1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under the
6 // University of Illinois Open Source 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/Type.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/ScalarEvolutionExpander.h"
27 #include "llvm/Support/CFG.h"
28 #include "llvm/Support/GetElementPtrTypeIterator.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Target/TargetLowering.h"
41 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
42 Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
43 Statistic<> NumVariable("loop-reduce","Number of PHIs with variable strides");
45 /// IVStrideUse - Keep track of one use of a strided induction variable, where
46 /// the stride is stored externally. The Offset member keeps track of the
47 /// offset from the IV, User is the actual user of the operand, and 'Operand'
48 /// is the operand # of the User that is the use.
52 Value *OperandValToReplace;
54 // isUseOfPostIncrementedValue - True if this should use the
55 // post-incremented version of this IV, not the preincremented version.
56 // This can only be set in special cases, such as the terminating setcc
57 // instruction for a loop or uses dominated by the loop.
58 bool isUseOfPostIncrementedValue;
60 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
61 : Offset(Offs), User(U), OperandValToReplace(O),
62 isUseOfPostIncrementedValue(false) {}
65 /// IVUsersOfOneStride - This structure keeps track of all instructions that
66 /// have an operand that is based on the trip count multiplied by some stride.
67 /// The stride for all of these users is common and kept external to this
69 struct IVUsersOfOneStride {
70 /// Users - Keep track of all of the users of this stride as well as the
71 /// initial value and the operand that uses the IV.
72 std::vector<IVStrideUse> Users;
74 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
75 Users.push_back(IVStrideUse(Offset, User, Operand));
79 /// IVInfo - This structure keeps track of one IV expression inserted during
80 /// StrengthReduceStridedIVUsers. It contains the base value, as well as the
81 /// PHI node and increment value created for rewrite.
87 IVExpr(const SCEVHandle &base, PHINode *phi, Value *incv)
88 : Base(base), PHI(phi), IncV(incv) {}
91 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
92 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
93 struct IVsOfOneStride {
94 std::vector<IVExpr> IVs;
96 void addIV(const SCEVHandle &Base, PHINode *PHI, Value *IncV) {
97 IVs.push_back(IVExpr(Base, PHI, IncV));
101 class LoopStrengthReduce : public FunctionPass {
105 const TargetData *TD;
106 const Type *UIntPtrTy;
109 /// IVUsesByStride - Keep track of all uses of induction variables that we
110 /// are interested in. The key of the map is the stride of the access.
111 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
113 /// IVsByStride - Keep track of all IVs that have been inserted for a
114 /// particular stride.
115 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
117 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
118 /// We use this to iterate over the IVUsesByStride collection without being
119 /// dependent on random ordering of pointers in the process.
120 std::vector<SCEVHandle> StrideOrder;
122 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
123 /// of the casted version of each value. This is accessed by
124 /// getCastedVersionOf.
125 std::map<Value*, Value*> CastedPointers;
127 /// DeadInsts - Keep track of instructions we may have made dead, so that
128 /// we can remove them after we are done working.
129 std::set<Instruction*> DeadInsts;
131 /// TLI - Keep a pointer of a TargetLowering to consult for determining
132 /// transformation profitability.
133 const TargetLowering *TLI;
136 LoopStrengthReduce(const TargetLowering *tli = NULL)
140 virtual bool runOnFunction(Function &) {
141 LI = &getAnalysis<LoopInfo>();
142 EF = &getAnalysis<ETForest>();
143 SE = &getAnalysis<ScalarEvolution>();
144 TD = &getAnalysis<TargetData>();
145 UIntPtrTy = TD->getIntPtrType();
148 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 // We split critical edges, so we change the CFG. However, we do update
156 // many analyses if they are around.
157 AU.addPreservedID(LoopSimplifyID);
158 AU.addPreserved<LoopInfo>();
159 AU.addPreserved<DominatorSet>();
160 AU.addPreserved<ETForest>();
161 AU.addPreserved<ImmediateDominators>();
162 AU.addPreserved<DominanceFrontier>();
163 AU.addPreserved<DominatorTree>();
165 AU.addRequiredID(LoopSimplifyID);
166 AU.addRequired<LoopInfo>();
167 AU.addRequired<ETForest>();
168 AU.addRequired<TargetData>();
169 AU.addRequired<ScalarEvolution>();
172 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
174 Value *getCastedVersionOf(Value *V);
176 void runOnLoop(Loop *L);
177 bool AddUsersIfInteresting(Instruction *I, Loop *L,
178 std::set<Instruction*> &Processed);
179 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
181 void OptimizeIndvars(Loop *L);
183 unsigned CheckForIVReuse(const SCEVHandle &Stride, IVExpr &IV);
185 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
186 IVUsersOfOneStride &Uses,
187 Loop *L, bool isOnlyStride);
188 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
190 RegisterOpt<LoopStrengthReduce> X("loop-reduce",
191 "Loop Strength Reduction");
194 FunctionPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
195 return new LoopStrengthReduce(TLI);
198 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
200 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
201 if (V->getType() == UIntPtrTy) return V;
202 if (Constant *CB = dyn_cast<Constant>(V))
203 return ConstantExpr::getCast(CB, UIntPtrTy);
205 Value *&New = CastedPointers[V];
208 New = SCEVExpander::InsertCastOfTo(V, UIntPtrTy);
209 DeadInsts.insert(cast<Instruction>(New));
214 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
215 /// specified set are trivially dead, delete them and see if this makes any of
216 /// their operands subsequently dead.
