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.
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 "Strength Reduce GEP Uses of Ind. Vars");
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) && User->getParent() == L->getHeader())
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 nested 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.
355 IVUsesByStride[Stride].addUser(Start, User, I);
362 /// BasedUser - For a particular base value, keep information about how we've
363 /// partitioned the expression so far.
365 /// Base - The Base value for the PHI node that needs to be inserted for
366 /// this use. As the use is processed, information gets moved from this
367 /// field to the Imm field (below). BasedUser values are sorted by this
371 /// Inst - The instruction using the induction variable.
374 /// OperandValToReplace - The operand value of Inst to replace with the
376 Value *OperandValToReplace;
378 /// Imm - The immediate value that should be added to the base immediately
379 /// before Inst, because it will be folded into the imm field of the
383 /// EmittedBase - The actual value* to use for the base value of this
384 /// operation. This is null if we should just use zero so far.
387 // isUseOfPostIncrementedValue - True if this should use the
388 // post-incremented version of this IV, not the preincremented version.
389 // This can only be set in special cases, such as the terminating setcc
390 // instruction for a loop.
391 bool isUseOfPostIncrementedValue;
393 BasedUser(IVStrideUse &IVSU)
394 : Base(IVSU.Offset), Inst(IVSU.User),
395 OperandValToReplace(IVSU.OperandValToReplace),
396 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
397 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
399 // Once we rewrite the code to insert the new IVs we want, update the
400 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
402 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
403 SCEVExpander &Rewriter, Loop *L,
406 // Sort by the Base field.
407 bool operator<(const BasedUser &BU) const { return Base < BU.Base; }
413 void BasedUser::dump() const {
414 std::cerr << " Base=" << *Base;
415 std::cerr << " Imm=" << *Imm;
417 std::cerr << " EB=" << *EmittedBase;
419 std::cerr << " Inst: " << *Inst;
422 // Once we rewrite the code to insert the new IVs we want, update the
423 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
425 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
426 SCEVExpander &Rewriter,
428 if (!isa<PHINode>(Inst)) {
429 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
430 Value *NewVal = Rewriter.expandCodeFor(NewValSCEV, Inst,
431 OperandValToReplace->getType());
432 // Replace the use of the operand Value with the new Phi we just created.
433 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
434 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
438 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
439 // expression into each operand block that uses it. Note that PHI nodes can
440 // have multiple entries for the same predecessor. We use a map to make sure
441 // that a PHI node only has a single Value* for each predecessor (which also
442 // prevents us from inserting duplicate code in some blocks).
443 std::map<BasicBlock*, Value*> InsertedCode;
444 PHINode *PN = cast<PHINode>(Inst);
445 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
446 if (PN->getIncomingValue(i) == OperandValToReplace) {
447 // If this is a critical edge, split the edge so that we do not insert the
448 // code on all predecessor/successor paths.
450 PN->getIncomingBlock(i)->getTerminator()->getNumSuccessors() > 1) {
452 // First step, split the critical edge.
453 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P);
455 // Next step: move the basic block. In particular, if the PHI node
456 // is outside of the loop, and PredTI is in the loop, we want to
457 // move the block to be immediately before the PHI block, not
458 // immediately after PredTI.
459 if (L->contains(PN->getIncomingBlock(i)) &&
460 !L->contains(PN->getParent())) {
461 BasicBlock *NewBB = PN->getIncomingBlock(i);
462 NewBB->moveBefore(PN->getParent());
467 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
469 // Insert the code into the end of the predecessor block.
470 BasicBlock::iterator InsertPt =PN->getIncomingBlock(i)->getTerminator();
472 SCEVHandle NewValSCEV = SCEVAddExpr::get(NewBase, Imm);
473 Code = Rewriter.expandCodeFor(NewValSCEV, InsertPt,
474 OperandValToReplace->getType());
477 // Replace the use of the operand Value with the new Phi we just created.
478 PN->setIncomingValue(i, Code);
482 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
486 /// isTargetConstant - Return true if the following can be referenced by the
487 /// immediate field of a target instruction.
