1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs a strength reduction on array references inside loops that
11 // have as one or more of their components the loop induction variable. This is
12 // accomplished by creating a new Value to hold the initial value of the array
13 // access for the first iteration, and then creating a new GEP instruction in
14 // the loop to increment the value by the appropriate amount.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "loop-reduce"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Type.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/ScalarEvolutionExpander.h"
29 #include "llvm/Support/CFG.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Target/TargetLowering.h"
43 STATISTIC(NumReduced , "Number of GEPs strength reduced");
44 STATISTIC(NumInserted, "Number of PHIs inserted");
45 STATISTIC(NumVariable, "Number of PHIs with variable strides");
46 STATISTIC(NumEliminated , "Number of strides eliminated");
52 /// IVStrideUse - Keep track of one use of a strided induction variable, where
53 /// the stride is stored externally. The Offset member keeps track of the
54 /// offset from the IV, User is the actual user of the operand, and
55 /// 'OperandValToReplace' is the operand of the User that is the use.
56 struct VISIBILITY_HIDDEN IVStrideUse {
59 Value *OperandValToReplace;
61 // isUseOfPostIncrementedValue - True if this should use the
62 // post-incremented version of this IV, not the preincremented version.
63 // This can only be set in special cases, such as the terminating setcc
64 // instruction for a loop or uses dominated by the loop.
65 bool isUseOfPostIncrementedValue;
67 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
68 : Offset(Offs), User(U), OperandValToReplace(O),
69 isUseOfPostIncrementedValue(false) {}
72 /// IVUsersOfOneStride - This structure keeps track of all instructions that
73 /// have an operand that is based on the trip count multiplied by some stride.
74 /// The stride for all of these users is common and kept external to this
76 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
77 /// Users - Keep track of all of the users of this stride as well as the
78 /// initial value and the operand that uses the IV.
79 std::vector<IVStrideUse> Users;
81 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
82 Users.push_back(IVStrideUse(Offset, User, Operand));
86 /// IVInfo - This structure keeps track of one IV expression inserted during
87 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
88 /// well as the PHI node and increment value created for rewrite.
89 struct VISIBILITY_HIDDEN IVExpr {
95 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
97 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
100 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
101 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
102 struct VISIBILITY_HIDDEN IVsOfOneStride {
103 std::vector<IVExpr> IVs;
105 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
107 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
111 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
115 const TargetData *TD;
116 const Type *UIntPtrTy;
119 /// IVUsesByStride - Keep track of all uses of induction variables that we
120 /// are interested in. The key of the map is the stride of the access.
121 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
123 /// IVsByStride - Keep track of all IVs that have been inserted for a
124 /// particular stride.
125 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
127 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
128 /// We use this to iterate over the IVUsesByStride collection without being
129 /// dependent on random ordering of pointers in the process.
130 SmallVector<SCEVHandle, 16> StrideOrder;
132 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
133 /// of the casted version of each value. This is accessed by
134 /// getCastedVersionOf.
135 DenseMap<Value*, Value*> CastedPointers;
137 /// DeadInsts - Keep track of instructions we may have made dead, so that
138 /// we can remove them after we are done working.
139 SmallPtrSet<Instruction*,16> DeadInsts;
141 /// TLI - Keep a pointer of a TargetLowering to consult for determining
142 /// transformation profitability.
143 const TargetLowering *TLI;
146 static char ID; // Pass ID, replacement for typeid
147 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
148 LoopPass((intptr_t)&ID), TLI(tli) {
151 bool runOnLoop(Loop *L, LPPassManager &LPM);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 // We split critical edges, so we change the CFG. However, we do update
155 // many analyses if they are around.
156 AU.addPreservedID(LoopSimplifyID);
157 AU.addPreserved<LoopInfo>();
158 AU.addPreserved<DominanceFrontier>();
159 AU.addPreserved<DominatorTree>();
161 AU.addRequiredID(LoopSimplifyID);
162 AU.addRequired<LoopInfo>();
163 AU.addRequired<DominatorTree>();
164 AU.addRequired<TargetData>();
165 AU.addRequired<ScalarEvolution>();
168 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
170 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
172 bool AddUsersIfInteresting(Instruction *I, Loop *L,
173 SmallPtrSet<Instruction*,16> &Processed);
174 SCEVHandle GetExpressionSCEV(Instruction *E);
175 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
176 IVStrideUse* &CondUse,
177 const SCEVHandle* &CondStride);
178 void OptimizeIndvars(Loop *L);
179 bool FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
180 const SCEVHandle *&CondStride);
181 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
182 unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
183 IVExpr&, const Type*,
184 const std::vector<BasedUser>& UsersToProcess);
185 bool ValidStride(bool, int64_t,
186 const std::vector<BasedUser>& UsersToProcess);
187 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
188 IVUsersOfOneStride &Uses,
190 bool &AllUsesAreAddresses,
191 std::vector<BasedUser> &UsersToProcess);
192 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
193 IVUsersOfOneStride &Uses,
194 Loop *L, bool isOnlyStride);
195 void DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*,16> &Insts);
197 char LoopStrengthReduce::ID = 0;
198 RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
201 LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
202 return new LoopStrengthReduce(TLI);
205 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
206 /// assumes that the Value* V is of integer or pointer type only.
208 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
210 if (V->getType() == UIntPtrTy) return V;
211 if (Constant *CB = dyn_cast<Constant>(V))
212 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
214 Value *&New = CastedPointers[V];
217 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
218 DeadInsts.insert(cast<Instruction>(New));
223 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
224 /// specified set are trivially dead, delete them and see if this makes any of
225 /// their operands subsequently dead.
226 void LoopStrengthReduce::
227 DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*,16> &Insts) {
228 while (!Insts.empty()) {
229 Instruction *I = *Insts.begin();
232 if (PHINode *PN = dyn_cast<PHINode>(I)) {
233 // If all incoming values to the Phi are the same, we can replace the Phi
235 if (Value *PNV = PN->hasConstantValue()) {
236 if (Instruction *U = dyn_cast<Instruction>(PNV))
238 PN->replaceAllUsesWith(PNV);
239 SE->deleteValueFromRecords(PN);
240 PN->eraseFromParent();
246 if (isInstructionTriviallyDead(I)) {
247 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
248 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
250 SE->deleteValueFromRecords(I);
251 I->eraseFromParent();
258 /// GetExpressionSCEV - Compute and return the SCEV for the specified
260 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
261 // Pointer to pointer bitcast instructions return the same value as their
263 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
264 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
265 return SE->getSCEV(BCI);
266 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
271 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
272 // If this is a GEP that SE doesn't know about, compute it now and insert it.
