1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs a strength reduction on array references inside loops that
11 // have as one or more of their components the loop induction variable. This is
12 // accomplished by creating a new Value to hold the initial value of the array
13 // access for the first iteration, and then creating a new GEP instruction in
14 // the loop to increment the value by the appropriate amount.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "loop-reduce"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/Type.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/ScalarEvolutionExpander.h"
27 #include "llvm/Support/CFG.h"
28 #include "llvm/Support/GetElementPtrTypeIterator.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 #include "llvm/Target/TargetData.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Support/Debug.h"
38 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
39 Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
41 /// IVStrideUse - Keep track of one use of a strided induction variable, where
42 /// the stride is stored externally. The Offset member keeps track of the
43 /// offset from the IV, User is the actual user of the operand, and 'Operand'
44 /// is the operand # of the User that is the use.
48 Value *OperandValToReplace;
50 // isUseOfPostIncrementedValue - True if this should use the
51 // post-incremented version of this IV, not the preincremented version.
52 // This can only be set in special cases, such as the terminating setcc
53 // instruction for a loop.
54 bool isUseOfPostIncrementedValue;
56 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
57 : Offset(Offs), User(U), OperandValToReplace(O),
58 isUseOfPostIncrementedValue(false) {}
61 /// IVUsersOfOneStride - This structure keeps track of all instructions that
62 /// have an operand that is based on the trip count multiplied by some stride.
63 /// The stride for all of these users is common and kept external to this
65 struct IVUsersOfOneStride {
66 /// Users - Keep track of all of the users of this stride as well as the
67 /// initial value and the operand that uses the IV.
68 std::vector<IVStrideUse> Users;
70 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
71 Users.push_back(IVStrideUse(Offset, User, Operand));
76 class LoopStrengthReduce : public FunctionPass {
81 const Type *UIntPtrTy;
84 /// MaxTargetAMSize - This is the maximum power-of-two scale value that the
85 /// target can handle for free with its addressing modes.
86 unsigned MaxTargetAMSize;
88 /// IVUsesByStride - Keep track of all uses of induction variables that we
89 /// are interested in. The key of the map is the stride of the access.
90 std::map<Value*, IVUsersOfOneStride> IVUsesByStride;
92 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
93 /// of the casted version of each value. This is accessed by
94 /// getCastedVersionOf.
95 std::map<Value*, Value*> CastedPointers;
97 /// DeadInsts - Keep track of instructions we may have made dead, so that
98 /// we can remove them after we are done working.
99 std::set<Instruction*> DeadInsts;
101 LoopStrengthReduce(unsigned MTAMS = 1)
102 : MaxTargetAMSize(MTAMS) {
105 virtual bool runOnFunction(Function &) {
106 LI = &getAnalysis<LoopInfo>();
107 DS = &getAnalysis<DominatorSet>();
108 SE = &getAnalysis<ScalarEvolution>();
109 TD = &getAnalysis<TargetData>();
110 UIntPtrTy = TD->getIntPtrType();
113 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
119 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
120 AU.setPreservesCFG();
121 AU.addRequiredID(LoopSimplifyID);
122 AU.addRequired<LoopInfo>();
123 AU.addRequired<DominatorSet>();
124 AU.addRequired<TargetData>();
125 AU.addRequired<ScalarEvolution>();
128 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
130 Value *getCastedVersionOf(Value *V);
132 void runOnLoop(Loop *L);
133 bool AddUsersIfInteresting(Instruction *I, Loop *L,
134 std::set<Instruction*> &Processed);
135 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
137 void OptimizeIndvars(Loop *L);
139 void StrengthReduceStridedIVUsers(Value *Stride, IVUsersOfOneStride &Uses,
140 Loop *L, bool isOnlyStride);
141 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
143 RegisterOpt<LoopStrengthReduce> X("loop-reduce",
144 "Strength Reduce GEP Uses of Ind. Vars");
147 FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize) {
148 return new LoopStrengthReduce(MaxTargetAMSize);
151 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
153 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
154 if (V->getType() == UIntPtrTy) return V;
155 if (Constant *CB = dyn_cast<Constant>(V))
156 return ConstantExpr::getCast(CB, UIntPtrTy);
158 Value *&New = CastedPointers[V];
161 BasicBlock::iterator InsertPt;
162 if (Argument *Arg = dyn_cast<Argument>(V)) {
163 // Insert into the entry of the function, after any allocas.
