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 'Operand'
55 /// 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 std::vector<SCEVHandle> 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 std::map<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, Loop *L);
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();
231 if (isInstructionTriviallyDead(I)) {
232 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
233 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
235 SE->deleteValueFromRecords(I);
236 I->eraseFromParent();
243 /// GetExpressionSCEV - Compute and return the SCEV for the specified
245 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
246 // Pointer to pointer bitcast instructions return the same value as their
248 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
249 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
250 return SE->getSCEV(BCI);
251 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)), L);
256 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
257 // If this is a GEP that SE doesn't know about, compute it now and insert it.
258 // If this is not a GEP, or if we have already done this computation, just let
260 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
261 if (!GEP || SE->hasSCEV(GEP))
262 return SE->getSCEV(Exp);
264 // Analyze all of the subscripts of this getelementptr instruction, looking
265 // for uses that are determined by the trip count of L. First, skip all
266 // operands the are not dependent on the IV.
268 // Build up the base expression. Insert an LLVM cast of the pointer to
270 SCEVHandle GEPVal = SE->getUnknown(
271 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
273 gep_type_iterator GTI = gep_type_begin(GEP);
275 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
276 // If this is a use of a recurrence that we can analyze, and it comes before
277 // Op does in the GEP operand list, we will handle this when we process this
279 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
280 const StructLayout *SL = TD->getStructLayout(STy);
281 unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
282 uint64_t Offset = SL->getElementOffset(Idx);
283 GEPVal = SE->getAddExpr(GEPVal,
284 SE->getIntegerSCEV(Offset, UIntPtrTy));
286 unsigned GEPOpiBits =
287 GEP->getOperand(i)->getType()->getPrimitiveSizeInBits();
288 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
289 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
290 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
291 Instruction::BitCast));
292 Value *OpVal = getCastedVersionOf(opcode, GEP->getOperand(i));
293 SCEVHandle Idx = SE->getSCEV(OpVal);
295 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
297 Idx = SE->getMulExpr(Idx,
298 SE->getConstant(ConstantInt::get(UIntPtrTy,
300 GEPVal = SE->getAddExpr(GEPVal, Idx);
304 SE->setSCEV(GEP, GEPVal);
308 /// getSCEVStartAndStride - Compute the start and stride of this expression,
309 /// returning false if the expression is not a start/stride pair, or true if it
310 /// is. The stride must be a loop invariant expression, but the start may be
311 /// a mix of loop invariant and loop variant expressions.
312 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
313 SCEVHandle &Start, SCEVHandle &Stride,
314 ScalarEvolution *SE) {
315 SCEVHandle TheAddRec = Start; // Initialize to zero.
317 // If the outer level is an AddExpr, the operands are all start values except
318 // for a nested AddRecExpr.
319 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
320 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
321 if (SCEVAddRecExpr *AddRec =
322 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
323 if (AddRec->getLoop() == L)
324 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
326 return false; // Nested IV of some sort?
328 Start = SE->getAddExpr(Start, AE->getOperand(i));
331 } else if (isa<SCEVAddRecExpr>(SH)) {
334 return false; // not analyzable.
337 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
338 if (!AddRec || AddRec->getLoop() != L) return false;
340 // FIXME: Generalize to non-affine IV's.
341 if (!AddRec->isAffine()) return false;
343 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
345 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
346 DOUT << "[" << L->getHeader()->getName()
347 << "] Variable stride: " << *AddRec << "\n";
349 Stride = AddRec->getOperand(1);
353 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
354 /// and now we need to decide whether the user should use the preinc or post-inc
355 /// value. If this user should use the post-inc version of the IV, return true.
357 /// Choosing wrong here can break dominance properties (if we choose to use the
358 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
359 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
360 /// should use the post-inc value).
361 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
362 Loop *L, DominatorTree *DT, Pass *P) {
363 // If the user is in the loop, use the preinc value.
364 if (L->contains(User->getParent())) return false;
366 BasicBlock *LatchBlock = L->getLoopLatch();
368 // Ok, the user is outside of the loop. If it is dominated by the latch
369 // block, use the post-inc value.
370 if (DT->dominates(LatchBlock, User->getParent()))
373 // There is one case we have to be careful of: PHI nodes. These little guys
374 // can live in blocks that do not dominate the latch block, but (since their
375 // uses occur in the predecessor block, not the block the PHI lives in) should
376 // still use the post-inc value. Check for this case now.
377 PHINode *PN = dyn_cast<PHINode>(User);
378 if (!PN) return false; // not a phi, not dominated by latch block.
380 // Look at all of the uses of IV by the PHI node. If any use corresponds to
381 // a block that is not dominated by the latch block, give up and use the
382 // preincremented value.
383 unsigned NumUses = 0;
384 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
385 if (PN->getIncomingValue(i) == IV) {
387 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
391 // Okay, all uses of IV by PN are in predecessor blocks that really are
392 // dominated by the latch block. Split the critical edges and use the
393 // post-incremented value.
394 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
395 if (PN->getIncomingValue(i) == IV) {
396 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P,
398 // Splitting the critical edge can reduce the number of entries in this
400 e = PN->getNumIncomingValues();
401 if (--NumUses == 0) break;
409 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
410 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
411 /// return true. Otherwise, return false.
412 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
413 SmallPtrSet<Instruction*,16> &Processed) {
414 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
415 return false; // Void and FP expressions cannot be reduced.
416 if (!Processed.insert(I))
417 return true; // Instruction already handled.
419 // Get the symbolic expression for this instruction.
420 SCEVHandle ISE = GetExpressionSCEV(I, L);
421 if (isa<SCEVCouldNotCompute>(ISE)) return false;
423 // Get the start and stride for this expression.
