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/Support/Compiler.h"
35 #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 stride, the common base, as
81 /// well as the PHI node and increment value created for rewrite.
89 : Stride(SCEVUnknown::getIntegerSCEV(0, Type::UIntTy)),
90 Base (SCEVUnknown::getIntegerSCEV(0, Type::UIntTy)) {}
91 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
93 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
96 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
97 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
98 struct IVsOfOneStride {
99 std::vector<IVExpr> IVs;
101 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
103 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
107 class VISIBILITY_HIDDEN LoopStrengthReduce : public FunctionPass {
111 const TargetData *TD;
112 const Type *UIntPtrTy;
115 /// IVUsesByStride - Keep track of all uses of induction variables that we
116 /// are interested in. The key of the map is the stride of the access.
117 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
119 /// IVsByStride - Keep track of all IVs that have been inserted for a
120 /// particular stride.
121 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
123 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
124 /// We use this to iterate over the IVUsesByStride collection without being
125 /// dependent on random ordering of pointers in the process.
126 std::vector<SCEVHandle> StrideOrder;
128 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
129 /// of the casted version of each value. This is accessed by
130 /// getCastedVersionOf.
131 std::map<Value*, Value*> CastedPointers;
133 /// DeadInsts - Keep track of instructions we may have made dead, so that
134 /// we can remove them after we are done working.
135 std::set<Instruction*> DeadInsts;
137 /// TLI - Keep a pointer of a TargetLowering to consult for determining
138 /// transformation profitability.
139 const TargetLowering *TLI;
142 LoopStrengthReduce(const TargetLowering *tli = NULL)
146 virtual bool runOnFunction(Function &) {
147 LI = &getAnalysis<LoopInfo>();
148 EF = &getAnalysis<ETForest>();
149 SE = &getAnalysis<ScalarEvolution>();
150 TD = &getAnalysis<TargetData>();
151 UIntPtrTy = TD->getIntPtrType();
154 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
160 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
161 // We split critical edges, so we change the CFG. However, we do update
162 // many analyses if they are around.
163 AU.addPreservedID(LoopSimplifyID);
164 AU.addPreserved<LoopInfo>();
165 AU.addPreserved<DominatorSet>();
166 AU.addPreserved<ETForest>();
167 AU.addPreserved<ImmediateDominators>();
168 AU.addPreserved<DominanceFrontier>();
169 AU.addPreserved<DominatorTree>();
171 AU.addRequiredID(LoopSimplifyID);
172 AU.addRequired<LoopInfo>();
173 AU.addRequired<ETForest>();
174 AU.addRequired<TargetData>();
175 AU.addRequired<ScalarEvolution>();
178 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
180 Value *getCastedVersionOf(Value *V);
182 void runOnLoop(Loop *L);
183 bool AddUsersIfInteresting(Instruction *I, Loop *L,
184 std::set<Instruction*> &Processed);
185 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
187 void OptimizeIndvars(Loop *L);
189 unsigned CheckForIVReuse(const SCEVHandle&, IVExpr&, const Type*);
191 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
192 IVUsersOfOneStride &Uses,
193 Loop *L, bool isOnlyStride);
194 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
196 RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
199 FunctionPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
200 return new LoopStrengthReduce(TLI);
203 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
205 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
206 if (V->getType() == UIntPtrTy) return V;
207 if (Constant *CB = dyn_cast<Constant>(V))
208 return ConstantExpr::getCast(CB, UIntPtrTy);
210 Value *&New = CastedPointers[V];
213 New = SCEVExpander::InsertCastOfTo(V, UIntPtrTy);
214 DeadInsts.insert(cast<Instruction>(New));
219 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
220 /// specified set are trivially dead, delete them and see if this makes any of
221 /// their operands subsequently dead.
222 void LoopStrengthReduce::
223 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
224 while (!Insts.empty()) {
225 Instruction *I = *Insts.begin();
226 Insts.erase(Insts.begin());
227 if (isInstructionTriviallyDead(I)) {
228 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
229 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
231 SE->deleteInstructionFromRecords(I);
232 I->eraseFromParent();
239 /// GetExpressionSCEV - Compute and return the SCEV for the specified
241 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
242 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
243 // If this is a GEP that SE doesn't know about, compute it now and insert it.
244 // If this is not a GEP, or if we have already done this computation, just let
246 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
247 if (!GEP || SE->hasSCEV(GEP))
248 return SE->getSCEV(Exp);
250 // Analyze all of the subscripts of this getelementptr instruction, looking
251 // for uses that are determined by the trip count of L. First, skip all
252 // operands the are not dependent on the IV.