488 static bool isTargetConstant(const SCEVHandle &V) {
490 // FIXME: Look at the target to decide if &GV is a legal constant immediate.
491 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
492 // PPC allows a sign-extended 16-bit immediate field.
493 if ((int64_t)SC->getValue()->getRawValue() > -(1 << 16) &&
494 (int64_t)SC->getValue()->getRawValue() < (1 << 16)-1)
499 return false; // ENABLE this for x86
501 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
502 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
503 if (CE->getOpcode() == Instruction::Cast)
504 if (isa<GlobalValue>(CE->getOperand(0)))
505 // FIXME: should check to see that the dest is uintptr_t!
510 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
511 /// loop varying to the Imm operand.
512 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
514 if (Val->isLoopInvariant(L)) return; // Nothing to do.
516 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
517 std::vector<SCEVHandle> NewOps;
518 NewOps.reserve(SAE->getNumOperands());
520 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
521 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
522 // If this is a loop-variant expression, it must stay in the immediate
523 // field of the expression.
524 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
526 NewOps.push_back(SAE->getOperand(i));
530 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
532 Val = SCEVAddExpr::get(NewOps);
533 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
534 // Try to pull immediates out of the start value of nested addrec's.
535 SCEVHandle Start = SARE->getStart();
536 MoveLoopVariantsToImediateField(Start, Imm, L);
538 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
540 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
542 // Otherwise, all of Val is variant, move the whole thing over.
543 Imm = SCEVAddExpr::get(Imm, Val);
544 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
549 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
550 /// that can fit into the immediate field of instructions in the target.
551 /// Accumulate these immediate values into the Imm value.
552 static void MoveImmediateValues(SCEVHandle &Val, SCEVHandle &Imm,
553 bool isAddress, Loop *L) {
554 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
555 std::vector<SCEVHandle> NewOps;
556 NewOps.reserve(SAE->getNumOperands());
558 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
559 if (isAddress && isTargetConstant(SAE->getOperand(i))) {
560 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
561 } else if (!SAE->getOperand(i)->isLoopInvariant(L)) {
562 // If this is a loop-variant expression, it must stay in the immediate
563 // field of the expression.
564 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
566 NewOps.push_back(SAE->getOperand(i));
570 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
572 Val = SCEVAddExpr::get(NewOps);
574 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
575 // Try to pull immediates out of the start value of nested addrec's.
576 SCEVHandle Start = SARE->getStart();
577 MoveImmediateValues(Start, Imm, isAddress, L);
579 if (Start != SARE->getStart()) {
580 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
582 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
587 // Loop-variant expressions must stay in the immediate field of the
589 if ((isAddress && isTargetConstant(Val)) ||
590 !Val->isLoopInvariant(L)) {
591 Imm = SCEVAddExpr::get(Imm, Val);
592 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
596 // Otherwise, no immediates to move.
600 /// IncrementAddExprUses - Decompose the specified expression into its added
601 /// subexpressions, and increment SubExpressionUseCounts for each of these
602 /// decomposed parts.
603 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
605 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
606 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
607 SeparateSubExprs(SubExprs, AE->getOperand(j));
608 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
609 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
610 if (SARE->getOperand(0) == Zero) {
611 SubExprs.push_back(Expr);
613 // Compute the addrec with zero as its base.
614 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
615 Ops[0] = Zero; // Start with zero base.
616 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
619 SeparateSubExprs(SubExprs, SARE->getOperand(0));
621 } else if (!isa<SCEVConstant>(Expr) ||
622 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
624 SubExprs.push_back(Expr);
629 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
630 /// removing any common subexpressions from it. Anything truly common is
631 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
632 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
634 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
635 unsigned NumUses = Uses.size();
637 // Only one use? Use its base, regardless of what it is!
638 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
639 SCEVHandle Result = Zero;
641 std::swap(Result, Uses[0].Base);
645 // To find common subexpressions, count how many of Uses use each expression.
646 // If any subexpressions are used Uses.size() times, they are common.
647 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
649 std::vector<SCEVHandle> SubExprs;
650 for (unsigned i = 0; i != NumUses; ++i) {
651 // If the base is zero (which is common), return zero now, there are no
653 if (Uses[i].Base == Zero) return Zero;
655 // Split the expression into subexprs.
