mul.getSExtValue() != NewCmpVal) {
+ NewCmpVal = CmpVal;
+ continue;
+ }
+
+ // Watch out for overflow.
+ if (ICmpInst::isSignedPredicate(Predicate) &&
+ (CmpVal & SignBit) != (NewCmpVal & SignBit))
+ NewCmpVal = CmpVal;
+
+ if (NewCmpVal != CmpVal) {
+ // Pick the best iv to use trying to avoid a cast.
+ NewIncV = NULL;
+ for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
+ E = SI->second.Users.end(); UI != E; ++UI) {
+ NewIncV = UI->OperandValToReplace;
+ if (NewIncV->getType() == CmpTy)
+ break;
+ }
+ if (!NewIncV) {
+ NewCmpVal = CmpVal;
+ continue;
+ }
+
+ NewCmpTy = NewIncV->getType();
+ NewTyBits = isa<PointerType>(NewCmpTy)
+ ? UIntPtrTy->getPrimitiveSizeInBits()
+ : NewCmpTy->getPrimitiveSizeInBits();
+ if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
+ // Check if it is possible to rewrite it using
+ // an iv / stride of a smaller integer type.
+ bool TruncOk = false;
+ if (NewCmpTy->isInteger()) {
+ unsigned Bits = NewTyBits;
+ if (ICmpInst::isSignedPredicate(Predicate))
+ --Bits;
+ uint64_t Mask = (1ULL << Bits) - 1;
+ if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
+ TruncOk = true;
+ }
+ if (!TruncOk) {
+ NewCmpVal = CmpVal;
+ continue;
+ }
+ }
+
+ // Don't rewrite if use offset is non-constant and the new type is
+ // of a different type.
+ // FIXME: too conservative?
+ if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
+ NewCmpVal = CmpVal;
+ continue;
+ }
+
+ bool AllUsesAreAddresses = true;
+ std::vector<BasedUser> UsersToProcess;
+ SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
+ AllUsesAreAddresses,
+ UsersToProcess);
+ // Avoid rewriting the compare instruction with an iv of new stride
+ // if it's likely the new stride uses will be rewritten using the
+ if (AllUsesAreAddresses &&
+ ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
+ NewCmpVal = CmpVal;
+ continue;
+ }
+
+ // If scale is negative, use swapped predicate unless it's testing
+ // for equality.
+ if (Scale < 0 && !Cond->isEquality())
+ Predicate = ICmpInst::getSwappedPredicate(Predicate);
+
+ NewStride = &StrideOrder[i];
+ break;
+ }
+ }
+
+ // Forgo this transformation if it the increment happens to be
+ // unfortunately positioned after the condition, and the condition
+ // has multiple uses which prevent it from being moved immediately
+ // before the branch. See
+ // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
+ // for an example of this situation.
+ if (!Cond->hasOneUse()) {
+ for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
+ I != E; ++I)
+ if (I == NewIncV)
+ return Cond;
+ }
+
+ if (NewCmpVal != CmpVal) {
+ // Create a new compare instruction using new stride / iv.
+ ICmpInst *OldCond = Cond;
+ Value *RHS;
+ if (!isa<PointerType>(NewCmpTy))
+ RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
+ else {
+ RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
+ RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
+ }
+ // Insert new compare instruction.
+ Cond = new ICmpInst(Predicate, NewIncV, RHS,
+ L->getHeader()->getName() + ".termcond",
+ OldCond);
+
+ // Remove the old compare instruction. The old indvar is probably dead too.
+ DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
+ SE->deleteValueFromRecords(OldCond);
+ OldCond->replaceAllUsesWith(Cond);
+ OldCond->eraseFromParent();
+
+ IVUsesByStride[*CondStride].Users.pop_back();
+ SCEVHandle NewOffset = TyBits == NewTyBits
+ ? SE->getMulExpr(CondUse->Offset,
+ SE->getConstant(ConstantInt::get(CmpTy, Scale)))
+ : SE->getConstant(ConstantInt::get(NewCmpTy,
+ cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
+ IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
+ CondUse = &IVUsesByStride[*NewStride].Users.back();
+ CondStride = NewStride;
+ ++NumEliminated;
+ }
+
+ return Cond;
+}
+
// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
// uses in the loop, look to see if we can eliminate some, in favor of using
// common indvars for the different uses.
void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
// TODO: implement optzns here.