254 // Build up the base expression. Insert an LLVM cast of the pointer to
256 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
258 gep_type_iterator GTI = gep_type_begin(GEP);
260 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
261 // If this is a use of a recurrence that we can analyze, and it comes before
262 // Op does in the GEP operand list, we will handle this when we process this
264 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
265 const StructLayout *SL = TD->getStructLayout(STy);
266 unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
267 uint64_t Offset = SL->MemberOffsets[Idx];
268 GEPVal = SCEVAddExpr::get(GEPVal,
269 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
271 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
272 SCEVHandle Idx = SE->getSCEV(OpVal);
274 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
276 Idx = SCEVMulExpr::get(Idx,
277 SCEVConstant::get(ConstantInt::get(UIntPtrTy,
279 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
283 SE->setSCEV(GEP, GEPVal);
287 /// getSCEVStartAndStride - Compute the start and stride of this expression,
288 /// returning false if the expression is not a start/stride pair, or true if it
289 /// is. The stride must be a loop invariant expression, but the start may be
290 /// a mix of loop invariant and loop variant expressions.
291 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
292 SCEVHandle &Start, SCEVHandle &Stride) {
293 SCEVHandle TheAddRec = Start; // Initialize to zero.
295 // If the outer level is an AddExpr, the operands are all start values except
296 // for a nested AddRecExpr.
297 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
298 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
299 if (SCEVAddRecExpr *AddRec =
300 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
301 if (AddRec->getLoop() == L)
302 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
304 return false; // Nested IV of some sort?
306 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
309 } else if (isa<SCEVAddRecExpr>(SH)) {
312 return false; // not analyzable.
315 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
316 if (!AddRec || AddRec->getLoop() != L) return false;
318 // FIXME: Generalize to non-affine IV's.
319 if (!AddRec->isAffine()) return false;
321 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
323 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
324 DOUT << "[" << L->getHeader()->getName()
325 << "] Variable stride: " << *AddRec << "\n";
327 Stride = AddRec->getOperand(1);
328 // Check that all constant strides are the unsigned type, we don't want to
329 // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
331 assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
332 "Constants should be canonicalized to unsigned!");
337 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
338 /// and now we need to decide whether the user should use the preinc or post-inc
339 /// value. If this user should use the post-inc version of the IV, return true.
341 /// Choosing wrong here can break dominance properties (if we choose to use the
342 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
343 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
344 /// should use the post-inc value).
345 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
346 Loop *L, ETForest *EF, Pass *P) {
347 // If the user is in the loop, use the preinc value.
348 if (L->contains(User->getParent())) return false;
350 BasicBlock *LatchBlock = L->getLoopLatch();
352 // Ok, the user is outside of the loop. If it is dominated by the latch
353 // block, use the post-inc value.
354 if (EF->dominates(LatchBlock, User->getParent()))
357 // There is one case we have to be careful of: PHI nodes. These little guys
358 // can live in blocks that do not dominate the latch block, but (since their
359 // uses occur in the predecessor block, not the block the PHI lives in) should
360 // still use the post-inc value. Check for this case now.
361 PHINode *PN = dyn_cast<PHINode>(User);
362 if (!PN) return false; // not a phi, not dominated by latch block.
364 // Look at all of the uses of IV by the PHI node. If any use corresponds to
365 // a block that is not dominated by the latch block, give up and use the
366 // preincremented value.
367 unsigned NumUses = 0;
368 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
369 if (PN->getIncomingValue(i) == IV) {
371 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
375 // Okay, all uses of IV by PN are in predecessor blocks that really are
376 // dominated by the latch block. Split the critical edges and use the
377 // post-incremented value.
378 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
379 if (PN->getIncomingValue(i) == IV) {
380 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P,
382 // Splitting the critical edge can reduce the number of entries in this
384 e = PN->getNumIncomingValues();
385 if (--NumUses == 0) break;
393 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
394 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
395 /// return true. Otherwise, return false.
396 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
397 std::set<Instruction*> &Processed) {
398 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
399 return false; // Void and FP expressions cannot be reduced.
400 if (!Processed.insert(I).second)
401 return true; // Instruction already handled.
403 // Get the symbolic expression for this instruction.
404 SCEVHandle ISE = GetExpressionSCEV(I, L);
405 if (isa<SCEVCouldNotCompute>(ISE)) return false;
407 // Get the start and stride for this expression.
408 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
409 SCEVHandle Stride = Start;
410 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
411 return false; // Non-reducible symbolic expression, bail out.
413 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
414 Instruction *User = cast<Instruction>(*UI);
416 // Do not infinitely recurse on PHI nodes.
417 if (isa<PHINode>(User) && Processed.count(User))
420 // If this is an instruction defined in a nested loop, or outside this loop,
421 // don't recurse into it.