656 SeparateSubExprs(SubExprs, Uses[i].Base);
657 // Add one to SubExpressionUseCounts for each subexpr present.
658 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
659 SubExpressionUseCounts[SubExprs[j]]++;
664 // Now that we know how many times each is used, build Result.
665 for (std::map<SCEVHandle, unsigned>::iterator I =
666 SubExpressionUseCounts.begin(), E = SubExpressionUseCounts.end();
668 if (I->second == NumUses) { // Found CSE!
669 Result = SCEVAddExpr::get(Result, I->first);
672 // Remove non-cse's from SubExpressionUseCounts.
673 SubExpressionUseCounts.erase(I++);
676 // If we found no CSE's, return now.
677 if (Result == Zero) return Result;
679 // Otherwise, remove all of the CSE's we found from each of the base values.
680 for (unsigned i = 0; i != NumUses; ++i) {
681 // Split the expression into subexprs.
682 SeparateSubExprs(SubExprs, Uses[i].Base);
684 // Remove any common subexpressions.
685 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
686 if (SubExpressionUseCounts.count(SubExprs[j])) {
687 SubExprs.erase(SubExprs.begin()+j);
691 // Finally, the non-shared expressions together.
692 if (SubExprs.empty())
695 Uses[i].Base = SCEVAddExpr::get(SubExprs);
703 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
704 /// stride of IV. All of the users may have different starting values, and this
705 /// may not be the only stride (we know it is if isOnlyStride is true).
706 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
707 IVUsersOfOneStride &Uses,
710 // Transform our list of users and offsets to a bit more complex table. In
711 // this new vector, each 'BasedUser' contains 'Base' the base of the
712 // strided accessas well as the old information from Uses. We progressively
713 // move information from the Base field to the Imm field, until we eventually
714 // have the full access expression to rewrite the use.
715 std::vector<BasedUser> UsersToProcess;
716 UsersToProcess.reserve(Uses.Users.size());
717 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
718 UsersToProcess.push_back(Uses.Users[i]);
720 // Move any loop invariant operands from the offset field to the immediate
721 // field of the use, so that we don't try to use something before it is
723 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
724 UsersToProcess.back().Imm, L);
725 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
726 "Base value is not loop invariant!");
729 // We now have a whole bunch of uses of like-strided induction variables, but
730 // they might all have different bases. We want to emit one PHI node for this
731 // stride which we fold as many common expressions (between the IVs) into as
732 // possible. Start by identifying the common expressions in the base values
733 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
734 // "A+B"), emit it to the preheader, then remove the expression from the
735 // UsersToProcess base values.
736 SCEVHandle CommonExprs = RemoveCommonExpressionsFromUseBases(UsersToProcess);
738 // Next, figure out what we can represent in the immediate fields of
739 // instructions. If we can represent anything there, move it to the imm
740 // fields of the BasedUsers. We do this so that it increases the commonality
741 // of the remaining uses.
742 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
743 // If the user is not in the current loop, this means it is using the exit
744 // value of the IV. Do not put anything in the base, make sure it's all in
745 // the immediate field to allow as much factoring as possible.
746 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
747 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
748 UsersToProcess[i].Base);
749 UsersToProcess[i].Base =
750 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
753 // Addressing modes can be folded into loads and stores. Be careful that
754 // the store is through the expression, not of the expression though.
755 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
756 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
757 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
760 MoveImmediateValues(UsersToProcess[i].Base, UsersToProcess[i].Imm,
765 // Now that we know what we need to do, insert the PHI node itself.
767 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
768 << *CommonExprs << " :\n");
770 SCEVExpander Rewriter(*SE, *LI);
771 SCEVExpander PreheaderRewriter(*SE, *LI);
773 BasicBlock *Preheader = L->getLoopPreheader();
774 Instruction *PreInsertPt = Preheader->getTerminator();
775 Instruction *PhiInsertBefore = L->getHeader()->begin();
777 assert(isa<PHINode>(PhiInsertBefore) &&
778 "How could this loop have IV's without any phis?");
779 PHINode *SomeLoopPHI = cast<PHINode>(PhiInsertBefore);
780 assert(SomeLoopPHI->getNumIncomingValues() == 2 &&
781 "This loop isn't canonicalized right");
782 BasicBlock *LatchBlock =
783 SomeLoopPHI->getIncomingBlock(SomeLoopPHI->getIncomingBlock(0) == Preheader);
785 // Create a new Phi for this base, and stick it in the loop header.
