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));
80 class LoopStrengthReduce : public FunctionPass {
85 const Type *UIntPtrTy;
88 /// MaxTargetAMSize - This is the maximum power-of-two scale value that the
89 /// target can handle for free with its addressing modes.
90 unsigned MaxTargetAMSize;
92 /// IVUsesByStride - Keep track of all uses of induction variables that we
93 /// are interested in. The key of the map is the stride of the access.
94 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
96 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
97 /// We use this to iterate over the IVUsesByStride collection without being
98 /// dependent on random ordering of pointers in the process.
99 std::vector<SCEVHandle> StrideOrder;
101 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
102 /// of the casted version of each value. This is accessed by
103 /// getCastedVersionOf.
104 std::map<Value*, Value*> CastedPointers;
106 /// DeadInsts - Keep track of instructions we may have made dead, so that
107 /// we can remove them after we are done working.
108 std::set<Instruction*> DeadInsts;
110 /// TLI - Keep a pointer of a TargetLowering to consult for determining
111 /// transformation profitability.
112 const TargetLowering *TLI;
115 LoopStrengthReduce(unsigned MTAMS = 1, const TargetLowering *tli = NULL)
116 : MaxTargetAMSize(MTAMS), TLI(tli) {
119 virtual bool runOnFunction(Function &) {
120 LI = &getAnalysis<LoopInfo>();
121 EF = &getAnalysis<ETForest>();
122 SE = &getAnalysis<ScalarEvolution>();
123 TD = &getAnalysis<TargetData>();
124 UIntPtrTy = TD->getIntPtrType();
127 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
133 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
134 // We split critical edges, so we change the CFG. However, we do update
135 // many analyses if they are around.
136 AU.addPreservedID(LoopSimplifyID);
137 AU.addPreserved<LoopInfo>();
138 AU.addPreserved<DominatorSet>();
139 AU.addPreserved<ETForest>();
140 AU.addPreserved<ImmediateDominators>();
141 AU.addPreserved<DominanceFrontier>();
142 AU.addPreserved<DominatorTree>();
144 AU.addRequiredID(LoopSimplifyID);
145 AU.addRequired<LoopInfo>();
146 AU.addRequired<ETForest>();
147 AU.addRequired<TargetData>();
148 AU.addRequired<ScalarEvolution>();
151 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
153 Value *getCastedVersionOf(Value *V);
155 void runOnLoop(Loop *L);
156 bool AddUsersIfInteresting(Instruction *I, Loop *L,
157 std::set<Instruction*> &Processed);
158 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
160 void OptimizeIndvars(Loop *L);
162 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
163 IVUsersOfOneStride &Uses,
164 Loop *L, bool isOnlyStride);
165 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
167 RegisterOpt<LoopStrengthReduce> X("loop-reduce",
168 "Loop Strength Reduction");
171 FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize,
172 const TargetLowering *TLI) {
173 return new LoopStrengthReduce(MaxTargetAMSize, TLI);
176 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
178 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
179 if (V->getType() == UIntPtrTy) return V;
180 if (Constant *CB = dyn_cast<Constant>(V))
181 return ConstantExpr::getCast(CB, UIntPtrTy);
183 Value *&New = CastedPointers[V];
186 New = SCEVExpander::InsertCastOfTo(V, UIntPtrTy);
187 DeadInsts.insert(cast<Instruction>(New));
192 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
193 /// specified set are trivially dead, delete them and see if this makes any of
194 /// their operands subsequently dead.
195 void LoopStrengthReduce::
196 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
197 while (!Insts.empty()) {
198 Instruction *I = *Insts.begin();
199 Insts.erase(Insts.begin());
200 if (isInstructionTriviallyDead(I)) {
201 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
202 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
204 SE->deleteInstructionFromRecords(I);
205 I->eraseFromParent();
212 /// GetExpressionSCEV - Compute and return the SCEV for the specified
214 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
215 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
216 // If this is a GEP that SE doesn't know about, compute it now and insert it.
