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/Transforms/Utils/AddrModeMatcher.h"
30 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Support/CFG.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/GetElementPtrTypeIterator.h"
40 #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");
48 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
49 STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses");
51 static cl::opt<bool> EnableFullLSRMode("enable-full-lsr",
59 /// IVStrideUse - Keep track of one use of a strided induction variable, where
60 /// the stride is stored externally. The Offset member keeps track of the
61 /// offset from the IV, User is the actual user of the operand, and
62 /// 'OperandValToReplace' is the operand of the User that is the use.
63 struct VISIBILITY_HIDDEN IVStrideUse {
66 Value *OperandValToReplace;
68 // isUseOfPostIncrementedValue - True if this should use the
69 // post-incremented version of this IV, not the preincremented version.
70 // This can only be set in special cases, such as the terminating setcc
71 // instruction for a loop or uses dominated by the loop.
72 bool isUseOfPostIncrementedValue;
74 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
75 : Offset(Offs), User(U), OperandValToReplace(O),
76 isUseOfPostIncrementedValue(false) {}
79 /// IVUsersOfOneStride - This structure keeps track of all instructions that
80 /// have an operand that is based on the trip count multiplied by some stride.
81 /// The stride for all of these users is common and kept external to this
83 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
84 /// Users - Keep track of all of the users of this stride as well as the
85 /// initial value and the operand that uses the IV.
86 std::vector<IVStrideUse> Users;
88 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
89 Users.push_back(IVStrideUse(Offset, User, Operand));
93 /// IVInfo - This structure keeps track of one IV expression inserted during
94 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
95 /// well as the PHI node and increment value created for rewrite.
96 struct VISIBILITY_HIDDEN IVExpr {
102 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
104 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
107 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
108 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
109 struct VISIBILITY_HIDDEN IVsOfOneStride {
110 std::vector<IVExpr> IVs;
112 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
114 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
118 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
122 const TargetData *TD;
123 const Type *UIntPtrTy;
126 /// IVUsesByStride - Keep track of all uses of induction variables that we
127 /// are interested in. The key of the map is the stride of the access.
128 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
130 /// IVsByStride - Keep track of all IVs that have been inserted for a
131 /// particular stride.
132 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
134 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
135 /// We use this to iterate over the IVUsesByStride collection without being
136 /// dependent on random ordering of pointers in the process.
137 SmallVector<SCEVHandle, 16> StrideOrder;
139 /// GEPlist - A list of the GEP's that have been remembered in the SCEV
140 /// data structures. SCEV does not know to update these when the operands
141 /// of the GEP are changed, which means we cannot leave them live across
143 SmallVector<GetElementPtrInst *, 16> GEPlist;
145 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
146 /// of the casted version of each value. This is accessed by
147 /// getCastedVersionOf.
148 DenseMap<Value*, Value*> CastedPointers;
150 /// DeadInsts - Keep track of instructions we may have made dead, so that
151 /// we can remove them after we are done working.
152 SmallVector<Instruction*, 16> DeadInsts;
154 /// TLI - Keep a pointer of a TargetLowering to consult for determining
155 /// transformation profitability.
156 const TargetLowering *TLI;
159 static char ID; // Pass ID, replacement for typeid
160 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
161 LoopPass(&ID), TLI(tli) {
164 bool runOnLoop(Loop *L, LPPassManager &LPM);
166 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
167 // We split critical edges, so we change the CFG. However, we do update
168 // many analyses if they are around.
169 AU.addPreservedID(LoopSimplifyID);
170 AU.addPreserved<LoopInfo>();
171 AU.addPreserved<DominanceFrontier>();
172 AU.addPreserved<DominatorTree>();
174 AU.addRequiredID(LoopSimplifyID);
175 AU.addRequired<LoopInfo>();
176 AU.addRequired<DominatorTree>();
177 AU.addRequired<TargetData>();
178 AU.addRequired<ScalarEvolution>();
179 AU.addPreserved<ScalarEvolution>();
182 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
184 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
186 bool AddUsersIfInteresting(Instruction *I, Loop *L,
187 SmallPtrSet<Instruction*,16> &Processed);
188 SCEVHandle GetExpressionSCEV(Instruction *E);
189 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
190 IVStrideUse* &CondUse,
191 const SCEVHandle* &CondStride);
192 void OptimizeIndvars(Loop *L);
194 /// OptimizeShadowIV - If IV is used in a int-to-float cast
195 /// inside the loop then try to eliminate the cast opeation.
196 void OptimizeShadowIV(Loop *L);
198 /// OptimizeSMax - Rewrite the loop's terminating condition
199 /// if it uses an smax computation.
200 ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
201 IVStrideUse* &CondUse);
203 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
204 const SCEVHandle *&CondStride);
205 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
206 SCEVHandle CheckForIVReuse(bool, bool, bool, const SCEVHandle&,
207 IVExpr&, const Type*,
208 const std::vector<BasedUser>& UsersToProcess);
209 bool ValidStride(bool, int64_t,
210 const std::vector<BasedUser>& UsersToProcess);
211 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
212 IVUsersOfOneStride &Uses,
214 bool &AllUsesAreAddresses,
215 bool &AllUsesAreOutsideLoop,
216 std::vector<BasedUser> &UsersToProcess);
217 bool ShouldUseFullStrengthReductionMode(
218 const std::vector<BasedUser> &UsersToProcess,
220 bool AllUsesAreAddresses,
222 void PrepareToStrengthReduceFully(
223 std::vector<BasedUser> &UsersToProcess,
225 SCEVHandle CommonExprs,
227 SCEVExpander &PreheaderRewriter);
228 void PrepareToStrengthReduceFromSmallerStride(
229 std::vector<BasedUser> &UsersToProcess,
231 const IVExpr &ReuseIV,
232 Instruction *PreInsertPt);
233 void PrepareToStrengthReduceWithNewPhi(
234 std::vector<BasedUser> &UsersToProcess,
236 SCEVHandle CommonExprs,
239 SCEVExpander &PreheaderRewriter);
240 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
241 IVUsersOfOneStride &Uses,
243 void DeleteTriviallyDeadInstructions();
247 char LoopStrengthReduce::ID = 0;
248 static RegisterPass<LoopStrengthReduce>
249 X("loop-reduce", "Loop Strength Reduction");
251 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
252 return new LoopStrengthReduce(TLI);
255 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
256 /// assumes that the Value* V is of integer or pointer type only.
258 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
260 if (V->getType() == UIntPtrTy) return V;
261 if (Constant *CB = dyn_cast<Constant>(V))
262 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
264 Value *&New = CastedPointers[V];
267 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
268 DeadInsts.push_back(cast<Instruction>(New));
273 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
274 /// specified set are trivially dead, delete them and see if this makes any of
275 /// their operands subsequently dead.
276 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
277 if (DeadInsts.empty()) return;
279 // Sort the deadinsts list so that we can trivially eliminate duplicates as we
280 // go. The code below never adds a non-dead instruction to the worklist, but
281 // callers may not be so careful.
282 array_pod_sort(DeadInsts.begin(), DeadInsts.end());
284 // Drop duplicate instructions and those with uses.
285 for (unsigned i = 0, e = DeadInsts.size()-1; i < e; ++i) {
286 Instruction *I = DeadInsts[i];
287 if (!I->use_empty()) DeadInsts[i] = 0;
288 while (i != e && DeadInsts[i+1] == I)
292 while (!DeadInsts.empty()) {
293 Instruction *I = DeadInsts.back();
294 DeadInsts.pop_back();
296 if (I == 0 || !isInstructionTriviallyDead(I))
299 SE->deleteValueFromRecords(I);
301 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
302 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
305 DeadInsts.push_back(U);
309 I->eraseFromParent();
315 /// GetExpressionSCEV - Compute and return the SCEV for the specified
317 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
318 // Pointer to pointer bitcast instructions return the same value as their
320 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
321 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
322 return SE->getSCEV(BCI);
323 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
328 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
329 // If this is a GEP that SE doesn't know about, compute it now and insert it.
330 // If this is not a GEP, or if we have already done this computation, just let
332 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
333 if (!GEP || SE->hasSCEV(GEP))
334 return SE->getSCEV(Exp);
336 // Analyze all of the subscripts of this getelementptr instruction, looking
337 // for uses that are determined by the trip count of the loop. First, skip
338 // all operands the are not dependent on the IV.
340 // Build up the base expression. Insert an LLVM cast of the pointer to
342 SCEVHandle GEPVal = SE->getUnknown(
343 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
345 gep_type_iterator GTI = gep_type_begin(GEP);
347 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
348 i != e; ++i, ++GTI) {
349 // If this is a use of a recurrence that we can analyze, and it comes before
350 // Op does in the GEP operand list, we will handle this when we process this
352 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
353 const StructLayout *SL = TD->getStructLayout(STy);
354 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
355 uint64_t Offset = SL->getElementOffset(Idx);
356 GEPVal = SE->getAddExpr(GEPVal,
357 SE->getIntegerSCEV(Offset, UIntPtrTy));
359 unsigned GEPOpiBits =
360 (*i)->getType()->getPrimitiveSizeInBits();
361 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
362 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
363 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
364 Instruction::BitCast));
365 Value *OpVal = getCastedVersionOf(opcode, *i);
366 SCEVHandle Idx = SE->getSCEV(OpVal);
368 uint64_t TypeSize = TD->getTypePaddedSize(GTI.getIndexedType());
370 Idx = SE->getMulExpr(Idx,
371 SE->getConstant(ConstantInt::get(UIntPtrTy,
373 GEPVal = SE->getAddExpr(GEPVal, Idx);
377 SE->setSCEV(GEP, GEPVal);
378 GEPlist.push_back(GEP);
382 /// containsAddRecFromDifferentLoop - Determine whether expression S involves a
383 /// subexpression that is an AddRec from a loop other than L. An outer loop
384 /// of L is OK, but not an inner loop nor a disjoint loop.
385 static bool containsAddRecFromDifferentLoop(SCEVHandle S, Loop *L) {
386 // This is very common, put it first.
