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 /// OptimizeSMax - Rewrite the loop's terminating condition
188 /// if it uses an smax computation.
189 ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
190 IVStrideUse* &CondUse);
192 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
193 const SCEVHandle *&CondStride);
194 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
195 unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
196 IVExpr&, const Type*,
197 const std::vector<BasedUser>& UsersToProcess);
198 bool ValidStride(bool, int64_t,
199 const std::vector<BasedUser>& UsersToProcess);
200 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
201 IVUsersOfOneStride &Uses,
203 bool &AllUsesAreAddresses,
204 std::vector<BasedUser> &UsersToProcess);
205 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
206 IVUsersOfOneStride &Uses,
207 Loop *L, bool isOnlyStride);
208 void DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts);
212 char LoopStrengthReduce::ID = 0;
213 static RegisterPass<LoopStrengthReduce>
214 X("loop-reduce", "Loop Strength Reduction");
216 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
217 return new LoopStrengthReduce(TLI);
220 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
221 /// assumes that the Value* V is of integer or pointer type only.
223 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
225 if (V->getType() == UIntPtrTy) return V;
226 if (Constant *CB = dyn_cast<Constant>(V))
227 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
229 Value *&New = CastedPointers[V];
232 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
233 DeadInsts.insert(cast<Instruction>(New));
238 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
239 /// specified set are trivially dead, delete them and see if this makes any of
240 /// their operands subsequently dead.
241 void LoopStrengthReduce::
242 DeleteTriviallyDeadInstructions(SetVector<Instruction*> &Insts) {
243 while (!Insts.empty()) {
244 Instruction *I = Insts.back();
247 if (PHINode *PN = dyn_cast<PHINode>(I)) {
248 // If all incoming values to the Phi are the same, we can replace the Phi
250 if (Value *PNV = PN->hasConstantValue()) {
251 if (Instruction *U = dyn_cast<Instruction>(PNV))
253 SE->deleteValueFromRecords(PN);
254 PN->replaceAllUsesWith(PNV);
255 PN->eraseFromParent();
261 if (isInstructionTriviallyDead(I)) {
262 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
263 if (Instruction *U = dyn_cast<Instruction>(*i))
265 SE->deleteValueFromRecords(I);
266 I->eraseFromParent();
273 /// GetExpressionSCEV - Compute and return the SCEV for the specified
275 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
276 // Pointer to pointer bitcast instructions return the same value as their
278 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
279 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
280 return SE->getSCEV(BCI);
281 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
286 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
287 // If this is a GEP that SE doesn't know about, compute it now and insert it.
288 // If this is not a GEP, or if we have already done this computation, just let
290 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
291 if (!GEP || SE->hasSCEV(GEP))
292 return SE->getSCEV(Exp);
294 // Analyze all of the subscripts of this getelementptr instruction, looking
295 // for uses that are determined by the trip count of the loop. First, skip
296 // all operands the are not dependent on the IV.
298 // Build up the base expression. Insert an LLVM cast of the pointer to
300 SCEVHandle GEPVal = SE->getUnknown(
301 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
303 gep_type_iterator GTI = gep_type_begin(GEP);
305 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
306 i != e; ++i, ++GTI) {
307 // If this is a use of a recurrence that we can analyze, and it comes before
308 // Op does in the GEP operand list, we will handle this when we process this
310 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
311 const StructLayout *SL = TD->getStructLayout(STy);
312 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
313 uint64_t Offset = SL->getElementOffset(Idx);
314 GEPVal = SE->getAddExpr(GEPVal,
315 SE->getIntegerSCEV(Offset, UIntPtrTy));
317 unsigned GEPOpiBits =
318 (*i)->getType()->getPrimitiveSizeInBits();
319 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
320 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
321 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
322 Instruction::BitCast));
323 Value *OpVal = getCastedVersionOf(opcode, *i);
324 SCEVHandle Idx = SE->getSCEV(OpVal);
326 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
328 Idx = SE->getMulExpr(Idx,
329 SE->getConstant(ConstantInt::get(UIntPtrTy,
331 GEPVal = SE->getAddExpr(GEPVal, Idx);
335 SE->setSCEV(GEP, GEPVal);
339 /// getSCEVStartAndStride - Compute the start and stride of this expression,
340 /// returning false if the expression is not a start/stride pair, or true if it
341 /// is. The stride must be a loop invariant expression, but the start may be
342 /// a mix of loop invariant and loop variant expressions.
343 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
344 SCEVHandle &Start, SCEVHandle &Stride,
345 ScalarEvolution *SE) {
346 SCEVHandle TheAddRec = Start; // Initialize to zero.
348 // If the outer level is an AddExpr, the operands are all start values except
349 // for a nested AddRecExpr.
350 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
351 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
352 if (SCEVAddRecExpr *AddRec =
353 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
354 if (AddRec->getLoop() == L)
355 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
357 return false; // Nested IV of some sort?
359 Start = SE->getAddExpr(Start, AE->getOperand(i));
362 } else if (isa<SCEVAddRecExpr>(SH)) {
365 return false; // not analyzable.
368 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
369 if (!AddRec || AddRec->getLoop() != L) return false;
371 // FIXME: Generalize to non-affine IV's.
372 if (!AddRec->isAffine()) return false;
374 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
376 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
377 DOUT << "[" << L->getHeader()->getName()
378 << "] Variable stride: " << *AddRec << "\n";
380 Stride = AddRec->getOperand(1);
384 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
385 /// and now we need to decide whether the user should use the preinc or post-inc
386 /// value. If this user should use the post-inc version of the IV, return true.
388 /// Choosing wrong here can break dominance properties (if we choose to use the
389 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
390 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
391 /// should use the post-inc value).
392 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
393 Loop *L, DominatorTree *DT, Pass *P,
394 SetVector<Instruction*> &DeadInsts){
395 // If the user is in the loop, use the preinc value.
396 if (L->contains(User->getParent())) return false;
398 BasicBlock *LatchBlock = L->getLoopLatch();
400 // Ok, the user is outside of the loop. If it is dominated by the latch
401 // block, use the post-inc value.
402 if (DT->dominates(LatchBlock, User->getParent()))
405 // There is one case we have to be careful of: PHI nodes. These little guys
406 // can live in blocks that do not dominate the latch block, but (since their
407 // uses occur in the predecessor block, not the block the PHI lives in) should
408 // still use the post-inc value. Check for this case now.
409 PHINode *PN = dyn_cast<PHINode>(User);
410 if (!PN) return false; // not a phi, not dominated by latch block.
412 // Look at all of the uses of IV by the PHI node. If any use corresponds to
413 // a block that is not dominated by the latch block, give up and use the
414 // preincremented value.
415 unsigned NumUses = 0;
416 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
417 if (PN->getIncomingValue(i) == IV) {
419 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
423 // Okay, all uses of IV by PN are in predecessor blocks that really are
424 // dominated by the latch block. Split the critical edges and use the
425 // post-incremented value.
