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"
39 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
40 Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
41 Statistic<> NumVariable("loop-reduce","Number of PHIs with variable strides");
43 /// IVStrideUse - Keep track of one use of a strided induction variable, where
44 /// the stride is stored externally. The Offset member keeps track of the
45 /// offset from the IV, User is the actual user of the operand, and 'Operand'
46 /// is the operand # of the User that is the use.
50 Value *OperandValToReplace;
52 // isUseOfPostIncrementedValue - True if this should use the
53 // post-incremented version of this IV, not the preincremented version.
54 // This can only be set in special cases, such as the terminating setcc
55 // instruction for a loop or uses dominated by the loop.
56 bool isUseOfPostIncrementedValue;
58 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
59 : Offset(Offs), User(U), OperandValToReplace(O),
60 isUseOfPostIncrementedValue(false) {}
63 /// IVUsersOfOneStride - This structure keeps track of all instructions that
64 /// have an operand that is based on the trip count multiplied by some stride.
65 /// The stride for all of these users is common and kept external to this
67 struct IVUsersOfOneStride {
68 /// Users - Keep track of all of the users of this stride as well as the
69 /// initial value and the operand that uses the IV.
70 std::vector<IVStrideUse> Users;
72 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
73 Users.push_back(IVStrideUse(Offset, User, Operand));
78 class LoopStrengthReduce : public FunctionPass {
83 const Type *UIntPtrTy;
86 /// MaxTargetAMSize - This is the maximum power-of-two scale value that the
87 /// target can handle for free with its addressing modes.
88 unsigned MaxTargetAMSize;
90 /// IVUsesByStride - Keep track of all uses of induction variables that we
91 /// are interested in. The key of the map is the stride of the access.
92 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
94 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
95 /// of the casted version of each value. This is accessed by
96 /// getCastedVersionOf.
97 std::map<Value*, Value*> CastedPointers;
99 /// DeadInsts - Keep track of instructions we may have made dead, so that
100 /// we can remove them after we are done working.
101 std::set<Instruction*> DeadInsts;
103 LoopStrengthReduce(unsigned MTAMS = 1)
104 : MaxTargetAMSize(MTAMS) {
107 virtual bool runOnFunction(Function &) {
108 LI = &getAnalysis<LoopInfo>();
109 DS = &getAnalysis<DominatorSet>();
110 SE = &getAnalysis<ScalarEvolution>();
111 TD = &getAnalysis<TargetData>();
112 UIntPtrTy = TD->getIntPtrType();
115 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
121 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
122 // We split critical edges, so we change the CFG. However, we do update
123 // many analyses if they are around.
124 AU.addPreservedID(LoopSimplifyID);
125 AU.addPreserved<LoopInfo>();
126 AU.addPreserved<DominatorSet>();
127 AU.addPreserved<ImmediateDominators>();
128 AU.addPreserved<DominanceFrontier>();
129 AU.addPreserved<DominatorTree>();
131 AU.addRequiredID(LoopSimplifyID);
132 AU.addRequired<LoopInfo>();
133 AU.addRequired<DominatorSet>();
134 AU.addRequired<TargetData>();
135 AU.addRequired<ScalarEvolution>();
138 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
140 Value *getCastedVersionOf(Value *V);
142 void runOnLoop(Loop *L);
143 bool AddUsersIfInteresting(Instruction *I, Loop *L,
144 std::set<Instruction*> &Processed);
145 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
147 void OptimizeIndvars(Loop *L);
149 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
150 IVUsersOfOneStride &Uses,
151 Loop *L, bool isOnlyStride);
152 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
154 RegisterOpt<LoopStrengthReduce> X("loop-reduce",
155 "Loop Strength Reduction");
158 FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize) {
159 return new LoopStrengthReduce(MaxTargetAMSize);
162 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
164 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
165 if (V->getType() == UIntPtrTy) return V;
166 if (Constant *CB = dyn_cast<Constant>(V))
167 return ConstantExpr::getCast(CB, UIntPtrTy);
169 Value *&New = CastedPointers[V];
172 BasicBlock::iterator InsertPt;
173 if (Argument *Arg = dyn_cast<Argument>(V)) {
174 // Insert into the entry of the function, after any allocas.