422 bool AddUserToIVUsers = false;
423 if (LI->getLoopFor(User->getParent()) != L) {
424 DOUT << "FOUND USER in other loop: " << *User
425 << " OF SCEV: " << *ISE << "\n";
426 AddUserToIVUsers = true;
427 } else if (!AddUsersIfInteresting(User, L, Processed)) {
428 DOUT << "FOUND USER: " << *User
429 << " OF SCEV: " << *ISE << "\n";
430 AddUserToIVUsers = true;
433 if (AddUserToIVUsers) {
434 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
435 if (StrideUses.Users.empty()) // First occurance of this stride?
436 StrideOrder.push_back(Stride);
438 // Okay, we found a user that we cannot reduce. Analyze the instruction
439 // and decide what to do with it. If we are a use inside of the loop, use
440 // the value before incrementation, otherwise use it after incrementation.
441 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
442 // The value used will be incremented by the stride more than we are
443 // expecting, so subtract this off.
444 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
445 StrideUses.addUser(NewStart, User, I);
446 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
447 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
449 StrideUses.addUser(Start, User, I);
457 /// BasedUser - For a particular base value, keep information about how we've
458 /// partitioned the expression so far.
460 /// Base - The Base value for the PHI node that needs to be inserted for
461 /// this use. As the use is processed, information gets moved from this
462 /// field to the Imm field (below). BasedUser values are sorted by this
466 /// Inst - The instruction using the induction variable.
469 /// OperandValToReplace - The operand value of Inst to replace with the
471 Value *OperandValToReplace;
473 /// Imm - The immediate value that should be added to the base immediately
474 /// before Inst, because it will be folded into the imm field of the
478 /// EmittedBase - The actual value* to use for the base value of this
479 /// operation. This is null if we should just use zero so far.
482 // isUseOfPostIncrementedValue - True if this should use the
483 // post-incremented version of this IV, not the preincremented version.
484 // This can only be set in special cases, such as the terminating setcc
485 // instruction for a loop and uses outside the loop that are dominated by
487 bool isUseOfPostIncrementedValue;
489 BasedUser(IVStrideUse &IVSU)
490 : Base(IVSU.Offset), Inst(IVSU.User),
491 OperandValToReplace(IVSU.OperandValToReplace),
492 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
493 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
495 // Once we rewrite the code to insert the new IVs we want, update the
496 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
498 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
499 SCEVExpander &Rewriter, Loop *L,
502 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
503 SCEVExpander &Rewriter,
504 Instruction *IP, Loop *L);
509 void BasedUser::dump() const {
510 llvm_cerr << " Base=" << *Base;
511 llvm_cerr << " Imm=" << *Imm;
513 llvm_cerr << " EB=" << *EmittedBase;
515 llvm_cerr << " Inst: " << *Inst;
518 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
519 SCEVExpander &Rewriter,
520 Instruction *IP, Loop *L) {
521 // Figure out where we *really* want to insert this code. In particular, if
522 // the user is inside of a loop that is nested inside of L, we really don't
523 // want to insert this expression before the user, we'd rather pull it out as
524 // many loops as possible.
525 LoopInfo &LI = Rewriter.getLoopInfo();
526 Instruction *BaseInsertPt = IP;
528 // Figure out the most-nested loop that IP is in.
529 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
531 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
532 // the preheader of the outer-most loop where NewBase is not loop invariant.
533 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
534 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
535 InsertLoop = InsertLoop->getParentLoop();
538 // If there is no immediate value, skip the next part.
539 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
540 if (SC->getValue()->isNullValue())
541 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
542 OperandValToReplace->getType());
544 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
546 // Always emit the immediate (if non-zero) into the same block as the user.
547 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
548 return Rewriter.expandCodeFor(NewValSCEV, IP,
549 OperandValToReplace->getType());
553 // Once we rewrite the code to insert the new IVs we want, update the
554 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
556 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
557 SCEVExpander &Rewriter,
559 if (!isa<PHINode>(Inst)) {
560 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
561 // Replace the use of the operand Value with the new Phi we just created.
562 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
563 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
567 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
568 // expression into each operand block that uses it. Note that PHI nodes can
569 // have multiple entries for the same predecessor. We use a map to make sure
570 // that a PHI node only has a single Value* for each predecessor (which also
571 // prevents us from inserting duplicate code in some blocks).
572 std::map<BasicBlock*, Value*> InsertedCode;
573 PHINode *PN = cast<PHINode>(Inst);
574 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
575 if (PN->getIncomingValue(i) == OperandValToReplace) {
576 // If this is a critical edge, split the edge so that we do not insert the
577 // code on all predecessor/successor paths. We do this unless this is the
578 // canonical backedge for this loop, as this can make some inserted code
579 // be in an illegal position.
