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((intptr_t)&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>();
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, ICmpInst *Cond,
185 const SCEVHandle *&CondStride);
186 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
187 const SCEVHandle *&CondStride);
188 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
189 unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
190 IVExpr&, const Type*,
191 const std::vector<BasedUser>& UsersToProcess);
192 bool ValidStride(bool, int64_t,
193 const std::vector<BasedUser>& UsersToProcess);
194 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
195 IVUsersOfOneStride &Uses,
197 bool &AllUsesAreAddresses,
198 std::vector<BasedUser> &UsersToProcess);
199 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
200 IVUsersOfOneStride &Uses,
201 Loop *L, bool isOnlyStride);
202 void DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts);
206 char LoopStrengthReduce::ID = 0;
207 static RegisterPass<LoopStrengthReduce>
208 X("loop-reduce", "Loop Strength Reduction");
210 LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
211 return new LoopStrengthReduce(TLI);
214 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
215 /// assumes that the Value* V is of integer or pointer type only.
217 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
219 if (V->getType() == UIntPtrTy) return V;
220 if (Constant *CB = dyn_cast<Constant>(V))
221 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
223 Value *&New = CastedPointers[V];
226 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
227 DeadInsts.insert(cast<Instruction>(New));
232 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
233 /// specified set are trivially dead, delete them and see if this makes any of
234 /// their operands subsequently dead.
235 void LoopStrengthReduce::
236 DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts) {
237 while (!Insts.empty()) {
238 Instruction *I = Insts.back();
241 if (PHINode *PN = dyn_cast<PHINode>(I)) {
242 // If all incoming values to the Phi are the same, we can replace the Phi
244 if (Value *PNV = PN->hasConstantValue()) {
245 if (Instruction *U = dyn_cast<Instruction>(PNV))
247 SE->deleteValueFromRecords(PN);
248 PN->replaceAllUsesWith(PNV);
249 PN->eraseFromParent();
255 if (isInstructionTriviallyDead(I)) {
256 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
257 if (Instruction *U = dyn_cast<Instruction>(*i))
259 SE->deleteValueFromRecords(I);
260 I->eraseFromParent();
267 /// GetExpressionSCEV - Compute and return the SCEV for the specified
269 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
270 // Pointer to pointer bitcast instructions return the same value as their
272 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
273 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
274 return SE->getSCEV(BCI);
275 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
280 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
281 // If this is a GEP that SE doesn't know about, compute it now and insert it.
282 // If this is not a GEP, or if we have already done this computation, just let
284 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
285 if (!GEP || SE->hasSCEV(GEP))
286 return SE->getSCEV(Exp);
288 // Analyze all of the subscripts of this getelementptr instruction, looking
289 // for uses that are determined by the trip count of the loop. First, skip
290 // all operands the are not dependent on the IV.
292 // Build up the base expression. Insert an LLVM cast of the pointer to
294 SCEVHandle GEPVal = SE->getUnknown(
295 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
297 gep_type_iterator GTI = gep_type_begin(GEP);
299 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
300 i != e; ++i, ++GTI) {
301 // If this is a use of a recurrence that we can analyze, and it comes before
302 // Op does in the GEP operand list, we will handle this when we process this
304 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
305 const StructLayout *SL = TD->getStructLayout(STy);
306 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
307 uint64_t Offset = SL->getElementOffset(Idx);
308 GEPVal = SE->getAddExpr(GEPVal,
309 SE->getIntegerSCEV(Offset, UIntPtrTy));
311 unsigned GEPOpiBits =
312 (*i)->getType()->getPrimitiveSizeInBits();
313 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
314 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
315 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
316 Instruction::BitCast));
317 Value *OpVal = getCastedVersionOf(opcode, *i);
318 SCEVHandle Idx = SE->getSCEV(OpVal);
320 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
322 Idx = SE->getMulExpr(Idx,
323 SE->getConstant(ConstantInt::get(UIntPtrTy,
325 GEPVal = SE->getAddExpr(GEPVal, Idx);
329 SE->setSCEV(GEP, GEPVal);
333 /// getSCEVStartAndStride - Compute the start and stride of this expression,
334 /// returning false if the expression is not a start/stride pair, or true if it
335 /// is. The stride must be a loop invariant expression, but the start may be
336 /// a mix of loop invariant and loop variant expressions.
337 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
338 SCEVHandle &Start, SCEVHandle &Stride,
339 ScalarEvolution *SE) {
340 SCEVHandle TheAddRec = Start; // Initialize to zero.
342 // If the outer level is an AddExpr, the operands are all start values except
343 // for a nested AddRecExpr.
344 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
345 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
346 if (SCEVAddRecExpr *AddRec =
347 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
348 if (AddRec->getLoop() == L)
349 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
351 return false; // Nested IV of some sort?
353 Start = SE->getAddExpr(Start, AE->getOperand(i));
356 } else if (isa<SCEVAddRecExpr>(SH)) {
359 return false; // not analyzable.
362 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
363 if (!AddRec || AddRec->getLoop() != L) return false;
365 // FIXME: Generalize to non-affine IV's.
366 if (!AddRec->isAffine()) return false;
368 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
370 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
371 DOUT << "[" << L->getHeader()->getName()
372 << "] Variable stride: " << *AddRec << "\n";
374 Stride = AddRec->getOperand(1);
378 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
379 /// and now we need to decide whether the user should use the preinc or post-inc
380 /// value. If this user should use the post-inc version of the IV, return true.
382 /// Choosing wrong here can break dominance properties (if we choose to use the
383 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
384 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
385 /// should use the post-inc value).
386 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
387 Loop *L, DominatorTree *DT, Pass *P,
388 SetVector<Instruction*> &DeadInsts){
389 // If the user is in the loop, use the preinc value.
390 if (L->contains(User->getParent())) return false;
392 BasicBlock *LatchBlock = L->getLoopLatch();
394 // Ok, the user is outside of the loop. If it is dominated by the latch
395 // block, use the post-inc value.
396 if (DT->dominates(LatchBlock, User->getParent()))
399 // There is one case we have to be careful of: PHI nodes. These little guys
400 // can live in blocks that do not dominate the latch block, but (since their
401 // uses occur in the predecessor block, not the block the PHI lives in) should
402 // still use the post-inc value. Check for this case now.
403 PHINode *PN = dyn_cast<PHINode>(User);
404 if (!PN) return false; // not a phi, not dominated by latch block.