273 // If this is not a GEP, or if we have already done this computation, just let
275 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
276 if (!GEP || SE->hasSCEV(GEP))
277 return SE->getSCEV(Exp);
279 // Analyze all of the subscripts of this getelementptr instruction, looking
280 // for uses that are determined by the trip count of the loop. First, skip
281 // all operands the are not dependent on the IV.
283 // Build up the base expression. Insert an LLVM cast of the pointer to
285 SCEVHandle GEPVal = SE->getUnknown(
286 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
288 gep_type_iterator GTI = gep_type_begin(GEP);
290 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
291 // If this is a use of a recurrence that we can analyze, and it comes before
292 // Op does in the GEP operand list, we will handle this when we process this
294 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
295 const StructLayout *SL = TD->getStructLayout(STy);
296 unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
297 uint64_t Offset = SL->getElementOffset(Idx);
298 GEPVal = SE->getAddExpr(GEPVal,
299 SE->getIntegerSCEV(Offset, UIntPtrTy));
301 unsigned GEPOpiBits =
302 GEP->getOperand(i)->getType()->getPrimitiveSizeInBits();
303 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
304 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
305 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
306 Instruction::BitCast));
307 Value *OpVal = getCastedVersionOf(opcode, GEP->getOperand(i));
308 SCEVHandle Idx = SE->getSCEV(OpVal);
310 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
312 Idx = SE->getMulExpr(Idx,
313 SE->getConstant(ConstantInt::get(UIntPtrTy,
315 GEPVal = SE->getAddExpr(GEPVal, Idx);
319 SE->setSCEV(GEP, GEPVal);
323 /// getSCEVStartAndStride - Compute the start and stride of this expression,
324 /// returning false if the expression is not a start/stride pair, or true if it
325 /// is. The stride must be a loop invariant expression, but the start may be
326 /// a mix of loop invariant and loop variant expressions.
327 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
328 SCEVHandle &Start, SCEVHandle &Stride,
329 ScalarEvolution *SE) {
330 SCEVHandle TheAddRec = Start; // Initialize to zero.
332 // If the outer level is an AddExpr, the operands are all start values except
333 // for a nested AddRecExpr.
334 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
335 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
336 if (SCEVAddRecExpr *AddRec =
337 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
338 if (AddRec->getLoop() == L)
339 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
341 return false; // Nested IV of some sort?
343 Start = SE->getAddExpr(Start, AE->getOperand(i));
346 } else if (isa<SCEVAddRecExpr>(SH)) {
349 return false; // not analyzable.
352 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
353 if (!AddRec || AddRec->getLoop() != L) return false;
355 // FIXME: Generalize to non-affine IV's.
356 if (!AddRec->isAffine()) return false;
358 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
360 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
361 DOUT << "[" << L->getHeader()->getName()
362 << "] Variable stride: " << *AddRec << "\n";
364 Stride = AddRec->getOperand(1);
368 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
369 /// and now we need to decide whether the user should use the preinc or post-inc
370 /// value. If this user should use the post-inc version of the IV, return true.
372 /// Choosing wrong here can break dominance properties (if we choose to use the
373 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
374 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
375 /// should use the post-inc value).
376 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
377 Loop *L, DominatorTree *DT, Pass *P,
378 SmallPtrSet<Instruction*,16> &DeadInsts){
379 // If the user is in the loop, use the preinc value.
380 if (L->contains(User->getParent())) return false;
382 BasicBlock *LatchBlock = L->getLoopLatch();
384 // Ok, the user is outside of the loop. If it is dominated by the latch
385 // block, use the post-inc value.
386 if (DT->dominates(LatchBlock, User->getParent()))
389 // There is one case we have to be careful of: PHI nodes. These little guys
390 // can live in blocks that do not dominate the latch block, but (since their
391 // uses occur in the predecessor block, not the block the PHI lives in) should
392 // still use the post-inc value. Check for this case now.
393 PHINode *PN = dyn_cast<PHINode>(User);
394 if (!PN) return false; // not a phi, not dominated by latch block.
396 // Look at all of the uses of IV by the PHI node. If any use corresponds to
397 // a block that is not dominated by the latch block, give up and use the
398 // preincremented value.
399 unsigned NumUses = 0;
400 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
401 if (PN->getIncomingValue(i) == IV) {
403 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
407 // Okay, all uses of IV by PN are in predecessor blocks that really are
408 // dominated by the latch block. Split the critical edges and use the
409 // post-incremented value.
410 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
411 if (PN->getIncomingValue(i) == IV) {
412 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
413 // Splitting the critical edge can reduce the number of entries in this
415 e = PN->getNumIncomingValues();
416 if (--NumUses == 0) break;
419 // PHI node might have become a constant value after SplitCriticalEdge.
420 DeadInsts.insert(User);
427 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
428 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
429 /// return true. Otherwise, return false.
430 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
431 SmallPtrSet<Instruction*,16> &Processed) {
432 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
433 return false; // Void and FP expressions cannot be reduced.
434 if (!Processed.insert(I))
435 return true; // Instruction already handled.
437 // Get the symbolic expression for this instruction.
438 SCEVHandle ISE = GetExpressionSCEV(I);
439 if (isa<SCEVCouldNotCompute>(ISE)) return false;
441 // Get the start and stride for this expression.
442 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
443 SCEVHandle Stride = Start;
444 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
445 return false; // Non-reducible symbolic expression, bail out.
447 std::vector<Instruction *> IUsers;
448 // Collect all I uses now because IVUseShouldUsePostIncValue may
449 // invalidate use_iterator.
450 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
451 IUsers.push_back(cast<Instruction>(*UI));
453 for (unsigned iused_index = 0, iused_size = IUsers.size();
454 iused_index != iused_size; ++iused_index) {
456 Instruction *User = IUsers[iused_index];
458 // Do not infinitely recurse on PHI nodes.
459 if (isa<PHINode>(User) && Processed.count(User))
462 // If this is an instruction defined in a nested loop, or outside this loop,
463 // don't recurse into it.
