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"
40 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
41 Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
42 Statistic<> NumVariable("loop-reduce","Number of PHIs with variable strides");
44 /// IVStrideUse - Keep track of one use of a strided induction variable, where
45 /// the stride is stored externally. The Offset member keeps track of the
46 /// offset from the IV, User is the actual user of the operand, and 'Operand'
47 /// is the operand # of the User that is the use.
51 Value *OperandValToReplace;
53 // isUseOfPostIncrementedValue - True if this should use the
54 // post-incremented version of this IV, not the preincremented version.
55 // This can only be set in special cases, such as the terminating setcc
56 // instruction for a loop or uses dominated by the loop.
57 bool isUseOfPostIncrementedValue;
59 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
60 : Offset(Offs), User(U), OperandValToReplace(O),
61 isUseOfPostIncrementedValue(false) {}
64 /// IVUsersOfOneStride - This structure keeps track of all instructions that
65 /// have an operand that is based on the trip count multiplied by some stride.
66 /// The stride for all of these users is common and kept external to this
68 struct IVUsersOfOneStride {
69 /// Users - Keep track of all of the users of this stride as well as the
70 /// initial value and the operand that uses the IV.
71 std::vector<IVStrideUse> Users;
73 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
74 Users.push_back(IVStrideUse(Offset, User, Operand));
79 class LoopStrengthReduce : public FunctionPass {
84 const Type *UIntPtrTy;
87 /// MaxTargetAMSize - This is the maximum power-of-two scale value that the
88 /// target can handle for free with its addressing modes.
89 unsigned MaxTargetAMSize;
91 /// IVUsesByStride - Keep track of all uses of induction variables that we
92 /// are interested in. The key of the map is the stride of the access.
93 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
95 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
96 /// We use this to iterate over the IVUsesByStride collection without being
97 /// dependent on random ordering of pointers in the process.
98 std::vector<SCEVHandle> StrideOrder;
100 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
101 /// of the casted version of each value. This is accessed by
102 /// getCastedVersionOf.
103 std::map<Value*, Value*> CastedPointers;
105 /// DeadInsts - Keep track of instructions we may have made dead, so that
106 /// we can remove them after we are done working.
107 std::set<Instruction*> DeadInsts;
109 LoopStrengthReduce(unsigned MTAMS = 1)
110 : MaxTargetAMSize(MTAMS) {
113 virtual bool runOnFunction(Function &) {
114 LI = &getAnalysis<LoopInfo>();
115 EF = &getAnalysis<ETForest>();
116 SE = &getAnalysis<ScalarEvolution>();
117 TD = &getAnalysis<TargetData>();
118 UIntPtrTy = TD->getIntPtrType();
121 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
127 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
128 // We split critical edges, so we change the CFG. However, we do update
129 // many analyses if they are around.
130 AU.addPreservedID(LoopSimplifyID);
131 AU.addPreserved<LoopInfo>();
132 AU.addPreserved<DominatorSet>();
133 AU.addPreserved<ETForest>();
134 AU.addPreserved<ImmediateDominators>();
135 AU.addPreserved<DominanceFrontier>();
136 AU.addPreserved<DominatorTree>();
138 AU.addRequiredID(LoopSimplifyID);
139 AU.addRequired<LoopInfo>();
140 AU.addRequired<ETForest>();
141 AU.addRequired<TargetData>();
142 AU.addRequired<ScalarEvolution>();
145 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
147 Value *getCastedVersionOf(Value *V);
149 void runOnLoop(Loop *L);
150 bool AddUsersIfInteresting(Instruction *I, Loop *L,
151 std::set<Instruction*> &Processed);
152 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
154 void OptimizeIndvars(Loop *L);
156 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
157 IVUsersOfOneStride &Uses,
158 Loop *L, bool isOnlyStride);
159 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
161 RegisterOpt<LoopStrengthReduce> X("loop-reduce",
162 "Loop Strength Reduction");
165 FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize) {
166 return new LoopStrengthReduce(MaxTargetAMSize);
169 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
171 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
172 if (V->getType() == UIntPtrTy) return V;
173 if (Constant *CB = dyn_cast<Constant>(V))
174 return ConstantExpr::getCast(CB, UIntPtrTy);
176 Value *&New = CastedPointers[V];
179 BasicBlock::iterator InsertPt;
180 if (Argument *Arg = dyn_cast<Argument>(V)) {
181 // Insert into the entry of the function, after any allocas.