164 InsertPt = Arg->getParent()->begin()->begin();
165 while (isa<AllocaInst>(InsertPt)) ++InsertPt;
167 if (InvokeInst *II = dyn_cast<InvokeInst>(V)) {
168 InsertPt = II->getNormalDest()->begin();
170 InsertPt = cast<Instruction>(V);
174 // Do not insert casts into the middle of PHI node blocks.
175 while (isa<PHINode>(InsertPt)) ++InsertPt;
178 New = new CastInst(V, UIntPtrTy, V->getName(), InsertPt);
179 DeadInsts.insert(cast<Instruction>(New));
184 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
185 /// specified set are trivially dead, delete them and see if this makes any of
186 /// their operands subsequently dead.
187 void LoopStrengthReduce::
188 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
189 while (!Insts.empty()) {
190 Instruction *I = *Insts.begin();
191 Insts.erase(Insts.begin());
192 if (isInstructionTriviallyDead(I)) {
193 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
194 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
196 SE->deleteInstructionFromRecords(I);
197 I->eraseFromParent();
204 /// GetExpressionSCEV - Compute and return the SCEV for the specified
206 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
207 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
209 return SE->getSCEV(Exp);
211 // Analyze all of the subscripts of this getelementptr instruction, looking
212 // for uses that are determined by the trip count of L. First, skip all
213 // operands the are not dependent on the IV.
215 // Build up the base expression. Insert an LLVM cast of the pointer to
217 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
219 gep_type_iterator GTI = gep_type_begin(GEP);
221 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
222 // If this is a use of a recurrence that we can analyze, and it comes before
223 // Op does in the GEP operand list, we will handle this when we process this
225 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
226 const StructLayout *SL = TD->getStructLayout(STy);
227 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue();
228 uint64_t Offset = SL->MemberOffsets[Idx];
229 GEPVal = SCEVAddExpr::get(GEPVal,
230 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
232 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
233 SCEVHandle Idx = SE->getSCEV(OpVal);
235 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
237 Idx = SCEVMulExpr::get(Idx,
238 SCEVConstant::get(ConstantUInt::get(UIntPtrTy,
240 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
247 /// getSCEVStartAndStride - Compute the start and stride of this expression,
248 /// returning false if the expression is not a start/stride pair, or true if it
249 /// is. The stride must be a loop invariant expression, but the start may be
250 /// a mix of loop invariant and loop variant expressions.
251 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
252 SCEVHandle &Start, Value *&Stride) {
253 SCEVHandle TheAddRec = Start; // Initialize to zero.
255 // If the outer level is an AddExpr, the operands are all start values except
256 // for a nested AddRecExpr.
257 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
258 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
259 if (SCEVAddRecExpr *AddRec =
260 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
261 if (AddRec->getLoop() == L)
262 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
264 return false; // Nested IV of some sort?
266 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
269 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
272 return false; // not analyzable.
275 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
276 if (!AddRec || AddRec->getLoop() != L) return false;
278 // FIXME: Generalize to non-affine IV's.
279 if (!AddRec->isAffine()) return false;
281 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
283 // FIXME: generalize to IV's with more complex strides (must emit stride
284 // expression outside of loop!)
285 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
288 SCEVConstant *StrideC = cast<SCEVConstant>(AddRec->getOperand(1));
289 Stride = StrideC->getValue();
291 assert(Stride->getType()->isUnsigned() &&
292 "Constants should be canonicalized to unsigned!");
296 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
297 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
298 /// return true. Otherwise, return false.