464 bool AddUserToIVUsers = false;
465 if (LI->getLoopFor(User->getParent()) != L) {
466 DOUT << "FOUND USER in other loop: " << *User
467 << " OF SCEV: " << *ISE << "\n";
468 AddUserToIVUsers = true;
469 } else if (!AddUsersIfInteresting(User, L, Processed)) {
470 DOUT << "FOUND USER: " << *User
471 << " OF SCEV: " << *ISE << "\n";
472 AddUserToIVUsers = true;
475 if (AddUserToIVUsers) {
476 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
477 if (StrideUses.Users.empty()) // First occurance of this stride?
478 StrideOrder.push_back(Stride);
480 // Okay, we found a user that we cannot reduce. Analyze the instruction
481 // and decide what to do with it. If we are a use inside of the loop, use
482 // the value before incrementation, otherwise use it after incrementation.
483 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
484 // The value used will be incremented by the stride more than we are
485 // expecting, so subtract this off.
486 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
487 StrideUses.addUser(NewStart, User, I);
488 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
489 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
491 StrideUses.addUser(Start, User, I);
499 /// BasedUser - For a particular base value, keep information about how we've
500 /// partitioned the expression so far.
502 /// SE - The current ScalarEvolution object.
505 /// Base - The Base value for the PHI node that needs to be inserted for
506 /// this use. As the use is processed, information gets moved from this
507 /// field to the Imm field (below). BasedUser values are sorted by this
511 /// Inst - The instruction using the induction variable.
514 /// OperandValToReplace - The operand value of Inst to replace with the
516 Value *OperandValToReplace;
518 /// Imm - The immediate value that should be added to the base immediately
519 /// before Inst, because it will be folded into the imm field of the
523 /// EmittedBase - The actual value* to use for the base value of this
524 /// operation. This is null if we should just use zero so far.
527 // isUseOfPostIncrementedValue - True if this should use the
528 // post-incremented version of this IV, not the preincremented version.
529 // This can only be set in special cases, such as the terminating setcc
530 // instruction for a loop and uses outside the loop that are dominated by
532 bool isUseOfPostIncrementedValue;
534 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
535 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
536 OperandValToReplace(IVSU.OperandValToReplace),
537 Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
538 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
540 // Once we rewrite the code to insert the new IVs we want, update the
541 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
543 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
544 SCEVExpander &Rewriter, Loop *L, Pass *P,
545 SmallPtrSet<Instruction*,16> &DeadInsts);
547 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
548 SCEVExpander &Rewriter,
549 Instruction *IP, Loop *L);
554 void BasedUser::dump() const {
555 cerr << " Base=" << *Base;
556 cerr << " Imm=" << *Imm;
558 cerr << " EB=" << *EmittedBase;
560 cerr << " Inst: " << *Inst;
563 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
564 SCEVExpander &Rewriter,
565 Instruction *IP, Loop *L) {
566 // Figure out where we *really* want to insert this code. In particular, if
567 // the user is inside of a loop that is nested inside of L, we really don't
568 // want to insert this expression before the user, we'd rather pull it out as
569 // many loops as possible.
570 LoopInfo &LI = Rewriter.getLoopInfo();
571 Instruction *BaseInsertPt = IP;
573 // Figure out the most-nested loop that IP is in.
574 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
576 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
577 // the preheader of the outer-most loop where NewBase is not loop invariant.
578 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
579 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
580 InsertLoop = InsertLoop->getParentLoop();
583 // If there is no immediate value, skip the next part.
584 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
585 if (SC->getValue()->isZero())
586 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
588 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
590 // If we are inserting the base and imm values in the same block, make sure to
591 // adjust the IP position if insertion reused a result.
592 if (IP == BaseInsertPt)
593 IP = Rewriter.getInsertionPoint();
595 // Always emit the immediate (if non-zero) into the same block as the user.
596 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
597 return Rewriter.expandCodeFor(NewValSCEV, IP);
602 // Once we rewrite the code to insert the new IVs we want, update the
603 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
605 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
606 SCEVExpander &Rewriter, Loop *L, Pass *P,
607 SmallPtrSet<Instruction*,16> &DeadInsts) {
608 if (!isa<PHINode>(Inst)) {
609 // By default, insert code at the user instruction.
610 BasicBlock::iterator InsertPt = Inst;
612 // However, if the Operand is itself an instruction, the (potentially
613 // complex) inserted code may be shared by many users. Because of this, we
614 // want to emit code for the computation of the operand right before its old
615 // computation. This is usually safe, because we obviously used to use the
616 // computation when it was computed in its current block. However, in some
617 // cases (e.g. use of a post-incremented induction variable) the NewBase
618 // value will be pinned to live somewhere after the original computation.
619 // In this case, we have to back off.
620 if (!isUseOfPostIncrementedValue) {
621 if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) {
623 while (isa<PHINode>(InsertPt)) ++InsertPt;
626 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
627 // Adjust the type back to match the Inst. Note that we can't use InsertPt
628 // here because the SCEVExpander may have inserted the instructions after
629 // that point, in its efforts to avoid inserting redundant expressions.
630 if (isa<PointerType>(OperandValToReplace->getType())) {
631 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
633 OperandValToReplace->getType());
635 // Replace the use of the operand Value with the new Phi we just created.
636 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
637 DOUT << " CHANGED: IMM =" << *Imm;
638 DOUT << " \tNEWBASE =" << *NewBase;
639 DOUT << " \tInst = " << *Inst;
643 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
644 // expression into each operand block that uses it. Note that PHI nodes can
645 // have multiple entries for the same predecessor. We use a map to make sure
646 // that a PHI node only has a single Value* for each predecessor (which also
647 // prevents us from inserting duplicate code in some blocks).
648 DenseMap<BasicBlock*, Value*> InsertedCode;
649 PHINode *PN = cast<PHINode>(Inst);
650 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
651 if (PN->getIncomingValue(i) == OperandValToReplace) {
652 // If this is a critical edge, split the edge so that we do not insert the
653 // code on all predecessor/successor paths. We do this unless this is the
654 // canonical backedge for this loop, as this can make some inserted code
655 // be in an illegal position.
656 BasicBlock *PHIPred = PN->getIncomingBlock(i);
657 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
658 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
660 // First step, split the critical edge.
661 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
663 // Next step: move the basic block. In particular, if the PHI node
664 // is outside of the loop, and PredTI is in the loop, we want to
665 // move the block to be immediately before the PHI block, not
666 // immediately after PredTI.
667 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
668 BasicBlock *NewBB = PN->getIncomingBlock(i);
669 NewBB->moveBefore(PN->getParent());
672 // Splitting the edge can reduce the number of PHI entries we have.