182 InsertPt = Arg->getParent()->begin()->begin();
183 while (isa<AllocaInst>(InsertPt)) ++InsertPt;
185 if (InvokeInst *II = dyn_cast<InvokeInst>(V)) {
186 InsertPt = II->getNormalDest()->begin();
188 InsertPt = cast<Instruction>(V);
192 // Do not insert casts into the middle of PHI node blocks.
193 while (isa<PHINode>(InsertPt)) ++InsertPt;
196 New = new CastInst(V, UIntPtrTy, V->getName(), InsertPt);
197 DeadInsts.insert(cast<Instruction>(New));
202 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
203 /// specified set are trivially dead, delete them and see if this makes any of
204 /// their operands subsequently dead.
205 void LoopStrengthReduce::
206 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
207 while (!Insts.empty()) {
208 Instruction *I = *Insts.begin();
209 Insts.erase(Insts.begin());
210 if (isInstructionTriviallyDead(I)) {
211 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
212 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
214 SE->deleteInstructionFromRecords(I);
215 I->eraseFromParent();
222 /// GetExpressionSCEV - Compute and return the SCEV for the specified
224 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
225 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
226 // If this is a GEP that SE doesn't know about, compute it now and insert it.
227 // If this is not a GEP, or if we have already done this computation, just let
229 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
230 if (!GEP || SE->hasSCEV(GEP))
231 return SE->getSCEV(Exp);
233 // Analyze all of the subscripts of this getelementptr instruction, looking
234 // for uses that are determined by the trip count of L. First, skip all
235 // operands the are not dependent on the IV.
237 // Build up the base expression. Insert an LLVM cast of the pointer to
239 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
241 gep_type_iterator GTI = gep_type_begin(GEP);
243 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
244 // If this is a use of a recurrence that we can analyze, and it comes before
245 // Op does in the GEP operand list, we will handle this when we process this
247 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
248 const StructLayout *SL = TD->getStructLayout(STy);
249 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
250 uint64_t Offset = SL->MemberOffsets[Idx];
251 GEPVal = SCEVAddExpr::get(GEPVal,
252 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
254 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
255 SCEVHandle Idx = SE->getSCEV(OpVal);
257 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
259 Idx = SCEVMulExpr::get(Idx,
260 SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
262 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
266 SE->setSCEV(GEP, GEPVal);
270 /// getSCEVStartAndStride - Compute the start and stride of this expression,
271 /// returning false if the expression is not a start/stride pair, or true if it
272 /// is. The stride must be a loop invariant expression, but the start may be
273 /// a mix of loop invariant and loop variant expressions.
274 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
275 SCEVHandle &Start, SCEVHandle &Stride) {
276 SCEVHandle TheAddRec = Start; // Initialize to zero.
278 // If the outer level is an AddExpr, the operands are all start values except
279 // for a nested AddRecExpr.
280 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
281 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
282 if (SCEVAddRecExpr *AddRec =
283 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
284 if (AddRec->getLoop() == L)
285 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
287 return false; // Nested IV of some sort?
289 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
292 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
295 return false; // not analyzable.
298 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
299 if (!AddRec || AddRec->getLoop() != L) return false;
301 // FIXME: Generalize to non-affine IV's.
302 if (!AddRec->isAffine()) return false;
304 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
306 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
307 DEBUG(std::cerr << "[" << L->getHeader()->getName()
308 << "] Variable stride: " << *AddRec << "\n");
310 Stride = AddRec->getOperand(1);
311 // Check that all constant strides are the unsigned type, we don't want to
312 // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
314 assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
315 "Constants should be canonicalized to unsigned!");
320 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
321 /// and now we need to decide whether the user should use the preinc or post-inc
322 /// value. If this user should use the post-inc version of the IV, return true.