-
-
-
// Finally, get the terminating condition for the loop if possible. If we
// can, we want to change it to use a post-incremented version of its
// induction variable, to allow coalescing the live ranges for the IV into
BasicBlock *LatchBlock =
SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
- if (!TermBr || TermBr->isUnconditional() ||
- !isa<SetCondInst>(TermBr->getCondition()))
+ if (!TermBr || TermBr->isUnconditional() ||
+ !isa<ICmpInst>(TermBr->getCondition()))
return;
- SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
+ ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
// Search IVUsesByStride to find Cond's IVUse if there is one.
IVStrideUse *CondUse = 0;
const SCEVHandle *CondStride = 0;
- for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
- ++Stride) {
- std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
- IVUsesByStride.find(StrideOrder[Stride]);
- assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
-
- for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
- E = SI->second.Users.end(); UI != E; ++UI)
- if (UI->User == Cond) {
- CondUse = &*UI;
- CondStride = &SI->first;
- // NOTE: we could handle setcc instructions with multiple uses here, but
- // InstCombine does it as well for simple uses, it's not clear that it
- // occurs enough in real life to handle.
- break;
- }
- }
- if (!CondUse) return; // setcc doesn't use the IV.
+ if (!FindIVUserForCond(Cond, CondUse, CondStride))
+ return; // setcc doesn't use the IV.
+
+ // If possible, change stride and operands of the compare instruction to
+ // eliminate one stride.
+ Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
// It's possible for the setcc instruction to be anywhere in the loop, and
// possible for it to have multiple users. If it is not immediately before
Cond->moveBefore(TermBr);
} else {
// Otherwise, clone the terminating condition and insert into the loopend.
- Cond = cast<SetCondInst>(Cond->clone());
+ Cond = cast<ICmpInst>(Cond->clone());
Cond->setName(L->getHeader()->getName() + ".termcond");
LatchBlock->getInstList().insert(TermBr, Cond);
// If we get to here, we know that we can transform the setcc instruction to
// use the post-incremented version of the IV, allowing us to coalesce the
// live ranges for the IV correctly.
- CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
+ CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
CondUse->isUseOfPostIncrementedValue = true;
+ Changed = true;
}
-namespace {
- // Constant strides come first which in turns are sorted by their absolute
- // values. If absolute values are the same, then positive strides comes first.
- // e.g.
- // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
- struct StrideCompare {
- bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
- SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
- SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
- if (LHSC && RHSC) {
- int64_t LV = LHSC->getValue()->getSExtValue();
- int64_t RV = RHSC->getValue()->getSExtValue();
- uint64_t ALV = (LV < 0) ? -LV : LV;
- uint64_t ARV = (RV < 0) ? -RV : RV;
- if (ALV == ARV)
- return LV > RV;
- else
- return ALV < ARV;
- }
- return (LHSC && !RHSC);
- }
- };
-}
+bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
-void LoopStrengthReduce::runOnLoop(Loop *L) {
- // First step, transform all loops nesting inside of this loop.
- for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
- runOnLoop(*I);
+ LI = &getAnalysis<LoopInfo>();
+ DT = &getAnalysis<DominatorTree>();
+ SE = &getAnalysis<ScalarEvolution>();
+ TD = &getAnalysis<TargetData>();
+ UIntPtrTy = TD->getIntPtrType();
+ Changed = false;
- // Next, find all uses of induction variables in this loop, and catagorize
+ // Find all uses of induction variables in this loop, and catagorize
// them by stride. Start by finding all of the PHI nodes in the header for
// this loop. If they are induction variables, inspect their uses.
- std::set<Instruction*> Processed; // Don't reprocess instructions.
+ SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
AddUsersIfInteresting(I, L, Processed);
- // If we have nothing to do, return.
- if (IVUsesByStride.empty()) return;
-
- // Optimize induction variables. Some indvar uses can be transformed to use
- // strides that will be needed for other purposes. A common example of this
- // is the exit test for the loop, which can often be rewritten to use the
- // computation of some other indvar to decide when to terminate the loop.
- OptimizeIndvars(L);
-
+ if (!IVUsesByStride.empty()) {
+ // Optimize induction variables. Some indvar uses can be transformed to use
+ // strides that will be needed for other purposes. A common example of this
+ // is the exit test for the loop, which can often be rewritten to use the
+ // computation of some other indvar to decide when to terminate the loop.
+ OptimizeIndvars(L);
- // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
- // doing computation in byte values, promote to 32-bit values if safe.
+ // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
+ // doing computation in byte values, promote to 32-bit values if safe.