299 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
300 std::set<Instruction*> &Processed) {
301 if (I->getType() == Type::VoidTy) return false;
302 if (!Processed.insert(I).second)
303 return true; // Instruction already handled.
305 // Get the symbolic expression for this instruction.
306 SCEVHandle ISE = GetExpressionSCEV(I, L);
307 if (isa<SCEVCouldNotCompute>(ISE)) return false;
309 // Get the start and stride for this expression.
310 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
312 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
313 return false; // Non-reducible symbolic expression, bail out.
315 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
316 Instruction *User = cast<Instruction>(*UI);
318 // Do not infinitely recurse on PHI nodes.
319 if (isa<PHINode>(User) && User->getParent() == L->getHeader())
322 // If this is an instruction defined in a nested loop, or outside this loop,
323 // don't recurse into it.
324 bool AddUserToIVUsers = false;
325 if (LI->getLoopFor(User->getParent()) != L) {
326 DEBUG(std::cerr << "FOUND USER in nested loop: " << *User
327 << " OF SCEV: " << *ISE << "\n");
328 AddUserToIVUsers = true;
329 } else if (!AddUsersIfInteresting(User, L, Processed)) {
330 DEBUG(std::cerr << "FOUND USER: " << *User
331 << " OF SCEV: " << *ISE << "\n");
332 AddUserToIVUsers = true;
335 if (AddUserToIVUsers) {
336 // Okay, we found a user that we cannot reduce. Analyze the instruction
337 // and decide what to do with it.
338 IVUsesByStride[Stride].addUser(Start, User, I);
345 /// BasedUser - For a particular base value, keep information about how we've
346 /// partitioned the expression so far.
348 /// Base - The Base value for the PHI node that needs to be inserted for
349 /// this use. As the use is processed, information gets moved from this
350 /// field to the Imm field (below). BasedUser values are sorted by this
354 /// Inst - The instruction using the induction variable.
357 /// OperandValToReplace - The operand value of Inst to replace with the
359 Value *OperandValToReplace;
361 /// Imm - The immediate value that should be added to the base immediately
362 /// before Inst, because it will be folded into the imm field of the
366 /// EmittedBase - The actual value* to use for the base value of this
367 /// operation. This is null if we should just use zero so far.
370 // isUseOfPostIncrementedValue - True if this should use the
371 // post-incremented version of this IV, not the preincremented version.
372 // This can only be set in special cases, such as the terminating setcc
373 // instruction for a loop.
374 bool isUseOfPostIncrementedValue;
376 BasedUser(IVStrideUse &IVSU)
377 : Base(IVSU.Offset), Inst(IVSU.User),
378 OperandValToReplace(IVSU.OperandValToReplace),
379 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
380 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
382 // Once we rewrite the code to insert the new IVs we want, update the
383 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
385 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
386 SCEVExpander &Rewriter);
388 // Sort by the Base field.
389 bool operator<(const BasedUser &BU) const { return Base < BU.Base; }
395 void BasedUser::dump() const {
396 std::cerr << " Base=" << *Base;
397 std::cerr << " Imm=" << *Imm;
399 std::cerr << " EB=" << *EmittedBase;
401 std::cerr << " Inst: " << *Inst;
404 // Once we rewrite the code to insert the new IVs we want, update the
405 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
407 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
408 SCEVExpander &Rewriter) {
409 if (!isa<PHINode>(Inst)) {
410 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
411 Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, Inst,
412 OperandValToReplace->getType());
414 // Replace the use of the operand Value with the new Phi we just created.
415 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
416 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
420 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
421 // expression into each operand block that uses it.
422 PHINode *PN = cast<PHINode>(Inst);
423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
424 if (PN->getIncomingValue(i) == OperandValToReplace) {
425 // FIXME: this should split any critical edges.
427 // Insert the code into the end of the predecessor block.