673 e = PN->getNumIncomingValues();
676 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
678 // Insert the code into the end of the predecessor block.
679 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
680 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
682 // Adjust the type back to match the PHI. Note that we can't use
683 // InsertPt here because the SCEVExpander may have inserted its
684 // instructions after that point, in its efforts to avoid inserting
685 // redundant expressions.
686 if (isa<PointerType>(PN->getType())) {
687 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
693 // Replace the use of the operand Value with the new Phi we just created.
694 PN->setIncomingValue(i, Code);
699 // PHI node might have become a constant value after SplitCriticalEdge.
700 DeadInsts.insert(Inst);
702 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
706 /// isTargetConstant - Return true if the following can be referenced by the
707 /// immediate field of a target instruction.
708 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
709 const TargetLowering *TLI) {
710 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
711 int64_t VC = SC->getValue()->getSExtValue();
713 TargetLowering::AddrMode AM;
715 return TLI->isLegalAddressingMode(AM, UseTy);
717 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
718 return (VC > -(1 << 16) && VC < (1 << 16)-1);
722 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
723 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
724 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
725 Constant *Op0 = CE->getOperand(0);
726 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
727 TargetLowering::AddrMode AM;
729 return TLI->isLegalAddressingMode(AM, UseTy);
735 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
736 /// loop varying to the Imm operand.
737 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
738 Loop *L, ScalarEvolution *SE) {
739 if (Val->isLoopInvariant(L)) return; // Nothing to do.
741 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
742 std::vector<SCEVHandle> NewOps;
743 NewOps.reserve(SAE->getNumOperands());
745 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
746 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
747 // If this is a loop-variant expression, it must stay in the immediate
748 // field of the expression.
749 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
751 NewOps.push_back(SAE->getOperand(i));
755 Val = SE->getIntegerSCEV(0, Val->getType());
757 Val = SE->getAddExpr(NewOps);
758 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
759 // Try to pull immediates out of the start value of nested addrec's.
760 SCEVHandle Start = SARE->getStart();
761 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
763 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
765 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
767 // Otherwise, all of Val is variant, move the whole thing over.
768 Imm = SE->getAddExpr(Imm, Val);
769 Val = SE->getIntegerSCEV(0, Val->getType());
774 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
775 /// that can fit into the immediate field of instructions in the target.
776 /// Accumulate these immediate values into the Imm value.
777 static void MoveImmediateValues(const TargetLowering *TLI,
779 SCEVHandle &Val, SCEVHandle &Imm,
780 bool isAddress, Loop *L,
781 ScalarEvolution *SE) {
782 const Type *UseTy = User->getType();
783 if (StoreInst *SI = dyn_cast<StoreInst>(User))
784 UseTy = SI->getOperand(0)->getType();
786 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
787 std::vector<SCEVHandle> NewOps;
788 NewOps.reserve(SAE->getNumOperands());
790 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
791 SCEVHandle NewOp = SAE->getOperand(i);
792 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
794 if (!NewOp->isLoopInvariant(L)) {
795 // If this is a loop-variant expression, it must stay in the immediate
796 // field of the expression.
797 Imm = SE->getAddExpr(Imm, NewOp);
799 NewOps.push_back(NewOp);
804 Val = SE->getIntegerSCEV(0, Val->getType());
806 Val = SE->getAddExpr(NewOps);
808 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
809 // Try to pull immediates out of the start value of nested addrec's.
810 SCEVHandle Start = SARE->getStart();
811 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
813 if (Start != SARE->getStart()) {
814 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
816 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
819 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
820 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
821 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
822 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
824 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
825 SCEVHandle NewOp = SME->getOperand(1);
826 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
828 // If we extracted something out of the subexpressions, see if we can
830 if (NewOp != SME->getOperand(1)) {
831 // Scale SubImm up by "8". If the result is a target constant, we are
833 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
834 if (isTargetConstant(SubImm, UseTy, TLI)) {
835 // Accumulate the immediate.
836 Imm = SE->getAddExpr(Imm, SubImm);
838 // Update what is left of 'Val'.
839 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
846 // Loop-variant expressions must stay in the immediate field of the
848 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
849 !Val->isLoopInvariant(L)) {
850 Imm = SE->getAddExpr(Imm, Val);
851 Val = SE->getIntegerSCEV(0, Val->getType());
855 // Otherwise, no immediates to move.
859 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
860 /// added together. This is used to reassociate common addition subexprs
861 /// together for maximal sharing when rewriting bases.
862 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
864 ScalarEvolution *SE) {
865 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
866 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
867 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
868 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
869 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
870 if (SARE->getOperand(0) == Zero) {
871 SubExprs.push_back(Expr);
873 // Compute the addrec with zero as its base.
874 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
875 Ops[0] = Zero; // Start with zero base.
876 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
879 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
881 } else if (!isa<SCEVConstant>(Expr) ||
882 !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
884 SubExprs.push_back(Expr);
889 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
890 /// removing any common subexpressions from it. Anything truly common is
891 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
892 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
894 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
895 ScalarEvolution *SE) {
896 unsigned NumUses = Uses.size();
898 // Only one use? Use its base, regardless of what it is!
899 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
900 SCEVHandle Result = Zero;
902 std::swap(Result, Uses[0].Base);
906 // To find common subexpressions, count how many of Uses use each expression.
907 // If any subexpressions are used Uses.size() times, they are common.
908 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
910 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
911 // order we see them.
912 std::vector<SCEVHandle> UniqueSubExprs;
914 std::vector<SCEVHandle> SubExprs;
915 for (unsigned i = 0; i != NumUses; ++i) {
916 // If the base is zero (which is common), return zero now, there are no
918 if (Uses[i].Base == Zero) return Zero;
920 // Split the expression into subexprs.
921 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
922 // Add one to SubExpressionUseCounts for each subexpr present.
923 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
924 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
925 UniqueSubExprs.push_back(SubExprs[j]);
929 // Now that we know how many times each is used, build Result. Iterate over
930 // UniqueSubexprs so that we have a stable ordering.
931 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
932 std::map<SCEVHandle, unsigned>::iterator I =
933 SubExpressionUseCounts.find(UniqueSubExprs[i]);
934 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
935 if (I->second == NumUses) { // Found CSE!
936 Result = SE->getAddExpr(Result, I->first);
938 // Remove non-cse's from SubExpressionUseCounts.
939 SubExpressionUseCounts.erase(I);
943 // If we found no CSE's, return now.
944 if (Result == Zero) return Result;
946 // Otherwise, remove all of the CSE's we found from each of the base values.