426 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
427 if (PN->getIncomingValue(i) == IV) {
428 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
429 // Splitting the critical edge can reduce the number of entries in this
431 e = PN->getNumIncomingValues();
432 if (--NumUses == 0) break;
435 // PHI node might have become a constant value after SplitCriticalEdge.
436 DeadInsts.insert(User);
443 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
444 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
445 /// return true. Otherwise, return false.
446 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
447 SmallPtrSet<Instruction*,16> &Processed) {
448 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
449 return false; // Void and FP expressions cannot be reduced.
450 if (!Processed.insert(I))
451 return true; // Instruction already handled.
453 // Get the symbolic expression for this instruction.
454 SCEVHandle ISE = GetExpressionSCEV(I);
455 if (isa<SCEVCouldNotCompute>(ISE)) return false;
457 // Get the start and stride for this expression.
458 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
459 SCEVHandle Stride = Start;
460 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
461 return false; // Non-reducible symbolic expression, bail out.
463 std::vector<Instruction *> IUsers;
464 // Collect all I uses now because IVUseShouldUsePostIncValue may
465 // invalidate use_iterator.
466 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
467 IUsers.push_back(cast<Instruction>(*UI));
469 for (unsigned iused_index = 0, iused_size = IUsers.size();
470 iused_index != iused_size; ++iused_index) {
472 Instruction *User = IUsers[iused_index];
474 // Do not infinitely recurse on PHI nodes.
475 if (isa<PHINode>(User) && Processed.count(User))
478 // If this is an instruction defined in a nested loop, or outside this loop,
479 // don't recurse into it.
480 bool AddUserToIVUsers = false;
481 if (LI->getLoopFor(User->getParent()) != L) {
482 DOUT << "FOUND USER in other loop: " << *User
483 << " OF SCEV: " << *ISE << "\n";
484 AddUserToIVUsers = true;
485 } else if (!AddUsersIfInteresting(User, L, Processed)) {
486 DOUT << "FOUND USER: " << *User
487 << " OF SCEV: " << *ISE << "\n";
488 AddUserToIVUsers = true;
491 if (AddUserToIVUsers) {
492 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
493 if (StrideUses.Users.empty()) // First occurance of this stride?
494 StrideOrder.push_back(Stride);
496 // Okay, we found a user that we cannot reduce. Analyze the instruction
497 // and decide what to do with it. If we are a use inside of the loop, use
498 // the value before incrementation, otherwise use it after incrementation.
499 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
500 // The value used will be incremented by the stride more than we are
501 // expecting, so subtract this off.
502 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
503 StrideUses.addUser(NewStart, User, I);
504 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
505 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
507 StrideUses.addUser(Start, User, I);
515 /// BasedUser - For a particular base value, keep information about how we've
516 /// partitioned the expression so far.
518 /// SE - The current ScalarEvolution object.
521 /// Base - The Base value for the PHI node that needs to be inserted for
522 /// this use. As the use is processed, information gets moved from this
523 /// field to the Imm field (below). BasedUser values are sorted by this
527 /// Inst - The instruction using the induction variable.
530 /// OperandValToReplace - The operand value of Inst to replace with the
532 Value *OperandValToReplace;
534 /// Imm - The immediate value that should be added to the base immediately
535 /// before Inst, because it will be folded into the imm field of the
539 /// EmittedBase - The actual value* to use for the base value of this
540 /// operation. This is null if we should just use zero so far.
543 // isUseOfPostIncrementedValue - True if this should use the
544 // post-incremented version of this IV, not the preincremented version.
545 // This can only be set in special cases, such as the terminating setcc
546 // instruction for a loop and uses outside the loop that are dominated by
548 bool isUseOfPostIncrementedValue;
550 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
551 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
552 OperandValToReplace(IVSU.OperandValToReplace),
553 Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
554 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
556 // Once we rewrite the code to insert the new IVs we want, update the
557 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
559 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
560 Instruction *InsertPt,
561 SCEVExpander &Rewriter, Loop *L, Pass *P,
562 SetVector<Instruction*> &DeadInsts);
564 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
565 SCEVExpander &Rewriter,
566 Instruction *IP, Loop *L);
571 void BasedUser::dump() const {
572 cerr << " Base=" << *Base;
573 cerr << " Imm=" << *Imm;
575 cerr << " EB=" << *EmittedBase;
577 cerr << " Inst: " << *Inst;
580 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
581 SCEVExpander &Rewriter,
582 Instruction *IP, Loop *L) {
583 // Figure out where we *really* want to insert this code. In particular, if
584 // the user is inside of a loop that is nested inside of L, we really don't
585 // want to insert this expression before the user, we'd rather pull it out as
586 // many loops as possible.
587 LoopInfo &LI = Rewriter.getLoopInfo();
588 Instruction *BaseInsertPt = IP;
590 // Figure out the most-nested loop that IP is in.
591 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
593 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
594 // the preheader of the outer-most loop where NewBase is not loop invariant.
595 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
596 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
597 InsertLoop = InsertLoop->getParentLoop();
600 // If there is no immediate value, skip the next part.
602 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
604 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
606 // If we are inserting the base and imm values in the same block, make sure to
607 // adjust the IP position if insertion reused a result.
608 if (IP == BaseInsertPt)
609 IP = Rewriter.getInsertionPoint();
611 // Always emit the immediate (if non-zero) into the same block as the user.
612 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
613 return Rewriter.expandCodeFor(NewValSCEV, IP);
618 // Once we rewrite the code to insert the new IVs we want, update the
619 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
620 // to it. NewBasePt is the last instruction which contributes to the
621 // value of NewBase in the case that it's a diffferent instruction from
622 // the PHI that NewBase is computed from, or null otherwise.
624 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
625 Instruction *NewBasePt,
626 SCEVExpander &Rewriter, Loop *L, Pass *P,
627 SetVector<Instruction*> &DeadInsts) {
628 if (!isa<PHINode>(Inst)) {
629 // By default, insert code at the user instruction.
630 BasicBlock::iterator InsertPt = Inst;
632 // However, if the Operand is itself an instruction, the (potentially
633 // complex) inserted code may be shared by many users. Because of this, we
634 // want to emit code for the computation of the operand right before its old
635 // computation. This is usually safe, because we obviously used to use the
636 // computation when it was computed in its current block. However, in some
637 // cases (e.g. use of a post-incremented induction variable) the NewBase
638 // value will be pinned to live somewhere after the original computation.
639 // In this case, we have to back off.
640 if (!isUseOfPostIncrementedValue) {
641 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
642 InsertPt = NewBasePt;
644 } else if (Instruction *OpInst
645 = dyn_cast<Instruction>(OperandValToReplace)) {
647 while (isa<PHINode>(InsertPt)) ++InsertPt;
650 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
651 // Adjust the type back to match the Inst. Note that we can't use InsertPt
652 // here because the SCEVExpander may have inserted the instructions after
653 // that point, in its efforts to avoid inserting redundant expressions.