406 // Look at all of the uses of IV by the PHI node. If any use corresponds to
407 // a block that is not dominated by the latch block, give up and use the
408 // preincremented value.
409 unsigned NumUses = 0;
410 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
411 if (PN->getIncomingValue(i) == IV) {
413 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
417 // Okay, all uses of IV by PN are in predecessor blocks that really are
418 // dominated by the latch block. Split the critical edges and use the
419 // post-incremented value.
420 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
421 if (PN->getIncomingValue(i) == IV) {
422 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
423 // Splitting the critical edge can reduce the number of entries in this
425 e = PN->getNumIncomingValues();
426 if (--NumUses == 0) break;
429 // PHI node might have become a constant value after SplitCriticalEdge.
430 DeadInsts.insert(User);
437 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
438 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
439 /// return true. Otherwise, return false.
440 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
441 SmallPtrSet<Instruction*,16> &Processed) {
442 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
443 return false; // Void and FP expressions cannot be reduced.
444 if (!Processed.insert(I))
445 return true; // Instruction already handled.
447 // Get the symbolic expression for this instruction.
448 SCEVHandle ISE = GetExpressionSCEV(I);
449 if (isa<SCEVCouldNotCompute>(ISE)) return false;
451 // Get the start and stride for this expression.
452 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
453 SCEVHandle Stride = Start;
454 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
455 return false; // Non-reducible symbolic expression, bail out.
457 std::vector<Instruction *> IUsers;
458 // Collect all I uses now because IVUseShouldUsePostIncValue may
459 // invalidate use_iterator.
460 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
461 IUsers.push_back(cast<Instruction>(*UI));
463 for (unsigned iused_index = 0, iused_size = IUsers.size();
464 iused_index != iused_size; ++iused_index) {
466 Instruction *User = IUsers[iused_index];
468 // Do not infinitely recurse on PHI nodes.
469 if (isa<PHINode>(User) && Processed.count(User))
472 // If this is an instruction defined in a nested loop, or outside this loop,
473 // don't recurse into it.
474 bool AddUserToIVUsers = false;
475 if (LI->getLoopFor(User->getParent()) != L) {
476 DOUT << "FOUND USER in other loop: " << *User
477 << " OF SCEV: " << *ISE << "\n";
478 AddUserToIVUsers = true;
479 } else if (!AddUsersIfInteresting(User, L, Processed)) {
480 DOUT << "FOUND USER: " << *User
481 << " OF SCEV: " << *ISE << "\n";
482 AddUserToIVUsers = true;
485 if (AddUserToIVUsers) {
486 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
487 if (StrideUses.Users.empty()) // First occurance of this stride?
488 StrideOrder.push_back(Stride);
490 // Okay, we found a user that we cannot reduce. Analyze the instruction
491 // and decide what to do with it. If we are a use inside of the loop, use
492 // the value before incrementation, otherwise use it after incrementation.
493 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
494 // The value used will be incremented by the stride more than we are
495 // expecting, so subtract this off.
496 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
497 StrideUses.addUser(NewStart, User, I);
498 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
499 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
501 StrideUses.addUser(Start, User, I);
509 /// BasedUser - For a particular base value, keep information about how we've
510 /// partitioned the expression so far.
512 /// SE - The current ScalarEvolution object.
515 /// Base - The Base value for the PHI node that needs to be inserted for
516 /// this use. As the use is processed, information gets moved from this
517 /// field to the Imm field (below). BasedUser values are sorted by this
521 /// Inst - The instruction using the induction variable.
524 /// OperandValToReplace - The operand value of Inst to replace with the
526 Value *OperandValToReplace;
528 /// Imm - The immediate value that should be added to the base immediately
529 /// before Inst, because it will be folded into the imm field of the
533 /// EmittedBase - The actual value* to use for the base value of this
534 /// operation. This is null if we should just use zero so far.
537 // isUseOfPostIncrementedValue - True if this should use the
538 // post-incremented version of this IV, not the preincremented version.
539 // This can only be set in special cases, such as the terminating setcc
540 // instruction for a loop and uses outside the loop that are dominated by
542 bool isUseOfPostIncrementedValue;
544 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
545 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
546 OperandValToReplace(IVSU.OperandValToReplace),
547 Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
548 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
550 // Once we rewrite the code to insert the new IVs we want, update the
551 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
553 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
554 Instruction *InsertPt,
555 SCEVExpander &Rewriter, Loop *L, Pass *P,
556 SetVector<Instruction*> &DeadInsts);
558 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
559 SCEVExpander &Rewriter,
560 Instruction *IP, Loop *L);
565 void BasedUser::dump() const {
566 cerr << " Base=" << *Base;
567 cerr << " Imm=" << *Imm;
569 cerr << " EB=" << *EmittedBase;
571 cerr << " Inst: " << *Inst;
574 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
575 SCEVExpander &Rewriter,
576 Instruction *IP, Loop *L) {
577 // Figure out where we *really* want to insert this code. In particular, if
578 // the user is inside of a loop that is nested inside of L, we really don't
579 // want to insert this expression before the user, we'd rather pull it out as
580 // many loops as possible.
581 LoopInfo &LI = Rewriter.getLoopInfo();
582 Instruction *BaseInsertPt = IP;
584 // Figure out the most-nested loop that IP is in.
585 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
587 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
588 // the preheader of the outer-most loop where NewBase is not loop invariant.
589 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
590 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
591 InsertLoop = InsertLoop->getParentLoop();
594 // If there is no immediate value, skip the next part.
596 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
598 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
600 // If we are inserting the base and imm values in the same block, make sure to
601 // adjust the IP position if insertion reused a result.
602 if (IP == BaseInsertPt)
603 IP = Rewriter.getInsertionPoint();
605 // Always emit the immediate (if non-zero) into the same block as the user.
606 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
607 return Rewriter.expandCodeFor(NewValSCEV, IP);
612 // Once we rewrite the code to insert the new IVs we want, update the
613 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
614 // to it. NewBasePt is the last instruction which contributes to the
615 // value of NewBase in the case that it's a diffferent instruction from
616 // the PHI that NewBase is computed from, or null otherwise.
618 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
619 Instruction *NewBasePt,
620 SCEVExpander &Rewriter, Loop *L, Pass *P,
621 SetVector<Instruction*> &DeadInsts) {
622 if (!isa<PHINode>(Inst)) {
623 // By default, insert code at the user instruction.
