1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs a strength reduction on array references inside loops that
11 // have as one or more of their components the loop induction variable. This is
12 // accomplished by creating a new Value to hold the initial value of the array
13 // access for the first iteration, and then creating a new GEP instruction in
14 // the loop to increment the value by the appropriate amount.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "loop-reduce"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/IntrinsicInst.h"
23 #include "llvm/Type.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/ScalarEvolutionExpander.h"
29 #include "llvm/Support/CFG.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/ADT/SetVector.h"
35 #include "llvm/ADT/SmallPtrSet.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/Compiler.h"
39 #include "llvm/Target/TargetLowering.h"
44 STATISTIC(NumReduced , "Number of GEPs strength reduced");
45 STATISTIC(NumInserted, "Number of PHIs inserted");
46 STATISTIC(NumVariable, "Number of PHIs with variable strides");
47 STATISTIC(NumEliminated, "Number of strides eliminated");
48 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
54 /// IVStrideUse - Keep track of one use of a strided induction variable, where
55 /// the stride is stored externally. The Offset member keeps track of the
56 /// offset from the IV, User is the actual user of the operand, and
57 /// 'OperandValToReplace' is the operand of the User that is the use.
58 struct VISIBILITY_HIDDEN IVStrideUse {
61 Value *OperandValToReplace;
63 // isUseOfPostIncrementedValue - True if this should use the
64 // post-incremented version of this IV, not the preincremented version.
65 // This can only be set in special cases, such as the terminating setcc
66 // instruction for a loop or uses dominated by the loop.
67 bool isUseOfPostIncrementedValue;
69 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
70 : Offset(Offs), User(U), OperandValToReplace(O),
71 isUseOfPostIncrementedValue(false) {}
74 /// IVUsersOfOneStride - This structure keeps track of all instructions that
75 /// have an operand that is based on the trip count multiplied by some stride.
76 /// The stride for all of these users is common and kept external to this
78 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
79 /// Users - Keep track of all of the users of this stride as well as the
80 /// initial value and the operand that uses the IV.
81 std::vector<IVStrideUse> Users;
83 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
84 Users.push_back(IVStrideUse(Offset, User, Operand));
88 /// IVInfo - This structure keeps track of one IV expression inserted during
89 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
90 /// well as the PHI node and increment value created for rewrite.
91 struct VISIBILITY_HIDDEN IVExpr {
97 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
99 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
102 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
103 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
104 struct VISIBILITY_HIDDEN IVsOfOneStride {
105 std::vector<IVExpr> IVs;
107 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
109 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
113 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
117 const TargetData *TD;
118 const Type *UIntPtrTy;
121 /// IVUsesByStride - Keep track of all uses of induction variables that we
122 /// are interested in. The key of the map is the stride of the access.
123 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
125 /// IVsByStride - Keep track of all IVs that have been inserted for a
126 /// particular stride.
127 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
129 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
130 /// We use this to iterate over the IVUsesByStride collection without being
131 /// dependent on random ordering of pointers in the process.
132 SmallVector<SCEVHandle, 16> StrideOrder;
134 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
135 /// of the casted version of each value. This is accessed by
136 /// getCastedVersionOf.
137 DenseMap<Value*, Value*> CastedPointers;
139 /// DeadInsts - Keep track of instructions we may have made dead, so that
140 /// we can remove them after we are done working.
141 SetVector<Instruction*> DeadInsts;
143 /// TLI - Keep a pointer of a TargetLowering to consult for determining
144 /// transformation profitability.
145 const TargetLowering *TLI;
148 static char ID; // Pass ID, replacement for typeid
149 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
150 LoopPass(&ID), TLI(tli) {
153 bool runOnLoop(Loop *L, LPPassManager &LPM);
155 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
156 // We split critical edges, so we change the CFG. However, we do update
157 // many analyses if they are around.
158 AU.addPreservedID(LoopSimplifyID);
159 AU.addPreserved<LoopInfo>();
160 AU.addPreserved<DominanceFrontier>();
161 AU.addPreserved<DominatorTree>();
163 AU.addRequiredID(LoopSimplifyID);
164 AU.addRequired<LoopInfo>();
165 AU.addRequired<DominatorTree>();
166 AU.addRequired<TargetData>();
167 AU.addRequired<ScalarEvolution>();
168 AU.addPreserved<ScalarEvolution>();
171 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
173 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
175 bool AddUsersIfInteresting(Instruction *I, Loop *L,
176 SmallPtrSet<Instruction*,16> &Processed);
177 SCEVHandle GetExpressionSCEV(Instruction *E);
178 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
179 IVStrideUse* &CondUse,
180 const SCEVHandle* &CondStride);
181 void OptimizeIndvars(Loop *L);
183 /// OptimizeShadowIV - If IV is used in a int-to-float cast
184 /// inside the loop then try to eliminate the cast opeation.
185 void OptimizeShadowIV(Loop *L);
187 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
188 const SCEVHandle *&CondStride);
189 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
190 unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
191 IVExpr&, const Type*,
192 const std::vector<BasedUser>& UsersToProcess);
193 bool ValidStride(bool, int64_t,
194 const std::vector<BasedUser>& UsersToProcess);
195 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
196 IVUsersOfOneStride &Uses,
198 bool &AllUsesAreAddresses,
199 std::vector<BasedUser> &UsersToProcess);
200 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
201 IVUsersOfOneStride &Uses,
202 Loop *L, bool isOnlyStride);
203 void DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts);
207 char LoopStrengthReduce::ID = 0;
208 static RegisterPass<LoopStrengthReduce>
209 X("loop-reduce", "Loop Strength Reduction");
211 LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
212 return new LoopStrengthReduce(TLI);
215 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
216 /// assumes that the Value* V is of integer or pointer type only.
218 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
220 if (V->getType() == UIntPtrTy) return V;
221 if (Constant *CB = dyn_cast<Constant>(V))
222 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
224 Value *&New = CastedPointers[V];
227 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
228 DeadInsts.insert(cast<Instruction>(New));
233 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
234 /// specified set are trivially dead, delete them and see if this makes any of
235 /// their operands subsequently dead.
236 void LoopStrengthReduce::
237 DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts) {
238 while (!Insts.empty()) {
239 Instruction *I = Insts.back();
242 if (PHINode *PN = dyn_cast<PHINode>(I)) {
243 // If all incoming values to the Phi are the same, we can replace the Phi
245 if (Value *PNV = PN->hasConstantValue()) {
246 if (Instruction *U = dyn_cast<Instruction>(PNV))
248 SE->deleteValueFromRecords(PN);
249 PN->replaceAllUsesWith(PNV);
250 PN->eraseFromParent();
256 if (isInstructionTriviallyDead(I)) {
257 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
258 if (Instruction *U = dyn_cast<Instruction>(*i))
260 SE->deleteValueFromRecords(I);
261 I->eraseFromParent();
268 /// GetExpressionSCEV - Compute and return the SCEV for the specified
270 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
271 // Pointer to pointer bitcast instructions return the same value as their
273 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
274 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
275 return SE->getSCEV(BCI);
276 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
281 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
282 // If this is a GEP that SE doesn't know about, compute it now and insert it.
