//
// 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/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Target/TargetData.h"
+#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Target/TargetLowering.h"
#include <algorithm>
#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(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 {
+ // Hidden options for help debugging.
+ cl::opt<bool> AllowPHIIVReuse("lsr-allow-phi-iv-reuse",
+ cl::init(true), cl::Hidden);
+}
namespace {
/// 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.
+ /// 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;
/// 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;
+ SmallPtrSet<Instruction*,16> DeadInsts;
/// TLI - Keep a pointer of a TargetLowering to consult for determining
/// transformation profitability.
Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
private:
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);
bool FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
const SCEVHandle *&CondStride);
-
- unsigned CheckForIVReuse(bool, const SCEVHandle&,
+ 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(SmallPtrSet<Instruction*,16> &Insts);
};
char LoopStrengthReduce::ID = 0;
RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
/// 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(SmallPtrSet<Instruction*,16> &Insts) {
while (!Insts.empty()) {
Instruction *I = *Insts.begin();
- Insts.erase(Insts.begin());
+ Insts.erase(I);
+
+ 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);
+ PN->replaceAllUsesWith(PNV);
+ SE->deleteValueFromRecords(PN);
+ 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)))
/// 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)), L);
+ SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
SE->setSCEV(BCI, R);
return R;
}
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.
/// 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, DominatorTree *DT, Pass *P) {
+ Loop *L, DominatorTree *DT, Pass *P,
+ SmallPtrSet<Instruction*,16> &DeadInsts){
// If the user is in the loop, use the preinc value.
if (L->contains(User->getParent())) 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)
+ 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.
// 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, DT, 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 = SE->getMinusSCEV(Start, Stride);
// 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);
+ SCEVExpander &Rewriter, Loop *L, Pass *P,
+ SmallPtrSet<Instruction*,16> &DeadInsts);
Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
SCEVExpander &Rewriter,
// operands of Inst to use the new expression 'NewBase', with 'Imm' added
// to it.
void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
- SCEVExpander &Rewriter,
- Loop *L, Pass *P) {
+ SCEVExpander &Rewriter, Loop *L, Pass *P,
+ SmallPtrSet<Instruction*,16> &DeadInsts) {
if (!isa<PHINode>(Inst)) {
// By default, insert code at the user instruction.
BasicBlock::iterator InsertPt = Inst;
// 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
Rewriter.clear();
}
}
+
+ // PHI node might have become a constant value after SplitCriticalEdge.
+ DeadInsts.insert(Inst);
+
DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
}
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;
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)) {
+ if (AllowPHIIVReuse)
+ continue;
+ return false;
+ }
TargetLowering::AddrMode AM;
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
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 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 (!TLI) return 0;
-
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
int64_t SInt = SC->getValue()->getSExtValue();
- if (SInt == 1) return 0;
-
- for (std::map<SCEVHandle, IVsOfOneStride>::iterator SI= IVsByStride.begin(),
- SE = IVsByStride.end(); SI != SE; ++SI) {
+ 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;
int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
- if (SInt != -SSInt &&
+ if (SI->first != Stride &&
(unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
continue;
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 (ValidStride(HasBaseReg, Scale, UsersToProcess))
+ // 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 (isZero(II->Base) && II->Base->getType() == Ty) {
+ if (isZero(II->Base) &&
+ !RequiresTypeConversion(II->Base->getType(), Ty)) {
IV = *II;
return Scale;
}
return SC->getValue()->getValue().isNegative();
}
-/// 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;
+/// 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;
+ case Intrinsic::x86_sse2_loadh_pd:
+ case Intrinsic::x86_sse2_loadl_pd:
+ if (II->getOperand(2) == 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 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.
+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(BasedUser(Uses.Users[i], SE));
SCEVHandle CommonExprs =
RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
- // 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 = !isZero(CommonExprs);
-
- // 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;
-
// 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
// 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;
- } else if (IntrinsicInst *II =
- dyn_cast<IntrinsicInst>(UsersToProcess[i].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) == UsersToProcess[i].OperandValToReplace)
- isAddress = true;
- break;
- case Intrinsic::x86_sse2_loadh_pd:
- case Intrinsic::x86_sse2_loadl_pd:
- if (II->getOperand(2) == UsersToProcess[i].OperandValToReplace)
- isAddress = true;
- break;
- }
+ 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)
+ if (!isAddress && !(AllowPHIIVReuse && isPHI))
AllUsesAreAddresses = false;
MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
}
}
+ // 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 = !isZero(CommonExprs);
+
// 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
SE->getIntegerSCEV(0, Type::Int32Ty),
0, 0);
unsigned RewriteFactor = 0;
- if (AllUsesAreAddresses)
- RewriteFactor = CheckForIVReuse(HaveCommonExprs, Stride, ReuseIV,
- CommonExprs->getType(),
- UsersToProcess);
+ 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";
// 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]);
// 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 =
- SE->getMulExpr(SE->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 BaseV to the PHI value if needed.
RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
- User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
+ User.RewriteInstructionToUseNewBase(RewriteExpr, 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.
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;
+
+ // 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 a 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(!isZero(CommonExprs), Scale, UsersToProcess)) {
+ NewCmpVal = CmpVal;
+ continue;
+ }
+
+ // If scale is negative, use inverse predicate unless it's testing
+ // for equality.
+ if (Scale < 0 && !Cond->isEquality())
+ Predicate = ICmpInst::getInversePredicate(Predicate);
+
+ NewStride = &StrideOrder[i];
+ break;
+ }
+ }
+
+ 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);
+ Cond->setName(L->getHeader()->getName() + ".termcond");
+ OldCond->getParent()->getInstList().insert(OldCond, Cond);
+
+ // Remove the old compare instruction. The old indvar is probably dead too.
+ DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
+ OldCond->replaceAllUsesWith(Cond);
+ SE->deleteValueFromRecords(OldCond);
+ 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.
if (!FindIVForUser(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
CondUse->isUseOfPostIncrementedValue = 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) {
LI = &getAnalysis<LoopInfo>();
// 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);
// 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.
+ // 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) {
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