428 BasicBlock::iterator InsertPt = PN->getIncomingBlock(i)->getTerminator();
430 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
431 Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, InsertPt,
432 OperandValToReplace->getType());
434 // Replace the use of the operand Value with the new Phi we just created.
435 PN->setIncomingValue(i, NewVal);
439 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
443 /// isTargetConstant - Return true if the following can be referenced by the
444 /// immediate field of a target instruction.
445 static bool isTargetConstant(const SCEVHandle &V) {
447 // FIXME: Look at the target to decide if &GV is a legal constant immediate.
448 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
449 // PPC allows a sign-extended 16-bit immediate field.
450 if ((int64_t)SC->getValue()->getRawValue() > -(1 << 16) &&
451 (int64_t)SC->getValue()->getRawValue() < (1 << 16)-1)
456 return false; // ENABLE this for x86
458 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
459 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
460 if (CE->getOpcode() == Instruction::Cast)
461 if (isa<GlobalValue>(CE->getOperand(0)))
462 // FIXME: should check to see that the dest is uintptr_t!
467 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
468 /// loop varying to the Imm operand.
469 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
471 if (Val->isLoopInvariant(L)) return; // Nothing to do.
473 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
474 std::vector<SCEVHandle> NewOps;
475 NewOps.reserve(SAE->getNumOperands());
477 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
478 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
479 // If this is a loop-variant expression, it must stay in the immediate
480 // field of the expression.
481 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
483 NewOps.push_back(SAE->getOperand(i));
487 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
489 Val = SCEVAddExpr::get(NewOps);
490 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
491 // Try to pull immediates out of the start value of nested addrec's.
492 SCEVHandle Start = SARE->getStart();
493 MoveLoopVariantsToImediateField(Start, Imm, L);
495 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
497 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
499 // Otherwise, all of Val is variant, move the whole thing over.
500 Imm = SCEVAddExpr::get(Imm, Val);
501 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
506 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
507 /// that can fit into the immediate field of instructions in the target.
508 /// Accumulate these immediate values into the Imm value.
509 static void MoveImmediateValues(SCEVHandle &Val, SCEVHandle &Imm,
510 bool isAddress, Loop *L) {
511 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
512 std::vector<SCEVHandle> NewOps;
513 NewOps.reserve(SAE->getNumOperands());
515 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
516 if (isAddress && isTargetConstant(SAE->getOperand(i))) {
517 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
518 } else if (!SAE->getOperand(i)->isLoopInvariant(L)) {
519 // If this is a loop-variant expression, it must stay in the immediate
520 // field of the expression.
521 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
523 NewOps.push_back(SAE->getOperand(i));
527 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
529 Val = SCEVAddExpr::get(NewOps);
531 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
532 // Try to pull immediates out of the start value of nested addrec's.
533 SCEVHandle Start = SARE->getStart();
534 MoveImmediateValues(Start, Imm, isAddress, L);
536 if (Start != SARE->getStart()) {
537 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
539 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
544 // Loop-variant expressions must stay in the immediate field of the
546 if ((isAddress && isTargetConstant(Val)) ||
547 !Val->isLoopInvariant(L)) {
548 Imm = SCEVAddExpr::get(Imm, Val);
549 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
553 // Otherwise, no immediates to move.
556 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
557 /// removing any common subexpressions from it. Anything truly common is
558 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
559 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
561 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
562 unsigned NumUses = Uses.size();
564 // Only one use? Use its base, regardless of what it is!
565 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
566 SCEVHandle Result = Zero;
568 std::swap(Result, Uses[0].Base);
572 // To find common subexpressions, count how many of Uses use each expression.
573 // If any subexpressions are used Uses.size() times, they are common.
574 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
576 for (unsigned i = 0; i != NumUses; ++i)
577 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Uses[i].Base)) {
578 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
579 SubExpressionUseCounts[AE->getOperand(j)]++;
581 // If the base is zero (which is common), return zero now, there are no
583 if (Uses[i].Base == Zero) return Result;
584 SubExpressionUseCounts[Uses[i].Base]++;
587 // Now that we know how many times each is used, build Result.