387 if (isa<SCEVConstant>(S))
389 if (SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) {
390 for (unsigned int i=0; i< AE->getNumOperands(); i++)
391 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L))
395 if (SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) {
396 if (const Loop *newLoop = AE->getLoop()) {
399 // if newLoop is an outer loop of L, this is OK.
400 if (!LoopInfoBase<BasicBlock>::isNotAlreadyContainedIn(L, newLoop))
405 if (SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S))
406 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
407 containsAddRecFromDifferentLoop(DE->getRHS(), L);
409 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll
410 // need this when it is.
411 if (SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S))
412 return containsAddRecFromDifferentLoop(DE->getLHS(), L) ||
413 containsAddRecFromDifferentLoop(DE->getRHS(), L);
415 if (SCEVTruncateExpr *TE = dyn_cast<SCEVTruncateExpr>(S))
416 return containsAddRecFromDifferentLoop(TE->getOperand(), L);
417 if (SCEVZeroExtendExpr *ZE = dyn_cast<SCEVZeroExtendExpr>(S))
418 return containsAddRecFromDifferentLoop(ZE->getOperand(), L);
419 if (SCEVSignExtendExpr *SE = dyn_cast<SCEVSignExtendExpr>(S))
420 return containsAddRecFromDifferentLoop(SE->getOperand(), L);
424 /// getSCEVStartAndStride - Compute the start and stride of this expression,
425 /// returning false if the expression is not a start/stride pair, or true if it
426 /// is. The stride must be a loop invariant expression, but the start may be
427 /// a mix of loop invariant and loop variant expressions. The start cannot,
428 /// however, contain an AddRec from a different loop, unless that loop is an
429 /// outer loop of the current loop.
430 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
431 SCEVHandle &Start, SCEVHandle &Stride,
432 ScalarEvolution *SE, DominatorTree *DT) {
433 SCEVHandle TheAddRec = Start; // Initialize to zero.
435 // If the outer level is an AddExpr, the operands are all start values except
436 // for a nested AddRecExpr.
437 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
438 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
439 if (SCEVAddRecExpr *AddRec =
440 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
441 if (AddRec->getLoop() == L)
442 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
444 return false; // Nested IV of some sort?
446 Start = SE->getAddExpr(Start, AE->getOperand(i));
449 } else if (isa<SCEVAddRecExpr>(SH)) {
452 return false; // not analyzable.
455 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
456 if (!AddRec || AddRec->getLoop() != L) return false;
458 // FIXME: Generalize to non-affine IV's.
459 if (!AddRec->isAffine()) return false;
461 // If Start contains an SCEVAddRecExpr from a different loop, other than an
462 // outer loop of the current loop, reject it. SCEV has no concept of
463 // operating on more than one loop at a time so don't confuse it with such
465 if (containsAddRecFromDifferentLoop(AddRec->getOperand(0), L))
468 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
470 if (!isa<SCEVConstant>(AddRec->getOperand(1))) {
471 // If stride is an instruction, make sure it dominates the loop preheader.
472 // Otherwise we could end up with a use before def situation.
473 BasicBlock *Preheader = L->getLoopPreheader();
474 if (!AddRec->getOperand(1)->dominates(Preheader, DT))
477 DOUT << "[" << L->getHeader()->getName()
478 << "] Variable stride: " << *AddRec << "\n";
481 Stride = AddRec->getOperand(1);
485 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
486 /// and now we need to decide whether the user should use the preinc or post-inc
487 /// value. If this user should use the post-inc version of the IV, return true.
489 /// Choosing wrong here can break dominance properties (if we choose to use the
490 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
491 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
492 /// should use the post-inc value).
493 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
494 Loop *L, DominatorTree *DT, Pass *P,
495 SmallVectorImpl<Instruction*> &DeadInsts){
496 // If the user is in the loop, use the preinc value.
497 if (L->contains(User->getParent())) return false;
499 BasicBlock *LatchBlock = L->getLoopLatch();
501 // Ok, the user is outside of the loop. If it is dominated by the latch
502 // block, use the post-inc value.
503 if (DT->dominates(LatchBlock, User->getParent()))
506 // There is one case we have to be careful of: PHI nodes. These little guys
507 // can live in blocks that do not dominate the latch block, but (since their
508 // uses occur in the predecessor block, not the block the PHI lives in) should
509 // still use the post-inc value. Check for this case now.
510 PHINode *PN = dyn_cast<PHINode>(User);
511 if (!PN) return false; // not a phi, not dominated by latch block.
513 // Look at all of the uses of IV by the PHI node. If any use corresponds to
514 // a block that is not dominated by the latch block, give up and use the
515 // preincremented value.
516 unsigned NumUses = 0;
517 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
518 if (PN->getIncomingValue(i) == IV) {
520 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
524 // Okay, all uses of IV by PN are in predecessor blocks that really are
525 // dominated by the latch block. Split the critical edges and use the
526 // post-incremented value.
527 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
528 if (PN->getIncomingValue(i) == IV) {
529 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
530 // Splitting the critical edge can reduce the number of entries in this
532 e = PN->getNumIncomingValues();
533 if (--NumUses == 0) break;
536 // PHI node might have become a constant value after SplitCriticalEdge.
537 DeadInsts.push_back(User);
542 /// isAddressUse - Returns true if the specified instruction is using the
543 /// specified value as an address.
544 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
545 bool isAddress = isa<LoadInst>(Inst);
546 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
547 if (SI->getOperand(1) == OperandVal)
549 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
550 // Addressing modes can also be folded into prefetches and a variety
552 switch (II->getIntrinsicID()) {
554 case Intrinsic::prefetch:
555 case Intrinsic::x86_sse2_loadu_dq:
556 case Intrinsic::x86_sse2_loadu_pd:
557 case Intrinsic::x86_sse_loadu_ps:
558 case Intrinsic::x86_sse_storeu_ps:
559 case Intrinsic::x86_sse2_storeu_pd:
560 case Intrinsic::x86_sse2_storeu_dq:
561 case Intrinsic::x86_sse2_storel_dq:
562 if (II->getOperand(1) == OperandVal)
570 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
571 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
572 /// return true. Otherwise, return false.
573 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
574 SmallPtrSet<Instruction*,16> &Processed) {
575 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
576 return false; // Void and FP expressions cannot be reduced.
577 if (!Processed.insert(I))
578 return true; // Instruction already handled.
580 // Get the symbolic expression for this instruction.
581 SCEVHandle ISE = GetExpressionSCEV(I);
582 if (isa<SCEVCouldNotCompute>(ISE)) return false;
584 // Get the start and stride for this expression.
585 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
586 SCEVHandle Stride = Start;
587 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE, DT))
588 return false; // Non-reducible symbolic expression, bail out.
590 std::vector<Instruction *> IUsers;
591 // Collect all I uses now because IVUseShouldUsePostIncValue may
592 // invalidate use_iterator.
593 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
594 IUsers.push_back(cast<Instruction>(*UI));
596 for (unsigned iused_index = 0, iused_size = IUsers.size();
597 iused_index != iused_size; ++iused_index) {
599 Instruction *User = IUsers[iused_index];
601 // Do not infinitely recurse on PHI nodes.
602 if (isa<PHINode>(User) && Processed.count(User))
605 // Descend recursively, but not into PHI nodes outside the current loop.
606 // It's important to see the entire expression outside the loop to get
607 // choices that depend on addressing mode use right, although we won't
608 // consider references ouside the loop in all cases.
609 // If User is already in Processed, we don't want to recurse into it again,
610 // but do want to record a second reference in the same instruction.
611 bool AddUserToIVUsers = false;
612 if (LI->getLoopFor(User->getParent()) != L) {
613 if (isa<PHINode>(User) || Processed.count(User) ||
614 !AddUsersIfInteresting(User, L, Processed)) {
615 DOUT << "FOUND USER in other loop: " << *User
616 << " OF SCEV: " << *ISE << "\n";
617 AddUserToIVUsers = true;
619 } else if (Processed.count(User) ||
620 !AddUsersIfInteresting(User, L, Processed)) {
621 DOUT << "FOUND USER: " << *User
622 << " OF SCEV: " << *ISE << "\n";
623 AddUserToIVUsers = true;
626 if (AddUserToIVUsers) {
627 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
628 if (StrideUses.Users.empty()) // First occurrence of this stride?
629 StrideOrder.push_back(Stride);
631 // Okay, we found a user that we cannot reduce. Analyze the instruction
632 // and decide what to do with it. If we are a use inside of the loop, use
633 // the value before incrementation, otherwise use it after incrementation.
634 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
635 // The value used will be incremented by the stride more than we are
636 // expecting, so subtract this off.
637 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
638 StrideUses.addUser(NewStart, User, I);
639 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
640 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
642 StrideUses.addUser(Start, User, I);
650 /// BasedUser - For a particular base value, keep information about how we've
651 /// partitioned the expression so far.
653 /// SE - The current ScalarEvolution object.
656 /// Base - The Base value for the PHI node that needs to be inserted for
657 /// this use. As the use is processed, information gets moved from this
658 /// field to the Imm field (below). BasedUser values are sorted by this
662 /// Inst - The instruction using the induction variable.
665 /// OperandValToReplace - The operand value of Inst to replace with the
667 Value *OperandValToReplace;
669 /// Imm - The immediate value that should be added to the base immediately
670 /// before Inst, because it will be folded into the imm field of the
671 /// instruction. This is also sometimes used for loop-variant values that
672 /// must be added inside the loop.
675 /// Phi - The induction variable that performs the striding that
676 /// should be used for this user.
679 /// IncV - The post-incremented value of Phi.
682 // isUseOfPostIncrementedValue - True if this should use the
683 // post-incremented version of this IV, not the preincremented version.
