1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Nate Begeman and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs a strength reduction on array references inside loops that
11 // have as one or more of their components the loop induction variable. This is
12 // accomplished by creating a new Value to hold the initial value of the array
13 // access for the first iteration, and then creating a new GEP instruction in
14 // the loop to increment the value by the appropriate amount.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "loop-reduce"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/Constants.h"
21 #include "llvm/Instructions.h"
22 #include "llvm/Type.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Analysis/Dominators.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/ScalarEvolutionExpander.h"
27 #include "llvm/Support/CFG.h"
28 #include "llvm/Support/GetElementPtrTypeIterator.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Target/TargetData.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/Compiler.h"
35 #include "llvm/Target/TargetLowering.h"
42 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced");
43 Statistic<> NumInserted("loop-reduce", "Number of PHIs inserted");
44 Statistic<> NumVariable("loop-reduce","Number of PHIs with variable strides");
46 /// IVStrideUse - Keep track of one use of a strided induction variable, where
47 /// the stride is stored externally. The Offset member keeps track of the
48 /// offset from the IV, User is the actual user of the operand, and 'Operand'
49 /// is the operand # of the User that is the use.
53 Value *OperandValToReplace;
55 // isUseOfPostIncrementedValue - True if this should use the
56 // post-incremented version of this IV, not the preincremented version.
57 // This can only be set in special cases, such as the terminating setcc
58 // instruction for a loop or uses dominated by the loop.
59 bool isUseOfPostIncrementedValue;
61 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
62 : Offset(Offs), User(U), OperandValToReplace(O),
63 isUseOfPostIncrementedValue(false) {}
66 /// IVUsersOfOneStride - This structure keeps track of all instructions that
67 /// have an operand that is based on the trip count multiplied by some stride.
68 /// The stride for all of these users is common and kept external to this
70 struct IVUsersOfOneStride {
71 /// Users - Keep track of all of the users of this stride as well as the
72 /// initial value and the operand that uses the IV.
73 std::vector<IVStrideUse> Users;
75 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
76 Users.push_back(IVStrideUse(Offset, User, Operand));
80 /// IVInfo - This structure keeps track of one IV expression inserted during
81 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
82 /// well as the PHI node and increment value created for rewrite.
90 : Stride(SCEVUnknown::getIntegerSCEV(0, Type::UIntTy)),
91 Base (SCEVUnknown::getIntegerSCEV(0, Type::UIntTy)) {}
92 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
94 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
97 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
98 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
99 struct IVsOfOneStride {
100 std::vector<IVExpr> IVs;
102 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
104 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
108 class VISIBILITY_HIDDEN LoopStrengthReduce : public FunctionPass {
112 const TargetData *TD;
113 const Type *UIntPtrTy;
116 /// IVUsesByStride - Keep track of all uses of induction variables that we
117 /// are interested in. The key of the map is the stride of the access.
118 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
120 /// IVsByStride - Keep track of all IVs that have been inserted for a
121 /// particular stride.
122 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
124 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
125 /// We use this to iterate over the IVUsesByStride collection without being
126 /// dependent on random ordering of pointers in the process.
127 std::vector<SCEVHandle> StrideOrder;
129 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
130 /// of the casted version of each value. This is accessed by
131 /// getCastedVersionOf.
132 std::map<Value*, Value*> CastedPointers;
134 /// DeadInsts - Keep track of instructions we may have made dead, so that
135 /// we can remove them after we are done working.
136 std::set<Instruction*> DeadInsts;
138 /// TLI - Keep a pointer of a TargetLowering to consult for determining
139 /// transformation profitability.
140 const TargetLowering *TLI;
143 LoopStrengthReduce(const TargetLowering *tli = NULL)
147 virtual bool runOnFunction(Function &) {
148 LI = &getAnalysis<LoopInfo>();
149 EF = &getAnalysis<ETForest>();
150 SE = &getAnalysis<ScalarEvolution>();
151 TD = &getAnalysis<TargetData>();
152 UIntPtrTy = TD->getIntPtrType();
155 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
161 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
162 // We split critical edges, so we change the CFG. However, we do update
163 // many analyses if they are around.
164 AU.addPreservedID(LoopSimplifyID);
165 AU.addPreserved<LoopInfo>();
166 AU.addPreserved<DominatorSet>();
167 AU.addPreserved<ETForest>();
168 AU.addPreserved<ImmediateDominators>();
169 AU.addPreserved<DominanceFrontier>();
170 AU.addPreserved<DominatorTree>();
172 AU.addRequiredID(LoopSimplifyID);
173 AU.addRequired<LoopInfo>();
174 AU.addRequired<ETForest>();
175 AU.addRequired<TargetData>();
176 AU.addRequired<ScalarEvolution>();
179 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
181 Value *getCastedVersionOf(Value *V);
183 void runOnLoop(Loop *L);
184 bool AddUsersIfInteresting(Instruction *I, Loop *L,
185 std::set<Instruction*> &Processed);
186 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
188 void OptimizeIndvars(Loop *L);
190 unsigned CheckForIVReuse(const SCEVHandle&, IVExpr&, const Type*);
192 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
193 IVUsersOfOneStride &Uses,
194 Loop *L, bool isOnlyStride);
195 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
197 RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
200 FunctionPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
201 return new LoopStrengthReduce(TLI);
204 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
206 Value *LoopStrengthReduce::getCastedVersionOf(Value *V) {
207 if (V->getType() == UIntPtrTy) return V;
208 if (Constant *CB = dyn_cast<Constant>(V))
209 return ConstantExpr::getCast(CB, UIntPtrTy);
211 Value *&New = CastedPointers[V];
214 New = SCEVExpander::InsertCastOfTo(V, UIntPtrTy);
215 DeadInsts.insert(cast<Instruction>(New));
220 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
221 /// specified set are trivially dead, delete them and see if this makes any of
222 /// their operands subsequently dead.