217 void LoopStrengthReduce::
218 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
219 while (!Insts.empty()) {
220 Instruction *I = *Insts.begin();
221 Insts.erase(Insts.begin());
222 if (isInstructionTriviallyDead(I)) {
223 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
224 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
226 SE->deleteInstructionFromRecords(I);
227 I->eraseFromParent();
234 /// GetExpressionSCEV - Compute and return the SCEV for the specified
236 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
237 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
238 // If this is a GEP that SE doesn't know about, compute it now and insert it.
239 // If this is not a GEP, or if we have already done this computation, just let
241 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
242 if (!GEP || SE->hasSCEV(GEP))
243 return SE->getSCEV(Exp);
245 // Analyze all of the subscripts of this getelementptr instruction, looking
246 // for uses that are determined by the trip count of L. First, skip all
247 // operands the are not dependent on the IV.
249 // Build up the base expression. Insert an LLVM cast of the pointer to
251 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
253 gep_type_iterator GTI = gep_type_begin(GEP);
255 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
256 // If this is a use of a recurrence that we can analyze, and it comes before
257 // Op does in the GEP operand list, we will handle this when we process this
259 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
260 const StructLayout *SL = TD->getStructLayout(STy);
261 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
262 uint64_t Offset = SL->MemberOffsets[Idx];
263 GEPVal = SCEVAddExpr::get(GEPVal,
264 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
266 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
267 SCEVHandle Idx = SE->getSCEV(OpVal);
269 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
271 Idx = SCEVMulExpr::get(Idx,
272 SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
274 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
278 SE->setSCEV(GEP, GEPVal);
282 /// getSCEVStartAndStride - Compute the start and stride of this expression,
283 /// returning false if the expression is not a start/stride pair, or true if it
284 /// is. The stride must be a loop invariant expression, but the start may be
285 /// a mix of loop invariant and loop variant expressions.
286 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
287 SCEVHandle &Start, SCEVHandle &Stride) {
288 SCEVHandle TheAddRec = Start; // Initialize to zero.
290 // If the outer level is an AddExpr, the operands are all start values except
291 // for a nested AddRecExpr.
292 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
293 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
294 if (SCEVAddRecExpr *AddRec =
295 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
296 if (AddRec->getLoop() == L)
297 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
299 return false; // Nested IV of some sort?
301 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
304 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
307 return false; // not analyzable.
310 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
311 if (!AddRec || AddRec->getLoop() != L) return false;
313 // FIXME: Generalize to non-affine IV's.
314 if (!AddRec->isAffine()) return false;
316 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
318 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
319 DEBUG(std::cerr << "[" << L->getHeader()->getName()
320 << "] Variable stride: " << *AddRec << "\n");
322 Stride = AddRec->getOperand(1);
323 // Check that all constant strides are the unsigned type, we don't want to
324 // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
326 assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
327 "Constants should be canonicalized to unsigned!");
332 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
333 /// and now we need to decide whether the user should use the preinc or post-inc
334 /// value. If this user should use the post-inc version of the IV, return true.
336 /// Choosing wrong here can break dominance properties (if we choose to use the
337 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
338 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
339 /// should use the post-inc value).
340 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
341 Loop *L, ETForest *EF, Pass *P) {
342 // If the user is in the loop, use the preinc value.
343 if (L->contains(User->getParent())) return false;
345 BasicBlock *LatchBlock = L->getLoopLatch();
347 // Ok, the user is outside of the loop. If it is dominated by the latch
348 // block, use the post-inc value.
349 if (EF->dominates(LatchBlock, User->getParent()))
352 // There is one case we have to be careful of: PHI nodes. These little guys
353 // can live in blocks that do not dominate the latch block, but (since their
354 // uses occur in the predecessor block, not the block the PHI lives in) should
355 // still use the post-inc value. Check for this case now.