175 InsertPt = Arg->getParent()->begin()->begin();
176 while (isa<AllocaInst>(InsertPt)) ++InsertPt;
178 if (InvokeInst *II = dyn_cast<InvokeInst>(V)) {
179 InsertPt = II->getNormalDest()->begin();
181 InsertPt = cast<Instruction>(V);
185 // Do not insert casts into the middle of PHI node blocks.
186 while (isa<PHINode>(InsertPt)) ++InsertPt;
189 New = new CastInst(V, UIntPtrTy, V->getName(), InsertPt);
190 DeadInsts.insert(cast<Instruction>(New));
195 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
196 /// specified set are trivially dead, delete them and see if this makes any of
197 /// their operands subsequently dead.
198 void LoopStrengthReduce::
199 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
200 while (!Insts.empty()) {
201 Instruction *I = *Insts.begin();
202 Insts.erase(Insts.begin());
203 if (isInstructionTriviallyDead(I)) {
204 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
205 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
207 SE->deleteInstructionFromRecords(I);
208 I->eraseFromParent();
215 /// GetExpressionSCEV - Compute and return the SCEV for the specified
217 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
218 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
219 // If this is a GEP that SE doesn't know about, compute it now and insert it.
220 // If this is not a GEP, or if we have already done this computation, just let
222 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
223 if (!GEP || SE->hasSCEV(GEP))
224 return SE->getSCEV(Exp);
226 // Analyze all of the subscripts of this getelementptr instruction, looking
227 // for uses that are determined by the trip count of L. First, skip all
228 // operands the are not dependent on the IV.
230 // Build up the base expression. Insert an LLVM cast of the pointer to
232 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
234 gep_type_iterator GTI = gep_type_begin(GEP);
236 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
237 // If this is a use of a recurrence that we can analyze, and it comes before
238 // Op does in the GEP operand list, we will handle this when we process this
240 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
241 const StructLayout *SL = TD->getStructLayout(STy);
242 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
243 uint64_t Offset = SL->MemberOffsets[Idx];
244 GEPVal = SCEVAddExpr::get(GEPVal,
245 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
247 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
248 SCEVHandle Idx = SE->getSCEV(OpVal);
250 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
252 Idx = SCEVMulExpr::get(Idx,
253 SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
255 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
259 SE->setSCEV(GEP, GEPVal);
263 /// getSCEVStartAndStride - Compute the start and stride of this expression,
264 /// returning false if the expression is not a start/stride pair, or true if it
265 /// is. The stride must be a loop invariant expression, but the start may be
266 /// a mix of loop invariant and loop variant expressions.
267 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
268 SCEVHandle &Start, SCEVHandle &Stride) {
269 SCEVHandle TheAddRec = Start; // Initialize to zero.
271 // If the outer level is an AddExpr, the operands are all start values except
272 // for a nested AddRecExpr.
273 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
274 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
275 if (SCEVAddRecExpr *AddRec =
276 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
277 if (AddRec->getLoop() == L)
278 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
280 return false; // Nested IV of some sort?
282 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
285 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
288 return false; // not analyzable.
291 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
292 if (!AddRec || AddRec->getLoop() != L) return false;
294 // FIXME: Generalize to non-affine IV's.
295 if (!AddRec->isAffine()) return false;
297 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
299 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
300 DEBUG(std::cerr << "[" << L->getHeader()->getName()
301 << "] Variable stride: " << *AddRec << "\n");
303 Stride = AddRec->getOperand(1);
304 // Check that all constant strides are the unsigned type, we don't want to
305 // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
307 assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
308 "Constants should be canonicalized to unsigned!");
313 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
314 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
315 /// return true. Otherwise, return false.
316 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
317 std::set<Instruction*> &Processed) {
318 if (I->getType() == Type::VoidTy) return false;
319 if (!Processed.insert(I).second)
320 return true; // Instruction already handled.
322 // Get the symbolic expression for this instruction.
323 SCEVHandle ISE = GetExpressionSCEV(I, L);
324 if (isa<SCEVCouldNotCompute>(ISE)) return false;
326 // Get the start and stride for this expression.
327 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
328 SCEVHandle Stride = Start;
329 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
330 return false; // Non-reducible symbolic expression, bail out.
332 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
333 Instruction *User = cast<Instruction>(*UI);
335 // Do not infinitely recurse on PHI nodes.