588 for (std::map<SCEVHandle, unsigned>::iterator I =
589 SubExpressionUseCounts.begin(), E = SubExpressionUseCounts.end();
591 if (I->second == NumUses) { // Found CSE!
592 Result = SCEVAddExpr::get(Result, I->first);
595 // Remove non-cse's from SubExpressionUseCounts.
596 SubExpressionUseCounts.erase(I++);
599 // If we found no CSE's, return now.
600 if (Result == Zero) return Result;
602 // Otherwise, remove all of the CSE's we found from each of the base values.
603 for (unsigned i = 0; i != NumUses; ++i)
604 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Uses[i].Base)) {
605 std::vector<SCEVHandle> NewOps;
607 // Remove all of the values that are now in SubExpressionUseCounts.
608 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
609 if (!SubExpressionUseCounts.count(AE->getOperand(j)))
610 NewOps.push_back(AE->getOperand(j));
611 if (NewOps.size() == 0)
614 Uses[i].Base = SCEVAddExpr::get(NewOps);
616 // If the base is zero (which is common), return zero now, there are no
618 assert(Uses[i].Base == Result);
626 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
627 /// stride of IV. All of the users may have different starting values, and this
628 /// may not be the only stride (we know it is if isOnlyStride is true).
629 void LoopStrengthReduce::StrengthReduceStridedIVUsers(Value *Stride,
630 IVUsersOfOneStride &Uses,
633 // Transform our list of users and offsets to a bit more complex table. In
634 // this new vector, each 'BasedUser' contains 'Base' the base of the
635 // strided accessas well as the old information from Uses. We progressively
636 // move information from the Base field to the Imm field, until we eventually
637 // have the full access expression to rewrite the use.
638 std::vector<BasedUser> UsersToProcess;
639 UsersToProcess.reserve(Uses.Users.size());
640 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
641 UsersToProcess.push_back(Uses.Users[i]);
643 // Move any loop invariant operands from the offset field to the immediate
644 // field of the use, so that we don't try to use something before it is
646 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
647 UsersToProcess.back().Imm, L);
648 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
649 "Base value is not loop invariant!");
652 // We now have a whole bunch of uses of like-strided induction variables, but
653 // they might all have different bases. We want to emit one PHI node for this
654 // stride which we fold as many common expressions (between the IVs) into as
655 // possible. Start by identifying the common expressions in the base values
656 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
657 // "A+B"), emit it to the preheader, then remove the expression from the
658 // UsersToProcess base values.
659 SCEVHandle CommonExprs = RemoveCommonExpressionsFromUseBases(UsersToProcess);
661 // Next, figure out what we can represent in the immediate fields of
662 // instructions. If we can represent anything there, move it to the imm
663 // fields of the BasedUsers. We do this so that it increases the commonality
664 // of the remaining uses.
665 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
666 // Addressing modes can be folded into loads and stores. Be careful that
667 // the store is through the expression, not of the expression though.
668 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
669 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
670 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
673 MoveImmediateValues(UsersToProcess[i].Base, UsersToProcess[i].Imm,
677 // Now that we know what we need to do, insert the PHI node itself.
679 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
680 << *CommonExprs << " :\n");
682 SCEVExpander Rewriter(*SE, *LI);
683 SCEVExpander PreheaderRewriter(*SE, *LI);
685 BasicBlock *Preheader = L->getLoopPreheader();
686 Instruction *PreInsertPt = Preheader->getTerminator();
687 Instruction *PhiInsertBefore = L->getHeader()->begin();
689 assert(isa<PHINode>(PhiInsertBefore) &&
690 "How could this loop have IV's without any phis?");
691 PHINode *SomeLoopPHI = cast<PHINode>(PhiInsertBefore);
692 assert(SomeLoopPHI->getNumIncomingValues() == 2 &&
693 "This loop isn't canonicalized right");
694 BasicBlock *LatchBlock =
695 SomeLoopPHI->getIncomingBlock(SomeLoopPHI->getIncomingBlock(0) == Preheader);
697 // Create a new Phi for this base, and stick it in the loop header.