324 /// Choosing wrong here can break dominance properties (if we choose to use the
325 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
326 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
327 /// should use the post-inc value).
328 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
329 Loop *L, ETForest *EF, Pass *P) {
330 // If the user is in the loop, use the preinc value.
331 if (L->contains(User->getParent())) return false;
333 BasicBlock *LatchBlock = L->getLoopLatch();
335 // Ok, the user is outside of the loop. If it is dominated by the latch
336 // block, use the post-inc value.
337 if (EF->dominates(LatchBlock, User->getParent()))
340 // There is one case we have to be careful of: PHI nodes. These little guys
341 // can live in blocks that do not dominate the latch block, but (since their
342 // uses occur in the predecessor block, not the block the PHI lives in) should
343 // still use the post-inc value. Check for this case now.
344 PHINode *PN = dyn_cast<PHINode>(User);
345 if (!PN) return false; // not a phi, not dominated by latch block.
347 // Look at all of the uses of IV by the PHI node. If any use corresponds to
348 // a block that is not dominated by the latch block, give up and use the
349 // preincremented value.
350 unsigned NumUses = 0;
351 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
352 if (PN->getIncomingValue(i) == IV) {
354 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
358 // Okay, all uses of IV by PN are in predecessor blocks that really are
359 // dominated by the latch block. Split the critical edges and use the
360 // post-incremented value.
361 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
362 if (PN->getIncomingValue(i) == IV) {
363 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P);
364 if (--NumUses == 0) break;
372 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
373 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
374 /// return true. Otherwise, return false.
375 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
376 std::set<Instruction*> &Processed) {
377 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
378 return false; // Void and FP expressions cannot be reduced.
379 if (!Processed.insert(I).second)
380 return true; // Instruction already handled.
382 // Get the symbolic expression for this instruction.
383 SCEVHandle ISE = GetExpressionSCEV(I, L);
384 if (isa<SCEVCouldNotCompute>(ISE)) return false;
386 // Get the start and stride for this expression.
387 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
388 SCEVHandle Stride = Start;
389 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
390 return false; // Non-reducible symbolic expression, bail out.
392 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
393 Instruction *User = cast<Instruction>(*UI);
395 // Do not infinitely recurse on PHI nodes.
396 if (isa<PHINode>(User) && Processed.count(User))
399 // If this is an instruction defined in a nested loop, or outside this loop,
400 // don't recurse into it.
401 bool AddUserToIVUsers = false;
402 if (LI->getLoopFor(User->getParent()) != L) {
403 DEBUG(std::cerr << "FOUND USER in other loop: " << *User
404 << " OF SCEV: " << *ISE << "\n");
405 AddUserToIVUsers = true;
406 } else if (!AddUsersIfInteresting(User, L, Processed)) {
407 DEBUG(std::cerr << "FOUND USER: " << *User
408 << " OF SCEV: " << *ISE << "\n");
409 AddUserToIVUsers = true;
412 if (AddUserToIVUsers) {
413 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
414 if (StrideUses.Users.empty()) // First occurance of this stride?
415 StrideOrder.push_back(Stride);
417 // Okay, we found a user that we cannot reduce. Analyze the instruction
418 // and decide what to do with it. If we are a use inside of the loop, use
419 // the value before incrementation, otherwise use it after incrementation.
420 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
421 // The value used will be incremented by the stride more than we are
422 // expecting, so subtract this off.
423 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
424 StrideUses.addUser(NewStart, User, I);
425 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
426 DEBUG(std::cerr << " USING POSTINC SCEV, START=" << *NewStart<< "\n");
428 StrideUses.addUser(Start, User, I);
436 /// BasedUser - For a particular base value, keep information about how we've
437 /// partitioned the expression so far.
439 /// Base - The Base value for the PHI node that needs to be inserted for
440 /// this use. As the use is processed, information gets moved from this
441 /// field to the Imm field (below). BasedUser values are sorted by this
445 /// Inst - The instruction using the induction variable.
