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/LoopPass.h"
27 #include "llvm/Analysis/ScalarEvolutionExpander.h"
28 #include "llvm/Support/CFG.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/Compiler.h"
36 #include "llvm/Target/TargetLowering.h"
41 STATISTIC(NumReduced , "Number of GEPs strength reduced");
42 STATISTIC(NumInserted, "Number of PHIs inserted");
43 STATISTIC(NumVariable, "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.
50 struct VISIBILITY_HIDDEN IVStrideUse {
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 VISIBILITY_HIDDEN 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.
83 struct VISIBILITY_HIDDEN IVExpr {
90 : Stride(SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)),
91 Base (SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)) {}
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 VISIBILITY_HIDDEN 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 LoopPass {
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 bool runOnLoop(Loop *L, LPPassManager &LPM);
149 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
150 // We split critical edges, so we change the CFG. However, we do update
151 // many analyses if they are around.
152 AU.addPreservedID(LoopSimplifyID);
153 AU.addPreserved<LoopInfo>();
154 AU.addPreserved<DominatorSet>();
155 AU.addPreserved<ETForest>();
156 AU.addPreserved<ImmediateDominators>();
157 AU.addPreserved<DominanceFrontier>();
158 AU.addPreserved<DominatorTree>();
160 AU.addRequiredID(LoopSimplifyID);
161 AU.addRequired<LoopInfo>();
162 AU.addRequired<ETForest>();
163 AU.addRequired<TargetData>();
164 AU.addRequired<ScalarEvolution>();
167 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
169 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
171 bool AddUsersIfInteresting(Instruction *I, Loop *L,
172 std::set<Instruction*> &Processed);
173 SCEVHandle GetExpressionSCEV(Instruction *E, Loop *L);
175 void OptimizeIndvars(Loop *L);
177 unsigned CheckForIVReuse(const SCEVHandle&, IVExpr&, const Type*);
179 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
180 IVUsersOfOneStride &Uses,
181 Loop *L, bool isOnlyStride);
182 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
184 RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
187 LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
188 return new LoopStrengthReduce(TLI);
191 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
192 /// assumes that the Value* V is of integer or pointer type only.
194 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
196 if (V->getType() == UIntPtrTy) return V;
197 if (Constant *CB = dyn_cast<Constant>(V))
198 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
200 Value *&New = CastedPointers[V];
203 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
204 DeadInsts.insert(cast<Instruction>(New));
209 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
210 /// specified set are trivially dead, delete them and see if this makes any of
211 /// their operands subsequently dead.
212 void LoopStrengthReduce::
213 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
214 while (!Insts.empty()) {
215 Instruction *I = *Insts.begin();
216 Insts.erase(Insts.begin());
217 if (isInstructionTriviallyDead(I)) {
218 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
219 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
221 SE->deleteInstructionFromRecords(I);
222 I->eraseFromParent();
229 /// GetExpressionSCEV - Compute and return the SCEV for the specified
231 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
232 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
233 // If this is a GEP that SE doesn't know about, compute it now and insert it.
234 // If this is not a GEP, or if we have already done this computation, just let
236 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
237 if (!GEP || SE->hasSCEV(GEP))
238 return SE->getSCEV(Exp);
240 // Analyze all of the subscripts of this getelementptr instruction, looking
241 // for uses that are determined by the trip count of L. First, skip all
242 // operands the are not dependent on the IV.
244 // Build up the base expression. Insert an LLVM cast of the pointer to
246 SCEVHandle GEPVal = SCEVUnknown::get(
247 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
249 gep_type_iterator GTI = gep_type_begin(GEP);
251 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
252 // If this is a use of a recurrence that we can analyze, and it comes before
253 // Op does in the GEP operand list, we will handle this when we process this
255 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
256 const StructLayout *SL = TD->getStructLayout(STy);
257 unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
258 uint64_t Offset = SL->getElementOffset(Idx);
259 GEPVal = SCEVAddExpr::get(GEPVal,
260 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
262 unsigned GEPOpiBits =
263 GEP->getOperand(i)->getType()->getPrimitiveSizeInBits();
264 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
265 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
266 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
267 Instruction::BitCast));
268 Value *OpVal = getCastedVersionOf(opcode, GEP->getOperand(i));
269 SCEVHandle Idx = SE->getSCEV(OpVal);
271 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
273 Idx = SCEVMulExpr::get(Idx,
274 SCEVConstant::get(ConstantInt::get(UIntPtrTy,
276 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
280 SE->setSCEV(GEP, GEPVal);
284 /// getSCEVStartAndStride - Compute the start and stride of this expression,
285 /// returning false if the expression is not a start/stride pair, or true if it
286 /// is. The stride must be a loop invariant expression, but the start may be
287 /// a mix of loop invariant and loop variant expressions.
