1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs a strength reduction on array references inside loops that
11 // have as one or more of their components the loop induction variable. This is
12 // accomplished by creating a new Value to hold the initial value of the array
13 // access for the first iteration, and then creating a new GEP instruction in
14 // the loop to increment the value by the appropriate amount.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "loop-reduce"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Type.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/ScalarEvolutionExpander.h"
29 #include "llvm/Support/CFG.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Target/TargetLowering.h"
43 STATISTIC(NumReduced , "Number of GEPs strength reduced");
44 STATISTIC(NumInserted, "Number of PHIs inserted");
45 STATISTIC(NumVariable, "Number of PHIs with variable strides");
46 STATISTIC(NumEliminated, "Number of strides eliminated");
47 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
53 /// IVStrideUse - Keep track of one use of a strided induction variable, where
54 /// the stride is stored externally. The Offset member keeps track of the
55 /// offset from the IV, User is the actual user of the operand, and
56 /// 'OperandValToReplace' is the operand of the User that is the use.
57 struct VISIBILITY_HIDDEN IVStrideUse {
60 Value *OperandValToReplace;
62 // isUseOfPostIncrementedValue - True if this should use the
63 // post-incremented version of this IV, not the preincremented version.
64 // This can only be set in special cases, such as the terminating setcc
65 // instruction for a loop or uses dominated by the loop.
66 bool isUseOfPostIncrementedValue;
68 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
69 : Offset(Offs), User(U), OperandValToReplace(O),
70 isUseOfPostIncrementedValue(false) {}
73 /// IVUsersOfOneStride - This structure keeps track of all instructions that
74 /// have an operand that is based on the trip count multiplied by some stride.
75 /// The stride for all of these users is common and kept external to this
77 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
78 /// Users - Keep track of all of the users of this stride as well as the
79 /// initial value and the operand that uses the IV.
80 std::vector<IVStrideUse> Users;
82 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
83 Users.push_back(IVStrideUse(Offset, User, Operand));
87 /// IVInfo - This structure keeps track of one IV expression inserted during
88 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
89 /// well as the PHI node and increment value created for rewrite.
90 struct VISIBILITY_HIDDEN IVExpr {
96 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
98 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
101 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
102 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
103 struct VISIBILITY_HIDDEN IVsOfOneStride {
104 std::vector<IVExpr> IVs;
106 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
108 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
112 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
116 const TargetData *TD;
117 const Type *UIntPtrTy;
120 /// IVUsesByStride - Keep track of all uses of induction variables that we
121 /// are interested in. The key of the map is the stride of the access.
122 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
124 /// IVsByStride - Keep track of all IVs that have been inserted for a
125 /// particular stride.
126 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
128 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
129 /// We use this to iterate over the IVUsesByStride collection without being
130 /// dependent on random ordering of pointers in the process.
131 SmallVector<SCEVHandle, 16> StrideOrder;
133 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
134 /// of the casted version of each value. This is accessed by
135 /// getCastedVersionOf.
136 DenseMap<Value*, Value*> CastedPointers;
138 /// DeadInsts - Keep track of instructions we may have made dead, so that
139 /// we can remove them after we are done working.
140 SmallVector<Instruction*, 16> DeadInsts;
142 /// TLI - Keep a pointer of a TargetLowering to consult for determining
143 /// transformation profitability.
144 const TargetLowering *TLI;
147 static char ID; // Pass ID, replacement for typeid
148 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
149 LoopPass(&ID), TLI(tli) {
152 bool runOnLoop(Loop *L, LPPassManager &LPM);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 // We split critical edges, so we change the CFG. However, we do update
156 // many analyses if they are around.
157 AU.addPreservedID(LoopSimplifyID);
158 AU.addPreserved<LoopInfo>();
159 AU.addPreserved<DominanceFrontier>();
160 AU.addPreserved<DominatorTree>();
162 AU.addRequiredID(LoopSimplifyID);
163 AU.addRequired<LoopInfo>();
164 AU.addRequired<DominatorTree>();
165 AU.addRequired<TargetData>();
166 AU.addRequired<ScalarEvolution>();
167 AU.addPreserved<ScalarEvolution>();
170 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
172 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
174 bool AddUsersIfInteresting(Instruction *I, Loop *L,
175 SmallPtrSet<Instruction*,16> &Processed);
176 SCEVHandle GetExpressionSCEV(Instruction *E);
177 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
178 IVStrideUse* &CondUse,
179 const SCEVHandle* &CondStride);
180 void OptimizeIndvars(Loop *L);
182 /// OptimizeShadowIV - If IV is used in a int-to-float cast
183 /// inside the loop then try to eliminate the cast opeation.
184 void OptimizeShadowIV(Loop *L);
186 /// OptimizeSMax - Rewrite the loop's terminating condition
187 /// if it uses an smax computation.
188 ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
189 IVStrideUse* &CondUse);
191 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
192 const SCEVHandle *&CondStride);
193 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
194 int64_t CheckForIVReuse(bool, bool, bool, const SCEVHandle&,
195 IVExpr&, const Type*,
196 const std::vector<BasedUser>& UsersToProcess);
197 bool ValidStride(bool, int64_t,
198 const std::vector<BasedUser>& UsersToProcess);
199 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
200 IVUsersOfOneStride &Uses,
202 bool &AllUsesAreAddresses,
203 bool &AllUsesAreOutsideLoop,
204 std::vector<BasedUser> &UsersToProcess);
205 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
206 IVUsersOfOneStride &Uses,
207 Loop *L, bool isOnlyStride);
208 void DeleteTriviallyDeadInstructions();
212 char LoopStrengthReduce::ID = 0;
213 static RegisterPass<LoopStrengthReduce>
214 X("loop-reduce", "Loop Strength Reduction");
216 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
217 return new LoopStrengthReduce(TLI);
220 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
221 /// assumes that the Value* V is of integer or pointer type only.
223 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
225 if (V->getType() == UIntPtrTy) return V;
226 if (Constant *CB = dyn_cast<Constant>(V))
227 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
229 Value *&New = CastedPointers[V];
232 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
233 DeadInsts.push_back(cast<Instruction>(New));
238 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
239 /// specified set are trivially dead, delete them and see if this makes any of
240 /// their operands subsequently dead.
241 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
242 if (DeadInsts.empty()) return;
244 // Sort the deadinsts list so that we can trivially eliminate duplicates as we
245 // go. The code below never adds a non-dead instruction to the worklist, but
246 // callers may not be so careful.
247 array_pod_sort(DeadInsts.begin(), DeadInsts.end());
249 // Drop duplicate instructions and those with uses.
250 for (unsigned i = 0, e = DeadInsts.size()-1; i < e; ++i) {
251 Instruction *I = DeadInsts[i];
252 if (!I->use_empty()) DeadInsts[i] = 0;
253 while (i != e && DeadInsts[i+1] == I)
257 while (!DeadInsts.empty()) {
258 Instruction *I = DeadInsts.back();
259 DeadInsts.pop_back();
261 if (I == 0 || !isInstructionTriviallyDead(I))
264 SE->deleteValueFromRecords(I);
266 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
267 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
270 DeadInsts.push_back(U);
274 I->eraseFromParent();
280 /// GetExpressionSCEV - Compute and return the SCEV for the specified
282 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
283 // Pointer to pointer bitcast instructions return the same value as their
285 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
286 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
287 return SE->getSCEV(BCI);
288 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
293 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
294 // If this is a GEP that SE doesn't know about, compute it now and insert it.
295 // If this is not a GEP, or if we have already done this computation, just let
297 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
298 if (!GEP || SE->hasSCEV(GEP))
299 return SE->getSCEV(Exp);
301 // Analyze all of the subscripts of this getelementptr instruction, looking
302 // for uses that are determined by the trip count of the loop. First, skip
303 // all operands the are not dependent on the IV.