698 const Type *ReplacedTy = CommonExprs->getType();
699 PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
702 // Emit the initial base value into the loop preheader, and add it to the
704 Value *PHIBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
706 NewPHI->addIncoming(PHIBaseV, Preheader);
708 // Emit the increment of the base value before the terminator of the loop
709 // latch block, and add it to the Phi node.
710 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
711 SCEVUnknown::get(Stride));
713 Value *IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
715 IncV->setName(NewPHI->getName()+".inc");
716 NewPHI->addIncoming(IncV, LatchBlock);
718 // Sort by the base value, so that all IVs with identical bases are next to
720 std::sort(UsersToProcess.begin(), UsersToProcess.end());
721 while (!UsersToProcess.empty()) {
722 SCEVHandle Base = UsersToProcess.front().Base;
724 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
726 // Emit the code for Base into the preheader.
727 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
730 // If BaseV is a constant other than 0, make sure that it gets inserted into
731 // the preheader, instead of being forward substituted into the uses. We do
732 // this by forcing a noop cast to be inserted into the preheader in this
734 if (Constant *C = dyn_cast<Constant>(BaseV))
735 if (!C->isNullValue()) {
736 // We want this constant emitted into the preheader!
737 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
741 // Emit the code to add the immediate offset to the Phi value, just before
742 // the instructions that we identified as using this stride and base.
743 while (!UsersToProcess.empty() && UsersToProcess.front().Base == Base) {
744 BasedUser &User = UsersToProcess.front();
746 // If this instruction wants to use the post-incremented value, move it
747 // after the post-inc and use its value instead of the PHI.
748 Value *RewriteOp = NewPHI;
749 if (User.isUseOfPostIncrementedValue) {
751 User.Inst->moveBefore(LatchBlock->getTerminator());
753 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
755 // Clear the SCEVExpander's expression map so that we are guaranteed
756 // to have the code emitted where we expect it.
759 // Now that we know what we need to do, insert code before User for the
760 // immediate and any loop-variant expressions.
761 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
762 // Add BaseV to the PHI value if needed.
763 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
765 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter);
767 // Mark old value we replaced as possibly dead, so that it is elminated
768 // if we just replaced the last use of that value.
769 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
771 UsersToProcess.erase(UsersToProcess.begin());
774 // TODO: Next, find out which base index is the most common, pull it out.
777 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
778 // different starting values, into different PHIs.
781 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
782 // uses in the loop, look to see if we can eliminate some, in favor of using
783 // common indvars for the different uses.
784 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
785 // TODO: implement optzns here.
790 // Finally, get the terminating condition for the loop if possible. If we
791 // can, we want to change it to use a post-incremented version of its
792 // induction variable, to allow coallescing the live ranges for the IV into
793 // one register value.
794 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
795 BasicBlock *Preheader = L->getLoopPreheader();
796 BasicBlock *LatchBlock =
797 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
798 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
799 if (!TermBr || TermBr->isUnconditional() ||
800 !isa<SetCondInst>(TermBr->getCondition()))
802 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
804 // Search IVUsesByStride to find Cond's IVUse if there is one.
805 IVStrideUse *CondUse = 0;
806 Value *CondStride = 0;
808 for (std::map<Value*, IVUsersOfOneStride>::iterator I =IVUsesByStride.begin(),
809 E = IVUsesByStride.end(); I != E && !CondUse; ++I)
810 for (std::vector<IVStrideUse>::iterator UI = I->second.Users.begin(),
811 E = I->second.Users.end(); UI != E; ++UI)
812 if (UI->User == Cond) {
814 CondStride = I->first;
815 // NOTE: we could handle setcc instructions with multiple uses here, but
816 // InstCombine does it as well for simple uses, it's not clear that it
817 // occurs enough in real life to handle.