947 for (unsigned i = 0; i != NumUses; ++i) {
948 // Split the expression into subexprs.
949 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
951 // Remove any common subexpressions.
952 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
953 if (SubExpressionUseCounts.count(SubExprs[j])) {
954 SubExprs.erase(SubExprs.begin()+j);
958 // Finally, the non-shared expressions together.
959 if (SubExprs.empty())
962 Uses[i].Base = SE->getAddExpr(SubExprs);
969 /// isZero - returns true if the scalar evolution expression is zero.
971 static bool isZero(const SCEVHandle &V) {
972 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
973 return SC->getValue()->isZero();
977 /// ValidStride - Check whether the given Scale is valid for all loads and
978 /// stores in UsersToProcess.
980 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
982 const std::vector<BasedUser>& UsersToProcess) {
983 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
984 // If this is a load or other access, pass the type of the access in.
985 const Type *AccessTy = Type::VoidTy;
986 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
987 AccessTy = SI->getOperand(0)->getType();
988 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
989 AccessTy = LI->getType();
991 TargetLowering::AddrMode AM;
992 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
993 AM.BaseOffs = SC->getValue()->getSExtValue();
994 AM.HasBaseReg = HasBaseReg || !isZero(UsersToProcess[i].Base);
997 // If load[imm+r*scale] is illegal, bail out.
998 if (TLI && !TLI->isLegalAddressingMode(AM, AccessTy))
1004 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1006 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1010 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1012 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1013 !(isa<PointerType>(Ty2) &&
1014 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1015 !(isa<PointerType>(Ty1) &&
1016 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1019 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1020 /// of a previous stride and it is a legal value for the target addressing
1021 /// mode scale component and optional base reg. This allows the users of
1022 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1023 /// reuse is possible.
1024 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1025 bool AllUsesAreAddresses,
1026 const SCEVHandle &Stride,
1027 IVExpr &IV, const Type *Ty,
1028 const std::vector<BasedUser>& UsersToProcess) {
1029 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1030 int64_t SInt = SC->getValue()->getSExtValue();
1031 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1033 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1034 IVsByStride.find(StrideOrder[NewStride]);
1035 if (SI == IVsByStride.end())
1037 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1038 if (SI->first != Stride &&
1039 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1041 int64_t Scale = SInt / SSInt;
1042 // Check that this stride is valid for all the types used for loads and
1043 // stores; if it can be used for some and not others, we might as well use
1044 // the original stride everywhere, since we have to create the IV for it
1045 // anyway. If the scale is 1, then we don't need to worry about folding
1048 (AllUsesAreAddresses &&
1049 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1050 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1051 IE = SI->second.IVs.end(); II != IE; ++II)
1052 // FIXME: Only handle base == 0 for now.
1053 // Only reuse previous IV if it would not require a type conversion.
1054 if (isZero(II->Base) &&
1055 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1064 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1065 /// returns true if Val's isUseOfPostIncrementedValue is true.
1066 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1067 return Val.isUseOfPostIncrementedValue;
1070 /// isNonConstantNegative - REturn true if the specified scev is negated, but
1072 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1073 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1074 if (!Mul) return false;
1076 // If there is a constant factor, it will be first.
1077 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1078 if (!SC) return false;
1080 // Return true if the value is negative, this matches things like (-42 * V).
1081 return SC->getValue()->getValue().isNegative();
1084 // CollectIVUsers - Transform our list of users and offsets to a bit more
1085 // complex table. In this new vector, each 'BasedUser' contains 'Base' the base
1086 // of the strided accessas well as the old information from Uses. We
1087 // progressively move information from the Base field to the Imm field, until
1088 // we eventually have the full access expression to rewrite the use.
1089 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1090 IVUsersOfOneStride &Uses,
1092 bool &AllUsesAreAddresses,
1093 std::vector<BasedUser> &UsersToProcess) {
1094 UsersToProcess.reserve(Uses.Users.size());
1095 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1096 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1098 // Move any loop invariant operands from the offset field to the immediate
1099 // field of the use, so that we don't try to use something before it is
1101 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1102 UsersToProcess.back().Imm, L, SE);
1103 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1104 "Base value is not loop invariant!");
1107 // We now have a whole bunch of uses of like-strided induction variables, but
1108 // they might all have different bases. We want to emit one PHI node for this
1109 // stride which we fold as many common expressions (between the IVs) into as
1110 // possible. Start by identifying the common expressions in the base values
1111 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1112 // "A+B"), emit it to the preheader, then remove the expression from the
1113 // UsersToProcess base values.
1114 SCEVHandle CommonExprs =
1115 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
1117 // Next, figure out what we can represent in the immediate fields of
1118 // instructions. If we can represent anything there, move it to the imm
1119 // fields of the BasedUsers. We do this so that it increases the commonality
1120 // of the remaining uses.
1121 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1122 // If the user is not in the current loop, this means it is using the exit
1123 // value of the IV. Do not put anything in the base, make sure it's all in
1124 // the immediate field to allow as much factoring as possible.
1125 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1126 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1127 UsersToProcess[i].Base);
1128 UsersToProcess[i].Base =
1129 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1132 // Addressing modes can be folded into loads and stores. Be careful that
1133 // the store is through the expression, not of the expression though.
1134 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
1135 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst)) {
1136 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1138 } else if (IntrinsicInst *II =
1139 dyn_cast<IntrinsicInst>(UsersToProcess[i].Inst)) {
1140 // Addressing modes can also be folded into prefetches and a variety
1142 switch (II->getIntrinsicID()) {
1144 case Intrinsic::prefetch:
1145 case Intrinsic::x86_sse2_loadu_dq:
1146 case Intrinsic::x86_sse2_loadu_pd:
1147 case Intrinsic::x86_sse_loadu_ps:
1148 case Intrinsic::x86_sse_storeu_ps:
1149 case Intrinsic::x86_sse2_storeu_pd:
1150 case Intrinsic::x86_sse2_storeu_dq:
1151 case Intrinsic::x86_sse2_storel_dq:
1152 if (II->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1155 case Intrinsic::x86_sse2_loadh_pd:
1156 case Intrinsic::x86_sse2_loadl_pd:
1157 if (II->getOperand(2) == UsersToProcess[i].OperandValToReplace)
1163 // If this use isn't an address, then not all uses are addresses.