305 // Build up the base expression. Insert an LLVM cast of the pointer to
307 SCEVHandle GEPVal = SE->getUnknown(
308 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
310 gep_type_iterator GTI = gep_type_begin(GEP);
312 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
313 i != e; ++i, ++GTI) {
314 // If this is a use of a recurrence that we can analyze, and it comes before
315 // Op does in the GEP operand list, we will handle this when we process this
317 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
318 const StructLayout *SL = TD->getStructLayout(STy);
319 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
320 uint64_t Offset = SL->getElementOffset(Idx);
321 GEPVal = SE->getAddExpr(GEPVal,
322 SE->getIntegerSCEV(Offset, UIntPtrTy));
324 unsigned GEPOpiBits =
325 (*i)->getType()->getPrimitiveSizeInBits();
326 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
327 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
328 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
329 Instruction::BitCast));
330 Value *OpVal = getCastedVersionOf(opcode, *i);
331 SCEVHandle Idx = SE->getSCEV(OpVal);
333 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
335 Idx = SE->getMulExpr(Idx,
336 SE->getConstant(ConstantInt::get(UIntPtrTy,
338 GEPVal = SE->getAddExpr(GEPVal, Idx);
342 SE->setSCEV(GEP, GEPVal);
346 /// getSCEVStartAndStride - Compute the start and stride of this expression,
347 /// returning false if the expression is not a start/stride pair, or true if it
348 /// is. The stride must be a loop invariant expression, but the start may be
349 /// a mix of loop invariant and loop variant expressions.
350 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
351 SCEVHandle &Start, SCEVHandle &Stride,
352 ScalarEvolution *SE) {
353 SCEVHandle TheAddRec = Start; // Initialize to zero.
355 // If the outer level is an AddExpr, the operands are all start values except
356 // for a nested AddRecExpr.
357 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
358 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
359 if (SCEVAddRecExpr *AddRec =
360 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
361 if (AddRec->getLoop() == L)
362 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
364 return false; // Nested IV of some sort?
366 Start = SE->getAddExpr(Start, AE->getOperand(i));
369 } else if (isa<SCEVAddRecExpr>(SH)) {
372 return false; // not analyzable.
375 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
376 if (!AddRec || AddRec->getLoop() != L) return false;
378 // FIXME: Generalize to non-affine IV's.
379 if (!AddRec->isAffine()) return false;
381 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
383 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
384 DOUT << "[" << L->getHeader()->getName()
385 << "] Variable stride: " << *AddRec << "\n";
387 Stride = AddRec->getOperand(1);
391 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
392 /// and now we need to decide whether the user should use the preinc or post-inc
393 /// value. If this user should use the post-inc version of the IV, return true.
395 /// Choosing wrong here can break dominance properties (if we choose to use the
396 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
397 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
398 /// should use the post-inc value).
399 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
400 Loop *L, DominatorTree *DT, Pass *P,
401 SmallVectorImpl<Instruction*> &DeadInsts){
402 // If the user is in the loop, use the preinc value.
403 if (L->contains(User->getParent())) return false;
405 BasicBlock *LatchBlock = L->getLoopLatch();
407 // Ok, the user is outside of the loop. If it is dominated by the latch
408 // block, use the post-inc value.
409 if (DT->dominates(LatchBlock, User->getParent()))
412 // There is one case we have to be careful of: PHI nodes. These little guys
413 // can live in blocks that do not dominate the latch block, but (since their
414 // uses occur in the predecessor block, not the block the PHI lives in) should
415 // still use the post-inc value. Check for this case now.
416 PHINode *PN = dyn_cast<PHINode>(User);
417 if (!PN) return false; // not a phi, not dominated by latch block.
419 // Look at all of the uses of IV by the PHI node. If any use corresponds to
420 // a block that is not dominated by the latch block, give up and use the
421 // preincremented value.
422 unsigned NumUses = 0;
423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
424 if (PN->getIncomingValue(i) == IV) {
426 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
430 // Okay, all uses of IV by PN are in predecessor blocks that really are
431 // dominated by the latch block. Split the critical edges and use the
432 // post-incremented value.
433 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
434 if (PN->getIncomingValue(i) == IV) {
435 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
436 // Splitting the critical edge can reduce the number of entries in this
438 e = PN->getNumIncomingValues();
439 if (--NumUses == 0) break;
442 // PHI node might have become a constant value after SplitCriticalEdge.
443 DeadInsts.push_back(User);
448 /// isAddress - Returns true if the specified instruction is using the
449 /// specified value as an address.
450 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
451 bool isAddress = isa<LoadInst>(Inst);
452 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
453 if (SI->getOperand(1) == OperandVal)
455 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
456 // Addressing modes can also be folded into prefetches and a variety
458 switch (II->getIntrinsicID()) {
460 case Intrinsic::prefetch:
461 case Intrinsic::x86_sse2_loadu_dq:
462 case Intrinsic::x86_sse2_loadu_pd:
463 case Intrinsic::x86_sse_loadu_ps:
464 case Intrinsic::x86_sse_storeu_ps:
465 case Intrinsic::x86_sse2_storeu_pd:
466 case Intrinsic::x86_sse2_storeu_dq:
467 case Intrinsic::x86_sse2_storel_dq:
468 if (II->getOperand(1) == OperandVal)
476 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
477 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
478 /// return true. Otherwise, return false.
479 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
480 SmallPtrSet<Instruction*,16> &Processed) {
481 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
482 return false; // Void and FP expressions cannot be reduced.
483 if (!Processed.insert(I))
484 return true; // Instruction already handled.
486 // Get the symbolic expression for this instruction.
487 SCEVHandle ISE = GetExpressionSCEV(I);
488 if (isa<SCEVCouldNotCompute>(ISE)) return false;
490 // Get the start and stride for this expression.
491 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
492 SCEVHandle Stride = Start;
493 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
494 return false; // Non-reducible symbolic expression, bail out.
496 std::vector<Instruction *> IUsers;
497 // Collect all I uses now because IVUseShouldUsePostIncValue may
498 // invalidate use_iterator.
499 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
500 IUsers.push_back(cast<Instruction>(*UI));
502 for (unsigned iused_index = 0, iused_size = IUsers.size();
503 iused_index != iused_size; ++iused_index) {
505 Instruction *User = IUsers[iused_index];
507 // Do not infinitely recurse on PHI nodes.
508 if (isa<PHINode>(User) && Processed.count(User))
511 // If this is an instruction defined in a nested loop, or outside this loop,
512 // don't recurse into it.
513 bool AddUserToIVUsers = false;
514 if (LI->getLoopFor(User->getParent()) != L) {
515 DOUT << "FOUND USER in other loop: " << *User
516 << " OF SCEV: " << *ISE << "\n";
517 AddUserToIVUsers = true;
518 } else if (!AddUsersIfInteresting(User, L, Processed)) {
519 DOUT << "FOUND USER: " << *User
520 << " OF SCEV: " << *ISE << "\n";
521 AddUserToIVUsers = true;
524 if (AddUserToIVUsers) {
525 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
526 if (StrideUses.Users.empty()) // First occurrence of this stride?
527 StrideOrder.push_back(Stride);
529 // Okay, we found a user that we cannot reduce. Analyze the instruction
530 // and decide what to do with it. If we are a use inside of the loop, use
531 // the value before incrementation, otherwise use it after incrementation.
532 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
533 // The value used will be incremented by the stride more than we are
534 // expecting, so subtract this off.
535 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
536 StrideUses.addUser(NewStart, User, I);
537 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
538 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
540 StrideUses.addUser(Start, User, I);
548 /// BasedUser - For a particular base value, keep information about how we've
549 /// partitioned the expression so far.
551 /// SE - The current ScalarEvolution object.
554 /// Base - The Base value for the PHI node that needs to be inserted for
555 /// this use. As the use is processed, information gets moved from this
556 /// field to the Imm field (below). BasedUser values are sorted by this
560 /// Inst - The instruction using the induction variable.
563 /// OperandValToReplace - The operand value of Inst to replace with the
565 Value *OperandValToReplace;
567 /// Imm - The immediate value that should be added to the base immediately
568 /// before Inst, because it will be folded into the imm field of the
572 // isUseOfPostIncrementedValue - True if this should use the
573 // post-incremented version of this IV, not the preincremented version.