356 PHINode *PN = dyn_cast<PHINode>(User);
357 if (!PN) return false; // not a phi, not dominated by latch block.
359 // Look at all of the uses of IV by the PHI node. If any use corresponds to
360 // a block that is not dominated by the latch block, give up and use the
361 // preincremented value.
362 unsigned NumUses = 0;
363 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
364 if (PN->getIncomingValue(i) == IV) {
366 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
370 // Okay, all uses of IV by PN are in predecessor blocks that really are
371 // dominated by the latch block. Split the critical edges and use the
372 // post-incremented value.
373 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
374 if (PN->getIncomingValue(i) == IV) {
375 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P);
376 if (--NumUses == 0) break;
384 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
385 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
386 /// return true. Otherwise, return false.
387 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
388 std::set<Instruction*> &Processed) {
389 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
390 return false; // Void and FP expressions cannot be reduced.
391 if (!Processed.insert(I).second)
392 return true; // Instruction already handled.
394 // Get the symbolic expression for this instruction.
395 SCEVHandle ISE = GetExpressionSCEV(I, L);
396 if (isa<SCEVCouldNotCompute>(ISE)) return false;
398 // Get the start and stride for this expression.
399 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
400 SCEVHandle Stride = Start;
401 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
402 return false; // Non-reducible symbolic expression, bail out.
404 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
405 Instruction *User = cast<Instruction>(*UI);
407 // Do not infinitely recurse on PHI nodes.
408 if (isa<PHINode>(User) && Processed.count(User))
411 // If this is an instruction defined in a nested loop, or outside this loop,
412 // don't recurse into it.
413 bool AddUserToIVUsers = false;
414 if (LI->getLoopFor(User->getParent()) != L) {
415 DEBUG(std::cerr << "FOUND USER in other loop: " << *User
416 << " OF SCEV: " << *ISE << "\n");
417 AddUserToIVUsers = true;
418 } else if (!AddUsersIfInteresting(User, L, Processed)) {
419 DEBUG(std::cerr << "FOUND USER: " << *User
420 << " OF SCEV: " << *ISE << "\n");
421 AddUserToIVUsers = true;
424 if (AddUserToIVUsers) {
425 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
426 if (StrideUses.Users.empty()) // First occurance of this stride?
427 StrideOrder.push_back(Stride);
429 // Okay, we found a user that we cannot reduce. Analyze the instruction
430 // and decide what to do with it. If we are a use inside of the loop, use
431 // the value before incrementation, otherwise use it after incrementation.
432 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
433 // The value used will be incremented by the stride more than we are
434 // expecting, so subtract this off.
435 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
436 StrideUses.addUser(NewStart, User, I);
437 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
438 DEBUG(std::cerr << " USING POSTINC SCEV, START=" << *NewStart<< "\n");
440 StrideUses.addUser(Start, User, I);
448 /// BasedUser - For a particular base value, keep information about how we've
449 /// partitioned the expression so far.
451 /// Base - The Base value for the PHI node that needs to be inserted for
452 /// this use. As the use is processed, information gets moved from this
453 /// field to the Imm field (below). BasedUser values are sorted by this
457 /// Inst - The instruction using the induction variable.
460 /// OperandValToReplace - The operand value of Inst to replace with the
462 Value *OperandValToReplace;
464 /// Imm - The immediate value that should be added to the base immediately
465 /// before Inst, because it will be folded into the imm field of the
469 /// EmittedBase - The actual value* to use for the base value of this
470 /// operation. This is null if we should just use zero so far.
473 // isUseOfPostIncrementedValue - True if this should use the
474 // post-incremented version of this IV, not the preincremented version.
475 // This can only be set in special cases, such as the terminating setcc
476 // instruction for a loop and uses outside the loop that are dominated by
478 bool isUseOfPostIncrementedValue;
480 BasedUser(IVStrideUse &IVSU)
481 : Base(IVSU.Offset), Inst(IVSU.User),
482 OperandValToReplace(IVSU.OperandValToReplace),
483 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
484 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
486 // Once we rewrite the code to insert the new IVs we want, update the
487 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
489 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
490 SCEVExpander &Rewriter, Loop *L,
493 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
494 SCEVExpander &Rewriter,
495 Instruction *IP, Loop *L);
500 void BasedUser::dump() const {
501 std::cerr << " Base=" << *Base;
502 std::cerr << " Imm=" << *Imm;
504 std::cerr << " EB=" << *EmittedBase;
506 std::cerr << " Inst: " << *Inst;
509 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
510 SCEVExpander &Rewriter,
511 Instruction *IP, Loop *L) {
512 // Figure out where we *really* want to insert this code. In particular, if
513 // the user is inside of a loop that is nested inside of L, we really don't
514 // want to insert this expression before the user, we'd rather pull it out as
515 // many loops as possible.