336 if (isa<PHINode>(User) && Processed.count(User))
339 // If this is an instruction defined in a nested loop, or outside this loop,
340 // don't recurse into it.
341 bool AddUserToIVUsers = false;
342 if (LI->getLoopFor(User->getParent()) != L) {
343 DEBUG(std::cerr << "FOUND USER in other loop: " << *User
344 << " OF SCEV: " << *ISE << "\n");
345 AddUserToIVUsers = true;
346 } else if (!AddUsersIfInteresting(User, L, Processed)) {
347 DEBUG(std::cerr << "FOUND USER: " << *User
348 << " OF SCEV: " << *ISE << "\n");
349 AddUserToIVUsers = true;
352 if (AddUserToIVUsers) {
353 // Okay, we found a user that we cannot reduce. Analyze the instruction
354 // and decide what to do with it. If we are a use inside of the loop, use
355 // the value before incrementation, otherwise use it after incrementation.
356 if (L->contains(User->getParent()) ||
357 // Alternatively, if we are a use outside of the loop, but is not
358 // dominated by the latch block, we have to use the preincremented
360 !DS->dominates(L->getLoopLatch(), User->getParent())) {
361 IVUsesByStride[Stride].addUser(Start, User, I);
363 // The value used will be incremented by the stride more than we are
364 // expecting, so subtract this off.
365 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
366 IVUsesByStride[Stride].addUser(NewStart, User, I);
367 IVUsesByStride[Stride].Users.back().isUseOfPostIncrementedValue = true;
375 /// BasedUser - For a particular base value, keep information about how we've
376 /// partitioned the expression so far.
378 /// Base - The Base value for the PHI node that needs to be inserted for
379 /// this use. As the use is processed, information gets moved from this
380 /// field to the Imm field (below). BasedUser values are sorted by this
384 /// Inst - The instruction using the induction variable.
387 /// OperandValToReplace - The operand value of Inst to replace with the
389 Value *OperandValToReplace;
391 /// Imm - The immediate value that should be added to the base immediately
392 /// before Inst, because it will be folded into the imm field of the
396 /// EmittedBase - The actual value* to use for the base value of this
397 /// operation. This is null if we should just use zero so far.
400 // isUseOfPostIncrementedValue - True if this should use the
401 // post-incremented version of this IV, not the preincremented version.
402 // This can only be set in special cases, such as the terminating setcc
403 // instruction for a loop and uses outside the loop that are dominated by
405 bool isUseOfPostIncrementedValue;
407 BasedUser(IVStrideUse &IVSU)
408 : Base(IVSU.Offset), Inst(IVSU.User),
409 OperandValToReplace(IVSU.OperandValToReplace),
410 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
411 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
413 // Once we rewrite the code to insert the new IVs we want, update the
414 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
416 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
417 SCEVExpander &Rewriter, Loop *L,
420 // Sort by the Base field.
421 bool operator<(const BasedUser &BU) const { return Base < BU.Base; }
427 void BasedUser::dump() const {
428 std::cerr << " Base=" << *Base;
429 std::cerr << " Imm=" << *Imm;
431 std::cerr << " EB=" << *EmittedBase;
433 std::cerr << " Inst: " << *Inst;
436 // Once we rewrite the code to insert the new IVs we want, update the
437 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
439 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
440 SCEVExpander &Rewriter,
442 if (!isa<PHINode>(Inst)) {
443 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
444 Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, Inst,
445 OperandValToReplace->getType());
446 // Replace the use of the operand Value with the new Phi we just created.
447 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
448 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
452 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
453 // expression into each operand block that uses it. Note that PHI nodes can
454 // have multiple entries for the same predecessor. We use a map to make sure
455 // that a PHI node only has a single Value* for each predecessor (which also
456 // prevents us from inserting duplicate code in some blocks).
457 std::map<BasicBlock*, Value*> InsertedCode;
458 PHINode *PN = cast<PHINode>(Inst);
459 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
460 if (PN->getIncomingValue(i) == OperandValToReplace) {
461 // If this is a critical edge, split the edge so that we do not insert the
462 // code on all predecessor/successor paths. We do this unless this is the
463 // canonical backedge for this loop, as this can make some inserted code
464 // be in an illegal position.
465 BasicBlock *PHIPred = PN->getIncomingBlock(i);
466 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
467 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
470 // First step, split the critical edge.