574 // This can only be set in special cases, such as the terminating setcc
575 // instruction for a loop and uses outside the loop that are dominated by
577 bool isUseOfPostIncrementedValue;
579 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
580 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
581 OperandValToReplace(IVSU.OperandValToReplace),
582 Imm(SE->getIntegerSCEV(0, Base->getType())),
583 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
585 // Once we rewrite the code to insert the new IVs we want, update the
586 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
588 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
589 Instruction *InsertPt,
590 SCEVExpander &Rewriter, Loop *L, Pass *P,
591 SmallVectorImpl<Instruction*> &DeadInsts);
593 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
594 SCEVExpander &Rewriter,
595 Instruction *IP, Loop *L);
600 void BasedUser::dump() const {
601 cerr << " Base=" << *Base;
602 cerr << " Imm=" << *Imm;
603 cerr << " Inst: " << *Inst;
606 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
607 SCEVExpander &Rewriter,
608 Instruction *IP, Loop *L) {
609 // Figure out where we *really* want to insert this code. In particular, if
610 // the user is inside of a loop that is nested inside of L, we really don't
611 // want to insert this expression before the user, we'd rather pull it out as
612 // many loops as possible.
613 LoopInfo &LI = Rewriter.getLoopInfo();
614 Instruction *BaseInsertPt = IP;
616 // Figure out the most-nested loop that IP is in.
617 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
619 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
620 // the preheader of the outer-most loop where NewBase is not loop invariant.
621 if (L->contains(IP->getParent()))
622 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
623 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
624 InsertLoop = InsertLoop->getParentLoop();
627 // If there is no immediate value, skip the next part.
629 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
631 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
633 // If we are inserting the base and imm values in the same block, make sure to
634 // adjust the IP position if insertion reused a result.
635 if (IP == BaseInsertPt)
636 IP = Rewriter.getInsertionPoint();
638 // Always emit the immediate (if non-zero) into the same block as the user.
639 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
640 return Rewriter.expandCodeFor(NewValSCEV, IP);
645 // Once we rewrite the code to insert the new IVs we want, update the
646 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
647 // to it. NewBasePt is the last instruction which contributes to the
648 // value of NewBase in the case that it's a diffferent instruction from
649 // the PHI that NewBase is computed from, or null otherwise.
651 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
652 Instruction *NewBasePt,
653 SCEVExpander &Rewriter, Loop *L, Pass *P,
654 SmallVectorImpl<Instruction*> &DeadInsts){
655 if (!isa<PHINode>(Inst)) {
656 // By default, insert code at the user instruction.
657 BasicBlock::iterator InsertPt = Inst;
659 // However, if the Operand is itself an instruction, the (potentially
660 // complex) inserted code may be shared by many users. Because of this, we
661 // want to emit code for the computation of the operand right before its old
662 // computation. This is usually safe, because we obviously used to use the
663 // computation when it was computed in its current block. However, in some
664 // cases (e.g. use of a post-incremented induction variable) the NewBase
665 // value will be pinned to live somewhere after the original computation.
666 // In this case, we have to back off.
668 // If this is a use outside the loop (which means after, since it is based
669 // on a loop indvar) we use the post-incremented value, so that we don't
670 // artificially make the preinc value live out the bottom of the loop.
671 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
672 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
673 InsertPt = NewBasePt;
675 } else if (Instruction *OpInst
676 = dyn_cast<Instruction>(OperandValToReplace)) {
678 while (isa<PHINode>(InsertPt)) ++InsertPt;
681 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
682 // Adjust the type back to match the Inst. Note that we can't use InsertPt
683 // here because the SCEVExpander may have inserted the instructions after
684 // that point, in its efforts to avoid inserting redundant expressions.
685 if (isa<PointerType>(OperandValToReplace->getType())) {
686 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
688 OperandValToReplace->getType());
690 // Replace the use of the operand Value with the new Phi we just created.
691 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
692 DOUT << " CHANGED: IMM =" << *Imm;
693 DOUT << " \tNEWBASE =" << *NewBase;
694 DOUT << " \tInst = " << *Inst;
698 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
699 // expression into each operand block that uses it. Note that PHI nodes can
700 // have multiple entries for the same predecessor. We use a map to make sure
701 // that a PHI node only has a single Value* for each predecessor (which also
702 // prevents us from inserting duplicate code in some blocks).
703 DenseMap<BasicBlock*, Value*> InsertedCode;
704 PHINode *PN = cast<PHINode>(Inst);
705 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
706 if (PN->getIncomingValue(i) == OperandValToReplace) {
707 // If this is a critical edge, split the edge so that we do not insert the
708 // code on all predecessor/successor paths. We do this unless this is the
709 // canonical backedge for this loop, as this can make some inserted code
710 // be in an illegal position.
711 BasicBlock *PHIPred = PN->getIncomingBlock(i);
712 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
713 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
715 // First step, split the critical edge.
716 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
718 // Next step: move the basic block. In particular, if the PHI node
719 // is outside of the loop, and PredTI is in the loop, we want to
720 // move the block to be immediately before the PHI block, not
721 // immediately after PredTI.
722 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
723 BasicBlock *NewBB = PN->getIncomingBlock(i);
724 NewBB->moveBefore(PN->getParent());
727 // Splitting the edge can reduce the number of PHI entries we have.
728 e = PN->getNumIncomingValues();
731 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
733 // Insert the code into the end of the predecessor block.
734 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
735 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
737 // Adjust the type back to match the PHI. Note that we can't use
738 // InsertPt here because the SCEVExpander may have inserted its
739 // instructions after that point, in its efforts to avoid inserting
740 // redundant expressions.
741 if (isa<PointerType>(PN->getType())) {
742 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
748 // Replace the use of the operand Value with the new Phi we just created.
749 PN->setIncomingValue(i, Code);
754 // PHI node might have become a constant value after SplitCriticalEdge.
755 DeadInsts.push_back(Inst);
757 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
761 /// fitsInAddressMode - Return true if V can be subsumed within an addressing
762 /// mode, and does not need to be put in a register first.
763 static bool fitsInAddressMode(const SCEVHandle &V, const Type *UseTy,
764 const TargetLowering *TLI, bool HasBaseReg) {
765 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
766 int64_t VC = SC->getValue()->getSExtValue();
768 TargetLowering::AddrMode AM;
770 AM.HasBaseReg = HasBaseReg;
771 return TLI->isLegalAddressingMode(AM, UseTy);
773 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
774 return (VC > -(1 << 16) && VC < (1 << 16)-1);
778 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
779 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
780 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
781 Constant *Op0 = CE->getOperand(0);
782 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
783 TargetLowering::AddrMode AM;
785 AM.HasBaseReg = HasBaseReg;
786 return TLI->isLegalAddressingMode(AM, UseTy);
792 /// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are
793 /// loop varying to the Imm operand.
794 static void MoveLoopVariantsToImmediateField(SCEVHandle &Val, SCEVHandle &Imm,
795 Loop *L, ScalarEvolution *SE) {
796 if (Val->isLoopInvariant(L)) return; // Nothing to do.
798 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
799 std::vector<SCEVHandle> NewOps;
800 NewOps.reserve(SAE->getNumOperands());
802 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
803 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
804 // If this is a loop-variant expression, it must stay in the immediate
805 // field of the expression.
806 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
808 NewOps.push_back(SAE->getOperand(i));
812 Val = SE->getIntegerSCEV(0, Val->getType());
814 Val = SE->getAddExpr(NewOps);
815 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
816 // Try to pull immediates out of the start value of nested addrec's.
817 SCEVHandle Start = SARE->getStart();
818 MoveLoopVariantsToImmediateField(Start, Imm, L, SE);
820 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
822 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
824 // Otherwise, all of Val is variant, move the whole thing over.
825 Imm = SE->getAddExpr(Imm, Val);
826 Val = SE->getIntegerSCEV(0, Val->getType());
831 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
832 /// that can fit into the immediate field of instructions in the target.
833 /// Accumulate these immediate values into the Imm value.