283 // If this is not a GEP, or if we have already done this computation, just let
285 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
286 if (!GEP || SE->hasSCEV(GEP))
287 return SE->getSCEV(Exp);
289 // Analyze all of the subscripts of this getelementptr instruction, looking
290 // for uses that are determined by the trip count of the loop. First, skip
291 // all operands the are not dependent on the IV.
293 // Build up the base expression. Insert an LLVM cast of the pointer to
295 SCEVHandle GEPVal = SE->getUnknown(
296 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
298 gep_type_iterator GTI = gep_type_begin(GEP);
300 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
301 i != e; ++i, ++GTI) {
302 // If this is a use of a recurrence that we can analyze, and it comes before
303 // Op does in the GEP operand list, we will handle this when we process this
305 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
306 const StructLayout *SL = TD->getStructLayout(STy);
307 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
308 uint64_t Offset = SL->getElementOffset(Idx);
309 GEPVal = SE->getAddExpr(GEPVal,
310 SE->getIntegerSCEV(Offset, UIntPtrTy));
312 unsigned GEPOpiBits =
313 (*i)->getType()->getPrimitiveSizeInBits();
314 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
315 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
316 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
317 Instruction::BitCast));
318 Value *OpVal = getCastedVersionOf(opcode, *i);
319 SCEVHandle Idx = SE->getSCEV(OpVal);
321 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
323 Idx = SE->getMulExpr(Idx,
324 SE->getConstant(ConstantInt::get(UIntPtrTy,
326 GEPVal = SE->getAddExpr(GEPVal, Idx);
330 SE->setSCEV(GEP, GEPVal);
334 /// getSCEVStartAndStride - Compute the start and stride of this expression,
335 /// returning false if the expression is not a start/stride pair, or true if it
336 /// is. The stride must be a loop invariant expression, but the start may be
337 /// a mix of loop invariant and loop variant expressions.
338 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
339 SCEVHandle &Start, SCEVHandle &Stride,
340 ScalarEvolution *SE) {
341 SCEVHandle TheAddRec = Start; // Initialize to zero.
343 // If the outer level is an AddExpr, the operands are all start values except
344 // for a nested AddRecExpr.
345 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
346 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
347 if (SCEVAddRecExpr *AddRec =
348 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
349 if (AddRec->getLoop() == L)
350 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
352 return false; // Nested IV of some sort?
354 Start = SE->getAddExpr(Start, AE->getOperand(i));
357 } else if (isa<SCEVAddRecExpr>(SH)) {
360 return false; // not analyzable.
363 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
364 if (!AddRec || AddRec->getLoop() != L) return false;
366 // FIXME: Generalize to non-affine IV's.
367 if (!AddRec->isAffine()) return false;
369 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
371 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
372 DOUT << "[" << L->getHeader()->getName()
373 << "] Variable stride: " << *AddRec << "\n";
375 Stride = AddRec->getOperand(1);
379 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
380 /// and now we need to decide whether the user should use the preinc or post-inc
381 /// value. If this user should use the post-inc version of the IV, return true.
383 /// Choosing wrong here can break dominance properties (if we choose to use the
384 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
385 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
386 /// should use the post-inc value).
387 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
388 Loop *L, DominatorTree *DT, Pass *P,
389 SetVector<Instruction*> &DeadInsts){
390 // If the user is in the loop, use the preinc value.
391 if (L->contains(User->getParent())) return false;
393 BasicBlock *LatchBlock = L->getLoopLatch();
395 // Ok, the user is outside of the loop. If it is dominated by the latch
396 // block, use the post-inc value.
397 if (DT->dominates(LatchBlock, User->getParent()))
400 // There is one case we have to be careful of: PHI nodes. These little guys
401 // can live in blocks that do not dominate the latch block, but (since their
402 // uses occur in the predecessor block, not the block the PHI lives in) should
403 // still use the post-inc value. Check for this case now.
404 PHINode *PN = dyn_cast<PHINode>(User);
405 if (!PN) return false; // not a phi, not dominated by latch block.
407 // Look at all of the uses of IV by the PHI node. If any use corresponds to
408 // a block that is not dominated by the latch block, give up and use the
409 // preincremented value.
410 unsigned NumUses = 0;
411 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
412 if (PN->getIncomingValue(i) == IV) {
414 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
418 // Okay, all uses of IV by PN are in predecessor blocks that really are
419 // dominated by the latch block. Split the critical edges and use the
420 // post-incremented value.
421 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
422 if (PN->getIncomingValue(i) == IV) {
423 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
424 // Splitting the critical edge can reduce the number of entries in this
426 e = PN->getNumIncomingValues();
427 if (--NumUses == 0) break;
430 // PHI node might have become a constant value after SplitCriticalEdge.
431 DeadInsts.insert(User);
438 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
439 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
440 /// return true. Otherwise, return false.
441 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
442 SmallPtrSet<Instruction*,16> &Processed) {
443 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
444 return false; // Void and FP expressions cannot be reduced.
445 if (!Processed.insert(I))
446 return true; // Instruction already handled.
448 // Get the symbolic expression for this instruction.
449 SCEVHandle ISE = GetExpressionSCEV(I);
450 if (isa<SCEVCouldNotCompute>(ISE)) return false;
452 // Get the start and stride for this expression.
453 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
454 SCEVHandle Stride = Start;
455 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
456 return false; // Non-reducible symbolic expression, bail out.
458 std::vector<Instruction *> IUsers;
459 // Collect all I uses now because IVUseShouldUsePostIncValue may
460 // invalidate use_iterator.
461 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
462 IUsers.push_back(cast<Instruction>(*UI));
464 for (unsigned iused_index = 0, iused_size = IUsers.size();
465 iused_index != iused_size; ++iused_index) {
467 Instruction *User = IUsers[iused_index];
469 // Do not infinitely recurse on PHI nodes.
470 if (isa<PHINode>(User) && Processed.count(User))
473 // If this is an instruction defined in a nested loop, or outside this loop,
474 // don't recurse into it.
475 bool AddUserToIVUsers = false;
476 if (LI->getLoopFor(User->getParent()) != L) {
477 DOUT << "FOUND USER in other loop: " << *User
478 << " OF SCEV: " << *ISE << "\n";
479 AddUserToIVUsers = true;
480 } else if (!AddUsersIfInteresting(User, L, Processed)) {
481 DOUT << "FOUND USER: " << *User
482 << " OF SCEV: " << *ISE << "\n";
483 AddUserToIVUsers = true;
486 if (AddUserToIVUsers) {
487 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
488 if (StrideUses.Users.empty()) // First occurance of this stride?
489 StrideOrder.push_back(Stride);
491 // Okay, we found a user that we cannot reduce. Analyze the instruction
492 // and decide what to do with it. If we are a use inside of the loop, use
493 // the value before incrementation, otherwise use it after incrementation.