217 // If this is not a GEP, or if we have already done this computation, just let
219 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
220 if (!GEP || SE->hasSCEV(GEP))
221 return SE->getSCEV(Exp);
223 // Analyze all of the subscripts of this getelementptr instruction, looking
224 // for uses that are determined by the trip count of L. First, skip all
225 // operands the are not dependent on the IV.
227 // Build up the base expression. Insert an LLVM cast of the pointer to
229 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
231 gep_type_iterator GTI = gep_type_begin(GEP);
233 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
234 // If this is a use of a recurrence that we can analyze, and it comes before
235 // Op does in the GEP operand list, we will handle this when we process this
237 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
238 const StructLayout *SL = TD->getStructLayout(STy);
239 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
240 uint64_t Offset = SL->MemberOffsets[Idx];
241 GEPVal = SCEVAddExpr::get(GEPVal,
242 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
244 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
245 SCEVHandle Idx = SE->getSCEV(OpVal);
247 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
249 Idx = SCEVMulExpr::get(Idx,
250 SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
252 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
256 SE->setSCEV(GEP, GEPVal);
260 /// getSCEVStartAndStride - Compute the start and stride of this expression,
261 /// returning false if the expression is not a start/stride pair, or true if it
262 /// is. The stride must be a loop invariant expression, but the start may be
263 /// a mix of loop invariant and loop variant expressions.
264 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
265 SCEVHandle &Start, SCEVHandle &Stride) {
266 SCEVHandle TheAddRec = Start; // Initialize to zero.
268 // If the outer level is an AddExpr, the operands are all start values except
269 // for a nested AddRecExpr.
270 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
271 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
272 if (SCEVAddRecExpr *AddRec =
273 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
274 if (AddRec->getLoop() == L)
275 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
277 return false; // Nested IV of some sort?
279 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
282 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
285 return false; // not analyzable.
288 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
289 if (!AddRec || AddRec->getLoop() != L) return false;
291 // FIXME: Generalize to non-affine IV's.
292 if (!AddRec->isAffine()) return false;
294 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
296 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
297 DEBUG(std::cerr << "[" << L->getHeader()->getName()
298 << "] Variable stride: " << *AddRec << "\n");
300 Stride = AddRec->getOperand(1);
301 // Check that all constant strides are the unsigned type, we don't want to
302 // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
304 assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
305 "Constants should be canonicalized to unsigned!");
310 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
311 /// and now we need to decide whether the user should use the preinc or post-inc
312 /// value. If this user should use the post-inc version of the IV, return true.
314 /// Choosing wrong here can break dominance properties (if we choose to use the
315 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
316 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
317 /// should use the post-inc value).
318 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
319 Loop *L, ETForest *EF, Pass *P) {
320 // If the user is in the loop, use the preinc value.
321 if (L->contains(User->getParent())) return false;
323 BasicBlock *LatchBlock = L->getLoopLatch();
325 // Ok, the user is outside of the loop. If it is dominated by the latch
326 // block, use the post-inc value.
327 if (EF->dominates(LatchBlock, User->getParent()))
330 // There is one case we have to be careful of: PHI nodes. These little guys
331 // can live in blocks that do not dominate the latch block, but (since their
332 // uses occur in the predecessor block, not the block the PHI lives in) should
333 // still use the post-inc value. Check for this case now.
334 PHINode *PN = dyn_cast<PHINode>(User);
335 if (!PN) return false; // not a phi, not dominated by latch block.
337 // Look at all of the uses of IV by the PHI node. If any use corresponds to
338 // a block that is not dominated by the latch block, give up and use the
339 // preincremented value.
340 unsigned NumUses = 0;
341 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
342 if (PN->getIncomingValue(i) == IV) {
344 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
348 // Okay, all uses of IV by PN are in predecessor blocks that really are
349 // dominated by the latch block. Split the critical edges and use the
350 // post-incremented value.
351 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
352 if (PN->getIncomingValue(i) == IV) {
353 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P);
354 if (--NumUses == 0) break;
362 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
363 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
364 /// return true. Otherwise, return false.