424 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
425 SCEVHandle Stride = Start;
426 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
427 return false; // Non-reducible symbolic expression, bail out.
429 std::vector<Instruction *> IUsers;
430 // Collect all I uses now because IVUseShouldUsePostIncValue may
431 // invalidate use_iterator.
432 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
433 IUsers.push_back(cast<Instruction>(*UI));
435 for (unsigned iused_index = 0, iused_size = IUsers.size();
436 iused_index != iused_size; ++iused_index) {
438 Instruction *User = IUsers[iused_index];
440 // Do not infinitely recurse on PHI nodes.
441 if (isa<PHINode>(User) && Processed.count(User))
444 // If this is an instruction defined in a nested loop, or outside this loop,
445 // don't recurse into it.
446 bool AddUserToIVUsers = false;
447 if (LI->getLoopFor(User->getParent()) != L) {
448 DOUT << "FOUND USER in other loop: " << *User
449 << " OF SCEV: " << *ISE << "\n";
450 AddUserToIVUsers = true;
451 } else if (!AddUsersIfInteresting(User, L, Processed)) {
452 DOUT << "FOUND USER: " << *User
453 << " OF SCEV: " << *ISE << "\n";
454 AddUserToIVUsers = true;
457 if (AddUserToIVUsers) {
458 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
459 if (StrideUses.Users.empty()) // First occurance of this stride?
460 StrideOrder.push_back(Stride);
462 // Okay, we found a user that we cannot reduce. Analyze the instruction
463 // and decide what to do with it. If we are a use inside of the loop, use
464 // the value before incrementation, otherwise use it after incrementation.
465 if (IVUseShouldUsePostIncValue(User, I, L, DT, this)) {
466 // The value used will be incremented by the stride more than we are
467 // expecting, so subtract this off.
468 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
469 StrideUses.addUser(NewStart, User, I);
470 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
471 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
473 StrideUses.addUser(Start, User, I);
481 /// BasedUser - For a particular base value, keep information about how we've
482 /// partitioned the expression so far.
484 /// SE - The current ScalarEvolution object.
487 /// Base - The Base value for the PHI node that needs to be inserted for
488 /// this use. As the use is processed, information gets moved from this
489 /// field to the Imm field (below). BasedUser values are sorted by this
493 /// Inst - The instruction using the induction variable.
496 /// OperandValToReplace - The operand value of Inst to replace with the
498 Value *OperandValToReplace;
500 /// Imm - The immediate value that should be added to the base immediately
501 /// before Inst, because it will be folded into the imm field of the
505 /// EmittedBase - The actual value* to use for the base value of this
506 /// operation. This is null if we should just use zero so far.
509 // isUseOfPostIncrementedValue - True if this should use the
510 // post-incremented version of this IV, not the preincremented version.
511 // This can only be set in special cases, such as the terminating setcc
512 // instruction for a loop and uses outside the loop that are dominated by
514 bool isUseOfPostIncrementedValue;
516 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
517 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
518 OperandValToReplace(IVSU.OperandValToReplace),
519 Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
520 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
522 // Once we rewrite the code to insert the new IVs we want, update the
523 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
525 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
526 SCEVExpander &Rewriter, Loop *L,
529 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
530 SCEVExpander &Rewriter,
531 Instruction *IP, Loop *L);
536 void BasedUser::dump() const {
537 cerr << " Base=" << *Base;
538 cerr << " Imm=" << *Imm;
540 cerr << " EB=" << *EmittedBase;
542 cerr << " Inst: " << *Inst;
545 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
546 SCEVExpander &Rewriter,
547 Instruction *IP, Loop *L) {
548 // Figure out where we *really* want to insert this code. In particular, if
549 // the user is inside of a loop that is nested inside of L, we really don't
550 // want to insert this expression before the user, we'd rather pull it out as
551 // many loops as possible.
552 LoopInfo &LI = Rewriter.getLoopInfo();
553 Instruction *BaseInsertPt = IP;
555 // Figure out the most-nested loop that IP is in.
556 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
558 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
559 // the preheader of the outer-most loop where NewBase is not loop invariant.
560 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
561 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
562 InsertLoop = InsertLoop->getParentLoop();
565 // If there is no immediate value, skip the next part.
566 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
567 if (SC->getValue()->isZero())
568 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
570 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
572 // If we are inserting the base and imm values in the same block, make sure to
573 // adjust the IP position if insertion reused a result.
574 if (IP == BaseInsertPt)
575 IP = Rewriter.getInsertionPoint();
577 // Always emit the immediate (if non-zero) into the same block as the user.
578 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
579 return Rewriter.expandCodeFor(NewValSCEV, IP);
584 // Once we rewrite the code to insert the new IVs we want, update the
585 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
587 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
588 SCEVExpander &Rewriter,
590 if (!isa<PHINode>(Inst)) {
591 // By default, insert code at the user instruction.
592 BasicBlock::iterator InsertPt = Inst;
594 // However, if the Operand is itself an instruction, the (potentially
595 // complex) inserted code may be shared by many users. Because of this, we
596 // want to emit code for the computation of the operand right before its old
597 // computation. This is usually safe, because we obviously used to use the
598 // computation when it was computed in its current block. However, in some
599 // cases (e.g. use of a post-incremented induction variable) the NewBase
600 // value will be pinned to live somewhere after the original computation.
601 // In this case, we have to back off.