448 /// OperandValToReplace - The operand value of Inst to replace with the
450 Value *OperandValToReplace;
452 /// Imm - The immediate value that should be added to the base immediately
453 /// before Inst, because it will be folded into the imm field of the
457 /// EmittedBase - The actual value* to use for the base value of this
458 /// operation. This is null if we should just use zero so far.
461 // isUseOfPostIncrementedValue - True if this should use the
462 // post-incremented version of this IV, not the preincremented version.
463 // This can only be set in special cases, such as the terminating setcc
464 // instruction for a loop and uses outside the loop that are dominated by
466 bool isUseOfPostIncrementedValue;
468 BasedUser(IVStrideUse &IVSU)
469 : Base(IVSU.Offset), Inst(IVSU.User),
470 OperandValToReplace(IVSU.OperandValToReplace),
471 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
472 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
474 // Once we rewrite the code to insert the new IVs we want, update the
475 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
477 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
478 SCEVExpander &Rewriter, Loop *L,
484 void BasedUser::dump() const {
485 std::cerr << " Base=" << *Base;
486 std::cerr << " Imm=" << *Imm;
488 std::cerr << " EB=" << *EmittedBase;
490 std::cerr << " Inst: " << *Inst;
493 // Once we rewrite the code to insert the new IVs we want, update the
494 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
496 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
497 SCEVExpander &Rewriter,
499 if (!isa<PHINode>(Inst)) {
500 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
501 Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, Inst,
502 OperandValToReplace->getType());
503 // Replace the use of the operand Value with the new Phi we just created.
504 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
505 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
509 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
510 // expression into each operand block that uses it. Note that PHI nodes can
511 // have multiple entries for the same predecessor. We use a map to make sure
512 // that a PHI node only has a single Value* for each predecessor (which also
513 // prevents us from inserting duplicate code in some blocks).
514 std::map<BasicBlock*, Value*> InsertedCode;
515 PHINode *PN = cast<PHINode>(Inst);
516 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
517 if (PN->getIncomingValue(i) == OperandValToReplace) {
518 // If this is a critical edge, split the edge so that we do not insert the
519 // code on all predecessor/successor paths. We do this unless this is the
520 // canonical backedge for this loop, as this can make some inserted code
521 // be in an illegal position.
522 BasicBlock *PHIPred = PN->getIncomingBlock(i);
523 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
524 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
526 // First step, split the critical edge.
527 SplitCriticalEdge(PHIPred, PN->getParent(), P);
529 // Next step: move the basic block. In particular, if the PHI node
530 // is outside of the loop, and PredTI is in the loop, we want to
531 // move the block to be immediately before the PHI block, not
532 // immediately after PredTI.
533 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
534 BasicBlock *NewBB = PN->getIncomingBlock(i);
535 NewBB->moveBefore(PN->getParent());
539 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
541 // Insert the code into the end of the predecessor block.
542 BasicBlock::iterator InsertPt =PN->getIncomingBlock(i)->getTerminator();
544 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
545 Code = Rewriter.expandCodeFor(NewValSCEV, InsertPt,
546 OperandValToReplace->getType());
549 // Replace the use of the operand Value with the new Phi we just created.
550 PN->setIncomingValue(i, Code);
554 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
558 /// isTargetConstant - Return true if the following can be referenced by the
559 /// immediate field of a target instruction.
560 static bool isTargetConstant(const SCEVHandle &V) {
562 // FIXME: Look at the target to decide if &GV is a legal constant immediate.
563 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
564 // PPC allows a sign-extended 16-bit immediate field.
565 int64_t V = SC->getValue()->getSExtValue();
566 if (V > -(1 << 16) && V < (1 << 16)-1)
571 return false; // ENABLE this for x86
573 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
574 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
575 if (CE->getOpcode() == Instruction::Cast)
576 if (isa<GlobalValue>(CE->getOperand(0)))
577 // FIXME: should check to see that the dest is uintptr_t!