516 LoopInfo &LI = Rewriter.getLoopInfo();
517 Instruction *BaseInsertPt = IP;
519 // Figure out the most-nested loop that IP is in.
520 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
522 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
523 // the preheader of the outer-most loop where NewBase is not loop invariant.
524 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
525 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
526 InsertLoop = InsertLoop->getParentLoop();
529 // If there is no immediate value, skip the next part.
530 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
531 if (SC->getValue()->isNullValue())
532 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
533 OperandValToReplace->getType());
535 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
537 // Always emit the immediate (if non-zero) into the same block as the user.
538 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
539 return Rewriter.expandCodeFor(NewValSCEV, IP,
540 OperandValToReplace->getType());
544 // Once we rewrite the code to insert the new IVs we want, update the
545 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
547 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
548 SCEVExpander &Rewriter,
550 if (!isa<PHINode>(Inst)) {
551 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
552 // Replace the use of the operand Value with the new Phi we just created.
553 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
554 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
558 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
559 // expression into each operand block that uses it. Note that PHI nodes can
560 // have multiple entries for the same predecessor. We use a map to make sure
561 // that a PHI node only has a single Value* for each predecessor (which also
562 // prevents us from inserting duplicate code in some blocks).
563 std::map<BasicBlock*, Value*> InsertedCode;
564 PHINode *PN = cast<PHINode>(Inst);
565 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
566 if (PN->getIncomingValue(i) == OperandValToReplace) {
567 // If this is a critical edge, split the edge so that we do not insert the
568 // code on all predecessor/successor paths. We do this unless this is the
569 // canonical backedge for this loop, as this can make some inserted code
570 // be in an illegal position.
571 BasicBlock *PHIPred = PN->getIncomingBlock(i);
572 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
573 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
575 // First step, split the critical edge.
576 SplitCriticalEdge(PHIPred, PN->getParent(), P);
578 // Next step: move the basic block. In particular, if the PHI node
579 // is outside of the loop, and PredTI is in the loop, we want to
580 // move the block to be immediately before the PHI block, not
581 // immediately after PredTI.
582 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
583 BasicBlock *NewBB = PN->getIncomingBlock(i);
584 NewBB->moveBefore(PN->getParent());
588 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
590 // Insert the code into the end of the predecessor block.
591 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
592 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
595 // Replace the use of the operand Value with the new Phi we just created.
596 PN->setIncomingValue(i, Code);
600 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
604 /// isTargetConstant - Return true if the following can be referenced by the
605 /// immediate field of a target instruction.
606 static bool isTargetConstant(const SCEVHandle &V, const TargetLowering *TLI) {
607 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
608 int64_t V = SC->getValue()->getSExtValue();
610 return TLI->isLegalAddressImmediate(V);
612 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
613 return (V > -(1 << 16) && V < (1 << 16)-1);
616 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
617 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
618 if (CE->getOpcode() == Instruction::Cast) {
619 Constant *Op0 = CE->getOperand(0);
620 if (isa<GlobalValue>(Op0) &&
622 TLI->isLegalAddressImmediate(cast<GlobalValue>(Op0)))
628 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
629 /// loop varying to the Imm operand.
630 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
632 if (Val->isLoopInvariant(L)) return; // Nothing to do.
634 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
635 std::vector<SCEVHandle> NewOps;
636 NewOps.reserve(SAE->getNumOperands());
638 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
639 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
640 // If this is a loop-variant expression, it must stay in the immediate
641 // field of the expression.
642 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
644 NewOps.push_back(SAE->getOperand(i));
648 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
650 Val = SCEVAddExpr::get(NewOps);
651 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
652 // Try to pull immediates out of the start value of nested addrec's.
653 SCEVHandle Start = SARE->getStart();
654 MoveLoopVariantsToImediateField(Start, Imm, L);
656 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
658 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
660 // Otherwise, all of Val is variant, move the whole thing over.
661 Imm = SCEVAddExpr::get(Imm, Val);
662 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
667 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
668 /// that can fit into the immediate field of instructions in the target.
669 /// Accumulate these immediate values into the Imm value.
670 static void MoveImmediateValues(const TargetLowering *TLI,
671 SCEVHandle &Val, SCEVHandle &Imm,
672 bool isAddress, Loop *L) {
673 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
674 std::vector<SCEVHandle> NewOps;
675 NewOps.reserve(SAE->getNumOperands());
677 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
678 SCEVHandle NewOp = SAE->getOperand(i);
679 MoveImmediateValues(TLI, NewOp, Imm, isAddress, L);
681 if (!NewOp->isLoopInvariant(L)) {
682 // If this is a loop-variant expression, it must stay in the immediate
683 // field of the expression.