1165 AllUsesAreAddresses = false;
1167 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1168 UsersToProcess[i].Imm, isAddress, L, SE);
1175 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1176 /// stride of IV. All of the users may have different starting values, and this
1177 /// may not be the only stride (we know it is if isOnlyStride is true).
1178 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1179 IVUsersOfOneStride &Uses,
1181 bool isOnlyStride) {
1182 // If all the users are moved to another stride, then there is nothing to do.
1183 if (Uses.Users.size() == 0)
1186 // Keep track if every use in UsersToProcess is an address. If they all are,
1187 // we may be able to rewrite the entire collection of them in terms of a
1188 // smaller-stride IV.
1189 bool AllUsesAreAddresses = true;
1191 // Transform our list of users and offsets to a bit more complex table. In
1192 // this new vector, each 'BasedUser' contains 'Base' the base of the
1193 // strided accessas well as the old information from Uses. We progressively
1194 // move information from the Base field to the Imm field, until we eventually
1195 // have the full access expression to rewrite the use.
1196 std::vector<BasedUser> UsersToProcess;
1197 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1200 // If we managed to find some expressions in common, we'll need to carry
1201 // their value in a register and add it in for each use. This will take up
1202 // a register operand, which potentially restricts what stride values are
1204 bool HaveCommonExprs = !isZero(CommonExprs);
1206 // If all uses are addresses, check if it is possible to reuse an IV with a
1207 // stride that is a factor of this stride. And that the multiple is a number
1208 // that can be encoded in the scale field of the target addressing mode. And
1209 // that we will have a valid instruction after this substition, including the
1210 // immediate field, if any.
1211 PHINode *NewPHI = NULL;
1213 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1214 SE->getIntegerSCEV(0, Type::Int32Ty),
1216 unsigned RewriteFactor = 0;
1217 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1218 Stride, ReuseIV, CommonExprs->getType(),
1220 if (RewriteFactor != 0) {
1221 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1222 << " and BASE " << *ReuseIV.Base << " :\n";
1223 NewPHI = ReuseIV.PHI;
1224 IncV = ReuseIV.IncV;
1227 const Type *ReplacedTy = CommonExprs->getType();
1229 // Now that we know what we need to do, insert the PHI node itself.
1231 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1232 << *Stride << " and BASE " << *CommonExprs << ": ";
1234 SCEVExpander Rewriter(*SE, *LI);
1235 SCEVExpander PreheaderRewriter(*SE, *LI);
1237 BasicBlock *Preheader = L->getLoopPreheader();
1238 Instruction *PreInsertPt = Preheader->getTerminator();
1239 Instruction *PhiInsertBefore = L->getHeader()->begin();
1241 BasicBlock *LatchBlock = L->getLoopLatch();
1244 // Emit the initial base value into the loop preheader.
1246 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1248 if (RewriteFactor == 0) {
1249 // Create a new Phi for this base, and stick it in the loop header.
1250 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1253 // Add common base to the new Phi node.
1254 NewPHI->addIncoming(CommonBaseV, Preheader);
1256 // If the stride is negative, insert a sub instead of an add for the
1258 bool isNegative = isNonConstantNegative(Stride);
1259 SCEVHandle IncAmount = Stride;
1261 IncAmount = SE->getNegativeSCEV(Stride);
1263 // Insert the stride into the preheader.
1264 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1265 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1267 // Emit the increment of the base value before the terminator of the loop
1268 // latch block, and add it to the Phi node.
1269 SCEVHandle IncExp = SE->getUnknown(StrideV);
1271 IncExp = SE->getNegativeSCEV(IncExp);
1272 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1274 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1275 IncV->setName(NewPHI->getName()+".inc");
1276 NewPHI->addIncoming(IncV, LatchBlock);
1278 // Remember this in case a later stride is multiple of this.
1279 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1281 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1283 Constant *C = dyn_cast<Constant>(CommonBaseV);
1285 (!C->isNullValue() &&
1286 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1287 // We want the common base emitted into the preheader! This is just
1288 // using cast as a copy so BitCast (no-op cast) is appropriate
1289 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1290 "commonbase", PreInsertPt);
1294 // We want to emit code for users inside the loop first. To do this, we
1295 // rearrange BasedUser so that the entries at the end have
1296 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1297 // vector (so we handle them first).
1298 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1299 PartitionByIsUseOfPostIncrementedValue);
1301 // Sort this by base, so that things with the same base are handled
1302 // together. By partitioning first and stable-sorting later, we are
1303 // guaranteed that within each base we will pop off users from within the
1304 // loop before users outside of the loop with a particular base.
1306 // We would like to use stable_sort here, but we can't. The problem is that
1307 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1308 // we don't have anything to do a '<' comparison on. Because we think the
1309 // number of uses is small, do a horrible bubble sort which just relies on
1311 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1312 // Get a base value.
1313 SCEVHandle Base = UsersToProcess[i].Base;
1315 // Compact everything with this base to be consequtive with this one.
1316 for (unsigned j = i+1; j != e; ++j) {
1317 if (UsersToProcess[j].Base == Base) {
1318 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1324 // Process all the users now. This outer loop handles all bases, the inner
1325 // loop handles all users of a particular base.
1326 while (!UsersToProcess.empty()) {
1327 SCEVHandle Base = UsersToProcess.back().Base;
1329 // Emit the code for Base into the preheader.
1330 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1332 DOUT << " INSERTING code for BASE = " << *Base << ":";
1333 if (BaseV->hasName())
1334 DOUT << " Result value name = %" << BaseV->getNameStr();
1337 // If BaseV is a constant other than 0, make sure that it gets inserted into
1338 // the preheader, instead of being forward substituted into the uses. We do
1339 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1341 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1342 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1343 // We want this constant emitted into the preheader! This is just
1344 // using cast as a copy so BitCast (no-op cast) is appropriate
1345 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1350 // Emit the code to add the immediate offset to the Phi value, just before
1351 // the instructions that we identified as using this stride and base.
1353 // FIXME: Use emitted users to emit other users.
1354 BasedUser &User = UsersToProcess.back();
1356 // If this instruction wants to use the post-incremented value, move it
1357 // after the post-inc and use its value instead of the PHI.
1358 Value *RewriteOp = NewPHI;
1359 if (User.isUseOfPostIncrementedValue) {
1362 // If this user is in the loop, make sure it is the last thing in the
1363 // loop to ensure it is dominated by the increment.