654 if (isa<PointerType>(OperandValToReplace->getType())) {
655 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
657 OperandValToReplace->getType());
659 // Replace the use of the operand Value with the new Phi we just created.
660 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
661 DOUT << " CHANGED: IMM =" << *Imm;
662 DOUT << " \tNEWBASE =" << *NewBase;
663 DOUT << " \tInst = " << *Inst;
667 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
668 // expression into each operand block that uses it. Note that PHI nodes can
669 // have multiple entries for the same predecessor. We use a map to make sure
670 // that a PHI node only has a single Value* for each predecessor (which also
671 // prevents us from inserting duplicate code in some blocks).
672 DenseMap<BasicBlock*, Value*> InsertedCode;
673 PHINode *PN = cast<PHINode>(Inst);
674 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
675 if (PN->getIncomingValue(i) == OperandValToReplace) {
676 // If this is a critical edge, split the edge so that we do not insert the
677 // code on all predecessor/successor paths. We do this unless this is the
678 // canonical backedge for this loop, as this can make some inserted code
679 // be in an illegal position.
680 BasicBlock *PHIPred = PN->getIncomingBlock(i);
681 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
682 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
684 // First step, split the critical edge.
685 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
687 // Next step: move the basic block. In particular, if the PHI node
688 // is outside of the loop, and PredTI is in the loop, we want to
689 // move the block to be immediately before the PHI block, not
690 // immediately after PredTI.
691 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
692 BasicBlock *NewBB = PN->getIncomingBlock(i);
693 NewBB->moveBefore(PN->getParent());
696 // Splitting the edge can reduce the number of PHI entries we have.
697 e = PN->getNumIncomingValues();
700 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
702 // Insert the code into the end of the predecessor block.
703 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
704 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
706 // Adjust the type back to match the PHI. Note that we can't use
707 // InsertPt here because the SCEVExpander may have inserted its
708 // instructions after that point, in its efforts to avoid inserting
709 // redundant expressions.
710 if (isa<PointerType>(PN->getType())) {
711 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
717 // Replace the use of the operand Value with the new Phi we just created.
718 PN->setIncomingValue(i, Code);
723 // PHI node might have become a constant value after SplitCriticalEdge.
724 DeadInsts.insert(Inst);
726 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
730 /// isTargetConstant - Return true if the following can be referenced by the
731 /// immediate field of a target instruction.
732 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
733 const TargetLowering *TLI) {
734 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
735 int64_t VC = SC->getValue()->getSExtValue();
737 TargetLowering::AddrMode AM;
739 return TLI->isLegalAddressingMode(AM, UseTy);
741 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
742 return (VC > -(1 << 16) && VC < (1 << 16)-1);
746 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
747 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
748 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
749 Constant *Op0 = CE->getOperand(0);
750 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
751 TargetLowering::AddrMode AM;
753 return TLI->isLegalAddressingMode(AM, UseTy);
759 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
760 /// loop varying to the Imm operand.
761 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
762 Loop *L, ScalarEvolution *SE) {
763 if (Val->isLoopInvariant(L)) return; // Nothing to do.
765 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
766 std::vector<SCEVHandle> NewOps;
767 NewOps.reserve(SAE->getNumOperands());
769 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
770 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
771 // If this is a loop-variant expression, it must stay in the immediate
772 // field of the expression.
773 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
775 NewOps.push_back(SAE->getOperand(i));
779 Val = SE->getIntegerSCEV(0, Val->getType());
781 Val = SE->getAddExpr(NewOps);
782 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
783 // Try to pull immediates out of the start value of nested addrec's.
784 SCEVHandle Start = SARE->getStart();
785 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
787 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
789 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
791 // Otherwise, all of Val is variant, move the whole thing over.
792 Imm = SE->getAddExpr(Imm, Val);
793 Val = SE->getIntegerSCEV(0, Val->getType());
798 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
799 /// that can fit into the immediate field of instructions in the target.
800 /// Accumulate these immediate values into the Imm value.
801 static void MoveImmediateValues(const TargetLowering *TLI,
803 SCEVHandle &Val, SCEVHandle &Imm,
804 bool isAddress, Loop *L,
805 ScalarEvolution *SE) {
806 const Type *UseTy = User->getType();
807 if (StoreInst *SI = dyn_cast<StoreInst>(User))
808 UseTy = SI->getOperand(0)->getType();
810 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
811 std::vector<SCEVHandle> NewOps;
812 NewOps.reserve(SAE->getNumOperands());
814 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
815 SCEVHandle NewOp = SAE->getOperand(i);
816 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
818 if (!NewOp->isLoopInvariant(L)) {
819 // If this is a loop-variant expression, it must stay in the immediate
820 // field of the expression.
821 Imm = SE->getAddExpr(Imm, NewOp);
823 NewOps.push_back(NewOp);
828 Val = SE->getIntegerSCEV(0, Val->getType());
830 Val = SE->getAddExpr(NewOps);
832 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
833 // Try to pull immediates out of the start value of nested addrec's.
834 SCEVHandle Start = SARE->getStart();
835 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
837 if (Start != SARE->getStart()) {
838 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
840 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
843 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
844 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
845 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
846 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
848 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
849 SCEVHandle NewOp = SME->getOperand(1);
850 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
852 // If we extracted something out of the subexpressions, see if we can
854 if (NewOp != SME->getOperand(1)) {
855 // Scale SubImm up by "8". If the result is a target constant, we are
857 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
858 if (isTargetConstant(SubImm, UseTy, TLI)) {
859 // Accumulate the immediate.
860 Imm = SE->getAddExpr(Imm, SubImm);
862 // Update what is left of 'Val'.
863 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
870 // Loop-variant expressions must stay in the immediate field of the
872 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
873 !Val->isLoopInvariant(L)) {
874 Imm = SE->getAddExpr(Imm, Val);
875 Val = SE->getIntegerSCEV(0, Val->getType());
879 // Otherwise, no immediates to move.
883 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
884 /// added together. This is used to reassociate common addition subexprs
885 /// together for maximal sharing when rewriting bases.
886 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
888 ScalarEvolution *SE) {
889 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
890 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
891 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
892 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
893 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
894 if (SARE->getOperand(0) == Zero) {
895 SubExprs.push_back(Expr);
897 // Compute the addrec with zero as its base.
898 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
899 Ops[0] = Zero; // Start with zero base.
900 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
903 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
905 } else if (!Expr->isZero()) {
907 SubExprs.push_back(Expr);
912 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
913 /// removing any common subexpressions from it. Anything truly common is
914 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
915 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
917 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
918 ScalarEvolution *SE) {
919 unsigned NumUses = Uses.size();
921 // Only one use? Use its base, regardless of what it is!