494 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
495 // The value used will be incremented by the stride more than we are
496 // expecting, so subtract this off.
497 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
498 StrideUses.addUser(NewStart, User, I);
499 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
500 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
502 StrideUses.addUser(Start, User, I);
510 /// BasedUser - For a particular base value, keep information about how we've
511 /// partitioned the expression so far.
513 /// SE - The current ScalarEvolution object.
516 /// Base - The Base value for the PHI node that needs to be inserted for
517 /// this use. As the use is processed, information gets moved from this
518 /// field to the Imm field (below). BasedUser values are sorted by this
522 /// Inst - The instruction using the induction variable.
525 /// OperandValToReplace - The operand value of Inst to replace with the
527 Value *OperandValToReplace;
529 /// Imm - The immediate value that should be added to the base immediately
530 /// before Inst, because it will be folded into the imm field of the
534 /// EmittedBase - The actual value* to use for the base value of this
535 /// operation. This is null if we should just use zero so far.
538 // isUseOfPostIncrementedValue - True if this should use the
539 // post-incremented version of this IV, not the preincremented version.
540 // This can only be set in special cases, such as the terminating setcc
541 // instruction for a loop and uses outside the loop that are dominated by
543 bool isUseOfPostIncrementedValue;
545 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
546 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
547 OperandValToReplace(IVSU.OperandValToReplace),
548 Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
549 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
551 // Once we rewrite the code to insert the new IVs we want, update the
552 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
554 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
555 Instruction *InsertPt,
556 SCEVExpander &Rewriter, Loop *L, Pass *P,
557 SetVector<Instruction*> &DeadInsts);
559 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
560 SCEVExpander &Rewriter,
561 Instruction *IP, Loop *L);
566 void BasedUser::dump() const {
567 cerr << " Base=" << *Base;
568 cerr << " Imm=" << *Imm;
570 cerr << " EB=" << *EmittedBase;
572 cerr << " Inst: " << *Inst;
575 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
576 SCEVExpander &Rewriter,
577 Instruction *IP, Loop *L) {
578 // Figure out where we *really* want to insert this code. In particular, if
579 // the user is inside of a loop that is nested inside of L, we really don't
580 // want to insert this expression before the user, we'd rather pull it out as
581 // many loops as possible.
582 LoopInfo &LI = Rewriter.getLoopInfo();
583 Instruction *BaseInsertPt = IP;
585 // Figure out the most-nested loop that IP is in.
586 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
588 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
589 // the preheader of the outer-most loop where NewBase is not loop invariant.
590 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
591 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
592 InsertLoop = InsertLoop->getParentLoop();
595 // If there is no immediate value, skip the next part.
597 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
599 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
601 // If we are inserting the base and imm values in the same block, make sure to
602 // adjust the IP position if insertion reused a result.
603 if (IP == BaseInsertPt)
604 IP = Rewriter.getInsertionPoint();
606 // Always emit the immediate (if non-zero) into the same block as the user.
607 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
608 return Rewriter.expandCodeFor(NewValSCEV, IP);
613 // Once we rewrite the code to insert the new IVs we want, update the
614 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
615 // to it. NewBasePt is the last instruction which contributes to the
616 // value of NewBase in the case that it's a diffferent instruction from
617 // the PHI that NewBase is computed from, or null otherwise.
619 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
620 Instruction *NewBasePt,
621 SCEVExpander &Rewriter, Loop *L, Pass *P,
622 SetVector<Instruction*> &DeadInsts) {
623 if (!isa<PHINode>(Inst)) {
624 // By default, insert code at the user instruction.
625 BasicBlock::iterator InsertPt = Inst;
627 // However, if the Operand is itself an instruction, the (potentially
628 // complex) inserted code may be shared by many users. Because of this, we
629 // want to emit code for the computation of the operand right before its old
630 // computation. This is usually safe, because we obviously used to use the
631 // computation when it was computed in its current block. However, in some
632 // cases (e.g. use of a post-incremented induction variable) the NewBase
633 // value will be pinned to live somewhere after the original computation.
634 // In this case, we have to back off.
635 if (!isUseOfPostIncrementedValue) {
636 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
637 InsertPt = NewBasePt;
639 } else if (Instruction *OpInst
640 = dyn_cast<Instruction>(OperandValToReplace)) {
642 while (isa<PHINode>(InsertPt)) ++InsertPt;
645 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
646 // Adjust the type back to match the Inst. Note that we can't use InsertPt
647 // here because the SCEVExpander may have inserted the instructions after
648 // that point, in its efforts to avoid inserting redundant expressions.
649 if (isa<PointerType>(OperandValToReplace->getType())) {
650 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
652 OperandValToReplace->getType());
654 // Replace the use of the operand Value with the new Phi we just created.
655 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
656 DOUT << " CHANGED: IMM =" << *Imm;
657 DOUT << " \tNEWBASE =" << *NewBase;
658 DOUT << " \tInst = " << *Inst;
662 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
663 // expression into each operand block that uses it. Note that PHI nodes can
664 // have multiple entries for the same predecessor. We use a map to make sure
665 // that a PHI node only has a single Value* for each predecessor (which also
666 // prevents us from inserting duplicate code in some blocks).
667 DenseMap<BasicBlock*, Value*> InsertedCode;
668 PHINode *PN = cast<PHINode>(Inst);
669 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
670 if (PN->getIncomingValue(i) == OperandValToReplace) {
671 // If this is a critical edge, split the edge so that we do not insert the
672 // code on all predecessor/successor paths. We do this unless this is the
673 // canonical backedge for this loop, as this can make some inserted code
674 // be in an illegal position.
675 BasicBlock *PHIPred = PN->getIncomingBlock(i);
676 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
677 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
679 // First step, split the critical edge.
680 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
682 // Next step: move the basic block. In particular, if the PHI node
683 // is outside of the loop, and PredTI is in the loop, we want to
684 // move the block to be immediately before the PHI block, not
685 // immediately after PredTI.
686 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
687 BasicBlock *NewBB = PN->getIncomingBlock(i);
688 NewBB->moveBefore(PN->getParent());
691 // Splitting the edge can reduce the number of PHI entries we have.
692 e = PN->getNumIncomingValues();
695 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
697 // Insert the code into the end of the predecessor block.
698 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
699 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
701 // Adjust the type back to match the PHI. Note that we can't use
702 // InsertPt here because the SCEVExpander may have inserted its
703 // instructions after that point, in its efforts to avoid inserting
704 // redundant expressions.
705 if (isa<PointerType>(PN->getType())) {
706 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
712 // Replace the use of the operand Value with the new Phi we just created.
713 PN->setIncomingValue(i, Code);
718 // PHI node might have become a constant value after SplitCriticalEdge.
719 DeadInsts.insert(Inst);
721 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
725 /// isTargetConstant - Return true if the following can be referenced by the
726 /// immediate field of a target instruction.
