1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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/SetVector.h"
35 #include "llvm/ADT/SmallPtrSet.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/Compiler.h"
39 #include "llvm/Target/TargetLowering.h"
44 STATISTIC(NumReduced , "Number of GEPs strength reduced");
45 STATISTIC(NumInserted, "Number of PHIs inserted");
46 STATISTIC(NumVariable, "Number of PHIs with variable strides");
47 STATISTIC(NumEliminated , "Number of strides eliminated");
53 /// IVStrideUse - Keep track of one use of a strided induction variable, where
54 /// the stride is stored externally. The Offset member keeps track of the
55 /// offset from the IV, User is the actual user of the operand, and
56 /// 'OperandValToReplace' is the operand of the User that is the use.
57 struct VISIBILITY_HIDDEN IVStrideUse {
60 Value *OperandValToReplace;
62 // isUseOfPostIncrementedValue - True if this should use the
63 // post-incremented version of this IV, not the preincremented version.
64 // This can only be set in special cases, such as the terminating setcc
65 // instruction for a loop or uses dominated by the loop.
66 bool isUseOfPostIncrementedValue;
68 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
69 : Offset(Offs), User(U), OperandValToReplace(O),
70 isUseOfPostIncrementedValue(false) {}
73 /// IVUsersOfOneStride - This structure keeps track of all instructions that
74 /// have an operand that is based on the trip count multiplied by some stride.
75 /// The stride for all of these users is common and kept external to this
77 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
78 /// Users - Keep track of all of the users of this stride as well as the
79 /// initial value and the operand that uses the IV.
80 std::vector<IVStrideUse> Users;
82 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
83 Users.push_back(IVStrideUse(Offset, User, Operand));
87 /// IVInfo - This structure keeps track of one IV expression inserted during
88 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
89 /// well as the PHI node and increment value created for rewrite.
90 struct VISIBILITY_HIDDEN IVExpr {
96 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
98 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
101 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
102 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
103 struct VISIBILITY_HIDDEN IVsOfOneStride {
104 std::vector<IVExpr> IVs;
106 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
108 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
112 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
116 const TargetData *TD;
117 const Type *UIntPtrTy;
120 /// IVUsesByStride - Keep track of all uses of induction variables that we
121 /// are interested in. The key of the map is the stride of the access.
122 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
124 /// IVsByStride - Keep track of all IVs that have been inserted for a
125 /// particular stride.
126 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
128 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
129 /// We use this to iterate over the IVUsesByStride collection without being
130 /// dependent on random ordering of pointers in the process.
131 SmallVector<SCEVHandle, 16> StrideOrder;
133 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
134 /// of the casted version of each value. This is accessed by
135 /// getCastedVersionOf.
136 DenseMap<Value*, Value*> CastedPointers;
138 /// DeadInsts - Keep track of instructions we may have made dead, so that
139 /// we can remove them after we are done working.
140 SetVector<Instruction*> DeadInsts;
142 /// TLI - Keep a pointer of a TargetLowering to consult for determining
143 /// transformation profitability.
144 const TargetLowering *TLI;
147 static char ID; // Pass ID, replacement for typeid
148 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
149 LoopPass((intptr_t)&ID), TLI(tli) {
152 bool runOnLoop(Loop *L, LPPassManager &LPM);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 // We split critical edges, so we change the CFG. However, we do update
156 // many analyses if they are around.
157 AU.addPreservedID(LoopSimplifyID);
158 AU.addPreserved<LoopInfo>();
159 AU.addPreserved<DominanceFrontier>();
160 AU.addPreserved<DominatorTree>();
162 AU.addRequiredID(LoopSimplifyID);
163 AU.addRequired<LoopInfo>();
164 AU.addRequired<DominatorTree>();
165 AU.addRequired<TargetData>();
166 AU.addRequired<ScalarEvolution>();
169 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
171 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
173 bool AddUsersIfInteresting(Instruction *I, Loop *L,
174 SmallPtrSet<Instruction*,16> &Processed);
175 SCEVHandle GetExpressionSCEV(Instruction *E);
176 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
177 IVStrideUse* &CondUse,
178 const SCEVHandle* &CondStride);
179 void OptimizeIndvars(Loop *L);
180 bool FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
181 const SCEVHandle *&CondStride);
182 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
183 unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
184 IVExpr&, const Type*,
185 const std::vector<BasedUser>& UsersToProcess);
186 bool ValidStride(bool, int64_t,
187 const std::vector<BasedUser>& UsersToProcess);
188 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
189 IVUsersOfOneStride &Uses,
191 bool &AllUsesAreAddresses,
192 std::vector<BasedUser> &UsersToProcess);
193 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
194 IVUsersOfOneStride &Uses,
195 Loop *L, bool isOnlyStride);
196 void DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts);
200 char LoopStrengthReduce::ID = 0;
201 static RegisterPass<LoopStrengthReduce>
202 X("loop-reduce", "Loop Strength Reduction");
204 LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
205 return new LoopStrengthReduce(TLI);
208 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
209 /// assumes that the Value* V is of integer or pointer type only.
211 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
213 if (V->getType() == UIntPtrTy) return V;
214 if (Constant *CB = dyn_cast<Constant>(V))
215 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
217 Value *&New = CastedPointers[V];
220 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
221 DeadInsts.insert(cast<Instruction>(New));
226 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
227 /// specified set are trivially dead, delete them and see if this makes any of
228 /// their operands subsequently dead.
229 void LoopStrengthReduce::
230 DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts) {
231 while (!Insts.empty()) {
232 Instruction *I = Insts.back();
235 if (PHINode *PN = dyn_cast<PHINode>(I)) {
236 // If all incoming values to the Phi are the same, we can replace the Phi
238 if (Value *PNV = PN->hasConstantValue()) {
239 if (Instruction *U = dyn_cast<Instruction>(PNV))
241 SE->deleteValueFromRecords(PN);
242 PN->replaceAllUsesWith(PNV);
243 PN->eraseFromParent();
249 if (isInstructionTriviallyDead(I)) {
250 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
251 if (Instruction *U = dyn_cast<Instruction>(*i))
253 SE->deleteValueFromRecords(I);
254 I->eraseFromParent();
261 /// GetExpressionSCEV - Compute and return the SCEV for the specified
263 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
264 // Pointer to pointer bitcast instructions return the same value as their
266 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
267 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
268 return SE->getSCEV(BCI);
269 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
274 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
275 // If this is a GEP that SE doesn't know about, compute it now and insert it.