922 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
923 SCEVHandle Result = Zero;
925 std::swap(Result, Uses[0].Base);
929 // To find common subexpressions, count how many of Uses use each expression.
930 // If any subexpressions are used Uses.size() times, they are common.
931 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
933 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
934 // order we see them.
935 std::vector<SCEVHandle> UniqueSubExprs;
937 std::vector<SCEVHandle> SubExprs;
938 for (unsigned i = 0; i != NumUses; ++i) {
939 // If the base is zero (which is common), return zero now, there are no
941 if (Uses[i].Base == Zero) return Zero;
943 // Split the expression into subexprs.
944 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
945 // Add one to SubExpressionUseCounts for each subexpr present.
946 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
947 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
948 UniqueSubExprs.push_back(SubExprs[j]);
952 // Now that we know how many times each is used, build Result. Iterate over
953 // UniqueSubexprs so that we have a stable ordering.
954 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
955 std::map<SCEVHandle, unsigned>::iterator I =
956 SubExpressionUseCounts.find(UniqueSubExprs[i]);
957 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
958 if (I->second == NumUses) { // Found CSE!
959 Result = SE->getAddExpr(Result, I->first);
961 // Remove non-cse's from SubExpressionUseCounts.
962 SubExpressionUseCounts.erase(I);
966 // If we found no CSE's, return now.
967 if (Result == Zero) return Result;
969 // Otherwise, remove all of the CSE's we found from each of the base values.
970 for (unsigned i = 0; i != NumUses; ++i) {
971 // Split the expression into subexprs.
972 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
974 // Remove any common subexpressions.
975 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
976 if (SubExpressionUseCounts.count(SubExprs[j])) {
977 SubExprs.erase(SubExprs.begin()+j);
981 // Finally, the non-shared expressions together.
982 if (SubExprs.empty())
985 Uses[i].Base = SE->getAddExpr(SubExprs);
992 /// ValidStride - Check whether the given Scale is valid for all loads and
993 /// stores in UsersToProcess.
995 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
997 const std::vector<BasedUser>& UsersToProcess) {
1001 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1002 // If this is a load or other access, pass the type of the access in.
1003 const Type *AccessTy = Type::VoidTy;
1004 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1005 AccessTy = SI->getOperand(0)->getType();
1006 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1007 AccessTy = LI->getType();
1008 else if (isa<PHINode>(UsersToProcess[i].Inst))
1011 TargetLowering::AddrMode AM;
1012 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1013 AM.BaseOffs = SC->getValue()->getSExtValue();
1014 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1017 // If load[imm+r*scale] is illegal, bail out.
1018 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1024 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1026 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1030 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1032 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1033 !(isa<PointerType>(Ty2) &&
1034 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1035 !(isa<PointerType>(Ty1) &&
1036 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1039 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1040 /// of a previous stride and it is a legal value for the target addressing
1041 /// mode scale component and optional base reg. This allows the users of
1042 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1043 /// reuse is possible.
1044 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1045 bool AllUsesAreAddresses,
1046 const SCEVHandle &Stride,
1047 IVExpr &IV, const Type *Ty,
1048 const std::vector<BasedUser>& UsersToProcess) {
1049 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1050 int64_t SInt = SC->getValue()->getSExtValue();
1051 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1053 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1054 IVsByStride.find(StrideOrder[NewStride]);
1055 if (SI == IVsByStride.end())
1057 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1058 if (SI->first != Stride &&
1059 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1061 int64_t Scale = SInt / SSInt;
1062 // Check that this stride is valid for all the types used for loads and
1063 // stores; if it can be used for some and not others, we might as well use
1064 // the original stride everywhere, since we have to create the IV for it
1065 // anyway. If the scale is 1, then we don't need to worry about folding
1068 (AllUsesAreAddresses &&
1069 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1070 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1071 IE = SI->second.IVs.end(); II != IE; ++II)
1072 // FIXME: Only handle base == 0 for now.
1073 // Only reuse previous IV if it would not require a type conversion.
1074 if (II->Base->isZero() &&
1075 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1084 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1085 /// returns true if Val's isUseOfPostIncrementedValue is true.
1086 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1087 return Val.isUseOfPostIncrementedValue;
1090 /// isNonConstantNegative - Return true if the specified scev is negated, but
1092 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1093 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1094 if (!Mul) return false;
1096 // If there is a constant factor, it will be first.
1097 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1098 if (!SC) return false;
1100 // Return true if the value is negative, this matches things like (-42 * V).
1101 return SC->getValue()->getValue().isNegative();
1104 /// isAddress - Returns true if the specified instruction is using the
1105 /// specified value as an address.
1106 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
1107 bool isAddress = isa<LoadInst>(Inst);
1108 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1109 if (SI->getOperand(1) == OperandVal)
1111 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1112 // Addressing modes can also be folded into prefetches and a variety
1114 switch (II->getIntrinsicID()) {
1116 case Intrinsic::prefetch:
1117 case Intrinsic::x86_sse2_loadu_dq:
1118 case Intrinsic::x86_sse2_loadu_pd:
1119 case Intrinsic::x86_sse_loadu_ps:
1120 case Intrinsic::x86_sse_storeu_ps:
1121 case Intrinsic::x86_sse2_storeu_pd:
1122 case Intrinsic::x86_sse2_storeu_dq:
1123 case Intrinsic::x86_sse2_storel_dq:
1124 if (II->getOperand(1) == OperandVal)
1132 // CollectIVUsers - Transform our list of users and offsets to a bit more
1133 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1134 // of the strided accesses, as well as the old information from Uses. We
1135 // progressively move information from the Base field to the Imm field, until
1136 // we eventually have the full access expression to rewrite the use.
1137 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1138 IVUsersOfOneStride &Uses,
1140 bool &AllUsesAreAddresses,
1141 std::vector<BasedUser> &UsersToProcess) {
1142 UsersToProcess.reserve(Uses.Users.size());
1143 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1144 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1146 // Move any loop invariant operands from the offset field to the immediate
1147 // field of the use, so that we don't try to use something before it is
1149 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1150 UsersToProcess.back().Imm, L, SE);
1151 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1152 "Base value is not loop invariant!");
1155 // We now have a whole bunch of uses of like-strided induction variables, but
1156 // they might all have different bases. We want to emit one PHI node for this
1157 // stride which we fold as many common expressions (between the IVs) into as
1158 // possible. Start by identifying the common expressions in the base values
1159 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1160 // "A+B"), emit it to the preheader, then remove the expression from the
1161 // UsersToProcess base values.
1162 SCEVHandle CommonExprs =
1163 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
1165 // Next, figure out what we can represent in the immediate fields of
1166 // instructions. If we can represent anything there, move it to the imm
1167 // fields of the BasedUsers. We do this so that it increases the commonality
1168 // of the remaining uses.