834 static void MoveImmediateValues(const TargetLowering *TLI,
836 SCEVHandle &Val, SCEVHandle &Imm,
837 bool isAddress, Loop *L,
838 ScalarEvolution *SE) {
839 const Type *UseTy = User->getType();
840 if (StoreInst *SI = dyn_cast<StoreInst>(User))
841 UseTy = SI->getOperand(0)->getType();
843 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
844 std::vector<SCEVHandle> NewOps;
845 NewOps.reserve(SAE->getNumOperands());
847 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
848 SCEVHandle NewOp = SAE->getOperand(i);
849 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
851 if (!NewOp->isLoopInvariant(L)) {
852 // If this is a loop-variant expression, it must stay in the immediate
853 // field of the expression.
854 Imm = SE->getAddExpr(Imm, NewOp);
856 NewOps.push_back(NewOp);
861 Val = SE->getIntegerSCEV(0, Val->getType());
863 Val = SE->getAddExpr(NewOps);
865 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
866 // Try to pull immediates out of the start value of nested addrec's.
867 SCEVHandle Start = SARE->getStart();
868 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
870 if (Start != SARE->getStart()) {
871 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
873 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
876 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
877 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
878 if (isAddress && fitsInAddressMode(SME->getOperand(0), UseTy, TLI, false) &&
879 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
881 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
882 SCEVHandle NewOp = SME->getOperand(1);
883 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
885 // If we extracted something out of the subexpressions, see if we can
887 if (NewOp != SME->getOperand(1)) {
888 // Scale SubImm up by "8". If the result is a target constant, we are
890 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
891 if (fitsInAddressMode(SubImm, UseTy, TLI, false)) {
892 // Accumulate the immediate.
893 Imm = SE->getAddExpr(Imm, SubImm);
895 // Update what is left of 'Val'.
896 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
903 // Loop-variant expressions must stay in the immediate field of the
905 if ((isAddress && fitsInAddressMode(Val, UseTy, TLI, false)) ||
906 !Val->isLoopInvariant(L)) {
907 Imm = SE->getAddExpr(Imm, Val);
908 Val = SE->getIntegerSCEV(0, Val->getType());
912 // Otherwise, no immediates to move.
916 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
917 /// added together. This is used to reassociate common addition subexprs
918 /// together for maximal sharing when rewriting bases.
919 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
921 ScalarEvolution *SE) {
922 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
923 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
924 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
925 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
926 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
927 if (SARE->getOperand(0) == Zero) {
928 SubExprs.push_back(Expr);
930 // Compute the addrec with zero as its base.
931 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
932 Ops[0] = Zero; // Start with zero base.
933 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
936 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
938 } else if (!Expr->isZero()) {
940 SubExprs.push_back(Expr);
944 // This is logically local to the following function, but C++ says we have
945 // to make it file scope.
946 struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
948 /// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all
949 /// the Uses, removing any common subexpressions, except that if all such
950 /// subexpressions can be folded into an addressing mode for all uses inside
951 /// the loop (this case is referred to as "free" in comments herein) we do
952 /// not remove anything. This looks for things like (a+b+c) and
953 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
954 /// is *removed* from the Bases and returned.
956 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
957 ScalarEvolution *SE, Loop *L,
958 const TargetLowering *TLI) {
959 unsigned NumUses = Uses.size();
961 // Only one use? This is a very common case, so we handle it specially and
963 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
964 SCEVHandle Result = Zero;
965 SCEVHandle FreeResult = Zero;
967 // If the use is inside the loop, use its base, regardless of what it is:
968 // it is clearly shared across all the IV's. If the use is outside the loop
969 // (which means after it) we don't want to factor anything *into* the loop,
970 // so just use 0 as the base.
971 if (L->contains(Uses[0].Inst->getParent()))
972 std::swap(Result, Uses[0].Base);
976 // To find common subexpressions, count how many of Uses use each expression.
977 // If any subexpressions are used Uses.size() times, they are common.
978 // Also track whether all uses of each expression can be moved into an
979 // an addressing mode "for free"; such expressions are left within the loop.
980 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; };
981 std::map<SCEVHandle, SubExprUseData> SubExpressionUseData;
983 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
984 // order we see them.
985 std::vector<SCEVHandle> UniqueSubExprs;
987 std::vector<SCEVHandle> SubExprs;
988 unsigned NumUsesInsideLoop = 0;
989 for (unsigned i = 0; i != NumUses; ++i) {
990 // If the user is outside the loop, just ignore it for base computation.
991 // Since the user is outside the loop, it must be *after* the loop (if it
992 // were before, it could not be based on the loop IV). We don't want users
993 // after the loop to affect base computation of values *inside* the loop,
994 // because we can always add their offsets to the result IV after the loop
995 // is done, ensuring we get good code inside the loop.
996 if (!L->contains(Uses[i].Inst->getParent()))
1000 // If the base is zero (which is common), return zero now, there are no
1001 // CSEs we can find.
1002 if (Uses[i].Base == Zero) return Zero;
1004 // If this use is as an address we may be able to put CSEs in the addressing
1005 // mode rather than hoisting them.
1006 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace);
1007 // We may need the UseTy below, but only when isAddrUse, so compute it
1008 // only in that case.
1009 const Type *UseTy = 0;
1011 UseTy = Uses[i].Inst->getType();
1012 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1013 UseTy = SI->getOperand(0)->getType();
1016 // Split the expression into subexprs.
1017 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1018 // Add one to SubExpressionUseData.Count for each subexpr present, and
1019 // if the subexpr is not a valid immediate within an addressing mode use,
1020 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to
1021 // hoist these out of the loop (if they are common to all uses).
1022 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1023 if (++SubExpressionUseData[SubExprs[j]].Count == 1)
1024 UniqueSubExprs.push_back(SubExprs[j]);
1025 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], UseTy, TLI, false))
1026 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true;
1031 // Now that we know how many times each is used, build Result. Iterate over
1032 // UniqueSubexprs so that we have a stable ordering.
1033 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
1034 std::map<SCEVHandle, SubExprUseData>::iterator I =
1035 SubExpressionUseData.find(UniqueSubExprs[i]);
1036 assert(I != SubExpressionUseData.end() && "Entry not found?");
1037 if (I->second.Count == NumUsesInsideLoop) { // Found CSE!
1038 if (I->second.notAllUsesAreFree)
1039 Result = SE->getAddExpr(Result, I->first);
1041 FreeResult = SE->getAddExpr(FreeResult, I->first);
1043 // Remove non-cse's from SubExpressionUseData.
1044 SubExpressionUseData.erase(I);
1047 if (FreeResult != Zero) {
1048 // We have some subexpressions that can be subsumed into addressing
1049 // modes in every use inside the loop. However, it's possible that
1050 // there are so many of them that the combined FreeResult cannot
1051 // be subsumed, or that the target cannot handle both a FreeResult
1052 // and a Result in the same instruction (for example because it would
1053 // require too many registers). Check this.
1054 for (unsigned i=0; i<NumUses; ++i) {
1055 if (!L->contains(Uses[i].Inst->getParent()))
1057 // We know this is an addressing mode use; if there are any uses that
1058 // are not, FreeResult would be Zero.
1059 const Type *UseTy = Uses[i].Inst->getType();
1060 if (StoreInst *SI = dyn_cast<StoreInst>(Uses[i].Inst))
1061 UseTy = SI->getOperand(0)->getType();
1062 if (!fitsInAddressMode(FreeResult, UseTy, TLI, Result!=Zero)) {
1063 // FIXME: could split up FreeResult into pieces here, some hoisted
1064 // and some not. There is no obvious advantage to this.
1065 Result = SE->getAddExpr(Result, FreeResult);
1072 // If we found no CSE's, return now.
1073 if (Result == Zero) return Result;
1075 // If we still have a FreeResult, remove its subexpressions from
1076 // SubExpressionUseData. This means they will remain in the use Bases.
1077 if (FreeResult != Zero) {
1078 SeparateSubExprs(SubExprs, FreeResult, SE);
1079 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) {
1080 std::map<SCEVHandle, SubExprUseData>::iterator I =
1081 SubExpressionUseData.find(SubExprs[j]);
1082 SubExpressionUseData.erase(I);
1087 // Otherwise, remove all of the CSE's we found from each of the base values.
1088 for (unsigned i = 0; i != NumUses; ++i) {
1089 // Uses outside the loop don't necessarily include the common base, but
1090 // the final IV value coming into those uses does. Instead of trying to
1091 // remove the pieces of the common base, which might not be there,
1092 // subtract off the base to compensate for this.