276 // If this is not a GEP, or if we have already done this computation, just let
278 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
279 if (!GEP || SE->hasSCEV(GEP))
280 return SE->getSCEV(Exp);
282 // Analyze all of the subscripts of this getelementptr instruction, looking
283 // for uses that are determined by the trip count of the loop. First, skip
284 // all operands the are not dependent on the IV.
286 // Build up the base expression. Insert an LLVM cast of the pointer to
288 SCEVHandle GEPVal = SE->getUnknown(
289 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
291 gep_type_iterator GTI = gep_type_begin(GEP);
293 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
294 i != e; ++i, ++GTI) {
295 // If this is a use of a recurrence that we can analyze, and it comes before
296 // Op does in the GEP operand list, we will handle this when we process this
298 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
299 const StructLayout *SL = TD->getStructLayout(STy);
300 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
301 uint64_t Offset = SL->getElementOffset(Idx);
302 GEPVal = SE->getAddExpr(GEPVal,
303 SE->getIntegerSCEV(Offset, UIntPtrTy));
305 unsigned GEPOpiBits =
306 (*i)->getType()->getPrimitiveSizeInBits();
307 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
308 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
309 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
310 Instruction::BitCast));
311 Value *OpVal = getCastedVersionOf(opcode, *i);
312 SCEVHandle Idx = SE->getSCEV(OpVal);
314 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
316 Idx = SE->getMulExpr(Idx,
317 SE->getConstant(ConstantInt::get(UIntPtrTy,
319 GEPVal = SE->getAddExpr(GEPVal, Idx);
323 SE->setSCEV(GEP, GEPVal);
327 /// getSCEVStartAndStride - Compute the start and stride of this expression,
328 /// returning false if the expression is not a start/stride pair, or true if it
329 /// is. The stride must be a loop invariant expression, but the start may be
330 /// a mix of loop invariant and loop variant expressions.
331 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
332 SCEVHandle &Start, SCEVHandle &Stride,
333 ScalarEvolution *SE) {
334 SCEVHandle TheAddRec = Start; // Initialize to zero.
336 // If the outer level is an AddExpr, the operands are all start values except
337 // for a nested AddRecExpr.
338 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
339 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
340 if (SCEVAddRecExpr *AddRec =
341 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
342 if (AddRec->getLoop() == L)
343 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
345 return false; // Nested IV of some sort?
347 Start = SE->getAddExpr(Start, AE->getOperand(i));
350 } else if (isa<SCEVAddRecExpr>(SH)) {
353 return false; // not analyzable.
356 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
357 if (!AddRec || AddRec->getLoop() != L) return false;
359 // FIXME: Generalize to non-affine IV's.
360 if (!AddRec->isAffine()) return false;
362 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
364 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
365 DOUT << "[" << L->getHeader()->getName()
366 << "] Variable stride: " << *AddRec << "\n";
368 Stride = AddRec->getOperand(1);
372 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
373 /// and now we need to decide whether the user should use the preinc or post-inc
374 /// value. If this user should use the post-inc version of the IV, return true.
376 /// Choosing wrong here can break dominance properties (if we choose to use the
377 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
378 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
379 /// should use the post-inc value).
380 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
381 Loop *L, DominatorTree *DT, Pass *P,
382 SetVector<Instruction*> &DeadInsts){
383 // If the user is in the loop, use the preinc value.
384 if (L->contains(User->getParent())) return false;
386 BasicBlock *LatchBlock = L->getLoopLatch();
388 // Ok, the user is outside of the loop. If it is dominated by the latch
389 // block, use the post-inc value.
390 if (DT->dominates(LatchBlock, User->getParent()))
393 // There is one case we have to be careful of: PHI nodes. These little guys
394 // can live in blocks that do not dominate the latch block, but (since their
395 // uses occur in the predecessor block, not the block the PHI lives in) should
396 // still use the post-inc value. Check for this case now.
397 PHINode *PN = dyn_cast<PHINode>(User);
398 if (!PN) return false; // not a phi, not dominated by latch block.
400 // Look at all of the uses of IV by the PHI node. If any use corresponds to
401 // a block that is not dominated by the latch block, give up and use the
402 // preincremented value.
403 unsigned NumUses = 0;
404 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
405 if (PN->getIncomingValue(i) == IV) {
407 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
411 // Okay, all uses of IV by PN are in predecessor blocks that really are
412 // dominated by the latch block. Split the critical edges and use the
413 // post-incremented value.
414 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
415 if (PN->getIncomingValue(i) == IV) {
416 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
417 // Splitting the critical edge can reduce the number of entries in this
419 e = PN->getNumIncomingValues();
420 if (--NumUses == 0) break;
423 // PHI node might have become a constant value after SplitCriticalEdge.
424 DeadInsts.insert(User);
431 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
432 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
433 /// return true. Otherwise, return false.
434 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
435 SmallPtrSet<Instruction*,16> &Processed) {
436 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
437 return false; // Void and FP expressions cannot be reduced.
438 if (!Processed.insert(I))
439 return true; // Instruction already handled.
441 // Get the symbolic expression for this instruction.
442 SCEVHandle ISE = GetExpressionSCEV(I);
443 if (isa<SCEVCouldNotCompute>(ISE)) return false;
445 // Get the start and stride for this expression.
446 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
447 SCEVHandle Stride = Start;
448 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
449 return false; // Non-reducible symbolic expression, bail out.
451 std::vector<Instruction *> IUsers;
452 // Collect all I uses now because IVUseShouldUsePostIncValue may
453 // invalidate use_iterator.
454 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
455 IUsers.push_back(cast<Instruction>(*UI));
457 for (unsigned iused_index = 0, iused_size = IUsers.size();
458 iused_index != iused_size; ++iused_index) {
460 Instruction *User = IUsers[iused_index];
462 // Do not infinitely recurse on PHI nodes.
463 if (isa<PHINode>(User) && Processed.count(User))
466 // If this is an instruction defined in a nested loop, or outside this loop,
467 // don't recurse into it.
468 bool AddUserToIVUsers = false;
469 if (LI->getLoopFor(User->getParent()) != L) {
470 DOUT << "FOUND USER in other loop: " << *User
471 << " OF SCEV: " << *ISE << "\n";
472 AddUserToIVUsers = true;
473 } else if (!AddUsersIfInteresting(User, L, Processed)) {
474 DOUT << "FOUND USER: " << *User
475 << " OF SCEV: " << *ISE << "\n";
476 AddUserToIVUsers = true;
479 if (AddUserToIVUsers) {
480 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
481 if (StrideUses.Users.empty()) // First occurance of this stride?