365 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
366 std::set<Instruction*> &Processed) {
367 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
368 return false; // Void and FP expressions cannot be reduced.
369 if (!Processed.insert(I).second)
370 return true; // Instruction already handled.
372 // Get the symbolic expression for this instruction.
373 SCEVHandle ISE = GetExpressionSCEV(I, L);
374 if (isa<SCEVCouldNotCompute>(ISE)) return false;
376 // Get the start and stride for this expression.
377 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
378 SCEVHandle Stride = Start;
379 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
380 return false; // Non-reducible symbolic expression, bail out.
382 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
383 Instruction *User = cast<Instruction>(*UI);
385 // Do not infinitely recurse on PHI nodes.
386 if (isa<PHINode>(User) && Processed.count(User))
389 // If this is an instruction defined in a nested loop, or outside this loop,
390 // don't recurse into it.
391 bool AddUserToIVUsers = false;
392 if (LI->getLoopFor(User->getParent()) != L) {
393 DEBUG(std::cerr << "FOUND USER in other loop: " << *User
394 << " OF SCEV: " << *ISE << "\n");
395 AddUserToIVUsers = true;
396 } else if (!AddUsersIfInteresting(User, L, Processed)) {
397 DEBUG(std::cerr << "FOUND USER: " << *User
398 << " OF SCEV: " << *ISE << "\n");
399 AddUserToIVUsers = true;
402 if (AddUserToIVUsers) {
403 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
404 if (StrideUses.Users.empty()) // First occurance of this stride?
405 StrideOrder.push_back(Stride);
407 // Okay, we found a user that we cannot reduce. Analyze the instruction
408 // and decide what to do with it. If we are a use inside of the loop, use
409 // the value before incrementation, otherwise use it after incrementation.
410 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
411 // The value used will be incremented by the stride more than we are
412 // expecting, so subtract this off.
413 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
414 StrideUses.addUser(NewStart, User, I);
415 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
416 DEBUG(std::cerr << " USING POSTINC SCEV, START=" << *NewStart<< "\n");
418 StrideUses.addUser(Start, User, I);
426 /// BasedUser - For a particular base value, keep information about how we've
427 /// partitioned the expression so far.
429 /// Base - The Base value for the PHI node that needs to be inserted for
430 /// this use. As the use is processed, information gets moved from this
431 /// field to the Imm field (below). BasedUser values are sorted by this
435 /// Inst - The instruction using the induction variable.
438 /// OperandValToReplace - The operand value of Inst to replace with the
440 Value *OperandValToReplace;
442 /// Imm - The immediate value that should be added to the base immediately
443 /// before Inst, because it will be folded into the imm field of the
447 /// EmittedBase - The actual value* to use for the base value of this
448 /// operation. This is null if we should just use zero so far.
451 // isUseOfPostIncrementedValue - True if this should use the
452 // post-incremented version of this IV, not the preincremented version.
453 // This can only be set in special cases, such as the terminating setcc
454 // instruction for a loop and uses outside the loop that are dominated by
456 bool isUseOfPostIncrementedValue;
458 BasedUser(IVStrideUse &IVSU)
459 : Base(IVSU.Offset), Inst(IVSU.User),
460 OperandValToReplace(IVSU.OperandValToReplace),
461 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
462 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
464 // Once we rewrite the code to insert the new IVs we want, update the
465 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
467 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
468 SCEVExpander &Rewriter, Loop *L,
471 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
472 SCEVExpander &Rewriter,
473 Instruction *IP, Loop *L);
478 void BasedUser::dump() const {
479 std::cerr << " Base=" << *Base;
480 std::cerr << " Imm=" << *Imm;
482 std::cerr << " EB=" << *EmittedBase;
484 std::cerr << " Inst: " << *Inst;
487 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
488 SCEVExpander &Rewriter,
489 Instruction *IP, Loop *L) {
490 // Figure out where we *really* want to insert this code. In particular, if
491 // the user is inside of a loop that is nested inside of L, we really don't
492 // want to insert this expression before the user, we'd rather pull it out as
493 // many loops as possible.