1169 unsigned NumPHI = 0;
1170 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1171 // If the user is not in the current loop, this means it is using the exit
1172 // value of the IV. Do not put anything in the base, make sure it's all in
1173 // the immediate field to allow as much factoring as possible.
1174 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1175 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1176 UsersToProcess[i].Base);
1177 UsersToProcess[i].Base =
1178 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1181 // Addressing modes can be folded into loads and stores. Be careful that
1182 // the store is through the expression, not of the expression though.
1184 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1185 UsersToProcess[i].OperandValToReplace);
1186 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1191 // If this use isn't an address, then not all uses are addresses.
1192 if (!isAddress && !isPHI)
1193 AllUsesAreAddresses = false;
1195 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1196 UsersToProcess[i].Imm, isAddress, L, SE);
1200 // If one of the use if a PHI node and all other uses are addresses, still
1201 // allow iv reuse. Essentially we are trading one constant multiplication
1202 // for one fewer iv.
1204 AllUsesAreAddresses = false;
1209 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1210 /// stride of IV. All of the users may have different starting values, and this
1211 /// may not be the only stride (we know it is if isOnlyStride is true).
1212 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1213 IVUsersOfOneStride &Uses,
1215 bool isOnlyStride) {
1216 // If all the users are moved to another stride, then there is nothing to do.
1217 if (Uses.Users.empty())
1220 // Keep track if every use in UsersToProcess is an address. If they all are,
1221 // we may be able to rewrite the entire collection of them in terms of a
1222 // smaller-stride IV.
1223 bool AllUsesAreAddresses = true;
1225 // Transform our list of users and offsets to a bit more complex table. In
1226 // this new vector, each 'BasedUser' contains 'Base' the base of the
1227 // strided accessas well as the old information from Uses. We progressively
1228 // move information from the Base field to the Imm field, until we eventually
1229 // have the full access expression to rewrite the use.
1230 std::vector<BasedUser> UsersToProcess;
1231 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1234 // If we managed to find some expressions in common, we'll need to carry
1235 // their value in a register and add it in for each use. This will take up
1236 // a register operand, which potentially restricts what stride values are
1238 bool HaveCommonExprs = !CommonExprs->isZero();
1240 // If all uses are addresses, check if it is possible to reuse an IV with a
1241 // stride that is a factor of this stride. And that the multiple is a number
1242 // that can be encoded in the scale field of the target addressing mode. And
1243 // that we will have a valid instruction after this substition, including the
1244 // immediate field, if any.
1245 PHINode *NewPHI = NULL;
1247 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1248 SE->getIntegerSCEV(0, Type::Int32Ty),
1250 unsigned RewriteFactor = 0;
1251 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1252 Stride, ReuseIV, CommonExprs->getType(),
1254 if (RewriteFactor != 0) {
1255 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1256 << " and BASE " << *ReuseIV.Base << " :\n";
1257 NewPHI = ReuseIV.PHI;
1258 IncV = ReuseIV.IncV;
1261 const Type *ReplacedTy = CommonExprs->getType();
1263 // Now that we know what we need to do, insert the PHI node itself.
1265 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1266 << *Stride << " and BASE " << *CommonExprs << ": ";
1268 SCEVExpander Rewriter(*SE, *LI);
1269 SCEVExpander PreheaderRewriter(*SE, *LI);
1271 BasicBlock *Preheader = L->getLoopPreheader();
1272 Instruction *PreInsertPt = Preheader->getTerminator();
1273 Instruction *PhiInsertBefore = L->getHeader()->begin();
1275 BasicBlock *LatchBlock = L->getLoopLatch();
1278 // Emit the initial base value into the loop preheader.
1280 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1282 if (RewriteFactor == 0) {
1283 // Create a new Phi for this base, and stick it in the loop header.
1284 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1287 // Add common base to the new Phi node.
1288 NewPHI->addIncoming(CommonBaseV, Preheader);
1290 // If the stride is negative, insert a sub instead of an add for the
1292 bool isNegative = isNonConstantNegative(Stride);
1293 SCEVHandle IncAmount = Stride;
1295 IncAmount = SE->getNegativeSCEV(Stride);
1297 // Insert the stride into the preheader.
1298 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1299 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1301 // Emit the increment of the base value before the terminator of the loop
1302 // latch block, and add it to the Phi node.
1303 SCEVHandle IncExp = SE->getUnknown(StrideV);
1305 IncExp = SE->getNegativeSCEV(IncExp);
1306 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1308 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1309 IncV->setName(NewPHI->getName()+".inc");
1310 NewPHI->addIncoming(IncV, LatchBlock);
1312 // Remember this in case a later stride is multiple of this.
1313 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1315 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1317 Constant *C = dyn_cast<Constant>(CommonBaseV);
1319 (!C->isNullValue() &&
1320 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1321 // We want the common base emitted into the preheader! This is just
1322 // using cast as a copy so BitCast (no-op cast) is appropriate
1323 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1324 "commonbase", PreInsertPt);
1328 // We want to emit code for users inside the loop first. To do this, we
1329 // rearrange BasedUser so that the entries at the end have
1330 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1331 // vector (so we handle them first).
1332 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1333 PartitionByIsUseOfPostIncrementedValue);
1335 // Sort this by base, so that things with the same base are handled
1336 // together. By partitioning first and stable-sorting later, we are
1337 // guaranteed that within each base we will pop off users from within the
1338 // loop before users outside of the loop with a particular base.
1340 // We would like to use stable_sort here, but we can't. The problem is that
1341 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1342 // we don't have anything to do a '<' comparison on. Because we think the
1343 // number of uses is small, do a horrible bubble sort which just relies on
1345 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1346 // Get a base value.
1347 SCEVHandle Base = UsersToProcess[i].Base;
1349 // Compact everything with this base to be consequtive with this one.
1350 for (unsigned j = i+1; j != e; ++j) {
1351 if (UsersToProcess[j].Base == Base) {
1352 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1358 // Process all the users now. This outer loop handles all bases, the inner
1359 // loop handles all users of a particular base.
1360 while (!UsersToProcess.empty()) {
1361 SCEVHandle Base = UsersToProcess.back().Base;
1363 // Emit the code for Base into the preheader.
1364 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1366 DOUT << " INSERTING code for BASE = " << *Base << ":";
1367 if (BaseV->hasName())
1368 DOUT << " Result value name = %" << BaseV->getNameStr();
1371 // If BaseV is a constant other than 0, make sure that it gets inserted into
1372 // the preheader, instead of being forward substituted into the uses. We do
1373 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1375 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1376 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1377 // We want this constant emitted into the preheader! This is just
1378 // using cast as a copy so BitCast (no-op cast) is appropriate
1379 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1384 // Emit the code to add the immediate offset to the Phi value, just before
1385 // the instructions that we identified as using this stride and base.