684 // This can only be set in special cases, such as the terminating setcc
685 // instruction for a loop and uses outside the loop that are dominated by
687 bool isUseOfPostIncrementedValue;
689 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
690 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
691 OperandValToReplace(IVSU.OperandValToReplace),
692 Imm(SE->getIntegerSCEV(0, Base->getType())),
693 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
695 // Once we rewrite the code to insert the new IVs we want, update the
696 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
698 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
699 Instruction *InsertPt,
700 SCEVExpander &Rewriter, Loop *L, Pass *P,
701 SmallVectorImpl<Instruction*> &DeadInsts);
703 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
704 SCEVExpander &Rewriter,
705 Instruction *IP, Loop *L);
710 void BasedUser::dump() const {
711 cerr << " Base=" << *Base;
712 cerr << " Imm=" << *Imm;
713 cerr << " Inst: " << *Inst;
716 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
717 SCEVExpander &Rewriter,
718 Instruction *IP, Loop *L) {
719 // Figure out where we *really* want to insert this code. In particular, if
720 // the user is inside of a loop that is nested inside of L, we really don't
721 // want to insert this expression before the user, we'd rather pull it out as
722 // many loops as possible.
723 LoopInfo &LI = Rewriter.getLoopInfo();
724 Instruction *BaseInsertPt = IP;
726 // Figure out the most-nested loop that IP is in.
727 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
729 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
730 // the preheader of the outer-most loop where NewBase is not loop invariant.
731 if (L->contains(IP->getParent()))
732 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
733 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
734 InsertLoop = InsertLoop->getParentLoop();
737 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
739 // If there is no immediate value, skip the next part.
743 // If we are inserting the base and imm values in the same block, make sure to
744 // adjust the IP position if insertion reused a result.
745 if (IP == BaseInsertPt)
746 IP = Rewriter.getInsertionPoint();
748 // Always emit the immediate (if non-zero) into the same block as the user.
749 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
750 return Rewriter.expandCodeFor(NewValSCEV, IP);
755 // Once we rewrite the code to insert the new IVs we want, update the
756 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
757 // to it. NewBasePt is the last instruction which contributes to the
758 // value of NewBase in the case that it's a diffferent instruction from
759 // the PHI that NewBase is computed from, or null otherwise.
761 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
762 Instruction *NewBasePt,
763 SCEVExpander &Rewriter, Loop *L, Pass *P,
764 SmallVectorImpl<Instruction*> &DeadInsts){
765 if (!isa<PHINode>(Inst)) {
766 // By default, insert code at the user instruction.
767 BasicBlock::iterator InsertPt = Inst;
769 // However, if the Operand is itself an instruction, the (potentially
770 // complex) inserted code may be shared by many users. Because of this, we
771 // want to emit code for the computation of the operand right before its old
772 // computation. This is usually safe, because we obviously used to use the
773 // computation when it was computed in its current block. However, in some
774 // cases (e.g. use of a post-incremented induction variable) the NewBase
775 // value will be pinned to live somewhere after the original computation.
776 // In this case, we have to back off.
778 // If this is a use outside the loop (which means after, since it is based
779 // on a loop indvar) we use the post-incremented value, so that we don't
780 // artificially make the preinc value live out the bottom of the loop.
781 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
782 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
783 InsertPt = NewBasePt;
785 } else if (Instruction *OpInst
786 = dyn_cast<Instruction>(OperandValToReplace)) {
788 while (isa<PHINode>(InsertPt)) ++InsertPt;
791 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
792 // Adjust the type back to match the Inst. Note that we can't use InsertPt
793 // here because the SCEVExpander may have inserted the instructions after
794 // that point, in its efforts to avoid inserting redundant expressions.
795 if (isa<PointerType>(OperandValToReplace->getType())) {
796 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
798 OperandValToReplace->getType());
800 // Replace the use of the operand Value with the new Phi we just created.
801 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
803 DOUT << " Replacing with ";
804 DEBUG(WriteAsOperand(*DOUT, NewVal, /*PrintType=*/false));
805 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
809 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
810 // expression into each operand block that uses it. Note that PHI nodes can
811 // have multiple entries for the same predecessor. We use a map to make sure
812 // that a PHI node only has a single Value* for each predecessor (which also
813 // prevents us from inserting duplicate code in some blocks).
814 DenseMap<BasicBlock*, Value*> InsertedCode;
815 PHINode *PN = cast<PHINode>(Inst);
816 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
817 if (PN->getIncomingValue(i) == OperandValToReplace) {
818 // If the original expression is outside the loop, put the replacement
819 // code in the same place as the original expression,
820 // which need not be an immediate predecessor of this PHI. This way we
821 // need only one copy of it even if it is referenced multiple times in
822 // the PHI. We don't do this when the original expression is inside the
823 // loop because multiple copies sometimes do useful sinking of code in
825 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
826 if (L->contains(OldLoc->getParent())) {
827 // If this is a critical edge, split the edge so that we do not insert
828 // the code on all predecessor/successor paths. We do this unless this
829 // is the canonical backedge for this loop, as this can make some
830 // inserted code be in an illegal position.
831 BasicBlock *PHIPred = PN->getIncomingBlock(i);
832 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
833 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
835 // First step, split the critical edge.
836 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
838 // Next step: move the basic block. In particular, if the PHI node
839 // is outside of the loop, and PredTI is in the loop, we want to
840 // move the block to be immediately before the PHI block, not
841 // immediately after PredTI.
842 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
843 BasicBlock *NewBB = PN->getIncomingBlock(i);
844 NewBB->moveBefore(PN->getParent());
847 // Splitting the edge can reduce the number of PHI entries we have.
848 e = PN->getNumIncomingValues();
851 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
853 // Insert the code into the end of the predecessor block.
854 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
855 PN->getIncomingBlock(i)->getTerminator() :
856 OldLoc->getParent()->getTerminator();
857 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
859 // Adjust the type back to match the PHI. Note that we can't use
860 // InsertPt here because the SCEVExpander may have inserted its
861 // instructions after that point, in its efforts to avoid inserting
862 // redundant expressions.
863 if (isa<PointerType>(PN->getType())) {
864 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
869 DOUT << " Changing PHI use to ";
870 DEBUG(WriteAsOperand(*DOUT, Code, /*PrintType=*/false));
871 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
874 // Replace the use of the operand Value with the new Phi we just created.
875 PN->setIncomingValue(i, Code);
880 // PHI node might have become a constant value after SplitCriticalEdge.
881 DeadInsts.push_back(Inst);
885 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
886 /// mode, and does not need to be put in a register first.
887 static bool fitsInAddressMode(const SCEVHandle &V, const Type *UseTy,
888 const TargetLowering *TLI, bool HasBaseReg) {
889 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
890 int64_t VC = SC->getValue()->getSExtValue();
892 TargetLowering::AddrMode AM;
894 AM.HasBaseReg = HasBaseReg;
895 return TLI->isLegalAddressingMode(AM, UseTy);
897 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
898 return (VC > -(1 << 16) && VC < (1 << 16)-1);
902 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
903 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
904 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
905 Constant *Op0 = CE->getOperand(0);
906 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
907 TargetLowering::AddrMode AM;
909 AM.HasBaseReg = HasBaseReg;
910 return TLI->isLegalAddressingMode(AM, UseTy);
916 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
917 /// loop varying to the Imm operand.
918 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
919 Loop *L, ScalarEvolution *SE) {
920 if (Val->isLoopInvariant(L)) return; // Nothing to do.
922 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
923 std::vector<SCEVHandle> NewOps;
924 NewOps.reserve(SAE->getNumOperands());
926 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
927 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
928 // If this is a loop-variant expression, it must stay in the immediate
929 // field of the expression.
930 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
932 NewOps.push_back(SAE->getOperand(i));
936 Val = SE->getIntegerSCEV(0, Val->getType());
938 Val = SE->getAddExpr(NewOps);
939 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
940 // Try to pull immediates out of the start value of nested addrec's.
941 SCEVHandle Start = SARE->getStart();
942 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
944 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
946 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
948 // Otherwise, all of Val is variant, move the whole thing over.
949 Imm = SE->getAddExpr(Imm, Val);
950 Val = SE->getIntegerSCEV(0, Val->getType());
955 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
956 /// that can fit into the immediate field of instructions in the target.
957 /// Accumulate these immediate values into the Imm value.
958 static void MoveImmediateValues(const TargetLowering *TLI,
960 SCEVHandle &Val, SCEVHandle &Imm,
961 bool isAddress, Loop *L,
962 ScalarEvolution *SE) {
963 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
964 std::vector<SCEVHandle> NewOps;
965 NewOps.reserve(SAE->getNumOperands());
967 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
968 SCEVHandle NewOp = SAE->getOperand(i);
969 MoveImmediateValues(TLI, UseTy, NewOp, Imm, isAddress, L, SE);
971 if (!NewOp->isLoopInvariant(L)) {
972 // If this is a loop-variant expression, it must stay in the immediate
973 // field of the expression.
974 Imm = SE->getAddExpr(Imm, NewOp);
976 NewOps.push_back(NewOp);
981 Val = SE->getIntegerSCEV(0, Val->getType());
983 Val = SE->getAddExpr(NewOps);
985 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
986 // Try to pull immediates out of the start value of nested addrec's.
987 SCEVHandle Start = SARE->getStart();
988 MoveImmediateValues(TLI, UseTy, Start, Imm, isAddress, L, SE);
990 if (Start != SARE->getStart()) {
991 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
993 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
996 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
997 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
998 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
999 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
1001 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
1002 SCEVHandle NewOp = SME->getOperand(1);
1003 MoveImmediateValues(TLI, UseTy, NewOp, SubImm, isAddress, L, SE);
1005 // If we extracted something out of the subexpressions, see if we can
1007 if (NewOp != SME->getOperand(1)) {
1008 // Scale SubImm up by "8". If the result is a target constant, we are
1010 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
1011 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
1012 // Accumulate the immediate.
1013 Imm = SE->getAddExpr(Imm, SubImm);
1015 // Update what is left of 'Val'.
1016 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
1023 // Loop-variant expressions must stay in the immediate field of the
1025 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
1026 !Val->isLoopInvariant(L)) {
1027 Imm = SE->getAddExpr(Imm, Val);
1028 Val = SE->getIntegerSCEV(0, Val->getType());
1032 // Otherwise, no immediates to move.