494 LoopInfo &LI = Rewriter.getLoopInfo();
495 Instruction *BaseInsertPt = IP;
497 // Figure out the most-nested loop that IP is in.
498 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
500 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
501 // the preheader of the outer-most loop where NewBase is not loop invariant.
502 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
503 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
504 InsertLoop = InsertLoop->getParentLoop();
507 // If there is no immediate value, skip the next part.
508 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
509 if (SC->getValue()->isNullValue())
510 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
511 OperandValToReplace->getType());
513 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
515 // Always emit the immediate (if non-zero) into the same block as the user.
516 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
517 return Rewriter.expandCodeFor(NewValSCEV, IP,
518 OperandValToReplace->getType());
522 // Once we rewrite the code to insert the new IVs we want, update the
523 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
525 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
526 SCEVExpander &Rewriter,
528 if (!isa<PHINode>(Inst)) {
529 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
530 // Replace the use of the operand Value with the new Phi we just created.
531 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
532 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
536 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
537 // expression into each operand block that uses it. Note that PHI nodes can
538 // have multiple entries for the same predecessor. We use a map to make sure
539 // that a PHI node only has a single Value* for each predecessor (which also
540 // prevents us from inserting duplicate code in some blocks).
541 std::map<BasicBlock*, Value*> InsertedCode;
542 PHINode *PN = cast<PHINode>(Inst);
543 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
544 if (PN->getIncomingValue(i) == OperandValToReplace) {
545 // If this is a critical edge, split the edge so that we do not insert the
546 // code on all predecessor/successor paths. We do this unless this is the
547 // canonical backedge for this loop, as this can make some inserted code
548 // be in an illegal position.
549 BasicBlock *PHIPred = PN->getIncomingBlock(i);
550 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
551 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
553 // First step, split the critical edge.
554 SplitCriticalEdge(PHIPred, PN->getParent(), P);
556 // Next step: move the basic block. In particular, if the PHI node
557 // is outside of the loop, and PredTI is in the loop, we want to
558 // move the block to be immediately before the PHI block, not
559 // immediately after PredTI.
560 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
561 BasicBlock *NewBB = PN->getIncomingBlock(i);
562 NewBB->moveBefore(PN->getParent());
566 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
568 // Insert the code into the end of the predecessor block.
569 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
570 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
573 // Replace the use of the operand Value with the new Phi we just created.
574 PN->setIncomingValue(i, Code);
578 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
582 /// isTargetConstant - Return true if the following can be referenced by the
583 /// immediate field of a target instruction.
584 static bool isTargetConstant(const SCEVHandle &V, const TargetLowering *TLI) {
585 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
586 int64_t V = SC->getValue()->getSExtValue();
588 return TLI->isLegalAddressImmediate(V);
590 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
591 return (V > -(1 << 16) && V < (1 << 16)-1);
594 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
595 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
596 if (CE->getOpcode() == Instruction::Cast) {
597 Constant *Op0 = CE->getOperand(0);
598 if (isa<GlobalValue>(Op0) &&
600 TLI->isLegalAddressImmediate(cast<GlobalValue>(Op0)))
606 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
607 /// loop varying to the Imm operand.
608 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
610 if (Val->isLoopInvariant(L)) return; // Nothing to do.
612 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
613 std::vector<SCEVHandle> NewOps;
614 NewOps.reserve(SAE->getNumOperands());
616 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
617 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
618 // If this is a loop-variant expression, it must stay in the immediate
619 // field of the expression.
620 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
622 NewOps.push_back(SAE->getOperand(i));
626 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
628 Val = SCEVAddExpr::get(NewOps);
629 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
630 // Try to pull immediates out of the start value of nested addrec's.
631 SCEVHandle Start = SARE->getStart();
632 MoveLoopVariantsToImediateField(Start, Imm, L);
634 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
636 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
638 // Otherwise, all of Val is variant, move the whole thing over.
639 Imm = SCEVAddExpr::get(Imm, Val);
640 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
645 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
646 /// that can fit into the immediate field of instructions in the target.