602 if (!isUseOfPostIncrementedValue) {
603 if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) {
605 while (isa<PHINode>(InsertPt)) ++InsertPt;
608 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
609 // Adjust the type back to match the Inst. Note that we can't use InsertPt
610 // here because the SCEVExpander may have inserted the instructions after
611 // that point, in its efforts to avoid inserting redundant expressions.
612 if (isa<PointerType>(OperandValToReplace->getType())) {
613 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
615 OperandValToReplace->getType());
617 // Replace the use of the operand Value with the new Phi we just created.
618 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
619 DOUT << " CHANGED: IMM =" << *Imm;
620 DOUT << " \tNEWBASE =" << *NewBase;
621 DOUT << " \tInst = " << *Inst;
625 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
626 // expression into each operand block that uses it. Note that PHI nodes can
627 // have multiple entries for the same predecessor. We use a map to make sure
628 // that a PHI node only has a single Value* for each predecessor (which also
629 // prevents us from inserting duplicate code in some blocks).
630 std::map<BasicBlock*, Value*> InsertedCode;
631 PHINode *PN = cast<PHINode>(Inst);
632 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
633 if (PN->getIncomingValue(i) == OperandValToReplace) {
634 // If this is a critical edge, split the edge so that we do not insert the
635 // code on all predecessor/successor paths. We do this unless this is the
636 // canonical backedge for this loop, as this can make some inserted code
637 // be in an illegal position.
638 BasicBlock *PHIPred = PN->getIncomingBlock(i);
639 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
640 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
642 // First step, split the critical edge.
643 SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
645 // Next step: move the basic block. In particular, if the PHI node
646 // is outside of the loop, and PredTI is in the loop, we want to
647 // move the block to be immediately before the PHI block, not
648 // immediately after PredTI.
649 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
650 BasicBlock *NewBB = PN->getIncomingBlock(i);
651 NewBB->moveBefore(PN->getParent());
654 // Splitting the edge can reduce the number of PHI entries we have.
655 e = PN->getNumIncomingValues();
658 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
660 // Insert the code into the end of the predecessor block.
661 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
662 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
664 // Adjust the type back to match the PHI. Note that we can't use
665 // InsertPt here because the SCEVExpander may have inserted its
666 // instructions after that point, in its efforts to avoid inserting
667 // redundant expressions.
668 if (isa<PointerType>(PN->getType())) {
669 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
675 // Replace the use of the operand Value with the new Phi we just created.
676 PN->setIncomingValue(i, Code);
680 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
684 /// isTargetConstant - Return true if the following can be referenced by the
685 /// immediate field of a target instruction.
686 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
687 const TargetLowering *TLI) {
688 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
689 int64_t VC = SC->getValue()->getSExtValue();
691 TargetLowering::AddrMode AM;
693 return TLI->isLegalAddressingMode(AM, UseTy);
695 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
696 return (VC > -(1 << 16) && VC < (1 << 16)-1);
700 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
701 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
702 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
703 Constant *Op0 = CE->getOperand(0);
704 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
705 TargetLowering::AddrMode AM;
707 return TLI->isLegalAddressingMode(AM, UseTy);
713 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
714 /// loop varying to the Imm operand.
715 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
716 Loop *L, ScalarEvolution *SE) {
717 if (Val->isLoopInvariant(L)) return; // Nothing to do.
719 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
720 std::vector<SCEVHandle> NewOps;
721 NewOps.reserve(SAE->getNumOperands());
723 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
724 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
725 // If this is a loop-variant expression, it must stay in the immediate
726 // field of the expression.
727 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
729 NewOps.push_back(SAE->getOperand(i));
733 Val = SE->getIntegerSCEV(0, Val->getType());
735 Val = SE->getAddExpr(NewOps);
736 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
737 // Try to pull immediates out of the start value of nested addrec's.
738 SCEVHandle Start = SARE->getStart();
739 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
741 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
743 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
745 // Otherwise, all of Val is variant, move the whole thing over.
746 Imm = SE->getAddExpr(Imm, Val);
747 Val = SE->getIntegerSCEV(0, Val->getType());
752 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
753 /// that can fit into the immediate field of instructions in the target.
754 /// Accumulate these immediate values into the Imm value.
755 static void MoveImmediateValues(const TargetLowering *TLI,
757 SCEVHandle &Val, SCEVHandle &Imm,
758 bool isAddress, Loop *L,
759 ScalarEvolution *SE) {
760 const Type *UseTy = User->getType();
761 if (StoreInst *SI = dyn_cast<StoreInst>(User))
762 UseTy = SI->getOperand(0)->getType();
764 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
765 std::vector<SCEVHandle> NewOps;
766 NewOps.reserve(SAE->getNumOperands());
768 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
769 SCEVHandle NewOp = SAE->getOperand(i);
770 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
772 if (!NewOp->isLoopInvariant(L)) {
773 // If this is a loop-variant expression, it must stay in the immediate
774 // field of the expression.
775 Imm = SE->getAddExpr(Imm, NewOp);
777 NewOps.push_back(NewOp);
782 Val = SE->getIntegerSCEV(0, Val->getType());
784 Val = SE->getAddExpr(NewOps);
786 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
787 // Try to pull immediates out of the start value of nested addrec's.
788 SCEVHandle Start = SARE->getStart();
789 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
791 if (Start != SARE->getStart()) {
792 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
794 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
797 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
798 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
799 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
800 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
802 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
803 SCEVHandle NewOp = SME->getOperand(1);
804 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
806 // If we extracted something out of the subexpressions, see if we can
808 if (NewOp != SME->getOperand(1)) {
809 // Scale SubImm up by "8". If the result is a target constant, we are
811 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
812 if (isTargetConstant(SubImm, UseTy, TLI)) {
813 // Accumulate the immediate.