580 BasicBlock *PHIPred = PN->getIncomingBlock(i);
581 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
582 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
584 // First step, split the critical edge.
585 SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
587 // Next step: move the basic block. In particular, if the PHI node
588 // is outside of the loop, and PredTI is in the loop, we want to
589 // move the block to be immediately before the PHI block, not
590 // immediately after PredTI.
591 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
592 BasicBlock *NewBB = PN->getIncomingBlock(i);
593 NewBB->moveBefore(PN->getParent());
596 // Splitting the edge can reduce the number of PHI entries we have.
597 e = PN->getNumIncomingValues();
600 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
602 // Insert the code into the end of the predecessor block.
603 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
604 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
607 // Replace the use of the operand Value with the new Phi we just created.
608 PN->setIncomingValue(i, Code);
612 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
616 /// isTargetConstant - Return true if the following can be referenced by the
617 /// immediate field of a target instruction.
618 static bool isTargetConstant(const SCEVHandle &V, const TargetLowering *TLI) {
619 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
620 int64_t V = SC->getValue()->getSExtValue();
622 return TLI->isLegalAddressImmediate(V);
624 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
625 return (V > -(1 << 16) && V < (1 << 16)-1);
628 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
629 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
630 if (CE->getOpcode() == Instruction::Cast) {
631 Constant *Op0 = CE->getOperand(0);
632 if (isa<GlobalValue>(Op0) &&
634 TLI->isLegalAddressImmediate(cast<GlobalValue>(Op0)))
640 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
641 /// loop varying to the Imm operand.
642 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
644 if (Val->isLoopInvariant(L)) return; // Nothing to do.
646 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
647 std::vector<SCEVHandle> NewOps;
648 NewOps.reserve(SAE->getNumOperands());
650 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
651 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
652 // If this is a loop-variant expression, it must stay in the immediate
653 // field of the expression.
654 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
656 NewOps.push_back(SAE->getOperand(i));
660 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
662 Val = SCEVAddExpr::get(NewOps);
663 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
664 // Try to pull immediates out of the start value of nested addrec's.
665 SCEVHandle Start = SARE->getStart();
666 MoveLoopVariantsToImediateField(Start, Imm, L);
668 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
670 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
672 // Otherwise, all of Val is variant, move the whole thing over.
673 Imm = SCEVAddExpr::get(Imm, Val);
674 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
679 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
680 /// that can fit into the immediate field of instructions in the target.
681 /// Accumulate these immediate values into the Imm value.
682 static void MoveImmediateValues(const TargetLowering *TLI,
683 SCEVHandle &Val, SCEVHandle &Imm,
684 bool isAddress, Loop *L) {
685 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
686 std::vector<SCEVHandle> NewOps;
687 NewOps.reserve(SAE->getNumOperands());
689 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
690 SCEVHandle NewOp = SAE->getOperand(i);
691 MoveImmediateValues(TLI, NewOp, Imm, isAddress, L);
693 if (!NewOp->isLoopInvariant(L)) {
694 // If this is a loop-variant expression, it must stay in the immediate
695 // field of the expression.
696 Imm = SCEVAddExpr::get(Imm, NewOp);
698 NewOps.push_back(NewOp);
703 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
705 Val = SCEVAddExpr::get(NewOps);
707 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
708 // Try to pull immediates out of the start value of nested addrec's.
709 SCEVHandle Start = SARE->getStart();
710 MoveImmediateValues(TLI, Start, Imm, isAddress, L);
712 if (Start != SARE->getStart()) {
713 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
715 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
718 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
719 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
720 if (isAddress && isTargetConstant(SME->getOperand(0), TLI) &&
721 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
723 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
724 SCEVHandle NewOp = SME->getOperand(1);
725 MoveImmediateValues(TLI, NewOp, SubImm, isAddress, L);
727 // If we extracted something out of the subexpressions, see if we can
729 if (NewOp != SME->getOperand(1)) {
730 // Scale SubImm up by "8". If the result is a target constant, we are
732 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
733 if (isTargetConstant(SubImm, TLI)) {
734 // Accumulate the immediate.
735 Imm = SCEVAddExpr::get(Imm, SubImm);
737 // Update what is left of 'Val'.
738 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
745 // Loop-variant expressions must stay in the immediate field of the
747 if ((isAddress && isTargetConstant(Val, TLI)) ||
748 !Val->isLoopInvariant(L)) {
749 Imm = SCEVAddExpr::get(Imm, Val);
750 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
754 // Otherwise, no immediates to move.