471 SplitCriticalEdge(PHIPred, PN->getParent(), P);
473 // Next step: move the basic block. In particular, if the PHI node
474 // is outside of the loop, and PredTI is in the loop, we want to
475 // move the block to be immediately before the PHI block, not
476 // immediately after PredTI.
477 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
478 BasicBlock *NewBB = PN->getIncomingBlock(i);
479 NewBB->moveBefore(PN->getParent());
483 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
485 // Insert the code into the end of the predecessor block.
486 BasicBlock::iterator InsertPt =PN->getIncomingBlock(i)->getTerminator();
488 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
489 Code = Rewriter.expandCodeFor(NewValSCEV, InsertPt,
490 OperandValToReplace->getType());
493 // Replace the use of the operand Value with the new Phi we just created.
494 PN->setIncomingValue(i, Code);
498 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
502 /// isTargetConstant - Return true if the following can be referenced by the
503 /// immediate field of a target instruction.
504 static bool isTargetConstant(const SCEVHandle &V) {
506 // FIXME: Look at the target to decide if &GV is a legal constant immediate.
507 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
508 // PPC allows a sign-extended 16-bit immediate field.
509 if ((int64_t)SC->getValue()->getRawValue() > -(1 << 16) &&
510 (int64_t)SC->getValue()->getRawValue() < (1 << 16)-1)
515 return false; // ENABLE this for x86
517 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
518 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
519 if (CE->getOpcode() == Instruction::Cast)
520 if (isa<GlobalValue>(CE->getOperand(0)))
521 // FIXME: should check to see that the dest is uintptr_t!
526 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
527 /// loop varying to the Imm operand.
528 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
530 if (Val->isLoopInvariant(L)) return; // Nothing to do.
532 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
533 std::vector<SCEVHandle> NewOps;
534 NewOps.reserve(SAE->getNumOperands());
536 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
537 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
538 // If this is a loop-variant expression, it must stay in the immediate
539 // field of the expression.
540 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
542 NewOps.push_back(SAE->getOperand(i));
546 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
548 Val = SCEVAddExpr::get(NewOps);
549 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
550 // Try to pull immediates out of the start value of nested addrec's.
551 SCEVHandle Start = SARE->getStart();
552 MoveLoopVariantsToImediateField(Start, Imm, L);
554 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
556 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
558 // Otherwise, all of Val is variant, move the whole thing over.
559 Imm = SCEVAddExpr::get(Imm, Val);
560 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
565 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
566 /// that can fit into the immediate field of instructions in the target.
567 /// Accumulate these immediate values into the Imm value.
568 static void MoveImmediateValues(SCEVHandle &Val, SCEVHandle &Imm,
569 bool isAddress, Loop *L) {
570 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
571 std::vector<SCEVHandle> NewOps;
572 NewOps.reserve(SAE->getNumOperands());
574 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
575 if (isAddress && isTargetConstant(SAE->getOperand(i))) {
576 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
577 } else if (!SAE->getOperand(i)->isLoopInvariant(L)) {
578 // If this is a loop-variant expression, it must stay in the immediate
579 // field of the expression.
580 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
582 NewOps.push_back(SAE->getOperand(i));
586 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
588 Val = SCEVAddExpr::get(NewOps);
590 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
591 // Try to pull immediates out of the start value of nested addrec's.
592 SCEVHandle Start = SARE->getStart();
593 MoveImmediateValues(Start, Imm, isAddress, L);
595 if (Start != SARE->getStart()) {
596 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
598 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
603 // Loop-variant expressions must stay in the immediate field of the
605 if ((isAddress && isTargetConstant(Val)) ||
606 !Val->isLoopInvariant(L)) {
607 Imm = SCEVAddExpr::get(Imm, Val);
608 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
612 // Otherwise, no immediates to move.
616 /// IncrementAddExprUses - Decompose the specified expression into its added
617 /// subexpressions, and increment SubExpressionUseCounts for each of these
618 /// decomposed parts.
619 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
621 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
622 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
623 SeparateSubExprs(SubExprs, AE->getOperand(j));
624 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
625 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
626 if (SARE->getOperand(0) == Zero) {
627 SubExprs.push_back(Expr);
629 // Compute the addrec with zero as its base.