814 Imm = SE->getAddExpr(Imm, SubImm);
816 // Update what is left of 'Val'.
817 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
824 // Loop-variant expressions must stay in the immediate field of the
826 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
827 !Val->isLoopInvariant(L)) {
828 Imm = SE->getAddExpr(Imm, Val);
829 Val = SE->getIntegerSCEV(0, Val->getType());
833 // Otherwise, no immediates to move.
837 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
838 /// added together. This is used to reassociate common addition subexprs
839 /// together for maximal sharing when rewriting bases.
840 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
842 ScalarEvolution *SE) {
843 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
844 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
845 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
846 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
847 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
848 if (SARE->getOperand(0) == Zero) {
849 SubExprs.push_back(Expr);
851 // Compute the addrec with zero as its base.
852 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
853 Ops[0] = Zero; // Start with zero base.
854 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
857 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
859 } else if (!isa<SCEVConstant>(Expr) ||
860 !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
862 SubExprs.push_back(Expr);
867 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
868 /// removing any common subexpressions from it. Anything truly common is
869 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
870 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
872 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
873 ScalarEvolution *SE) {
874 unsigned NumUses = Uses.size();
876 // Only one use? Use its base, regardless of what it is!
877 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
878 SCEVHandle Result = Zero;
880 std::swap(Result, Uses[0].Base);
884 // To find common subexpressions, count how many of Uses use each expression.
885 // If any subexpressions are used Uses.size() times, they are common.
886 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
888 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
889 // order we see them.
890 std::vector<SCEVHandle> UniqueSubExprs;
892 std::vector<SCEVHandle> SubExprs;
893 for (unsigned i = 0; i != NumUses; ++i) {
894 // If the base is zero (which is common), return zero now, there are no
896 if (Uses[i].Base == Zero) return Zero;
898 // Split the expression into subexprs.
899 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
900 // Add one to SubExpressionUseCounts for each subexpr present.
901 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
902 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
903 UniqueSubExprs.push_back(SubExprs[j]);
907 // Now that we know how many times each is used, build Result. Iterate over
908 // UniqueSubexprs so that we have a stable ordering.
909 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
910 std::map<SCEVHandle, unsigned>::iterator I =
911 SubExpressionUseCounts.find(UniqueSubExprs[i]);
912 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
913 if (I->second == NumUses) { // Found CSE!
914 Result = SE->getAddExpr(Result, I->first);
916 // Remove non-cse's from SubExpressionUseCounts.
917 SubExpressionUseCounts.erase(I);
921 // If we found no CSE's, return now.
922 if (Result == Zero) return Result;
924 // Otherwise, remove all of the CSE's we found from each of the base values.
925 for (unsigned i = 0; i != NumUses; ++i) {
926 // Split the expression into subexprs.
927 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
929 // Remove any common subexpressions.
930 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
931 if (SubExpressionUseCounts.count(SubExprs[j])) {
932 SubExprs.erase(SubExprs.begin()+j);
936 // Finally, the non-shared expressions together.
937 if (SubExprs.empty())
940 Uses[i].Base = SE->getAddExpr(SubExprs);
947 /// isZero - returns true if the scalar evolution expression is zero.
949 static bool isZero(const SCEVHandle &V) {
950 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
951 return SC->getValue()->isZero();
955 /// ValidStride - Check whether the given Scale is valid for all loads and
956 /// stores in UsersToProcess.
958 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
960 const std::vector<BasedUser>& UsersToProcess) {
961 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
962 // If this is a load or other access, pass the type of the access in.
963 const Type *AccessTy = Type::VoidTy;
964 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
965 AccessTy = SI->getOperand(0)->getType();
966 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
967 AccessTy = LI->getType();
969 TargetLowering::AddrMode AM;
970 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
971 AM.BaseOffs = SC->getValue()->getSExtValue();
972 AM.HasBaseReg = HasBaseReg || !isZero(UsersToProcess[i].Base);
975 // If load[imm+r*scale] is illegal, bail out.
976 if (TLI && !TLI->isLegalAddressingMode(AM, AccessTy))
982 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
984 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
988 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
990 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
991 !(isa<PointerType>(Ty2) &&
992 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
993 !(isa<PointerType>(Ty1) &&
994 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
997 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
998 /// of a previous stride and it is a legal value for the target addressing
999 /// mode scale component and optional base reg. This allows the users of
1000 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1001 /// reuse is possible.
1002 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1003 bool AllUsesAreAddresses,
1004 const SCEVHandle &Stride,
1005 IVExpr &IV, const Type *Ty,
1006 const std::vector<BasedUser>& UsersToProcess) {
1007 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1008 int64_t SInt = SC->getValue()->getSExtValue();
1009 for (std::map<SCEVHandle, IVsOfOneStride>::iterator SI= IVsByStride.begin(),
1010 SE = IVsByStride.end(); SI != SE; ++SI) {
1011 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1012 if (SI->first != Stride &&
1013 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1015 int64_t Scale = SInt / SSInt;
1016 // Check that this stride is valid for all the types used for loads and
1017 // stores; if it can be used for some and not others, we might as well use
1018 // the original stride everywhere, since we have to create the IV for it
1019 // anyway. If the scale is 1, then we don't need to worry about folding
1022 (AllUsesAreAddresses &&
1023 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1024 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1025 IE = SI->second.IVs.end(); II != IE; ++II)
1026 // FIXME: Only handle base == 0 for now.
1027 // Only reuse previous IV if it would not require a type conversion.
1028 if (isZero(II->Base) &&
1029 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1038 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1039 /// returns true if Val's isUseOfPostIncrementedValue is true.