482 StrideOrder.push_back(Stride);
484 // Okay, we found a user that we cannot reduce. Analyze the instruction
485 // and decide what to do with it. If we are a use inside of the loop, use
486 // the value before incrementation, otherwise use it after incrementation.
487 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
488 // The value used will be incremented by the stride more than we are
489 // expecting, so subtract this off.
490 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
491 StrideUses.addUser(NewStart, User, I);
492 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
493 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
495 StrideUses.addUser(Start, User, I);
503 /// BasedUser - For a particular base value, keep information about how we've
504 /// partitioned the expression so far.
506 /// SE - The current ScalarEvolution object.
509 /// Base - The Base value for the PHI node that needs to be inserted for
510 /// this use. As the use is processed, information gets moved from this
511 /// field to the Imm field (below). BasedUser values are sorted by this
515 /// Inst - The instruction using the induction variable.
518 /// OperandValToReplace - The operand value of Inst to replace with the
520 Value *OperandValToReplace;
522 /// Imm - The immediate value that should be added to the base immediately
523 /// before Inst, because it will be folded into the imm field of the
527 /// EmittedBase - The actual value* to use for the base value of this
528 /// operation. This is null if we should just use zero so far.
531 // isUseOfPostIncrementedValue - True if this should use the
532 // post-incremented version of this IV, not the preincremented version.
533 // This can only be set in special cases, such as the terminating setcc
534 // instruction for a loop and uses outside the loop that are dominated by
536 bool isUseOfPostIncrementedValue;
538 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
539 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
540 OperandValToReplace(IVSU.OperandValToReplace),
541 Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
542 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
544 // Once we rewrite the code to insert the new IVs we want, update the
545 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
547 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
548 Instruction *InsertPt,
549 SCEVExpander &Rewriter, Loop *L, Pass *P,
550 SetVector<Instruction*> &DeadInsts);
552 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
553 SCEVExpander &Rewriter,
554 Instruction *IP, Loop *L);
559 void BasedUser::dump() const {
560 cerr << " Base=" << *Base;
561 cerr << " Imm=" << *Imm;
563 cerr << " EB=" << *EmittedBase;
565 cerr << " Inst: " << *Inst;
568 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
569 SCEVExpander &Rewriter,
570 Instruction *IP, Loop *L) {
571 // Figure out where we *really* want to insert this code. In particular, if
572 // the user is inside of a loop that is nested inside of L, we really don't
573 // want to insert this expression before the user, we'd rather pull it out as
574 // many loops as possible.
575 LoopInfo &LI = Rewriter.getLoopInfo();
576 Instruction *BaseInsertPt = IP;
578 // Figure out the most-nested loop that IP is in.
579 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
581 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
582 // the preheader of the outer-most loop where NewBase is not loop invariant.
583 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
584 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
585 InsertLoop = InsertLoop->getParentLoop();
588 // If there is no immediate value, skip the next part.
589 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
590 if (SC->getValue()->isZero())
591 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
593 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
595 // If we are inserting the base and imm values in the same block, make sure to
596 // adjust the IP position if insertion reused a result.
597 if (IP == BaseInsertPt)
598 IP = Rewriter.getInsertionPoint();
600 // Always emit the immediate (if non-zero) into the same block as the user.
601 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
602 return Rewriter.expandCodeFor(NewValSCEV, IP);
607 // Once we rewrite the code to insert the new IVs we want, update the
608 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
609 // to it. NewBasePt is the last instruction which contributes to the
610 // value of NewBase in the case that it's a diffferent instruction from
611 // the PHI that NewBase is computed from, or null otherwise.
613 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
614 Instruction *NewBasePt,
615 SCEVExpander &Rewriter, Loop *L, Pass *P,
616 SetVector<Instruction*> &DeadInsts) {
617 if (!isa<PHINode>(Inst)) {
618 // By default, insert code at the user instruction.
619 BasicBlock::iterator InsertPt = Inst;
621 // However, if the Operand is itself an instruction, the (potentially
622 // complex) inserted code may be shared by many users. Because of this, we
623 // want to emit code for the computation of the operand right before its old
624 // computation. This is usually safe, because we obviously used to use the
625 // computation when it was computed in its current block. However, in some
626 // cases (e.g. use of a post-incremented induction variable) the NewBase
627 // value will be pinned to live somewhere after the original computation.
628 // In this case, we have to back off.
629 if (!isUseOfPostIncrementedValue) {
630 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
631 InsertPt = NewBasePt;
633 } else if (Instruction *OpInst
634 = dyn_cast<Instruction>(OperandValToReplace)) {
636 while (isa<PHINode>(InsertPt)) ++InsertPt;
639 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
640 // Adjust the type back to match the Inst. Note that we can't use InsertPt
641 // here because the SCEVExpander may have inserted the instructions after
642 // that point, in its efforts to avoid inserting redundant expressions.
643 if (isa<PointerType>(OperandValToReplace->getType())) {
644 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
646 OperandValToReplace->getType());
648 // Replace the use of the operand Value with the new Phi we just created.
649 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
650 DOUT << " CHANGED: IMM =" << *Imm;
651 DOUT << " \tNEWBASE =" << *NewBase;
652 DOUT << " \tInst = " << *Inst;
656 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
657 // expression into each operand block that uses it. Note that PHI nodes can
658 // have multiple entries for the same predecessor. We use a map to make sure
659 // that a PHI node only has a single Value* for each predecessor (which also
660 // prevents us from inserting duplicate code in some blocks).
661 DenseMap<BasicBlock*, Value*> InsertedCode;
662 PHINode *PN = cast<PHINode>(Inst);
663 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
664 if (PN->getIncomingValue(i) == OperandValToReplace) {
665 // If this is a critical edge, split the edge so that we do not insert the
666 // code on all predecessor/successor paths. We do this unless this is the
667 // canonical backedge for this loop, as this can make some inserted code
668 // be in an illegal position.
669 BasicBlock *PHIPred = PN->getIncomingBlock(i);
670 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
671 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
673 // First step, split the critical edge.
674 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
676 // Next step: move the basic block. In particular, if the PHI node
677 // is outside of the loop, and PredTI is in the loop, we want to
678 // move the block to be immediately before the PHI block, not
679 // immediately after PredTI.
680 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
681 BasicBlock *NewBB = PN->getIncomingBlock(i);
682 NewBB->moveBefore(PN->getParent());
685 // Splitting the edge can reduce the number of PHI entries we have.
686 e = PN->getNumIncomingValues();
689 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
691 // Insert the code into the end of the predecessor block.
692 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
693 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
695 // Adjust the type back to match the PHI. Note that we can't use
696 // InsertPt here because the SCEVExpander may have inserted its
697 // instructions after that point, in its efforts to avoid inserting
698 // redundant expressions.
