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 /// getAccessType - Return the type of the memory being accessed.
571 static const Type *getAccessType(const Instruction *Inst) {
572 const Type *UseTy = Inst->getType();
573 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
574 UseTy = SI->getOperand(0)->getType();
575 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
576 // Addressing modes can also be folded into prefetches and a variety
578 switch (II->getIntrinsicID()) {
580 case Intrinsic::x86_sse_storeu_ps:
581 case Intrinsic::x86_sse2_storeu_pd:
582 case Intrinsic::x86_sse2_storeu_dq:
583 case Intrinsic::x86_sse2_storel_dq:
584 UseTy = II->getOperand(1)->getType();
591 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
592 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
593 /// return true. Otherwise, return false.
594 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
595 SmallPtrSet<Instruction*,16> &Processed) {
596 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
597 return false; // Void and FP expressions cannot be reduced.
598 if (!Processed.insert(I))
599 return true; // Instruction already handled.
601 // Get the symbolic expression for this instruction.
602 SCEVHandle ISE = GetExpressionSCEV(I);
603 if (isa<SCEVCouldNotCompute>(ISE)) return false;
605 // Get the start and stride for this expression.
606 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
607 SCEVHandle Stride = Start;
608 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE, DT))
609 return false; // Non-reducible symbolic expression, bail out.
611 std::vector<Instruction *> IUsers;
612 // Collect all I uses now because IVUseShouldUsePostIncValue may
613 // invalidate use_iterator.
614 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
615 IUsers.push_back(cast<Instruction>(*UI));
617 for (unsigned iused_index = 0, iused_size = IUsers.size();
618 iused_index != iused_size; ++iused_index) {
620 Instruction *User = IUsers[iused_index];
622 // Do not infinitely recurse on PHI nodes.
623 if (isa<PHINode>(User) && Processed.count(User))
626 // Descend recursively, but not into PHI nodes outside the current loop.
627 // It's important to see the entire expression outside the loop to get
628 // choices that depend on addressing mode use right, although we won't
629 // consider references ouside the loop in all cases.
630 // If User is already in Processed, we don't want to recurse into it again,
631 // but do want to record a second reference in the same instruction.
632 bool AddUserToIVUsers = false;
633 if (LI->getLoopFor(User->getParent()) != L) {
634 if (isa<PHINode>(User) || Processed.count(User) ||
635 !AddUsersIfInteresting(User, L, Processed)) {
636 DOUT << "FOUND USER in other loop: " << *User
637 << " OF SCEV: " << *ISE << "\n";
638 AddUserToIVUsers = true;
640 } else if (Processed.count(User) ||
641 !AddUsersIfInteresting(User, L, Processed)) {
642 DOUT << "FOUND USER: " << *User
643 << " OF SCEV: " << *ISE << "\n";
644 AddUserToIVUsers = true;
647 if (AddUserToIVUsers) {
648 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
649 if (StrideUses.Users.empty()) // First occurrence of this stride?
650 StrideOrder.push_back(Stride);
652 // Okay, we found a user that we cannot reduce. Analyze the instruction
653 // and decide what to do with it. If we are a use inside of the loop, use
654 // the value before incrementation, otherwise use it after incrementation.
655 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
656 // The value used will be incremented by the stride more than we are
657 // expecting, so subtract this off.
658 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
659 StrideUses.addUser(NewStart, User, I);
660 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
661 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
663 StrideUses.addUser(Start, User, I);
671 /// BasedUser - For a particular base value, keep information about how we've
672 /// partitioned the expression so far.
674 /// SE - The current ScalarEvolution object.
677 /// Base - The Base value for the PHI node that needs to be inserted for
678 /// this use. As the use is processed, information gets moved from this
679 /// field to the Imm field (below). BasedUser values are sorted by this
683 /// Inst - The instruction using the induction variable.
686 /// OperandValToReplace - The operand value of Inst to replace with the
688 Value *OperandValToReplace;
690 /// Imm - The immediate value that should be added to the base immediately
691 /// before Inst, because it will be folded into the imm field of the
692 /// instruction. This is also sometimes used for loop-variant values that
693 /// must be added inside the loop.
696 /// Phi - The induction variable that performs the striding that
697 /// should be used for this user.
700 /// IncV - The post-incremented value of Phi.
703 // isUseOfPostIncrementedValue - True if this should use the
704 // post-incremented version of this IV, not the preincremented version.
705 // This can only be set in special cases, such as the terminating setcc
706 // instruction for a loop and uses outside the loop that are dominated by
708 bool isUseOfPostIncrementedValue;
710 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
711 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
712 OperandValToReplace(IVSU.OperandValToReplace),
713 Imm(SE->getIntegerSCEV(0, Base->getType())),
714 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
716 // Once we rewrite the code to insert the new IVs we want, update the
717 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
719 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
720 Instruction *InsertPt,
721 SCEVExpander &Rewriter, Loop *L, Pass *P,
722 SmallVectorImpl<Instruction*> &DeadInsts);
724 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
725 SCEVExpander &Rewriter,
726 Instruction *IP, Loop *L);
731 void BasedUser::dump() const {
732 cerr << " Base=" << *Base;
733 cerr << " Imm=" << *Imm;
734 cerr << " Inst: " << *Inst;
737 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
738 SCEVExpander &Rewriter,
739 Instruction *IP, Loop *L) {
740 // Figure out where we *really* want to insert this code. In particular, if
741 // the user is inside of a loop that is nested inside of L, we really don't
742 // want to insert this expression before the user, we'd rather pull it out as
743 // many loops as possible.
744 LoopInfo &LI = Rewriter.getLoopInfo();
745 Instruction *BaseInsertPt = IP;
747 // Figure out the most-nested loop that IP is in.
748 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
750 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
751 // the preheader of the outer-most loop where NewBase is not loop invariant.
752 if (L->contains(IP->getParent()))
753 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
754 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
755 InsertLoop = InsertLoop->getParentLoop();
758 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
760 // If there is no immediate value, skip the next part.
764 // If we are inserting the base and imm values in the same block, make sure to
765 // adjust the IP position if insertion reused a result.
766 if (IP == BaseInsertPt)
767 IP = Rewriter.getInsertionPoint();
769 // Always emit the immediate (if non-zero) into the same block as the user.
770 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
771 return Rewriter.expandCodeFor(NewValSCEV, IP);
776 // Once we rewrite the code to insert the new IVs we want, update the
777 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
778 // to it. NewBasePt is the last instruction which contributes to the
779 // value of NewBase in the case that it's a diffferent instruction from
780 // the PHI that NewBase is computed from, or null otherwise.
782 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
783 Instruction *NewBasePt,
784 SCEVExpander &Rewriter, Loop *L, Pass *P,
785 SmallVectorImpl<Instruction*> &DeadInsts){
786 if (!isa<PHINode>(Inst)) {
787 // By default, insert code at the user instruction.
788 BasicBlock::iterator InsertPt = Inst;
790 // However, if the Operand is itself an instruction, the (potentially
791 // complex) inserted code may be shared by many users. Because of this, we
792 // want to emit code for the computation of the operand right before its old
793 // computation. This is usually safe, because we obviously used to use the
794 // computation when it was computed in its current block. However, in some
795 // cases (e.g. use of a post-incremented induction variable) the NewBase
796 // value will be pinned to live somewhere after the original computation.
797 // In this case, we have to back off.
799 // If this is a use outside the loop (which means after, since it is based
800 // on a loop indvar) we use the post-incremented value, so that we don't
801 // artificially make the preinc value live out the bottom of the loop.
802 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
803 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
804 InsertPt = NewBasePt;
806 } else if (Instruction *OpInst
807 = dyn_cast<Instruction>(OperandValToReplace)) {
809 while (isa<PHINode>(InsertPt)) ++InsertPt;
812 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
813 // Adjust the type back to match the Inst. Note that we can't use InsertPt
814 // here because the SCEVExpander may have inserted the instructions after
815 // that point, in its efforts to avoid inserting redundant expressions.
816 if (isa<PointerType>(OperandValToReplace->getType())) {
817 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
819 OperandValToReplace->getType());
821 // Replace the use of the operand Value with the new Phi we just created.