630 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
631 Ops[0] = Zero; // Start with zero base.
632 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
635 SeparateSubExprs(SubExprs, SARE->getOperand(0));
637 } else if (!isa<SCEVConstant>(Expr) ||
638 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
640 SubExprs.push_back(Expr);
645 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
646 /// removing any common subexpressions from it. Anything truly common is
647 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
648 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
650 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
651 unsigned NumUses = Uses.size();
653 // Only one use? Use its base, regardless of what it is!
654 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
655 SCEVHandle Result = Zero;
657 std::swap(Result, Uses[0].Base);
661 // To find common subexpressions, count how many of Uses use each expression.
662 // If any subexpressions are used Uses.size() times, they are common.
663 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
665 std::vector<SCEVHandle> SubExprs;
666 for (unsigned i = 0; i != NumUses; ++i) {
667 // If the base is zero (which is common), return zero now, there are no
669 if (Uses[i].Base == Zero) return Zero;
671 // Split the expression into subexprs.
672 SeparateSubExprs(SubExprs, Uses[i].Base);
673 // Add one to SubExpressionUseCounts for each subexpr present.
674 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
675 SubExpressionUseCounts[SubExprs[j]]++;
680 // Now that we know how many times each is used, build Result.
681 for (std::map<SCEVHandle, unsigned>::iterator I =
682 SubExpressionUseCounts.begin(), E = SubExpressionUseCounts.end();
684 if (I->second == NumUses) { // Found CSE!
685 Result = SCEVAddExpr::get(Result, I->first);
688 // Remove non-cse's from SubExpressionUseCounts.
689 SubExpressionUseCounts.erase(I++);
692 // If we found no CSE's, return now.
693 if (Result == Zero) return Result;
695 // Otherwise, remove all of the CSE's we found from each of the base values.
696 for (unsigned i = 0; i != NumUses; ++i) {
697 // Split the expression into subexprs.
698 SeparateSubExprs(SubExprs, Uses[i].Base);
700 // Remove any common subexpressions.
701 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
702 if (SubExpressionUseCounts.count(SubExprs[j])) {
703 SubExprs.erase(SubExprs.begin()+j);
707 // Finally, the non-shared expressions together.
708 if (SubExprs.empty())
711 Uses[i].Base = SCEVAddExpr::get(SubExprs);
719 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
720 /// stride of IV. All of the users may have different starting values, and this
721 /// may not be the only stride (we know it is if isOnlyStride is true).
722 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
723 IVUsersOfOneStride &Uses,
726 // Transform our list of users and offsets to a bit more complex table. In
727 // this new vector, each 'BasedUser' contains 'Base' the base of the
728 // strided accessas well as the old information from Uses. We progressively
729 // move information from the Base field to the Imm field, until we eventually
730 // have the full access expression to rewrite the use.
731 std::vector<BasedUser> UsersToProcess;
732 UsersToProcess.reserve(Uses.Users.size());
733 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
734 UsersToProcess.push_back(Uses.Users[i]);
736 // Move any loop invariant operands from the offset field to the immediate
737 // field of the use, so that we don't try to use something before it is
739 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
740 UsersToProcess.back().Imm, L);
741 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
742 "Base value is not loop invariant!");
745 // We now have a whole bunch of uses of like-strided induction variables, but
746 // they might all have different bases. We want to emit one PHI node for this
747 // stride which we fold as many common expressions (between the IVs) into as
748 // possible. Start by identifying the common expressions in the base values
749 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
750 // "A+B"), emit it to the preheader, then remove the expression from the
751 // UsersToProcess base values.
752 SCEVHandle CommonExprs = RemoveCommonExpressionsFromUseBases(UsersToProcess);
754 // Next, figure out what we can represent in the immediate fields of
755 // instructions. If we can represent anything there, move it to the imm
756 // fields of the BasedUsers. We do this so that it increases the commonality
757 // of the remaining uses.
758 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
759 // If the user is not in the current loop, this means it is using the exit
760 // value of the IV. Do not put anything in the base, make sure it's all in
761 // the immediate field to allow as much factoring as possible.
762 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
763 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
764 UsersToProcess[i].Base);
765 UsersToProcess[i].Base =
766 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
769 // Addressing modes can be folded into loads and stores. Be careful that
770 // the store is through the expression, not of the expression though.