684 Imm = SCEVAddExpr::get(Imm, NewOp);
686 NewOps.push_back(NewOp);
691 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
693 Val = SCEVAddExpr::get(NewOps);
695 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
696 // Try to pull immediates out of the start value of nested addrec's.
697 SCEVHandle Start = SARE->getStart();
698 MoveImmediateValues(TLI, Start, Imm, isAddress, L);
700 if (Start != SARE->getStart()) {
701 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
703 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
706 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
707 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
708 if (isAddress && isTargetConstant(SME->getOperand(0), TLI) &&
709 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
711 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
712 SCEVHandle NewOp = SME->getOperand(1);
713 MoveImmediateValues(TLI, NewOp, SubImm, isAddress, L);
715 // If we extracted something out of the subexpressions, see if we can
717 if (NewOp != SME->getOperand(1)) {
718 // Scale SubImm up by "8". If the result is a target constant, we are
720 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
721 if (isTargetConstant(SubImm, TLI)) {
722 // Accumulate the immediate.
723 Imm = SCEVAddExpr::get(Imm, SubImm);
725 // Update what is left of 'Val'.
726 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
733 // Loop-variant expressions must stay in the immediate field of the
735 if ((isAddress && isTargetConstant(Val, TLI)) ||
736 !Val->isLoopInvariant(L)) {
737 Imm = SCEVAddExpr::get(Imm, Val);
738 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
742 // Otherwise, no immediates to move.
746 /// IncrementAddExprUses - Decompose the specified expression into its added
747 /// subexpressions, and increment SubExpressionUseCounts for each of these
748 /// decomposed parts.
749 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
751 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
752 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
753 SeparateSubExprs(SubExprs, AE->getOperand(j));
754 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
755 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
756 if (SARE->getOperand(0) == Zero) {
757 SubExprs.push_back(Expr);
759 // Compute the addrec with zero as its base.
760 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
761 Ops[0] = Zero; // Start with zero base.
762 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
765 SeparateSubExprs(SubExprs, SARE->getOperand(0));
767 } else if (!isa<SCEVConstant>(Expr) ||
768 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
770 SubExprs.push_back(Expr);
775 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
776 /// removing any common subexpressions from it. Anything truly common is
777 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
778 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
780 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
781 unsigned NumUses = Uses.size();
783 // Only one use? Use its base, regardless of what it is!
784 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
785 SCEVHandle Result = Zero;
787 std::swap(Result, Uses[0].Base);
791 // To find common subexpressions, count how many of Uses use each expression.
792 // If any subexpressions are used Uses.size() times, they are common.
793 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
795 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
796 // order we see them.
797 std::vector<SCEVHandle> UniqueSubExprs;
799 std::vector<SCEVHandle> SubExprs;
800 for (unsigned i = 0; i != NumUses; ++i) {
801 // If the base is zero (which is common), return zero now, there are no
803 if (Uses[i].Base == Zero) return Zero;
805 // Split the expression into subexprs.
806 SeparateSubExprs(SubExprs, Uses[i].Base);
807 // Add one to SubExpressionUseCounts for each subexpr present.
808 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
809 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
810 UniqueSubExprs.push_back(SubExprs[j]);
814 // Now that we know how many times each is used, build Result. Iterate over
815 // UniqueSubexprs so that we have a stable ordering.
816 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
817 std::map<SCEVHandle, unsigned>::iterator I =
818 SubExpressionUseCounts.find(UniqueSubExprs[i]);
819 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
820 if (I->second == NumUses) { // Found CSE!
821 Result = SCEVAddExpr::get(Result, I->first);
823 // Remove non-cse's from SubExpressionUseCounts.
824 SubExpressionUseCounts.erase(I);
828 // If we found no CSE's, return now.
829 if (Result == Zero) return Result;
831 // Otherwise, remove all of the CSE's we found from each of the base values.
832 for (unsigned i = 0; i != NumUses; ++i) {
833 // Split the expression into subexprs.
834 SeparateSubExprs(SubExprs, Uses[i].Base);
836 // Remove any common subexpressions.
837 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
838 if (SubExpressionUseCounts.count(SubExprs[j])) {
839 SubExprs.erase(SubExprs.begin()+j);
843 // Finally, the non-shared expressions together.
844 if (SubExprs.empty())
847 Uses[i].Base = SCEVAddExpr::get(SubExprs);
854 /// isZero - returns true if the scalar evolution expression is zero.