822 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
824 DOUT << " Replacing with ";
825 DEBUG(WriteAsOperand(*DOUT, NewVal, /*PrintType=*/false));
826 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
830 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
831 // expression into each operand block that uses it. Note that PHI nodes can
832 // have multiple entries for the same predecessor. We use a map to make sure
833 // that a PHI node only has a single Value* for each predecessor (which also
834 // prevents us from inserting duplicate code in some blocks).
835 DenseMap<BasicBlock*, Value*> InsertedCode;
836 PHINode *PN = cast<PHINode>(Inst);
837 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
838 if (PN->getIncomingValue(i) == OperandValToReplace) {
839 // If the original expression is outside the loop, put the replacement
840 // code in the same place as the original expression,
841 // which need not be an immediate predecessor of this PHI. This way we
842 // need only one copy of it even if it is referenced multiple times in
843 // the PHI. We don't do this when the original expression is inside the
844 // loop because multiple copies sometimes do useful sinking of code in
846 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace);
847 if (L->contains(OldLoc->getParent())) {
848 // If this is a critical edge, split the edge so that we do not insert
849 // the code on all predecessor/successor paths. We do this unless this
850 // is the canonical backedge for this loop, as this can make some
851 // inserted code be in an illegal position.
852 BasicBlock *PHIPred = PN->getIncomingBlock(i);
853 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
854 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
856 // First step, split the critical edge.
857 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
859 // Next step: move the basic block. In particular, if the PHI node
860 // is outside of the loop, and PredTI is in the loop, we want to
861 // move the block to be immediately before the PHI block, not
862 // immediately after PredTI.
863 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
864 BasicBlock *NewBB = PN->getIncomingBlock(i);
865 NewBB->moveBefore(PN->getParent());
868 // Splitting the edge can reduce the number of PHI entries we have.
869 e = PN->getNumIncomingValues();
872 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
874 // Insert the code into the end of the predecessor block.
875 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ?
876 PN->getIncomingBlock(i)->getTerminator() :
877 OldLoc->getParent()->getTerminator();
878 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
880 // Adjust the type back to match the PHI. Note that we can't use
881 // InsertPt here because the SCEVExpander may have inserted its
882 // instructions after that point, in its efforts to avoid inserting
883 // redundant expressions.
884 if (isa<PointerType>(PN->getType())) {
885 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
890 DOUT << " Changing PHI use to ";
891 DEBUG(WriteAsOperand(*DOUT, Code, /*PrintType=*/false));
892 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n";
895 // Replace the use of the operand Value with the new Phi we just created.
896 PN->setIncomingValue(i, Code);
901 // PHI node might have become a constant value after SplitCriticalEdge.
902 DeadInsts.push_back(Inst);
906 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
907 /// mode, and does not need to be put in a register first.
908 static bool fitsInAddressMode(const SCEVHandle &V, const Type *UseTy,
909 const TargetLowering *TLI, bool HasBaseReg) {
910 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
911 int64_t VC = SC->getValue()->getSExtValue();
913 TargetLowering::AddrMode AM;
915 AM.HasBaseReg = HasBaseReg;
916 return TLI->isLegalAddressingMode(AM, UseTy);
918 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
919 return (VC > -(1 << 16) && VC < (1 << 16)-1);
923 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
924 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
925 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
926 Constant *Op0 = CE->getOperand(0);
927 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
928 TargetLowering::AddrMode AM;
930 AM.HasBaseReg = HasBaseReg;
931 return TLI->isLegalAddressingMode(AM, UseTy);
937 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
938 /// loop varying to the Imm operand.
939 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
940 Loop *L, ScalarEvolution *SE) {
941 if (Val->isLoopInvariant(L)) return; // Nothing to do.
943 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
944 std::vector<SCEVHandle> NewOps;
945 NewOps.reserve(SAE->getNumOperands());
947 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
948 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
949 // If this is a loop-variant expression, it must stay in the immediate
950 // field of the expression.
951 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
953 NewOps.push_back(SAE->getOperand(i));
957 Val = SE->getIntegerSCEV(0, Val->getType());
959 Val = SE->getAddExpr(NewOps);
960 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
961 // Try to pull immediates out of the start value of nested addrec's.
962 SCEVHandle Start = SARE->getStart();
963 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
965 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
967 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
969 // Otherwise, all of Val is variant, move the whole thing over.
970 Imm = SE->getAddExpr(Imm, Val);
971 Val = SE->getIntegerSCEV(0, Val->getType());
976 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
977 /// that can fit into the immediate field of instructions in the target.
978 /// Accumulate these immediate values into the Imm value.
979 static void MoveImmediateValues(const TargetLowering *TLI,
981 SCEVHandle &Val, SCEVHandle &Imm,
982 bool isAddress, Loop *L,
983 ScalarEvolution *SE) {
984 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
985 std::vector<SCEVHandle> NewOps;
986 NewOps.reserve(SAE->getNumOperands());
988 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
989 SCEVHandle NewOp = SAE->getOperand(i);
990 MoveImmediateValues(TLI, UseTy, NewOp, Imm, isAddress, L, SE);
992 if (!NewOp->isLoopInvariant(L)) {
993 // If this is a loop-variant expression, it must stay in the immediate
994 // field of the expression.
995 Imm = SE->getAddExpr(Imm, NewOp);
997 NewOps.push_back(NewOp);
1002 Val = SE->getIntegerSCEV(0, Val->getType());
1004 Val = SE->getAddExpr(NewOps);
1006 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
1007 // Try to pull immediates out of the start value of nested addrec's.
1008 SCEVHandle Start = SARE->getStart();
1009 MoveImmediateValues(TLI, UseTy, Start, Imm, isAddress, L, SE);
1011 if (Start != SARE->getStart()) {
1012 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
1014 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
1017 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
1018 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
1019 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
1020 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
1022 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
1023 SCEVHandle NewOp = SME->getOperand(1);
1024 MoveImmediateValues(TLI, UseTy, NewOp, SubImm, isAddress, L, SE);
1026 // If we extracted something out of the subexpressions, see if we can
1028 if (NewOp != SME->getOperand(1)) {
1029 // Scale SubImm up by "8". If the result is a target constant, we are
1031 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
1032 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
1033 // Accumulate the immediate.
1034 Imm = SE->getAddExpr(Imm, SubImm);
1036 // Update what is left of 'Val'.
1037 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
1044 // Loop-variant expressions must stay in the immediate field of the
1046 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
1047 !Val->isLoopInvariant(L)) {
1048 Imm = SE->getAddExpr(Imm, Val);
1049 Val = SE->getIntegerSCEV(0, Val->getType());
1053 // Otherwise, no immediates to move.
1056 static void MoveImmediateValues(const TargetLowering *TLI,
1058 SCEVHandle &Val, SCEVHandle &Imm,
1059 bool isAddress, Loop *L,
1060 ScalarEvolution *SE) {
1061 const Type *UseTy = getAccessType(User);
1062 MoveImmediateValues(TLI, UseTy, Val, Imm, isAddress, L, SE);
1065 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
1066 /// added together. This is used to reassociate common addition subexprs
1067 /// together for maximal sharing when rewriting bases.
1068 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
1070 ScalarEvolution *SE) {
1071 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
1072 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
1073 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
1074 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
1075 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
1076 if (SARE->getOperand(0) == Zero) {
1077 SubExprs.push_back(Expr);
1079 // Compute the addrec with zero as its base.
1080 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
1081 Ops[0] = Zero; // Start with zero base.
1082 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
1085 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
1087 } else if (!Expr->isZero()) {
1089 SubExprs.push_back(Expr);
1093 // This is logically local to the following function, but C++ says we have
1094 // to make it file scope.
1095 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
1097 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
1098 /// the Uses, removing any common subexpressions, except that if all such
1099 /// subexpressions can be folded into an addressing mode for all uses inside
1100 /// the loop (this case is referred to as "free" in comments herein) we do
1101 /// not remove anything. This looks for things like (a+b+c) and
1102 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
1103 /// is *removed* from the Bases and returned.
1105 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
1106 ScalarEvolution *SE, Loop *L,
1107 const TargetLowering *TLI) {
1108 unsigned NumUses = Uses.size();
1110 // Only one use? This is a very common case, so we handle it specially and
1112 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
1113 SCEVHandle Result = Zero;
1114 SCEVHandle FreeResult = Zero;
1116 // If the use is inside the loop, use its base, regardless of what it is:
1117 // it is clearly shared across all the IV's. If the use is outside the loop
1118 // (which means after it) we don't want to factor anything *into* the loop,
1119 // so just use 0 as the base.