1040 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1041 return Val.isUseOfPostIncrementedValue;
1044 /// isNonConstantNegative - REturn true if the specified scev is negated, but
1046 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1047 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1048 if (!Mul) return false;
1050 // If there is a constant factor, it will be first.
1051 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1052 if (!SC) return false;
1054 // Return true if the value is negative, this matches things like (-42 * V).
1055 return SC->getValue()->getValue().isNegative();
1058 // CollectIVUsers - Transform our list of users and offsets to a bit more
1059 // complex table. In this new vector, each 'BasedUser' contains 'Base' the base
1060 // of the strided accessas well as the old information from Uses. We
1061 // progressively move information from the Base field to the Imm field, until
1062 // we eventually have the full access expression to rewrite the use.
1063 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1064 IVUsersOfOneStride &Uses,
1066 bool &AllUsesAreAddresses,
1067 std::vector<BasedUser> &UsersToProcess) {
1068 UsersToProcess.reserve(Uses.Users.size());
1069 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1070 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1072 // Move any loop invariant operands from the offset field to the immediate
1073 // field of the use, so that we don't try to use something before it is
1075 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1076 UsersToProcess.back().Imm, L, SE);
1077 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1078 "Base value is not loop invariant!");
1081 // We now have a whole bunch of uses of like-strided induction variables, but
1082 // they might all have different bases. We want to emit one PHI node for this
1083 // stride which we fold as many common expressions (between the IVs) into as
1084 // possible. Start by identifying the common expressions in the base values
1085 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1086 // "A+B"), emit it to the preheader, then remove the expression from the
1087 // UsersToProcess base values.
1088 SCEVHandle CommonExprs =
1089 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
1091 // Next, figure out what we can represent in the immediate fields of
1092 // instructions. If we can represent anything there, move it to the imm
1093 // fields of the BasedUsers. We do this so that it increases the commonality
1094 // of the remaining uses.
1095 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1096 // If the user is not in the current loop, this means it is using the exit
1097 // value of the IV. Do not put anything in the base, make sure it's all in
1098 // the immediate field to allow as much factoring as possible.
1099 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1100 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1101 UsersToProcess[i].Base);
1102 UsersToProcess[i].Base =
1103 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1106 // Addressing modes can be folded into loads and stores. Be careful that
1107 // the store is through the expression, not of the expression though.
1108 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
1109 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst)) {
1110 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1112 } else if (IntrinsicInst *II =
1113 dyn_cast<IntrinsicInst>(UsersToProcess[i].Inst)) {
1114 // Addressing modes can also be folded into prefetches and a variety
1116 switch (II->getIntrinsicID()) {
1118 case Intrinsic::prefetch:
1119 case Intrinsic::x86_sse2_loadu_dq:
1120 case Intrinsic::x86_sse2_loadu_pd:
1121 case Intrinsic::x86_sse_loadu_ps:
1122 case Intrinsic::x86_sse_storeu_ps:
1123 case Intrinsic::x86_sse2_storeu_pd:
1124 case Intrinsic::x86_sse2_storeu_dq:
1125 case Intrinsic::x86_sse2_storel_dq:
1126 if (II->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1129 case Intrinsic::x86_sse2_loadh_pd:
1130 case Intrinsic::x86_sse2_loadl_pd:
1131 if (II->getOperand(2) == UsersToProcess[i].OperandValToReplace)
1137 // If this use isn't an address, then not all uses are addresses.
1139 AllUsesAreAddresses = false;
1141 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1142 UsersToProcess[i].Imm, isAddress, L, SE);
1149 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1150 /// stride of IV. All of the users may have different starting values, and this
1151 /// may not be the only stride (we know it is if isOnlyStride is true).
1152 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1153 IVUsersOfOneStride &Uses,
1155 bool isOnlyStride) {
1156 // If all the users are moved to another stride, then there is nothing to do.
1157 if (Uses.Users.size() == 0)
1160 // Keep track if every use in UsersToProcess is an address. If they all are,
1161 // we may be able to rewrite the entire collection of them in terms of a
1162 // smaller-stride IV.
1163 bool AllUsesAreAddresses = true;
1165 // Transform our list of users and offsets to a bit more complex table. In
1166 // this new vector, each 'BasedUser' contains 'Base' the base of the
1167 // strided accessas well as the old information from Uses. We progressively
1168 // move information from the Base field to the Imm field, until we eventually
1169 // have the full access expression to rewrite the use.
1170 std::vector<BasedUser> UsersToProcess;
1171 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1174 // If we managed to find some expressions in common, we'll need to carry
1175 // their value in a register and add it in for each use. This will take up
1176 // a register operand, which potentially restricts what stride values are
1178 bool HaveCommonExprs = !isZero(CommonExprs);
1180 // If all uses are addresses, check if it is possible to reuse an IV with a
1181 // stride that is a factor of this stride. And that the multiple is a number
1182 // that can be encoded in the scale field of the target addressing mode. And
1183 // that we will have a valid instruction after this substition, including the
1184 // immediate field, if any.
1185 PHINode *NewPHI = NULL;
1187 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1188 SE->getIntegerSCEV(0, Type::Int32Ty),
1190 unsigned RewriteFactor = 0;
1191 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1192 Stride, ReuseIV, CommonExprs->getType(),
1194 if (RewriteFactor != 0) {
1195 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1196 << " and BASE " << *ReuseIV.Base << " :\n";
1197 NewPHI = ReuseIV.PHI;
1198 IncV = ReuseIV.IncV;
1201 const Type *ReplacedTy = CommonExprs->getType();
1203 // Now that we know what we need to do, insert the PHI node itself.