647 /// Accumulate these immediate values into the Imm value.
648 static void MoveImmediateValues(const TargetLowering *TLI,
649 SCEVHandle &Val, SCEVHandle &Imm,
650 bool isAddress, Loop *L) {
651 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
652 std::vector<SCEVHandle> NewOps;
653 NewOps.reserve(SAE->getNumOperands());
655 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
656 SCEVHandle NewOp = SAE->getOperand(i);
657 MoveImmediateValues(TLI, NewOp, Imm, isAddress, L);
659 if (!NewOp->isLoopInvariant(L)) {
660 // If this is a loop-variant expression, it must stay in the immediate
661 // field of the expression.
662 Imm = SCEVAddExpr::get(Imm, NewOp);
664 NewOps.push_back(NewOp);
669 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
671 Val = SCEVAddExpr::get(NewOps);
673 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
674 // Try to pull immediates out of the start value of nested addrec's.
675 SCEVHandle Start = SARE->getStart();
676 MoveImmediateValues(TLI, Start, Imm, isAddress, L);
678 if (Start != SARE->getStart()) {
679 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
681 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
684 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
685 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
686 if (isAddress && isTargetConstant(SME->getOperand(0), TLI) &&
687 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
689 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
690 SCEVHandle NewOp = SME->getOperand(1);
691 MoveImmediateValues(TLI, NewOp, SubImm, isAddress, L);
693 // If we extracted something out of the subexpressions, see if we can
695 if (NewOp != SME->getOperand(1)) {
696 // Scale SubImm up by "8". If the result is a target constant, we are
698 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
699 if (isTargetConstant(SubImm, TLI)) {
700 // Accumulate the immediate.
701 Imm = SCEVAddExpr::get(Imm, SubImm);
703 // Update what is left of 'Val'.
704 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
711 // Loop-variant expressions must stay in the immediate field of the
713 if ((isAddress && isTargetConstant(Val, TLI)) ||
714 !Val->isLoopInvariant(L)) {
715 Imm = SCEVAddExpr::get(Imm, Val);
716 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
720 // Otherwise, no immediates to move.
724 /// IncrementAddExprUses - Decompose the specified expression into its added
725 /// subexpressions, and increment SubExpressionUseCounts for each of these
726 /// decomposed parts.
727 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
729 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
730 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
731 SeparateSubExprs(SubExprs, AE->getOperand(j));
732 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
733 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
734 if (SARE->getOperand(0) == Zero) {
735 SubExprs.push_back(Expr);
737 // Compute the addrec with zero as its base.
738 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
739 Ops[0] = Zero; // Start with zero base.
740 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
743 SeparateSubExprs(SubExprs, SARE->getOperand(0));
745 } else if (!isa<SCEVConstant>(Expr) ||
746 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
748 SubExprs.push_back(Expr);
753 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
754 /// removing any common subexpressions from it. Anything truly common is
755 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
756 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
758 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
759 unsigned NumUses = Uses.size();
761 // Only one use? Use its base, regardless of what it is!
762 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
763 SCEVHandle Result = Zero;
765 std::swap(Result, Uses[0].Base);
769 // To find common subexpressions, count how many of Uses use each expression.
770 // If any subexpressions are used Uses.size() times, they are common.
771 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
773 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
774 // order we see them.
775 std::vector<SCEVHandle> UniqueSubExprs;
777 std::vector<SCEVHandle> SubExprs;
778 for (unsigned i = 0; i != NumUses; ++i) {
779 // If the base is zero (which is common), return zero now, there are no
781 if (Uses[i].Base == Zero) return Zero;
783 // Split the expression into subexprs.
784 SeparateSubExprs(SubExprs, Uses[i].Base);
785 // Add one to SubExpressionUseCounts for each subexpr present.
786 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
787 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
788 UniqueSubExprs.push_back(SubExprs[j]);
792 // Now that we know how many times each is used, build Result. Iterate over
793 // UniqueSubexprs so that we have a stable ordering.