1387 // FIXME: Use emitted users to emit other users.
1388 BasedUser &User = UsersToProcess.back();
1390 // If this instruction wants to use the post-incremented value, move it
1391 // after the post-inc and use its value instead of the PHI.
1392 Value *RewriteOp = NewPHI;
1393 if (User.isUseOfPostIncrementedValue) {
1396 // If this user is in the loop, make sure it is the last thing in the
1397 // loop to ensure it is dominated by the increment.
1398 if (L->contains(User.Inst->getParent()))
1399 User.Inst->moveBefore(LatchBlock->getTerminator());
1401 if (RewriteOp->getType() != ReplacedTy) {
1402 Instruction::CastOps opcode = Instruction::Trunc;
1403 if (ReplacedTy->getPrimitiveSizeInBits() ==
1404 RewriteOp->getType()->getPrimitiveSizeInBits())
1405 opcode = Instruction::BitCast;
1406 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1409 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1411 // If we had to insert new instrutions for RewriteOp, we have to
1412 // consider that they may not have been able to end up immediately
1413 // next to RewriteOp, because non-PHI instructions may never precede
1414 // PHI instructions in a block. In this case, remember where the last
1415 // instruction was inserted so that if we're replacing a different
1416 // PHI node, we can use the later point to expand the final
1418 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1419 if (RewriteOp == NewPHI) NewBasePt = 0;
1421 // Clear the SCEVExpander's expression map so that we are guaranteed
1422 // to have the code emitted where we expect it.
1425 // If we are reusing the iv, then it must be multiplied by a constant
1426 // factor take advantage of addressing mode scale component.
1427 if (RewriteFactor != 0) {
1428 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1429 RewriteExpr->getType()),
1432 // The common base is emitted in the loop preheader. But since we
1433 // are reusing an IV, it has not been used to initialize the PHI node.
1434 // Add it to the expression used to rewrite the uses.
1435 if (!isa<ConstantInt>(CommonBaseV) ||
1436 !cast<ConstantInt>(CommonBaseV)->isZero())
1437 RewriteExpr = SE->getAddExpr(RewriteExpr,
1438 SE->getUnknown(CommonBaseV));
1441 // Now that we know what we need to do, insert code before User for the
1442 // immediate and any loop-variant expressions.
1443 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1444 // Add BaseV to the PHI value if needed.
1445 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1447 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1451 // Mark old value we replaced as possibly dead, so that it is elminated
1452 // if we just replaced the last use of that value.
1453 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1455 UsersToProcess.pop_back();
1458 // If there are any more users to process with the same base, process them
1459 // now. We sorted by base above, so we just have to check the last elt.
1460 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1461 // TODO: Next, find out which base index is the most common, pull it out.
1464 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1465 // different starting values, into different PHIs.
1468 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1469 /// set the IV user and stride information and return true, otherwise return
1471 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1472 const SCEVHandle *&CondStride) {
1473 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1475 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1476 IVUsesByStride.find(StrideOrder[Stride]);
1477 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1479 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1480 E = SI->second.Users.end(); UI != E; ++UI)
1481 if (UI->User == Cond) {
1482 // NOTE: we could handle setcc instructions with multiple uses here, but
1483 // InstCombine does it as well for simple uses, it's not clear that it
1484 // occurs enough in real life to handle.
1486 CondStride = &SI->first;
1494 // Constant strides come first which in turns are sorted by their absolute
1495 // values. If absolute values are the same, then positive strides comes first.
1497 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1498 struct StrideCompare {
1499 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1500 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1501 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1503 int64_t LV = LHSC->getValue()->getSExtValue();
1504 int64_t RV = RHSC->getValue()->getSExtValue();
1505 uint64_t ALV = (LV < 0) ? -LV : LV;
1506 uint64_t ARV = (RV < 0) ? -RV : RV;
1512 return (LHSC && !RHSC);
1517 /// ChangeCompareStride - If a loop termination compare instruction is the
1518 /// only use of its stride, and the compaison is against a constant value,
1519 /// try eliminate the stride by moving the compare instruction to another
1520 /// stride and change its constant operand accordingly. e.g.
1526 /// if (v2 < 10) goto loop
1531 /// if (v1 < 30) goto loop
1532 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1533 IVStrideUse* &CondUse,
1534 const SCEVHandle* &CondStride) {
1535 if (StrideOrder.size() < 2 ||
1536 IVUsesByStride[*CondStride].Users.size() != 1)
1538 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1539 if (!SC) return Cond;
1540 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1541 if (!C) return Cond;
1543 ICmpInst::Predicate Predicate = Cond->getPredicate();
1544 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1545 int64_t CmpVal = C->getValue().getSExtValue();
1546 unsigned BitWidth = C->getValue().getBitWidth();
1547 uint64_t SignBit = 1ULL << (BitWidth-1);
1548 const Type *CmpTy = C->getType();
1549 const Type *NewCmpTy = NULL;
1550 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1551 unsigned NewTyBits = 0;
1552 int64_t NewCmpVal = CmpVal;
1553 SCEVHandle *NewStride = NULL;
1554 Value *NewIncV = NULL;
1557 // Check stride constant and the comparision constant signs to detect
1559 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1562 // Look for a suitable stride / iv as replacement.
1563 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1564 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1565 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1566 IVUsesByStride.find(StrideOrder[i]);
1567 if (!isa<SCEVConstant>(SI->first))
1569 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1570 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1573 Scale = SSInt / CmpSSInt;
1574 NewCmpVal = CmpVal * Scale;
1575 APInt Mul = APInt(BitWidth, NewCmpVal);
1576 // Check for overflow.
1577 if (Mul.getSExtValue() != NewCmpVal) {
1582 // Watch out for overflow.
1583 if (ICmpInst::isSignedPredicate(Predicate) &&
1584 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1587 if (NewCmpVal != CmpVal) {
1588 // Pick the best iv to use trying to avoid a cast.
1590 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1591 E = SI->second.Users.end(); UI != E; ++UI) {
1592 NewIncV = UI->OperandValToReplace;
1593 if (NewIncV->getType() == CmpTy)
1601 NewCmpTy = NewIncV->getType();
1602 NewTyBits = isa<PointerType>(NewCmpTy)
1603 ? UIntPtrTy->getPrimitiveSizeInBits()
1604 : NewCmpTy->getPrimitiveSizeInBits();
1605 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1606 // Check if it is possible to rewrite it using
1607 // an iv / stride of a smaller integer type.
1608 bool TruncOk = false;
1609 if (NewCmpTy->isInteger()) {
1610 unsigned Bits = NewTyBits;
1611 if (ICmpInst::isSignedPredicate(Predicate))
1613 uint64_t Mask = (1ULL << Bits) - 1;
1614 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1623 // Don't rewrite if use offset is non-constant and the new type is
1624 // of a different type.