1364 if (L->contains(User.Inst->getParent()))
1365 User.Inst->moveBefore(LatchBlock->getTerminator());
1367 if (RewriteOp->getType() != ReplacedTy) {
1368 Instruction::CastOps opcode = Instruction::Trunc;
1369 if (ReplacedTy->getPrimitiveSizeInBits() ==
1370 RewriteOp->getType()->getPrimitiveSizeInBits())
1371 opcode = Instruction::BitCast;
1372 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1375 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1377 // Clear the SCEVExpander's expression map so that we are guaranteed
1378 // to have the code emitted where we expect it.
1381 // If we are reusing the iv, then it must be multiplied by a constant
1382 // factor take advantage of addressing mode scale component.
1383 if (RewriteFactor != 0) {
1384 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1385 RewriteExpr->getType()),
1388 // The common base is emitted in the loop preheader. But since we
1389 // are reusing an IV, it has not been used to initialize the PHI node.
1390 // Add it to the expression used to rewrite the uses.
1391 if (!isa<ConstantInt>(CommonBaseV) ||
1392 !cast<ConstantInt>(CommonBaseV)->isZero())
1393 RewriteExpr = SE->getAddExpr(RewriteExpr,
1394 SE->getUnknown(CommonBaseV));
1397 // Now that we know what we need to do, insert code before User for the
1398 // immediate and any loop-variant expressions.
1399 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1400 // Add BaseV to the PHI value if needed.
1401 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1403 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this,
1406 // Mark old value we replaced as possibly dead, so that it is elminated
1407 // if we just replaced the last use of that value.
1408 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1410 UsersToProcess.pop_back();
1413 // If there are any more users to process with the same base, process them
1414 // now. We sorted by base above, so we just have to check the last elt.
1415 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1416 // TODO: Next, find out which base index is the most common, pull it out.
1419 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1420 // different starting values, into different PHIs.
1423 /// FindIVForUser - If Cond has an operand that is an expression of an IV,
1424 /// set the IV user and stride information and return true, otherwise return
1426 bool LoopStrengthReduce::FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
1427 const SCEVHandle *&CondStride) {
1428 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1430 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1431 IVUsesByStride.find(StrideOrder[Stride]);
1432 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1434 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1435 E = SI->second.Users.end(); UI != E; ++UI)
1436 if (UI->User == Cond) {
1437 // NOTE: we could handle setcc instructions with multiple uses here, but
1438 // InstCombine does it as well for simple uses, it's not clear that it
1439 // occurs enough in real life to handle.
1441 CondStride = &SI->first;
1449 // Constant strides come first which in turns are sorted by their absolute
1450 // values. If absolute values are the same, then positive strides comes first.
1452 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1453 struct StrideCompare {
1454 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1455 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1456 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1458 int64_t LV = LHSC->getValue()->getSExtValue();
1459 int64_t RV = RHSC->getValue()->getSExtValue();
1460 uint64_t ALV = (LV < 0) ? -LV : LV;
1461 uint64_t ARV = (RV < 0) ? -RV : RV;
1467 return (LHSC && !RHSC);
1472 /// ChangeCompareStride - If a loop termination compare instruction is the
1473 /// only use of its stride, and the compaison is against a constant value,
1474 /// try eliminate the stride by moving the compare instruction to another
1475 /// stride and change its constant operand accordingly. e.g.
1481 /// if (v2 < 10) goto loop
1486 /// if (v1 < 30) goto loop
1487 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1488 IVStrideUse* &CondUse,
1489 const SCEVHandle* &CondStride) {
1490 if (StrideOrder.size() < 2 ||
1491 IVUsesByStride[*CondStride].Users.size() != 1)
1493 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1494 if (!SC) return Cond;
1495 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1496 if (!C) return Cond;
1498 ICmpInst::Predicate Predicate = Cond->getPredicate();
1499 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1500 int64_t CmpVal = C->getValue().getSExtValue();
1501 unsigned BitWidth = C->getValue().getBitWidth();
1502 uint64_t SignBit = 1ULL << (BitWidth-1);
1503 const Type *CmpTy = C->getType();
1504 const Type *NewCmpTy = NULL;
1505 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1506 unsigned NewTyBits = 0;
1507 int64_t NewCmpVal = CmpVal;
1508 SCEVHandle *NewStride = NULL;
1509 Value *NewIncV = NULL;
1512 // Look for a suitable stride / iv as replacement.
1513 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1514 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1515 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1516 IVUsesByStride.find(StrideOrder[i]);
1517 if (!isa<SCEVConstant>(SI->first))
1519 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1520 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1523 Scale = SSInt / CmpSSInt;
1524 NewCmpVal = CmpVal * Scale;
1525 APInt Mul = APInt(BitWidth, NewCmpVal);
1526 // Check for overflow.
1527 if (Mul.getSExtValue() != NewCmpVal) {
1532 // Watch out for overflow.
1533 if (ICmpInst::isSignedPredicate(Predicate) &&
1534 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1537 if (NewCmpVal != CmpVal) {
1538 // Pick the best iv to use trying to avoid a cast.
1540 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1541 E = SI->second.Users.end(); UI != E; ++UI) {
1542 NewIncV = UI->OperandValToReplace;
1543 if (NewIncV->getType() == CmpTy)
1551 NewCmpTy = NewIncV->getType();
1552 NewTyBits = isa<PointerType>(NewCmpTy)
1553 ? UIntPtrTy->getPrimitiveSizeInBits()
1554 : NewCmpTy->getPrimitiveSizeInBits();
1555 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1556 // Check if it is possible to rewrite it using a iv / stride of a smaller
1558 bool TruncOk = false;
1559 if (NewCmpTy->isInteger()) {
1560 unsigned Bits = NewTyBits;
1561 if (ICmpInst::isSignedPredicate(Predicate))
1563 uint64_t Mask = (1ULL << Bits) - 1;
1564 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1573 // Don't rewrite if use offset is non-constant and the new type is
1574 // of a different type.
1575 // FIXME: too conservative?
1576 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1581 bool AllUsesAreAddresses = true;
1582 std::vector<BasedUser> UsersToProcess;
1583 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1584 AllUsesAreAddresses,
1586 // Avoid rewriting the compare instruction with an iv of new stride
1587 // if it's likely the new stride uses will be rewritten using the
1588 if (AllUsesAreAddresses &&
1589 ValidStride(!isZero(CommonExprs), Scale, UsersToProcess)) {
1594 // If scale is negative, use inverse predicate unless it's testing
1596 if (Scale < 0 && !Cond->isEquality())
1597 Predicate = ICmpInst::getInversePredicate(Predicate);
1599 NewStride = &StrideOrder[i];
1604 if (NewCmpVal != CmpVal) {
1605 // Create a new compare instruction using new stride / iv.