1035 static void MoveImmediateValues(const TargetLowering *TLI,
1037 SCEVHandle &Val, SCEVHandle &Imm,
1038 bool isAddress, Loop *L,
1039 ScalarEvolution *SE) {
1040 const Type *UseTy = User->getType();
1041 if (StoreInst *SI = dyn_cast<StoreInst>(User))
1042 UseTy = SI->getOperand(0)->getType();
1043 MoveImmediateValues(TLI, UseTy, Val, Imm, isAddress, L, SE);
1046 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
1047 /// added together. This is used to reassociate common addition subexprs
1048 /// together for maximal sharing when rewriting bases.
1049 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
1051 ScalarEvolution *SE) {
1052 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
1053 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
1054 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
1055 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
1056 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
1057 if (SARE->getOperand(0) == Zero) {
1058 SubExprs.push_back(Expr);
1060 // Compute the addrec with zero as its base.
1061 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
1062 Ops[0] = Zero; // Start with zero base.
1063 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
1066 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
1068 } else if (!Expr->isZero()) {
1070 SubExprs.push_back(Expr);
1074 // This is logically local to the following function, but C++ says we have
1075 // to make it file scope.
1076 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
1078 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
1079 /// the Uses, removing any common subexpressions, except that if all such
1080 /// subexpressions can be folded into an addressing mode for all uses inside
1081 /// the loop (this case is referred to as "free" in comments herein) we do
1082 /// not remove anything. This looks for things like (a+b+c) and
1083 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
1084 /// is *removed* from the Bases and returned.
1086 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
1087 ScalarEvolution *SE, Loop *L,
1088 const TargetLowering *TLI) {
1089 unsigned NumUses = Uses.size();
1091 // Only one use? This is a very common case, so we handle it specially and
1093 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
1094 SCEVHandle Result = Zero;
1095 SCEVHandle FreeResult = Zero;
1097 // If the use is inside the loop, use its base, regardless of what it is:
1098 // it is clearly shared across all the IV's. If the use is outside the loop
1099 // (which means after it) we don't want to factor anything *into* the loop,
1100 // so just use 0 as the base.
1101 if (L->contains(Uses[0].Inst->getParent()))
1102 std::swap(Result, Uses[0].Base);
1106 // To find common subexpressions, count how many of Uses use each expression.
1107 // If any subexpressions are used Uses.size() times, they are common.
1108 // Also track whether all uses of each expression can be moved into an
1109 // an addressing mode "for free"; such expressions are left within the loop.
1110 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
1111 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
1113 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
1114 // order we see them.
1115 std::vector<SCEVHandle> UniqueSubExprs;
1117 std::vector<SCEVHandle> SubExprs;
1118 unsigned NumUsesInsideLoop = 0;
1119 for (unsigned i = 0; i != NumUses; ++i) {
1120 // If the user is outside the loop, just ignore it for base computation.
1121 // Since the user is outside the loop, it must be *after* the loop (if it
1122 // were before, it could not be based on the loop IV). We don't want users
1123 // after the loop to affect base computation of values *inside* the loop,
1124 // because we can always add their offsets to the result IV after the loop
1125 // is done, ensuring we get good code inside the loop.
1126 if (!L->contains(Uses[i].Inst->getParent()))
1128 NumUsesInsideLoop++;
1130 // If the base is zero (which is common), return zero now, there are no
1131 // CSEs we can find.
1132 if (Uses[i].Base == Zero) return Zero;
1134 // If this use is as an address we may be able to put CSEs in the addressing
1135 // mode rather than hoisting them.
1136 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1137 // We may need the UseTy below, but only when isAddrUse, so compute it
1138 // only in that case.
1139 const Type *UseTy = 0;
1141 UseTy = Uses[i].Inst->getType();
1142 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1143 UseTy = SI->getOperand(0)->getType();
1146 // Split the expression into subexprs.
1147 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1148 // Add one to SubExpressionUseData.Count for each subexpr present, and
1149 // if the subexpr is not a valid immediate within an addressing mode use,
1150 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1151 // hoist these out of the loop (if they are common to all uses).
1152 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1153 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1154 UniqueSubExprs.push_back(SubExprs[j]);
1155 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1156 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1161 // Now that we know how many times each is used, build Result. Iterate over
1162 // UniqueSubexprs so that we have a stable ordering.
1163 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1164 std::map<SCEVHandle, SubExprUseData>::iterator I =
1165 SubExpressionUseData.find(UniqueSubExprs[i]);
1166 assert(I != SubExpressionUseData.end() && "Entry not found?");
1167 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1168 if (I->second.notAllUsesAreFree)
1169 Result = SE->getAddExpr(Result, I->first);
1171 FreeResult = SE->getAddExpr(FreeResult, I->first);
1173 // Remove non-cse's from SubExpressionUseData.
1174 SubExpressionUseData.erase(I);
1177 if (FreeResult != Zero) {
1178 // We have some subexpressions that can be subsumed into addressing
1179 // modes in every use inside the loop. However, it's possible that
1180 // there are so many of them that the combined FreeResult cannot
1181 // be subsumed, or that the target cannot handle both a FreeResult
1182 // and a Result in the same instruction (for example because it would
1183 // require too many registers). Check this.
1184 for (unsigned i=0; i<NumUses; ++i) {
1185 if (!L->contains(Uses[i].Inst->getParent()))
1187 // We know this is an addressing mode use; if there are any uses that
1188 // are not, FreeResult would be Zero.
1189 const Type *UseTy = Uses[i].Inst->getType();
1190 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1191 UseTy = SI->getOperand(0)->getType();
1192 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1193 // FIXME: could split up FreeResult into pieces here, some hoisted
1194 // and some not. There is no obvious advantage to this.
1195 Result = SE->getAddExpr(Result, FreeResult);
1202 // If we found no CSE's, return now.
1203 if (Result == Zero) return Result;
1205 // If we still have a FreeResult, remove its subexpressions from
1206 // SubExpressionUseData. This means they will remain in the use Bases.
1207 if (FreeResult != Zero) {
1208 SeparateSubExprs(SubExprs, FreeResult, SE);
1209 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1210 std::map<SCEVHandle, SubExprUseData>::iterator I =
1211 SubExpressionUseData.find(SubExprs[j]);
1212 SubExpressionUseData.erase(I);
1217 // Otherwise, remove all of the CSE's we found from each of the base values.
1218 for (unsigned i = 0; i != NumUses; ++i) {
1219 // Uses outside the loop don't necessarily include the common base, but
1220 // the final IV value coming into those uses does. Instead of trying to
1221 // remove the pieces of the common base, which might not be there,
1222 // subtract off the base to compensate for this.
1223 if (!L->contains(Uses[i].Inst->getParent())) {
1224 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1228 // Split the expression into subexprs.
1229 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1231 // Remove any common subexpressions.
1232 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1233 if (SubExpressionUseData.count(SubExprs[j])) {
1234 SubExprs.erase(SubExprs.begin()+j);
1238 // Finally, add the non-shared expressions together.
1239 if (SubExprs.empty())
1240 Uses[i].Base = Zero;
1242 Uses[i].Base = SE->getAddExpr(SubExprs);
1249 /// ValidStride - Check whether the given Scale is valid for all loads and
1250 /// stores in UsersToProcess.
1252 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1254 const std::vector<BasedUser>& UsersToProcess) {
1258 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1259 // If this is a load or other access, pass the type of the access in.
1260 const Type *AccessTy = Type::VoidTy;
1261 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1262 AccessTy = SI->getOperand(0)->getType();
1263 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1264 AccessTy = LI->getType();
1265 else if (isa<PHINode>(UsersToProcess[i].Inst))
1268 TargetLowering::AddrMode AM;
1269 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1270 AM.BaseOffs = SC->getValue()->getSExtValue();
1271 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1274 // If load[imm+r*scale] is illegal, bail out.
1275 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1281 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
1283 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1287 if (Ty1->canLosslesslyBitCastTo(Ty2))
1289 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1291 if (isa<PointerType>(Ty2) && Ty1->canLosslesslyBitCastTo(UIntPtrTy))
1293 if (isa<PointerType>(Ty1) && Ty2->canLosslesslyBitCastTo(UIntPtrTy))
1298 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1299 /// of a previous stride and it is a legal value for the target addressing
1300 /// mode scale component and optional base reg. This allows the users of
1301 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1302 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1304 /// If all uses are outside the loop, we don't require that all multiplies
1305 /// be folded into the addressing mode, nor even that the factor be constant;
1306 /// a multiply (executed once) outside the loop is better than another IV
1307 /// within. Well, usually.
1308 SCEVHandle LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1309 bool AllUsesAreAddresses,
1310 bool AllUsesAreOutsideLoop,
1311 const SCEVHandle &Stride,
1312 IVExpr &IV, const Type *Ty,
1313 const std::vector<BasedUser>& UsersToProcess) {
1314 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1315 int64_t SInt = SC->getValue()->getSExtValue();
1316 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1318 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1319 IVsByStride.find(StrideOrder[NewStride]);
1320 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1322 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1323 if (SI->first != Stride &&
1324 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1326 int64_t Scale = SInt / SSInt;
1327 // Check that this stride is valid for all the types used for loads and
1328 // stores; if it can be used for some and not others, we might as well use
1329 // the original stride everywhere, since we have to create the IV for it
1330 // anyway. If the scale is 1, then we don't need to worry about folding
1333 (AllUsesAreAddresses &&
1334 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1335 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1336 IE = SI->second.IVs.end(); II != IE; ++II)
1337 // FIXME: Only handle base == 0 for now.
1338 // Only reuse previous IV if it would not require a type conversion.
1339 if (II->Base->isZero() &&
1340 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1342 return SE->getIntegerSCEV(Scale, Stride->getType());
1345 } else if (AllUsesAreOutsideLoop) {
1346 // Accept nonconstant strides here; it is really really right to substitute
1347 // an existing IV if we can.
1348 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1350 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1351 IVsByStride.find(StrideOrder[NewStride]);
1352 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1354 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1355 if (SI->first != Stride && SSInt != 1)
1357 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1358 IE = SI->second.IVs.end(); II != IE; ++II)
1359 // Accept nonzero base here.
1360 // Only reuse previous IV if it would not require a type conversion.