771 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
772 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
773 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
776 MoveImmediateValues(UsersToProcess[i].Base, UsersToProcess[i].Imm,
781 // Now that we know what we need to do, insert the PHI node itself.
783 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
784 << *CommonExprs << " :\n");
786 SCEVExpander Rewriter(*SE, *LI);
787 SCEVExpander PreheaderRewriter(*SE, *LI);
789 BasicBlock *Preheader = L->getLoopPreheader();
790 Instruction *PreInsertPt = Preheader->getTerminator();
791 Instruction *PhiInsertBefore = L->getHeader()->begin();
793 BasicBlock *LatchBlock = L->getLoopLatch();
795 // Create a new Phi for this base, and stick it in the loop header.
796 const Type *ReplacedTy = CommonExprs->getType();
797 PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
800 // Insert the stride into the preheader.
801 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
803 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
806 // Emit the initial base value into the loop preheader, and add it to the
808 Value *PHIBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
810 NewPHI->addIncoming(PHIBaseV, Preheader);
812 // Emit the increment of the base value before the terminator of the loop
813 // latch block, and add it to the Phi node.
814 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
815 SCEVUnknown::get(StrideV));
817 Value *IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
819 IncV->setName(NewPHI->getName()+".inc");
820 NewPHI->addIncoming(IncV, LatchBlock);
822 // Sort by the base value, so that all IVs with identical bases are next to
824 std::sort(UsersToProcess.begin(), UsersToProcess.end());
825 while (!UsersToProcess.empty()) {
826 SCEVHandle Base = UsersToProcess.front().Base;
828 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
830 // Emit the code for Base into the preheader.
831 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
834 // If BaseV is a constant other than 0, make sure that it gets inserted into
835 // the preheader, instead of being forward substituted into the uses. We do
836 // this by forcing a noop cast to be inserted into the preheader in this
838 if (Constant *C = dyn_cast<Constant>(BaseV))
839 if (!C->isNullValue() && !isTargetConstant(Base)) {
840 // We want this constant emitted into the preheader!
841 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
845 // Emit the code to add the immediate offset to the Phi value, just before
846 // the instructions that we identified as using this stride and base.
847 while (!UsersToProcess.empty() && UsersToProcess.front().Base == Base) {
848 BasedUser &User = UsersToProcess.front();
850 // If this instruction wants to use the post-incremented value, move it
851 // after the post-inc and use its value instead of the PHI.
852 Value *RewriteOp = NewPHI;
853 if (User.isUseOfPostIncrementedValue) {
856 // If this user is in the loop, make sure it is the last thing in the
857 // loop to ensure it is dominated by the increment.
858 if (L->contains(User.Inst->getParent()))
859 User.Inst->moveBefore(LatchBlock->getTerminator());
861 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
863 // Clear the SCEVExpander's expression map so that we are guaranteed
864 // to have the code emitted where we expect it.
867 // Now that we know what we need to do, insert code before User for the
868 // immediate and any loop-variant expressions.
869 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
870 // Add BaseV to the PHI value if needed.
871 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
873 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
875 // Mark old value we replaced as possibly dead, so that it is elminated
876 // if we just replaced the last use of that value.
877 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
879 UsersToProcess.erase(UsersToProcess.begin());
882 // TODO: Next, find out which base index is the most common, pull it out.
885 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
886 // different starting values, into different PHIs.
889 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
890 // uses in the loop, look to see if we can eliminate some, in favor of using
891 // common indvars for the different uses.
892 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
893 // TODO: implement optzns here.
898 // Finally, get the terminating condition for the loop if possible. If we
899 // can, we want to change it to use a post-incremented version of its
900 // induction variable, to allow coallescing the live ranges for the IV into
901 // one register value.
902 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
903 BasicBlock *Preheader = L->getLoopPreheader();
904 BasicBlock *LatchBlock =
905 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
906 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
907 if (!TermBr || TermBr->isUnconditional() ||
908 !isa<SetCondInst>(TermBr->getCondition()))
910 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
912 // Search IVUsesByStride to find Cond's IVUse if there is one.