758 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
759 /// added together. This is used to reassociate common addition subexprs
760 /// together for maximal sharing when rewriting bases.
761 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
763 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
764 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
765 SeparateSubExprs(SubExprs, AE->getOperand(j));
766 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
767 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
768 if (SARE->getOperand(0) == Zero) {
769 SubExprs.push_back(Expr);
771 // Compute the addrec with zero as its base.
772 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
773 Ops[0] = Zero; // Start with zero base.
774 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
777 SeparateSubExprs(SubExprs, SARE->getOperand(0));
779 } else if (!isa<SCEVConstant>(Expr) ||
780 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
782 SubExprs.push_back(Expr);
787 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
788 /// removing any common subexpressions from it. Anything truly common is
789 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
790 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
792 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
793 unsigned NumUses = Uses.size();
795 // Only one use? Use its base, regardless of what it is!
796 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
797 SCEVHandle Result = Zero;
799 std::swap(Result, Uses[0].Base);
803 // To find common subexpressions, count how many of Uses use each expression.
804 // If any subexpressions are used Uses.size() times, they are common.
805 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
807 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
808 // order we see them.
809 std::vector<SCEVHandle> UniqueSubExprs;
811 std::vector<SCEVHandle> SubExprs;
812 for (unsigned i = 0; i != NumUses; ++i) {
813 // If the base is zero (which is common), return zero now, there are no
815 if (Uses[i].Base == Zero) return Zero;
817 // Split the expression into subexprs.
818 SeparateSubExprs(SubExprs, Uses[i].Base);
819 // Add one to SubExpressionUseCounts for each subexpr present.
820 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
821 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
822 UniqueSubExprs.push_back(SubExprs[j]);
826 // Now that we know how many times each is used, build Result. Iterate over
827 // UniqueSubexprs so that we have a stable ordering.
828 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
829 std::map<SCEVHandle, unsigned>::iterator I =
830 SubExpressionUseCounts.find(UniqueSubExprs[i]);
831 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
832 if (I->second == NumUses) { // Found CSE!
833 Result = SCEVAddExpr::get(Result, I->first);
835 // Remove non-cse's from SubExpressionUseCounts.
836 SubExpressionUseCounts.erase(I);
840 // If we found no CSE's, return now.
841 if (Result == Zero) return Result;
843 // Otherwise, remove all of the CSE's we found from each of the base values.
844 for (unsigned i = 0; i != NumUses; ++i) {
845 // Split the expression into subexprs.
846 SeparateSubExprs(SubExprs, Uses[i].Base);
848 // Remove any common subexpressions.
849 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
850 if (SubExpressionUseCounts.count(SubExprs[j])) {
851 SubExprs.erase(SubExprs.begin()+j);
855 // Finally, the non-shared expressions together.
856 if (SubExprs.empty())
859 Uses[i].Base = SCEVAddExpr::get(SubExprs);
866 /// isZero - returns true if the scalar evolution expression is zero.
868 static bool isZero(SCEVHandle &V) {
869 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
870 return SC->getValue()->getZExtValue() == 0;
875 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
876 /// of a previous stride and it is a legal value for the target addressing
877 /// mode scale component. This allows the users of this stride to be rewritten
878 /// as prev iv * factor. It returns 0 if no reuse is possible.
879 unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride,
880 IVExpr &IV, const Type *Ty) {
883 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
884 int64_t SInt = SC->getValue()->getSExtValue();
885 if (SInt == 1) return 0;
887 for (TargetLowering::legal_am_scale_iterator
888 I = TLI->legal_am_scale_begin(), E = TLI->legal_am_scale_end();
891 if (unsigned(abs(SInt)) < Scale || (SInt % Scale) != 0)
893 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
894 IVsByStride.find(SCEVUnknown::getIntegerSCEV(SInt/Scale, Type::UIntTy));
895 if (SI == IVsByStride.end())
897 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
898 IE = SI->second.IVs.end(); II != IE; ++II)
899 // FIXME: Only handle base == 0 for now.
900 // Only reuse previous IV if it would not require a type conversion.
901 if (isZero(II->Base) &&
902 II->Base->getType()->isLosslesslyConvertibleTo(Ty)) {
912 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
913 /// returns true if Val's isUseOfPostIncrementedValue is true.
914 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
915 return Val.isUseOfPostIncrementedValue;
918 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
919 /// stride of IV. All of the users may have different starting values, and this
920 /// may not be the only stride (we know it is if isOnlyStride is true).
921 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
922 IVUsersOfOneStride &Uses,
925 // Transform our list of users and offsets to a bit more complex table. In
926 // this new vector, each 'BasedUser' contains 'Base' the base of the
927 // strided accessas well as the old information from Uses. We progressively
928 // move information from the Base field to the Imm field, until we eventually
929 // have the full access expression to rewrite the use.