- // FIXME: Attempt to reuse values across multiple IV's. In particular, we
- // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
- // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
- // to be careful that IV's are all the same type. Only works for intptr_t
- // indvars.
+ // FIXME: Attempt to reuse values across multiple IV's. In particular, we
+ // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
+ // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
+ // Need to be careful that IV's are all the same type. Only works for
+ // intptr_t indvars.
- // If we only have one stride, we can more aggressively eliminate some things.
- bool HasOneStride = IVUsesByStride.size() == 1;
+ // If we only have one stride, we can more aggressively eliminate some
+ // things.
+ bool HasOneStride = IVUsesByStride.size() == 1;
#ifndef NDEBUG
- DOUT << "\nLSR on ";
- DEBUG(L->dump());
+ DOUT << "\nLSR on ";
+ DEBUG(L->dump());
#endif
- // IVsByStride keeps IVs for one particular loop.
- IVsByStride.clear();
-
- // Sort the StrideOrder so we process larger strides first.
- std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
-
- // Note: this processes each stride/type pair individually. All users passed
- // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
- // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
- // This extra layer of indirection makes the ordering of strides deterministic
- // - not dependent on map order.
- for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
- std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
- IVUsesByStride.find(StrideOrder[Stride]);
- assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
- StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
+ // IVsByStride keeps IVs for one particular loop.
+ assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
+
+ // Sort the StrideOrder so we process larger strides first.
+ std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
+
+ // Note: this processes each stride/type pair individually. All users
+ // passed into StrengthReduceStridedIVUsers have the same type AND stride.
+ // Also, note that we iterate over IVUsesByStride indirectly by using
+ // StrideOrder. This extra layer of indirection makes the ordering of
+ // strides deterministic - not dependent on map order.
+ for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
+ std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
+ IVUsesByStride.find(StrideOrder[Stride]);
+ assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
+ StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
+ }
}
+ // We're done analyzing this loop; release all the state we built up for it.
+ CastedPointers.clear();
+ IVUsesByStride.clear();
+ IVsByStride.clear();
+ StrideOrder.clear();
+
// Clean up after ourselves
if (!DeadInsts.empty()) {
DeleteTriviallyDeadInstructions(DeadInsts);
BasicBlock::iterator I = L->getHeader()->begin();
- PHINode *PN;
- while ((PN = dyn_cast<PHINode>(I))) {
- ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
-
- // At this point, we know that we have killed one or more GEP
- // instructions. It is worth checking to see if the cann indvar is also
- // dead, so that we can remove it as well. The requirements for the cann
- // indvar to be considered dead are:
- // 1. the cann indvar has one use
- // 2. the use is an add instruction
- // 3. the add has one use
- // 4. the add is used by the cann indvar
- // If all four cases above are true, then we can remove both the add and
- // the cann indvar.
+ while (PHINode *PN = dyn_cast<PHINode>(I++)) {
+ // At this point, we know that we have killed one or more IV users.
+ // It is worth checking to see if the cann indvar is also
+ // dead, so that we can remove it as well.
+ //
+ // We can remove a PHI if it is on a cycle in the def-use graph
+ // where each node in the cycle has degree one, i.e. only one use,
+ // and is an instruction with no side effects.
+ //
// FIXME: this needs to eliminate an induction variable even if it's being
// compared against some value to decide loop termination.
if (PN->hasOneUse()) {
- BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
- if (BO && BO->hasOneUse()) {
- if (PN == *(BO->use_begin())) {
- DeadInsts.insert(BO);
- // Break the cycle, then delete the PHI.
+ SmallPtrSet<PHINode *, 2> PHIs;
+ for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
+ J && J->hasOneUse() && !J->mayWriteToMemory();
+ J = dyn_cast<Instruction>(*J->use_begin())) {
+ // If we find the original PHI, we've discovered a cycle.
+ if (J == PN) {
+ // Break the cycle and mark the PHI for deletion.
+ SE->deleteValueFromRecords(PN);
PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
- SE->deleteInstructionFromRecords(PN);
- PN->eraseFromParent();
+ DeadInsts.insert(PN);
+ Changed = true;
+ break;
}
+ // If we find a PHI more than once, we're on a cycle that
+ // won't prove fruitful.
+ if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
+ break;
}
}
}
DeleteTriviallyDeadInstructions(DeadInsts);
}
- CastedPointers.clear();
- IVUsesByStride.clear();
- StrideOrder.clear();
- return;
+ return Changed;
}
projects / oota-llvm.git / blobdiff