699 if (isa<PointerType>(PN->getType())) {
700 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
706 // Replace the use of the operand Value with the new Phi we just created.
707 PN->setIncomingValue(i, Code);
712 // PHI node might have become a constant value after SplitCriticalEdge.
713 DeadInsts.insert(Inst);
715 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
719 /// isTargetConstant - Return true if the following can be referenced by the
720 /// immediate field of a target instruction.
721 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
722 const TargetLowering *TLI) {
723 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
724 int64_t VC = SC->getValue()->getSExtValue();
726 TargetLowering::AddrMode AM;
728 return TLI->isLegalAddressingMode(AM, UseTy);
730 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
731 return (VC > -(1 << 16) && VC < (1 << 16)-1);
735 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
736 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
737 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
738 Constant *Op0 = CE->getOperand(0);
739 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
740 TargetLowering::AddrMode AM;
742 return TLI->isLegalAddressingMode(AM, UseTy);
748 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
749 /// loop varying to the Imm operand.
750 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
751 Loop *L, ScalarEvolution *SE) {
752 if (Val->isLoopInvariant(L)) return; // Nothing to do.
754 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
755 std::vector<SCEVHandle> NewOps;
756 NewOps.reserve(SAE->getNumOperands());
758 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
759 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
760 // If this is a loop-variant expression, it must stay in the immediate
761 // field of the expression.
762 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
764 NewOps.push_back(SAE->getOperand(i));
768 Val = SE->getIntegerSCEV(0, Val->getType());
770 Val = SE->getAddExpr(NewOps);
771 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
772 // Try to pull immediates out of the start value of nested addrec's.
773 SCEVHandle Start = SARE->getStart();
774 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
776 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
778 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
780 // Otherwise, all of Val is variant, move the whole thing over.
781 Imm = SE->getAddExpr(Imm, Val);
782 Val = SE->getIntegerSCEV(0, Val->getType());
787 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
788 /// that can fit into the immediate field of instructions in the target.
789 /// Accumulate these immediate values into the Imm value.
790 static void MoveImmediateValues(const TargetLowering *TLI,
792 SCEVHandle &Val, SCEVHandle &Imm,
793 bool isAddress, Loop *L,
794 ScalarEvolution *SE) {
795 const Type *UseTy = User->getType();
796 if (StoreInst *SI = dyn_cast<StoreInst>(User))
797 UseTy = SI->getOperand(0)->getType();
799 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
800 std::vector<SCEVHandle> NewOps;
801 NewOps.reserve(SAE->getNumOperands());
803 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
804 SCEVHandle NewOp = SAE->getOperand(i);
805 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
807 if (!NewOp->isLoopInvariant(L)) {
808 // If this is a loop-variant expression, it must stay in the immediate
809 // field of the expression.
810 Imm = SE->getAddExpr(Imm, NewOp);
812 NewOps.push_back(NewOp);
817 Val = SE->getIntegerSCEV(0, Val->getType());
819 Val = SE->getAddExpr(NewOps);
821 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
822 // Try to pull immediates out of the start value of nested addrec's.
823 SCEVHandle Start = SARE->getStart();
824 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
826 if (Start != SARE->getStart()) {
827 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
829 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
832 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
833 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
834 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
835 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
837 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
838 SCEVHandle NewOp = SME->getOperand(1);
839 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
841 // If we extracted something out of the subexpressions, see if we can
843 if (NewOp != SME->getOperand(1)) {
844 // Scale SubImm up by "8". If the result is a target constant, we are
846 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
847 if (isTargetConstant(SubImm, UseTy, TLI)) {
848 // Accumulate the immediate.
849 Imm = SE->getAddExpr(Imm, SubImm);
851 // Update what is left of 'Val'.
852 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
859 // Loop-variant expressions must stay in the immediate field of the
861 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
862 !Val->isLoopInvariant(L)) {
863 Imm = SE->getAddExpr(Imm, Val);
864 Val = SE->getIntegerSCEV(0, Val->getType());
868 // Otherwise, no immediates to move.
872 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
873 /// added together. This is used to reassociate common addition subexprs
874 /// together for maximal sharing when rewriting bases.
875 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
877 ScalarEvolution *SE) {
878 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
879 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
880 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
881 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
882 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
883 if (SARE->getOperand(0) == Zero) {
884 SubExprs.push_back(Expr);
886 // Compute the addrec with zero as its base.
887 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
888 Ops[0] = Zero; // Start with zero base.
889 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
892 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
894 } else if (!isa<SCEVConstant>(Expr) ||
895 !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
897 SubExprs.push_back(Expr);
902 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
903 /// removing any common subexpressions from it. Anything truly common is
904 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
905 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
907 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
908 ScalarEvolution *SE) {
909 unsigned NumUses = Uses.size();
911 // Only one use? Use its base, regardless of what it is!
912 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
913 SCEVHandle Result = Zero;
915 std::swap(Result, Uses[0].Base);
919 // To find common subexpressions, count how many of Uses use each expression.
920 // If any subexpressions are used Uses.size() times, they are common.
921 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
923 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
924 // order we see them.
925 std::vector<SCEVHandle> UniqueSubExprs;
927 std::vector<SCEVHandle> SubExprs;
928 for (unsigned i = 0; i != NumUses; ++i) {
929 // If the base is zero (which is common), return zero now, there are no
931 if (Uses[i].Base == Zero) return Zero;
933 // Split the expression into subexprs.
934 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
935 // Add one to SubExpressionUseCounts for each subexpr present.
936 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
937 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
938 UniqueSubExprs.push_back(SubExprs[j]);
942 // Now that we know how many times each is used, build Result. Iterate over
943 // UniqueSubexprs so that we have a stable ordering.
944 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
945 std::map<SCEVHandle, unsigned>::iterator I =
946 SubExpressionUseCounts.find(UniqueSubExprs[i]);
947 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
948 if (I->second == NumUses) { // Found CSE!
949 Result = SE->getAddExpr(Result, I->first);
951 // Remove non-cse's from SubExpressionUseCounts.
952 SubExpressionUseCounts.erase(I);
956 // If we found no CSE's, return now.
957 if (Result == Zero) return Result;
959 // Otherwise, remove all of the CSE's we found from each of the base values.
960 for (unsigned i = 0; i != NumUses; ++i) {
961 // Split the expression into subexprs.
962 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
964 // Remove any common subexpressions.
965 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
966 if (SubExpressionUseCounts.count(SubExprs[j])) {
967 SubExprs.erase(SubExprs.begin()+j);
971 // Finally, the non-shared expressions together.