930 std::vector<BasedUser> UsersToProcess;
931 UsersToProcess.reserve(Uses.Users.size());
932 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
933 UsersToProcess.push_back(Uses.Users[i]);
935 // Move any loop invariant operands from the offset field to the immediate
936 // field of the use, so that we don't try to use something before it is
938 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
939 UsersToProcess.back().Imm, L);
940 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
941 "Base value is not loop invariant!");
944 // We now have a whole bunch of uses of like-strided induction variables, but
945 // they might all have different bases. We want to emit one PHI node for this
946 // stride which we fold as many common expressions (between the IVs) into as
947 // possible. Start by identifying the common expressions in the base values
948 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
949 // "A+B"), emit it to the preheader, then remove the expression from the
950 // UsersToProcess base values.
951 SCEVHandle CommonExprs =
952 RemoveCommonExpressionsFromUseBases(UsersToProcess);
954 // Check if it is possible to reuse a IV with stride that is factor of this
955 // stride. And the multiple is a number that can be encoded in the scale
956 // field of the target addressing mode.
957 PHINode *NewPHI = NULL;
960 unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
961 CommonExprs->getType());
962 if (RewriteFactor != 0) {
963 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
964 << " and BASE " << *ReuseIV.Base << " :\n";
965 NewPHI = ReuseIV.PHI;
969 // Next, figure out what we can represent in the immediate fields of
970 // instructions. If we can represent anything there, move it to the imm
971 // fields of the BasedUsers. We do this so that it increases the commonality
972 // of the remaining uses.
973 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
974 // If the user is not in the current loop, this means it is using the exit
975 // value of the IV. Do not put anything in the base, make sure it's all in
976 // the immediate field to allow as much factoring as possible.
977 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
978 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
979 UsersToProcess[i].Base);
980 UsersToProcess[i].Base =
981 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
984 // Addressing modes can be folded into loads and stores. Be careful that
985 // the store is through the expression, not of the expression though.
986 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
987 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
988 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
991 MoveImmediateValues(TLI, UsersToProcess[i].Base, UsersToProcess[i].Imm,
996 // Now that we know what we need to do, insert the PHI node itself.
998 DOUT << "INSERTING IV of STRIDE " << *Stride << " and BASE "
999 << *CommonExprs << " :\n";
1001 SCEVExpander Rewriter(*SE, *LI);
1002 SCEVExpander PreheaderRewriter(*SE, *LI);
1004 BasicBlock *Preheader = L->getLoopPreheader();
1005 Instruction *PreInsertPt = Preheader->getTerminator();
1006 Instruction *PhiInsertBefore = L->getHeader()->begin();
1008 BasicBlock *LatchBlock = L->getLoopLatch();
1010 const Type *ReplacedTy = CommonExprs->getType();
1012 // Emit the initial base value into the loop preheader.
1014 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
1017 if (RewriteFactor == 0) {
1018 // Create a new Phi for this base, and stick it in the loop header.
1019 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1022 // Add common base to the new Phi node.
1023 NewPHI->addIncoming(CommonBaseV, Preheader);
1025 // Insert the stride into the preheader.
1026 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
1028 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1030 // Emit the increment of the base value before the terminator of the loop
1031 // latch block, and add it to the Phi node.
1032 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
1033 SCEVUnknown::get(StrideV));
1035 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
1037 IncV->setName(NewPHI->getName()+".inc");
1038 NewPHI->addIncoming(IncV, LatchBlock);
1040 // Remember this in case a later stride is multiple of this.
1041 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1043 Constant *C = dyn_cast<Constant>(CommonBaseV);
1045 (!C->isNullValue() &&
1046 !isTargetConstant(SCEVUnknown::get(CommonBaseV), TLI)))
1047 // We want the common base emitted into the preheader!
1048 CommonBaseV = new CastInst(CommonBaseV, CommonBaseV->getType(),
1049 "commonbase", PreInsertPt);
1052 // We want to emit code for users inside the loop first. To do this, we
1053 // rearrange BasedUser so that the entries at the end have
1054 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1055 // vector (so we handle them first).
1056 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1057 PartitionByIsUseOfPostIncrementedValue);
1059 // Sort this by base, so that things with the same base are handled
1060 // together. By partitioning first and stable-sorting later, we are
1061 // guaranteed that within each base we will pop off users from within the
1062 // loop before users outside of the loop with a particular base.
1064 // We would like to use stable_sort here, but we can't. The problem is that
1065 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1066 // we don't have anything to do a '<' comparison on. Because we think the
1067 // number of uses is small, do a horrible bubble sort which just relies on
1069 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1070 // Get a base value.
