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/Local.h"
30 #include "llvm/Target/TargetData.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Support/Debug.h"
38 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
39 Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
41 /// IVStrideUse - Keep track of one use of a strided induction variable, where
42 /// the stride is stored externally. The Offset member keeps track of the
43 /// offset from the IV, User is the actual user of the operand, and 'Operand'
44 /// is the operand # of the User that is the use.
48 Value *OperandValToReplace;
50 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
51 : Offset(Offs), User(U), OperandValToReplace(O) {}
54 /// IVUsersOfOneStride - This structure keeps track of all instructions that
55 /// have an operand that is based on the trip count multiplied by some stride.
56 /// The stride for all of these users is common and kept external to this
58 struct IVUsersOfOneStride {
59 /// Users - Keep track of all of the users of this stride as well as the
60 /// initial value and the operand that uses the IV.
61 std::vector<IVStrideUse> Users;
63 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
64 Users.push_back(IVStrideUse(Offset, User, Operand));
69 class LoopStrengthReduce : public FunctionPass {
74 const Type *UIntPtrTy;
77 /// MaxTargetAMSize - This is the maximum power-of-two scale value that the
78 /// target can handle for free with its addressing modes.
79 unsigned MaxTargetAMSize;
81 /// IVUsesByStride - Keep track of all uses of induction variables that we
82 /// are interested in. The key of the map is the stride of the access.
83 std::map<Value*, IVUsersOfOneStride> IVUsesByStride;
85 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
86 /// of the casted version of each value. This is accessed by
87 /// getCastedVersionOf.
88 std::map<Value*, Value*> CastedPointers;
90 /// DeadInsts - Keep track of instructions we may have made dead, so that
91 /// we can remove them after we are done working.
92 std::set<Instruction*> DeadInsts;
94 LoopStrengthReduce(unsigned MTAMS = 1)
95 : MaxTargetAMSize(MTAMS) {
98 virtual bool runOnFunction(Function &) {
99 LI = &getAnalysis<LoopInfo>();
100 DS = &getAnalysis<DominatorSet>();
101 SE = &getAnalysis<ScalarEvolution>();
102 TD = &getAnalysis<TargetData>();
103 UIntPtrTy = TD->getIntPtrType();
106 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
112 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
113 AU.setPreservesCFG();
114 AU.addRequiredID(LoopSimplifyID);
115 AU.addRequired<LoopInfo>();
116 AU.addRequired<DominatorSet>();
117 AU.addRequired<TargetData>();
118 AU.addRequired<ScalarEvolution>();
121 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
123 Value *getCastedVersionOf(Value *V);
125 void runOnLoop(Loop *L);
126 bool AddUsersIfInteresting(Instruction *I, Loop *L,
127 std::set<Instruction*> &Processed);
128 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
131 void StrengthReduceStridedIVUsers(Value *Stride, IVUsersOfOneStride &Uses,
132 Loop *L, bool isOnlyStride);
133 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
135 RegisterOpt<LoopStrengthReduce> X("loop-reduce",
136 "Strength Reduce GEP Uses of Ind. Vars");
139 FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize) {
140 return new LoopStrengthReduce(MaxTargetAMSize);
143 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
145 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
146 if (V->getType() == UIntPtrTy) return V;
147 if (Constant *CB = dyn_cast<Constant>(V))
148 return ConstantExpr::getCast(CB, UIntPtrTy);
150 Value *&New = CastedPointers[V];
153 BasicBlock::iterator InsertPt;
154 if (Argument *Arg = dyn_cast<Argument>(V)) {
155 // Insert into the entry of the function, after any allocas.
156 InsertPt = Arg->getParent()->begin()->begin();
157 while (isa<AllocaInst>(InsertPt)) ++InsertPt;
159 if (InvokeInst *II = dyn_cast<InvokeInst>(V)) {
160 InsertPt = II->getNormalDest()->begin();
162 InsertPt = cast<Instruction>(V);
166 // Do not insert casts into the middle of PHI node blocks.