582 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
583 /// loop varying to the Imm operand.
584 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
586 if (Val->isLoopInvariant(L)) return; // Nothing to do.
588 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
589 std::vector<SCEVHandle> NewOps;
590 NewOps.reserve(SAE->getNumOperands());
592 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
593 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
594 // If this is a loop-variant expression, it must stay in the immediate
595 // field of the expression.
596 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
598 NewOps.push_back(SAE->getOperand(i));
602 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
604 Val = SCEVAddExpr::get(NewOps);
605 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
606 // Try to pull immediates out of the start value of nested addrec's.
607 SCEVHandle Start = SARE->getStart();
608 MoveLoopVariantsToImediateField(Start, Imm, L);
610 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
612 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
614 // Otherwise, all of Val is variant, move the whole thing over.
615 Imm = SCEVAddExpr::get(Imm, Val);
616 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
621 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
622 /// that can fit into the immediate field of instructions in the target.
623 /// Accumulate these immediate values into the Imm value.
624 static void MoveImmediateValues(SCEVHandle &Val, SCEVHandle &Imm,
625 bool isAddress, Loop *L) {
626 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
627 std::vector<SCEVHandle> NewOps;
628 NewOps.reserve(SAE->getNumOperands());
630 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
631 if (isAddress && isTargetConstant(SAE->getOperand(i))) {
632 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
633 } else if (!SAE->getOperand(i)->isLoopInvariant(L)) {
634 // If this is a loop-variant expression, it must stay in the immediate
635 // field of the expression.
636 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
638 NewOps.push_back(SAE->getOperand(i));
642 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
644 Val = SCEVAddExpr::get(NewOps);
646 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
647 // Try to pull immediates out of the start value of nested addrec's.
648 SCEVHandle Start = SARE->getStart();
649 MoveImmediateValues(Start, Imm, isAddress, L);
651 if (Start != SARE->getStart()) {
652 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
654 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
659 // Loop-variant expressions must stay in the immediate field of the
661 if ((isAddress && isTargetConstant(Val)) ||
662 !Val->isLoopInvariant(L)) {
663 Imm = SCEVAddExpr::get(Imm, Val);
664 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
668 // Otherwise, no immediates to move.
672 /// IncrementAddExprUses - Decompose the specified expression into its added
673 /// subexpressions, and increment SubExpressionUseCounts for each of these
674 /// decomposed parts.
675 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
677 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
678 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
679 SeparateSubExprs(SubExprs, AE->getOperand(j));
680 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
681 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
682 if (SARE->getOperand(0) == Zero) {
683 SubExprs.push_back(Expr);
685 // Compute the addrec with zero as its base.
686 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
687 Ops[0] = Zero; // Start with zero base.
688 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
691 SeparateSubExprs(SubExprs, SARE->getOperand(0));
693 } else if (!isa<SCEVConstant>(Expr) ||
694 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
696 SubExprs.push_back(Expr);
701 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
702 /// removing any common subexpressions from it. Anything truly common is
703 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
704 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
706 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
707 unsigned NumUses = Uses.size();
709 // Only one use? Use its base, regardless of what it is!
710 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
711 SCEVHandle Result = Zero;
713 std::swap(Result, Uses[0].Base);
717 // To find common subexpressions, count how many of Uses use each expression.
718 // If any subexpressions are used Uses.size() times, they are common.
719 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
721 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
722 // order we see them.
723 std::vector<SCEVHandle> UniqueSubExprs;
725 std::vector<SCEVHandle> SubExprs;
726 for (unsigned i = 0; i != NumUses; ++i) {
727 // If the base is zero (which is common), return zero now, there are no
729 if (Uses[i].Base == Zero) return Zero;
731 // Split the expression into subexprs.
732 SeparateSubExprs(SubExprs, Uses[i].Base);
733 // Add one to SubExpressionUseCounts for each subexpr present.
734 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
735 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
736 UniqueSubExprs.push_back(SubExprs[j]);
740 // Now that we know how many times each is used, build Result. Iterate over
741 // UniqueSubexprs so that we have a stable ordering.