794 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
795 std::map<SCEVHandle, unsigned>::iterator I =
796 SubExpressionUseCounts.find(UniqueSubExprs[i]);
797 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
798 if (I->second == NumUses) { // Found CSE!
799 Result = SCEVAddExpr::get(Result, I->first);
801 // Remove non-cse's from SubExpressionUseCounts.
802 SubExpressionUseCounts.erase(I);
806 // If we found no CSE's, return now.
807 if (Result == Zero) return Result;
809 // Otherwise, remove all of the CSE's we found from each of the base values.
810 for (unsigned i = 0; i != NumUses; ++i) {
811 // Split the expression into subexprs.
812 SeparateSubExprs(SubExprs, Uses[i].Base);
814 // Remove any common subexpressions.
815 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
816 if (SubExpressionUseCounts.count(SubExprs[j])) {
817 SubExprs.erase(SubExprs.begin()+j);
821 // Finally, the non-shared expressions together.
822 if (SubExprs.empty())
825 Uses[i].Base = SCEVAddExpr::get(SubExprs);
833 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
834 /// stride of IV. All of the users may have different starting values, and this
835 /// may not be the only stride (we know it is if isOnlyStride is true).
836 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
837 IVUsersOfOneStride &Uses,
840 // Transform our list of users and offsets to a bit more complex table. In
841 // this new vector, each 'BasedUser' contains 'Base' the base of the
842 // strided accessas well as the old information from Uses. We progressively
843 // move information from the Base field to the Imm field, until we eventually
844 // have the full access expression to rewrite the use.
845 std::vector<BasedUser> UsersToProcess;
846 UsersToProcess.reserve(Uses.Users.size());
847 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
848 UsersToProcess.push_back(Uses.Users[i]);
850 // Move any loop invariant operands from the offset field to the immediate
851 // field of the use, so that we don't try to use something before it is
853 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
854 UsersToProcess.back().Imm, L);
855 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
856 "Base value is not loop invariant!");
859 // We now have a whole bunch of uses of like-strided induction variables, but
860 // they might all have different bases. We want to emit one PHI node for this
861 // stride which we fold as many common expressions (between the IVs) into as
862 // possible. Start by identifying the common expressions in the base values
863 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
864 // "A+B"), emit it to the preheader, then remove the expression from the
865 // UsersToProcess base values.
866 SCEVHandle CommonExprs = RemoveCommonExpressionsFromUseBases(UsersToProcess);
868 // Next, figure out what we can represent in the immediate fields of
869 // instructions. If we can represent anything there, move it to the imm
870 // fields of the BasedUsers. We do this so that it increases the commonality
871 // of the remaining uses.
872 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
873 // If the user is not in the current loop, this means it is using the exit
874 // value of the IV. Do not put anything in the base, make sure it's all in
875 // the immediate field to allow as much factoring as possible.
876 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
877 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
878 UsersToProcess[i].Base);
879 UsersToProcess[i].Base =
880 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
883 // Addressing modes can be folded into loads and stores. Be careful that
884 // the store is through the expression, not of the expression though.
885 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
886 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
887 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
890 MoveImmediateValues(TLI, UsersToProcess[i].Base, UsersToProcess[i].Imm,
895 // Now that we know what we need to do, insert the PHI node itself.
897 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
898 << *CommonExprs << " :\n");
900 SCEVExpander Rewriter(*SE, *LI);
901 SCEVExpander PreheaderRewriter(*SE, *LI);
903 BasicBlock *Preheader = L->getLoopPreheader();
904 Instruction *PreInsertPt = Preheader->getTerminator();
905 Instruction *PhiInsertBefore = L->getHeader()->begin();
907 BasicBlock *LatchBlock = L->getLoopLatch();
909 // Create a new Phi for this base, and stick it in the loop header.
910 const Type *ReplacedTy = CommonExprs->getType();
911 PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
914 // Insert the stride into the preheader.