856 static bool isZero(SCEVHandle &V) {
857 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
858 return SC->getValue()->getRawValue() == 0;
863 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
864 /// of a previous stride and it is a legal value for the target addressing
865 /// mode scale component. This allows the users of this stride to be rewritten
866 /// as prev iv * factor. It returns 1 if no reuse is possible.
867 unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride,
872 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
873 unsigned SInt = SC->getValue()->getRawValue();
877 for (TargetLowering::legal_am_scale_iterator
878 I = TLI->legal_am_scale_begin(), E = TLI->legal_am_scale_end();
881 if (SInt >= Scale && (SInt % Scale) != 0)
883 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
884 IVsByStride.find(SCEVUnknown::getIntegerSCEV(SInt/Scale, Type::UIntTy));
885 if (SI == IVsByStride.end())
887 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
888 IE = SI->second.IVs.end(); II != IE; ++II)
889 // FIXME: Only handle base == 0 for now.
890 if (isZero(II->Base)) {
901 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
902 /// stride of IV. All of the users may have different starting values, and this
903 /// may not be the only stride (we know it is if isOnlyStride is true).
904 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
905 IVUsersOfOneStride &Uses,
908 // Transform our list of users and offsets to a bit more complex table. In
909 // this new vector, each 'BasedUser' contains 'Base' the base of the
910 // strided accessas well as the old information from Uses. We progressively
911 // move information from the Base field to the Imm field, until we eventually
912 // have the full access expression to rewrite the use.
913 std::vector<BasedUser> UsersToProcess;
914 UsersToProcess.reserve(Uses.Users.size());
915 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
916 UsersToProcess.push_back(Uses.Users[i]);
918 // Move any loop invariant operands from the offset field to the immediate
919 // field of the use, so that we don't try to use something before it is
921 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
922 UsersToProcess.back().Imm, L);
923 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
924 "Base value is not loop invariant!");
927 // Check if it is possible to reuse a IV with stride that is factor of this
928 // stride. And the multiple is a number that can be encoded in the scale
929 // field of the target addressing mode.
930 PHINode *NewPHI = NULL;
932 IVExpr ReuseIV(Stride, NULL, NULL);
933 unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV);
934 if (RewriteFactor > 1) {
935 NewPHI = ReuseIV.PHI;
939 // We now have a whole bunch of uses of like-strided induction variables, but
940 // they might all have different bases. We want to emit one PHI node for this
941 // stride which we fold as many common expressions (between the IVs) into as
942 // possible. Start by identifying the common expressions in the base values
943 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
944 // "A+B"), emit it to the preheader, then remove the expression from the
945 // UsersToProcess base values.
946 SCEVHandle CommonExprs =
947 RemoveCommonExpressionsFromUseBases(UsersToProcess);
949 // Next, figure out what we can represent in the immediate fields of
950 // instructions. If we can represent anything there, move it to the imm
951 // fields of the BasedUsers. We do this so that it increases the commonality
952 // of the remaining uses.
953 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
954 // If the user is not in the current loop, this means it is using the exit
955 // value of the IV. Do not put anything in the base, make sure it's all in
956 // the immediate field to allow as much factoring as possible.
957 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
958 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
959 UsersToProcess[i].Base);
960 UsersToProcess[i].Base =
961 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
964 // Addressing modes can be folded into loads and stores. Be careful that
965 // the store is through the expression, not of the expression though.
966 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
967 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
968 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
971 MoveImmediateValues(TLI, UsersToProcess[i].Base, UsersToProcess[i].Imm,
976 // Now that we know what we need to do, insert the PHI node itself.
978 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
979 << *CommonExprs << " :\n");
981 SCEVExpander Rewriter(*SE, *LI);
982 SCEVExpander PreheaderRewriter(*SE, *LI);
984 BasicBlock *Preheader = L->getLoopPreheader();
985 Instruction *PreInsertPt = Preheader->getTerminator();
986 Instruction *PhiInsertBefore = L->getHeader()->begin();
988 BasicBlock *LatchBlock = L->getLoopLatch();
990 const Type *ReplacedTy = CommonExprs->getType();
992 // Emit the initial base value into the loop preheader.
994 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
997 if (RewriteFactor == 1) {
998 // Create a new Phi for this base, and stick it in the loop header.
999 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1002 // Add common base to the new Phi node.
1003 NewPHI->addIncoming(CommonBaseV, Preheader);
1005 // Insert the stride into the preheader.
1006 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
1008 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1010 // Emit the increment of the base value before the terminator of the loop
1011 // latch block, and add it to the Phi node.