1120 if (L->contains(Uses[0].Inst->getParent()))
1121 std::swap(Result, Uses[0].Base);
1125 // To find common subexpressions, count how many of Uses use each expression.
1126 // If any subexpressions are used Uses.size() times, they are common.
1127 // Also track whether all uses of each expression can be moved into an
1128 // an addressing mode "for free"; such expressions are left within the loop.
1129 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
1130 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
1132 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
1133 // order we see them.
1134 std::vector<SCEVHandle> UniqueSubExprs;
1136 std::vector<SCEVHandle> SubExprs;
1137 unsigned NumUsesInsideLoop = 0;
1138 for (unsigned i = 0; i != NumUses; ++i) {
1139 // If the user is outside the loop, just ignore it for base computation.
1140 // Since the user is outside the loop, it must be *after* the loop (if it
1141 // were before, it could not be based on the loop IV). We don't want users
1142 // after the loop to affect base computation of values *inside* the loop,
1143 // because we can always add their offsets to the result IV after the loop
1144 // is done, ensuring we get good code inside the loop.
1145 if (!L->contains(Uses[i].Inst->getParent()))
1147 NumUsesInsideLoop++;
1149 // If the base is zero (which is common), return zero now, there are no
1150 // CSEs we can find.
1151 if (Uses[i].Base == Zero) return Zero;
1153 // If this use is as an address we may be able to put CSEs in the addressing
1154 // mode rather than hoisting them.
1155 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1156 // We may need the UseTy below, but only when isAddrUse, so compute it
1157 // only in that case.
1158 const Type *UseTy = 0;
1160 UseTy = getAccessType(Uses[i].Inst);
1162 // Split the expression into subexprs.
1163 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1164 // Add one to SubExpressionUseData.Count for each subexpr present, and
1165 // if the subexpr is not a valid immediate within an addressing mode use,
1166 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1167 // hoist these out of the loop (if they are common to all uses).
1168 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1169 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1170 UniqueSubExprs.push_back(SubExprs[j]);
1171 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1172 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1177 // Now that we know how many times each is used, build Result. Iterate over
1178 // UniqueSubexprs so that we have a stable ordering.
1179 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1180 std::map<SCEVHandle, SubExprUseData>::iterator I =
1181 SubExpressionUseData.find(UniqueSubExprs[i]);
1182 assert(I != SubExpressionUseData.end() && "Entry not found?");
1183 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1184 if (I->second.notAllUsesAreFree)
1185 Result = SE->getAddExpr(Result, I->first);
1187 FreeResult = SE->getAddExpr(FreeResult, I->first);
1189 // Remove non-cse's from SubExpressionUseData.
1190 SubExpressionUseData.erase(I);
1193 if (FreeResult != Zero) {
1194 // We have some subexpressions that can be subsumed into addressing
1195 // modes in every use inside the loop. However, it's possible that
1196 // there are so many of them that the combined FreeResult cannot
1197 // be subsumed, or that the target cannot handle both a FreeResult
1198 // and a Result in the same instruction (for example because it would
1199 // require too many registers). Check this.
1200 for (unsigned i=0; i<NumUses; ++i) {
1201 if (!L->contains(Uses[i].Inst->getParent()))
1203 // We know this is an addressing mode use; if there are any uses that
1204 // are not, FreeResult would be Zero.
1205 const Type *UseTy = getAccessType(Uses[i].Inst);
1206 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1207 // FIXME: could split up FreeResult into pieces here, some hoisted
1208 // and some not. There is no obvious advantage to this.
1209 Result = SE->getAddExpr(Result, FreeResult);
1216 // If we found no CSE's, return now.
1217 if (Result == Zero) return Result;
1219 // If we still have a FreeResult, remove its subexpressions from
1220 // SubExpressionUseData. This means they will remain in the use Bases.
1221 if (FreeResult != Zero) {
1222 SeparateSubExprs(SubExprs, FreeResult, SE);
1223 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1224 std::map<SCEVHandle, SubExprUseData>::iterator I =
1225 SubExpressionUseData.find(SubExprs[j]);
1226 SubExpressionUseData.erase(I);
1231 // Otherwise, remove all of the CSE's we found from each of the base values.
1232 for (unsigned i = 0; i != NumUses; ++i) {
1233 // Uses outside the loop don't necessarily include the common base, but
1234 // the final IV value coming into those uses does. Instead of trying to
1235 // remove the pieces of the common base, which might not be there,
1236 // subtract off the base to compensate for this.
1237 if (!L->contains(Uses[i].Inst->getParent())) {
1238 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1242 // Split the expression into subexprs.
1243 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1245 // Remove any common subexpressions.
1246 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1247 if (SubExpressionUseData.count(SubExprs[j])) {
1248 SubExprs.erase(SubExprs.begin()+j);
1252 // Finally, add the non-shared expressions together.
1253 if (SubExprs.empty())
1254 Uses[i].Base = Zero;
1256 Uses[i].Base = SE->getAddExpr(SubExprs);
1263 /// ValidStride - Check whether the given Scale is valid for all loads and
1264 /// stores in UsersToProcess.
1266 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1268 const std::vector<BasedUser>& UsersToProcess) {
1272 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1273 // If this is a load or other access, pass the type of the access in.
1274 const Type *AccessTy = Type::VoidTy;
1275 if (isAddressUse(UsersToProcess[i].Inst,
1276 UsersToProcess[i].OperandValToReplace))
1277 AccessTy = getAccessType(UsersToProcess[i].Inst);
1278 else if (isa<PHINode>(UsersToProcess[i].Inst))
1281 TargetLowering::AddrMode AM;
1282 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1283 AM.BaseOffs = SC->getValue()->getSExtValue();
1284 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1287 // If load[imm+r*scale] is illegal, bail out.
1288 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1294 /// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not
1296 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1300 if (Ty1->canLosslesslyBitCastTo(Ty2))
1302 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1304 if (isa<PointerType>(Ty2) && Ty1->canLosslesslyBitCastTo(UIntPtrTy))
1306 if (isa<PointerType>(Ty1) && Ty2->canLosslesslyBitCastTo(UIntPtrTy))
1311 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1312 /// of a previous stride and it is a legal value for the target addressing
1313 /// mode scale component and optional base reg. This allows the users of
1314 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1315 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1317 /// If all uses are outside the loop, we don't require that all multiplies
1318 /// be folded into the addressing mode, nor even that the factor be constant;
1319 /// a multiply (executed once) outside the loop is better than another IV
1320 /// within. Well, usually.
1321 SCEVHandle LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1322 bool AllUsesAreAddresses,
1323 bool AllUsesAreOutsideLoop,
1324 const SCEVHandle &Stride,
1325 IVExpr &IV, const Type *Ty,
1326 const std::vector<BasedUser>& UsersToProcess) {
1327 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1328 int64_t SInt = SC->getValue()->getSExtValue();
1329 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1331 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1332 IVsByStride.find(StrideOrder[NewStride]);
1333 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1335 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1336 if (SI->first != Stride &&
1337 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1339 int64_t Scale = SInt / SSInt;
1340 // Check that this stride is valid for all the types used for loads and
1341 // stores; if it can be used for some and not others, we might as well use
1342 // the original stride everywhere, since we have to create the IV for it
1343 // anyway. If the scale is 1, then we don't need to worry about folding
1346 (AllUsesAreAddresses &&
1347 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1348 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1349 IE = SI->second.IVs.end(); II != IE; ++II)
1350 // FIXME: Only handle base == 0 for now.
1351 // Only reuse previous IV if it would not require a type conversion.
1352 if (II->Base->isZero() &&
1353 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1355 return SE->getIntegerSCEV(Scale, Stride->getType());
1358 } else if (AllUsesAreOutsideLoop) {
1359 // Accept nonconstant strides here; it is really really right to substitute
1360 // an existing IV if we can.
1361 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1363 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1364 IVsByStride.find(StrideOrder[NewStride]);
1365 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first))
1367 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1368 if (SI->first != Stride && SSInt != 1)
1370 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1371 IE = SI->second.IVs.end(); II != IE; ++II)
1372 // Accept nonzero base here.
1373 // Only reuse previous IV if it would not require a type conversion.