972 if (SubExprs.empty())
975 Uses[i].Base = SE->getAddExpr(SubExprs);
982 /// isZero - returns true if the scalar evolution expression is zero.
984 static bool isZero(const SCEVHandle &V) {
985 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
986 return SC->getValue()->isZero();
990 /// ValidStride - Check whether the given Scale is valid for all loads and
991 /// stores in UsersToProcess.
993 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
995 const std::vector<BasedUser>& UsersToProcess) {
999 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1000 // If this is a load or other access, pass the type of the access in.
1001 const Type *AccessTy = Type::VoidTy;
1002 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1003 AccessTy = SI->getOperand(0)->getType();
1004 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1005 AccessTy = LI->getType();
1006 else if (isa<PHINode>(UsersToProcess[i].Inst))
1009 TargetLowering::AddrMode AM;
1010 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1011 AM.BaseOffs = SC->getValue()->getSExtValue();
1012 AM.HasBaseReg = HasBaseReg || !isZero(UsersToProcess[i].Base);
1015 // If load[imm+r*scale] is illegal, bail out.
1016 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1022 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1024 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1028 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1030 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1031 !(isa<PointerType>(Ty2) &&
1032 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1033 !(isa<PointerType>(Ty1) &&
1034 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1037 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1038 /// of a previous stride and it is a legal value for the target addressing
1039 /// mode scale component and optional base reg. This allows the users of
1040 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1041 /// reuse is possible.
1042 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1043 bool AllUsesAreAddresses,
1044 const SCEVHandle &Stride,
1045 IVExpr &IV, const Type *Ty,
1046 const std::vector<BasedUser>& UsersToProcess) {
1047 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1048 int64_t SInt = SC->getValue()->getSExtValue();
1049 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1051 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1052 IVsByStride.find(StrideOrder[NewStride]);
1053 if (SI == IVsByStride.end())
1055 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1056 if (SI->first != Stride &&
1057 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1059 int64_t Scale = SInt / SSInt;
1060 // Check that this stride is valid for all the types used for loads and
1061 // stores; if it can be used for some and not others, we might as well use
1062 // the original stride everywhere, since we have to create the IV for it
1063 // anyway. If the scale is 1, then we don't need to worry about folding
1066 (AllUsesAreAddresses &&
1067 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1068 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1069 IE = SI->second.IVs.end(); II != IE; ++II)
1070 // FIXME: Only handle base == 0 for now.
1071 // Only reuse previous IV if it would not require a type conversion.
1072 if (isZero(II->Base) &&
1073 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1082 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1083 /// returns true if Val's isUseOfPostIncrementedValue is true.
1084 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1085 return Val.isUseOfPostIncrementedValue;
1088 /// isNonConstantNegative - Return true if the specified scev is negated, but
1090 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1091 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1092 if (!Mul) return false;
1094 // If there is a constant factor, it will be first.
1095 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1096 if (!SC) return false;
1098 // Return true if the value is negative, this matches things like (-42 * V).
1099 return SC->getValue()->getValue().isNegative();
1102 /// isAddress - Returns true if the specified instruction is using the
1103 /// specified value as an address.
1104 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
1105 bool isAddress = isa<LoadInst>(Inst);
1106 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1107 if (SI->getOperand(1) == OperandVal)
1109 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1110 // Addressing modes can also be folded into prefetches and a variety
1112 switch (II->getIntrinsicID()) {
1114 case Intrinsic::prefetch:
1115 case Intrinsic::x86_sse2_loadu_dq:
1116 case Intrinsic::x86_sse2_loadu_pd:
1117 case Intrinsic::x86_sse_loadu_ps:
1118 case Intrinsic::x86_sse_storeu_ps:
1119 case Intrinsic::x86_sse2_storeu_pd:
1120 case Intrinsic::x86_sse2_storeu_dq:
1121 case Intrinsic::x86_sse2_storel_dq:
1122 if (II->getOperand(1) == OperandVal)
1130 // CollectIVUsers - Transform our list of users and offsets to a bit more
1131 // complex table. In this new vector, each 'BasedUser' contains 'Base' the base
1132 // of the strided accessas well as the old information from Uses. We
1133 // progressively move information from the Base field to the Imm field, until
1134 // we eventually have the full access expression to rewrite the use.
1135 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1136 IVUsersOfOneStride &Uses,
1138 bool &AllUsesAreAddresses,
1139 std::vector<BasedUser> &UsersToProcess) {
1140 UsersToProcess.reserve(Uses.Users.size());
1141 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1142 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1144 // Move any loop invariant operands from the offset field to the immediate
1145 // field of the use, so that we don't try to use something before it is
1147 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1148 UsersToProcess.back().Imm, L, SE);
1149 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1150 "Base value is not loop invariant!");
1153 // We now have a whole bunch of uses of like-strided induction variables, but
1154 // they might all have different bases. We want to emit one PHI node for this
1155 // stride which we fold as many common expressions (between the IVs) into as
1156 // possible. Start by identifying the common expressions in the base values
1157 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1158 // "A+B"), emit it to the preheader, then remove the expression from the
1159 // UsersToProcess base values.
1160 SCEVHandle CommonExprs =
1161 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
1163 // Next, figure out what we can represent in the immediate fields of
1164 // instructions. If we can represent anything there, move it to the imm
1165 // fields of the BasedUsers. We do this so that it increases the commonality
1166 // of the remaining uses.
1167 unsigned NumPHI = 0;
1168 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1169 // If the user is not in the current loop, this means it is using the exit
1170 // value of the IV. Do not put anything in the base, make sure it's all in
1171 // the immediate field to allow as much factoring as possible.
1172 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1173 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1174 UsersToProcess[i].Base);
1175 UsersToProcess[i].Base =
1176 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1179 // Addressing modes can be folded into loads and stores. Be careful that
1180 // the store is through the expression, not of the expression though.
1182 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1183 UsersToProcess[i].OperandValToReplace);
1184 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1189 // If this use isn't an address, then not all uses are addresses.
1190 if (!isAddress && !isPHI)
1191 AllUsesAreAddresses = false;
1193 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1194 UsersToProcess[i].Imm, isAddress, L, SE);
1198 // If one of the use if a PHI node and all other uses are addresses, still
1199 // allow iv reuse. Essentially we are trading one constant multiplication
1200 // for one fewer iv.
1202 AllUsesAreAddresses = false;
1207 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1208 /// stride of IV. All of the users may have different starting values, and this
1209 /// may not be the only stride (we know it is if isOnlyStride is true).