786 const Type *ReplacedTy = CommonExprs->getType();
787 PHINode *NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
790 // Insert the stride into the preheader.
791 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
793 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
796 // Emit the initial base value into the loop preheader, and add it to the
798 Value *PHIBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
800 NewPHI->addIncoming(PHIBaseV, Preheader);
802 // Emit the increment of the base value before the terminator of the loop
803 // latch block, and add it to the Phi node.
804 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
805 SCEVUnknown::get(StrideV));
807 Value *IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
809 IncV->setName(NewPHI->getName()+".inc");
810 NewPHI->addIncoming(IncV, LatchBlock);
812 // Sort by the base value, so that all IVs with identical bases are next to
814 std::sort(UsersToProcess.begin(), UsersToProcess.end());
815 while (!UsersToProcess.empty()) {
816 SCEVHandle Base = UsersToProcess.front().Base;
818 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
820 // Emit the code for Base into the preheader.
821 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
824 // If BaseV is a constant other than 0, make sure that it gets inserted into
825 // the preheader, instead of being forward substituted into the uses. We do
826 // this by forcing a noop cast to be inserted into the preheader in this
828 if (Constant *C = dyn_cast<Constant>(BaseV))
829 if (!C->isNullValue() && !isTargetConstant(Base)) {
830 // We want this constant emitted into the preheader!
831 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
835 // Emit the code to add the immediate offset to the Phi value, just before
836 // the instructions that we identified as using this stride and base.
837 while (!UsersToProcess.empty() && UsersToProcess.front().Base == Base) {
838 BasedUser &User = UsersToProcess.front();
840 // If this instruction wants to use the post-incremented value, move it
841 // after the post-inc and use its value instead of the PHI.
842 Value *RewriteOp = NewPHI;
843 if (User.isUseOfPostIncrementedValue) {
845 User.Inst->moveBefore(LatchBlock->getTerminator());
847 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
849 // Clear the SCEVExpander's expression map so that we are guaranteed
850 // to have the code emitted where we expect it.
853 // Now that we know what we need to do, insert code before User for the
854 // immediate and any loop-variant expressions.
855 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
856 // Add BaseV to the PHI value if needed.
857 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
859 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
861 // Mark old value we replaced as possibly dead, so that it is elminated
862 // if we just replaced the last use of that value.
863 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
865 UsersToProcess.erase(UsersToProcess.begin());
868 // TODO: Next, find out which base index is the most common, pull it out.
871 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
872 // different starting values, into different PHIs.
875 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
876 // uses in the loop, look to see if we can eliminate some, in favor of using
877 // common indvars for the different uses.
878 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
879 // TODO: implement optzns here.
884 // Finally, get the terminating condition for the loop if possible. If we
885 // can, we want to change it to use a post-incremented version of its
886 // induction variable, to allow coallescing the live ranges for the IV into
887 // one register value.
888 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
889 BasicBlock *Preheader = L->getLoopPreheader();
890 BasicBlock *LatchBlock =
891 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
892 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
893 if (!TermBr || TermBr->isUnconditional() ||
894 !isa<SetCondInst>(TermBr->getCondition()))
896 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
898 // Search IVUsesByStride to find Cond's IVUse if there is one.
899 IVStrideUse *CondUse = 0;
900 const SCEVHandle *CondStride = 0;
902 for (std::map<SCEVHandle, IVUsersOfOneStride>::iterator
903 I = IVUsesByStride.begin(), E = IVUsesByStride.end();
904 I != E && !CondUse; ++I)
905 for (std::vector<IVStrideUse>::iterator UI = I->second.Users.begin(),
906 E = I->second.Users.end(); UI != E; ++UI)
907 if (UI->User == Cond) {
909 CondStride = &I->first;
910 // NOTE: we could handle setcc instructions with multiple uses here, but
911 // InstCombine does it as well for simple uses, it's not clear that it
912 // occurs enough in real life to handle.