1071 SCEVHandle Base = UsersToProcess[i].Base;
1073 // Compact everything with this base to be consequetive with this one.
1074 for (unsigned j = i+1; j != e; ++j) {
1075 if (UsersToProcess[j].Base == Base) {
1076 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1082 // Process all the users now. This outer loop handles all bases, the inner
1083 // loop handles all users of a particular base.
1084 while (!UsersToProcess.empty()) {
1085 SCEVHandle Base = UsersToProcess.back().Base;
1087 DOUT << " INSERTING code for BASE = " << *Base << ":\n";
1089 // Emit the code for Base into the preheader.
1090 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
1093 // If BaseV is a constant other than 0, make sure that it gets inserted into
1094 // the preheader, instead of being forward substituted into the uses. We do
1095 // this by forcing a noop cast to be inserted into the preheader in this
1097 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1098 if (!C->isNullValue() && !isTargetConstant(Base, TLI)) {
1099 // We want this constant emitted into the preheader!
1100 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
1105 // Emit the code to add the immediate offset to the Phi value, just before
1106 // the instructions that we identified as using this stride and base.
1108 // FIXME: Use emitted users to emit other users.
1109 BasedUser &User = UsersToProcess.back();
1111 // If this instruction wants to use the post-incremented value, move it
1112 // after the post-inc and use its value instead of the PHI.
1113 Value *RewriteOp = NewPHI;
1114 if (User.isUseOfPostIncrementedValue) {
1117 // If this user is in the loop, make sure it is the last thing in the
1118 // loop to ensure it is dominated by the increment.
1119 if (L->contains(User.Inst->getParent()))
1120 User.Inst->moveBefore(LatchBlock->getTerminator());
1122 if (RewriteOp->getType() != ReplacedTy)
1123 RewriteOp = SCEVExpander::InsertCastOfTo(RewriteOp, ReplacedTy);
1125 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
1127 // Clear the SCEVExpander's expression map so that we are guaranteed
1128 // to have the code emitted where we expect it.
1131 // If we are reusing the iv, then it must be multiplied by a constant
1132 // factor take advantage of addressing mode scale component.
1133 if (RewriteFactor != 0) {
1135 SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
1136 RewriteExpr->getType()),
1139 // The common base is emitted in the loop preheader. But since we
1140 // are reusing an IV, it has not been used to initialize the PHI node.
1141 // Add it to the expression used to rewrite the uses.
1142 if (!isa<ConstantInt>(CommonBaseV) ||
1143 !cast<ConstantInt>(CommonBaseV)->isNullValue())
1144 RewriteExpr = SCEVAddExpr::get(RewriteExpr,
1145 SCEVUnknown::get(CommonBaseV));
1148 // Now that we know what we need to do, insert code before User for the
1149 // immediate and any loop-variant expressions.
1150 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
1151 // Add BaseV to the PHI value if needed.
1152 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
1154 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1156 // Mark old value we replaced as possibly dead, so that it is elminated
1157 // if we just replaced the last use of that value.
1158 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1160 UsersToProcess.pop_back();
1163 // If there are any more users to process with the same base, process them
1164 // now. We sorted by base above, so we just have to check the last elt.
1165 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1166 // TODO: Next, find out which base index is the most common, pull it out.
1169 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1170 // different starting values, into different PHIs.
1173 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1174 // uses in the loop, look to see if we can eliminate some, in favor of using
1175 // common indvars for the different uses.
1176 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1177 // TODO: implement optzns here.
1182 // Finally, get the terminating condition for the loop if possible. If we
1183 // can, we want to change it to use a post-incremented version of its
1184 // induction variable, to allow coalescing the live ranges for the IV into
1185 // one register value.
1186 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1187 BasicBlock *Preheader = L->getLoopPreheader();
1188 BasicBlock *LatchBlock =
1189 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1190 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1191 if (!TermBr || TermBr->isUnconditional() ||
1192 !isa<SetCondInst>(TermBr->getCondition()))
1194 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
1196 // Search IVUsesByStride to find Cond's IVUse if there is one.
1197 IVStrideUse *CondUse = 0;
1198 const SCEVHandle *CondStride = 0;
1200 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1202 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1203 IVUsesByStride.find(StrideOrder[Stride]);
1204 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1206 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1207 E = SI->second.Users.end(); UI != E; ++UI)
1208 if (UI->User == Cond) {
1210 CondStride = &SI->first;
1211 // NOTE: we could handle setcc instructions with multiple uses here, but
1212 // InstCombine does it as well for simple uses, it's not clear that it
1213 // occurs enough in real life to handle.