624 BasicBlock::iterator InsertPt = Inst;
626 // However, if the Operand is itself an instruction, the (potentially
627 // complex) inserted code may be shared by many users. Because of this, we
628 // want to emit code for the computation of the operand right before its old
629 // computation. This is usually safe, because we obviously used to use the
630 // computation when it was computed in its current block. However, in some
631 // cases (e.g. use of a post-incremented induction variable) the NewBase
632 // value will be pinned to live somewhere after the original computation.
633 // In this case, we have to back off.
634 if (!isUseOfPostIncrementedValue) {
635 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
636 InsertPt = NewBasePt;
638 } else if (Instruction *OpInst
639 = dyn_cast<Instruction>(OperandValToReplace)) {
641 while (isa<PHINode>(InsertPt)) ++InsertPt;
644 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
645 // Adjust the type back to match the Inst. Note that we can't use InsertPt
646 // here because the SCEVExpander may have inserted the instructions after
647 // that point, in its efforts to avoid inserting redundant expressions.
648 if (isa<PointerType>(OperandValToReplace->getType())) {
649 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
651 OperandValToReplace->getType());
653 // Replace the use of the operand Value with the new Phi we just created.
654 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
655 DOUT << " CHANGED: IMM =" << *Imm;
656 DOUT << " \tNEWBASE =" << *NewBase;
657 DOUT << " \tInst = " << *Inst;
661 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
662 // expression into each operand block that uses it. Note that PHI nodes can
663 // have multiple entries for the same predecessor. We use a map to make sure
664 // that a PHI node only has a single Value* for each predecessor (which also
665 // prevents us from inserting duplicate code in some blocks).
666 DenseMap<BasicBlock*, Value*> InsertedCode;
667 PHINode *PN = cast<PHINode>(Inst);
668 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
669 if (PN->getIncomingValue(i) == OperandValToReplace) {
670 // If this is a critical edge, split the edge so that we do not insert the
671 // code on all predecessor/successor paths. We do this unless this is the
672 // canonical backedge for this loop, as this can make some inserted code
673 // be in an illegal position.
674 BasicBlock *PHIPred = PN->getIncomingBlock(i);
675 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
676 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
678 // First step, split the critical edge.
679 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
681 // Next step: move the basic block. In particular, if the PHI node
682 // is outside of the loop, and PredTI is in the loop, we want to
683 // move the block to be immediately before the PHI block, not
684 // immediately after PredTI.
685 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
686 BasicBlock *NewBB = PN->getIncomingBlock(i);
687 NewBB->moveBefore(PN->getParent());
690 // Splitting the edge can reduce the number of PHI entries we have.
691 e = PN->getNumIncomingValues();
694 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
696 // Insert the code into the end of the predecessor block.
697 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
698 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
700 // Adjust the type back to match the PHI. Note that we can't use
701 // InsertPt here because the SCEVExpander may have inserted its
702 // instructions after that point, in its efforts to avoid inserting
703 // redundant expressions.
704 if (isa<PointerType>(PN->getType())) {
705 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
711 // Replace the use of the operand Value with the new Phi we just created.
712 PN->setIncomingValue(i, Code);
717 // PHI node might have become a constant value after SplitCriticalEdge.
718 DeadInsts.insert(Inst);
720 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
724 /// isTargetConstant - Return true if the following can be referenced by the
725 /// immediate field of a target instruction.
726 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
727 const TargetLowering *TLI) {
728 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
729 int64_t VC = SC->getValue()->getSExtValue();
731 TargetLowering::AddrMode AM;
733 return TLI->isLegalAddressingMode(AM, UseTy);
735 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
736 return (VC > -(1 << 16) && VC < (1 << 16)-1);
740 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
741 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
742 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
743 Constant *Op0 = CE->getOperand(0);
744 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
745 TargetLowering::AddrMode AM;
747 return TLI->isLegalAddressingMode(AM, UseTy);
753 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
754 /// loop varying to the Imm operand.
755 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
756 Loop *L, ScalarEvolution *SE) {
757 if (Val->isLoopInvariant(L)) return; // Nothing to do.
759 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
760 std::vector<SCEVHandle> NewOps;
761 NewOps.reserve(SAE->getNumOperands());
763 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
764 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
765 // If this is a loop-variant expression, it must stay in the immediate
766 // field of the expression.
767 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
769 NewOps.push_back(SAE->getOperand(i));
773 Val = SE->getIntegerSCEV(0, Val->getType());
775 Val = SE->getAddExpr(NewOps);
776 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
777 // Try to pull immediates out of the start value of nested addrec's.
778 SCEVHandle Start = SARE->getStart();
779 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
781 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
783 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
785 // Otherwise, all of Val is variant, move the whole thing over.
786 Imm = SE->getAddExpr(Imm, Val);
787 Val = SE->getIntegerSCEV(0, Val->getType());
792 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
793 /// that can fit into the immediate field of instructions in the target.
794 /// Accumulate these immediate values into the Imm value.
795 static void MoveImmediateValues(const TargetLowering *TLI,
797 SCEVHandle &Val, SCEVHandle &Imm,
798 bool isAddress, Loop *L,
799 ScalarEvolution *SE) {
800 const Type *UseTy = User->getType();
801 if (StoreInst *SI = dyn_cast<StoreInst>(User))
802 UseTy = SI->getOperand(0)->getType();
804 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
805 std::vector<SCEVHandle> NewOps;
806 NewOps.reserve(SAE->getNumOperands());
808 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
809 SCEVHandle NewOp = SAE->getOperand(i);
810 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
812 if (!NewOp->isLoopInvariant(L)) {
813 // If this is a loop-variant expression, it must stay in the immediate
814 // field of the expression.
815 Imm = SE->getAddExpr(Imm, NewOp);
817 NewOps.push_back(NewOp);
822 Val = SE->getIntegerSCEV(0, Val->getType());
824 Val = SE->getAddExpr(NewOps);
826 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
827 // Try to pull immediates out of the start value of nested addrec's.
828 SCEVHandle Start = SARE->getStart();
829 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
831 if (Start != SARE->getStart()) {
832 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
834 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
837 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
838 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
839 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
840 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
842 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
843 SCEVHandle NewOp = SME->getOperand(1);
844 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
846 // If we extracted something out of the subexpressions, see if we can
848 if (NewOp != SME->getOperand(1)) {
849 // Scale SubImm up by "8". If the result is a target constant, we are
851 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
852 if (isTargetConstant(SubImm, UseTy, TLI)) {
853 // Accumulate the immediate.