1093 if (!L->contains(Uses[i].Inst->getParent())) {
1094 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1098 // Split the expression into subexprs.
1099 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1101 // Remove any common subexpressions.
1102 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1103 if (SubExpressionUseData.count(SubExprs[j])) {
1104 SubExprs.erase(SubExprs.begin()+j);
1108 // Finally, add the non-shared expressions together.
1109 if (SubExprs.empty())
1110 Uses[i].Base = Zero;
1112 Uses[i].Base = SE->getAddExpr(SubExprs);
1119 /// ValidStride - Check whether the given Scale is valid for all loads and
1120 /// stores in UsersToProcess.
1122 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1124 const std::vector<BasedUser>& UsersToProcess) {
1128 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1129 // If this is a load or other access, pass the type of the access in.
1130 const Type *AccessTy = Type::VoidTy;
1131 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1132 AccessTy = SI->getOperand(0)->getType();
1133 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1134 AccessTy = LI->getType();
1135 else if (isa<PHINode>(UsersToProcess[i].Inst))
1138 TargetLowering::AddrMode AM;
1139 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1140 AM.BaseOffs = SC->getValue()->getSExtValue();
1141 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1144 // If load[imm+r*scale] is illegal, bail out.
1145 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1151 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1153 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1157 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1159 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1160 !(isa<PointerType>(Ty2) &&
1161 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1162 !(isa<PointerType>(Ty1) &&
1163 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1166 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1167 /// of a previous stride and it is a legal value for the target addressing
1168 /// mode scale component and optional base reg. This allows the users of
1169 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1170 /// reuse is possible. Factors can be negative on same targets, e.g. ARM.
1171 int64_t LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1172 bool AllUsesAreAddresses,
1173 bool AllUsesAreOutsideLoop,
1174 const SCEVHandle &Stride,
1175 IVExpr &IV, const Type *Ty,
1176 const std::vector<BasedUser>& UsersToProcess) {
1177 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1178 int64_t SInt = SC->getValue()->getSExtValue();
1179 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1181 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1182 IVsByStride.find(StrideOrder[NewStride]);
1183 if (SI == IVsByStride.end())
1185 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1186 if (SI->first != Stride &&
1187 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1189 int64_t Scale = SInt / SSInt;
1190 // Check that this stride is valid for all the types used for loads and
1191 // stores; if it can be used for some and not others, we might as well use
1192 // the original stride everywhere, since we have to create the IV for it
1193 // anyway. If the scale is 1, then we don't need to worry about folding
1196 (AllUsesAreAddresses &&
1197 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1198 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1199 IE = SI->second.IVs.end(); II != IE; ++II)
1200 // FIXME: Only handle base == 0 for now.
1201 // Only reuse previous IV if it would not require a type conversion.
1202 if (II->Base->isZero() &&
1203 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1212 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1213 /// returns true if Val's isUseOfPostIncrementedValue is true.
1214 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1215 return Val.isUseOfPostIncrementedValue;
1218 /// isNonConstantNegative - Return true if the specified scev is negated, but
1220 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1221 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1222 if (!Mul) return false;
1224 // If there is a constant factor, it will be first.
1225 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1226 if (!SC) return false;
1228 // Return true if the value is negative, this matches things like (-42 * V).
1229 return SC->getValue()->getValue().isNegative();
1232 // CollectIVUsers - Transform our list of users and offsets to a bit more
1233 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1234 // of the strided accesses, as well as the old information from Uses. We
1235 // progressively move information from the Base field to the Imm field, until
1236 // we eventually have the full access expression to rewrite the use.
1237 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1238 IVUsersOfOneStride &Uses,
1240 bool &AllUsesAreAddresses,
1241 bool &AllUsesAreOutsideLoop,
1242 std::vector<BasedUser> &UsersToProcess) {
1243 UsersToProcess.reserve(Uses.Users.size());
1244 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1245 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1247 // Move any loop variant operands from the offset field to the immediate
1248 // field of the use, so that we don't try to use something before it is
1250 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base,
1251 UsersToProcess.back().Imm, L, SE);
1252 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1253 "Base value is not loop invariant!");
1256 // We now have a whole bunch of uses of like-strided induction variables, but
1257 // they might all have different bases. We want to emit one PHI node for this
1258 // stride which we fold as many common expressions (between the IVs) into as
1259 // possible. Start by identifying the common expressions in the base values
1260 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1261 // "A+B"), emit it to the preheader, then remove the expression from the
1262 // UsersToProcess base values.
1263 SCEVHandle CommonExprs =
1264 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI);
1266 // Next, figure out what we can represent in the immediate fields of
1267 // instructions. If we can represent anything there, move it to the imm
1268 // fields of the BasedUsers. We do this so that it increases the commonality
1269 // of the remaining uses.
1270 unsigned NumPHI = 0;
1271 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1272 // If the user is not in the current loop, this means it is using the exit
1273 // value of the IV. Do not put anything in the base, make sure it's all in
1274 // the immediate field to allow as much factoring as possible.
1275 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1276 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1277 UsersToProcess[i].Base);
1278 UsersToProcess[i].Base =
1279 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1282 // Addressing modes can be folded into loads and stores. Be careful that
1283 // the store is through the expression, not of the expression though.
1285 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1286 UsersToProcess[i].OperandValToReplace);
1287 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1292 // Not all uses are outside the loop.
1293 AllUsesAreOutsideLoop = false;
1295 // If this use isn't an address, then not all uses are addresses.
1296 if (!isAddress && !isPHI)
1297 AllUsesAreAddresses = false;
1299 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1300 UsersToProcess[i].Imm, isAddress, L, SE);
1304 // If one of the use if a PHI node and all other uses are addresses, still
1305 // allow iv reuse. Essentially we are trading one constant multiplication
1306 // for one fewer iv.
1308 AllUsesAreAddresses = false;
1313 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1314 /// stride of IV. All of the users may have different starting values, and this
1315 /// may not be the only stride (we know it is if isOnlyStride is true).
1316 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1317 IVUsersOfOneStride &Uses,
1319 bool isOnlyStride) {
1320 // If all the users are moved to another stride, then there is nothing to do.
1321 if (Uses.Users.empty())
1324 // Keep track if every use in UsersToProcess is an address. If they all are,
1325 // we may be able to rewrite the entire collection of them in terms of a
1326 // smaller-stride IV.
1327 bool AllUsesAreAddresses = true;
1329 // Keep track if every use of a single stride is outside the loop. If so,
1330 // we want to be more aggressive about reusing a smaller-stride IV; a
1331 // multiply outside the loop is better than another IV inside. Well, usually.
1332 bool AllUsesAreOutsideLoop = true;
1334 // Transform our list of users and offsets to a bit more complex table. In
1335 // this new vector, each 'BasedUser' contains 'Base' the base of the
1336 // strided accessas well as the old information from Uses. We progressively
1337 // move information from the Base field to the Imm field, until we eventually
1338 // have the full access expression to rewrite the use.
1339 std::vector<BasedUser> UsersToProcess;
1340 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1341 AllUsesAreOutsideLoop,
1344 // If we managed to find some expressions in common, we'll need to carry
1345 // their value in a register and add it in for each use. This will take up
1346 // a register operand, which potentially restricts what stride values are
1348 bool HaveCommonExprs = !CommonExprs->isZero();
1350 // If all uses are addresses, check if it is possible to reuse an IV with a
1351 // stride that is a factor of this stride. And that the multiple is a number
1352 // that can be encoded in the scale field of the target addressing mode. And
1353 // that we will have a valid instruction after this substition, including the
1354 // immediate field, if any.
1355 PHINode *NewPHI = NULL;
1357 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1358 SE->getIntegerSCEV(0, Type::Int32Ty),
1360 int64_t RewriteFactor = 0;
1361 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1362 AllUsesAreOutsideLoop,
1363 Stride, ReuseIV, CommonExprs->getType(),
1365 if (RewriteFactor != 0) {
1366 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1367 << " and BASE " << *ReuseIV.Base << " :\n";
1368 NewPHI = ReuseIV.PHI;
1369 IncV = ReuseIV.IncV;
1372 const Type *ReplacedTy = CommonExprs->getType();
1374 // Now that we know what we need to do, insert the PHI node itself.