223 void LoopStrengthReduce::
224 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
225 while (!Insts.empty()) {
226 Instruction *I = *Insts.begin();
227 Insts.erase(Insts.begin());
228 if (isInstructionTriviallyDead(I)) {
229 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
230 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
232 SE->deleteInstructionFromRecords(I);
233 I->eraseFromParent();
240 /// GetExpressionSCEV - Compute and return the SCEV for the specified
242 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
243 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
244 // If this is a GEP that SE doesn't know about, compute it now and insert it.
245 // If this is not a GEP, or if we have already done this computation, just let
247 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
248 if (!GEP || SE->hasSCEV(GEP))
249 return SE->getSCEV(Exp);
251 // Analyze all of the subscripts of this getelementptr instruction, looking
252 // for uses that are determined by the trip count of L. First, skip all
253 // operands the are not dependent on the IV.
255 // Build up the base expression. Insert an LLVM cast of the pointer to
257 SCEVHandle GEPVal = SCEVUnknown::get(getCastedVersionOf(GEP->getOperand(0)));
259 gep_type_iterator GTI = gep_type_begin(GEP);
261 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
262 // If this is a use of a recurrence that we can analyze, and it comes before
263 // Op does in the GEP operand list, we will handle this when we process this
265 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
266 const StructLayout *SL = TD->getStructLayout(STy);
267 unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
268 uint64_t Offset = SL->MemberOffsets[Idx];
269 GEPVal = SCEVAddExpr::get(GEPVal,
270 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
272 Value *OpVal = getCastedVersionOf(GEP->getOperand(i));
273 SCEVHandle Idx = SE->getSCEV(OpVal);
275 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
277 Idx = SCEVMulExpr::get(Idx,
278 SCEVConstant::get(ConstantInt::get(UIntPtrTy,
280 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
284 SE->setSCEV(GEP, GEPVal);
288 /// getSCEVStartAndStride - Compute the start and stride of this expression,
289 /// returning false if the expression is not a start/stride pair, or true if it
290 /// is. The stride must be a loop invariant expression, but the start may be
291 /// a mix of loop invariant and loop variant expressions.
292 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
293 SCEVHandle &Start, SCEVHandle &Stride) {
294 SCEVHandle TheAddRec = Start; // Initialize to zero.
296 // If the outer level is an AddExpr, the operands are all start values except
297 // for a nested AddRecExpr.
298 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
299 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
300 if (SCEVAddRecExpr *AddRec =
301 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
302 if (AddRec->getLoop() == L)
303 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
305 return false; // Nested IV of some sort?
307 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
310 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH)) {
313 return false; // not analyzable.
316 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
317 if (!AddRec || AddRec->getLoop() != L) return false;
319 // FIXME: Generalize to non-affine IV's.
320 if (!AddRec->isAffine()) return false;
322 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
324 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
325 DEBUG(std::cerr << "[" << L->getHeader()->getName()
326 << "] Variable stride: " << *AddRec << "\n");
328 Stride = AddRec->getOperand(1);
329 // Check that all constant strides are the unsigned type, we don't want to
330 // have two IV's one of signed stride 4 and one of unsigned stride 4 to not be
332 assert((!isa<SCEVConstant>(Stride) || Stride->getType()->isUnsigned()) &&
333 "Constants should be canonicalized to unsigned!");
338 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
339 /// and now we need to decide whether the user should use the preinc or post-inc
340 /// value. If this user should use the post-inc version of the IV, return true.
342 /// Choosing wrong here can break dominance properties (if we choose to use the
343 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
344 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
345 /// should use the post-inc value).
346 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
347 Loop *L, ETForest *EF, Pass *P) {
348 // If the user is in the loop, use the preinc value.
349 if (L->contains(User->getParent())) return false;
351 BasicBlock *LatchBlock = L->getLoopLatch();
353 // Ok, the user is outside of the loop. If it is dominated by the latch
354 // block, use the post-inc value.
355 if (EF->dominates(LatchBlock, User->getParent()))
358 // There is one case we have to be careful of: PHI nodes. These little guys
359 // can live in blocks that do not dominate the latch block, but (since their
360 // uses occur in the predecessor block, not the block the PHI lives in) should
361 // still use the post-inc value. Check for this case now.