854 Imm = SE->getAddExpr(Imm, SubImm);
856 // Update what is left of 'Val'.
857 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
864 // Loop-variant expressions must stay in the immediate field of the
866 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
867 !Val->isLoopInvariant(L)) {
868 Imm = SE->getAddExpr(Imm, Val);
869 Val = SE->getIntegerSCEV(0, Val->getType());
873 // Otherwise, no immediates to move.
877 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
878 /// added together. This is used to reassociate common addition subexprs
879 /// together for maximal sharing when rewriting bases.
880 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
882 ScalarEvolution *SE) {
883 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
884 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
885 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
886 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
887 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
888 if (SARE->getOperand(0) == Zero) {
889 SubExprs.push_back(Expr);
891 // Compute the addrec with zero as its base.
892 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
893 Ops[0] = Zero; // Start with zero base.
894 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
897 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
899 } else if (!Expr->isZero()) {
901 SubExprs.push_back(Expr);
906 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
907 /// removing any common subexpressions from it. Anything truly common is
908 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
909 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
911 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
912 ScalarEvolution *SE) {
913 unsigned NumUses = Uses.size();
915 // Only one use? Use its base, regardless of what it is!
916 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
917 SCEVHandle Result = Zero;
919 std::swap(Result, Uses[0].Base);
923 // To find common subexpressions, count how many of Uses use each expression.
924 // If any subexpressions are used Uses.size() times, they are common.
925 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
927 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
928 // order we see them.
929 std::vector<SCEVHandle> UniqueSubExprs;
931 std::vector<SCEVHandle> SubExprs;
932 for (unsigned i = 0; i != NumUses; ++i) {
933 // If the base is zero (which is common), return zero now, there are no
935 if (Uses[i].Base == Zero) return Zero;
937 // Split the expression into subexprs.
938 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
939 // Add one to SubExpressionUseCounts for each subexpr present.
940 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
941 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
942 UniqueSubExprs.push_back(SubExprs[j]);
946 // Now that we know how many times each is used, build Result. Iterate over
947 // UniqueSubexprs so that we have a stable ordering.
948 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
949 std::map<SCEVHandle, unsigned>::iterator I =
950 SubExpressionUseCounts.find(UniqueSubExprs[i]);
951 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
952 if (I->second == NumUses) { // Found CSE!
953 Result = SE->getAddExpr(Result, I->first);
955 // Remove non-cse's from SubExpressionUseCounts.
956 SubExpressionUseCounts.erase(I);
960 // If we found no CSE's, return now.
961 if (Result == Zero) return Result;
963 // Otherwise, remove all of the CSE's we found from each of the base values.
964 for (unsigned i = 0; i != NumUses; ++i) {
965 // Split the expression into subexprs.
966 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
968 // Remove any common subexpressions.
969 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
970 if (SubExpressionUseCounts.count(SubExprs[j])) {
971 SubExprs.erase(SubExprs.begin()+j);
975 // Finally, the non-shared expressions together.
976 if (SubExprs.empty())
979 Uses[i].Base = SE->getAddExpr(SubExprs);
986 /// ValidStride - Check whether the given Scale is valid for all loads and
987 /// stores in UsersToProcess.
989 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
991 const std::vector<BasedUser>& UsersToProcess) {
995 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
996 // If this is a load or other access, pass the type of the access in.
997 const Type *AccessTy = Type::VoidTy;
998 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
999 AccessTy = SI->getOperand(0)->getType();
1000 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1001 AccessTy = LI->getType();
1002 else if (isa<PHINode>(UsersToProcess[i].Inst))
1005 TargetLowering::AddrMode AM;
1006 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1007 AM.BaseOffs = SC->getValue()->getSExtValue();
1008 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1011 // If load[imm+r*scale] is illegal, bail out.
1012 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1018 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1020 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1024 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1026 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1027 !(isa<PointerType>(Ty2) &&
1028 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1029 !(isa<PointerType>(Ty1) &&
1030 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1033 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1034 /// of a previous stride and it is a legal value for the target addressing
1035 /// mode scale component and optional base reg. This allows the users of
1036 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1037 /// reuse is possible.
1038 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1039 bool AllUsesAreAddresses,
1040 const SCEVHandle &Stride,
1041 IVExpr &IV, const Type *Ty,
1042 const std::vector<BasedUser>& UsersToProcess) {
1043 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1044 int64_t SInt = SC->getValue()->getSExtValue();
1045 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1047 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1048 IVsByStride.find(StrideOrder[NewStride]);
1049 if (SI == IVsByStride.end())
1051 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1052 if (SI->first != Stride &&
1053 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1055 int64_t Scale = SInt / SSInt;
1056 // Check that this stride is valid for all the types used for loads and
1057 // stores; if it can be used for some and not others, we might as well use
1058 // the original stride everywhere, since we have to create the IV for it
1059 // anyway. If the scale is 1, then we don't need to worry about folding
1062 (AllUsesAreAddresses &&
1063 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1064 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1065 IE = SI->second.IVs.end(); II != IE; ++II)
1066 // FIXME: Only handle base == 0 for now.
1067 // Only reuse previous IV if it would not require a type conversion.
1068 if (II->Base->isZero() &&
1069 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1078 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1079 /// returns true if Val's isUseOfPostIncrementedValue is true.
1080 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1081 return Val.isUseOfPostIncrementedValue;
1084 /// isNonConstantNegative - Return true if the specified scev is negated, but
1086 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1087 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1088 if (!Mul) return false;
1090 // If there is a constant factor, it will be first.
1091 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1092 if (!SC) return false;
1094 // Return true if the value is negative, this matches things like (-42 * V).
1095 return SC->getValue()->getValue().isNegative();
1098 /// isAddress - Returns true if the specified instruction is using the
1099 /// specified value as an address.