1376 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1377 << *Stride << " and BASE " << *CommonExprs << ": ";
1379 SCEVExpander Rewriter(*SE, *LI);
1380 SCEVExpander PreheaderRewriter(*SE, *LI);
1382 BasicBlock *Preheader = L->getLoopPreheader();
1383 Instruction *PreInsertPt = Preheader->getTerminator();
1384 Instruction *PhiInsertBefore = L->getHeader()->begin();
1386 BasicBlock *LatchBlock = L->getLoopLatch();
1389 // Emit the initial base value into the loop preheader.
1391 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1393 if (RewriteFactor == 0) {
1394 // Create a new Phi for this base, and stick it in the loop header.
1395 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1398 // Add common base to the new Phi node.
1399 NewPHI->addIncoming(CommonBaseV, Preheader);
1401 // If the stride is negative, insert a sub instead of an add for the
1403 bool isNegative = isNonConstantNegative(Stride);
1404 SCEVHandle IncAmount = Stride;
1406 IncAmount = SE->getNegativeSCEV(Stride);
1408 // Insert the stride into the preheader.
1409 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1410 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1412 // Emit the increment of the base value before the terminator of the loop
1413 // latch block, and add it to the Phi node.
1414 SCEVHandle IncExp = SE->getUnknown(StrideV);
1416 IncExp = SE->getNegativeSCEV(IncExp);
1417 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1419 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1420 IncV->setName(NewPHI->getName()+".inc");
1421 NewPHI->addIncoming(IncV, LatchBlock);
1423 // Remember this in case a later stride is multiple of this.
1424 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1426 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1428 Constant *C = dyn_cast<Constant>(CommonBaseV);
1430 (!C->isNullValue() &&
1431 !fitsInAddressMode(SE->getUnknown(CommonBaseV), ReplacedTy,
1433 // We want the common base emitted into the preheader! This is just
1434 // using cast as a copy so BitCast (no-op cast) is appropriate
1435 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1436 "commonbase", PreInsertPt);
1440 // We want to emit code for users inside the loop first. To do this, we
1441 // rearrange BasedUser so that the entries at the end have
1442 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1443 // vector (so we handle them first).
1444 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1445 PartitionByIsUseOfPostIncrementedValue);
1447 // Sort this by base, so that things with the same base are handled
1448 // together. By partitioning first and stable-sorting later, we are
1449 // guaranteed that within each base we will pop off users from within the
1450 // loop before users outside of the loop with a particular base.
1452 // We would like to use stable_sort here, but we can't. The problem is that
1453 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1454 // we don't have anything to do a '<' comparison on. Because we think the
1455 // number of uses is small, do a horrible bubble sort which just relies on
1457 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1458 // Get a base value.
1459 SCEVHandle Base = UsersToProcess[i].Base;
1461 // Compact everything with this base to be consecutive with this one.
1462 for (unsigned j = i+1; j != e; ++j) {
1463 if (UsersToProcess[j].Base == Base) {
1464 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1470 // Process all the users now. This outer loop handles all bases, the inner
1471 // loop handles all users of a particular base.
1472 while (!UsersToProcess.empty()) {
1473 SCEVHandle Base = UsersToProcess.back().Base;
1475 // Emit the code for Base into the preheader.
1476 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1478 DOUT << " INSERTING code for BASE = " << *Base << ":";
1479 if (BaseV->hasName())
1480 DOUT << " Result value name = %" << BaseV->getNameStr();
1483 // If BaseV is a constant other than 0, make sure that it gets inserted into
1484 // the preheader, instead of being forward substituted into the uses. We do
1485 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1487 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1488 if (!C->isNullValue() && !fitsInAddressMode(Base, ReplacedTy,
1490 // We want this constant emitted into the preheader! This is just
1491 // using cast as a copy so BitCast (no-op cast) is appropriate
1492 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1497 // Emit the code to add the immediate offset to the Phi value, just before
1498 // the instructions that we identified as using this stride and base.
1500 // FIXME: Use emitted users to emit other users.
1501 BasedUser &User = UsersToProcess.back();
1503 // If this instruction wants to use the post-incremented value, move it
1504 // after the post-inc and use its value instead of the PHI.
1505 Value *RewriteOp = NewPHI;
1506 if (User.isUseOfPostIncrementedValue) {
1509 // If this user is in the loop, make sure it is the last thing in the
1510 // loop to ensure it is dominated by the increment.
1511 if (L->contains(User.Inst->getParent()))
1512 User.Inst->moveBefore(LatchBlock->getTerminator());
1514 if (RewriteOp->getType() != ReplacedTy) {
1515 Instruction::CastOps opcode = Instruction::Trunc;
1516 if (ReplacedTy->getPrimitiveSizeInBits() ==
1517 RewriteOp->getType()->getPrimitiveSizeInBits())
1518 opcode = Instruction::BitCast;
1519 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1522 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1524 // If we had to insert new instructions for RewriteOp, we have to
1525 // consider that they may not have been able to end up immediately
1526 // next to RewriteOp, because non-PHI instructions may never precede
1527 // PHI instructions in a block. In this case, remember where the last
1528 // instruction was inserted so that if we're replacing a different
1529 // PHI node, we can use the later point to expand the final
1531 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1532 if (RewriteOp == NewPHI) NewBasePt = 0;
1534 // Clear the SCEVExpander's expression map so that we are guaranteed
1535 // to have the code emitted where we expect it.
1538 // If we are reusing the iv, then it must be multiplied by a constant
1539 // factor take advantage of addressing mode scale component.
1540 if (RewriteFactor != 0) {
1541 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1542 RewriteExpr->getType()),
1545 // The common base is emitted in the loop preheader. But since we
1546 // are reusing an IV, it has not been used to initialize the PHI node.
1547 // Add it to the expression used to rewrite the uses.
1548 if (!isa<ConstantInt>(CommonBaseV) ||
1549 !cast<ConstantInt>(CommonBaseV)->isZero())
1550 RewriteExpr = SE->getAddExpr(RewriteExpr,
1551 SE->getUnknown(CommonBaseV));
1554 // Now that we know what we need to do, insert code before User for the
1555 // immediate and any loop-variant expressions.
1556 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1557 // Add BaseV to the PHI value if needed.
1558 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1560 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1564 // Mark old value we replaced as possibly dead, so that it is eliminated
1565 // if we just replaced the last use of that value.
1566 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1568 UsersToProcess.pop_back();
1571 // If there are any more users to process with the same base, process them
1572 // now. We sorted by base above, so we just have to check the last elt.
1573 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1574 // TODO: Next, find out which base index is the most common, pull it out.
1577 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1578 // different starting values, into different PHIs.
1581 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1582 /// set the IV user and stride information and return true, otherwise return
1584 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1585 const SCEVHandle *&CondStride) {
1586 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1588 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1589 IVUsesByStride.find(StrideOrder[Stride]);
1590 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1592 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1593 E = SI->second.Users.end(); UI != E; ++UI)
1594 if (UI->User == Cond) {
1595 // NOTE: we could handle setcc instructions with multiple uses here, but
1596 // InstCombine does it as well for simple uses, it's not clear that it
1597 // occurs enough in real life to handle.
1599 CondStride = &SI->first;
1607 // Constant strides come first which in turns are sorted by their absolute
1608 // values. If absolute values are the same, then positive strides comes first.
1610 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1611 struct StrideCompare {
1612 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1613 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1614 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1616 int64_t LV = LHSC->getValue()->getSExtValue();
1617 int64_t RV = RHSC->getValue()->getSExtValue();
1618 uint64_t ALV = (LV < 0) ? -LV : LV;
1619 uint64_t ARV = (RV < 0) ? -RV : RV;
1625 return (LHSC && !RHSC);
1630 /// ChangeCompareStride - If a loop termination compare instruction is the
1631 /// only use of its stride, and the compaison is against a constant value,
1632 /// try eliminate the stride by moving the compare instruction to another
1633 /// stride and change its constant operand accordingly. e.g.