1205 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1206 << *Stride << " and BASE " << *CommonExprs << ": ";
1208 SCEVExpander Rewriter(*SE, *LI);
1209 SCEVExpander PreheaderRewriter(*SE, *LI);
1211 BasicBlock *Preheader = L->getLoopPreheader();
1212 Instruction *PreInsertPt = Preheader->getTerminator();
1213 Instruction *PhiInsertBefore = L->getHeader()->begin();
1215 BasicBlock *LatchBlock = L->getLoopLatch();
1218 // Emit the initial base value into the loop preheader.
1220 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1222 if (RewriteFactor == 0) {
1223 // Create a new Phi for this base, and stick it in the loop header.
1224 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1227 // Add common base to the new Phi node.
1228 NewPHI->addIncoming(CommonBaseV, Preheader);
1230 // If the stride is negative, insert a sub instead of an add for the
1232 bool isNegative = isNonConstantNegative(Stride);
1233 SCEVHandle IncAmount = Stride;
1235 IncAmount = SE->getNegativeSCEV(Stride);
1237 // Insert the stride into the preheader.
1238 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1239 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1241 // Emit the increment of the base value before the terminator of the loop
1242 // latch block, and add it to the Phi node.
1243 SCEVHandle IncExp = SE->getUnknown(StrideV);
1245 IncExp = SE->getNegativeSCEV(IncExp);
1246 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1248 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1249 IncV->setName(NewPHI->getName()+".inc");
1250 NewPHI->addIncoming(IncV, LatchBlock);
1252 // Remember this in case a later stride is multiple of this.
1253 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1255 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1257 Constant *C = dyn_cast<Constant>(CommonBaseV);
1259 (!C->isNullValue() &&
1260 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1261 // We want the common base emitted into the preheader! This is just
1262 // using cast as a copy so BitCast (no-op cast) is appropriate
1263 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1264 "commonbase", PreInsertPt);
1268 // We want to emit code for users inside the loop first. To do this, we
1269 // rearrange BasedUser so that the entries at the end have
1270 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1271 // vector (so we handle them first).
1272 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1273 PartitionByIsUseOfPostIncrementedValue);
1275 // Sort this by base, so that things with the same base are handled
1276 // together. By partitioning first and stable-sorting later, we are
1277 // guaranteed that within each base we will pop off users from within the
1278 // loop before users outside of the loop with a particular base.
1280 // We would like to use stable_sort here, but we can't. The problem is that
1281 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1282 // we don't have anything to do a '<' comparison on. Because we think the
1283 // number of uses is small, do a horrible bubble sort which just relies on
1285 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1286 // Get a base value.
1287 SCEVHandle Base = UsersToProcess[i].Base;
1289 // Compact everything with this base to be consequetive with this one.
1290 for (unsigned j = i+1; j != e; ++j) {
1291 if (UsersToProcess[j].Base == Base) {
1292 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1298 // Process all the users now. This outer loop handles all bases, the inner
1299 // loop handles all users of a particular base.
1300 while (!UsersToProcess.empty()) {
1301 SCEVHandle Base = UsersToProcess.back().Base;
1303 // Emit the code for Base into the preheader.
1304 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1306 DOUT << " INSERTING code for BASE = " << *Base << ":";
1307 if (BaseV->hasName())
1308 DOUT << " Result value name = %" << BaseV->getNameStr();
1311 // If BaseV is a constant other than 0, make sure that it gets inserted into
1312 // the preheader, instead of being forward substituted into the uses. We do
1313 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1315 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1316 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1317 // We want this constant emitted into the preheader! This is just
1318 // using cast as a copy so BitCast (no-op cast) is appropriate
1319 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1324 // Emit the code to add the immediate offset to the Phi value, just before
1325 // the instructions that we identified as using this stride and base.
1327 // FIXME: Use emitted users to emit other users.
1328 BasedUser &User = UsersToProcess.back();
1330 // If this instruction wants to use the post-incremented value, move it
1331 // after the post-inc and use its value instead of the PHI.
1332 Value *RewriteOp = NewPHI;
1333 if (User.isUseOfPostIncrementedValue) {
1336 // If this user is in the loop, make sure it is the last thing in the
1337 // loop to ensure it is dominated by the increment.
1338 if (L->contains(User.Inst->getParent()))
1339 User.Inst->moveBefore(LatchBlock->getTerminator());
1341 if (RewriteOp->getType() != ReplacedTy) {
1342 Instruction::CastOps opcode = Instruction::Trunc;
1343 if (ReplacedTy->getPrimitiveSizeInBits() ==
1344 RewriteOp->getType()->getPrimitiveSizeInBits())
1345 opcode = Instruction::BitCast;
1346 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1349 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1351 // Clear the SCEVExpander's expression map so that we are guaranteed
1352 // to have the code emitted where we expect it.
1355 // If we are reusing the iv, then it must be multiplied by a constant
1356 // factor take advantage of addressing mode scale component.
1357 if (RewriteFactor != 0) {
1359 SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1360 RewriteExpr->getType()),
1363 // The common base is emitted in the loop preheader. But since we
1364 // are reusing an IV, it has not been used to initialize the PHI node.
1365 // Add it to the expression used to rewrite the uses.
1366 if (!isa<ConstantInt>(CommonBaseV) ||
1367 !cast<ConstantInt>(CommonBaseV)->isZero())
1368 RewriteExpr = SE->getAddExpr(RewriteExpr,
1369 SE->getUnknown(CommonBaseV));
1372 // Now that we know what we need to do, insert code before User for the
1373 // immediate and any loop-variant expressions.
1374 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1375 // Add BaseV to the PHI value if needed.