167 while (isa<PHINode>(InsertPt)) ++InsertPt;
170 New = new CastInst(V, UIntPtrTy, V->getName(), InsertPt);
171 DeadInsts.insert(cast<Instruction>(New));
176 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
177 /// specified set are trivially dead, delete them and see if this makes any of
178 /// their operands subsequently dead.
179 void LoopStrengthReduce::
180 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
181 while (!Insts.empty()) {
182 Instruction *I = *Insts.begin();
183 Insts.erase(Insts.begin());
184 if (isInstructionTriviallyDead(I)) {
185 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
186 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
188 SE->deleteInstructionFromRecords(I);
189 I->eraseFromParent();
196 /// GetExpressionSCEV - Compute and return the SCEV for the specified
198 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
199 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
201 return SE->getSCEV(Exp);
203 // Analyze all of the subscripts of this getelementptr instruction, looking
204 // for uses that are determined by the trip count of L. First, skip all
205 // operands the are not dependent on the IV.
207 // Build up the base expression. Insert an LLVM cast of the pointer to
209 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
211 gep_type_iterator GTI = gep_type_begin(GEP);
213 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
214 // If this is a use of a recurrence that we can analyze, and it comes before
215 // Op does in the GEP operand list, we will handle this when we process this
217 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
218 const StructLayout *SL = TD->getStructLayout(STy);
219 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
220 uint64_t Offset = SL->MemberOffsets[Idx];
221 GEPVal = SCEVAddExpr::get(GEPVal,
222 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
224 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
225 SCEVHandle Idx = SE->getSCEV(OpVal);
227 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
229 Idx = SCEVMulExpr::get(Idx,
230 SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
232 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
239 /// getSCEVStartAndStride - Compute the start and stride of this expression,
240 /// returning false if the expression is not a start/stride pair, or true if it
241 /// is. The stride must be a loop invariant expression, but the start may be
242 /// a mix of loop invariant and loop variant expressions.
243 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
244 SCEVHandle &Start, Value *&Stride) {
245 SCEVHandle TheAddRec = Start; // Initialize to zero.
247 // If the outer level is an AddExpr, the operands are all start values except
248 // for a nested AddRecExpr.
249 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
250 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
251 if (SCEVAddRecExpr *AddRec =
252 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
253 if (AddRec->getLoop() == L)
254 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
256 return false; // Nested IV of some sort?
258 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
261 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
264 return false; // not analyzable.
267 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
268 if (!AddRec || AddRec->getLoop() != L) return false;
270 // FIXME: Generalize to non-affine IV's.
271 if (!AddRec->isAffine()) return false;
273 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
275 // FIXME: generalize to IV's with more complex strides (must emit stride
276 // expression outside of loop!)
277 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
280 SCEVConstant *StrideC = cast<SCEVConstant>(AddRec->getOperand(1));
281 Stride = StrideC->getValue();
283 assert(Stride->getType()->isUnsigned() &&
284 "Constants should be canonicalized to unsigned!");
288 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
289 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
290 /// return true. Otherwise, return false.
291 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
292 std::set<Instruction*> &Processed) {
293 if (I->getType() == Type::VoidTy) return false;
294 if (!Processed.insert(I).second)
295 return true; // Instruction already handled.
297 // Get the symbolic expression for this instruction.
298 SCEVHandle ISE = GetExpressionSCEV(I, L);
299 if (isa<SCEVCouldNotCompute>(ISE)) return false;
301 // Get the start and stride for this expression.
302 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
304 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
305 return false; // Non-reducible symbolic expression, bail out.
307 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
308 Instruction *User = cast<Instruction>(*UI);
310 // Do not infinitely recurse on PHI nodes.
311 if (isa<PHINode>(User) && User->getParent() == L->getHeader())
314 // If this is an instruction defined in a nested loop, or outside this loop,
315 // don't recurse into it.