362 PHINode *PN = dyn_cast<PHINode>(User);
363 if (!PN) return false; // not a phi, not dominated by latch block.
365 // Look at all of the uses of IV by the PHI node. If any use corresponds to
366 // a block that is not dominated by the latch block, give up and use the
367 // preincremented value.
368 unsigned NumUses = 0;
369 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
370 if (PN->getIncomingValue(i) == IV) {
372 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
376 // Okay, all uses of IV by PN are in predecessor blocks that really are
377 // dominated by the latch block. Split the critical edges and use the
378 // post-incremented value.
379 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
380 if (PN->getIncomingValue(i) == IV) {
381 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P);
382 if (--NumUses == 0) break;
390 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
391 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
392 /// return true. Otherwise, return false.
393 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
394 std::set<Instruction*> &Processed) {
395 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
396 return false; // Void and FP expressions cannot be reduced.
397 if (!Processed.insert(I).second)
398 return true; // Instruction already handled.
400 // Get the symbolic expression for this instruction.
401 SCEVHandle ISE = GetExpressionSCEV(I, L);
402 if (isa<SCEVCouldNotCompute>(ISE)) return false;
404 // Get the start and stride for this expression.
405 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
406 SCEVHandle Stride = Start;
407 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
408 return false; // Non-reducible symbolic expression, bail out.
410 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){
411 Instruction *User = cast<Instruction>(*UI);
413 // Do not infinitely recurse on PHI nodes.
414 if (isa<PHINode>(User) && Processed.count(User))
417 // If this is an instruction defined in a nested loop, or outside this loop,
418 // don't recurse into it.
419 bool AddUserToIVUsers = false;
420 if (LI->getLoopFor(User->getParent()) != L) {
421 DEBUG(std::cerr << "FOUND USER in other loop: " << *User
422 << " OF SCEV: " << *ISE << "\n");
423 AddUserToIVUsers = true;
424 } else if (!AddUsersIfInteresting(User, L, Processed)) {
425 DEBUG(std::cerr << "FOUND USER: " << *User
426 << " OF SCEV: " << *ISE << "\n");
427 AddUserToIVUsers = true;
430 if (AddUserToIVUsers) {
431 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
432 if (StrideUses.Users.empty()) // First occurance of this stride?
433 StrideOrder.push_back(Stride);
435 // Okay, we found a user that we cannot reduce. Analyze the instruction
436 // and decide what to do with it. If we are a use inside of the loop, use
437 // the value before incrementation, otherwise use it after incrementation.
438 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
439 // The value used will be incremented by the stride more than we are
440 // expecting, so subtract this off.
441 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
442 StrideUses.addUser(NewStart, User, I);
443 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
444 DEBUG(std::cerr << " USING POSTINC SCEV, START=" << *NewStart<< "\n");
446 StrideUses.addUser(Start, User, I);
454 /// BasedUser - For a particular base value, keep information about how we've
455 /// partitioned the expression so far.
457 /// Base - The Base value for the PHI node that needs to be inserted for
458 /// this use. As the use is processed, information gets moved from this
459 /// field to the Imm field (below). BasedUser values are sorted by this
463 /// Inst - The instruction using the induction variable.
466 /// OperandValToReplace - The operand value of Inst to replace with the
468 Value *OperandValToReplace;
470 /// Imm - The immediate value that should be added to the base immediately
471 /// before Inst, because it will be folded into the imm field of the
475 /// EmittedBase - The actual value* to use for the base value of this
476 /// operation. This is null if we should just use zero so far.
479 // isUseOfPostIncrementedValue - True if this should use the
480 // post-incremented version of this IV, not the preincremented version.
481 // This can only be set in special cases, such as the terminating setcc
482 // instruction for a loop and uses outside the loop that are dominated by
484 bool isUseOfPostIncrementedValue;
486 BasedUser(IVStrideUse &IVSU)
487 : Base(IVSU.Offset), Inst(IVSU.User),
488 OperandValToReplace(IVSU.OperandValToReplace),
489 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
490 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
492 // Once we rewrite the code to insert the new IVs we want, update the
493 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
495 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
496 SCEVExpander &Rewriter, Loop *L,
499 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
500 SCEVExpander &Rewriter,
501 Instruction *IP, Loop *L);
506 void BasedUser::dump() const {
507 std::cerr << " Base=" << *Base;
508 std::cerr << " Imm=" << *Imm;
510 std::cerr << " EB=" << *EmittedBase;
512 std::cerr << " Inst: " << *Inst;
515 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
516 SCEVExpander &Rewriter,
517 Instruction *IP, Loop *L) {
518 // Figure out where we *really* want to insert this code. In particular, if
519 // the user is inside of a loop that is nested inside of L, we really don't
520 // want to insert this expression before the user, we'd rather pull it out as
521 // many loops as possible.
522 LoopInfo &LI = Rewriter.getLoopInfo();
523 Instruction *BaseInsertPt = IP;
525 // Figure out the most-nested loop that IP is in.