1012 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
1013 SCEVUnknown::get(StrideV));
1015 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
1017 IncV->setName(NewPHI->getName()+".inc");
1018 NewPHI->addIncoming(IncV, LatchBlock);
1020 // Remember this in case a later stride is multiple of this.
1021 IVsByStride[Stride].addIV(CommonExprs, NewPHI, IncV);
1023 Constant *C = dyn_cast<Constant>(CommonBaseV);
1025 (!C->isNullValue() &&
1026 !isTargetConstant(SCEVUnknown::get(CommonBaseV), TLI)))
1027 // We want the common base emitted into the preheader!
1028 CommonBaseV = new CastInst(CommonBaseV, CommonBaseV->getType(),
1029 "commonbase", PreInsertPt);
1032 // Sort by the base value, so that all IVs with identical bases are next to
1034 while (!UsersToProcess.empty()) {
1035 SCEVHandle Base = UsersToProcess.back().Base;
1037 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
1039 // Emit the code for Base into the preheader.
1040 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
1043 // If BaseV is a constant other than 0, make sure that it gets inserted into
1044 // the preheader, instead of being forward substituted into the uses. We do
1045 // this by forcing a noop cast to be inserted into the preheader in this
1047 if (Constant *C = dyn_cast<Constant>(BaseV))
1048 if (!C->isNullValue() && !isTargetConstant(Base, TLI)) {
1049 // We want this constant emitted into the preheader!
1050 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
1054 // Emit the code to add the immediate offset to the Phi value, just before
1055 // the instructions that we identified as using this stride and base.
1056 unsigned ScanPos = 0;
1058 BasedUser &User = UsersToProcess.back();
1060 // If this instruction wants to use the post-incremented value, move it
1061 // after the post-inc and use its value instead of the PHI.
1062 Value *RewriteOp = NewPHI;
1063 if (User.isUseOfPostIncrementedValue) {
1066 // If this user is in the loop, make sure it is the last thing in the
1067 // loop to ensure it is dominated by the increment.
1068 if (L->contains(User.Inst->getParent()))
1069 User.Inst->moveBefore(LatchBlock->getTerminator());
1071 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
1073 // Clear the SCEVExpander's expression map so that we are guaranteed
1074 // to have the code emitted where we expect it.
1077 // If we are reusing the iv, then it must be multiplied by a constant
1078 // factor take advantage of addressing mode scale component.
1079 if (RewriteFactor != 1) {
1081 SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
1082 RewriteExpr->getType()),
1085 // The common base is emitted in the loop preheader. But since we
1086 // are reusing an IV, it has not been used to initialize the PHI node.
1087 // Add it to the expression used to rewrite the uses.
1088 if (!isa<ConstantInt>(CommonBaseV) ||
1089 !cast<ConstantInt>(CommonBaseV)->isNullValue())
1090 RewriteExpr = SCEVAddExpr::get(RewriteExpr,
1091 SCEVUnknown::get(CommonBaseV));
1094 // Now that we know what we need to do, insert code before User for the
1095 // immediate and any loop-variant expressions.
1096 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
1097 // Add BaseV to the PHI value if needed.
1098 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
1100 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1102 // Mark old value we replaced as possibly dead, so that it is elminated
1103 // if we just replaced the last use of that value.
1104 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1106 UsersToProcess.pop_back();
1109 // If there are any more users to process with the same base, move one of
1110 // them to the end of the list so that we will process it.
1111 if (!UsersToProcess.empty()) {
1112 for (unsigned e = UsersToProcess.size(); ScanPos != e; ++ScanPos)
1113 if (UsersToProcess[ScanPos].Base == Base) {
1114 std::swap(UsersToProcess[ScanPos], UsersToProcess.back());
1118 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1119 // TODO: Next, find out which base index is the most common, pull it out.
1122 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1123 // different starting values, into different PHIs.
1126 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1127 // uses in the loop, look to see if we can eliminate some, in favor of using
1128 // common indvars for the different uses.
1129 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1130 // TODO: implement optzns here.
1135 // Finally, get the terminating condition for the loop if possible. If we
1136 // can, we want to change it to use a post-incremented version of its
1137 // induction variable, to allow coallescing the live ranges for the IV into
1138 // one register value.
1139 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1140 BasicBlock *Preheader = L->getLoopPreheader();
1141 BasicBlock *LatchBlock =
1142 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1143 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1144 if (!TermBr || TermBr->isUnconditional() ||
1145 !isa<SetCondInst>(TermBr->getCondition()))
1147 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
1149 // Search IVUsesByStride to find Cond's IVUse if there is one.