316 bool AddUserToIVUsers = false;
317 if (LI->getLoopFor(User->getParent()) != L) {
318 DEBUG(std::cerr << "FOUND USER in nested loop: " << *User
319 << " OF SCEV: " << *ISE << "\n");
320 AddUserToIVUsers = true;
321 } else if (!AddUsersIfInteresting(User, L, Processed)) {
322 DEBUG(std::cerr << "FOUND USER: " << *User
323 << " OF SCEV: " << *ISE << "\n");
324 AddUserToIVUsers = true;
327 if (AddUserToIVUsers) {
328 // Okay, we found a user that we cannot reduce. Analyze the instruction
329 // and decide what to do with it.
330 IVUsesByStride[Stride].addUser(Start, User, I);
337 /// BasedUser - For a particular base value, keep information about how we've
338 /// partitioned the expression so far.
340 /// Inst - The instruction using the induction variable.
343 /// OperandValToReplace - The operand value of Inst to replace with the
345 Value *OperandValToReplace;
347 /// Imm - The immediate value that should be added to the base immediately
348 /// before Inst, because it will be folded into the imm field of the
352 /// EmittedBase - The actual value* to use for the base value of this
353 /// operation. This is null if we should just use zero so far.
356 BasedUser(Instruction *I, Value *Op, const SCEVHandle &IMM)
357 : Inst(I), OperandValToReplace(Op), Imm(IMM), EmittedBase(0) {}
359 // Once we rewrite the code to insert the new IVs we want, update the
360 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
362 void RewriteInstructionToUseNewBase(Value *NewBase, SCEVExpander &Rewriter);
364 // No need to compare these.
365 bool operator<(const BasedUser &BU) const { return 0; }
371 void BasedUser::dump() const {
372 std::cerr << " Imm=" << *Imm;
374 std::cerr << " EB=" << *EmittedBase;
376 std::cerr << " Inst: " << *Inst;
379 // Once we rewrite the code to insert the new IVs we want, update the
380 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
382 void BasedUser::RewriteInstructionToUseNewBase(Value *NewBase,
383 SCEVExpander &Rewriter) {
384 if (!isa<PHINode>(Inst)) {
385 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(NewBase), Imm);
386 Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, Inst,
387 OperandValToReplace->getType());
389 // Replace the use of the operand Value with the new Phi we just created.
390 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
391 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
395 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
396 // expression into each operand block that uses it.
397 PHINode *PN = cast<PHINode>(Inst);
398 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
399 if (PN->getIncomingValue(i) == OperandValToReplace) {
400 // FIXME: this should split any critical edges.
402 // Insert the code into the end of the predecessor block.
403 BasicBlock::iterator InsertPt = PN->getIncomingBlock(i)->getTerminator();
405 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(NewBase), Imm);
406 Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, InsertPt,
407 OperandValToReplace->getType());
409 // Replace the use of the operand Value with the new Phi we just created.
410 PN->setIncomingValue(i, NewVal);
414 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
418 /// isTargetConstant - Return true if the following can be referenced by the
419 /// immediate field of a target instruction.
420 static bool isTargetConstant(const SCEVHandle &V) {
422 // FIXME: Look at the target to decide if &GV is a legal constant immediate.
423 if (isa<SCEVConstant>(V)) return true;
425 return false; // ENABLE this for x86
427 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
428 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
429 if (CE->getOpcode() == Instruction::Cast)
430 if (isa<GlobalValue>(CE->getOperand(0)))
431 // FIXME: should check to see that the dest is uintptr_t!
436 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
437 /// that can fit into the immediate field of instructions in the target.
438 /// Accumulate these immediate values into the Imm value.
439 static void MoveImmediateValues(SCEVHandle &Val, SCEVHandle &Imm,
440 bool isAddress, Loop *L) {
441 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
442 std::vector<SCEVHandle> NewOps;
443 NewOps.reserve(SAE->getNumOperands());
445 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
446 if (isAddress && isTargetConstant(SAE->getOperand(i))) {
447 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
448 } else if (!SAE->getOperand(i)->isLoopInvariant(L)) {
449 // If this is a loop-variant expression, it must stay in the immediate
450 // field of the expression.