742 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
743 std::map<SCEVHandle, unsigned>::iterator I =
744 SubExpressionUseCounts.find(UniqueSubExprs[i]);
745 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
746 if (I->second == NumUses) { // Found CSE!
747 Result = SCEVAddExpr::get(Result, I->first);
749 // Remove non-cse's from SubExpressionUseCounts.
750 SubExpressionUseCounts.erase(I);
754 // If we found no CSE's, return now.
755 if (Result == Zero) return Result;
757 // Otherwise, remove all of the CSE's we found from each of the base values.
758 for (unsigned i = 0; i != NumUses; ++i) {
759 // Split the expression into subexprs.
760 SeparateSubExprs(SubExprs, Uses[i].Base);
762 // Remove any common subexpressions.
763 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
764 if (SubExpressionUseCounts.count(SubExprs[j])) {
765 SubExprs.erase(SubExprs.begin()+j);
769 // Finally, the non-shared expressions together.
770 if (SubExprs.empty())
773 Uses[i].Base = SCEVAddExpr::get(SubExprs);
781 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
782 /// stride of IV. All of the users may have different starting values, and this
783 /// may not be the only stride (we know it is if isOnlyStride is true).
784 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
785 IVUsersOfOneStride &Uses,
788 // Transform our list of users and offsets to a bit more complex table. In
789 // this new vector, each 'BasedUser' contains 'Base' the base of the
790 // strided accessas well as the old information from Uses. We progressively
791 // move information from the Base field to the Imm field, until we eventually
792 // have the full access expression to rewrite the use.
793 std::vector<BasedUser> UsersToProcess;
794 UsersToProcess.reserve(Uses.Users.size());
795 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
796 UsersToProcess.push_back(Uses.Users[i]);
798 // Move any loop invariant operands from the offset field to the immediate
799 // field of the use, so that we don't try to use something before it is
801 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
802 UsersToProcess.back().Imm, L);
803 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
804 "Base value is not loop invariant!");
807 // We now have a whole bunch of uses of like-strided induction variables, but
808 // they might all have different bases. We want to emit one PHI node for this
809 // stride which we fold as many common expressions (between the IVs) into as
810 // possible. Start by identifying the common expressions in the base values
811 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
812 // "A+B"), emit it to the preheader, then remove the expression from the
813 // UsersToProcess base values.
814 SCEVHandle CommonExprs = RemoveCommonExpressionsFromUseBases(UsersToProcess);
816 // Next, figure out what we can represent in the immediate fields of
817 // instructions. If we can represent anything there, move it to the imm
818 // fields of the BasedUsers. We do this so that it increases the commonality
819 // of the remaining uses.
820 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
821 // If the user is not in the current loop, this means it is using the exit
822 // value of the IV. Do not put anything in the base, make sure it's all in
823 // the immediate field to allow as much factoring as possible.
824 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
825 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
826 UsersToProcess[i].Base);
827 UsersToProcess[i].Base =
828 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
831 // Addressing modes can be folded into loads and stores. Be careful that
832 // the store is through the expression, not of the expression though.
833 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
834 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
835 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
838 MoveImmediateValues(UsersToProcess[i].Base, UsersToProcess[i].Imm,
843 // Now that we know what we need to do, insert the PHI node itself.
845 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
846 << *CommonExprs << " :\n");
848 SCEVExpander Rewriter(*SE, *LI);
849 SCEVExpander PreheaderRewriter(*SE, *LI);
851 BasicBlock *Preheader = L->getLoopPreheader();
852 Instruction *PreInsertPt = Preheader->getTerminator();
853 Instruction *PhiInsertBefore = L->getHeader()->begin();
855 BasicBlock *LatchBlock = L->getLoopLatch();
857 // Create a new Phi for this base, and stick it in the loop header.
858 const Type *ReplacedTy = CommonExprs->getType();
859 PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
862 // Insert the stride into the preheader.