1625 // FIXME: too conservative?
1626 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1631 bool AllUsesAreAddresses = true;
1632 std::vector<BasedUser> UsersToProcess;
1633 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1634 AllUsesAreAddresses,
1636 // Avoid rewriting the compare instruction with an iv of new stride
1637 // if it's likely the new stride uses will be rewritten using the
1638 if (AllUsesAreAddresses &&
1639 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1644 // If scale is negative, use swapped predicate unless it's testing
1646 if (Scale < 0 && !Cond->isEquality())
1647 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1649 NewStride = &StrideOrder[i];
1654 // Forgo this transformation if it the increment happens to be
1655 // unfortunately positioned after the condition, and the condition
1656 // has multiple uses which prevent it from being moved immediately
1657 // before the branch. See
1658 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1659 // for an example of this situation.
1660 if (!Cond->hasOneUse()) {
1661 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1667 if (NewCmpVal != CmpVal) {
1668 // Create a new compare instruction using new stride / iv.
1669 ICmpInst *OldCond = Cond;
1671 if (!isa<PointerType>(NewCmpTy))
1672 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1674 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1675 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1677 // Insert new compare instruction.
1678 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1679 L->getHeader()->getName() + ".termcond",
1682 // Remove the old compare instruction. The old indvar is probably dead too.
1683 DeadInsts.insert(cast<Instruction>(CondUse->OperandValToReplace));
1684 SE->deleteValueFromRecords(OldCond);
1685 OldCond->replaceAllUsesWith(Cond);
1686 OldCond->eraseFromParent();
1688 IVUsesByStride[*CondStride].Users.pop_back();
1689 SCEVHandle NewOffset = TyBits == NewTyBits
1690 ? SE->getMulExpr(CondUse->Offset,
1691 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1692 : SE->getConstant(ConstantInt::get(NewCmpTy,
1693 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1694 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1695 CondUse = &IVUsesByStride[*NewStride].Users.back();
1696 CondStride = NewStride;
1703 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
1704 /// an smax computation.
1706 /// This is a narrow solution to a specific, but acute, problem. For loops
1712 /// } while (++i < n);
1714 /// where the comparison is signed, the trip count isn't just 'n', because
1715 /// 'n' could be negative. And unfortunately this can come up even for loops
1716 /// where the user didn't use a C do-while loop. For example, seemingly
1717 /// well-behaved top-test loops will commonly be lowered like this:
1723 /// } while (++i < n);
1726 /// and then it's possible for subsequent optimization to obscure the if
1727 /// test in such a way that indvars can't find it.
1729 /// When indvars can't find the if test in loops like this, it creates a
1730 /// signed-max expression, which allows it to give the loop a canonical
1731 /// induction variable:
1734 /// smax = n < 1 ? 1 : n;
1737 /// } while (++i != smax);
1739 /// Canonical induction variables are necessary because the loop passes
1740 /// are designed around them. The most obvious example of this is the
1741 /// LoopInfo analysis, which doesn't remember trip count values. It
1742 /// expects to be able to rediscover the trip count each time it is
1743 /// needed, and it does this using a simple analyis that only succeeds if
1744 /// the loop has a canonical induction variable.
1746 /// However, when it comes time to generate code, the maximum operation
1747 /// can be quite costly, especially if it's inside of an outer loop.
1749 /// This function solves this problem by detecting this type of loop and
1750 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1751 /// the instructions for the maximum computation.
1753 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
1754 IVStrideUse* &CondUse) {
1755 // Check that the loop matches the pattern we're looking for.
1756 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1757 Cond->getPredicate() != CmpInst::ICMP_NE)
1760 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1761 if (!Sel || !Sel->hasOneUse()) return Cond;
1763 SCEVHandle IterationCount = SE->getIterationCount(L);
1764 if (isa<SCEVCouldNotCompute>(IterationCount))
1766 SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());
1768 // Adjust for an annoying getIterationCount quirk.
1769 IterationCount = SE->getAddExpr(IterationCount, One);
1771 // Check for a max calculation that matches the pattern.
1772 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
1773 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
1775 SCEVHandle SMaxLHS = SMax->getOperand(0);
1776 SCEVHandle SMaxRHS = SMax->getOperand(1);
1777 if (!SMaxLHS || SMaxLHS != One) return Cond;
1779 // Check the relevant induction variable for conformance to
1781 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
1782 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1783 if (!AR || !AR->isAffine() ||
1784 AR->getStart() != One ||
1785 AR->getStepRecurrence(*SE) != One)
1788 // Check the right operand of the select, and remember it, as it will
1789 // be used in the new comparison instruction.
1791 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
1792 NewRHS = Sel->getOperand(1);
1793 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
1794 NewRHS = Sel->getOperand(2);
1795 if (!NewRHS) return Cond;
1797 // Ok, everything looks ok to change the condition into an SLT or SGE and
1798 // delete the max calculation.
1800 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
1803 Cond->getOperand(0), NewRHS, "scmp", Cond);
1805 // Delete the max calculation instructions.
1806 SE->deleteValueFromRecords(Cond);
1807 Cond->replaceAllUsesWith(NewCond);
1808 Cond->eraseFromParent();
1809 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1810 SE->deleteValueFromRecords(Sel);
1811 Sel->eraseFromParent();
1812 if (Cmp->use_empty()) {
1813 SE->deleteValueFromRecords(Cmp);
1814 Cmp->eraseFromParent();
1816 CondUse->User = NewCond;
1820 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1821 /// inside the loop then try to eliminate the cast opeation.
1822 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
1824 SCEVHandle IterationCount = SE->getIterationCount(L);
1825 if (isa<SCEVCouldNotCompute>(IterationCount))
1828 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
1830 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1831 IVUsesByStride.find(StrideOrder[Stride]);
1832 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1833 if (!isa<SCEVConstant>(SI->first))
1836 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1837 E = SI->second.Users.end(); UI != E; /* empty */) {
1838 std::vector<IVStrideUse>::iterator CandidateUI = UI;
1840 Instruction *ShadowUse = CandidateUI->User;
1841 const Type *DestTy = NULL;
1843 /* If shadow use is a int->float cast then insert a second IV
1844 to eliminate this cast.