451 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
453 NewOps.push_back(SAE->getOperand(i));
457 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
459 Val = SCEVAddExpr::get(NewOps);
461 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
462 // Try to pull immediates out of the start value of nested addrec's.
463 SCEVHandle Start = SARE->getStart();
464 MoveImmediateValues(Start, Imm, isAddress, L);
466 if (Start != SARE->getStart()) {
467 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
469 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
474 // Loop-variant expressions must stay in the immediate field of the
476 if ((isAddress && isTargetConstant(Val)) ||
477 !Val->isLoopInvariant(L)) {
478 Imm = SCEVAddExpr::get(Imm, Val);
479 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
483 // Otherwise, no immediates to move.
486 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
487 /// stride of IV. All of the users may have different starting values, and this
488 /// may not be the only stride (we know it is if isOnlyStride is true).
489 void LoopStrengthReduce::StrengthReduceStridedIVUsers(Value *Stride,
490 IVUsersOfOneStride &Uses,
493 // Transform our list of users and offsets to a bit more complex table. In
494 // this new vector, the first entry for each element is the base of the
495 // strided access, and the second is the BasedUser object for the use. We
496 // progressively move information from the first to the second entry, until we
497 // eventually emit the object.
498 std::vector<std::pair<SCEVHandle, BasedUser> > UsersToProcess;
499 UsersToProcess.reserve(Uses.Users.size());
501 SCEVHandle ZeroBase = SCEVUnknown::getIntegerSCEV(0,
502 Uses.Users[0].Offset->getType());
504 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i)
505 UsersToProcess.push_back(std::make_pair(Uses.Users[i].Offset,
506 BasedUser(Uses.Users[i].User,
507 Uses.Users[i].OperandValToReplace,
510 // First pass, figure out what we can represent in the immediate fields of
511 // instructions. If we can represent anything there, move it to the imm
512 // fields of the BasedUsers.
513 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
514 // Addressing modes can be folded into loads and stores. Be careful that
515 // the store is through the expression, not of the expression though.
516 bool isAddress = isa<LoadInst>(UsersToProcess[i].second.Inst);
517 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].second.Inst))
518 if (SI->getOperand(1) == UsersToProcess[i].second.OperandValToReplace)
521 MoveImmediateValues(UsersToProcess[i].first, UsersToProcess[i].second.Imm,
524 assert(UsersToProcess[i].first->isLoopInvariant(L) &&
525 "Base value is not loop invariant!");
528 SCEVExpander Rewriter(*SE, *LI);
529 BasicBlock *Preheader = L->getLoopPreheader();
530 Instruction *PreInsertPt = Preheader->getTerminator();
531 Instruction *PhiInsertBefore = L->getHeader()->begin();
533 assert(isa<PHINode>(PhiInsertBefore) &&
534 "How could this loop have IV's without any phis?");
535 PHINode *SomeLoopPHI = cast<PHINode>(PhiInsertBefore);
536 assert(SomeLoopPHI->getNumIncomingValues() == 2 &&
537 "This loop isn't canonicalized right");
538 BasicBlock *LatchBlock =
539 SomeLoopPHI->getIncomingBlock(SomeLoopPHI->getIncomingBlock(0) == Preheader);
541 DEBUG(std::cerr << "INSERTING IVs of STRIDE " << *Stride << ":\n");
543 // FIXME: This loop needs increasing levels of intelligence.
544 // STAGE 0: just emit everything as its own base.
545 // STAGE 1: factor out common vars from bases, and try and push resulting
546 // constants into Imm field. <-- We are here
547 // STAGE 2: factor out large constants to try and make more constants
548 // acceptable for target loads and stores.
550 // Sort by the base value, so that all IVs with identical bases are next to
552 std::sort(UsersToProcess.begin(), UsersToProcess.end());
553 while (!UsersToProcess.empty()) {
554 SCEVHandle Base = UsersToProcess.front().first;
556 DEBUG(std::cerr << " INSERTING PHI with BASE = " << *Base << ":\n");
558 // Create a new Phi for this base, and stick it in the loop header.