915 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
917 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
920 // Emit the initial base value into the loop preheader, and add it to the
922 Value *PHIBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
924 NewPHI->addIncoming(PHIBaseV, Preheader);
926 // Emit the increment of the base value before the terminator of the loop
927 // latch block, and add it to the Phi node.
928 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
929 SCEVUnknown::get(StrideV));
931 Value *IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
933 IncV->setName(NewPHI->getName()+".inc");
934 NewPHI->addIncoming(IncV, LatchBlock);
936 // Sort by the base value, so that all IVs with identical bases are next to
938 while (!UsersToProcess.empty()) {
939 SCEVHandle Base = UsersToProcess.back().Base;
941 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
943 // Emit the code for Base into the preheader.
944 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
947 // If BaseV is a constant other than 0, make sure that it gets inserted into
948 // the preheader, instead of being forward substituted into the uses. We do
949 // this by forcing a noop cast to be inserted into the preheader in this
951 if (Constant *C = dyn_cast<Constant>(BaseV))
952 if (!C->isNullValue() && !isTargetConstant(Base, TLI)) {
953 // We want this constant emitted into the preheader!
954 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
958 // Emit the code to add the immediate offset to the Phi value, just before
959 // the instructions that we identified as using this stride and base.
960 unsigned ScanPos = 0;
962 BasedUser &User = UsersToProcess.back();
964 // If this instruction wants to use the post-incremented value, move it
965 // after the post-inc and use its value instead of the PHI.
966 Value *RewriteOp = NewPHI;
967 if (User.isUseOfPostIncrementedValue) {
970 // If this user is in the loop, make sure it is the last thing in the
971 // loop to ensure it is dominated by the increment.
972 if (L->contains(User.Inst->getParent()))
973 User.Inst->moveBefore(LatchBlock->getTerminator());
975 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
977 // Clear the SCEVExpander's expression map so that we are guaranteed
978 // to have the code emitted where we expect it.
981 // Now that we know what we need to do, insert code before User for the
982 // immediate and any loop-variant expressions.
983 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
984 // Add BaseV to the PHI value if needed.
985 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
987 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
989 // Mark old value we replaced as possibly dead, so that it is elminated
990 // if we just replaced the last use of that value.
991 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
993 UsersToProcess.pop_back();
996 // If there are any more users to process with the same base, move one of
997 // them to the end of the list so that we will process it.
998 if (!UsersToProcess.empty()) {
999 for (unsigned e = UsersToProcess.size(); ScanPos != e; ++ScanPos)
1000 if (UsersToProcess[ScanPos].Base == Base) {
1001 std::swap(UsersToProcess[ScanPos], UsersToProcess.back());
1005 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1006 // TODO: Next, find out which base index is the most common, pull it out.
1009 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1010 // different starting values, into different PHIs.
1013 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1014 // uses in the loop, look to see if we can eliminate some, in favor of using
1015 // common indvars for the different uses.
1016 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1017 // TODO: implement optzns here.
1022 // Finally, get the terminating condition for the loop if possible. If we
1023 // can, we want to change it to use a post-incremented version of its
1024 // induction variable, to allow coallescing the live ranges for the IV into
1025 // one register value.
1026 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1027 BasicBlock *Preheader = L->getLoopPreheader();
1028 BasicBlock *LatchBlock =
1029 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1030 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1031 if (!TermBr || TermBr->isUnconditional() ||
1032 !isa<SetCondInst>(TermBr->getCondition()))
1034 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
1036 // Search IVUsesByStride to find Cond's IVUse if there is one.
1037 IVStrideUse *CondUse = 0;
1038 const SCEVHandle *CondStride = 0;
1040 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1042 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1043 IVUsesByStride.find(StrideOrder[Stride]);
1044 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1046 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1047 E = SI->second.Users.end(); UI != E; ++UI)
1048 if (UI->User == Cond) {
1050 CondStride = &SI->first;
1051 // NOTE: we could handle setcc instructions with multiple uses here, but
1052 // InstCombine does it as well for simple uses, it's not clear that it
1053 // occurs enough in real life to handle.
1057 if (!CondUse) return; // setcc doesn't use the IV.