526 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
528 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
529 // the preheader of the outer-most loop where NewBase is not loop invariant.
530 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
531 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
532 InsertLoop = InsertLoop->getParentLoop();
535 // If there is no immediate value, skip the next part.
536 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
537 if (SC->getValue()->isNullValue())
538 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
539 OperandValToReplace->getType());
541 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
543 // Always emit the immediate (if non-zero) into the same block as the user.
544 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
545 return Rewriter.expandCodeFor(NewValSCEV, IP,
546 OperandValToReplace->getType());
550 // Once we rewrite the code to insert the new IVs we want, update the
551 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
553 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
554 SCEVExpander &Rewriter,
556 if (!isa<PHINode>(Inst)) {
557 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
558 // Replace the use of the operand Value with the new Phi we just created.
559 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
560 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
564 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
565 // expression into each operand block that uses it. Note that PHI nodes can
566 // have multiple entries for the same predecessor. We use a map to make sure
567 // that a PHI node only has a single Value* for each predecessor (which also
568 // prevents us from inserting duplicate code in some blocks).
569 std::map<BasicBlock*, Value*> InsertedCode;
570 PHINode *PN = cast<PHINode>(Inst);
571 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
572 if (PN->getIncomingValue(i) == OperandValToReplace) {
573 // If this is a critical edge, split the edge so that we do not insert the
574 // code on all predecessor/successor paths. We do this unless this is the
575 // canonical backedge for this loop, as this can make some inserted code
576 // be in an illegal position.
577 BasicBlock *PHIPred = PN->getIncomingBlock(i);
578 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
579 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
581 // First step, split the critical edge.
582 SplitCriticalEdge(PHIPred, PN->getParent(), P);
584 // Next step: move the basic block. In particular, if the PHI node
585 // is outside of the loop, and PredTI is in the loop, we want to
586 // move the block to be immediately before the PHI block, not
587 // immediately after PredTI.
588 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
589 BasicBlock *NewBB = PN->getIncomingBlock(i);
590 NewBB->moveBefore(PN->getParent());
594 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
596 // Insert the code into the end of the predecessor block.
597 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
598 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
601 // Replace the use of the operand Value with the new Phi we just created.
602 PN->setIncomingValue(i, Code);
606 DEBUG(std::cerr << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst);
610 /// isTargetConstant - Return true if the following can be referenced by the
611 /// immediate field of a target instruction.
612 static bool isTargetConstant(const SCEVHandle &V, const TargetLowering *TLI) {
613 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
614 int64_t V = SC->getValue()->getSExtValue();
616 return TLI->isLegalAddressImmediate(V);
618 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
619 return (V > -(1 << 16) && V < (1 << 16)-1);
622 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
623 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
624 if (CE->getOpcode() == Instruction::Cast) {
625 Constant *Op0 = CE->getOperand(0);
626 if (isa<GlobalValue>(Op0) &&
628 TLI->isLegalAddressImmediate(cast<GlobalValue>(Op0)))
634 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
635 /// loop varying to the Imm operand.
636 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
638 if (Val->isLoopInvariant(L)) return; // Nothing to do.
640 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
641 std::vector<SCEVHandle> NewOps;
642 NewOps.reserve(SAE->getNumOperands());
644 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
645 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
646 // If this is a loop-variant expression, it must stay in the immediate
647 // field of the expression.
648 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
650 NewOps.push_back(SAE->getOperand(i));
654 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
656 Val = SCEVAddExpr::get(NewOps);
657 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
658 // Try to pull immediates out of the start value of nested addrec's.
659 SCEVHandle Start = SARE->getStart();
660 MoveLoopVariantsToImediateField(Start, Imm, L);
662 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
664 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
666 // Otherwise, all of Val is variant, move the whole thing over.
667 Imm = SCEVAddExpr::get(Imm, Val);
668 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
673 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
674 /// that can fit into the immediate field of instructions in the target.
675 /// Accumulate these immediate values into the Imm value.
676 static void MoveImmediateValues(const TargetLowering *TLI,
677 SCEVHandle &Val, SCEVHandle &Imm,
678 bool isAddress, Loop *L) {
679 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
680 std::vector<SCEVHandle> NewOps;
681 NewOps.reserve(SAE->getNumOperands());
683 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
684 SCEVHandle NewOp = SAE->getOperand(i);
685 MoveImmediateValues(TLI, NewOp, Imm, isAddress, L);
687 if (!NewOp->isLoopInvariant(L)) {
688 // If this is a loop-variant expression, it must stay in the immediate
689 // field of the expression.
690 Imm = SCEVAddExpr::get(Imm, NewOp);
692 NewOps.push_back(NewOp);
697 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
699 Val = SCEVAddExpr::get(NewOps);
701 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
702 // Try to pull immediates out of the start value of nested addrec's.