1100 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
1101 bool isAddress = isa<LoadInst>(Inst);
1102 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1103 if (SI->getOperand(1) == OperandVal)
1105 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1106 // Addressing modes can also be folded into prefetches and a variety
1108 switch (II->getIntrinsicID()) {
1110 case Intrinsic::prefetch:
1111 case Intrinsic::x86_sse2_loadu_dq:
1112 case Intrinsic::x86_sse2_loadu_pd:
1113 case Intrinsic::x86_sse_loadu_ps:
1114 case Intrinsic::x86_sse_storeu_ps:
1115 case Intrinsic::x86_sse2_storeu_pd:
1116 case Intrinsic::x86_sse2_storeu_dq:
1117 case Intrinsic::x86_sse2_storel_dq:
1118 if (II->getOperand(1) == OperandVal)
1126 // CollectIVUsers - Transform our list of users and offsets to a bit more
1127 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1128 // of the strided accesses, as well as the old information from Uses. We
1129 // progressively move information from the Base field to the Imm field, until
1130 // we eventually have the full access expression to rewrite the use.
1131 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1132 IVUsersOfOneStride &Uses,
1134 bool &AllUsesAreAddresses,
1135 std::vector<BasedUser> &UsersToProcess) {
1136 UsersToProcess.reserve(Uses.Users.size());
1137 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1138 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1140 // Move any loop invariant operands from the offset field to the immediate
1141 // field of the use, so that we don't try to use something before it is
1143 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1144 UsersToProcess.back().Imm, L, SE);
1145 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1146 "Base value is not loop invariant!");
1149 // We now have a whole bunch of uses of like-strided induction variables, but
1150 // they might all have different bases. We want to emit one PHI node for this
1151 // stride which we fold as many common expressions (between the IVs) into as
1152 // possible. Start by identifying the common expressions in the base values
1153 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1154 // "A+B"), emit it to the preheader, then remove the expression from the
1155 // UsersToProcess base values.
1156 SCEVHandle CommonExprs =
1157 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
1159 // Next, figure out what we can represent in the immediate fields of
1160 // instructions. If we can represent anything there, move it to the imm
1161 // fields of the BasedUsers. We do this so that it increases the commonality
1162 // of the remaining uses.
1163 unsigned NumPHI = 0;
1164 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1165 // If the user is not in the current loop, this means it is using the exit
1166 // value of the IV. Do not put anything in the base, make sure it's all in
1167 // the immediate field to allow as much factoring as possible.
1168 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1169 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1170 UsersToProcess[i].Base);
1171 UsersToProcess[i].Base =
1172 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1175 // Addressing modes can be folded into loads and stores. Be careful that
1176 // the store is through the expression, not of the expression though.
1178 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1179 UsersToProcess[i].OperandValToReplace);
1180 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1185 // If this use isn't an address, then not all uses are addresses.
1186 if (!isAddress && !isPHI)
1187 AllUsesAreAddresses = false;
1189 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1190 UsersToProcess[i].Imm, isAddress, L, SE);
1194 // If one of the use if a PHI node and all other uses are addresses, still
1195 // allow iv reuse. Essentially we are trading one constant multiplication
1196 // for one fewer iv.
1198 AllUsesAreAddresses = false;
1203 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1204 /// stride of IV. All of the users may have different starting values, and this
1205 /// may not be the only stride (we know it is if isOnlyStride is true).
1206 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1207 IVUsersOfOneStride &Uses,
1209 bool isOnlyStride) {
1210 // If all the users are moved to another stride, then there is nothing to do.
1211 if (Uses.Users.empty())
1214 // Keep track if every use in UsersToProcess is an address. If they all are,
1215 // we may be able to rewrite the entire collection of them in terms of a
1216 // smaller-stride IV.
1217 bool AllUsesAreAddresses = true;
1219 // Transform our list of users and offsets to a bit more complex table. In
1220 // this new vector, each 'BasedUser' contains 'Base' the base of the
1221 // strided accessas well as the old information from Uses. We progressively
1222 // move information from the Base field to the Imm field, until we eventually
1223 // have the full access expression to rewrite the use.
1224 std::vector<BasedUser> UsersToProcess;
1225 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1228 // If we managed to find some expressions in common, we'll need to carry
1229 // their value in a register and add it in for each use. This will take up
1230 // a register operand, which potentially restricts what stride values are
1232 bool HaveCommonExprs = !CommonExprs->isZero();
1234 // If all uses are addresses, check if it is possible to reuse an IV with a
1235 // stride that is a factor of this stride. And that the multiple is a number
1236 // that can be encoded in the scale field of the target addressing mode. And
1237 // that we will have a valid instruction after this substition, including the
1238 // immediate field, if any.
1239 PHINode *NewPHI = NULL;
1241 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1242 SE->getIntegerSCEV(0, Type::Int32Ty),
1244 unsigned RewriteFactor = 0;
1245 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1246 Stride, ReuseIV, CommonExprs->getType(),
1248 if (RewriteFactor != 0) {
1249 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1250 << " and BASE " << *ReuseIV.Base << " :\n";
1251 NewPHI = ReuseIV.PHI;
1252 IncV = ReuseIV.IncV;
1255 const Type *ReplacedTy = CommonExprs->getType();
1257 // Now that we know what we need to do, insert the PHI node itself.
1259 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1260 << *Stride << " and BASE " << *CommonExprs << ": ";
1262 SCEVExpander Rewriter(*SE, *LI);
1263 SCEVExpander PreheaderRewriter(*SE, *LI);
1265 BasicBlock *Preheader = L->getLoopPreheader();
1266 Instruction *PreInsertPt = Preheader->getTerminator();
1267 Instruction *PhiInsertBefore = L->getHeader()->begin();
1269 BasicBlock *LatchBlock = L->getLoopLatch();
1272 // Emit the initial base value into the loop preheader.
1274 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1276 if (RewriteFactor == 0) {
1277 // Create a new Phi for this base, and stick it in the loop header.
1278 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1281 // Add common base to the new Phi node.
1282 NewPHI->addIncoming(CommonBaseV, Preheader);
1284 // If the stride is negative, insert a sub instead of an add for the
1286 bool isNegative = isNonConstantNegative(Stride);
1287 SCEVHandle IncAmount = Stride;
1289 IncAmount = SE->getNegativeSCEV(Stride);
1291 // Insert the stride into the preheader.
1292 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1293 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1295 // Emit the increment of the base value before the terminator of the loop
1296 // latch block, and add it to the Phi node.
1297 SCEVHandle IncExp = SE->getUnknown(StrideV);
1299 IncExp = SE->getNegativeSCEV(IncExp);
1300 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1302 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1303 IncV->setName(NewPHI->getName()+".inc");
1304 NewPHI->addIncoming(IncV, LatchBlock);
1306 // Remember this in case a later stride is multiple of this.