727 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
728 const TargetLowering *TLI) {
729 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
730 int64_t VC = SC->getValue()->getSExtValue();
732 TargetLowering::AddrMode AM;
734 return TLI->isLegalAddressingMode(AM, UseTy);
736 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
737 return (VC > -(1 << 16) && VC < (1 << 16)-1);
741 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
742 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
743 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
744 Constant *Op0 = CE->getOperand(0);
745 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
746 TargetLowering::AddrMode AM;
748 return TLI->isLegalAddressingMode(AM, UseTy);
754 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
755 /// loop varying to the Imm operand.
756 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
757 Loop *L, ScalarEvolution *SE) {
758 if (Val->isLoopInvariant(L)) return; // Nothing to do.
760 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
761 std::vector<SCEVHandle> NewOps;
762 NewOps.reserve(SAE->getNumOperands());
764 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
765 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
766 // If this is a loop-variant expression, it must stay in the immediate
767 // field of the expression.
768 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
770 NewOps.push_back(SAE->getOperand(i));
774 Val = SE->getIntegerSCEV(0, Val->getType());
776 Val = SE->getAddExpr(NewOps);
777 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
778 // Try to pull immediates out of the start value of nested addrec's.
779 SCEVHandle Start = SARE->getStart();
780 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
782 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
784 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
786 // Otherwise, all of Val is variant, move the whole thing over.
787 Imm = SE->getAddExpr(Imm, Val);
788 Val = SE->getIntegerSCEV(0, Val->getType());
793 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
794 /// that can fit into the immediate field of instructions in the target.
795 /// Accumulate these immediate values into the Imm value.
796 static void MoveImmediateValues(const TargetLowering *TLI,
798 SCEVHandle &Val, SCEVHandle &Imm,
799 bool isAddress, Loop *L,
800 ScalarEvolution *SE) {
801 const Type *UseTy = User->getType();
802 if (StoreInst *SI = dyn_cast<StoreInst>(User))
803 UseTy = SI->getOperand(0)->getType();
805 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
806 std::vector<SCEVHandle> NewOps;
807 NewOps.reserve(SAE->getNumOperands());
809 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
810 SCEVHandle NewOp = SAE->getOperand(i);
811 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
813 if (!NewOp->isLoopInvariant(L)) {
814 // If this is a loop-variant expression, it must stay in the immediate
815 // field of the expression.
816 Imm = SE->getAddExpr(Imm, NewOp);
818 NewOps.push_back(NewOp);
823 Val = SE->getIntegerSCEV(0, Val->getType());
825 Val = SE->getAddExpr(NewOps);
827 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
828 // Try to pull immediates out of the start value of nested addrec's.
829 SCEVHandle Start = SARE->getStart();
830 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
832 if (Start != SARE->getStart()) {
833 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
835 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
838 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
839 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
840 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
841 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
843 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
844 SCEVHandle NewOp = SME->getOperand(1);
845 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
847 // If we extracted something out of the subexpressions, see if we can
849 if (NewOp != SME->getOperand(1)) {
850 // Scale SubImm up by "8". If the result is a target constant, we are
852 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
853 if (isTargetConstant(SubImm, UseTy, TLI)) {
854 // Accumulate the immediate.
855 Imm = SE->getAddExpr(Imm, SubImm);
857 // Update what is left of 'Val'.
858 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
865 // Loop-variant expressions must stay in the immediate field of the
867 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
868 !Val->isLoopInvariant(L)) {
869 Imm = SE->getAddExpr(Imm, Val);
870 Val = SE->getIntegerSCEV(0, Val->getType());
874 // Otherwise, no immediates to move.
878 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
879 /// added together. This is used to reassociate common addition subexprs
880 /// together for maximal sharing when rewriting bases.
881 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
883 ScalarEvolution *SE) {
884 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
885 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
886 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
887 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
888 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
889 if (SARE->getOperand(0) == Zero) {
890 SubExprs.push_back(Expr);
892 // Compute the addrec with zero as its base.
893 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
894 Ops[0] = Zero; // Start with zero base.
895 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
898 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
900 } else if (!Expr->isZero()) {
902 SubExprs.push_back(Expr);
907 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
908 /// removing any common subexpressions from it. Anything truly common is
909 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
910 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
912 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
913 ScalarEvolution *SE) {
914 unsigned NumUses = Uses.size();
916 // Only one use? Use its base, regardless of what it is!
917 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
918 SCEVHandle Result = Zero;
920 std::swap(Result, Uses[0].Base);
924 // To find common subexpressions, count how many of Uses use each expression.
925 // If any subexpressions are used Uses.size() times, they are common.
926 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
928 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
929 // order we see them.
930 std::vector<SCEVHandle> UniqueSubExprs;
932 std::vector<SCEVHandle> SubExprs;
933 for (unsigned i = 0; i != NumUses; ++i) {
934 // If the base is zero (which is common), return zero now, there are no
936 if (Uses[i].Base == Zero) return Zero;
938 // Split the expression into subexprs.
939 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
940 // Add one to SubExpressionUseCounts for each subexpr present.
941 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
942 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
943 UniqueSubExprs.push_back(SubExprs[j]);
947 // Now that we know how many times each is used, build Result. Iterate over
948 // UniqueSubexprs so that we have a stable ordering.
949 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
950 std::map<SCEVHandle, unsigned>::iterator I =
951 SubExpressionUseCounts.find(UniqueSubExprs[i]);
952 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
953 if (I->second == NumUses) { // Found CSE!
954 Result = SE->getAddExpr(Result, I->first);
956 // Remove non-cse's from SubExpressionUseCounts.
957 SubExpressionUseCounts.erase(I);
961 // If we found no CSE's, return now.
962 if (Result == Zero) return Result;
964 // Otherwise, remove all of the CSE's we found from each of the base values.
965 for (unsigned i = 0; i != NumUses; ++i) {
966 // Split the expression into subexprs.
967 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
969 // Remove any common subexpressions.
970 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
971 if (SubExpressionUseCounts.count(SubExprs[j])) {
972 SubExprs.erase(SubExprs.begin()+j);
976 // Finally, the non-shared expressions together.
977 if (SubExprs.empty())
980 Uses[i].Base = SE->getAddExpr(SubExprs);
987 /// ValidStride - Check whether the given Scale is valid for all loads and
988 /// stores in UsersToProcess.
990 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
992 const std::vector<BasedUser>& UsersToProcess) {
996 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
997 // If this is a load or other access, pass the type of the access in.
998 const Type *AccessTy = Type::VoidTy;
999 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1000 AccessTy = SI->getOperand(0)->getType();
1001 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1002 AccessTy = LI->getType();
1003 else if (isa<PHINode>(UsersToProcess[i].Inst))
1006 TargetLowering::AddrMode AM;
1007 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1008 AM.BaseOffs = SC->getValue()->getSExtValue();
1009 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1012 // If load[imm+r*scale] is illegal, bail out.