1376 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1378 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1380 // Mark old value we replaced as possibly dead, so that it is elminated
1381 // if we just replaced the last use of that value.
1382 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1384 UsersToProcess.pop_back();
1387 // If there are any more users to process with the same base, process them
1388 // now. We sorted by base above, so we just have to check the last elt.
1389 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1390 // TODO: Next, find out which base index is the most common, pull it out.
1393 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1394 // different starting values, into different PHIs.
1397 /// FindIVForUser - If Cond has an operand that is an expression of an IV,
1398 /// set the IV user and stride information and return true, otherwise return
1400 bool LoopStrengthReduce::FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
1401 const SCEVHandle *&CondStride) {
1402 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1404 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1405 IVUsesByStride.find(StrideOrder[Stride]);
1406 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1408 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1409 E = SI->second.Users.end(); UI != E; ++UI)
1410 if (UI->User == Cond) {
1411 // NOTE: we could handle setcc instructions with multiple uses here, but
1412 // InstCombine does it as well for simple uses, it's not clear that it
1413 // occurs enough in real life to handle.
1415 CondStride = &SI->first;
1423 // Constant strides come first which in turns are sorted by their absolute
1424 // values. If absolute values are the same, then positive strides comes first.
1426 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1427 struct StrideCompare {
1428 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1429 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1430 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1432 int64_t LV = LHSC->getValue()->getSExtValue();
1433 int64_t RV = RHSC->getValue()->getSExtValue();
1434 uint64_t ALV = (LV < 0) ? -LV : LV;
1435 uint64_t ARV = (RV < 0) ? -RV : RV;
1441 return (LHSC && !RHSC);
1446 /// ChangeCompareStride - If a loop termination compare instruction is the
1447 /// only use of its stride, and the compaison is against a constant value,
1448 /// try eliminate the stride by moving the compare instruction to another
1449 /// stride and change its constant operand accordingly. e.g.
1455 /// if (v2 < 10) goto loop
1460 /// if (v1 < 30) goto loop
1461 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1462 IVStrideUse* &CondUse,
1463 const SCEVHandle* &CondStride) {
1464 if (StrideOrder.size() < 2 ||
1465 IVUsesByStride[*CondStride].Users.size() != 1)
1467 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1468 if (!SC) return Cond;
1469 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1470 if (!C) return Cond;
1472 ICmpInst::Predicate Predicate = Cond->getPredicate();
1473 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1474 int64_t CmpVal = C->getValue().getSExtValue();
1475 unsigned BitWidth = C->getValue().getBitWidth();
1476 uint64_t SignBit = 1ULL << (BitWidth-1);
1477 const Type *CmpTy = C->getType();
1478 const Type *NewCmpTy = NULL;
1479 int64_t NewCmpVal = CmpVal;
1480 SCEVHandle *NewStride = NULL;
1481 Value *NewIncV = NULL;
1484 // Look for a suitable stride / iv as replacement.
1485 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1486 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1487 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1488 IVUsesByStride.find(StrideOrder[i]);
1489 if (!isa<SCEVConstant>(SI->first))
1491 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1492 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1495 Scale = SSInt / CmpSSInt;
1496 NewCmpVal = CmpVal * Scale;
1497 APInt Mul = APInt(BitWidth, NewCmpVal);
1498 // Check for overflow.
1499 if (Mul.getSExtValue() != NewCmpVal) {
1504 // Watch out for overflow.
1505 if (ICmpInst::isSignedPredicate(Predicate) &&
1506 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1509 if (NewCmpVal != CmpVal) {
1510 // Pick the best iv to use trying to avoid a cast.
1512 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1513 E = SI->second.Users.end(); UI != E; ++UI) {
1514 NewIncV = UI->OperandValToReplace;
1515 if (NewIncV->getType() == CmpTy)
1523 NewCmpTy = NewIncV->getType();
1524 if (RequiresTypeConversion(CmpTy, NewCmpTy)) {
1529 bool AllUsesAreAddresses = true;
1530 std::vector<BasedUser> UsersToProcess;
1531 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1532 AllUsesAreAddresses,
1534 // Avoid rewriting the compare instruction with an iv of new stride
1535 // if it's likely the new stride uses will be rewritten using the
1536 if (AllUsesAreAddresses &&
1537 ValidStride(!isZero(CommonExprs), Scale, UsersToProcess)) {
1542 // If scale is negative, use inverse predicate unless it's testing
1544 if (Scale < 0 && !Cond->isEquality())
1545 Predicate = ICmpInst::getInversePredicate(Predicate);
1547 NewStride = &StrideOrder[i];
1552 if (NewCmpVal != CmpVal) {
1553 // Create a new compare instruction using new stride / iv.
1554 ICmpInst *OldCond = Cond;
1555 Value *RHS = ConstantInt::get(C->getType(), NewCmpVal);
1556 // Both sides of a ICmpInst must be of the same type.
1557 if (NewCmpTy != CmpTy) {
1558 if (isa<PointerType>(NewCmpTy) && !isa<PointerType>(CmpTy))
1559 RHS= SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1561 RHS = SCEVExpander::InsertCastOfTo(Instruction::BitCast, RHS, NewCmpTy);
1563 // Insert new compare instruction.
1564 Cond = new ICmpInst(Predicate, NewIncV, RHS);
1565 Cond->setName(L->getHeader()->getName() + ".termcond");
1566 OldCond->getParent()->getInstList().insert(OldCond, Cond);
1568 // Remove the old compare instruction. The old indvar is probably dead too.