1307 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1309 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1311 Constant *C = dyn_cast<Constant>(CommonBaseV);
1313 (!C->isNullValue() &&
1314 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1315 // We want the common base emitted into the preheader! This is just
1316 // using cast as a copy so BitCast (no-op cast) is appropriate
1317 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1318 "commonbase", PreInsertPt);
1322 // We want to emit code for users inside the loop first. To do this, we
1323 // rearrange BasedUser so that the entries at the end have
1324 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1325 // vector (so we handle them first).
1326 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1327 PartitionByIsUseOfPostIncrementedValue);
1329 // Sort this by base, so that things with the same base are handled
1330 // together. By partitioning first and stable-sorting later, we are
1331 // guaranteed that within each base we will pop off users from within the
1332 // loop before users outside of the loop with a particular base.
1334 // We would like to use stable_sort here, but we can't. The problem is that
1335 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1336 // we don't have anything to do a '<' comparison on. Because we think the
1337 // number of uses is small, do a horrible bubble sort which just relies on
1339 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1340 // Get a base value.
1341 SCEVHandle Base = UsersToProcess[i].Base;
1343 // Compact everything with this base to be consequtive with this one.
1344 for (unsigned j = i+1; j != e; ++j) {
1345 if (UsersToProcess[j].Base == Base) {
1346 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1352 // Process all the users now. This outer loop handles all bases, the inner
1353 // loop handles all users of a particular base.
1354 while (!UsersToProcess.empty()) {
1355 SCEVHandle Base = UsersToProcess.back().Base;
1357 // Emit the code for Base into the preheader.
1358 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1360 DOUT << " INSERTING code for BASE = " << *Base << ":";
1361 if (BaseV->hasName())
1362 DOUT << " Result value name = %" << BaseV->getNameStr();
1365 // If BaseV is a constant other than 0, make sure that it gets inserted into
1366 // the preheader, instead of being forward substituted into the uses. We do
1367 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1369 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1370 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1371 // We want this constant emitted into the preheader! This is just
1372 // using cast as a copy so BitCast (no-op cast) is appropriate
1373 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1378 // Emit the code to add the immediate offset to the Phi value, just before
1379 // the instructions that we identified as using this stride and base.
1381 // FIXME: Use emitted users to emit other users.
1382 BasedUser &User = UsersToProcess.back();
1384 // If this instruction wants to use the post-incremented value, move it
1385 // after the post-inc and use its value instead of the PHI.
1386 Value *RewriteOp = NewPHI;
1387 if (User.isUseOfPostIncrementedValue) {
1390 // If this user is in the loop, make sure it is the last thing in the
1391 // loop to ensure it is dominated by the increment.
1392 if (L->contains(User.Inst->getParent()))
1393 User.Inst->moveBefore(LatchBlock->getTerminator());
1395 if (RewriteOp->getType() != ReplacedTy) {
1396 Instruction::CastOps opcode = Instruction::Trunc;
1397 if (ReplacedTy->getPrimitiveSizeInBits() ==
1398 RewriteOp->getType()->getPrimitiveSizeInBits())
1399 opcode = Instruction::BitCast;
1400 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1403 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1405 // If we had to insert new instrutions for RewriteOp, we have to
1406 // consider that they may not have been able to end up immediately
1407 // next to RewriteOp, because non-PHI instructions may never precede
1408 // PHI instructions in a block. In this case, remember where the last
1409 // instruction was inserted so that if we're replacing a different
1410 // PHI node, we can use the later point to expand the final
1412 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1413 if (RewriteOp == NewPHI) NewBasePt = 0;
1415 // Clear the SCEVExpander's expression map so that we are guaranteed
1416 // to have the code emitted where we expect it.
1419 // If we are reusing the iv, then it must be multiplied by a constant
1420 // factor take advantage of addressing mode scale component.
1421 if (RewriteFactor != 0) {
1422 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1423 RewriteExpr->getType()),
1426 // The common base is emitted in the loop preheader. But since we
1427 // are reusing an IV, it has not been used to initialize the PHI node.
1428 // Add it to the expression used to rewrite the uses.
1429 if (!isa<ConstantInt>(CommonBaseV) ||
1430 !cast<ConstantInt>(CommonBaseV)->isZero())
1431 RewriteExpr = SE->getAddExpr(RewriteExpr,
1432 SE->getUnknown(CommonBaseV));
1435 // Now that we know what we need to do, insert code before User for the
1436 // immediate and any loop-variant expressions.
1437 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1438 // Add BaseV to the PHI value if needed.
1439 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1441 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1445 // Mark old value we replaced as possibly dead, so that it is elminated
1446 // if we just replaced the last use of that value.
1447 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1449 UsersToProcess.pop_back();
1452 // If there are any more users to process with the same base, process them
1453 // now. We sorted by base above, so we just have to check the last elt.
1454 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1455 // TODO: Next, find out which base index is the most common, pull it out.
1458 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1459 // different starting values, into different PHIs.
1462 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1463 /// set the IV user and stride information and return true, otherwise return
1465 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1466 const SCEVHandle *&CondStride) {
1467 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1469 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1470 IVUsesByStride.find(StrideOrder[Stride]);
1471 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1473 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1474 E = SI->second.Users.end(); UI != E; ++UI)
1475 if (UI->User == Cond) {
1476 // NOTE: we could handle setcc instructions with multiple uses here, but
1477 // InstCombine does it as well for simple uses, it's not clear that it
1478 // occurs enough in real life to handle.
1480 CondStride = &SI->first;
1488 // Constant strides come first which in turns are sorted by their absolute
1489 // values. If absolute values are the same, then positive strides comes first.
1491 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1492 struct StrideCompare {
1493 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1494 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1495 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1497 int64_t LV = LHSC->getValue()->getSExtValue();
1498 int64_t RV = RHSC->getValue()->getSExtValue();
1499 uint64_t ALV = (LV < 0) ? -LV : LV;
1500 uint64_t ARV = (RV < 0) ? -RV : RV;
1506 return (LHSC && !RHSC);
1511 /// ChangeCompareStride - If a loop termination compare instruction is the
1512 /// only use of its stride, and the compaison is against a constant value,
1513 /// try eliminate the stride by moving the compare instruction to another
1514 /// stride and change its constant operand accordingly. e.g.