863 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
865 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
868 // Emit the initial base value into the loop preheader, and add it to the
870 Value *PHIBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
872 NewPHI->addIncoming(PHIBaseV, Preheader);
874 // Emit the increment of the base value before the terminator of the loop
875 // latch block, and add it to the Phi node.
876 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
877 SCEVUnknown::get(StrideV));
879 Value *IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
881 IncV->setName(NewPHI->getName()+".inc");
882 NewPHI->addIncoming(IncV, LatchBlock);
884 // Sort by the base value, so that all IVs with identical bases are next to
886 while (!UsersToProcess.empty()) {
887 SCEVHandle Base = UsersToProcess.back().Base;
889 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
891 // Emit the code for Base into the preheader.
892 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
895 // If BaseV is a constant other than 0, make sure that it gets inserted into
896 // the preheader, instead of being forward substituted into the uses. We do
897 // this by forcing a noop cast to be inserted into the preheader in this
899 if (Constant *C = dyn_cast<Constant>(BaseV))
900 if (!C->isNullValue() && !isTargetConstant(Base)) {
901 // We want this constant emitted into the preheader!
902 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
906 // Emit the code to add the immediate offset to the Phi value, just before
907 // the instructions that we identified as using this stride and base.
908 unsigned ScanPos = 0;
910 BasedUser &User = UsersToProcess.back();
912 // If this instruction wants to use the post-incremented value, move it
913 // after the post-inc and use its value instead of the PHI.
914 Value *RewriteOp = NewPHI;
915 if (User.isUseOfPostIncrementedValue) {
918 // If this user is in the loop, make sure it is the last thing in the
919 // loop to ensure it is dominated by the increment.
920 if (L->contains(User.Inst->getParent()))
921 User.Inst->moveBefore(LatchBlock->getTerminator());
923 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
925 // Clear the SCEVExpander's expression map so that we are guaranteed
926 // to have the code emitted where we expect it.
929 // Now that we know what we need to do, insert code before User for the
930 // immediate and any loop-variant expressions.
931 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
932 // Add BaseV to the PHI value if needed.
933 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
935 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
937 // Mark old value we replaced as possibly dead, so that it is elminated
938 // if we just replaced the last use of that value.
939 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
941 UsersToProcess.pop_back();
944 // If there are any more users to process with the same base, move one of
945 // them to the end of the list so that we will process it.
946 if (!UsersToProcess.empty()) {
947 for (unsigned e = UsersToProcess.size(); ScanPos != e; ++ScanPos)
948 if (UsersToProcess[ScanPos].Base == Base) {
949 std::swap(UsersToProcess[ScanPos], UsersToProcess.back());
953 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
954 // TODO: Next, find out which base index is the most common, pull it out.
957 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
958 // different starting values, into different PHIs.
961 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
962 // uses in the loop, look to see if we can eliminate some, in favor of using
963 // common indvars for the different uses.
964 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
965 // TODO: implement optzns here.
970 // Finally, get the terminating condition for the loop if possible. If we
971 // can, we want to change it to use a post-incremented version of its
972 // induction variable, to allow coallescing the live ranges for the IV into
973 // one register value.
974 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
975 BasicBlock *Preheader = L->getLoopPreheader();
976 BasicBlock *LatchBlock =
977 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
978 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
979 if (!TermBr || TermBr->isUnconditional() ||
980 !isa<SetCondInst>(TermBr->getCondition()))
982 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
984 // Search IVUsesByStride to find Cond's IVUse if there is one.
985 IVStrideUse *CondUse = 0;
986 const SCEVHandle *CondStride = 0;
988 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
990 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
991 IVUsesByStride.find(StrideOrder[Stride]);
992 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
994 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
995 E = SI->second.Users.end(); UI != E; ++UI)
996 if (UI->User == Cond) {
998 CondStride = &SI->first;
999 // NOTE: we could handle setcc instructions with multiple uses here, but
1000 // InstCombine does it as well for simple uses, it's not clear that it
1001 // occurs enough in real life to handle.
1005 if (!CondUse) return; // setcc doesn't use the IV.