1361 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1366 // Special case, old IV is -1*x and this one is x. Can treat this one as
1368 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1370 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1371 IVsByStride.find(StrideOrder[NewStride]);
1372 if (SI == IVsByStride.end())
1374 if (SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1375 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1376 if (Stride == ME->getOperand(1) &&
1377 SC->getValue()->getSExtValue() == -1LL)
1378 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1379 IE = SI->second.IVs.end(); II != IE; ++II)
1380 // Accept nonzero base here.
1381 // Only reuse previous IV if it would not require type conversion.
1382 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1384 return SE->getIntegerSCEV(-1LL, Stride->getType());
1388 return SE->getIntegerSCEV(0, Stride->getType());
1391 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1392 /// returns true if Val's isUseOfPostIncrementedValue is true.
1393 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1394 return Val.isUseOfPostIncrementedValue;
1397 /// isNonConstantNegative - Return true if the specified scev is negated, but
1399 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1400 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1401 if (!Mul) return false;
1403 // If there is a constant factor, it will be first.
1404 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1405 if (!SC) return false;
1407 // Return true if the value is negative, this matches things like (-42 * V).
1408 return SC->getValue()->getValue().isNegative();
1411 // CollectIVUsers - Transform our list of users and offsets to a bit more
1412 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1413 // of the strided accesses, as well as the old information from Uses. We
1414 // progressively move information from the Base field to the Imm field, until
1415 // we eventually have the full access expression to rewrite the use.
1416 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1417 IVUsersOfOneStride &Uses,
1419 bool &AllUsesAreAddresses,
1420 bool &AllUsesAreOutsideLoop,
1421 std::vector<BasedUser> &UsersToProcess) {
1422 UsersToProcess.reserve(Uses.Users.size());
1423 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1424 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1426 // Move any loop variant operands from the offset field to the immediate
1427 // field of the use, so that we don't try to use something before it is
1429 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1430 UsersToProcess.back().Imm, L, SE);
1431 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1432 "Base value is not loop invariant!");
1435 // We now have a whole bunch of uses of like-strided induction variables, but
1436 // they might all have different bases. We want to emit one PHI node for this
1437 // stride which we fold as many common expressions (between the IVs) into as
1438 // possible. Start by identifying the common expressions in the base values
1439 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1440 // "A+B"), emit it to the preheader, then remove the expression from the
1441 // UsersToProcess base values.
1442 SCEVHandle CommonExprs =
1443 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1445 // Next, figure out what we can represent in the immediate fields of
1446 // instructions. If we can represent anything there, move it to the imm
1447 // fields of the BasedUsers. We do this so that it increases the commonality
1448 // of the remaining uses.
1449 unsigned NumPHI = 0;
1450 bool HasAddress = false;
1451 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1452 // If the user is not in the current loop, this means it is using the exit
1453 // value of the IV. Do not put anything in the base, make sure it's all in
1454 // the immediate field to allow as much factoring as possible.
1455 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1456 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1457 UsersToProcess[i].Base);
1458 UsersToProcess[i].Base =
1459 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1461 // Not all uses are outside the loop.
1462 AllUsesAreOutsideLoop = false;
1464 // Addressing modes can be folded into loads and stores. Be careful that
1465 // the store is through the expression, not of the expression though.
1467 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1468 UsersToProcess[i].OperandValToReplace);
1469 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1477 // If this use isn't an address, then not all uses are addresses.
1478 if (!isAddress && !isPHI)
1479 AllUsesAreAddresses = false;
1481 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1482 UsersToProcess[i].Imm, isAddress, L, SE);
1486 // If one of the use is a PHI node and all other uses are addresses, still
1487 // allow iv reuse. Essentially we are trading one constant multiplication
1488 // for one fewer iv.
1490 AllUsesAreAddresses = false;
1492 // There are no in-loop address uses.
1493 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1494 AllUsesAreAddresses = false;
1499 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1500 /// is valid and profitable for the given set of users of a stride. In
1501 /// full strength-reduction mode, all addresses at the current stride are
1502 /// strength-reduced all the way down to pointer arithmetic.
1504 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1505 const std::vector<BasedUser> &UsersToProcess,
1507 bool AllUsesAreAddresses,
1508 SCEVHandle Stride) {
1509 if (!EnableFullLSRMode)
1512 // The heuristics below aim to avoid increasing register pressure, but
1513 // fully strength-reducing all the addresses increases the number of
1514 // add instructions, so don't do this when optimizing for size.
1515 // TODO: If the loop is large, the savings due to simpler addresses
1516 // may oughtweight the costs of the extra increment instructions.
1517 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1520 // TODO: For now, don't do full strength reduction if there could
1521 // potentially be greater-stride multiples of the current stride
1522 // which could reuse the current stride IV.
1523 if (StrideOrder.back() != Stride)
1526 // Iterate through the uses to find conditions that automatically rule out
1528 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1529 SCEV *Base = UsersToProcess[i].Base;
1530 SCEV *Imm = UsersToProcess[i].Imm;
1531 // If any users have a loop-variant component, they can't be fully
1532 // strength-reduced.
1533 if (Imm && !Imm->isLoopInvariant(L))
1535 // If there are to users with the same base and the difference between
1536 // the two Imm values can't be folded into the address, full
1537 // strength reduction would increase register pressure.
1539 SCEV *CurImm = UsersToProcess[i].Imm;
1540 if ((CurImm || Imm) && CurImm != Imm) {
1541 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1542 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1543 const Instruction *Inst = UsersToProcess[i].Inst;
1544 const Type *UseTy = Inst->getType();
1545 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
1546 UseTy = SI->getOperand(0)->getType();
1547 SCEVHandle Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1548 if (!Diff->isZero() &&
1549 (!AllUsesAreAddresses ||
1550 !fitsInAddressMode(Diff, UseTy, TLI, /*HasBaseReg=*/true)))
1553 } while (++i != e && Base == UsersToProcess[i].Base);
1556 // If there's exactly one user in this stride, fully strength-reducing it
1557 // won't increase register pressure. If it's starting from a non-zero base,
1558 // it'll be simpler this way.
1559 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1562 // Otherwise, if there are any users in this stride that don't require
1563 // a register for their base, full strength-reduction will increase
1564 // register pressure.
1565 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1566 if (UsersToProcess[i].Base->isZero())
1569 // Otherwise, go for it.
1573 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1574 /// with the specified start and step values in the specified loop.
1576 /// If NegateStride is true, the stride should be negated by using a
1577 /// subtract instead of an add.
1579 /// Return the created phi node, and return the step instruction by
1580 /// reference in IncV.
1582 static PHINode *InsertAffinePhi(SCEVHandle Start, SCEVHandle Step,
1584 SCEVExpander &Rewriter,
1586 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1587 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1589 BasicBlock *Header = L->getHeader();
1590 BasicBlock *Preheader = L->getLoopPreheader();
1592 PHINode *PN = PHINode::Create(Start->getType(), "lsr.iv", Header->begin());
1593 PN->addIncoming(Rewriter.expandCodeFor(Start, Preheader->getTerminator()),
1596 pred_iterator HPI = pred_begin(Header);
1597 assert(HPI != pred_end(Header) && "Loop with zero preds???");
1598 if (!L->contains(*HPI)) ++HPI;
1599 assert(HPI != pred_end(Header) && L->contains(*HPI) &&
1600 "No backedge in loop?");
1602 // If the stride is negative, insert a sub instead of an add for the
1604 bool isNegative = isNonConstantNegative(Step);
1605 SCEVHandle IncAmount = Step;
1607 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1609 // Insert an add instruction right before the terminator corresponding
1610 // to the back-edge.
1611 Value *StepV = Rewriter.expandCodeFor(IncAmount, Preheader->getTerminator());
1613 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1614 (*HPI)->getTerminator());
1616 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1617 (*HPI)->getTerminator());
1619 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1621 pred_iterator PI = pred_begin(Header);
1622 if (*PI == L->getLoopPreheader())
1624 PN->addIncoming(IncV, *PI);
1630 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1631 // We want to emit code for users inside the loop first. To do this, we
1632 // rearrange BasedUser so that the entries at the end have
1633 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1634 // vector (so we handle them first).
1635 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1636 PartitionByIsUseOfPostIncrementedValue);
1638 // Sort this by base, so that things with the same base are handled
1639 // together. By partitioning first and stable-sorting later, we are
1640 // guaranteed that within each base we will pop off users from within the
1641 // loop before users outside of the loop with a particular base.
1643 // We would like to use stable_sort here, but we can't. The problem is that
1644 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1645 // we don't have anything to do a '<' comparison on. Because we think the
1646 // number of uses is small, do a horrible bubble sort which just relies on
1648 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1649 // Get a base value.
1650 SCEVHandle Base = UsersToProcess[i].Base;
1652 // Compact everything with this base to be consecutive with this one.
1653 for (unsigned j = i+1; j != e; ++j) {
1654 if (UsersToProcess[j].Base == Base) {
1655 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1662 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1663 /// UsersToProcess, meaning lowering addresses all the way down to direct
1664 /// pointer arithmetic.
1667 LoopStrengthReduce::PrepareToStrengthReduceFully(
1668 std::vector<BasedUser> &UsersToProcess,
1670 SCEVHandle CommonExprs,
1672 SCEVExpander &PreheaderRewriter) {
1673 DOUT << " Fully reducing all users\n";
1675 // Rewrite the UsersToProcess records, creating a separate PHI for each
1676 // unique Base value.
1677 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1678 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1679 // pick the first Imm value here to start with, and adjust it for the
1681 SCEVHandle Imm = UsersToProcess[i].Imm;
1682 SCEVHandle Base = UsersToProcess[i].Base;
1683 SCEVHandle Start = SE->getAddExpr(CommonExprs, Base, Imm);
1685 PHINode *Phi = InsertAffinePhi(Start, Stride, L,
1688 // Loop over all the users with the same base.
1690 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1691 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1692 UsersToProcess[i].Phi = Phi;
1693 UsersToProcess[i].IncV = IncV;
1694 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1695 "ShouldUseFullStrengthReductionMode should reject this!");
1696 } while (++i != e && Base == UsersToProcess[i].Base);
1700 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1701 /// given users to share.