1520 /// if (v2 < 10) goto loop
1525 /// if (v1 < 30) goto loop
1526 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1527 IVStrideUse* &CondUse,
1528 const SCEVHandle* &CondStride) {
1529 if (StrideOrder.size() < 2 ||
1530 IVUsesByStride[*CondStride].Users.size() != 1)
1532 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1533 if (!SC) return Cond;
1534 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1535 if (!C) return Cond;
1537 ICmpInst::Predicate Predicate = Cond->getPredicate();
1538 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1539 int64_t CmpVal = C->getValue().getSExtValue();
1540 unsigned BitWidth = C->getValue().getBitWidth();
1541 uint64_t SignBit = 1ULL << (BitWidth-1);
1542 const Type *CmpTy = C->getType();
1543 const Type *NewCmpTy = NULL;
1544 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1545 unsigned NewTyBits = 0;
1546 int64_t NewCmpVal = CmpVal;
1547 SCEVHandle *NewStride = NULL;
1548 Value *NewIncV = NULL;
1551 // Check stride constant and the comparision constant signs to detect
1553 if (ICmpInst::isSignedPredicate(Predicate) &&
1554 (CmpVal & SignBit) != (CmpSSInt & SignBit))
1557 // Look for a suitable stride / iv as replacement.
1558 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1559 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1560 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1561 IVUsesByStride.find(StrideOrder[i]);
1562 if (!isa<SCEVConstant>(SI->first))
1564 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1565 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1568 Scale = SSInt / CmpSSInt;
1569 NewCmpVal = CmpVal * Scale;
1570 APInt Mul = APInt(BitWidth, NewCmpVal);
1571 // Check for overflow.
1572 if (Mul.getSExtValue() != NewCmpVal) {
1577 // Watch out for overflow.
1578 if (ICmpInst::isSignedPredicate(Predicate) &&
1579 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1582 if (NewCmpVal != CmpVal) {
1583 // Pick the best iv to use trying to avoid a cast.
1585 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1586 E = SI->second.Users.end(); UI != E; ++UI) {
1587 NewIncV = UI->OperandValToReplace;
1588 if (NewIncV->getType() == CmpTy)
1596 NewCmpTy = NewIncV->getType();
1597 NewTyBits = isa<PointerType>(NewCmpTy)
1598 ? UIntPtrTy->getPrimitiveSizeInBits()
1599 : NewCmpTy->getPrimitiveSizeInBits();
1600 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1601 // Check if it is possible to rewrite it using
1602 // an iv / stride of a smaller integer type.
1603 bool TruncOk = false;
1604 if (NewCmpTy->isInteger()) {
1605 unsigned Bits = NewTyBits;
1606 if (ICmpInst::isSignedPredicate(Predicate))
1608 uint64_t Mask = (1ULL << Bits) - 1;
1609 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1618 // Don't rewrite if use offset is non-constant and the new type is
1619 // of a different type.
1620 // FIXME: too conservative?
1621 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1626 bool AllUsesAreAddresses = true;
1627 std::vector<BasedUser> UsersToProcess;
1628 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1629 AllUsesAreAddresses,
1631 // Avoid rewriting the compare instruction with an iv of new stride
1632 // if it's likely the new stride uses will be rewritten using the
1633 if (AllUsesAreAddresses &&
1634 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1639 // If scale is negative, use swapped predicate unless it's testing
1641 if (Scale < 0 && !Cond->isEquality())
1642 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1644 NewStride = &StrideOrder[i];
1649 // Forgo this transformation if it the increment happens to be
1650 // unfortunately positioned after the condition, and the condition
1651 // has multiple uses which prevent it from being moved immediately
1652 // before the branch. See
1653 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1654 // for an example of this situation.
1655 if (!Cond->hasOneUse()) {
1656 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1662 if (NewCmpVal != CmpVal) {
1663 // Create a new compare instruction using new stride / iv.
1664 ICmpInst *OldCond = Cond;
1666 if (!isa<PointerType>(NewCmpTy))
1667 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1669 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1670 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1672 // Insert new compare instruction.
1673 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1674 L->getHeader()->getName() + ".termcond",
1677 // Remove the old compare instruction. The old indvar is probably dead too.
1678 DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
1679 SE->deleteValueFromRecords(OldCond);
1680 OldCond->replaceAllUsesWith(Cond);
1681 OldCond->eraseFromParent();
1683 IVUsesByStride[*CondStride].Users.pop_back();
1684 SCEVHandle NewOffset = TyBits == NewTyBits
1685 ? SE->getMulExpr(CondUse->Offset,
1686 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1687 : SE->getConstant(ConstantInt::get(NewCmpTy,
1688 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1689 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1690 CondUse = &IVUsesByStride[*NewStride].Users.back();
1691 CondStride = NewStride;
1698 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1699 /// inside the loop then try to eliminate the cast opeation.
1700 void LoopStrengthReduce::OptimizeShadowIV(Loop *L, ICmpInst *Cond,
1701 const SCEVHandle *&CondStride) {
1703 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1706 SCEVHandle IterationCount = SE->getIterationCount(L);
1707 if (isa<SCEVCouldNotCompute>(IterationCount))
1710 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
1712 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1713 IVUsesByStride.find(StrideOrder[Stride]);
1714 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1716 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1717 E = SI->second.Users.end(); UI != E; /* empty */) {
1718 std::vector<IVStrideUse>::iterator CandidateUI = UI;
1720 Instruction *ShadowUse = CandidateUI->User;
1721 const Type *DestTy = NULL;
1723 /* If shadow use is a int->float cast then insert a second IV
1724 to elminate this cast.