1007 // setcc stride is complex, don't mess with users.
1008 // FIXME: Evaluate whether this is a good idea or not.
1009 if (!isa<SCEVConstant>(*CondStride)) return;
1011 // It's possible for the setcc instruction to be anywhere in the loop, and
1012 // possible for it to have multiple users. If it is not immediately before
1013 // the latch block branch, move it.
1014 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1015 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1016 Cond->moveBefore(TermBr);
1018 // Otherwise, clone the terminating condition and insert into the loopend.
1019 Cond = cast<SetCondInst>(Cond->clone());
1020 Cond->setName(L->getHeader()->getName() + ".termcond");
1021 LatchBlock->getInstList().insert(TermBr, Cond);
1023 // Clone the IVUse, as the old use still exists!
1024 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1025 CondUse->OperandValToReplace);
1026 CondUse = &IVUsesByStride[*CondStride].Users.back();
1030 // If we get to here, we know that we can transform the setcc instruction to
1031 // use the post-incremented version of the IV, allowing us to coallesce the
1032 // live ranges for the IV correctly.
1033 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1034 CondUse->isUseOfPostIncrementedValue = true;
1037 void LoopStrengthReduce::runOnLoop(Loop *L) {
1038 // First step, transform all loops nesting inside of this loop.
1039 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
1042 // Next, find all uses of induction variables in this loop, and catagorize
1043 // them by stride. Start by finding all of the PHI nodes in the header for
1044 // this loop. If they are induction variables, inspect their uses.
1045 std::set<Instruction*> Processed; // Don't reprocess instructions.
1046 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1047 AddUsersIfInteresting(I, L, Processed);
1049 // If we have nothing to do, return.
1050 if (IVUsesByStride.empty()) return;
1052 // Optimize induction variables. Some indvar uses can be transformed to use
1053 // strides that will be needed for other purposes. A common example of this
1054 // is the exit test for the loop, which can often be rewritten to use the
1055 // computation of some other indvar to decide when to terminate the loop.
1059 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1060 // doing computation in byte values, promote to 32-bit values if safe.
1062 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1063 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1064 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1065 // to be careful that IV's are all the same type. Only works for intptr_t
1068 // If we only have one stride, we can more aggressively eliminate some things.
1069 bool HasOneStride = IVUsesByStride.size() == 1;
1071 // Note: this processes each stride/type pair individually. All users passed
1072 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1073 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1074 // This extra layer of indirection makes the ordering of strides deterministic
1075 // - not dependent on map order.
1076 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1077 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1078 IVUsesByStride.find(StrideOrder[Stride]);
1079 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1080 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1083 // Clean up after ourselves
1084 if (!DeadInsts.empty()) {
1085 DeleteTriviallyDeadInstructions(DeadInsts);
1087 BasicBlock::iterator I = L->getHeader()->begin();
1089 while ((PN = dyn_cast<PHINode>(I))) {
1090 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1092 // At this point, we know that we have killed one or more GEP
1093 // instructions. It is worth checking to see if the cann indvar is also
1094 // dead, so that we can remove it as well. The requirements for the cann
1095 // indvar to be considered dead are:
1096 // 1. the cann indvar has one use
1097 // 2. the use is an add instruction
1098 // 3. the add has one use
1099 // 4. the add is used by the cann indvar
1100 // If all four cases above are true, then we can remove both the add and
1102 // FIXME: this needs to eliminate an induction variable even if it's being
1103 // compared against some value to decide loop termination.
1104 if (PN->hasOneUse()) {
1105 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1106 if (BO && BO->hasOneUse()) {
1107 if (PN == *(BO->use_begin())) {
1108 DeadInsts.insert(BO);
1109 // Break the cycle, then delete the PHI.
1110 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1111 SE->deleteInstructionFromRecords(PN);
1112 PN->eraseFromParent();
1117 DeleteTriviallyDeadInstructions(DeadInsts);
1120 CastedPointers.clear();
1121 IVUsesByStride.clear();
1122 StrideOrder.clear();