1374 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1379 // Special case, old IV is -1*x and this one is x. Can treat this one as
1381 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1383 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1384 IVsByStride.find(StrideOrder[NewStride]);
1385 if (SI == IVsByStride.end())
1387 if (SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first))
1388 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0)))
1389 if (Stride == ME->getOperand(1) &&
1390 SC->getValue()->getSExtValue() == -1LL)
1391 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1392 IE = SI->second.IVs.end(); II != IE; ++II)
1393 // Accept nonzero base here.
1394 // Only reuse previous IV if it would not require type conversion.
1395 if (!RequiresTypeConversion(II->Base->getType(), Ty)) {
1397 return SE->getIntegerSCEV(-1LL, Stride->getType());
1401 return SE->getIntegerSCEV(0, Stride->getType());
1404 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1405 /// returns true if Val's isUseOfPostIncrementedValue is true.
1406 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1407 return Val.isUseOfPostIncrementedValue;
1410 /// isNonConstantNegative - Return true if the specified scev is negated, but
1412 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1413 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1414 if (!Mul) return false;
1416 // If there is a constant factor, it will be first.
1417 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1418 if (!SC) return false;
1420 // Return true if the value is negative, this matches things like (-42 * V).
1421 return SC->getValue()->getValue().isNegative();
1424 // CollectIVUsers - Transform our list of users and offsets to a bit more
1425 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1426 // of the strided accesses, as well as the old information from Uses. We
1427 // progressively move information from the Base field to the Imm field, until
1428 // we eventually have the full access expression to rewrite the use.
1429 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1430 IVUsersOfOneStride &Uses,
1432 bool &AllUsesAreAddresses,
1433 bool &AllUsesAreOutsideLoop,
1434 std::vector<BasedUser> &UsersToProcess) {
1435 UsersToProcess.reserve(Uses.Users.size());
1436 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1437 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1439 // Move any loop variant operands from the offset field to the immediate
1440 // field of the use, so that we don't try to use something before it is
1442 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1443 UsersToProcess.back().Imm, L, SE);
1444 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1445 "Base value is not loop invariant!");
1448 // We now have a whole bunch of uses of like-strided induction variables, but
1449 // they might all have different bases. We want to emit one PHI node for this
1450 // stride which we fold as many common expressions (between the IVs) into as
1451 // possible. Start by identifying the common expressions in the base values
1452 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1453 // "A+B"), emit it to the preheader, then remove the expression from the
1454 // UsersToProcess base values.
1455 SCEVHandle CommonExprs =
1456 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1458 // Next, figure out what we can represent in the immediate fields of
1459 // instructions. If we can represent anything there, move it to the imm
1460 // fields of the BasedUsers. We do this so that it increases the commonality
1461 // of the remaining uses.
1462 unsigned NumPHI = 0;
1463 bool HasAddress = false;
1464 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1465 // If the user is not in the current loop, this means it is using the exit
1466 // value of the IV. Do not put anything in the base, make sure it's all in
1467 // the immediate field to allow as much factoring as possible.
1468 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1469 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1470 UsersToProcess[i].Base);
1471 UsersToProcess[i].Base =
1472 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1474 // Not all uses are outside the loop.
1475 AllUsesAreOutsideLoop = false;
1477 // Addressing modes can be folded into loads and stores. Be careful that
1478 // the store is through the expression, not of the expression though.
1480 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1481 UsersToProcess[i].OperandValToReplace);
1482 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1490 // If this use isn't an address, then not all uses are addresses.
1491 if (!isAddress && !isPHI)
1492 AllUsesAreAddresses = false;
1494 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1495 UsersToProcess[i].Imm, isAddress, L, SE);
1499 // If one of the use is a PHI node and all other uses are addresses, still
1500 // allow iv reuse. Essentially we are trading one constant multiplication
1501 // for one fewer iv.
1503 AllUsesAreAddresses = false;
1505 // There are no in-loop address uses.
1506 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop))
1507 AllUsesAreAddresses = false;
1512 /// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction
1513 /// is valid and profitable for the given set of users of a stride. In
1514 /// full strength-reduction mode, all addresses at the current stride are
1515 /// strength-reduced all the way down to pointer arithmetic.
1517 bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode(
1518 const std::vector<BasedUser> &UsersToProcess,
1520 bool AllUsesAreAddresses,
1521 SCEVHandle Stride) {
1522 if (!EnableFullLSRMode)
1525 // The heuristics below aim to avoid increasing register pressure, but
1526 // fully strength-reducing all the addresses increases the number of
1527 // add instructions, so don't do this when optimizing for size.
1528 // TODO: If the loop is large, the savings due to simpler addresses
1529 // may oughtweight the costs of the extra increment instructions.
1530 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize))
1533 // TODO: For now, don't do full strength reduction if there could
1534 // potentially be greater-stride multiples of the current stride
1535 // which could reuse the current stride IV.
1536 if (StrideOrder.back() != Stride)
1539 // Iterate through the uses to find conditions that automatically rule out
1541 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1542 SCEV *Base = UsersToProcess[i].Base;
1543 SCEV *Imm = UsersToProcess[i].Imm;
1544 // If any users have a loop-variant component, they can't be fully
1545 // strength-reduced.
1546 if (Imm && !Imm->isLoopInvariant(L))
1548 // If there are to users with the same base and the difference between
1549 // the two Imm values can't be folded into the address, full
1550 // strength reduction would increase register pressure.
1552 SCEV *CurImm = UsersToProcess[i].Imm;
1553 if ((CurImm || Imm) && CurImm != Imm) {
1554 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType());
1555 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType());
1556 const Instruction *Inst = UsersToProcess[i].Inst;
1557 const Type *UseTy = getAccessType(Inst);
1558 SCEVHandle Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1559 if (!Diff->isZero() &&
1560 (!AllUsesAreAddresses ||
1561 !fitsInAddressMode(Diff, UseTy, TLI, /*HasBaseReg=*/true)))
1564 } while (++i != e && Base == UsersToProcess[i].Base);
1567 // If there's exactly one user in this stride, fully strength-reducing it
1568 // won't increase register pressure. If it's starting from a non-zero base,
1569 // it'll be simpler this way.
1570 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero())
1573 // Otherwise, if there are any users in this stride that don't require
1574 // a register for their base, full strength-reduction will increase
1575 // register pressure.
1576 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1577 if (UsersToProcess[i].Base->isZero())
1580 // Otherwise, go for it.
1584 /// InsertAffinePhi Create and insert a PHI node for an induction variable
1585 /// with the specified start and step values in the specified loop.
1587 /// If NegateStride is true, the stride should be negated by using a
1588 /// subtract instead of an add.
1590 /// Return the created phi node, and return the step instruction by
1591 /// reference in IncV.
1593 static PHINode *InsertAffinePhi(SCEVHandle Start, SCEVHandle Step,
1595 SCEVExpander &Rewriter,
1597 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!");
1598 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!");
1600 BasicBlock *Header = L->getHeader();
1601 BasicBlock *Preheader = L->getLoopPreheader();
1602 BasicBlock *LatchBlock = L->getLoopLatch();
1604 PHINode *PN = PHINode::Create(Start->getType(), "lsr.iv", Header->begin());
1605 PN->addIncoming(Rewriter.expandCodeFor(Start, Preheader->getTerminator()),
1608 // If the stride is negative, insert a sub instead of an add for the
1610 bool isNegative = isNonConstantNegative(Step);
1611 SCEVHandle IncAmount = Step;
1613 IncAmount = Rewriter.SE.getNegativeSCEV(Step);
1615 // Insert an add instruction right before the terminator corresponding
1616 // to the back-edge.
1617 Value *StepV = Rewriter.expandCodeFor(IncAmount, Preheader->getTerminator());
1619 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next",
1620 LatchBlock->getTerminator());
1622 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next",
1623 LatchBlock->getTerminator());
1625 if (!isa<ConstantInt>(StepV)) ++NumVariable;
1627 PN->addIncoming(IncV, LatchBlock);
1633 static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) {
1634 // We want to emit code for users inside the loop first. To do this, we
1635 // rearrange BasedUser so that the entries at the end have
1636 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1637 // vector (so we handle them first).
1638 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1639 PartitionByIsUseOfPostIncrementedValue);
1641 // Sort this by base, so that things with the same base are handled
1642 // together. By partitioning first and stable-sorting later, we are
1643 // guaranteed that within each base we will pop off users from within the
1644 // loop before users outside of the loop with a particular base.