1210 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1211 IVUsersOfOneStride &Uses,
1213 bool isOnlyStride) {
1214 // If all the users are moved to another stride, then there is nothing to do.
1215 if (Uses.Users.empty())
1218 // Keep track if every use in UsersToProcess is an address. If they all are,
1219 // we may be able to rewrite the entire collection of them in terms of a
1220 // smaller-stride IV.
1221 bool AllUsesAreAddresses = true;
1223 // Transform our list of users and offsets to a bit more complex table. In
1224 // this new vector, each 'BasedUser' contains 'Base' the base of the
1225 // strided accessas well as the old information from Uses. We progressively
1226 // move information from the Base field to the Imm field, until we eventually
1227 // have the full access expression to rewrite the use.
1228 std::vector<BasedUser> UsersToProcess;
1229 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1232 // If we managed to find some expressions in common, we'll need to carry
1233 // their value in a register and add it in for each use. This will take up
1234 // a register operand, which potentially restricts what stride values are
1236 bool HaveCommonExprs = !isZero(CommonExprs);
1238 // If all uses are addresses, check if it is possible to reuse an IV with a
1239 // stride that is a factor of this stride. And that the multiple is a number
1240 // that can be encoded in the scale field of the target addressing mode. And
1241 // that we will have a valid instruction after this substition, including the
1242 // immediate field, if any.
1243 PHINode *NewPHI = NULL;
1245 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1246 SE->getIntegerSCEV(0, Type::Int32Ty),
1248 unsigned RewriteFactor = 0;
1249 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1250 Stride, ReuseIV, CommonExprs->getType(),
1252 if (RewriteFactor != 0) {
1253 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1254 << " and BASE " << *ReuseIV.Base << " :\n";
1255 NewPHI = ReuseIV.PHI;
1256 IncV = ReuseIV.IncV;
1259 const Type *ReplacedTy = CommonExprs->getType();
1261 // Now that we know what we need to do, insert the PHI node itself.
1263 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1264 << *Stride << " and BASE " << *CommonExprs << ": ";
1266 SCEVExpander Rewriter(*SE, *LI);
1267 SCEVExpander PreheaderRewriter(*SE, *LI);
1269 BasicBlock *Preheader = L->getLoopPreheader();
1270 Instruction *PreInsertPt = Preheader->getTerminator();
1271 Instruction *PhiInsertBefore = L->getHeader()->begin();
1273 BasicBlock *LatchBlock = L->getLoopLatch();
1276 // Emit the initial base value into the loop preheader.
1278 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1280 if (RewriteFactor == 0) {
1281 // Create a new Phi for this base, and stick it in the loop header.
1282 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1285 // Add common base to the new Phi node.
1286 NewPHI->addIncoming(CommonBaseV, Preheader);
1288 // If the stride is negative, insert a sub instead of an add for the
1290 bool isNegative = isNonConstantNegative(Stride);
1291 SCEVHandle IncAmount = Stride;
1293 IncAmount = SE->getNegativeSCEV(Stride);
1295 // Insert the stride into the preheader.
1296 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1297 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1299 // Emit the increment of the base value before the terminator of the loop
1300 // latch block, and add it to the Phi node.
1301 SCEVHandle IncExp = SE->getUnknown(StrideV);
1303 IncExp = SE->getNegativeSCEV(IncExp);
1304 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1306 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1307 IncV->setName(NewPHI->getName()+".inc");
1308 NewPHI->addIncoming(IncV, LatchBlock);
1310 // Remember this in case a later stride is multiple of this.
1311 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1313 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1315 Constant *C = dyn_cast<Constant>(CommonBaseV);
1317 (!C->isNullValue() &&
1318 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1319 // We want the common base emitted into the preheader! This is just
1320 // using cast as a copy so BitCast (no-op cast) is appropriate
1321 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1322 "commonbase", PreInsertPt);
1326 // We want to emit code for users inside the loop first. To do this, we
1327 // rearrange BasedUser so that the entries at the end have
1328 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1329 // vector (so we handle them first).
1330 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1331 PartitionByIsUseOfPostIncrementedValue);
1333 // Sort this by base, so that things with the same base are handled
1334 // together. By partitioning first and stable-sorting later, we are
1335 // guaranteed that within each base we will pop off users from within the
1336 // loop before users outside of the loop with a particular base.
1338 // We would like to use stable_sort here, but we can't. The problem is that
1339 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1340 // we don't have anything to do a '<' comparison on. Because we think the
1341 // number of uses is small, do a horrible bubble sort which just relies on
1343 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1344 // Get a base value.
1345 SCEVHandle Base = UsersToProcess[i].Base;
1347 // Compact everything with this base to be consequtive with this one.
1348 for (unsigned j = i+1; j != e; ++j) {
1349 if (UsersToProcess[j].Base == Base) {
1350 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1356 // Process all the users now. This outer loop handles all bases, the inner
1357 // loop handles all users of a particular base.
1358 while (!UsersToProcess.empty()) {
1359 SCEVHandle Base = UsersToProcess.back().Base;
1361 // Emit the code for Base into the preheader.
1362 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1364 DOUT << " INSERTING code for BASE = " << *Base << ":";
1365 if (BaseV->hasName())
1366 DOUT << " Result value name = %" << BaseV->getNameStr();
1369 // If BaseV is a constant other than 0, make sure that it gets inserted into
1370 // the preheader, instead of being forward substituted into the uses. We do
1371 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1373 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1374 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1375 // We want this constant emitted into the preheader! This is just
1376 // using cast as a copy so BitCast (no-op cast) is appropriate
1377 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1382 // Emit the code to add the immediate offset to the Phi value, just before
1383 // the instructions that we identified as using this stride and base.
1385 // FIXME: Use emitted users to emit other users.
1386 BasedUser &User = UsersToProcess.back();
1388 // If this instruction wants to use the post-incremented value, move it
1389 // after the post-inc and use its value instead of the PHI.
1390 Value *RewriteOp = NewPHI;
1391 if (User.isUseOfPostIncrementedValue) {
1394 // If this user is in the loop, make sure it is the last thing in the
1395 // loop to ensure it is dominated by the increment.
1396 if (L->contains(User.Inst->getParent()))
1397 User.Inst->moveBefore(LatchBlock->getTerminator());
1399 if (RewriteOp->getType() != ReplacedTy) {
1400 Instruction::CastOps opcode = Instruction::Trunc;
1401 if (ReplacedTy->getPrimitiveSizeInBits() ==
1402 RewriteOp->getType()->getPrimitiveSizeInBits())
1403 opcode = Instruction::BitCast;
1404 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1407 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1409 // If we had to insert new instrutions for RewriteOp, we have to
1410 // consider that they may not have been able to end up immediately
1411 // next to RewriteOp, because non-PHI instructions may never precede
1412 // PHI instructions in a block. In this case, remember where the last
1413 // instruction was inserted so that if we're replacing a different
1414 // PHI node, we can use the later point to expand the final
1416 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1417 if (RewriteOp == NewPHI) NewBasePt = 0;
1419 // Clear the SCEVExpander's expression map so that we are guaranteed
1420 // to have the code emitted where we expect it.