1639 /// if (v2 < 10) goto loop
1644 /// if (v1 < 30) goto loop
1645 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1646 IVStrideUse* &CondUse,
1647 const SCEVHandle* &CondStride) {
1648 if (StrideOrder.size() < 2 ||
1649 IVUsesByStride[*CondStride].Users.size() != 1)
1651 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1652 if (!SC) return Cond;
1653 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1654 if (!C) return Cond;
1656 ICmpInst::Predicate Predicate = Cond->getPredicate();
1657 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1658 int64_t CmpVal = C->getValue().getSExtValue();
1659 unsigned BitWidth = C->getValue().getBitWidth();
1660 uint64_t SignBit = 1ULL << (BitWidth-1);
1661 const Type *CmpTy = C->getType();
1662 const Type *NewCmpTy = NULL;
1663 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1664 unsigned NewTyBits = 0;
1665 int64_t NewCmpVal = CmpVal;
1666 SCEVHandle *NewStride = NULL;
1667 Value *NewIncV = NULL;
1670 // Check stride constant and the comparision constant signs to detect
1672 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1675 // Look for a suitable stride / iv as replacement.
1676 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1677 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1678 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1679 IVUsesByStride.find(StrideOrder[i]);
1680 if (!isa<SCEVConstant>(SI->first))
1682 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1683 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1686 Scale = SSInt / CmpSSInt;
1687 NewCmpVal = CmpVal * Scale;
1688 APInt Mul = APInt(BitWidth, NewCmpVal);
1689 // Check for overflow.
1690 if (Mul.getSExtValue() != NewCmpVal) {
1695 // Watch out for overflow.
1696 if (ICmpInst::isSignedPredicate(Predicate) &&
1697 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1700 if (NewCmpVal != CmpVal) {
1701 // Pick the best iv to use trying to avoid a cast.
1703 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1704 E = SI->second.Users.end(); UI != E; ++UI) {
1705 NewIncV = UI->OperandValToReplace;
1706 if (NewIncV->getType() == CmpTy)
1714 NewCmpTy = NewIncV->getType();
1715 NewTyBits = isa<PointerType>(NewCmpTy)
1716 ? UIntPtrTy->getPrimitiveSizeInBits()
1717 : NewCmpTy->getPrimitiveSizeInBits();
1718 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1719 // Check if it is possible to rewrite it using
1720 // an iv / stride of a smaller integer type.
1721 bool TruncOk = false;
1722 if (NewCmpTy->isInteger()) {
1723 unsigned Bits = NewTyBits;
1724 if (ICmpInst::isSignedPredicate(Predicate))
1726 uint64_t Mask = (1ULL << Bits) - 1;
1727 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1736 // Don't rewrite if use offset is non-constant and the new type is
1737 // of a different type.
1738 // FIXME: too conservative?
1739 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1744 bool AllUsesAreAddresses = true;
1745 bool AllUsesAreOutsideLoop = true;
1746 std::vector<BasedUser> UsersToProcess;
1747 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1748 AllUsesAreAddresses,
1749 AllUsesAreOutsideLoop,
1751 // Avoid rewriting the compare instruction with an iv of new stride
1752 // if it's likely the new stride uses will be rewritten using the
1753 if (AllUsesAreAddresses &&
1754 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1759 // If scale is negative, use swapped predicate unless it's testing
1761 if (Scale < 0 && !Cond->isEquality())
1762 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1764 NewStride = &StrideOrder[i];
1769 // Forgo this transformation if it the increment happens to be
1770 // unfortunately positioned after the condition, and the condition
1771 // has multiple uses which prevent it from being moved immediately
1772 // before the branch. See
1773 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1774 // for an example of this situation.
1775 if (!Cond->hasOneUse()) {
1776 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1782 if (NewCmpVal != CmpVal) {
1783 // Create a new compare instruction using new stride / iv.
1784 ICmpInst *OldCond = Cond;
1786 if (!isa<PointerType>(NewCmpTy))
1787 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1789 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1790 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1792 // Insert new compare instruction.
1793 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1794 L->getHeader()->getName() + ".termcond",
1797 // Remove the old compare instruction. The old indvar is probably dead too.
1798 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
1799 SE->deleteValueFromRecords(OldCond);
1800 OldCond->replaceAllUsesWith(Cond);
1801 OldCond->eraseFromParent();
1803 IVUsesByStride[*CondStride].Users.pop_back();
1804 SCEVHandle NewOffset = TyBits == NewTyBits
1805 ? SE->getMulExpr(CondUse->Offset,
1806 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1807 : SE->getConstant(ConstantInt::get(NewCmpTy,
1808 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1809 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1810 CondUse = &IVUsesByStride[*NewStride].Users.back();
1811 CondStride = NewStride;
1818 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
1819 /// an smax computation.
1821 /// This is a narrow solution to a specific, but acute, problem. For loops
1827 /// } while (++i < n);
1829 /// where the comparison is signed, the trip count isn't just 'n', because
1830 /// 'n' could be negative. And unfortunately this can come up even for loops
1831 /// where the user didn't use a C do-while loop. For example, seemingly
1832 /// well-behaved top-test loops will commonly be lowered like this:
1838 /// } while (++i < n);
1841 /// and then it's possible for subsequent optimization to obscure the if
1842 /// test in such a way that indvars can't find it.
1844 /// When indvars can't find the if test in loops like this, it creates a
1845 /// signed-max expression, which allows it to give the loop a canonical
1846 /// induction variable:
1849 /// smax = n < 1 ? 1 : n;
1852 /// } while (++i != smax);
1854 /// Canonical induction variables are necessary because the loop passes
1855 /// are designed around them. The most obvious example of this is the
1856 /// LoopInfo analysis, which doesn't remember trip count values. It
1857 /// expects to be able to rediscover the trip count each time it is
1858 /// needed, and it does this using a simple analyis that only succeeds if
1859 /// the loop has a canonical induction variable.
1861 /// However, when it comes time to generate code, the maximum operation
1862 /// can be quite costly, especially if it's inside of an outer loop.
1864 /// This function solves this problem by detecting this type of loop and
1865 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1866 /// the instructions for the maximum computation.
1868 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
1869 IVStrideUse* &CondUse) {
1870 // Check that the loop matches the pattern we're looking for.
1871 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1872 Cond->getPredicate() != CmpInst::ICMP_NE)
1875 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1876 if (!Sel || !Sel->hasOneUse()) return Cond;
1878 SCEVHandle IterationCount = SE->getIterationCount(L);
1879 if (isa<SCEVCouldNotCompute>(IterationCount))
1881 SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());
1883 // Adjust for an annoying getIterationCount quirk.
1884 IterationCount = SE->getAddExpr(IterationCount, One);
1886 // Check for a max calculation that matches the pattern.
1887 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
1888 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
1890 SCEVHandle SMaxLHS = SMax->getOperand(0);
1891 SCEVHandle SMaxRHS = SMax->getOperand(1);
1892 if (!SMaxLHS || SMaxLHS != One) return Cond;
1894 // Check the relevant induction variable for conformance to
1896 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
1897 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1898 if (!AR || !AR->isAffine() ||
1899 AR->getStart() != One ||
1900 AR->getStepRecurrence(*SE) != One)
1903 // Check the right operand of the select, and remember it, as it will
1904 // be used in the new comparison instruction.
1906 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
1907 NewRHS = Sel->getOperand(1);
1908 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
1909 NewRHS = Sel->getOperand(2);
1910 if (!NewRHS) return Cond;
1912 // Ok, everything looks ok to change the condition into an SLT or SGE and
1913 // delete the max calculation.
1915 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
1918 Cond->getOperand(0), NewRHS, "scmp", Cond);
1920 // Delete the max calculation instructions.
1921 SE->deleteValueFromRecords(Cond);
1922 Cond->replaceAllUsesWith(NewCond);
1923 Cond->eraseFromParent();
1924 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1925 SE->deleteValueFromRecords(Sel);
1926 Sel->eraseFromParent();
1927 if (Cmp->use_empty()) {
1928 SE->deleteValueFromRecords(Cmp);
1929 Cmp->eraseFromParent();
1931 CondUse->User = NewCond;
1935 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1936 /// inside the loop then try to eliminate the cast opeation.