915 if (!CondUse) return; // setcc doesn't use the IV.
917 // setcc stride is complex, don't mess with users.
918 // FIXME: Evaluate whether this is a good idea or not.
919 if (!isa<SCEVConstant>(*CondStride)) return;
921 // It's possible for the setcc instruction to be anywhere in the loop, and
922 // possible for it to have multiple users. If it is not immediately before
923 // the latch block branch, move it.
924 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
925 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
926 Cond->moveBefore(TermBr);
928 // Otherwise, clone the terminating condition and insert into the loopend.
929 Cond = cast<SetCondInst>(Cond->clone());
930 Cond->setName(L->getHeader()->getName() + ".termcond");
931 LatchBlock->getInstList().insert(TermBr, Cond);
933 // Clone the IVUse, as the old use still exists!
934 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
935 CondUse->OperandValToReplace);
936 CondUse = &IVUsesByStride[*CondStride].Users.back();
940 // If we get to here, we know that we can transform the setcc instruction to
941 // use the post-incremented version of the IV, allowing us to coallesce the
942 // live ranges for the IV correctly.
943 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
944 CondUse->isUseOfPostIncrementedValue = true;
947 void LoopStrengthReduce::runOnLoop(Loop *L) {
948 // First step, transform all loops nesting inside of this loop.
949 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
952 // Next, find all uses of induction variables in this loop, and catagorize
953 // them by stride. Start by finding all of the PHI nodes in the header for
954 // this loop. If they are induction variables, inspect their uses.
955 std::set<Instruction*> Processed; // Don't reprocess instructions.
956 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
957 AddUsersIfInteresting(I, L, Processed);
959 // If we have nothing to do, return.
960 if (IVUsesByStride.empty()) return;
962 // Optimize induction variables. Some indvar uses can be transformed to use
963 // strides that will be needed for other purposes. A common example of this
964 // is the exit test for the loop, which can often be rewritten to use the
965 // computation of some other indvar to decide when to terminate the loop.
969 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
970 // doing computation in byte values, promote to 32-bit values if safe.
972 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
973 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
974 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
975 // to be careful that IV's are all the same type. Only works for intptr_t
978 // If we only have one stride, we can more aggressively eliminate some things.
979 bool HasOneStride = IVUsesByStride.size() == 1;
981 // Note: this processes each stride/type pair individually. All users passed
982 // into StrengthReduceStridedIVUsers have the same type AND stride.
983 for (std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI
984 = IVUsesByStride.begin(), E = IVUsesByStride.end(); SI != E; ++SI)
985 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
987 // Clean up after ourselves
988 if (!DeadInsts.empty()) {
989 DeleteTriviallyDeadInstructions(DeadInsts);
991 BasicBlock::iterator I = L->getHeader()->begin();
993 while ((PN = dyn_cast<PHINode>(I))) {
994 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
996 // At this point, we know that we have killed one or more GEP
997 // instructions. It is worth checking to see if the cann indvar is also
998 // dead, so that we can remove it as well. The requirements for the cann
999 // indvar to be considered dead are:
1000 // 1. the cann indvar has one use
1001 // 2. the use is an add instruction
1002 // 3. the add has one use
1003 // 4. the add is used by the cann indvar
1004 // If all four cases above are true, then we can remove both the add and
1006 // FIXME: this needs to eliminate an induction variable even if it's being
1007 // compared against some value to decide loop termination.
1008 if (PN->hasOneUse()) {
1009 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1010 if (BO && BO->hasOneUse()) {
1011 if (PN == *(BO->use_begin())) {
1012 DeadInsts.insert(BO);
1013 // Break the cycle, then delete the PHI.
1014 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1015 SE->deleteInstructionFromRecords(PN);
1016 PN->eraseFromParent();
1021 DeleteTriviallyDeadInstructions(DeadInsts);
1024 CastedPointers.clear();
1025 IVUsesByStride.clear();