1013 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1019 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1021 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1025 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1027 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1028 !(isa<PointerType>(Ty2) &&
1029 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1030 !(isa<PointerType>(Ty1) &&
1031 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1034 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1035 /// of a previous stride and it is a legal value for the target addressing
1036 /// mode scale component and optional base reg. This allows the users of
1037 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1038 /// reuse is possible.
1039 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1040 bool AllUsesAreAddresses,
1041 const SCEVHandle &Stride,
1042 IVExpr &IV, const Type *Ty,
1043 const std::vector<BasedUser>& UsersToProcess) {
1044 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1045 int64_t SInt = SC->getValue()->getSExtValue();
1046 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1048 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1049 IVsByStride.find(StrideOrder[NewStride]);
1050 if (SI == IVsByStride.end())
1052 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1053 if (SI->first != Stride &&
1054 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1056 int64_t Scale = SInt / SSInt;
1057 // Check that this stride is valid for all the types used for loads and
1058 // stores; if it can be used for some and not others, we might as well use
1059 // the original stride everywhere, since we have to create the IV for it
1060 // anyway. If the scale is 1, then we don't need to worry about folding
1063 (AllUsesAreAddresses &&
1064 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1065 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1066 IE = SI->second.IVs.end(); II != IE; ++II)
1067 // FIXME: Only handle base == 0 for now.
1068 // Only reuse previous IV if it would not require a type conversion.
1069 if (II->Base->isZero() &&
1070 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1079 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1080 /// returns true if Val's isUseOfPostIncrementedValue is true.
1081 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1082 return Val.isUseOfPostIncrementedValue;
1085 /// isNonConstantNegative - Return true if the specified scev is negated, but
1087 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1088 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1089 if (!Mul) return false;
1091 // If there is a constant factor, it will be first.
1092 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1093 if (!SC) return false;
1095 // Return true if the value is negative, this matches things like (-42 * V).
1096 return SC->getValue()->getValue().isNegative();
1099 /// isAddress - Returns true if the specified instruction is using the
1100 /// specified value as an address.
1101 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
1102 bool isAddress = isa<LoadInst>(Inst);
1103 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1104 if (SI->getOperand(1) == OperandVal)
1106 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1107 // Addressing modes can also be folded into prefetches and a variety
1109 switch (II->getIntrinsicID()) {
1111 case Intrinsic::prefetch:
1112 case Intrinsic::x86_sse2_loadu_dq:
1113 case Intrinsic::x86_sse2_loadu_pd:
1114 case Intrinsic::x86_sse_loadu_ps:
1115 case Intrinsic::x86_sse_storeu_ps:
1116 case Intrinsic::x86_sse2_storeu_pd:
1117 case Intrinsic::x86_sse2_storeu_dq:
1118 case Intrinsic::x86_sse2_storel_dq:
1119 if (II->getOperand(1) == OperandVal)
1127 // CollectIVUsers - Transform our list of users and offsets to a bit more
1128 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1129 // of the strided accesses, as well as the old information from Uses. We
1130 // progressively move information from the Base field to the Imm field, until
1131 // we eventually have the full access expression to rewrite the use.
1132 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1133 IVUsersOfOneStride &Uses,
1135 bool &AllUsesAreAddresses,
1136 std::vector<BasedUser> &UsersToProcess) {
1137 UsersToProcess.reserve(Uses.Users.size());
1138 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1139 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1141 // Move any loop invariant operands from the offset field to the immediate
1142 // field of the use, so that we don't try to use something before it is
1144 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1145 UsersToProcess.back().Imm, L, SE);
1146 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1147 "Base value is not loop invariant!");
1150 // We now have a whole bunch of uses of like-strided induction variables, but
1151 // they might all have different bases. We want to emit one PHI node for this
1152 // stride which we fold as many common expressions (between the IVs) into as
1153 // possible. Start by identifying the common expressions in the base values
1154 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1155 // "A+B"), emit it to the preheader, then remove the expression from the
1156 // UsersToProcess base values.
1157 SCEVHandle CommonExprs =
1158 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
1160 // Next, figure out what we can represent in the immediate fields of
1161 // instructions. If we can represent anything there, move it to the imm
1162 // fields of the BasedUsers. We do this so that it increases the commonality
1163 // of the remaining uses.
1164 unsigned NumPHI = 0;
1165 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1166 // If the user is not in the current loop, this means it is using the exit
1167 // value of the IV. Do not put anything in the base, make sure it's all in
1168 // the immediate field to allow as much factoring as possible.
1169 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1170 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1171 UsersToProcess[i].Base);
1172 UsersToProcess[i].Base =
1173 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1176 // Addressing modes can be folded into loads and stores. Be careful that
1177 // the store is through the expression, not of the expression though.
1179 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1180 UsersToProcess[i].OperandValToReplace);
1181 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1186 // If this use isn't an address, then not all uses are addresses.
1187 if (!isAddress && !isPHI)
1188 AllUsesAreAddresses = false;
1190 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1191 UsersToProcess[i].Imm, isAddress, L, SE);
1195 // If one of the use if a PHI node and all other uses are addresses, still
1196 // allow iv reuse. Essentially we are trading one constant multiplication
1197 // for one fewer iv.
1199 AllUsesAreAddresses = false;
1204 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1205 /// stride of IV. All of the users may have different starting values, and this
1206 /// may not be the only stride (we know it is if isOnlyStride is true).
1207 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1208 IVUsersOfOneStride &Uses,
1210 bool isOnlyStride) {
1211 // If all the users are moved to another stride, then there is nothing to do.
1212 if (Uses.Users.empty())
1215 // Keep track if every use in UsersToProcess is an address. If they all are,
1216 // we may be able to rewrite the entire collection of them in terms of a
1217 // smaller-stride IV.
1218 bool AllUsesAreAddresses = true;
1220 // Transform our list of users and offsets to a bit more complex table. In
1221 // this new vector, each 'BasedUser' contains 'Base' the base of the
1222 // strided accessas well as the old information from Uses. We progressively
1223 // move information from the Base field to the Imm field, until we eventually
1224 // have the full access expression to rewrite the use.
1225 std::vector<BasedUser> UsersToProcess;
1226 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1229 // If we managed to find some expressions in common, we'll need to carry
1230 // their value in a register and add it in for each use. This will take up
1231 // a register operand, which potentially restricts what stride values are
1233 bool HaveCommonExprs = !CommonExprs->isZero();
1235 // If all uses are addresses, check if it is possible to reuse an IV with a
1236 // stride that is a factor of this stride. And that the multiple is a number
1237 // that can be encoded in the scale field of the target addressing mode. And
1238 // that we will have a valid instruction after this substition, including the
1239 // immediate field, if any.