1646 // We would like to use stable_sort here, but we can't. The problem is that
1647 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1648 // we don't have anything to do a '<' comparison on. Because we think the
1649 // number of uses is small, do a horrible bubble sort which just relies on
1651 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1652 // Get a base value.
1653 SCEVHandle Base = UsersToProcess[i].Base;
1655 // Compact everything with this base to be consecutive with this one.
1656 for (unsigned j = i+1; j != e; ++j) {
1657 if (UsersToProcess[j].Base == Base) {
1658 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1665 /// PrepareToStrengthReduceFully - Prepare to fully strength-reduce
1666 /// UsersToProcess, meaning lowering addresses all the way down to direct
1667 /// pointer arithmetic.
1670 LoopStrengthReduce::PrepareToStrengthReduceFully(
1671 std::vector<BasedUser> &UsersToProcess,
1673 SCEVHandle CommonExprs,
1675 SCEVExpander &PreheaderRewriter) {
1676 DOUT << " Fully reducing all users\n";
1678 // Rewrite the UsersToProcess records, creating a separate PHI for each
1679 // unique Base value.
1680 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) {
1681 // TODO: The uses are grouped by base, but not sorted. We arbitrarily
1682 // pick the first Imm value here to start with, and adjust it for the
1684 SCEVHandle Imm = UsersToProcess[i].Imm;
1685 SCEVHandle Base = UsersToProcess[i].Base;
1686 SCEVHandle Start = SE->getAddExpr(CommonExprs, Base, Imm);
1688 PHINode *Phi = InsertAffinePhi(Start, Stride, L,
1691 // Loop over all the users with the same base.
1693 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType());
1694 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm);
1695 UsersToProcess[i].Phi = Phi;
1696 UsersToProcess[i].IncV = IncV;
1697 assert(UsersToProcess[i].Imm->isLoopInvariant(L) &&
1698 "ShouldUseFullStrengthReductionMode should reject this!");
1699 } while (++i != e && Base == UsersToProcess[i].Base);
1703 /// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the
1704 /// given users to share.
1707 LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi(
1708 std::vector<BasedUser> &UsersToProcess,
1710 SCEVHandle CommonExprs,
1713 SCEVExpander &PreheaderRewriter) {
1714 DOUT << " Inserting new PHI:\n";
1717 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV),
1722 // Remember this in case a later stride is multiple of this.
1723 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi, IncV);
1725 // All the users will share this new IV.
1726 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1727 UsersToProcess[i].Phi = Phi;
1728 UsersToProcess[i].IncV = IncV;
1732 DEBUG(WriteAsOperand(*DOUT, Phi, /*PrintType=*/false));
1734 DEBUG(WriteAsOperand(*DOUT, IncV, /*PrintType=*/false));
1738 /// PrepareToStrengthReduceWithNewPhi - Prepare for the given users to reuse
1739 /// an induction variable with a stride that is a factor of the current
1740 /// induction variable.
1743 LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride(
1744 std::vector<BasedUser> &UsersToProcess,
1746 const IVExpr &ReuseIV,
1747 Instruction *PreInsertPt) {
1748 DOUT << " Rewriting in terms of existing IV of STRIDE " << *ReuseIV.Stride
1749 << " and BASE " << *ReuseIV.Base << "\n";
1751 // All the users will share the reused IV.
1752 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1753 UsersToProcess[i].Phi = ReuseIV.PHI;
1754 UsersToProcess[i].IncV = ReuseIV.IncV;
1757 Constant *C = dyn_cast<Constant>(CommonBaseV);
1759 (!C->isNullValue() &&
1760 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(),
1762 // We want the common base emitted into the preheader! This is just
1763 // using cast as a copy so BitCast (no-op cast) is appropriate
1764 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1765 "commonbase", PreInsertPt);
1768 static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset,
1769 const Type *AccessTy,
1770 std::vector<BasedUser> &UsersToProcess,
1771 const TargetLowering *TLI) {
1772 SmallVector<Instruction*, 16> AddrModeInsts;
1773 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1774 if (UsersToProcess[i].isUseOfPostIncrementedValue)
1776 ExtAddrMode AddrMode =
1777 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace,
1778 AccessTy, UsersToProcess[i].Inst,
1779 AddrModeInsts, *TLI);
1780 if (GV && GV != AddrMode.BaseGV)
1782 if (Offset && !AddrMode.BaseOffs)
1783 // FIXME: How to accurate check it's immediate offset is folded.
1785 AddrModeInsts.clear();
1790 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1791 /// stride of IV. All of the users may have different starting values, and this
1792 /// may not be the only stride.
1793 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1794 IVUsersOfOneStride &Uses,
1796 // If all the users are moved to another stride, then there is nothing to do.
1797 if (Uses.Users.empty())
1800 // Keep track if every use in UsersToProcess is an address. If they all are,
1801 // we may be able to rewrite the entire collection of them in terms of a
1802 // smaller-stride IV.
1803 bool AllUsesAreAddresses = true;
1805 // Keep track if every use of a single stride is outside the loop. If so,
1806 // we want to be more aggressive about reusing a smaller-stride IV; a
1807 // multiply outside the loop is better than another IV inside. Well, usually.
1808 bool AllUsesAreOutsideLoop = true;
1810 // Transform our list of users and offsets to a bit more complex table. In
1811 // this new vector, each 'BasedUser' contains 'Base' the base of the
1812 // strided accessas well as the old information from Uses. We progressively
1813 // move information from the Base field to the Imm field, until we eventually
1814 // have the full access expression to rewrite the use.
1815 std::vector<BasedUser> UsersToProcess;
1816 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1817 AllUsesAreOutsideLoop,
1820 // Sort the UsersToProcess array so that users with common bases are
1821 // next to each other.
1822 SortUsersToProcess(UsersToProcess);
1824 // If we managed to find some expressions in common, we'll need to carry
1825 // their value in a register and add it in for each use. This will take up
1826 // a register operand, which potentially restricts what stride values are
1828 bool HaveCommonExprs = !CommonExprs->isZero();
1830 const Type *ReplacedTy = CommonExprs->getType();
1832 // If all uses are addresses, consider sinking the immediate part of the
1833 // common expression back into uses if they can fit in the immediate fields.
1834 if (TLI && HaveCommonExprs && AllUsesAreAddresses) {
1835 SCEVHandle NewCommon = CommonExprs;
1836 SCEVHandle Imm = SE->getIntegerSCEV(0, ReplacedTy);
1837 MoveImmediateValues(TLI, Type::VoidTy, NewCommon, Imm, true, L, SE);
1838 if (!Imm->isZero()) {
1841 // If the immediate part of the common expression is a GV, check if it's
1842 // possible to fold it into the target addressing mode.
1843 GlobalValue *GV = 0;
1844 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm)) {
1845 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
1846 if (CE->getOpcode() == Instruction::PtrToInt)
1847 GV = dyn_cast<GlobalValue>(CE->getOperand(0));
1850 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
1851 Offset = SC->getValue()->getSExtValue();
1853 // Pass VoidTy as the AccessTy to be conservative, because
1854 // there could be multiple access types among all the uses.
1855 DoSink = IsImmFoldedIntoAddrMode(GV, Offset, Type::VoidTy,
1856 UsersToProcess, TLI);
1859 DOUT << " Sinking " << *Imm << " back down into uses\n";
1860 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i)
1861 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm);
1862 CommonExprs = NewCommon;
1863 HaveCommonExprs = !CommonExprs->isZero();
1869 // Now that we know what we need to do, insert the PHI node itself.
1871 DOUT << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE "
1873 << " Common base: " << *CommonExprs << "\n";
1875 SCEVExpander Rewriter(*SE, *LI);
1876 SCEVExpander PreheaderRewriter(*SE, *LI);
1878 BasicBlock *Preheader = L->getLoopPreheader();
1879 Instruction *PreInsertPt = Preheader->getTerminator();
1880 BasicBlock *LatchBlock = L->getLoopLatch();
1882 Value *CommonBaseV = ConstantInt::get(ReplacedTy, 0);
1884 SCEVHandle RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy);
1885 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1886 SE->getIntegerSCEV(0, Type::Int32Ty),
1889 /// Choose a strength-reduction strategy and prepare for it by creating
1890 /// the necessary PHIs and adjusting the bookkeeping.