1217 if (!CondUse) return; // setcc doesn't use the IV.
1219 // It's possible for the setcc instruction to be anywhere in the loop, and
1220 // possible for it to have multiple users. If it is not immediately before
1221 // the latch block branch, move it.
1222 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1223 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1224 Cond->moveBefore(TermBr);
1226 // Otherwise, clone the terminating condition and insert into the loopend.
1227 Cond = cast<SetCondInst>(Cond->clone());
1228 Cond->setName(L->getHeader()->getName() + ".termcond");
1229 LatchBlock->getInstList().insert(TermBr, Cond);
1231 // Clone the IVUse, as the old use still exists!
1232 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1233 CondUse->OperandValToReplace);
1234 CondUse = &IVUsesByStride[*CondStride].Users.back();
1238 // If we get to here, we know that we can transform the setcc instruction to
1239 // use the post-incremented version of the IV, allowing us to coalesce the
1240 // live ranges for the IV correctly.
1241 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1242 CondUse->isUseOfPostIncrementedValue = true;
1246 // Constant strides come first which in turns are sorted by their absolute
1247 // values. If absolute values are the same, then positive strides comes first.
1249 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1250 struct StrideCompare {
1251 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1252 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1253 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1255 int64_t LV = LHSC->getValue()->getSExtValue();
1256 int64_t RV = RHSC->getValue()->getSExtValue();
1257 uint64_t ALV = (LV < 0) ? -LV : LV;
1258 uint64_t ARV = (RV < 0) ? -RV : RV;
1264 return (LHSC && !RHSC);
1269 void LoopStrengthReduce::runOnLoop(Loop *L) {
1270 // First step, transform all loops nesting inside of this loop.
1271 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
1274 // Next, find all uses of induction variables in this loop, and catagorize
1275 // them by stride. Start by finding all of the PHI nodes in the header for
1276 // this loop. If they are induction variables, inspect their uses.
1277 std::set<Instruction*> Processed; // Don't reprocess instructions.
1278 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1279 AddUsersIfInteresting(I, L, Processed);
1281 // If we have nothing to do, return.
1282 if (IVUsesByStride.empty()) return;
1284 // Optimize induction variables. Some indvar uses can be transformed to use
1285 // strides that will be needed for other purposes. A common example of this
1286 // is the exit test for the loop, which can often be rewritten to use the
1287 // computation of some other indvar to decide when to terminate the loop.
1291 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1292 // doing computation in byte values, promote to 32-bit values if safe.
1294 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1295 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1296 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1297 // to be careful that IV's are all the same type. Only works for intptr_t
1300 // If we only have one stride, we can more aggressively eliminate some things.
1301 bool HasOneStride = IVUsesByStride.size() == 1;
1304 DOUT << "\nLSR on ";
1308 // IVsByStride keeps IVs for one particular loop.
1309 IVsByStride.clear();
1311 // Sort the StrideOrder so we process larger strides first.
1312 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1314 // Note: this processes each stride/type pair individually. All users passed
1315 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1316 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1317 // This extra layer of indirection makes the ordering of strides deterministic
1318 // - not dependent on map order.
1319 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1320 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1321 IVUsesByStride.find(StrideOrder[Stride]);
1322 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1323 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1326 // Clean up after ourselves
1327 if (!DeadInsts.empty()) {
1328 DeleteTriviallyDeadInstructions(DeadInsts);
1330 BasicBlock::iterator I = L->getHeader()->begin();
1332 while ((PN = dyn_cast<PHINode>(I))) {
1333 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1335 // At this point, we know that we have killed one or more GEP
1336 // instructions. It is worth checking to see if the cann indvar is also
1337 // dead, so that we can remove it as well. The requirements for the cann
1338 // indvar to be considered dead are:
1339 // 1. the cann indvar has one use
1340 // 2. the use is an add instruction
1341 // 3. the add has one use
1342 // 4. the add is used by the cann indvar
1343 // If all four cases above are true, then we can remove both the add and
1345 // FIXME: this needs to eliminate an induction variable even if it's being
1346 // compared against some value to decide loop termination.
1347 if (PN->hasOneUse()) {
1348 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1349 if (BO && BO->hasOneUse()) {
1350 if (PN == *(BO->use_begin())) {
1351 DeadInsts.insert(BO);
1352 // Break the cycle, then delete the PHI.
1353 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1354 SE->deleteInstructionFromRecords(PN);
1355 PN->eraseFromParent();
1360 DeleteTriviallyDeadInstructions(DeadInsts);
1363 CastedPointers.clear();
1364 IVUsesByStride.clear();
1365 StrideOrder.clear();