1937 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
1939 SCEVHandle IterationCount = SE->getIterationCount(L);
1940 if (isa<SCEVCouldNotCompute>(IterationCount))
1943 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
1945 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1946 IVUsesByStride.find(StrideOrder[Stride]);
1947 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1948 if (!isa<SCEVConstant>(SI->first))
1951 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1952 E = SI->second.Users.end(); UI != E; /* empty */) {
1953 std::vector<IVStrideUse>::iterator CandidateUI = UI;
1955 Instruction *ShadowUse = CandidateUI->User;
1956 const Type *DestTy = NULL;
1958 /* If shadow use is a int->float cast then insert a second IV
1959 to eliminate this cast.
1961 for (unsigned i = 0; i < n; ++i)
1967 for (unsigned i = 0; i < n; ++i, ++d)
1970 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
1971 DestTy = UCast->getDestTy();
1972 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
1973 DestTy = SCast->getDestTy();
1974 if (!DestTy) continue;
1977 /* If target does not support DestTy natively then do not apply
1978 this transformation. */
1979 MVT DVT = TLI->getValueType(DestTy);
1980 if (!TLI->isTypeLegal(DVT)) continue;
1983 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1985 if (PH->getNumIncomingValues() != 2) continue;
1987 const Type *SrcTy = PH->getType();
1988 int Mantissa = DestTy->getFPMantissaWidth();
1989 if (Mantissa == -1) continue;
1990 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
1993 unsigned Entry, Latch;
1994 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
2002 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
2003 if (!Init) continue;
2004 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
2006 BinaryOperator *Incr =
2007 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
2008 if (!Incr) continue;
2009 if (Incr->getOpcode() != Instruction::Add
2010 && Incr->getOpcode() != Instruction::Sub)
2013 /* Initialize new IV, double d = 0.0 in above example. */
2014 ConstantInt *C = NULL;
2015 if (Incr->getOperand(0) == PH)
2016 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
2017 else if (Incr->getOperand(1) == PH)
2018 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
2024 /* Add new PHINode. */
2025 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
2027 /* create new increment. '++d' in above example. */
2028 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
2029 BinaryOperator *NewIncr =
2030 BinaryOperator::Create(Incr->getOpcode(),
2031 NewPH, CFP, "IV.S.next.", Incr);
2033 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
2034 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
2036 /* Remove cast operation */
2037 SE->deleteValueFromRecords(ShadowUse);
2038 ShadowUse->replaceAllUsesWith(NewPH);
2039 ShadowUse->eraseFromParent();
2040 SI->second.Users.erase(CandidateUI);
2047 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
2048 // uses in the loop, look to see if we can eliminate some, in favor of using
2049 // common indvars for the different uses.
2050 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
2051 // TODO: implement optzns here.
2053 OptimizeShadowIV(L);
2055 // Finally, get the terminating condition for the loop if possible. If we
2056 // can, we want to change it to use a post-incremented version of its
2057 // induction variable, to allow coalescing the live ranges for the IV into
2058 // one register value.
2059 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
2060 BasicBlock *Preheader = L->getLoopPreheader();
2061 BasicBlock *LatchBlock =
2062 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
2063 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
2064 if (!TermBr || TermBr->isUnconditional() ||
2065 !isa<ICmpInst>(TermBr->getCondition()))
2067 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
2069 // Search IVUsesByStride to find Cond's IVUse if there is one.
2070 IVStrideUse *CondUse = 0;
2071 const SCEVHandle *CondStride = 0;
2073 if (!FindIVUserForCond(Cond, CondUse, CondStride))
2074 return; // setcc doesn't use the IV.
2076 // If the trip count is computed in terms of an smax (due to ScalarEvolution
2077 // being unable to find a sufficient guard, for example), change the loop
2078 // comparison to use SLT instead of NE.
2079 Cond = OptimizeSMax(L, Cond, CondUse);
2081 // If possible, change stride and operands of the compare instruction to
2082 // eliminate one stride.
2083 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2085 // It's possible for the setcc instruction to be anywhere in the loop, and
2086 // possible for it to have multiple users. If it is not immediately before
2087 // the latch block branch, move it.
2088 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2089 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2090 Cond->moveBefore(TermBr);
2092 // Otherwise, clone the terminating condition and insert into the loopend.
2093 Cond = cast<ICmpInst>(Cond->clone());
2094 Cond->setName(L->getHeader()->getName() + ".termcond");
2095 LatchBlock->getInstList().insert(TermBr, Cond);
2097 // Clone the IVUse, as the old use still exists!
2098 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2099 CondUse->OperandValToReplace);
2100 CondUse = &IVUsesByStride[*CondStride].Users.back();
2104 // If we get to here, we know that we can transform the setcc instruction to
2105 // use the post-incremented version of the IV, allowing us to coalesce the
2106 // live ranges for the IV correctly.
2107 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2108 CondUse->isUseOfPostIncrementedValue = true;
2112 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2114 LI = &getAnalysis<LoopInfo>();
2115 DT = &getAnalysis<DominatorTree>();
2116 SE = &getAnalysis<ScalarEvolution>();
2117 TD = &getAnalysis<TargetData>();
2118 UIntPtrTy = TD->getIntPtrType();
2121 // Find all uses of induction variables in this loop, and categorize
2122 // them by stride. Start by finding all of the PHI nodes in the header for
2123 // this loop. If they are induction variables, inspect their uses.
2124 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2125 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2126 AddUsersIfInteresting(I, L, Processed);
2128 if (!IVUsesByStride.empty()) {
2129 // Optimize induction variables. Some indvar uses can be transformed to use
2130 // strides that will be needed for other purposes. A common example of this
2131 // is the exit test for the loop, which can often be rewritten to use the
2132 // computation of some other indvar to decide when to terminate the loop.
2135 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2136 // doing computation in byte values, promote to 32-bit values if safe.
2138 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2139 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2140 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2141 // Need to be careful that IV's are all the same type. Only works for
2142 // intptr_t indvars.
2144 // If we only have one stride, we can more aggressively eliminate some
2146 bool HasOneStride = IVUsesByStride.size() == 1;
2149 DOUT << "\nLSR on ";
2153 // IVsByStride keeps IVs for one particular loop.
2154 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2156 // Sort the StrideOrder so we process larger strides first.
2157 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2159 // Note: this processes each stride/type pair individually. All users
2160 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2161 // Also, note that we iterate over IVUsesByStride indirectly by using
2162 // StrideOrder. This extra layer of indirection makes the ordering of
2163 // strides deterministic - not dependent on map order.
2164 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2165 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2166 IVUsesByStride.find(StrideOrder[Stride]);
2167 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2168 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
2172 // We're done analyzing this loop; release all the state we built up for it.
2173 CastedPointers.clear();
2174 IVUsesByStride.clear();
2175 IVsByStride.clear();
2176 StrideOrder.clear();
2178 // Clean up after ourselves
2179 if (!DeadInsts.empty()) {
2180 DeleteTriviallyDeadInstructions();
2182 BasicBlock::iterator I = L->getHeader()->begin();
2183 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2184 // At this point, we know that we have killed one or more IV users.
2185 // It is worth checking to see if the cannonical indvar is also
2186 // dead, so that we can remove it as well.
2188 // We can remove a PHI if it is on a cycle in the def-use graph
2189 // where each node in the cycle has degree one, i.e. only one use,
2190 // and is an instruction with no side effects.
2192 // FIXME: this needs to eliminate an induction variable even if it's being
2193 // compared against some value to decide loop termination.
2194 if (!PN->hasOneUse())
2197 SmallPtrSet<PHINode *, 4> PHIs;
2198 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2199 J && J->hasOneUse() && !J->mayWriteToMemory();
2200 J = dyn_cast<Instruction>(*J->use_begin())) {
2201 // If we find the original PHI, we've discovered a cycle.
2203 // Break the cycle and mark the PHI for deletion.
2204 SE->deleteValueFromRecords(PN);
2205 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2206 DeadInsts.push_back(PN);
2210 // If we find a PHI more than once, we're on a cycle that
2211 // won't prove fruitful.
2212 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2216 DeleteTriviallyDeadInstructions();