703 SCEVHandle Start = SARE->getStart();
704 MoveImmediateValues(TLI, Start, Imm, isAddress, L);
706 if (Start != SARE->getStart()) {
707 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
709 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
712 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
713 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
714 if (isAddress && isTargetConstant(SME->getOperand(0), TLI) &&
715 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
717 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
718 SCEVHandle NewOp = SME->getOperand(1);
719 MoveImmediateValues(TLI, NewOp, SubImm, isAddress, L);
721 // If we extracted something out of the subexpressions, see if we can
723 if (NewOp != SME->getOperand(1)) {
724 // Scale SubImm up by "8". If the result is a target constant, we are
726 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
727 if (isTargetConstant(SubImm, TLI)) {
728 // Accumulate the immediate.
729 Imm = SCEVAddExpr::get(Imm, SubImm);
731 // Update what is left of 'Val'.
732 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
739 // Loop-variant expressions must stay in the immediate field of the
741 if ((isAddress && isTargetConstant(Val, TLI)) ||
742 !Val->isLoopInvariant(L)) {
743 Imm = SCEVAddExpr::get(Imm, Val);
744 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
748 // Otherwise, no immediates to move.
752 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
753 /// added together. This is used to reassociate common addition subexprs
754 /// together for maximal sharing when rewriting bases.
755 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
757 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
758 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
759 SeparateSubExprs(SubExprs, AE->getOperand(j));
760 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
761 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
762 if (SARE->getOperand(0) == Zero) {
763 SubExprs.push_back(Expr);
765 // Compute the addrec with zero as its base.
766 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
767 Ops[0] = Zero; // Start with zero base.
768 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
771 SeparateSubExprs(SubExprs, SARE->getOperand(0));
773 } else if (!isa<SCEVConstant>(Expr) ||
774 !cast<SCEVConstant>(Expr)->getValue()->isNullValue()) {
776 SubExprs.push_back(Expr);
781 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
782 /// removing any common subexpressions from it. Anything truly common is
783 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
784 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
786 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
787 unsigned NumUses = Uses.size();
789 // Only one use? Use its base, regardless of what it is!
790 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
791 SCEVHandle Result = Zero;
793 std::swap(Result, Uses[0].Base);
797 // To find common subexpressions, count how many of Uses use each expression.
798 // If any subexpressions are used Uses.size() times, they are common.
799 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
801 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
802 // order we see them.
803 std::vector<SCEVHandle> UniqueSubExprs;
805 std::vector<SCEVHandle> SubExprs;
806 for (unsigned i = 0; i != NumUses; ++i) {
807 // If the base is zero (which is common), return zero now, there are no
809 if (Uses[i].Base == Zero) return Zero;
811 // Split the expression into subexprs.
812 SeparateSubExprs(SubExprs, Uses[i].Base);
813 // Add one to SubExpressionUseCounts for each subexpr present.
814 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
815 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
816 UniqueSubExprs.push_back(SubExprs[j]);
820 // Now that we know how many times each is used, build Result. Iterate over
821 // UniqueSubexprs so that we have a stable ordering.
822 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
823 std::map<SCEVHandle, unsigned>::iterator I =
824 SubExpressionUseCounts.find(UniqueSubExprs[i]);
825 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
826 if (I->second == NumUses) { // Found CSE!
827 Result = SCEVAddExpr::get(Result, I->first);
829 // Remove non-cse's from SubExpressionUseCounts.
830 SubExpressionUseCounts.erase(I);
834 // If we found no CSE's, return now.
835 if (Result == Zero) return Result;
837 // Otherwise, remove all of the CSE's we found from each of the base values.
838 for (unsigned i = 0; i != NumUses; ++i) {
839 // Split the expression into subexprs.
840 SeparateSubExprs(SubExprs, Uses[i].Base);
842 // Remove any common subexpressions.
843 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
844 if (SubExpressionUseCounts.count(SubExprs[j])) {
845 SubExprs.erase(SubExprs.begin()+j);
849 // Finally, the non-shared expressions together.
850 if (SubExprs.empty())
853 Uses[i].Base = SCEVAddExpr::get(SubExprs);
860 /// isZero - returns true if the scalar evolution expression is zero.
862 static bool isZero(SCEVHandle &V) {
863 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
864 return SC->getValue()->getZExtValue() == 0;
869 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
870 /// of a previous stride and it is a legal value for the target addressing
871 /// mode scale component. This allows the users of this stride to be rewritten
872 /// as prev iv * factor. It returns 0 if no reuse is possible.
873 unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride,
874 IVExpr &IV, const Type *Ty) {
877 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
878 int64_t SInt = SC->getValue()->getSExtValue();
879 if (SInt == 1) return 0;
881 for (TargetLowering::legal_am_scale_iterator
882 I = TLI->legal_am_scale_begin(), E = TLI->legal_am_scale_end();
885 if (unsigned(abs(SInt)) < Scale || (SInt % Scale) != 0)
887 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
888 IVsByStride.find(SCEVUnknown::getIntegerSCEV(SInt/Scale, Type::UIntTy));
889 if (SI == IVsByStride.end())
891 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
892 IE = SI->second.IVs.end(); II != IE; ++II)
893 // FIXME: Only handle base == 0 for now.