1704 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1705 std::vector<BasedUser> &UsersToProcess,
1707 SCEVHandle CommonExprs,
1710 SCEVExpander &PreheaderRewriter) {
1711 DOUT << " Inserting new PHI:\n";
1714 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1719 // Remember this in case a later stride is multiple of this.
1720 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi, IncV);
1722 // All the users will share this new IV.
1723 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1724 UsersToProcess[i].Phi = Phi;
1725 UsersToProcess[i].IncV = IncV;
1729 DEBUG(WriteAsOperand(*DOUT, Phi, /*PrintType=*/false));
1731 DEBUG(WriteAsOperand(*DOUT, IncV, /*PrintType=*/false));
1735 /// PrepareToStrengthReduceWithNewPhi - Prepare for the given users to reuse
1736 /// an induction variable with a stride that is a factor of the current
1737 /// induction variable.
1740 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1741 std::vector<BasedUser> &UsersToProcess,
1743 const IVExpr &ReuseIV,
1744 Instruction *PreInsertPt) {
1745 DOUT << " Rewriting in terms of existing IV of STRIDE " << *ReuseIV.Stride
1746 << " and BASE " << *ReuseIV.Base << "\n";
1748 // All the users will share the reused IV.
1749 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1750 UsersToProcess[i].Phi = ReuseIV.PHI;
1751 UsersToProcess[i].IncV = ReuseIV.IncV;
1754 Constant *C = dyn_cast<Constant>(CommonBaseV);
1756 (!C->isNullValue() &&
1757 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1759 // We want the common base emitted into the preheader! This is just
1760 // using cast as a copy so BitCast (no-op cast) is appropriate
1761 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1762 "commonbase", PreInsertPt);
1765 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1766 const Type *ReplacedTy,
1767 std::vector<BasedUser> &UsersToProcess,
1768 const TargetLowering *TLI) {
1769 SmallVector<Instruction*, 16> AddrModeInsts;
1770 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1771 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1773 ExtAddrMode AddrMode =
1774 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1775 ReplacedTy, UsersToProcess[i].Inst,
1776 AddrModeInsts, *TLI);
1777 if (GV && GV != AddrMode.BaseGV)
1779 if (Offset && !AddrMode.BaseOffs)
1780 // FIXME: How to accurate check it's immediate offset is folded.
1782 AddrModeInsts.clear();
1787 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1788 /// stride of IV. All of the users may have different starting values, and this
1789 /// may not be the only stride.
1790 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1791 IVUsersOfOneStride &Uses,
1793 // If all the users are moved to another stride, then there is nothing to do.
1794 if (Uses.Users.empty())
1797 // Keep track if every use in UsersToProcess is an address. If they all are,
1798 // we may be able to rewrite the entire collection of them in terms of a
1799 // smaller-stride IV.
1800 bool AllUsesAreAddresses = true;
1802 // Keep track if every use of a single stride is outside the loop. If so,
1803 // we want to be more aggressive about reusing a smaller-stride IV; a
1804 // multiply outside the loop is better than another IV inside. Well, usually.
1805 bool AllUsesAreOutsideLoop = true;
1807 // Transform our list of users and offsets to a bit more complex table. In
1808 // this new vector, each 'BasedUser' contains 'Base' the base of the
1809 // strided accessas well as the old information from Uses. We progressively
1810 // move information from the Base field to the Imm field, until we eventually
1811 // have the full access expression to rewrite the use.
1812 std::vector<BasedUser> UsersToProcess;
1813 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1814 AllUsesAreOutsideLoop,
1817 // Sort the UsersToProcess array so that users with common bases are
1818 // next to each other.
1819 SortUsersToProcess(UsersToProcess);
1821 // If we managed to find some expressions in common, we'll need to carry
1822 // their value in a register and add it in for each use. This will take up
1823 // a register operand, which potentially restricts what stride values are
1825 bool HaveCommonExprs = !CommonExprs->isZero();
1827 const Type *ReplacedTy = CommonExprs->getType();
1829 // If all uses are addresses, consider sinking the immediate part of the
1830 // common expression back into uses if they can fit in the immediate fields.
1831 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1832 SCEVHandle NewCommon = CommonExprs;
1833 SCEVHandle Imm = SE->getIntegerSCEV(0, ReplacedTy);
1834 MoveImmediateValues(TLI, ReplacedTy, NewCommon, Imm, true, L, SE);
1835 if (!Imm->isZero()) {
1838 // If the immediate part of the common expression is a GV, check if it's
1839 // possible to fold it into the target addressing mode.
1840 GlobalValue *GV = 0;
1841 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm)) {
1842 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
1843 if (CE->getOpcode() == Instruction::PtrToInt)
1844 GV = dyn_cast<GlobalValue>(CE->getOperand(0));
1847 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1848 Offset = SC->getValue()->getSExtValue();
1850 DoSink = IsImmFoldedIntoAddrMode(GV, Offset, ReplacedTy,
1851 UsersToProcess, TLI);
1854 DOUT << " Sinking " << *Imm << " back down into uses\n";
1855 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1856 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1857 CommonExprs = NewCommon;
1858 HaveCommonExprs = !CommonExprs->isZero();
1864 // Now that we know what we need to do, insert the PHI node itself.
1866 DOUT << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1868 << " Common base: " << *CommonExprs << "\n";
1870 SCEVExpander Rewriter(*SE, *LI);
1871 SCEVExpander PreheaderRewriter(*SE, *LI);
1873 BasicBlock *Preheader = L->getLoopPreheader();
1874 Instruction *PreInsertPt = Preheader->getTerminator();
1875 BasicBlock *LatchBlock = L->getLoopLatch();
1877 Value *CommonBaseV = ConstantInt::get(ReplacedTy, 0);
1879 SCEVHandle RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1880 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1881 SE->getIntegerSCEV(0, Type::Int32Ty),
1884 /// Choose a strength-reduction strategy and prepare for it by creating
1885 /// the necessary PHIs and adjusting the bookkeeping.
1886 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1887 AllUsesAreAddresses, Stride)) {
1888 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1891 // Emit the initial base value into the loop preheader.
1892 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1894 // If all uses are addresses, check if it is possible to reuse an IV with a
1895 // stride that is a factor of this stride. And that the multiple is a number
1896 // that can be encoded in the scale field of the target addressing mode. And
1897 // that we will have a valid instruction after this substition, including
1898 // the immediate field, if any.
1899 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1900 AllUsesAreOutsideLoop,
1901 Stride, ReuseIV, CommonExprs->getType(),
1903 if (isa<SCEVConstant>(RewriteFactor) &&
1904 cast<SCEVConstant>(RewriteFactor)->isZero())
1905 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1906 CommonBaseV, L, PreheaderRewriter);
1908 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1909 ReuseIV, PreInsertPt);
1912 // Process all the users now, replacing their strided uses with
1913 // strength-reduced forms. This outer loop handles all bases, the inner
1914 // loop handles all users of a particular base.
1915 while (!UsersToProcess.empty()) {
1916 SCEVHandle Base = UsersToProcess.back().Base;
1917 Instruction *Inst = UsersToProcess.back().Inst;
1919 // Emit the code for Base into the preheader.
1920 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1922 DOUT << " Examining uses with BASE ";
1923 DEBUG(WriteAsOperand(*DOUT, BaseV, /*PrintType=*/false));
1926 // If BaseV is a constant other than 0, make sure that it gets inserted into
1927 // the preheader, instead of being forward substituted into the uses. We do
1928 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1930 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1931 if (!C->isNullValue() && !fitsInAddressMode(Base, ReplacedTy,
1933 // We want this constant emitted into the preheader! This is just
1934 // using cast as a copy so BitCast (no-op cast) is appropriate
1935 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1940 // Emit the code to add the immediate offset to the Phi value, just before
1941 // the instructions that we identified as using this stride and base.
1943 // FIXME: Use emitted users to emit other users.
1944 BasedUser &User = UsersToProcess.back();
1946 DOUT << " Examining use ";
1947 DEBUG(WriteAsOperand(*DOUT, UsersToProcess.back().OperandValToReplace,
1948 /*PrintType=*/false));
1949 DOUT << " in Inst: " << *Inst;
1951 // If this instruction wants to use the post-incremented value, move it
1952 // after the post-inc and use its value instead of the PHI.
1953 Value *RewriteOp = User.Phi;
1954 if (User.isUseOfPostIncrementedValue) {
1955 RewriteOp = User.IncV;
1957 // If this user is in the loop, make sure it is the last thing in the
1958 // loop to ensure it is dominated by the increment.
1959 if (L->contains(User.Inst->getParent()))
1960 User.Inst->moveBefore(LatchBlock->getTerminator());
1962 if (RewriteOp->getType() != ReplacedTy) {
1963 Instruction::CastOps opcode = Instruction::Trunc;
1964 if (ReplacedTy->getPrimitiveSizeInBits() ==
1965 RewriteOp->getType()->getPrimitiveSizeInBits())
1966 opcode = Instruction::BitCast;
1967 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1970 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1972 // If we had to insert new instructions for RewriteOp, we have to
1973 // consider that they may not have been able to end up immediately
1974 // next to RewriteOp, because non-PHI instructions may never precede
1975 // PHI instructions in a block. In this case, remember where the last
1976 // instruction was inserted so that if we're replacing a different
1977 // PHI node, we can use the later point to expand the final
1979 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1980 if (RewriteOp == User.Phi) NewBasePt = 0;
1982 // Clear the SCEVExpander's expression map so that we are guaranteed
1983 // to have the code emitted where we expect it.
1986 // If we are reusing the iv, then it must be multiplied by a constant
1987 // factor to take advantage of the addressing mode scale component.
1988 if (!isa<SCEVConstant>(RewriteFactor) ||
1989 !cast<SCEVConstant>(RewriteFactor)->isZero()) {
1990 // If we're reusing an IV with a nonzero base (currently this happens
1991 // only when all reuses are outside the loop) subtract that base here.