1150 IVStrideUse *CondUse = 0;
1151 const SCEVHandle *CondStride = 0;
1153 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1155 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1156 IVUsesByStride.find(StrideOrder[Stride]);
1157 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1159 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1160 E = SI->second.Users.end(); UI != E; ++UI)
1161 if (UI->User == Cond) {
1163 CondStride = &SI->first;
1164 // NOTE: we could handle setcc instructions with multiple uses here, but
1165 // InstCombine does it as well for simple uses, it's not clear that it
1166 // occurs enough in real life to handle.
1170 if (!CondUse) return; // setcc doesn't use the IV.
1172 // setcc stride is complex, don't mess with users.
1173 // FIXME: Evaluate whether this is a good idea or not.
1174 if (!isa<SCEVConstant>(*CondStride)) return;
1176 // It's possible for the setcc instruction to be anywhere in the loop, and
1177 // possible for it to have multiple users. If it is not immediately before
1178 // the latch block branch, move it.
1179 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1180 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1181 Cond->moveBefore(TermBr);
1183 // Otherwise, clone the terminating condition and insert into the loopend.
1184 Cond = cast<SetCondInst>(Cond->clone());
1185 Cond->setName(L->getHeader()->getName() + ".termcond");
1186 LatchBlock->getInstList().insert(TermBr, Cond);
1188 // Clone the IVUse, as the old use still exists!
1189 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1190 CondUse->OperandValToReplace);
1191 CondUse = &IVUsesByStride[*CondStride].Users.back();
1195 // If we get to here, we know that we can transform the setcc instruction to
1196 // use the post-incremented version of the IV, allowing us to coallesce the
1197 // live ranges for the IV correctly.
1198 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1199 CondUse->isUseOfPostIncrementedValue = true;
1202 void LoopStrengthReduce::runOnLoop(Loop *L) {
1203 // First step, transform all loops nesting inside of this loop.
1204 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
1207 // Next, find all uses of induction variables in this loop, and catagorize
1208 // them by stride. Start by finding all of the PHI nodes in the header for
1209 // this loop. If they are induction variables, inspect their uses.
1210 std::set<Instruction*> Processed; // Don't reprocess instructions.
1211 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1212 AddUsersIfInteresting(I, L, Processed);
1214 // If we have nothing to do, return.
1215 if (IVUsesByStride.empty()) return;
1217 // Optimize induction variables. Some indvar uses can be transformed to use
1218 // strides that will be needed for other purposes. A common example of this
1219 // is the exit test for the loop, which can often be rewritten to use the
1220 // computation of some other indvar to decide when to terminate the loop.
1224 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1225 // doing computation in byte values, promote to 32-bit values if safe.
1227 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1228 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1229 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1230 // to be careful that IV's are all the same type. Only works for intptr_t
1233 // If we only have one stride, we can more aggressively eliminate some things.
1234 bool HasOneStride = IVUsesByStride.size() == 1;
1237 DEBUG(std::cerr << "\nLSR on ");
1241 // IVsByStride keeps IVs for one particular loop.
1242 IVsByStride.clear();
1244 // Note: this processes each stride/type pair individually. All users passed
1245 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1246 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1247 // This extra layer of indirection makes the ordering of strides deterministic
1248 // - not dependent on map order.
1249 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1250 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1251 IVUsesByStride.find(StrideOrder[Stride]);
1252 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1253 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1256 // Clean up after ourselves
1257 if (!DeadInsts.empty()) {
1258 DeleteTriviallyDeadInstructions(DeadInsts);
1260 BasicBlock::iterator I = L->getHeader()->begin();
1262 while ((PN = dyn_cast<PHINode>(I))) {
1263 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1265 // At this point, we know that we have killed one or more GEP
1266 // instructions. It is worth checking to see if the cann indvar is also
1267 // dead, so that we can remove it as well. The requirements for the cann
1268 // indvar to be considered dead are:
1269 // 1. the cann indvar has one use
1270 // 2. the use is an add instruction
1271 // 3. the add has one use
1272 // 4. the add is used by the cann indvar
1273 // If all four cases above are true, then we can remove both the add and
1275 // FIXME: this needs to eliminate an induction variable even if it's being
1276 // compared against some value to decide loop termination.
1277 if (PN->hasOneUse()) {
1278 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1279 if (BO && BO->hasOneUse()) {
1280 if (PN == *(BO->use_begin())) {
1281 DeadInsts.insert(BO);
1282 // Break the cycle, then delete the PHI.
1283 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1284 SE->deleteInstructionFromRecords(PN);
1285 PN->eraseFromParent();
1290 DeleteTriviallyDeadInstructions(DeadInsts);
1293 CastedPointers.clear();
1294 IVUsesByStride.clear();
1295 StrideOrder.clear();