894 // Only reuse previous IV if it would not require a type conversion.
895 if (isZero(II->Base) &&
896 II->Base->getType()->isLosslesslyConvertibleTo(Ty)) {
906 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
907 /// returns true if Val's isUseOfPostIncrementedValue is true.
908 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
909 return Val.isUseOfPostIncrementedValue;
912 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
913 /// stride of IV. All of the users may have different starting values, and this
914 /// may not be the only stride (we know it is if isOnlyStride is true).
915 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
916 IVUsersOfOneStride &Uses,
919 // Transform our list of users and offsets to a bit more complex table. In
920 // this new vector, each 'BasedUser' contains 'Base' the base of the
921 // strided accessas well as the old information from Uses. We progressively
922 // move information from the Base field to the Imm field, until we eventually
923 // have the full access expression to rewrite the use.
924 std::vector<BasedUser> UsersToProcess;
925 UsersToProcess.reserve(Uses.Users.size());
926 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
927 UsersToProcess.push_back(Uses.Users[i]);
929 // Move any loop invariant operands from the offset field to the immediate
930 // field of the use, so that we don't try to use something before it is
932 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
933 UsersToProcess.back().Imm, L);
934 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
935 "Base value is not loop invariant!");
938 // We now have a whole bunch of uses of like-strided induction variables, but
939 // they might all have different bases. We want to emit one PHI node for this
940 // stride which we fold as many common expressions (between the IVs) into as
941 // possible. Start by identifying the common expressions in the base values
942 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
943 // "A+B"), emit it to the preheader, then remove the expression from the
944 // UsersToProcess base values.
945 SCEVHandle CommonExprs =
946 RemoveCommonExpressionsFromUseBases(UsersToProcess);
948 // Check if it is possible to reuse a IV with stride that is factor of this
949 // stride. And the multiple is a number that can be encoded in the scale
950 // field of the target addressing mode.
951 PHINode *NewPHI = NULL;
954 unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
955 CommonExprs->getType());
956 if (RewriteFactor != 0) {
957 DEBUG(std::cerr << "BASED ON IV of STRIDE " << *ReuseIV.Stride
958 << " and BASE " << *ReuseIV.Base << " :\n");
959 NewPHI = ReuseIV.PHI;
963 // Next, figure out what we can represent in the immediate fields of
964 // instructions. If we can represent anything there, move it to the imm
965 // fields of the BasedUsers. We do this so that it increases the commonality
966 // of the remaining uses.
967 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
968 // If the user is not in the current loop, this means it is using the exit
969 // value of the IV. Do not put anything in the base, make sure it's all in
970 // the immediate field to allow as much factoring as possible.
971 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
972 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
973 UsersToProcess[i].Base);
974 UsersToProcess[i].Base =
975 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
978 // Addressing modes can be folded into loads and stores. Be careful that
979 // the store is through the expression, not of the expression though.
980 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
981 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
982 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
985 MoveImmediateValues(TLI, UsersToProcess[i].Base, UsersToProcess[i].Imm,
990 // Now that we know what we need to do, insert the PHI node itself.
992 DEBUG(std::cerr << "INSERTING IV of STRIDE " << *Stride << " and BASE "
993 << *CommonExprs << " :\n");
995 SCEVExpander Rewriter(*SE, *LI);
996 SCEVExpander PreheaderRewriter(*SE, *LI);
998 BasicBlock *Preheader = L->getLoopPreheader();
999 Instruction *PreInsertPt = Preheader->getTerminator();
1000 Instruction *PhiInsertBefore = L->getHeader()->begin();
1002 BasicBlock *LatchBlock = L->getLoopLatch();
1004 const Type *ReplacedTy = CommonExprs->getType();
1006 // Emit the initial base value into the loop preheader.
1008 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
1011 if (RewriteFactor == 0) {
1012 // Create a new Phi for this base, and stick it in the loop header.
1013 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1016 // Add common base to the new Phi node.
1017 NewPHI->addIncoming(CommonBaseV, Preheader);
1019 // Insert the stride into the preheader.
1020 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
1022 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1024 // Emit the increment of the base value before the terminator of the loop
1025 // latch block, and add it to the Phi node.
1026 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
1027 SCEVUnknown::get(StrideV));
1029 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
1031 IncV->setName(NewPHI->getName()+".inc");
1032 NewPHI->addIncoming(IncV, LatchBlock);
1034 // Remember this in case a later stride is multiple of this.
1035 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1037 Constant *C = dyn_cast<Constant>(CommonBaseV);
1039 (!C->isNullValue() &&
1040 !isTargetConstant(SCEVUnknown::get(CommonBaseV), TLI)))
1041 // We want the common base emitted into the preheader!
1042 CommonBaseV = new CastInst(CommonBaseV, CommonBaseV->getType(),
1043 "commonbase", PreInsertPt);
1046 // We want to emit code for users inside the loop first. To do this, we
1047 // rearrange BasedUser so that the entries at the end have
1048 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1049 // vector (so we handle them first).