1423 // If we are reusing the iv, then it must be multiplied by a constant
1424 // factor take advantage of addressing mode scale component.
1425 if (RewriteFactor != 0) {
1426 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1427 RewriteExpr->getType()),
1430 // The common base is emitted in the loop preheader. But since we
1431 // are reusing an IV, it has not been used to initialize the PHI node.
1432 // Add it to the expression used to rewrite the uses.
1433 if (!isa<ConstantInt>(CommonBaseV) ||
1434 !cast<ConstantInt>(CommonBaseV)->isZero())
1435 RewriteExpr = SE->getAddExpr(RewriteExpr,
1436 SE->getUnknown(CommonBaseV));
1439 // Now that we know what we need to do, insert code before User for the
1440 // immediate and any loop-variant expressions.
1441 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1442 // Add BaseV to the PHI value if needed.
1443 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1445 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1449 // Mark old value we replaced as possibly dead, so that it is elminated
1450 // if we just replaced the last use of that value.
1451 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1453 UsersToProcess.pop_back();
1456 // If there are any more users to process with the same base, process them
1457 // now. We sorted by base above, so we just have to check the last elt.
1458 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1459 // TODO: Next, find out which base index is the most common, pull it out.
1462 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1463 // different starting values, into different PHIs.
1466 /// FindIVForUser - If Cond has an operand that is an expression of an IV,
1467 /// set the IV user and stride information and return true, otherwise return
1469 bool LoopStrengthReduce::FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
1470 const SCEVHandle *&CondStride) {
1471 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1473 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1474 IVUsesByStride.find(StrideOrder[Stride]);
1475 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1477 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1478 E = SI->second.Users.end(); UI != E; ++UI)
1479 if (UI->User == Cond) {
1480 // NOTE: we could handle setcc instructions with multiple uses here, but
1481 // InstCombine does it as well for simple uses, it's not clear that it
1482 // occurs enough in real life to handle.
1484 CondStride = &SI->first;
1492 // Constant strides come first which in turns are sorted by their absolute
1493 // values. If absolute values are the same, then positive strides comes first.
1495 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1496 struct StrideCompare {
1497 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1498 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1499 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1501 int64_t LV = LHSC->getValue()->getSExtValue();
1502 int64_t RV = RHSC->getValue()->getSExtValue();
1503 uint64_t ALV = (LV < 0) ? -LV : LV;
1504 uint64_t ARV = (RV < 0) ? -RV : RV;
1510 return (LHSC && !RHSC);
1515 /// ChangeCompareStride - If a loop termination compare instruction is the
1516 /// only use of its stride, and the compaison is against a constant value,
1517 /// try eliminate the stride by moving the compare instruction to another
1518 /// stride and change its constant operand accordingly. e.g.
1524 /// if (v2 < 10) goto loop
1529 /// if (v1 < 30) goto loop
1530 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1531 IVStrideUse* &CondUse,
1532 const SCEVHandle* &CondStride) {
1533 // Forgo this transformation if the condition has multiple uses. This is
1534 // over-conservative, but simpler than alternatives. It guards against
1535 // comparisons with a use that occurs earlier than the add instruction for the
1536 // new stride index. See
1537 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness.ll
1538 // for an example of this situation.
1539 if (!Cond->hasOneUse())
1542 if (StrideOrder.size() < 2 ||
1543 IVUsesByStride[*CondStride].Users.size() != 1)
1545 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1546 if (!SC) return Cond;
1547 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1548 if (!C) return Cond;
1550 ICmpInst::Predicate Predicate = Cond->getPredicate();
1551 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1552 int64_t CmpVal = C->getValue().getSExtValue();
1553 unsigned BitWidth = C->getValue().getBitWidth();
1554 uint64_t SignBit = 1ULL << (BitWidth-1);
1555 const Type *CmpTy = C->getType();
1556 const Type *NewCmpTy = NULL;
1557 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1558 unsigned NewTyBits = 0;
1559 int64_t NewCmpVal = CmpVal;
1560 SCEVHandle *NewStride = NULL;
1561 Value *NewIncV = NULL;
1564 // Look for a suitable stride / iv as replacement.
1565 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1566 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1567 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1568 IVUsesByStride.find(StrideOrder[i]);
1569 if (!isa<SCEVConstant>(SI->first))
1571 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1572 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1575 Scale = SSInt / CmpSSInt;
1576 NewCmpVal = CmpVal * Scale;
1577 APInt Mul = APInt(BitWidth, NewCmpVal);
1578 // Check for overflow.
1579 if (Mul.getSExtValue() != NewCmpVal) {
1584 // Watch out for overflow.
1585 if (ICmpInst::isSignedPredicate(Predicate) &&
1586 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1589 if (NewCmpVal != CmpVal) {
1590 // Pick the best iv to use trying to avoid a cast.
1592 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1593 E = SI->second.Users.end(); UI != E; ++UI) {
1594 NewIncV = UI->OperandValToReplace;
1595 if (NewIncV->getType() == CmpTy)
1603 NewCmpTy = NewIncV->getType();
1604 NewTyBits = isa<PointerType>(NewCmpTy)
1605 ? UIntPtrTy->getPrimitiveSizeInBits()
1606 : NewCmpTy->getPrimitiveSizeInBits();
1607 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1608 // Check if it is possible to rewrite it using
1609 // an iv / stride of a smaller integer type.
1610 bool TruncOk = false;
1611 if (NewCmpTy->isInteger()) {
1612 unsigned Bits = NewTyBits;
1613 if (ICmpInst::isSignedPredicate(Predicate))
1615 uint64_t Mask = (1ULL << Bits) - 1;
1616 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1625 // Don't rewrite if use offset is non-constant and the new type is
1626 // of a different type.
1627 // FIXME: too conservative?