559 const Type *ReplacedTy = Base->getType();
560 PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
563 // Emit the initial base value into the loop preheader, and add it to the
565 Value *BaseV = Rewriter.expandCodeFor(Base, PreInsertPt, ReplacedTy);
566 NewPHI->addIncoming(BaseV, Preheader);
568 // Emit the increment of the base value before the terminator of the loop
569 // latch block, and add it to the Phi node.
570 SCEVHandle Inc = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
571 SCEVUnknown::get(Stride));
573 Value *IncV = Rewriter.expandCodeFor(Inc, LatchBlock->getTerminator(),
575 IncV->setName(NewPHI->getName()+".inc");
576 NewPHI->addIncoming(IncV, LatchBlock);
578 // Emit the code to add the immediate offset to the Phi value, just before
579 // the instructions that we identified as using this stride and base.
580 while (!UsersToProcess.empty() && UsersToProcess.front().first == Base) {
581 BasedUser &User = UsersToProcess.front().second;
583 // Clear the SCEVExpander's expression map so that we are guaranteed
584 // to have the code emitted where we expect it.
587 // Now that we know what we need to do, insert code before User for the
588 // immediate and any loop-variant expressions.
589 User.RewriteInstructionToUseNewBase(NewPHI, Rewriter);
591 // Mark old value we replaced as possibly dead, so that it is elminated
592 // if we just replaced the last use of that value.
593 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
595 UsersToProcess.erase(UsersToProcess.begin());
598 // TODO: Next, find out which base index is the most common, pull it out.
601 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
602 // different starting values, into different PHIs.
606 void LoopStrengthReduce::runOnLoop(Loop *L) {
607 // First step, transform all loops nesting inside of this loop.
608 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
611 // Next, find all uses of induction variables in this loop, and catagorize
612 // them by stride. Start by finding all of the PHI nodes in the header for
613 // this loop. If they are induction variables, inspect their uses.
614 std::set<Instruction*> Processed; // Don't reprocess instructions.
615 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
616 AddUsersIfInteresting(I, L, Processed);
618 // If we have nothing to do, return.
619 //if (IVUsesByStride.empty()) return;
621 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
622 // doing computation in byte values, promote to 32-bit values if safe.
624 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
625 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
626 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
627 // to be careful that IV's are all the same type. Only works for intptr_t
630 // If we only have one stride, we can more aggressively eliminate some things.
631 bool HasOneStride = IVUsesByStride.size() == 1;
633 for (std::map<Value*, IVUsersOfOneStride>::iterator SI
634 = IVUsesByStride.begin(), E = IVUsesByStride.end(); SI != E; ++SI)
635 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
637 // Clean up after ourselves
638 if (!DeadInsts.empty()) {
639 DeleteTriviallyDeadInstructions(DeadInsts);
641 BasicBlock::iterator I = L->getHeader()->begin();
643 while ((PN = dyn_cast<PHINode>(I))) {
644 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
646 // At this point, we know that we have killed one or more GEP instructions.
647 // It is worth checking to see if the cann indvar is also dead, so that we
648 // can remove it as well. The requirements for the cann indvar to be
649 // considered dead are:
650 // 1. the cann indvar has one use
651 // 2. the use is an add instruction
652 // 3. the add has one use
653 // 4. the add is used by the cann indvar
654 // If all four cases above are true, then we can remove both the add and
656 // FIXME: this needs to eliminate an induction variable even if it's being
657 // compared against some value to decide loop termination.
658 if (PN->hasOneUse()) {
659 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
660 if (BO && BO->hasOneUse()) {
661 if (PN == *(BO->use_begin())) {
662 DeadInsts.insert(BO);
663 // Break the cycle, then delete the PHI.
664 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
665 SE->deleteInstructionFromRecords(PN);
666 PN->eraseFromParent();
671 DeleteTriviallyDeadInstructions(DeadInsts);
674 CastedPointers.clear();
675 IVUsesByStride.clear();