1050 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1051 PartitionByIsUseOfPostIncrementedValue);
1053 // Sort this by base, so that things with the same base are handled
1054 // together. By partitioning first and stable-sorting later, we are
1055 // guaranteed that within each base we will pop off users from within the
1056 // loop before users outside of the loop with a particular base.
1058 // We would like to use stable_sort here, but we can't. The problem is that
1059 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1060 // we don't have anything to do a '<' comparison on. Because we think the
1061 // number of uses is small, do a horrible bubble sort which just relies on
1063 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1064 // Get a base value.
1065 SCEVHandle Base = UsersToProcess[i].Base;
1067 // Compact everything with this base to be consequetive with this one.
1068 for (unsigned j = i+1; j != e; ++j) {
1069 if (UsersToProcess[j].Base == Base) {
1070 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1076 // Process all the users now. This outer loop handles all bases, the inner
1077 // loop handles all users of a particular base.
1078 while (!UsersToProcess.empty()) {
1079 SCEVHandle Base = UsersToProcess.back().Base;
1081 DEBUG(std::cerr << " INSERTING code for BASE = " << *Base << ":\n");
1083 // Emit the code for Base into the preheader.
1084 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
1087 // If BaseV is a constant other than 0, make sure that it gets inserted into
1088 // the preheader, instead of being forward substituted into the uses. We do
1089 // this by forcing a noop cast to be inserted into the preheader in this
1091 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1092 if (!C->isNullValue() && !isTargetConstant(Base, TLI)) {
1093 // We want this constant emitted into the preheader!
1094 BaseV = new CastInst(BaseV, BaseV->getType(), "preheaderinsert",
1099 // Emit the code to add the immediate offset to the Phi value, just before
1100 // the instructions that we identified as using this stride and base.
1102 // FIXME: Use emitted users to emit other users.
1103 BasedUser &User = UsersToProcess.back();
1105 // If this instruction wants to use the post-incremented value, move it
1106 // after the post-inc and use its value instead of the PHI.
1107 Value *RewriteOp = NewPHI;
1108 if (User.isUseOfPostIncrementedValue) {
1111 // If this user is in the loop, make sure it is the last thing in the
1112 // loop to ensure it is dominated by the increment.
1113 if (L->contains(User.Inst->getParent()))
1114 User.Inst->moveBefore(LatchBlock->getTerminator());
1116 if (RewriteOp->getType() != ReplacedTy)
1117 RewriteOp = SCEVExpander::InsertCastOfTo(RewriteOp, ReplacedTy);
1119 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
1121 // Clear the SCEVExpander's expression map so that we are guaranteed
1122 // to have the code emitted where we expect it.
1125 // If we are reusing the iv, then it must be multiplied by a constant
1126 // factor take advantage of addressing mode scale component.
1127 if (RewriteFactor != 0) {
1129 SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
1130 RewriteExpr->getType()),
1133 // The common base is emitted in the loop preheader. But since we
1134 // are reusing an IV, it has not been used to initialize the PHI node.
1135 // Add it to the expression used to rewrite the uses.
1136 if (!isa<ConstantInt>(CommonBaseV) ||
1137 !cast<ConstantInt>(CommonBaseV)->isNullValue())
1138 RewriteExpr = SCEVAddExpr::get(RewriteExpr,
1139 SCEVUnknown::get(CommonBaseV));
1142 // Now that we know what we need to do, insert code before User for the
1143 // immediate and any loop-variant expressions.
1144 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isNullValue())
1145 // Add BaseV to the PHI value if needed.
1146 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
1148 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1150 // Mark old value we replaced as possibly dead, so that it is elminated
1151 // if we just replaced the last use of that value.
1152 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1154 UsersToProcess.pop_back();
1157 // If there are any more users to process with the same base, process them
1158 // now. We sorted by base above, so we just have to check the last elt.
1159 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1160 // TODO: Next, find out which base index is the most common, pull it out.
1163 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1164 // different starting values, into different PHIs.
1167 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1168 // uses in the loop, look to see if we can eliminate some, in favor of using
1169 // common indvars for the different uses.
1170 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1171 // TODO: implement optzns here.
1176 // Finally, get the terminating condition for the loop if possible. If we
1177 // can, we want to change it to use a post-incremented version of its
1178 // induction variable, to allow coalescing the live ranges for the IV into
1179 // one register value.
1180 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1181 BasicBlock *Preheader = L->getLoopPreheader();
1182 BasicBlock *LatchBlock =
1183 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1184 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1185 if (!TermBr || TermBr->isUnconditional() ||
1186 !isa<SetCondInst>(TermBr->getCondition()))
1188 SetCondInst *Cond = cast<SetCondInst>(TermBr->getCondition());
1190 // Search IVUsesByStride to find Cond's IVUse if there is one.
1191 IVStrideUse *CondUse = 0;
1192 const SCEVHandle *CondStride = 0;
1194 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1196 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1197 IVUsesByStride.find(StrideOrder[Stride]);
1198 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1200 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1201 E = SI->second.Users.end(); UI != E; ++UI)
1202 if (UI->User == Cond) {
1204 CondStride = &SI->first;
1205 // NOTE: we could handle setcc instructions with multiple uses here, but
1206 // InstCombine does it as well for simple uses, it's not clear that it
1207 // occurs enough in real life to handle.