1240 PHINode *NewPHI = NULL;
1242 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1243 SE->getIntegerSCEV(0, Type::Int32Ty),
1245 unsigned RewriteFactor = 0;
1246 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1247 Stride, ReuseIV, CommonExprs->getType(),
1249 if (RewriteFactor != 0) {
1250 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1251 << " and BASE " << *ReuseIV.Base << " :\n";
1252 NewPHI = ReuseIV.PHI;
1253 IncV = ReuseIV.IncV;
1256 const Type *ReplacedTy = CommonExprs->getType();
1258 // Now that we know what we need to do, insert the PHI node itself.
1260 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1261 << *Stride << " and BASE " << *CommonExprs << ": ";
1263 SCEVExpander Rewriter(*SE, *LI);
1264 SCEVExpander PreheaderRewriter(*SE, *LI);
1266 BasicBlock *Preheader = L->getLoopPreheader();
1267 Instruction *PreInsertPt = Preheader->getTerminator();
1268 Instruction *PhiInsertBefore = L->getHeader()->begin();
1270 BasicBlock *LatchBlock = L->getLoopLatch();
1273 // Emit the initial base value into the loop preheader.
1275 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1277 if (RewriteFactor == 0) {
1278 // Create a new Phi for this base, and stick it in the loop header.
1279 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1282 // Add common base to the new Phi node.
1283 NewPHI->addIncoming(CommonBaseV, Preheader);
1285 // If the stride is negative, insert a sub instead of an add for the
1287 bool isNegative = isNonConstantNegative(Stride);
1288 SCEVHandle IncAmount = Stride;
1290 IncAmount = SE->getNegativeSCEV(Stride);
1292 // Insert the stride into the preheader.
1293 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1294 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1296 // Emit the increment of the base value before the terminator of the loop
1297 // latch block, and add it to the Phi node.
1298 SCEVHandle IncExp = SE->getUnknown(StrideV);
1300 IncExp = SE->getNegativeSCEV(IncExp);
1301 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1303 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1304 IncV->setName(NewPHI->getName()+".inc");
1305 NewPHI->addIncoming(IncV, LatchBlock);
1307 // Remember this in case a later stride is multiple of this.
1308 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1310 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1312 Constant *C = dyn_cast<Constant>(CommonBaseV);
1314 (!C->isNullValue() &&
1315 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1316 // We want the common base emitted into the preheader! This is just
1317 // using cast as a copy so BitCast (no-op cast) is appropriate
1318 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1319 "commonbase", PreInsertPt);
1323 // We want to emit code for users inside the loop first. To do this, we
1324 // rearrange BasedUser so that the entries at the end have
1325 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1326 // vector (so we handle them first).
1327 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1328 PartitionByIsUseOfPostIncrementedValue);
1330 // Sort this by base, so that things with the same base are handled
1331 // together. By partitioning first and stable-sorting later, we are
1332 // guaranteed that within each base we will pop off users from within the
1333 // loop before users outside of the loop with a particular base.
1335 // We would like to use stable_sort here, but we can't. The problem is that
1336 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1337 // we don't have anything to do a '<' comparison on. Because we think the
1338 // number of uses is small, do a horrible bubble sort which just relies on
1340 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1341 // Get a base value.
1342 SCEVHandle Base = UsersToProcess[i].Base;
1344 // Compact everything with this base to be consequtive with this one.
1345 for (unsigned j = i+1; j != e; ++j) {
1346 if (UsersToProcess[j].Base == Base) {
1347 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1353 // Process all the users now. This outer loop handles all bases, the inner
1354 // loop handles all users of a particular base.
1355 while (!UsersToProcess.empty()) {
1356 SCEVHandle Base = UsersToProcess.back().Base;
1358 // Emit the code for Base into the preheader.
1359 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1361 DOUT << " INSERTING code for BASE = " << *Base << ":";
1362 if (BaseV->hasName())
1363 DOUT << " Result value name = %" << BaseV->getNameStr();
1366 // If BaseV is a constant other than 0, make sure that it gets inserted into
1367 // the preheader, instead of being forward substituted into the uses. We do
1368 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1370 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1371 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1372 // We want this constant emitted into the preheader! This is just
1373 // using cast as a copy so BitCast (no-op cast) is appropriate
1374 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1379 // Emit the code to add the immediate offset to the Phi value, just before
1380 // the instructions that we identified as using this stride and base.
1382 // FIXME: Use emitted users to emit other users.
1383 BasedUser &User = UsersToProcess.back();
1385 // If this instruction wants to use the post-incremented value, move it
1386 // after the post-inc and use its value instead of the PHI.
1387 Value *RewriteOp = NewPHI;
1388 if (User.isUseOfPostIncrementedValue) {
1391 // If this user is in the loop, make sure it is the last thing in the
1392 // loop to ensure it is dominated by the increment.
1393 if (L->contains(User.Inst->getParent()))
1394 User.Inst->moveBefore(LatchBlock->getTerminator());
1396 if (RewriteOp->getType() != ReplacedTy) {
1397 Instruction::CastOps opcode = Instruction::Trunc;
1398 if (ReplacedTy->getPrimitiveSizeInBits() ==
1399 RewriteOp->getType()->getPrimitiveSizeInBits())
1400 opcode = Instruction::BitCast;
1401 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1404 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1406 // If we had to insert new instrutions for RewriteOp, we have to
1407 // consider that they may not have been able to end up immediately
1408 // next to RewriteOp, because non-PHI instructions may never precede
1409 // PHI instructions in a block. In this case, remember where the last
1410 // instruction was inserted so that if we're replacing a different
1411 // PHI node, we can use the later point to expand the final
1413 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1414 if (RewriteOp == NewPHI) NewBasePt = 0;
1416 // Clear the SCEVExpander's expression map so that we are guaranteed
1417 // to have the code emitted where we expect it.
1420 // If we are reusing the iv, then it must be multiplied by a constant
1421 // factor take advantage of addressing mode scale component.
1422 if (RewriteFactor != 0) {
1423 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1424 RewriteExpr->getType()),
1427 // The common base is emitted in the loop preheader. But since we
1428 // are reusing an IV, it has not been used to initialize the PHI node.
1429 // Add it to the expression used to rewrite the uses.
1430 if (!isa<ConstantInt>(CommonBaseV) ||
1431 !cast<ConstantInt>(CommonBaseV)->isZero())
1432 RewriteExpr = SE->getAddExpr(RewriteExpr,
1433 SE->getUnknown(CommonBaseV));
1436 // Now that we know what we need to do, insert code before User for the
1437 // immediate and any loop-variant expressions.
1438 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1439 // Add BaseV to the PHI value if needed.
1440 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1442 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1446 // Mark old value we replaced as possibly dead, so that it is elminated
1447 // if we just replaced the last use of that value.
1448 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1450 UsersToProcess.pop_back();
1453 // If there are any more users to process with the same base, process them
1454 // now. We sorted by base above, so we just have to check the last elt.
1455 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1456 // TODO: Next, find out which base index is the most common, pull it out.
1459 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1460 // different starting values, into different PHIs.