820 if (!CondUse) return; // setcc doesn't use the IV.
822 // setcc stride is complex, don't mess with users.
823 if (!isa<ConstantInt>(CondStride)) return;
825 // It's possible for the setcc instruction to be anywhere in the loop, and
826 // possible for it to have multiple users. If it is not immediately before
827 // the latch block branch, move it.
828 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
829 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
830 Cond->moveBefore(TermBr);
832 // Otherwise, clone the terminating condition and insert into the loopend.
833 Cond = cast<SetCondInst>(Cond->clone());
834 Cond->setName(L->getHeader()->getName() + ".termcond");
835 LatchBlock->getInstList().insert(TermBr, Cond);
837 // Clone the IVUse, as the old use still exists!
838 IVUsesByStride[CondStride].addUser(CondUse->Offset, Cond,
839 CondUse->OperandValToReplace);
840 CondUse = &IVUsesByStride[CondStride].Users.back();
844 // If we get to here, we know that we can transform the setcc instruction to
845 // use the post-incremented version of the IV, allowing us to coallesce the
846 // live ranges for the IV correctly.
847 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset,
848 SCEVUnknown::get(CondStride));
849 CondUse->isUseOfPostIncrementedValue = true;
852 void LoopStrengthReduce::runOnLoop(Loop *L) {
853 // First step, transform all loops nesting inside of this loop.
854 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
857 // Next, find all uses of induction variables in this loop, and catagorize
858 // them by stride. Start by finding all of the PHI nodes in the header for
859 // this loop. If they are induction variables, inspect their uses.
860 std::set<Instruction*> Processed; // Don't reprocess instructions.
861 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
862 AddUsersIfInteresting(I, L, Processed);
864 // If we have nothing to do, return.
865 if (IVUsesByStride.empty()) return;
867 // Optimize induction variables. Some indvar uses can be transformed to use
868 // strides that will be needed for other purposes. A common example of this
869 // is the exit test for the loop, which can often be rewritten to use the
870 // computation of some other indvar to decide when to terminate the loop.
874 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
875 // doing computation in byte values, promote to 32-bit values if safe.
877 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
878 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
879 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
880 // to be careful that IV's are all the same type. Only works for intptr_t
883 // If we only have one stride, we can more aggressively eliminate some things.
884 bool HasOneStride = IVUsesByStride.size() == 1;
886 // Note: this processes each stride/type pair individually. All users passed
887 // into StrengthReduceStridedIVUsers have the same type AND stride.
888 for (std::map<Value*, IVUsersOfOneStride>::iterator SI
889 = IVUsesByStride.begin(), E = IVUsesByStride.end(); SI != E; ++SI)
890 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
892 // Clean up after ourselves
893 if (!DeadInsts.empty()) {
894 DeleteTriviallyDeadInstructions(DeadInsts);
896 BasicBlock::iterator I = L->getHeader()->begin();
898 while ((PN = dyn_cast<PHINode>(I))) {
899 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
901 // At this point, we know that we have killed one or more GEP instructions.
902 // It is worth checking to see if the cann indvar is also dead, so that we
903 // can remove it as well. The requirements for the cann indvar to be
904 // considered dead are:
905 // 1. the cann indvar has one use
906 // 2. the use is an add instruction
907 // 3. the add has one use
908 // 4. the add is used by the cann indvar
909 // If all four cases above are true, then we can remove both the add and
911 // FIXME: this needs to eliminate an induction variable even if it's being
912 // compared against some value to decide loop termination.
913 if (PN->hasOneUse()) {
914 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
915 if (BO && BO->hasOneUse()) {
916 if (PN == *(BO->use_begin())) {
917 DeadInsts.insert(BO);
918 // Break the cycle, then delete the PHI.
919 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
920 SE->deleteInstructionFromRecords(PN);
921 PN->eraseFromParent();
926 DeleteTriviallyDeadInstructions(DeadInsts);
929 CastedPointers.clear();
930 IVUsesByStride.clear();