1846 for (unsigned i = 0; i < n; ++i)
1852 for (unsigned i = 0; i < n; ++i, ++d)
1855 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
1856 DestTy = UCast->getDestTy();
1857 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
1858 DestTy = SCast->getDestTy();
1859 if (!DestTy) continue;
1862 /* If target does not support DestTy natively then do not apply
1863 this transformation. */
1864 MVT DVT = TLI->getValueType(DestTy);
1865 if (!TLI->isTypeLegal(DVT)) continue;
1868 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1870 if (PH->getNumIncomingValues() != 2) continue;
1872 const Type *SrcTy = PH->getType();
1873 int Mantissa = DestTy->getFPMantissaWidth();
1874 if (Mantissa == -1) continue;
1875 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
1878 unsigned Entry, Latch;
1879 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1887 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1888 if (!Init) continue;
1889 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1891 BinaryOperator *Incr =
1892 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1893 if (!Incr) continue;
1894 if (Incr->getOpcode() != Instruction::Add
1895 && Incr->getOpcode() != Instruction::Sub)
1898 /* Initialize new IV, double d = 0.0 in above example. */
1899 ConstantInt *C = NULL;
1900 if (Incr->getOperand(0) == PH)
1901 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1902 else if (Incr->getOperand(1) == PH)
1903 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1909 /* Add new PHINode. */
1910 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1912 /* create new increment. '++d' in above example. */
1913 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1914 BinaryOperator *NewIncr =
1915 BinaryOperator::Create(Incr->getOpcode(),
1916 NewPH, CFP, "IV.S.next.", Incr);
1918 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1919 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1921 /* Remove cast operation */
1922 SE->deleteValueFromRecords(ShadowUse);
1923 ShadowUse->replaceAllUsesWith(NewPH);
1924 ShadowUse->eraseFromParent();
1925 SI->second.Users.erase(CandidateUI);
1932 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1933 // uses in the loop, look to see if we can eliminate some, in favor of using
1934 // common indvars for the different uses.
1935 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1936 // TODO: implement optzns here.
1938 OptimizeShadowIV(L);
1940 // Finally, get the terminating condition for the loop if possible. If we
1941 // can, we want to change it to use a post-incremented version of its
1942 // induction variable, to allow coalescing the live ranges for the IV into
1943 // one register value.
1944 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1945 BasicBlock *Preheader = L->getLoopPreheader();
1946 BasicBlock *LatchBlock =
1947 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1948 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1949 if (!TermBr || TermBr->isUnconditional() ||
1950 !isa<ICmpInst>(TermBr->getCondition()))
1952 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1954 // Search IVUsesByStride to find Cond's IVUse if there is one.
1955 IVStrideUse *CondUse = 0;
1956 const SCEVHandle *CondStride = 0;
1958 if (!FindIVUserForCond(Cond, CondUse, CondStride))
1959 return; // setcc doesn't use the IV.
1961 // If the trip count is computed in terms of an smax (due to ScalarEvolution
1962 // being unable to find a sufficient guard, for example), change the loop
1963 // comparison to use SLT instead of NE.
1964 Cond = OptimizeSMax(L, Cond, CondUse);
1966 // If possible, change stride and operands of the compare instruction to
1967 // eliminate one stride.
1968 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
1970 // It's possible for the setcc instruction to be anywhere in the loop, and
1971 // possible for it to have multiple users. If it is not immediately before
1972 // the latch block branch, move it.
1973 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1974 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1975 Cond->moveBefore(TermBr);
1977 // Otherwise, clone the terminating condition and insert into the loopend.
1978 Cond = cast<ICmpInst>(Cond->clone());
1979 Cond->setName(L->getHeader()->getName() + ".termcond");
1980 LatchBlock->getInstList().insert(TermBr, Cond);
1982 // Clone the IVUse, as the old use still exists!
1983 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1984 CondUse->OperandValToReplace);
1985 CondUse = &IVUsesByStride[*CondStride].Users.back();
1989 // If we get to here, we know that we can transform the setcc instruction to
1990 // use the post-incremented version of the IV, allowing us to coalesce the
1991 // live ranges for the IV correctly.
1992 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
1993 CondUse->isUseOfPostIncrementedValue = true;
1997 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1999 LI = &getAnalysis<LoopInfo>();
2000 DT = &getAnalysis<DominatorTree>();
2001 SE = &getAnalysis<ScalarEvolution>();
2002 TD = &getAnalysis<TargetData>();
2003 UIntPtrTy = TD->getIntPtrType();
2006 // Find all uses of induction variables in this loop, and catagorize
2007 // them by stride. Start by finding all of the PHI nodes in the header for
2008 // this loop. If they are induction variables, inspect their uses.
2009 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2010 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2011 AddUsersIfInteresting(I, L, Processed);
2013 if (!IVUsesByStride.empty()) {
2014 // Optimize induction variables. Some indvar uses can be transformed to use
2015 // strides that will be needed for other purposes. A common example of this
2016 // is the exit test for the loop, which can often be rewritten to use the
2017 // computation of some other indvar to decide when to terminate the loop.
2020 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2021 // doing computation in byte values, promote to 32-bit values if safe.
2023 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2024 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2025 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2026 // Need to be careful that IV's are all the same type. Only works for
2027 // intptr_t indvars.
2029 // If we only have one stride, we can more aggressively eliminate some
2031 bool HasOneStride = IVUsesByStride.size() == 1;
2034 DOUT << "\nLSR on ";
2038 // IVsByStride keeps IVs for one particular loop.
2039 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2041 // Sort the StrideOrder so we process larger strides first.
2042 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2044 // Note: this processes each stride/type pair individually. All users
2045 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2046 // Also, note that we iterate over IVUsesByStride indirectly by using
2047 // StrideOrder. This extra layer of indirection makes the ordering of
2048 // strides deterministic - not dependent on map order.
2049 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2050 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2051 IVUsesByStride.find(StrideOrder[Stride]);
2052 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2053 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
2057 // We're done analyzing this loop; release all the state we built up for it.
2058 CastedPointers.clear();
2059 IVUsesByStride.clear();
2060 IVsByStride.clear();
2061 StrideOrder.clear();
2063 // Clean up after ourselves
2064 if (!DeadInsts.empty()) {
2065 DeleteTriviallyDeadInstructions(DeadInsts);
2067 BasicBlock::iterator I = L->getHeader()->begin();
2068 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2069 // At this point, we know that we have killed one or more IV users.
2070 // It is worth checking to see if the cann indvar is also
2071 // dead, so that we can remove it as well.
2073 // We can remove a PHI if it is on a cycle in the def-use graph
2074 // where each node in the cycle has degree one, i.e. only one use,
2075 // and is an instruction with no side effects.
2077 // FIXME: this needs to eliminate an induction variable even if it's being
2078 // compared against some value to decide loop termination.
2079 if (PN->hasOneUse()) {
2080 SmallPtrSet<PHINode *, 2> PHIs;
2081 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2082 J && J->hasOneUse() && !J->mayWriteToMemory();
2083 J = dyn_cast<Instruction>(*J->use_begin())) {
2084 // If we find the original PHI, we've discovered a cycle.
2086 // Break the cycle and mark the PHI for deletion.
2087 SE->deleteValueFromRecords(PN);
2088 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2089 DeadInsts.insert(PN);
2093 // If we find a PHI more than once, we're on a cycle that
2094 // won't prove fruitful.
2095 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2100 DeleteTriviallyDeadInstructions(DeadInsts);