1463 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1464 /// set the IV user and stride information and return true, otherwise return
1466 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1467 const SCEVHandle *&CondStride) {
1468 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1470 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1471 IVUsesByStride.find(StrideOrder[Stride]);
1472 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1474 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1475 E = SI->second.Users.end(); UI != E; ++UI)
1476 if (UI->User == Cond) {
1477 // NOTE: we could handle setcc instructions with multiple uses here, but
1478 // InstCombine does it as well for simple uses, it's not clear that it
1479 // occurs enough in real life to handle.
1481 CondStride = &SI->first;
1489 // Constant strides come first which in turns are sorted by their absolute
1490 // values. If absolute values are the same, then positive strides comes first.
1492 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1493 struct StrideCompare {
1494 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1495 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1496 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1498 int64_t LV = LHSC->getValue()->getSExtValue();
1499 int64_t RV = RHSC->getValue()->getSExtValue();
1500 uint64_t ALV = (LV < 0) ? -LV : LV;
1501 uint64_t ARV = (RV < 0) ? -RV : RV;
1507 return (LHSC && !RHSC);
1512 /// ChangeCompareStride - If a loop termination compare instruction is the
1513 /// only use of its stride, and the compaison is against a constant value,
1514 /// try eliminate the stride by moving the compare instruction to another
1515 /// stride and change its constant operand accordingly. e.g.
1521 /// if (v2 < 10) goto loop
1526 /// if (v1 < 30) goto loop
1527 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1528 IVStrideUse* &CondUse,
1529 const SCEVHandle* &CondStride) {
1530 if (StrideOrder.size() < 2 ||
1531 IVUsesByStride[*CondStride].Users.size() != 1)
1533 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1534 if (!SC) return Cond;
1535 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1536 if (!C) return Cond;
1538 ICmpInst::Predicate Predicate = Cond->getPredicate();
1539 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1540 int64_t CmpVal = C->getValue().getSExtValue();
1541 unsigned BitWidth = C->getValue().getBitWidth();
1542 uint64_t SignBit = 1ULL << (BitWidth-1);
1543 const Type *CmpTy = C->getType();
1544 const Type *NewCmpTy = NULL;
1545 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1546 unsigned NewTyBits = 0;
1547 int64_t NewCmpVal = CmpVal;
1548 SCEVHandle *NewStride = NULL;
1549 Value *NewIncV = NULL;
1552 // Check stride constant and the comparision constant signs to detect
1554 if (ICmpInst::isSignedPredicate(Predicate) &&
1555 (CmpVal & SignBit) != (CmpSSInt & SignBit))
1558 // Look for a suitable stride / iv as replacement.
1559 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1560 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1561 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1562 IVUsesByStride.find(StrideOrder[i]);
1563 if (!isa<SCEVConstant>(SI->first))
1565 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1566 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1569 Scale = SSInt / CmpSSInt;
1570 NewCmpVal = CmpVal * Scale;
1571 APInt Mul = APInt(BitWidth, NewCmpVal);
1572 // Check for overflow.
1573 if (Mul.getSExtValue() != NewCmpVal) {
1578 // Watch out for overflow.
1579 if (ICmpInst::isSignedPredicate(Predicate) &&
1580 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1583 if (NewCmpVal != CmpVal) {
1584 // Pick the best iv to use trying to avoid a cast.
1586 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1587 E = SI->second.Users.end(); UI != E; ++UI) {
1588 NewIncV = UI->OperandValToReplace;
1589 if (NewIncV->getType() == CmpTy)
1597 NewCmpTy = NewIncV->getType();
1598 NewTyBits = isa<PointerType>(NewCmpTy)
1599 ? UIntPtrTy->getPrimitiveSizeInBits()
1600 : NewCmpTy->getPrimitiveSizeInBits();
1601 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1602 // Check if it is possible to rewrite it using
1603 // an iv / stride of a smaller integer type.
1604 bool TruncOk = false;
1605 if (NewCmpTy->isInteger()) {
1606 unsigned Bits = NewTyBits;
1607 if (ICmpInst::isSignedPredicate(Predicate))
1609 uint64_t Mask = (1ULL << Bits) - 1;
1610 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1619 // Don't rewrite if use offset is non-constant and the new type is
1620 // of a different type.
1621 // FIXME: too conservative?
1622 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1627 bool AllUsesAreAddresses = true;
1628 std::vector<BasedUser> UsersToProcess;
1629 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1630 AllUsesAreAddresses,
1632 // Avoid rewriting the compare instruction with an iv of new stride
1633 // if it's likely the new stride uses will be rewritten using the
1634 if (AllUsesAreAddresses &&
1635 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1640 // If scale is negative, use swapped predicate unless it's testing
1642 if (Scale < 0 && !Cond->isEquality())
1643 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1645 NewStride = &StrideOrder[i];
1650 // Forgo this transformation if it the increment happens to be
1651 // unfortunately positioned after the condition, and the condition
1652 // has multiple uses which prevent it from being moved immediately
1653 // before the branch. See
1654 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1655 // for an example of this situation.
1656 if (!Cond->hasOneUse()) {
1657 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1663 if (NewCmpVal != CmpVal) {
1664 // Create a new compare instruction using new stride / iv.
1665 ICmpInst *OldCond = Cond;
1667 if (!isa<PointerType>(NewCmpTy))
1668 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1670 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1671 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1673 // Insert new compare instruction.
1674 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1675 L->getHeader()->getName() + ".termcond",
1678 // Remove the old compare instruction. The old indvar is probably dead too.
1679 DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
1680 SE->deleteValueFromRecords(OldCond);
1681 OldCond->replaceAllUsesWith(Cond);
1682 OldCond->eraseFromParent();
1684 IVUsesByStride[*CondStride].Users.pop_back();
1685 SCEVHandle NewOffset = TyBits == NewTyBits
1686 ? SE->getMulExpr(CondUse->Offset,
1687 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1688 : SE->getConstant(ConstantInt::get(NewCmpTy,
1689 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1690 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1691 CondUse = &IVUsesByStride[*NewStride].Users.back();
1692 CondStride = NewStride;
1699 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1700 /// inside the loop then try to eliminate the cast opeation.
1701 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
1703 SCEVHandle IterationCount = SE->getIterationCount(L);
1704 if (isa<SCEVCouldNotCompute>(IterationCount))
1707 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
1709 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1710 IVUsesByStride.find(StrideOrder[Stride]);
1711 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1712 if (!isa<SCEVConstant>(SI->first))
1715 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1716 E = SI->second.Users.end(); UI != E; /* empty */) {
1717 std::vector<IVStrideUse>::iterator CandidateUI = UI;
1719 Instruction *ShadowUse = CandidateUI->User;
1720 const Type *DestTy = NULL;
1722 /* If shadow use is a int->float cast then insert a second IV
1723 to eliminate this cast.