1891 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L,
1892 AllUsesAreAddresses, Stride)) {
1893 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L,
1896 // Emit the initial base value into the loop preheader.
1897 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1899 // If all uses are addresses, check if it is possible to reuse an IV with a
1900 // stride that is a factor of this stride. And that the multiple is a number
1901 // that can be encoded in the scale field of the target addressing mode. And
1902 // that we will have a valid instruction after this substition, including
1903 // the immediate field, if any.
1904 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1905 AllUsesAreOutsideLoop,
1906 Stride, ReuseIV, ReplacedTy,
1908 if (isa<SCEVConstant>(RewriteFactor) &&
1909 cast<SCEVConstant>(RewriteFactor)->isZero())
1910 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs,
1911 CommonBaseV, L, PreheaderRewriter);
1913 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV,
1914 ReuseIV, PreInsertPt);
1917 // Process all the users now, replacing their strided uses with
1918 // strength-reduced forms. This outer loop handles all bases, the inner
1919 // loop handles all users of a particular base.
1920 while (!UsersToProcess.empty()) {
1921 SCEVHandle Base = UsersToProcess.back().Base;
1922 Instruction *Inst = UsersToProcess.back().Inst;
1924 // Emit the code for Base into the preheader.
1925 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1927 DOUT << " Examining uses with BASE ";
1928 DEBUG(WriteAsOperand(*DOUT, BaseV, /*PrintType=*/false));
1931 // If BaseV is a constant other than 0, make sure that it gets inserted into
1932 // the preheader, instead of being forward substituted into the uses. We do
1933 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1935 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1936 if (!C->isNullValue() && !fitsInAddressMode(Base, getAccessType(Inst),
1938 // We want this constant emitted into the preheader! This is just
1939 // using cast as a copy so BitCast (no-op cast) is appropriate
1940 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1945 // Emit the code to add the immediate offset to the Phi value, just before
1946 // the instructions that we identified as using this stride and base.
1948 // FIXME: Use emitted users to emit other users.
1949 BasedUser &User = UsersToProcess.back();
1951 DOUT << " Examining use ";
1952 DEBUG(WriteAsOperand(*DOUT, UsersToProcess.back().OperandValToReplace,
1953 /*PrintType=*/false));
1954 DOUT << " in Inst: " << *Inst;
1956 // If this instruction wants to use the post-incremented value, move it
1957 // after the post-inc and use its value instead of the PHI.
1958 Value *RewriteOp = User.Phi;
1959 if (User.isUseOfPostIncrementedValue) {
1960 RewriteOp = User.IncV;
1962 // If this user is in the loop, make sure it is the last thing in the
1963 // loop to ensure it is dominated by the increment.
1964 if (L->contains(User.Inst->getParent()))
1965 User.Inst->moveBefore(LatchBlock->getTerminator());
1967 if (RewriteOp->getType() != ReplacedTy) {
1968 Instruction::CastOps opcode = Instruction::Trunc;
1969 if (ReplacedTy->getPrimitiveSizeInBits() ==
1970 RewriteOp->getType()->getPrimitiveSizeInBits())
1971 opcode = Instruction::BitCast;
1972 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1975 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1977 // If we had to insert new instructions for RewriteOp, we have to
1978 // consider that they may not have been able to end up immediately
1979 // next to RewriteOp, because non-PHI instructions may never precede
1980 // PHI instructions in a block. In this case, remember where the last
1981 // instruction was inserted so that if we're replacing a different
1982 // PHI node, we can use the later point to expand the final
1984 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1985 if (RewriteOp == User.Phi) NewBasePt = 0;
1987 // Clear the SCEVExpander's expression map so that we are guaranteed
1988 // to have the code emitted where we expect it.
1991 // If we are reusing the iv, then it must be multiplied by a constant
1992 // factor to take advantage of the addressing mode scale component.
1993 if (!isa<SCEVConstant>(RewriteFactor) ||
1994 !cast<SCEVConstant>(RewriteFactor)->isZero()) {
1995 // If we're reusing an IV with a nonzero base (currently this happens
1996 // only when all reuses are outside the loop) subtract that base here.
1997 // The base has been used to initialize the PHI node but we don't want
1999 if (!ReuseIV.Base->isZero()) {
2000 SCEVHandle typedBase = ReuseIV.Base;
2001 if (RewriteExpr->getType()->getPrimitiveSizeInBits() !=
2002 ReuseIV.Base->getType()->getPrimitiveSizeInBits()) {
2003 // It's possible the original IV is a larger type than the new IV,
2004 // in which case we have to truncate the Base. We checked in
2005 // RequiresTypeConversion that this is valid.
2006 assert (RewriteExpr->getType()->getPrimitiveSizeInBits() <
2007 ReuseIV.Base->getType()->getPrimitiveSizeInBits() &&
2008 "Unexpected lengthening conversion!");
2009 typedBase = SE->getTruncateExpr(ReuseIV.Base,
2010 RewriteExpr->getType());
2012 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase);
2015 // Multiply old variable, with base removed, by new scale factor.
2016 RewriteExpr = SE->getMulExpr(RewriteFactor,
2019 // The common base is emitted in the loop preheader. But since we
2020 // are reusing an IV, it has not been used to initialize the PHI node.
2021 // Add it to the expression used to rewrite the uses.
2022 // When this use is outside the loop, we earlier subtracted the
2023 // common base, and are adding it back here. Use the same expression
2024 // as before, rather than CommonBaseV, so DAGCombiner will zap it.
2025 if (!isa<ConstantInt>(CommonBaseV) ||
2026 !cast<ConstantInt>(CommonBaseV)->isZero()) {
2027 if (L->contains(User.Inst->getParent()))
2028 RewriteExpr = SE->getAddExpr(RewriteExpr,
2029 SE->getUnknown(CommonBaseV));
2031 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs);
2035 // Now that we know what we need to do, insert code before User for the
2036 // immediate and any loop-variant expressions.
2037 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
2038 // Add BaseV to the PHI value if needed.
2039 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
2041 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
2045 // Mark old value we replaced as possibly dead, so that it is eliminated
2046 // if we just replaced the last use of that value.
2047 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
2049 UsersToProcess.pop_back();
2052 // If there are any more users to process with the same base, process them
2053 // now. We sorted by base above, so we just have to check the last elt.
2054 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
2055 // TODO: Next, find out which base index is the most common, pull it out.
2058 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
2059 // different starting values, into different PHIs.
2062 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
2063 /// set the IV user and stride information and return true, otherwise return
2065 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
2066 const SCEVHandle *&CondStride) {
2067 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
2069 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2070 IVUsesByStride.find(StrideOrder[Stride]);
2071 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2073 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2074 E = SI->second.Users.end(); UI != E; ++UI)
2075 if (UI->User == Cond) {
2076 // NOTE: we could handle setcc instructions with multiple uses here, but
2077 // InstCombine does it as well for simple uses, it's not clear that it
2078 // occurs enough in real life to handle.
2080 CondStride = &SI->first;
2088 // Constant strides come first which in turns are sorted by their absolute
2089 // values. If absolute values are the same, then positive strides comes first.
2091 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
2092 struct StrideCompare {
2093 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
2094 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
2095 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
2097 int64_t LV = LHSC->getValue()->getSExtValue();
2098 int64_t RV = RHSC->getValue()->getSExtValue();
2099 uint64_t ALV = (LV < 0) ? -LV : LV;
2100 uint64_t ARV = (RV < 0) ? -RV : RV;
2108 // If it's the same value but different type, sort by bit width so
2109 // that we emit larger induction variables before smaller
2110 // ones, letting the smaller be re-written in terms of larger ones.
2111 return RHS->getBitWidth() < LHS->getBitWidth();
2113 return LHSC && !RHSC;
2118 /// ChangeCompareStride - If a loop termination compare instruction is the
2119 /// only use of its stride, and the compaison is against a constant value,
2120 /// try eliminate the stride by moving the compare instruction to another
2121 /// stride and change its constant operand accordingly. e.g.