1726 for (unsigned i = 0; i < n; ++i)
1732 for (unsigned i = 0; i < n; ++i, ++d)
1735 UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User);
1737 DestTy = UCast->getDestTy();
1739 SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User);
1740 if (!SCast) continue;
1741 DestTy = SCast->getDestTy();
1744 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1746 if (PH->getNumIncomingValues() != 2) continue;
1748 const Type *SrcTy = PH->getType();
1749 int Mantissa = DestTy->getFPMantissaWidth();
1750 if (Mantissa == -1) continue;
1751 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
1754 unsigned Entry, Latch;
1755 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1763 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1764 if (!Init) continue;
1765 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1767 BinaryOperator *Incr =
1768 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1769 if (!Incr) continue;
1770 if (Incr->getOpcode() != Instruction::Add
1771 && Incr->getOpcode() != Instruction::Sub)
1774 /* Initialize new IV, double d = 0.0 in above example. */
1775 ConstantInt *C = NULL;
1776 if (Incr->getOperand(0) == PH)
1777 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1778 else if (Incr->getOperand(1) == PH)
1779 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1785 /* Add new PHINode. */
1786 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1788 /* create new icnrement. '++d' in above example. */
1789 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1790 BinaryOperator *NewIncr =
1791 BinaryOperator::Create(Incr->getOpcode(),
1792 NewPH, CFP, "IV.S.next.", Incr);
1794 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1795 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1797 /* Remove cast operation */
1798 ShadowUse->replaceAllUsesWith(NewPH);
1799 ShadowUse->eraseFromParent();
1800 SI->second.Users.erase(CandidateUI);
1807 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1808 // uses in the loop, look to see if we can eliminate some, in favor of using
1809 // common indvars for the different uses.
1810 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1811 // TODO: implement optzns here.
1813 // Finally, get the terminating condition for the loop if possible. If we
1814 // can, we want to change it to use a post-incremented version of its
1815 // induction variable, to allow coalescing the live ranges for the IV into
1816 // one register value.
1817 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1818 BasicBlock *Preheader = L->getLoopPreheader();
1819 BasicBlock *LatchBlock =
1820 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1821 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1822 if (!TermBr || TermBr->isUnconditional() ||
1823 !isa<ICmpInst>(TermBr->getCondition()))
1825 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1827 // Search IVUsesByStride to find Cond's IVUse if there is one.
1828 IVStrideUse *CondUse = 0;
1829 const SCEVHandle *CondStride = 0;
1831 if (!FindIVUserForCond(Cond, CondUse, CondStride))
1832 return; // setcc doesn't use the IV.
1834 OptimizeShadowIV(L, Cond, CondStride);
1836 // If possible, change stride and operands of the compare instruction to
1837 // eliminate one stride.
1838 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
1840 // It's possible for the setcc instruction to be anywhere in the loop, and
1841 // possible for it to have multiple users. If it is not immediately before
1842 // the latch block branch, move it.
1843 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1844 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1845 Cond->moveBefore(TermBr);
1847 // Otherwise, clone the terminating condition and insert into the loopend.
1848 Cond = cast<ICmpInst>(Cond->clone());
1849 Cond->setName(L->getHeader()->getName() + ".termcond");
1850 LatchBlock->getInstList().insert(TermBr, Cond);
1852 // Clone the IVUse, as the old use still exists!
1853 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1854 CondUse->OperandValToReplace);
1855 CondUse = &IVUsesByStride[*CondStride].Users.back();
1859 // If we get to here, we know that we can transform the setcc instruction to
1860 // use the post-incremented version of the IV, allowing us to coalesce the
1861 // live ranges for the IV correctly.
1862 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
1863 CondUse->isUseOfPostIncrementedValue = true;
1867 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1869 LI = &getAnalysis<LoopInfo>();
1870 DT = &getAnalysis<DominatorTree>();
1871 SE = &getAnalysis<ScalarEvolution>();
1872 TD = &getAnalysis<TargetData>();
1873 UIntPtrTy = TD->getIntPtrType();
1876 // Find all uses of induction variables in this loop, and catagorize
1877 // them by stride. Start by finding all of the PHI nodes in the header for
1878 // this loop. If they are induction variables, inspect their uses.
1879 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
1880 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1881 AddUsersIfInteresting(I, L, Processed);
1883 if (!IVUsesByStride.empty()) {
1884 // Optimize induction variables. Some indvar uses can be transformed to use
1885 // strides that will be needed for other purposes. A common example of this
1886 // is the exit test for the loop, which can often be rewritten to use the
1887 // computation of some other indvar to decide when to terminate the loop.
1890 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1891 // doing computation in byte values, promote to 32-bit values if safe.
1893 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1894 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
1895 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
1896 // Need to be careful that IV's are all the same type. Only works for
1897 // intptr_t indvars.
1899 // If we only have one stride, we can more aggressively eliminate some
1901 bool HasOneStride = IVUsesByStride.size() == 1;
1904 DOUT << "\nLSR on ";
1908 // IVsByStride keeps IVs for one particular loop.
1909 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
1911 // Sort the StrideOrder so we process larger strides first.
1912 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1914 // Note: this processes each stride/type pair individually. All users
1915 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
1916 // Also, note that we iterate over IVUsesByStride indirectly by using
1917 // StrideOrder. This extra layer of indirection makes the ordering of
1918 // strides deterministic - not dependent on map order.
1919 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1920 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1921 IVUsesByStride.find(StrideOrder[Stride]);
1922 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1923 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1927 // We're done analyzing this loop; release all the state we built up for it.
1928 CastedPointers.clear();
1929 IVUsesByStride.clear();
1930 IVsByStride.clear();
1931 StrideOrder.clear();
1933 // Clean up after ourselves
1934 if (!DeadInsts.empty()) {
1935 DeleteTriviallyDeadInstructions(DeadInsts);
1937 BasicBlock::iterator I = L->getHeader()->begin();
1938 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
1939 // At this point, we know that we have killed one or more IV users.
1940 // It is worth checking to see if the cann indvar is also
1941 // dead, so that we can remove it as well.
1943 // We can remove a PHI if it is on a cycle in the def-use graph
1944 // where each node in the cycle has degree one, i.e. only one use,
1945 // and is an instruction with no side effects.
1947 // FIXME: this needs to eliminate an induction variable even if it's being
1948 // compared against some value to decide loop termination.
1949 if (PN->hasOneUse()) {
1950 SmallPtrSet<PHINode *, 2> PHIs;
1951 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
1952 J && J->hasOneUse() && !J->mayWriteToMemory();
1953 J = dyn_cast<Instruction>(*J->use_begin())) {
1954 // If we find the original PHI, we've discovered a cycle.
1956 // Break the cycle and mark the PHI for deletion.
1957 SE->deleteValueFromRecords(PN);
1958 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1959 DeadInsts.insert(PN);
1963 // If we find a PHI more than once, we're on a cycle that
1964 // won't prove fruitful.
1965 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
1970 DeleteTriviallyDeadInstructions(DeadInsts);