1211 if (!CondUse) return; // setcc doesn't use the IV.
1213 // setcc stride is complex, don't mess with users.
1214 // FIXME: Evaluate whether this is a good idea or not.
1215 if (!isa<SCEVConstant>(*CondStride)) return;
1217 // It's possible for the setcc instruction to be anywhere in the loop, and
1218 // possible for it to have multiple users. If it is not immediately before
1219 // the latch block branch, move it.
1220 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1221 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1222 Cond->moveBefore(TermBr);
1224 // Otherwise, clone the terminating condition and insert into the loopend.
1225 Cond = cast<SetCondInst>(Cond->clone());
1226 Cond->setName(L->getHeader()->getName() + ".termcond");
1227 LatchBlock->getInstList().insert(TermBr, Cond);
1229 // Clone the IVUse, as the old use still exists!
1230 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1231 CondUse->OperandValToReplace);
1232 CondUse = &IVUsesByStride[*CondStride].Users.back();
1236 // If we get to here, we know that we can transform the setcc instruction to
1237 // use the post-incremented version of the IV, allowing us to coalesce the
1238 // live ranges for the IV correctly.
1239 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1240 CondUse->isUseOfPostIncrementedValue = true;
1244 // Constant strides come first which in turns are sorted by their absolute
1245 // values. If absolute values are the same, then positive strides comes first.
1247 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1248 struct StrideCompare {
1249 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1250 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1251 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1253 int64_t LV = LHSC->getValue()->getSExtValue();
1254 int64_t RV = RHSC->getValue()->getSExtValue();
1255 uint64_t ALV = (LV < 0) ? -LV : LV;
1256 uint64_t ARV = (RV < 0) ? -RV : RV;
1262 return (LHSC && !RHSC);
1267 void LoopStrengthReduce::runOnLoop(Loop *L) {
1268 // First step, transform all loops nesting inside of this loop.
1269 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
1272 // Next, find all uses of induction variables in this loop, and catagorize
1273 // them by stride. Start by finding all of the PHI nodes in the header for
1274 // this loop. If they are induction variables, inspect their uses.
1275 std::set<Instruction*> Processed; // Don't reprocess instructions.
1276 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1277 AddUsersIfInteresting(I, L, Processed);
1279 // If we have nothing to do, return.
1280 if (IVUsesByStride.empty()) return;
1282 // Optimize induction variables. Some indvar uses can be transformed to use
1283 // strides that will be needed for other purposes. A common example of this
1284 // is the exit test for the loop, which can often be rewritten to use the
1285 // computation of some other indvar to decide when to terminate the loop.
1289 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1290 // doing computation in byte values, promote to 32-bit values if safe.
1292 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1293 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1294 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1295 // to be careful that IV's are all the same type. Only works for intptr_t
1298 // If we only have one stride, we can more aggressively eliminate some things.
1299 bool HasOneStride = IVUsesByStride.size() == 1;
1302 DEBUG(std::cerr << "\nLSR on ");
1306 // IVsByStride keeps IVs for one particular loop.
1307 IVsByStride.clear();
1309 // Sort the StrideOrder so we process larger strides first.
1310 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1312 // Note: this processes each stride/type pair individually. All users passed
1313 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1314 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1315 // This extra layer of indirection makes the ordering of strides deterministic
1316 // - not dependent on map order.
1317 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1318 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1319 IVUsesByStride.find(StrideOrder[Stride]);
1320 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1321 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1324 // Clean up after ourselves
1325 if (!DeadInsts.empty()) {
1326 DeleteTriviallyDeadInstructions(DeadInsts);
1328 BasicBlock::iterator I = L->getHeader()->begin();
1330 while ((PN = dyn_cast<PHINode>(I))) {
1331 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1333 // At this point, we know that we have killed one or more GEP
1334 // instructions. It is worth checking to see if the cann indvar is also
1335 // dead, so that we can remove it as well. The requirements for the cann
1336 // indvar to be considered dead are:
1337 // 1. the cann indvar has one use
1338 // 2. the use is an add instruction
1339 // 3. the add has one use
1340 // 4. the add is used by the cann indvar
1341 // If all four cases above are true, then we can remove both the add and
1343 // FIXME: this needs to eliminate an induction variable even if it's being
1344 // compared against some value to decide loop termination.
1345 if (PN->hasOneUse()) {
1346 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin()));
1347 if (BO && BO->hasOneUse()) {
1348 if (PN == *(BO->use_begin())) {
1349 DeadInsts.insert(BO);
1350 // Break the cycle, then delete the PHI.
1351 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1352 SE->deleteInstructionFromRecords(PN);
1353 PN->eraseFromParent();
1358 DeleteTriviallyDeadInstructions(DeadInsts);
1361 CastedPointers.clear();
1362 IVUsesByStride.clear();
1363 StrideOrder.clear();