1606 ICmpInst *OldCond = Cond;
1608 if (!isa<PointerType>(NewCmpTy))
1609 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1611 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1612 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1614 // Insert new compare instruction.
1615 Cond = new ICmpInst(Predicate, NewIncV, RHS);
1616 Cond->setName(L->getHeader()->getName() + ".termcond");
1617 OldCond->getParent()->getInstList().insert(OldCond, Cond);
1619 // Remove the old compare instruction. The old indvar is probably dead too.
1620 DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
1621 OldCond->replaceAllUsesWith(Cond);
1622 SE->deleteValueFromRecords(OldCond);
1623 OldCond->eraseFromParent();
1625 IVUsesByStride[*CondStride].Users.pop_back();
1626 SCEVHandle NewOffset = TyBits == NewTyBits
1627 ? SE->getMulExpr(CondUse->Offset,
1628 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1629 : SE->getConstant(ConstantInt::get(NewCmpTy,
1630 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1631 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1632 CondUse = &IVUsesByStride[*NewStride].Users.back();
1633 CondStride = NewStride;
1640 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1641 // uses in the loop, look to see if we can eliminate some, in favor of using
1642 // common indvars for the different uses.
1643 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1644 // TODO: implement optzns here.
1646 // Finally, get the terminating condition for the loop if possible. If we
1647 // can, we want to change it to use a post-incremented version of its
1648 // induction variable, to allow coalescing the live ranges for the IV into
1649 // one register value.
1650 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1651 BasicBlock *Preheader = L->getLoopPreheader();
1652 BasicBlock *LatchBlock =
1653 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1654 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1655 if (!TermBr || TermBr->isUnconditional() ||
1656 !isa<ICmpInst>(TermBr->getCondition()))
1658 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1660 // Search IVUsesByStride to find Cond's IVUse if there is one.
1661 IVStrideUse *CondUse = 0;
1662 const SCEVHandle *CondStride = 0;
1664 if (!FindIVForUser(Cond, CondUse, CondStride))
1665 return; // setcc doesn't use the IV.
1667 // If possible, change stride and operands of the compare instruction to
1668 // eliminate one stride.
1669 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
1671 // It's possible for the setcc instruction to be anywhere in the loop, and
1672 // possible for it to have multiple users. If it is not immediately before
1673 // the latch block branch, move it.
1674 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1675 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1676 Cond->moveBefore(TermBr);
1678 // Otherwise, clone the terminating condition and insert into the loopend.
1679 Cond = cast<ICmpInst>(Cond->clone());
1680 Cond->setName(L->getHeader()->getName() + ".termcond");
1681 LatchBlock->getInstList().insert(TermBr, Cond);
1683 // Clone the IVUse, as the old use still exists!
1684 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1685 CondUse->OperandValToReplace);
1686 CondUse = &IVUsesByStride[*CondStride].Users.back();
1690 // If we get to here, we know that we can transform the setcc instruction to
1691 // use the post-incremented version of the IV, allowing us to coalesce the
1692 // live ranges for the IV correctly.
1693 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
1694 CondUse->isUseOfPostIncrementedValue = true;
1697 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1699 LI = &getAnalysis<LoopInfo>();
1700 DT = &getAnalysis<DominatorTree>();
1701 SE = &getAnalysis<ScalarEvolution>();
1702 TD = &getAnalysis<TargetData>();
1703 UIntPtrTy = TD->getIntPtrType();
1705 // Find all uses of induction variables in this loop, and catagorize
1706 // them by stride. Start by finding all of the PHI nodes in the header for
1707 // this loop. If they are induction variables, inspect their uses.
1708 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
1709 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1710 AddUsersIfInteresting(I, L, Processed);
1712 // If we have nothing to do, return.
1713 if (IVUsesByStride.empty()) return false;
1715 // Optimize induction variables. Some indvar uses can be transformed to use
1716 // strides that will be needed for other purposes. A common example of this
1717 // is the exit test for the loop, which can often be rewritten to use the
1718 // computation of some other indvar to decide when to terminate the loop.
1722 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1723 // doing computation in byte values, promote to 32-bit values if safe.
1725 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1726 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1727 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1728 // to be careful that IV's are all the same type. Only works for intptr_t
1731 // If we only have one stride, we can more aggressively eliminate some things.
1732 bool HasOneStride = IVUsesByStride.size() == 1;
1735 DOUT << "\nLSR on ";
1739 // IVsByStride keeps IVs for one particular loop.
1740 IVsByStride.clear();
1742 // Sort the StrideOrder so we process larger strides first.
1743 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1745 // Note: this processes each stride/type pair individually. All users passed
1746 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1747 // note that we iterate over IVUsesByStride indirectly by using StrideOrder.
1748 // This extra layer of indirection makes the ordering of strides deterministic
1749 // - not dependent on map order.
1750 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1751 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1752 IVUsesByStride.find(StrideOrder[Stride]);
1753 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1754 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1757 // Clean up after ourselves
1758 if (!DeadInsts.empty()) {
1759 DeleteTriviallyDeadInstructions(DeadInsts);
1761 BasicBlock::iterator I = L->getHeader()->begin();
1763 while ((PN = dyn_cast<PHINode>(I))) {
1764 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1766 // At this point, we know that we have killed one or more GEP
1767 // instructions. It is worth checking to see if the cann indvar is also
1768 // dead, so that we can remove it as well. The requirements for the cann
1769 // indvar to be considered dead are:
1770 // 1. the cann indvar has one use
1771 // 2. the use is an add instruction
1772 // 3. the add has one use
1773 // 4. the add is used by the cann indvar
1774 // If all four cases above are true, then we can remove both the add and
1776 // FIXME: this needs to eliminate an induction variable even if it's being
1777 // compared against some value to decide loop termination.
1778 if (PN->hasOneUse()) {
1779 Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
1780 if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
1781 if (BO->hasOneUse() && PN == *(BO->use_begin())) {
1782 DeadInsts.insert(BO);
1783 // Break the cycle, then delete the PHI.
1784 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1785 SE->deleteValueFromRecords(PN);
1786 PN->eraseFromParent();
1791 DeleteTriviallyDeadInstructions(DeadInsts);
1794 CastedPointers.clear();
1795 IVUsesByStride.clear();
1796 StrideOrder.clear();