2127 /// if (v2 < 10) goto loop
2132 /// if (v1 < 30) goto loop
2133 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
2134 IVStrideUse* &CondUse,
2135 const SCEVHandle* &CondStride) {
2136 if (StrideOrder.size() < 2 ||
2137 IVUsesByStride[*CondStride].Users.size() != 1)
2139 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
2140 if (!SC) return Cond;
2142 ICmpInst::Predicate Predicate = Cond->getPredicate();
2143 int64_t CmpSSInt = SC->getValue()->getSExtValue();
2144 unsigned BitWidth = (*CondStride)->getBitWidth();
2145 uint64_t SignBit = 1ULL << (BitWidth-1);
2146 const Type *CmpTy = Cond->getOperand(0)->getType();
2147 const Type *NewCmpTy = NULL;
2148 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
2149 unsigned NewTyBits = 0;
2150 SCEVHandle *NewStride = NULL;
2151 Value *NewCmpLHS = NULL;
2152 Value *NewCmpRHS = NULL;
2154 SCEVHandle NewOffset = SE->getIntegerSCEV(0, UIntPtrTy);
2156 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) {
2157 int64_t CmpVal = C->getValue().getSExtValue();
2159 // Check stride constant and the comparision constant signs to detect
2161 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
2164 // Look for a suitable stride / iv as replacement.
2165 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
2166 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2167 IVUsesByStride.find(StrideOrder[i]);
2168 if (!isa<SCEVConstant>(SI->first))
2170 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
2171 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
2174 Scale = SSInt / CmpSSInt;
2175 int64_t NewCmpVal = CmpVal * Scale;
2176 APInt Mul = APInt(BitWidth, NewCmpVal);
2177 // Check for overflow.
2178 if (Mul.getSExtValue() != NewCmpVal)
2181 // Watch out for overflow.
2182 if (ICmpInst::isSignedPredicate(Predicate) &&
2183 (CmpVal & SignBit) != (NewCmpVal & SignBit))
2186 if (NewCmpVal == CmpVal)
2188 // Pick the best iv to use trying to avoid a cast.
2190 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2191 E = SI->second.Users.end(); UI != E; ++UI) {
2192 NewCmpLHS = UI->OperandValToReplace;
2193 if (NewCmpLHS->getType() == CmpTy)
2199 NewCmpTy = NewCmpLHS->getType();
2200 NewTyBits = isa<PointerType>(NewCmpTy)
2201 ? UIntPtrTy->getPrimitiveSizeInBits()
2202 : NewCmpTy->getPrimitiveSizeInBits();
2203 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
2204 // Check if it is possible to rewrite it using
2205 // an iv / stride of a smaller integer type.
2206 bool TruncOk = false;
2207 if (NewCmpTy->isInteger()) {
2208 unsigned Bits = NewTyBits;
2209 if (ICmpInst::isSignedPredicate(Predicate))
2211 uint64_t Mask = (1ULL << Bits) - 1;
2212 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
2219 // Don't rewrite if use offset is non-constant and the new type is
2220 // of a different type.
2221 // FIXME: too conservative?
2222 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset))
2225 bool AllUsesAreAddresses = true;
2226 bool AllUsesAreOutsideLoop = true;
2227 std::vector<BasedUser> UsersToProcess;
2228 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
2229 AllUsesAreAddresses,
2230 AllUsesAreOutsideLoop,
2232 // Avoid rewriting the compare instruction with an iv of new stride
2233 // if it's likely the new stride uses will be rewritten using the
2234 // stride of the compare instruction.
2235 if (AllUsesAreAddresses &&
2236 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess))
2239 // If scale is negative, use swapped predicate unless it's testing
2241 if (Scale < 0 && !Cond->isEquality())
2242 Predicate = ICmpInst::getSwappedPredicate(Predicate);
2244 NewStride = &StrideOrder[i];
2245 if (!isa<PointerType>(NewCmpTy))
2246 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal);
2248 NewCmpRHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
2249 NewCmpRHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
2250 NewCmpRHS, NewCmpTy);
2252 NewOffset = TyBits == NewTyBits
2253 ? SE->getMulExpr(CondUse->Offset,
2254 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
2255 : SE->getConstant(ConstantInt::get(NewCmpTy,
2256 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
2261 // Forgo this transformation if it the increment happens to be
2262 // unfortunately positioned after the condition, and the condition
2263 // has multiple uses which prevent it from being moved immediately
2264 // before the branch. See
2265 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
2266 // for an example of this situation.
2267 if (!Cond->hasOneUse()) {
2268 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
2275 // Create a new compare instruction using new stride / iv.
2276 ICmpInst *OldCond = Cond;
2277 // Insert new compare instruction.
2278 Cond = new ICmpInst(Predicate, NewCmpLHS, NewCmpRHS,
2279 L->getHeader()->getName() + ".termcond",
2282 // Remove the old compare instruction. The old indvar is probably dead too.
2283 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
2284 SE->deleteValueFromRecords(OldCond);
2285 OldCond->replaceAllUsesWith(Cond);
2286 OldCond->eraseFromParent();
2288 IVUsesByStride[*CondStride].Users.pop_back();
2289 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewCmpLHS);
2290 CondUse = &IVUsesByStride[*NewStride].Users.back();
2291 CondStride = NewStride;
2298 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
2299 /// an smax computation.
2301 /// This is a narrow solution to a specific, but acute, problem. For loops
2307 /// } while (++i < n);
2309 /// where the comparison is signed, the trip count isn't just 'n', because
2310 /// 'n' could be negative. And unfortunately this can come up even for loops
2311 /// where the user didn't use a C do-while loop. For example, seemingly
2312 /// well-behaved top-test loops will commonly be lowered like this:
2318 /// } while (++i < n);
2321 /// and then it's possible for subsequent optimization to obscure the if
2322 /// test in such a way that indvars can't find it.
2324 /// When indvars can't find the if test in loops like this, it creates a
2325 /// signed-max expression, which allows it to give the loop a canonical
2326 /// induction variable:
2329 /// smax = n < 1 ? 1 : n;
2332 /// } while (++i != smax);
2334 /// Canonical induction variables are necessary because the loop passes
2335 /// are designed around them. The most obvious example of this is the
2336 /// LoopInfo analysis, which doesn't remember trip count values. It
2337 /// expects to be able to rediscover the trip count each time it is
2338 /// needed, and it does this using a simple analyis that only succeeds if
2339 /// the loop has a canonical induction variable.
2341 /// However, when it comes time to generate code, the maximum operation
2342 /// can be quite costly, especially if it's inside of an outer loop.
2344 /// This function solves this problem by detecting this type of loop and
2345 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
2346 /// the instructions for the maximum computation.
2348 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
2349 IVStrideUse* &CondUse) {
2350 // Check that the loop matches the pattern we're looking for.
2351 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
2352 Cond->getPredicate() != CmpInst::ICMP_NE)
2355 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
2356 if (!Sel || !Sel->hasOneUse()) return Cond;
2358 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2359 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2361 SCEVHandle One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType());
2363 // Add one to the backedge-taken count to get the trip count.
2364 SCEVHandle IterationCount = SE->getAddExpr(BackedgeTakenCount, One);
2366 // Check for a max calculation that matches the pattern.
2367 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
2368 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
2370 SCEVHandle SMaxLHS = SMax->getOperand(0);
2371 SCEVHandle SMaxRHS = SMax->getOperand(1);
2372 if (!SMaxLHS || SMaxLHS != One) return Cond;
2374 // Check the relevant induction variable for conformance to
2376 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
2377 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
2378 if (!AR || !AR->isAffine() ||
2379 AR->getStart() != One ||
2380 AR->getStepRecurrence(*SE) != One)
2383 assert(AR->getLoop() == L &&
2384 "Loop condition operand is an addrec in a different loop!");
2386 // Check the right operand of the select, and remember it, as it will
2387 // be used in the new comparison instruction.
2389 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
2390 NewRHS = Sel->getOperand(1);
2391 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
2392 NewRHS = Sel->getOperand(2);
2393 if (!NewRHS) return Cond;
2395 // Ok, everything looks ok to change the condition into an SLT or SGE and
2396 // delete the max calculation.
2398 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
2401 Cond->getOperand(0), NewRHS, "scmp", Cond);
2403 // Delete the max calculation instructions.