1992 // The base has been used to initialize the PHI node but we don't want
1994 if (!ReuseIV.Base->isZero()) {
1995 SCEVHandle typedBase = ReuseIV.Base;
1996 if (RewriteExpr->getType()->getPrimitiveSizeInBits() !=
1997 ReuseIV.Base->getType()->getPrimitiveSizeInBits()) {
1998 // It's possible the original IV is a larger type than the new IV,
1999 // in which case we have to truncate the Base. We checked in
2000 // RequiresTypeConversion that this is valid.
2001 assert (RewriteExpr->getType()->getPrimitiveSizeInBits() <
2002 ReuseIV.Base->getType()->getPrimitiveSizeInBits() &&
2003 "Unexpected lengthening conversion!");
2004 typedBase = SE->getTruncateExpr(ReuseIV.Base,
2005 RewriteExpr->getType());
2007 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
2010 // Multiply old variable, with base removed, by new scale factor.
2011 RewriteExpr = SE->getMulExpr(RewriteFactor,
2014 // The common base is emitted in the loop preheader. But since we
2015 // are reusing an IV, it has not been used to initialize the PHI node.
2016 // Add it to the expression used to rewrite the uses.
2017 // When this use is outside the loop, we earlier subtracted the
2018 // common base, and are adding it back here. Use the same expression
2019 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
2020 if (!isa<ConstantInt>(CommonBaseV) ||
2021 !cast<ConstantInt>(CommonBaseV)->isZero()) {
2022 if (L->contains(User.Inst->getParent()))
2023 RewriteExpr = SE->getAddExpr(RewriteExpr,
2024 SE->getUnknown(CommonBaseV));
2026 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
2030 // Now that we know what we need to do, insert code before User for the
2031 // immediate and any loop-variant expressions.
2032 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
2033 // Add BaseV to the PHI value if needed.
2034 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
2036 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
2040 // Mark old value we replaced as possibly dead, so that it is eliminated
2041 // if we just replaced the last use of that value.
2042 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
2044 UsersToProcess.pop_back();
2047 // If there are any more users to process with the same base, process them
2048 // now. We sorted by base above, so we just have to check the last elt.
2049 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
2050 // TODO: Next, find out which base index is the most common, pull it out.
2053 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
2054 // different starting values, into different PHIs.
2057 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
2058 /// set the IV user and stride information and return true, otherwise return
2060 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
2061 const SCEVHandle *&CondStride) {
2062 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
2064 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2065 IVUsesByStride.find(StrideOrder[Stride]);
2066 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2068 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2069 E = SI->second.Users.end(); UI != E; ++UI)
2070 if (UI->User == Cond) {
2071 // NOTE: we could handle setcc instructions with multiple uses here, but
2072 // InstCombine does it as well for simple uses, it's not clear that it
2073 // occurs enough in real life to handle.
2075 CondStride = &SI->first;
2083 // Constant strides come first which in turns are sorted by their absolute
2084 // values. If absolute values are the same, then positive strides comes first.
2086 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
2087 struct StrideCompare {
2088 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
2089 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
2090 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
2092 int64_t LV = LHSC->getValue()->getSExtValue();
2093 int64_t RV = RHSC->getValue()->getSExtValue();
2094 uint64_t ALV = (LV < 0) ? -LV : LV;
2095 uint64_t ARV = (RV < 0) ? -RV : RV;
2103 // If it's the same value but different type, sort by bit width so
2104 // that we emit larger induction variables before smaller
2105 // ones, letting the smaller be re-written in terms of larger ones.
2106 return RHS->getBitWidth() < LHS->getBitWidth();
2108 return LHSC && !RHSC;
2113 /// ChangeCompareStride - If a loop termination compare instruction is the
2114 /// only use of its stride, and the compaison is against a constant value,
2115 /// try eliminate the stride by moving the compare instruction to another
2116 /// stride and change its constant operand accordingly. e.g.
2122 /// if (v2 < 10) goto loop
2127 /// if (v1 < 30) goto loop
2128 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
2129 IVStrideUse* &CondUse,
2130 const SCEVHandle* &CondStride) {
2131 if (StrideOrder.size() < 2 ||
2132 IVUsesByStride[*CondStride].Users.size() != 1)
2134 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
2135 if (!SC) return Cond;
2137 ICmpInst::Predicate Predicate = Cond->getPredicate();
2138 int64_t CmpSSInt = SC->getValue()->getSExtValue();
2139 unsigned BitWidth = (*CondStride)->getBitWidth();
2140 uint64_t SignBit = 1ULL << (BitWidth-1);
2141 const Type *CmpTy = Cond->getOperand(0)->getType();
2142 const Type *NewCmpTy = NULL;
2143 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
2144 unsigned NewTyBits = 0;
2145 SCEVHandle *NewStride = NULL;
2146 Value *NewCmpLHS = NULL;
2147 Value *NewCmpRHS = NULL;
2149 SCEVHandle NewOffset = SE->getIntegerSCEV(0, UIntPtrTy);
2150 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2152 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
2153 int64_t CmpVal = C->getValue().getSExtValue();
2155 // Check stride constant and the comparision constant signs to detect
2157 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
2160 // Look for a suitable stride / iv as replacement.
2161 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
2162 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2163 IVUsesByStride.find(StrideOrder[i]);
2164 if (!isa<SCEVConstant>(SI->first))
2166 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
2167 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
2170 Scale = SSInt / CmpSSInt;
2171 int64_t NewCmpVal = CmpVal * Scale;
2172 APInt Mul = APInt(BitWidth, NewCmpVal);
2173 // Check for overflow.
2174 if (Mul.getSExtValue() != NewCmpVal)
2177 // Watch out for overflow.
2178 if (ICmpInst::isSignedPredicate(Predicate) &&
2179 (CmpVal & SignBit) != (NewCmpVal & SignBit))
2182 if (NewCmpVal == CmpVal)
2184 // Pick the best iv to use trying to avoid a cast.
2186 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2187 E = SI->second.Users.end(); UI != E; ++UI) {
2188 NewCmpLHS = UI->OperandValToReplace;
2189 if (NewCmpLHS->getType() == CmpTy)
2195 NewCmpTy = NewCmpLHS->getType();
2196 NewTyBits = isa<PointerType>(NewCmpTy)
2197 ? UIntPtrTy->getPrimitiveSizeInBits()
2198 : NewCmpTy->getPrimitiveSizeInBits();
2199 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2200 // Check if it is possible to rewrite it using
2201 // an iv / stride of a smaller integer type.
2202 bool TruncOk = false;
2203 if (NewCmpTy->isInteger()) {
2204 unsigned Bits = NewTyBits;
2205 if (ICmpInst::isSignedPredicate(Predicate))
2207 uint64_t Mask = (1ULL << Bits) - 1;
2208 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
2215 // Don't rewrite if use offset is non-constant and the new type is
2216 // of a different type.
2217 // FIXME: too conservative?
2218 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset))
2221 bool AllUsesAreAddresses = true;
2222 bool AllUsesAreOutsideLoop = true;
2223 std::vector<BasedUser> UsersToProcess;
2224 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
2225 AllUsesAreAddresses,
2226 AllUsesAreOutsideLoop,
2228 // Avoid rewriting the compare instruction with an iv of new stride
2229 // if it's likely the new stride uses will be rewritten using the
2230 // stride of the compare instruction.
2231 if (AllUsesAreAddresses &&
2232 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess))
2235 // If scale is negative, use swapped predicate unless it's testing
2237 if (Scale < 0 && !Cond->isEquality())
2238 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2240 NewStride = &StrideOrder[i];
2241 if (!isa<PointerType>(NewCmpTy))
2242 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2244 NewCmpRHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
2245 NewCmpRHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
2246 NewCmpRHS, NewCmpTy);
2248 NewOffset = TyBits == NewTyBits
2249 ? SE->getMulExpr(CondUse->Offset,
2250 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
2251 : SE->getConstant(ConstantInt::get(NewCmpTy,
2252 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
2257 // Forgo this transformation if it the increment happens to be
2258 // unfortunately positioned after the condition, and the condition
2259 // has multiple uses which prevent it from being moved immediately
2260 // before the branch. See
2261 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2262 // for an example of this situation.
2263 if (!Cond->hasOneUse()) {
2264 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2271 // Create a new compare instruction using new stride / iv.
2272 ICmpInst *OldCond = Cond;
2273 // Insert new compare instruction.
2274 Cond = new ICmpInst(Predicate, NewCmpLHS, NewCmpRHS,
2275 L->getHeader()->getName() + ".termcond",
2278 // Remove the old compare instruction. The old indvar is probably dead too.
2279 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
2280 SE->deleteValueFromRecords(OldCond);
2281 OldCond->replaceAllUsesWith(Cond);
2282 OldCond->eraseFromParent();
2284 IVUsesByStride[*CondStride].Users.pop_back();
2285 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewCmpLHS);
2286 CondUse = &IVUsesByStride[*NewStride].Users.back();
2287 CondStride = NewStride;
2294 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
2295 /// an smax computation.
2297 /// This is a narrow solution to a specific, but acute, problem. For loops
2303 /// } while (++i < n);
2305 /// where the comparison is signed, the trip count isn't just 'n', because
2306 /// 'n' could be negative. And unfortunately this can come up even for loops
2307 /// where the user didn't use a C do-while loop. For example, seemingly
2308 /// well-behaved top-test loops will commonly be lowered like this:
2314 /// } while (++i < n);
2317 /// and then it's possible for subsequent optimization to obscure the if
2318 /// test in such a way that indvars can't find it.
2320 /// When indvars can't find the if test in loops like this, it creates a
2321 /// signed-max expression, which allows it to give the loop a canonical
2322 /// induction variable:
2325 /// smax = n < 1 ? 1 : n;
2328 /// } while (++i != smax);
2330 /// Canonical induction variables are necessary because the loop passes
2331 /// are designed around them. The most obvious example of this is the
2332 /// LoopInfo analysis, which doesn't remember trip count values. It
2333 /// expects to be able to rediscover the trip count each time it is
2334 /// needed, and it does this using a simple analyis that only succeeds if
2335 /// the loop has a canonical induction variable.
2337 /// However, when it comes time to generate code, the maximum operation
2338 /// can be quite costly, especially if it's inside of an outer loop.