1569 DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
1570 OldCond->replaceAllUsesWith(Cond);
1571 SE->deleteValueFromRecords(OldCond);
1572 OldCond->eraseFromParent();
1574 IVUsesByStride[*CondStride].Users.pop_back();
1575 SCEVHandle NewOffset = SE->getMulExpr(CondUse->Offset,
1576 SE->getConstant(ConstantInt::get(CondUse->Offset->getType(), Scale)));
1577 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1578 CondUse = &IVUsesByStride[*NewStride].Users.back();
1579 CondStride = NewStride;
1586 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1587 // uses in the loop, look to see if we can eliminate some, in favor of using
1588 // common indvars for the different uses.
1589 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1590 // TODO: implement optzns here.
1592 // Finally, get the terminating condition for the loop if possible. If we
1593 // can, we want to change it to use a post-incremented version of its
1594 // induction variable, to allow coalescing the live ranges for the IV into
1595 // one register value.
1596 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1597 BasicBlock *Preheader = L->getLoopPreheader();
1598 BasicBlock *LatchBlock =
1599 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1600 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1601 if (!TermBr || TermBr->isUnconditional() ||
1602 !isa<ICmpInst>(TermBr->getCondition()))
1604 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1606 // Search IVUsesByStride to find Cond's IVUse if there is one.
1607 IVStrideUse *CondUse = 0;
1608 const SCEVHandle *CondStride = 0;
1610 if (!FindIVForUser(Cond, CondUse, CondStride))
1611 return; // setcc doesn't use the IV.
1613 // If possible, change stride and operands of the compare instruction to
1614 // eliminate one stride.
1615 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
1617 // It's possible for the setcc instruction to be anywhere in the loop, and
1618 // possible for it to have multiple users. If it is not immediately before
1619 // the latch block branch, move it.
1620 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1621 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1622 Cond->moveBefore(TermBr);
1624 // Otherwise, clone the terminating condition and insert into the loopend.
1625 Cond = cast<ICmpInst>(Cond->clone());
1626 Cond->setName(L->getHeader()->getName() + ".termcond");
1627 LatchBlock->getInstList().insert(TermBr, Cond);
1629 // Clone the IVUse, as the old use still exists!
1630 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1631 CondUse->OperandValToReplace);
1632 CondUse = &IVUsesByStride[*CondStride].Users.back();
1636 // If we get to here, we know that we can transform the setcc instruction to
1637 // use the post-incremented version of the IV, allowing us to coalesce the
1638 // live ranges for the IV correctly.
1639 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
1640 CondUse->isUseOfPostIncrementedValue = true;
1643 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1645 LI = &getAnalysis<LoopInfo>();
1646 DT = &getAnalysis<DominatorTree>();
1647 SE = &getAnalysis<ScalarEvolution>();
1648 TD = &getAnalysis<TargetData>();
1649 UIntPtrTy = TD->getIntPtrType();
1651 // Find all uses of induction variables in this loop, and catagorize
1652 // them by stride. Start by finding all of the PHI nodes in the header for
1653 // this loop. If they are induction variables, inspect their uses.
1654 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
1655 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1656 AddUsersIfInteresting(I, L, Processed);
1658 // If we have nothing to do, return.
1659 if (IVUsesByStride.empty()) return false;
1661 // Optimize induction variables. Some indvar uses can be transformed to use
1662 // strides that will be needed for other purposes. A common example of this
1663 // is the exit test for the loop, which can often be rewritten to use the
1664 // computation of some other indvar to decide when to terminate the loop.
1668 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1669 // doing computation in byte values, promote to 32-bit values if safe.
1671 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1672 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1673 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1674 // to be careful that IV's are all the same type. Only works for intptr_t
1677 // If we only have one stride, we can more aggressively eliminate some things.
1678 bool HasOneStride = IVUsesByStride.size() == 1;
1681 DOUT << "\nLSR on ";
1685 // IVsByStride keeps IVs for one particular loop.
1686 IVsByStride.clear();
1688 // Sort the StrideOrder so we process larger strides first.
1689 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1691 // Note: this processes each stride/type pair individually. All users passed
1692 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1693 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1694 // This extra layer of indirection makes the ordering of strides deterministic
1695 // - not dependent on map order.
1696 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1697 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1698 IVUsesByStride.find(StrideOrder[Stride]);
1699 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1700 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1703 // Clean up after ourselves
1704 if (!DeadInsts.empty()) {
1705 DeleteTriviallyDeadInstructions(DeadInsts);
1707 BasicBlock::iterator I = L->getHeader()->begin();
1709 while ((PN = dyn_cast<PHINode>(I))) {
1710 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1712 // At this point, we know that we have killed one or more GEP
1713 // instructions. It is worth checking to see if the cann indvar is also
1714 // dead, so that we can remove it as well. The requirements for the cann
1715 // indvar to be considered dead are:
1716 // 1. the cann indvar has one use
1717 // 2. the use is an add instruction
1718 // 3. the add has one use
1719 // 4. the add is used by the cann indvar
1720 // If all four cases above are true, then we can remove both the add and
1722 // FIXME: this needs to eliminate an induction variable even if it's being
1723 // compared against some value to decide loop termination.
1724 if (PN->hasOneUse()) {
1725 Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
1726 if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
1727 if (BO->hasOneUse() && PN == *(BO->use_begin())) {
1728 DeadInsts.insert(BO);
1729 // Break the cycle, then delete the PHI.
1730 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1731 SE->deleteValueFromRecords(PN);
1732 PN->eraseFromParent();
1737 DeleteTriviallyDeadInstructions(DeadInsts);
1740 CastedPointers.clear();
1741 IVUsesByStride.clear();
1742 StrideOrder.clear();