2404 SE->deleteValueFromRecords(Cond);
2405 Cond->replaceAllUsesWith(NewCond);
2406 Cond->eraseFromParent();
2407 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
2408 SE->deleteValueFromRecords(Sel);
2409 Sel->eraseFromParent();
2410 if (Cmp->use_empty()) {
2411 SE->deleteValueFromRecords(Cmp);
2412 Cmp->eraseFromParent();
2414 CondUse->User = NewCond;
2418 /// OptimizeShadowIV - If IV is used in a int-to-float cast
2419 /// inside the loop then try to eliminate the cast opeation.
2420 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
2422 SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2423 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2426 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
2428 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2429 IVUsesByStride.find(StrideOrder[Stride]);
2430 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2431 if (!isa<SCEVConstant>(SI->first))
2434 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
2435 E = SI->second.Users.end(); UI != E; /* empty */) {
2436 std::vector<IVStrideUse>::iterator CandidateUI = UI;
2438 Instruction *ShadowUse = CandidateUI->User;
2439 const Type *DestTy = NULL;
2441 /* If shadow use is a int->float cast then insert a second IV
2442 to eliminate this cast.
2444 for (unsigned i = 0; i < n; ++i)
2450 for (unsigned i = 0; i < n; ++i, ++d)
2453 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
2454 DestTy = UCast->getDestTy();
2455 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
2456 DestTy = SCast->getDestTy();
2457 if (!DestTy) continue;
2460 /* If target does not support DestTy natively then do not apply
2461 this transformation. */
2462 MVT DVT = TLI->getValueType(DestTy);
2463 if (!TLI->isTypeLegal(DVT)) continue;
2466 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
2468 if (PH->getNumIncomingValues() != 2) continue;
2470 const Type *SrcTy = PH->getType();
2471 int Mantissa = DestTy->getFPMantissaWidth();
2472 if (Mantissa == -1) continue;
2473 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
2476 unsigned Entry, Latch;
2477 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2485 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2486 if (!Init) continue;
2487 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2489 BinaryOperator *Incr =
2490 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2491 if (!Incr) continue;
2492 if (Incr->getOpcode() != Instruction::Add
2493 && Incr->getOpcode() != Instruction::Sub)
2496 /* Initialize new IV, double d = 0.0 in above example. */
2497 ConstantInt *C = NULL;
2498 if (Incr->getOperand(0) == PH)
2499 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2500 else if (Incr->getOperand(1) == PH)
2501 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2507 /* Add new PHINode. */
2508 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2510 /* create new increment. '++d' in above example. */
2511 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2512 BinaryOperator *NewIncr =
2513 BinaryOperator::Create(Incr->getOpcode(),
2514 NewPH, CFP, "IV.S.next.", Incr);
2516 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2517 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2519 /* Remove cast operation */
2520 SE->deleteValueFromRecords(ShadowUse);
2521 ShadowUse->replaceAllUsesWith(NewPH);
2522 ShadowUse->eraseFromParent();
2523 SI->second.Users.erase(CandidateUI);
2530 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2531 // uses in the loop, look to see if we can eliminate some, in favor of using
2532 // common indvars for the different uses.
2533 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2534 // TODO: implement optzns here.
2536 OptimizeShadowIV(L);
2538 // Finally, get the terminating condition for the loop if possible. If we
2539 // can, we want to change it to use a post-incremented version of its
2540 // induction variable, to allow coalescing the live ranges for the IV into
2541 // one register value.
2542 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2543 BasicBlock *Preheader = L->getLoopPreheader();
2544 BasicBlock *LatchBlock =
2545 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2546 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2547 if (!TermBr || TermBr->isUnconditional() ||
2548 !isa<ICmpInst>(TermBr->getCondition()))
2550 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2552 // Search IVUsesByStride to find Cond's IVUse if there is one.
2553 IVStrideUse *CondUse = 0;
2554 const SCEVHandle *CondStride = 0;
2556 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2557 return; // setcc doesn't use the IV.
2559 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2560 // being unable to find a sufficient guard, for example), change the loop
2561 // comparison to use SLT instead of NE.
2562 Cond = OptimizeSMax(L, Cond, CondUse);
2564 // If possible, change stride and operands of the compare instruction to
2565 // eliminate one stride.
2566 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2568 // It's possible for the setcc instruction to be anywhere in the loop, and
2569 // possible for it to have multiple users. If it is not immediately before
2570 // the latch block branch, move it.
2571 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2572 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2573 Cond->moveBefore(TermBr);
2575 // Otherwise, clone the terminating condition and insert into the loopend.
2576 Cond = cast<ICmpInst>(Cond->clone());
2577 Cond->setName(L->getHeader()->getName() + ".termcond");
2578 LatchBlock->getInstList().insert(TermBr, Cond);
2580 // Clone the IVUse, as the old use still exists!
2581 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2582 CondUse->OperandValToReplace);
2583 CondUse = &IVUsesByStride[*CondStride].Users.back();
2587 // If we get to here, we know that we can transform the setcc instruction to
2588 // use the post-incremented version of the IV, allowing us to coalesce the
2589 // live ranges for the IV correctly.
2590 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2591 CondUse->isUseOfPostIncrementedValue = true;
2595 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2597 LI = &getAnalysis<LoopInfo>();
2598 DT = &getAnalysis<DominatorTree>();
2599 SE = &getAnalysis<ScalarEvolution>();
2600 TD = &getAnalysis<TargetData>();
2601 UIntPtrTy = TD->getIntPtrType();
2604 // Find all uses of induction variables in this loop, and categorize
2605 // them by stride. Start by finding all of the PHI nodes in the header for
2606 // this loop. If they are induction variables, inspect their uses.
2607 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2608 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2609 AddUsersIfInteresting(I, L, Processed);
2611 if (!IVUsesByStride.empty()) {
2613 DOUT << "\nLSR on \"" << L->getHeader()->getParent()->getNameStart()
2618 // Sort the StrideOrder so we process larger strides first.
2619 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2621 // Optimize induction variables. Some indvar uses can be transformed to use
2622 // strides that will be needed for other purposes. A common example of this
2623 // is the exit test for the loop, which can often be rewritten to use the
2624 // computation of some other indvar to decide when to terminate the loop.
2627 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2628 // doing computation in byte values, promote to 32-bit values if safe.
2630 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2631 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2632 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2633 // Need to be careful that IV's are all the same type. Only works for
2634 // intptr_t indvars.
2636 // IVsByStride keeps IVs for one particular loop.
2637 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2639 // Note: this processes each stride/type pair individually. All users
2640 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2641 // Also, note that we iterate over IVUsesByStride indirectly by using
2642 // StrideOrder. This extra layer of indirection makes the ordering of
2643 // strides deterministic - not dependent on map order.
2644 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2645 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2646 IVUsesByStride.find(StrideOrder[Stride]);
2647 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2648 StrengthReduceStridedIVUsers(SI->first, SI->second, L);
2652 // We're done analyzing this loop; release all the state we built up for it.
2653 CastedPointers.clear();
2654 IVUsesByStride.clear();
2655 IVsByStride.clear();
2656 StrideOrder.clear();
2657 for (unsigned i=0; i<GEPlist.size(); i++)
2658 SE->deleteValueFromRecords(GEPlist[i]);
2661 // Clean up after ourselves
2662 if (!DeadInsts.empty()) {
2663 DeleteTriviallyDeadInstructions();
2665 BasicBlock::iterator I = L->getHeader()->begin();
2666 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2667 // At this point, we know that we have killed one or more IV users.
2668 // It is worth checking to see if the cannonical indvar is also
2669 // dead, so that we can remove it as well.
2671 // We can remove a PHI if it is on a cycle in the def-use graph
2672 // where each node in the cycle has degree one, i.e. only one use,
2673 // and is an instruction with no side effects.
2675 // FIXME: this needs to eliminate an induction variable even if it's being
2676 // compared against some value to decide loop termination.
2677 if (!PN->hasOneUse())
2680 SmallPtrSet<PHINode *, 4> PHIs;
2681 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2682 J && J->hasOneUse() && !J->mayWriteToMemory();
2683 J = dyn_cast<Instruction>(*J->use_begin())) {
2684 // If we find the original PHI, we've discovered a cycle.
2686 // Break the cycle and mark the PHI for deletion.
2687 SE->deleteValueFromRecords(PN);
2688 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2689 DeadInsts.push_back(PN);
2693 // If we find a PHI more than once, we're on a cycle that
2694 // won't prove fruitful.
2695 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2699 DeleteTriviallyDeadInstructions();