913 IVStrideUse *CondUse = 0;
914 const SCEVHandle *CondStride = 0;
916 for (std::map<SCEVHandle, IVUsersOfOneStride>::iterator
917 I = IVUsesByStride.begin(), E = IVUsesByStride.end();
918 I != E && !CondUse; ++I)
919 for (std::vector<IVStrideUse>::iterator UI = I->second.Users.begin(),
920 E = I->second.Users.end(); UI != E; ++UI)
921 if (UI->User == Cond) {
923 CondStride = &I->first;
924 // NOTE: we could handle setcc instructions with multiple uses here, but
925 // InstCombine does it as well for simple uses, it's not clear that it
926 // occurs enough in real life to handle.
929 if (!CondUse) return; // setcc doesn't use the IV.
931 // setcc stride is complex, don't mess with users.
932 // FIXME: Evaluate whether this is a good idea or not.
933 if (!isa<SCEVConstant>(*CondStride)) return;
935 // It's possible for the setcc instruction to be anywhere in the loop, and
936 // possible for it to have multiple users. If it is not immediately before
937 // the latch block branch, move it.
938 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
939 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
940 Cond->moveBefore(TermBr);
942 // Otherwise, clone the terminating condition and insert into the loopend.
943 Cond = cast<SetCondInst>(Cond->clone());
944 Cond->setName(L->getHeader()->getName() + ".termcond");
945 LatchBlock->getInstList().insert(TermBr, Cond);
947 // Clone the IVUse, as the old use still exists!
948 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
949 CondUse->OperandValToReplace);
950 CondUse = &IVUsesByStride[*CondStride].Users.back();
954 // If we get to here, we know that we can transform the setcc instruction to
955 // use the post-incremented version of the IV, allowing us to coallesce the
956 // live ranges for the IV correctly.
957 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
958 CondUse->isUseOfPostIncrementedValue = true;
961 void LoopStrengthReduce::runOnLoop(Loop *L) {
962 // First step, transform all loops nesting inside of this loop.
963 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
966 // Next, find all uses of induction variables in this loop, and catagorize
967 // them by stride. Start by finding all of the PHI nodes in the header for
968 // this loop. If they are induction variables, inspect their uses.
969 std::set<Instruction*> Processed; // Don't reprocess instructions.
970 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
971 AddUsersIfInteresting(I, L, Processed);
973 // If we have nothing to do, return.
974 if (IVUsesByStride.empty()) return;
976 // Optimize induction variables. Some indvar uses can be transformed to use
977 // strides that will be needed for other purposes. A common example of this
978 // is the exit test for the loop, which can often be rewritten to use the
979 // computation of some other indvar to decide when to terminate the loop.
983 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
984 // doing computation in byte values, promote to 32-bit values if safe.
986 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
987 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
988 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
989 // to be careful that IV's are all the same type. Only works for intptr_t
992 // If we only have one stride, we can more aggressively eliminate some things.
993 bool HasOneStride = IVUsesByStride.size() == 1;
995 // Note: this processes each stride/type pair individually. All users passed
996 // into StrengthReduceStridedIVUsers have the same type AND stride.
997 for (std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI
998 = IVUsesByStride.begin(), E = IVUsesByStride.end(); SI != E; ++SI)
999 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1001 // Clean up after ourselves
1002 if (!DeadInsts.empty()) {
1003 DeleteTriviallyDeadInstructions(DeadInsts);
1005 BasicBlock::iterator I = L->getHeader()->begin();
1007 while ((PN = dyn_cast<PHINode>(I))) {
1008 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1010 // At this point, we know that we have killed one or more GEP
1011 // instructions. It is worth checking to see if the cann indvar is also
1012 // dead, so that we can remove it as well. The requirements for the cann
1013 // indvar to be considered dead are:
1014 // 1. the cann indvar has one use
1015 // 2. the use is an add instruction
1016 // 3. the add has one use
1017 // 4. the add is used by the cann indvar
1018 // If all four cases above are true, then we can remove both the add and
1020 // FIXME: this needs to eliminate an induction variable even if it's being
1021 // compared against some value to decide loop termination.
1022 if (PN->hasOneUse()) {
1023 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1024 if (BO && BO->hasOneUse()) {
1025 if (PN == *(BO->use_begin())) {
1026 DeadInsts.insert(BO);
1027 // Break the cycle, then delete the PHI.
1028 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1029 SE->deleteInstructionFromRecords(PN);
1030 PN->eraseFromParent();
1035 DeleteTriviallyDeadInstructions(DeadInsts);
1038 CastedPointers.clear();
1039 IVUsesByStride.clear();