1059 // setcc stride is complex, don't mess with users.
1060 // FIXME: Evaluate whether this is a good idea or not.
1061 if (!isa<SCEVConstant>(*CondStride)) return;
1063 // It's possible for the setcc instruction to be anywhere in the loop, and
1064 // possible for it to have multiple users. If it is not immediately before
1065 // the latch block branch, move it.
1066 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1067 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1068 Cond->moveBefore(TermBr);
1070 // Otherwise, clone the terminating condition and insert into the loopend.
1071 Cond = cast<SetCondInst>(Cond->clone());
1072 Cond->setName(L->getHeader()->getName() + ".termcond");
1073 LatchBlock->getInstList().insert(TermBr, Cond);
1075 // Clone the IVUse, as the old use still exists!
1076 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1077 CondUse->OperandValToReplace);
1078 CondUse = &IVUsesByStride[*CondStride].Users.back();
1082 // If we get to here, we know that we can transform the setcc instruction to
1083 // use the post-incremented version of the IV, allowing us to coallesce the
1084 // live ranges for the IV correctly.
1085 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1086 CondUse->isUseOfPostIncrementedValue = true;
1089 void LoopStrengthReduce::runOnLoop(Loop *L) {
1090 // First step, transform all loops nesting inside of this loop.
1091 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
1094 // Next, find all uses of induction variables in this loop, and catagorize
1095 // them by stride. Start by finding all of the PHI nodes in the header for
1096 // this loop. If they are induction variables, inspect their uses.
1097 std::set<Instruction*> Processed; // Don't reprocess instructions.
1098 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1099 AddUsersIfInteresting(I, L, Processed);
1101 // If we have nothing to do, return.
1102 if (IVUsesByStride.empty()) return;
1104 // Optimize induction variables. Some indvar uses can be transformed to use
1105 // strides that will be needed for other purposes. A common example of this
1106 // is the exit test for the loop, which can often be rewritten to use the
1107 // computation of some other indvar to decide when to terminate the loop.
1111 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1112 // doing computation in byte values, promote to 32-bit values if safe.
1114 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1115 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1116 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1117 // to be careful that IV's are all the same type. Only works for intptr_t
1120 // If we only have one stride, we can more aggressively eliminate some things.
1121 bool HasOneStride = IVUsesByStride.size() == 1;
1123 // Note: this processes each stride/type pair individually. All users passed
1124 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1125 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1126 // This extra layer of indirection makes the ordering of strides deterministic
1127 // - not dependent on map order.
1128 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1129 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1130 IVUsesByStride.find(StrideOrder[Stride]);
1131 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1132 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1135 // Clean up after ourselves
1136 if (!DeadInsts.empty()) {
1137 DeleteTriviallyDeadInstructions(DeadInsts);
1139 BasicBlock::iterator I = L->getHeader()->begin();
1141 while ((PN = dyn_cast<PHINode>(I))) {
1142 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1144 // At this point, we know that we have killed one or more GEP
1145 // instructions. It is worth checking to see if the cann indvar is also
1146 // dead, so that we can remove it as well. The requirements for the cann
1147 // indvar to be considered dead are:
1148 // 1. the cann indvar has one use
1149 // 2. the use is an add instruction
1150 // 3. the add has one use
1151 // 4. the add is used by the cann indvar
1152 // If all four cases above are true, then we can remove both the add and
1154 // FIXME: this needs to eliminate an induction variable even if it's being
1155 // compared against some value to decide loop termination.
1156 if (PN->hasOneUse()) {
1157 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1158 if (BO && BO->hasOneUse()) {
1159 if (PN == *(BO->use_begin())) {
1160 DeadInsts.insert(BO);
1161 // Break the cycle, then delete the PHI.
1162 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1163 SE->deleteInstructionFromRecords(PN);
1164 PN->eraseFromParent();
1169 DeleteTriviallyDeadInstructions(DeadInsts);
1172 CastedPointers.clear();
1173 IVUsesByStride.clear();
1174 StrideOrder.clear();