1628 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1633 bool AllUsesAreAddresses = true;
1634 std::vector<BasedUser> UsersToProcess;
1635 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1636 AllUsesAreAddresses,
1638 // Avoid rewriting the compare instruction with an iv of new stride
1639 // if it's likely the new stride uses will be rewritten using the
1640 if (AllUsesAreAddresses &&
1641 ValidStride(!isZero(CommonExprs), Scale, UsersToProcess)) {
1646 // If scale is negative, use inverse predicate unless it's testing
1648 if (Scale < 0 && !Cond->isEquality())
1649 Predicate = ICmpInst::getInversePredicate(Predicate);
1651 NewStride = &StrideOrder[i];
1656 if (NewCmpVal != CmpVal) {
1657 // Create a new compare instruction using new stride / iv.
1658 ICmpInst *OldCond = Cond;
1660 if (!isa<PointerType>(NewCmpTy))
1661 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1663 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1664 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1666 // Insert new compare instruction.
1667 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1668 L->getHeader()->getName() + ".termcond",
1671 // Remove the old compare instruction. The old indvar is probably dead too.
1672 DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
1673 SE->deleteValueFromRecords(OldCond);
1674 OldCond->replaceAllUsesWith(Cond);
1675 OldCond->eraseFromParent();
1677 IVUsesByStride[*CondStride].Users.pop_back();
1678 SCEVHandle NewOffset = TyBits == NewTyBits
1679 ? SE->getMulExpr(CondUse->Offset,
1680 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1681 : SE->getConstant(ConstantInt::get(NewCmpTy,
1682 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1683 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1684 CondUse = &IVUsesByStride[*NewStride].Users.back();
1685 CondStride = NewStride;
1692 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1693 // uses in the loop, look to see if we can eliminate some, in favor of using
1694 // common indvars for the different uses.
1695 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1696 // TODO: implement optzns here.
1698 // Finally, get the terminating condition for the loop if possible. If we
1699 // can, we want to change it to use a post-incremented version of its
1700 // induction variable, to allow coalescing the live ranges for the IV into
1701 // one register value.
1702 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1703 BasicBlock *Preheader = L->getLoopPreheader();
1704 BasicBlock *LatchBlock =
1705 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1706 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1707 if (!TermBr || TermBr->isUnconditional() ||
1708 !isa<ICmpInst>(TermBr->getCondition()))
1710 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1712 // Search IVUsesByStride to find Cond's IVUse if there is one.
1713 IVStrideUse *CondUse = 0;
1714 const SCEVHandle *CondStride = 0;
1716 if (!FindIVForUser(Cond, CondUse, CondStride))
1717 return; // setcc doesn't use the IV.
1719 // If possible, change stride and operands of the compare instruction to
1720 // eliminate one stride.
1721 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
1723 // It's possible for the setcc instruction to be anywhere in the loop, and
1724 // possible for it to have multiple users. If it is not immediately before
1725 // the latch block branch, move it.
1726 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1727 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1728 Cond->moveBefore(TermBr);
1730 // Otherwise, clone the terminating condition and insert into the loopend.
1731 Cond = cast<ICmpInst>(Cond->clone());
1732 Cond->setName(L->getHeader()->getName() + ".termcond");
1733 LatchBlock->getInstList().insert(TermBr, Cond);
1735 // Clone the IVUse, as the old use still exists!
1736 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1737 CondUse->OperandValToReplace);
1738 CondUse = &IVUsesByStride[*CondStride].Users.back();
1742 // If we get to here, we know that we can transform the setcc instruction to
1743 // use the post-incremented version of the IV, allowing us to coalesce the
1744 // live ranges for the IV correctly.
1745 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
1746 CondUse->isUseOfPostIncrementedValue = true;
1749 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1751 LI = &getAnalysis<LoopInfo>();
1752 DT = &getAnalysis<DominatorTree>();
1753 SE = &getAnalysis<ScalarEvolution>();
1754 TD = &getAnalysis<TargetData>();
1755 UIntPtrTy = TD->getIntPtrType();
1757 // Find all uses of induction variables in this loop, and catagorize
1758 // them by stride. Start by finding all of the PHI nodes in the header for
1759 // this loop. If they are induction variables, inspect their uses.
1760 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
1761 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1762 AddUsersIfInteresting(I, L, Processed);
1764 // If we have nothing to do, return.
1765 if (IVUsesByStride.empty()) return false;
1767 // Optimize induction variables. Some indvar uses can be transformed to use
1768 // strides that will be needed for other purposes. A common example of this
1769 // is the exit test for the loop, which can often be rewritten to use the
1770 // computation of some other indvar to decide when to terminate the loop.
1774 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1775 // doing computation in byte values, promote to 32-bit values if safe.
1777 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1778 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1779 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1780 // to be careful that IV's are all the same type. Only works for intptr_t
1783 // If we only have one stride, we can more aggressively eliminate some things.
1784 bool HasOneStride = IVUsesByStride.size() == 1;
1787 DOUT << "\nLSR on ";
1791 // IVsByStride keeps IVs for one particular loop.
1792 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
1794 // Sort the StrideOrder so we process larger strides first.
1795 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1797 // Note: this processes each stride/type pair individually. All users passed
1798 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1799 // note that we iterate over IVUsesByStride indirectly by using StrideOrder.
1800 // This extra layer of indirection makes the ordering of strides deterministic
1801 // - not dependent on map order.
1802 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1803 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1804 IVUsesByStride.find(StrideOrder[Stride]);
1805 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1806 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1809 // We're done analyzing this loop; release all the state we built up for it.
1810 CastedPointers.clear();
1811 IVUsesByStride.clear();
1812 IVsByStride.clear();
1813 StrideOrder.clear();
1815 // Clean up after ourselves
1816 if (!DeadInsts.empty()) {
1817 DeleteTriviallyDeadInstructions(DeadInsts);
1819 BasicBlock::iterator I = L->getHeader()->begin();
1821 while ((PN = dyn_cast<PHINode>(I))) {
1822 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1824 // At this point, we know that we have killed one or more GEP
1825 // instructions. It is worth checking to see if the cann indvar is also
1826 // dead, so that we can remove it as well. The requirements for the cann
1827 // indvar to be considered dead are:
1828 // 1. the cann indvar has one use
1829 // 2. the use is an add instruction
1830 // 3. the add has one use
1831 // 4. the add is used by the cann indvar
1832 // If all four cases above are true, then we can remove both the add and
1834 // FIXME: this needs to eliminate an induction variable even if it's being
1835 // compared against some value to decide loop termination.
1836 if (PN->hasOneUse()) {
1837 Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
1838 if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
1839 if (BO->hasOneUse() && PN == *(BO->use_begin())) {
1840 DeadInsts.insert(BO);
1841 // Break the cycle, then delete the PHI.
1842 SE->deleteValueFromRecords(PN);
1843 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1844 PN->eraseFromParent();
1849 DeleteTriviallyDeadInstructions(DeadInsts);