1725 for (unsigned i = 0; i < n; ++i)
1731 for (unsigned i = 0; i < n; ++i, ++d)
1734 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
1735 DestTy = UCast->getDestTy();
1736 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
1737 DestTy = SCast->getDestTy();
1738 if (!DestTy) continue;
1741 /* If target does not support DestTy natively then do not apply
1742 this transformation. */
1743 MVT DVT = TLI->getValueType(DestTy);
1744 if (!TLI->isTypeLegal(DVT)) continue;
1747 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1749 if (PH->getNumIncomingValues() != 2) continue;
1751 const Type *SrcTy = PH->getType();
1752 int Mantissa = DestTy->getFPMantissaWidth();
1753 if (Mantissa == -1) continue;
1754 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
1757 unsigned Entry, Latch;
1758 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1766 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1767 if (!Init) continue;
1768 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1770 BinaryOperator *Incr =
1771 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1772 if (!Incr) continue;
1773 if (Incr->getOpcode() != Instruction::Add
1774 && Incr->getOpcode() != Instruction::Sub)
1777 /* Initialize new IV, double d = 0.0 in above example. */
1778 ConstantInt *C = NULL;
1779 if (Incr->getOperand(0) == PH)
1780 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1781 else if (Incr->getOperand(1) == PH)
1782 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1788 /* Add new PHINode. */
1789 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1791 /* create new increment. '++d' in above example. */
1792 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1793 BinaryOperator *NewIncr =
1794 BinaryOperator::Create(Incr->getOpcode(),
1795 NewPH, CFP, "IV.S.next.", Incr);
1797 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1798 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1800 /* Remove cast operation */
1801 SE->deleteValueFromRecords(ShadowUse);
1802 ShadowUse->replaceAllUsesWith(NewPH);
1803 ShadowUse->eraseFromParent();
1804 SI->second.Users.erase(CandidateUI);
1811 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1812 // uses in the loop, look to see if we can eliminate some, in favor of using
1813 // common indvars for the different uses.
1814 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1815 // TODO: implement optzns here.
1817 OptimizeShadowIV(L);
1819 // Finally, get the terminating condition for the loop if possible. If we
1820 // can, we want to change it to use a post-incremented version of its
1821 // induction variable, to allow coalescing the live ranges for the IV into
1822 // one register value.
1823 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1824 BasicBlock *Preheader = L->getLoopPreheader();
1825 BasicBlock *LatchBlock =
1826 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1827 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1828 if (!TermBr || TermBr->isUnconditional() ||
1829 !isa<ICmpInst>(TermBr->getCondition()))
1831 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1833 // Search IVUsesByStride to find Cond's IVUse if there is one.
1834 IVStrideUse *CondUse = 0;
1835 const SCEVHandle *CondStride = 0;
1837 if (!FindIVUserForCond(Cond, CondUse, CondStride))
1838 return; // setcc doesn't use the IV.
1840 // If possible, change stride and operands of the compare instruction to
1841 // eliminate one stride.
1842 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
1844 // It's possible for the setcc instruction to be anywhere in the loop, and
1845 // possible for it to have multiple users. If it is not immediately before
1846 // the latch block branch, move it.
1847 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1848 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1849 Cond->moveBefore(TermBr);
1851 // Otherwise, clone the terminating condition and insert into the loopend.
1852 Cond = cast<ICmpInst>(Cond->clone());
1853 Cond->setName(L->getHeader()->getName() + ".termcond");
1854 LatchBlock->getInstList().insert(TermBr, Cond);
1856 // Clone the IVUse, as the old use still exists!
1857 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1858 CondUse->OperandValToReplace);
1859 CondUse = &IVUsesByStride[*CondStride].Users.back();
1863 // If we get to here, we know that we can transform the setcc instruction to
1864 // use the post-incremented version of the IV, allowing us to coalesce the
1865 // live ranges for the IV correctly.
1866 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
1867 CondUse->isUseOfPostIncrementedValue = true;
1871 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1873 LI = &getAnalysis<LoopInfo>();
1874 DT = &getAnalysis<DominatorTree>();
1875 SE = &getAnalysis<ScalarEvolution>();
1876 TD = &getAnalysis<TargetData>();
1877 UIntPtrTy = TD->getIntPtrType();
1880 // Find all uses of induction variables in this loop, and catagorize
1881 // them by stride. Start by finding all of the PHI nodes in the header for
1882 // this loop. If they are induction variables, inspect their uses.
1883 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
1884 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1885 AddUsersIfInteresting(I, L, Processed);
1887 if (!IVUsesByStride.empty()) {
1888 // Optimize induction variables. Some indvar uses can be transformed to use
1889 // strides that will be needed for other purposes. A common example of this
1890 // is the exit test for the loop, which can often be rewritten to use the
1891 // computation of some other indvar to decide when to terminate the loop.
1894 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1895 // doing computation in byte values, promote to 32-bit values if safe.
1897 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1898 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
1899 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
1900 // Need to be careful that IV's are all the same type. Only works for
1901 // intptr_t indvars.
1903 // If we only have one stride, we can more aggressively eliminate some
1905 bool HasOneStride = IVUsesByStride.size() == 1;
1908 DOUT << "\nLSR on ";
1912 // IVsByStride keeps IVs for one particular loop.
1913 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
1915 // Sort the StrideOrder so we process larger strides first.
1916 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1918 // Note: this processes each stride/type pair individually. All users
1919 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
1920 // Also, note that we iterate over IVUsesByStride indirectly by using
1921 // StrideOrder. This extra layer of indirection makes the ordering of
1922 // strides deterministic - not dependent on map order.
1923 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1924 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1925 IVUsesByStride.find(StrideOrder[Stride]);
1926 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1927 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1931 // We're done analyzing this loop; release all the state we built up for it.
1932 CastedPointers.clear();
1933 IVUsesByStride.clear();
1934 IVsByStride.clear();
1935 StrideOrder.clear();
1937 // Clean up after ourselves
1938 if (!DeadInsts.empty()) {
1939 DeleteTriviallyDeadInstructions(DeadInsts);
1941 BasicBlock::iterator I = L->getHeader()->begin();
1942 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
1943 // At this point, we know that we have killed one or more IV users.
1944 // It is worth checking to see if the cann indvar is also
1945 // dead, so that we can remove it as well.
1947 // We can remove a PHI if it is on a cycle in the def-use graph
1948 // where each node in the cycle has degree one, i.e. only one use,
1949 // and is an instruction with no side effects.
1951 // FIXME: this needs to eliminate an induction variable even if it's being
1952 // compared against some value to decide loop termination.
1953 if (PN->hasOneUse()) {
1954 SmallPtrSet<PHINode *, 2> PHIs;
1955 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
1956 J && J->hasOneUse() && !J->mayWriteToMemory();
1957 J = dyn_cast<Instruction>(*J->use_begin())) {
1958 // If we find the original PHI, we've discovered a cycle.
1960 // Break the cycle and mark the PHI for deletion.
1961 SE->deleteValueFromRecords(PN);
1962 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1963 DeadInsts.insert(PN);
1967 // If we find a PHI more than once, we're on a cycle that
1968 // won't prove fruitful.
1969 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
1974 DeleteTriviallyDeadInstructions(DeadInsts);