2340 /// This function solves this problem by detecting this type of loop and
2341 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2342 /// the instructions for the maximum computation.
2344 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
2345 IVStrideUse* &CondUse) {
2346 // Check that the loop matches the pattern we're looking for.
2347 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2348 Cond->getPredicate() != CmpInst::ICMP_NE)
2351 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2352 if (!Sel || !Sel->hasOneUse()) return Cond;
2354 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2355 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2357 SCEVHandle One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2359 // Add one to the backedge-taken count to get the trip count.
2360 SCEVHandle IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2362 // Check for a max calculation that matches the pattern.
2363 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
2364 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
2366 SCEVHandle SMaxLHS = SMax->getOperand(0);
2367 SCEVHandle SMaxRHS = SMax->getOperand(1);
2368 if (!SMaxLHS || SMaxLHS != One) return Cond;
2370 // Check the relevant induction variable for conformance to
2372 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
2373 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2374 if (!AR || !AR->isAffine() ||
2375 AR->getStart() != One ||
2376 AR->getStepRecurrence(*SE) != One)
2379 assert(AR->getLoop() == L &&
2380 "Loop condition operand is an addrec in a different loop!");
2382 // Check the right operand of the select, and remember it, as it will
2383 // be used in the new comparison instruction.
2385 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
2386 NewRHS = Sel->getOperand(1);
2387 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
2388 NewRHS = Sel->getOperand(2);
2389 if (!NewRHS) return Cond;
2391 // Ok, everything looks ok to change the condition into an SLT or SGE and
2392 // delete the max calculation.
2394 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
2397 Cond->getOperand(0), NewRHS, "scmp", Cond);
2399 // Delete the max calculation instructions.
2400 SE->deleteValueFromRecords(Cond);
2401 Cond->replaceAllUsesWith(NewCond);
2402 Cond->eraseFromParent();
2403 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2404 SE->deleteValueFromRecords(Sel);
2405 Sel->eraseFromParent();
2406 if (Cmp->use_empty()) {
2407 SE->deleteValueFromRecords(Cmp);
2408 Cmp->eraseFromParent();
2410 CondUse->User = NewCond;
2414 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2415 /// inside the loop then try to eliminate the cast opeation.
2416 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2418 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2419 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2422 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
2424 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2425 IVUsesByStride.find(StrideOrder[Stride]);
2426 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2427 if (!isa<SCEVConstant>(SI->first))
2430 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2431 E = SI->second.Users.end(); UI != E; /* empty */) {
2432 std::vector<IVStrideUse>::iterator CandidateUI = UI;
2434 Instruction *ShadowUse = CandidateUI->User;
2435 const Type *DestTy = NULL;
2437 /* If shadow use is a int->float cast then insert a second IV
2438 to eliminate this cast.
2440 for (unsigned i = 0; i < n; ++i)
2446 for (unsigned i = 0; i < n; ++i, ++d)
2449 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
2450 DestTy = UCast->getDestTy();
2451 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
2452 DestTy = SCast->getDestTy();
2453 if (!DestTy) continue;
2456 /* If target does not support DestTy natively then do not apply
2457 this transformation. */
2458 MVT DVT = TLI->getValueType(DestTy);
2459 if (!TLI->isTypeLegal(DVT)) continue;
2462 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2464 if (PH->getNumIncomingValues() != 2) continue;
2466 const Type *SrcTy = PH->getType();
2467 int Mantissa = DestTy->getFPMantissaWidth();
2468 if (Mantissa == -1) continue;
2469 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
2472 unsigned Entry, Latch;
2473 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2481 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2482 if (!Init) continue;
2483 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2485 BinaryOperator *Incr =
2486 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2487 if (!Incr) continue;
2488 if (Incr->getOpcode() != Instruction::Add
2489 && Incr->getOpcode() != Instruction::Sub)
2492 /* Initialize new IV, double d = 0.0 in above example. */
2493 ConstantInt *C = NULL;
2494 if (Incr->getOperand(0) == PH)
2495 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2496 else if (Incr->getOperand(1) == PH)
2497 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2503 /* Add new PHINode. */
2504 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2506 /* create new increment. '++d' in above example. */
2507 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2508 BinaryOperator *NewIncr =
2509 BinaryOperator::Create(Incr->getOpcode(),
2510 NewPH, CFP, "IV.S.next.", Incr);
2512 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2513 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2515 /* Remove cast operation */
2516 SE->deleteValueFromRecords(ShadowUse);
2517 ShadowUse->replaceAllUsesWith(NewPH);
2518 ShadowUse->eraseFromParent();
2519 SI->second.Users.erase(CandidateUI);
2526 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2527 // uses in the loop, look to see if we can eliminate some, in favor of using
2528 // common indvars for the different uses.
2529 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2530 // TODO: implement optzns here.
2532 OptimizeShadowIV(L);
2534 // Finally, get the terminating condition for the loop if possible. If we
2535 // can, we want to change it to use a post-incremented version of its
2536 // induction variable, to allow coalescing the live ranges for the IV into
2537 // one register value.
2538 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2539 BasicBlock *Preheader = L->getLoopPreheader();
2540 BasicBlock *LatchBlock =
2541 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2542 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2543 if (!TermBr || TermBr->isUnconditional() ||
2544 !isa<ICmpInst>(TermBr->getCondition()))
2546 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2548 // Search IVUsesByStride to find Cond's IVUse if there is one.
2549 IVStrideUse *CondUse = 0;
2550 const SCEVHandle *CondStride = 0;
2552 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2553 return; // setcc doesn't use the IV.
2555 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2556 // being unable to find a sufficient guard, for example), change the loop
2557 // comparison to use SLT instead of NE.
2558 Cond = OptimizeSMax(L, Cond, CondUse);
2560 // If possible, change stride and operands of the compare instruction to
2561 // eliminate one stride.
2562 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2564 // It's possible for the setcc instruction to be anywhere in the loop, and
2565 // possible for it to have multiple users. If it is not immediately before
2566 // the latch block branch, move it.
2567 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2568 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2569 Cond->moveBefore(TermBr);
2571 // Otherwise, clone the terminating condition and insert into the loopend.
2572 Cond = cast<ICmpInst>(Cond->clone());
2573 Cond->setName(L->getHeader()->getName() + ".termcond");
2574 LatchBlock->getInstList().insert(TermBr, Cond);
2576 // Clone the IVUse, as the old use still exists!
2577 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2578 CondUse->OperandValToReplace);
2579 CondUse = &IVUsesByStride[*CondStride].Users.back();
2583 // If we get to here, we know that we can transform the setcc instruction to
2584 // use the post-incremented version of the IV, allowing us to coalesce the
2585 // live ranges for the IV correctly.
2586 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2587 CondUse->isUseOfPostIncrementedValue = true;
2591 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2593 LI = &getAnalysis<LoopInfo>();
2594 DT = &getAnalysis<DominatorTree>();
2595 SE = &getAnalysis<ScalarEvolution>();
2596 TD = &getAnalysis<TargetData>();
2597 UIntPtrTy = TD->getIntPtrType();
2600 // Find all uses of induction variables in this loop, and categorize
2601 // them by stride. Start by finding all of the PHI nodes in the header for
2602 // this loop. If they are induction variables, inspect their uses.
2603 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2604 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2605 AddUsersIfInteresting(I, L, Processed);
2607 if (!IVUsesByStride.empty()) {
2609 DOUT << "\nLSR on \"" << L->getHeader()->getParent()->getNameStart()
2614 // Optimize induction variables. Some indvar uses can be transformed to use
2615 // strides that will be needed for other purposes. A common example of this
2616 // is the exit test for the loop, which can often be rewritten to use the
2617 // computation of some other indvar to decide when to terminate the loop.
2620 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2621 // doing computation in byte values, promote to 32-bit values if safe.
2623 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2624 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2625 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2626 // Need to be careful that IV's are all the same type. Only works for
2627 // intptr_t indvars.
2629 // IVsByStride keeps IVs for one particular loop.
2630 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2632 // Sort the StrideOrder so we process larger strides first.
2633 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2635 // Note: this processes each stride/type pair individually. All users
2636 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2637 // Also, note that we iterate over IVUsesByStride indirectly by using
2638 // StrideOrder. This extra layer of indirection makes the ordering of
2639 // strides deterministic - not dependent on map order.
2640 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2641 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2642 IVUsesByStride.find(StrideOrder[Stride]);
2643 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2644 StrengthReduceStridedIVUsers(SI->first, SI->second, L);
2648 // We're done analyzing this loop; release all the state we built up for it.
2649 CastedPointers.clear();
2650 IVUsesByStride.clear();
2651 IVsByStride.clear();
2652 StrideOrder.clear();
2653 for (unsigned i=0; i<GEPlist.size(); i++)
2654 SE->deleteValueFromRecords(GEPlist[i]);
2657 // Clean up after ourselves
2658 if (!DeadInsts.empty()) {
2659 DeleteTriviallyDeadInstructions();
2661 BasicBlock::iterator I = L->getHeader()->begin();
2662 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2663 // At this point, we know that we have killed one or more IV users.
2664 // It is worth checking to see if the cannonical indvar is also
2665 // dead, so that we can remove it as well.
2667 // We can remove a PHI if it is on a cycle in the def-use graph
2668 // where each node in the cycle has degree one, i.e. only one use,
2669 // and is an instruction with no side effects.
2671 // FIXME: this needs to eliminate an induction variable even if it's being
2672 // compared against some value to decide loop termination.
2673 if (!PN->hasOneUse())
2676 SmallPtrSet<PHINode *, 4> PHIs;
2677 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2678 J && J->hasOneUse() && !J->mayWriteToMemory();
2679 J = dyn_cast<Instruction>(*J->use_begin())) {
2680 // If we find the original PHI, we've discovered a cycle.
2682 // Break the cycle and mark the PHI for deletion.
2683 SE->deleteValueFromRecords(PN);
2684 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2685 DeadInsts.push_back(PN);
2689 // If we find a PHI more than once, we're on a cycle that
2690 // won't prove fruitful.
2691 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2695 DeleteTriviallyDeadInstructions();