288 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
289 SCEVHandle &Start, SCEVHandle &Stride) {
290 SCEVHandle TheAddRec = Start; // Initialize to zero.
292 // If the outer level is an AddExpr, the operands are all start values except
293 // for a nested AddRecExpr.
294 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
295 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
296 if (SCEVAddRecExpr *AddRec =
297 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
298 if (AddRec->getLoop() == L)
299 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
301 return false; // Nested IV of some sort?
303 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
306 } else if (isa<SCEVAddRecExpr>(SH)) {
309 return false; // not analyzable.
312 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
313 if (!AddRec || AddRec->getLoop() != L) return false;
315 // FIXME: Generalize to non-affine IV's.
316 if (!AddRec->isAffine()) return false;
318 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
320 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
321 DOUT << "[" << L->getHeader()->getName()
322 << "] Variable stride: " << *AddRec << "\n";
324 Stride = AddRec->getOperand(1);
328 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
329 /// and now we need to decide whether the user should use the preinc or post-inc
330 /// value. If this user should use the post-inc version of the IV, return true.
332 /// Choosing wrong here can break dominance properties (if we choose to use the
333 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
334 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
335 /// should use the post-inc value).
336 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
337 Loop *L, ETForest *EF, Pass *P) {
338 // If the user is in the loop, use the preinc value.
339 if (L->contains(User->getParent())) return false;
341 BasicBlock *LatchBlock = L->getLoopLatch();
343 // Ok, the user is outside of the loop. If it is dominated by the latch
344 // block, use the post-inc value.
345 if (EF->dominates(LatchBlock, User->getParent()))
348 // There is one case we have to be careful of: PHI nodes. These little guys
349 // can live in blocks that do not dominate the latch block, but (since their
350 // uses occur in the predecessor block, not the block the PHI lives in) should
351 // still use the post-inc value. Check for this case now.
352 PHINode *PN = dyn_cast<PHINode>(User);
353 if (!PN) return false; // not a phi, not dominated by latch block.
355 // Look at all of the uses of IV by the PHI node. If any use corresponds to
356 // a block that is not dominated by the latch block, give up and use the
357 // preincremented value.
358 unsigned NumUses = 0;
359 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
360 if (PN->getIncomingValue(i) == IV) {
362 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
366 // Okay, all uses of IV by PN are in predecessor blocks that really are
367 // dominated by the latch block. Split the critical edges and use the
368 // post-incremented value.
369 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
370 if (PN->getIncomingValue(i) == IV) {
371 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P,
373 // Splitting the critical edge can reduce the number of entries in this
375 e = PN->getNumIncomingValues();
376 if (--NumUses == 0) break;
384 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
385 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
386 /// return true. Otherwise, return false.
387 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
388 std::set<Instruction*> &Processed) {
389 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
390 return false; // Void and FP expressions cannot be reduced.
391 if (!Processed.insert(I).second)
392 return true; // Instruction already handled.
394 // Get the symbolic expression for this instruction.
395 SCEVHandle ISE = GetExpressionSCEV(I, L);
396 if (isa<SCEVCouldNotCompute>(ISE)) return false;
398 // Get the start and stride for this expression.
399 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
400 SCEVHandle Stride = Start;
401 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
402 return false; // Non-reducible symbolic expression, bail out.
404 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;) {
405 Instruction *User = cast<Instruction>(*UI);
407 // Increment iterator now because IVUseShouldUsePostIncValue may remove
408 // User from the list of I users.
411 // Do not infinitely recurse on PHI nodes.
412 if (isa<PHINode>(User) && Processed.count(User))
415 // If this is an instruction defined in a nested loop, or outside this loop,
416 // don't recurse into it.
417 bool AddUserToIVUsers = false;
418 if (LI->getLoopFor(User->getParent()) != L) {
419 DOUT << "FOUND USER in other loop: " << *User
420 << " OF SCEV: " << *ISE << "\n";
421 AddUserToIVUsers = true;
422 } else if (!AddUsersIfInteresting(User, L, Processed)) {
423 DOUT << "FOUND USER: " << *User
424 << " OF SCEV: " << *ISE << "\n";
425 AddUserToIVUsers = true;
428 if (AddUserToIVUsers) {
429 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
430 if (StrideUses.Users.empty()) // First occurance of this stride?
431 StrideOrder.push_back(Stride);
433 // Okay, we found a user that we cannot reduce. Analyze the instruction
434 // and decide what to do with it. If we are a use inside of the loop, use
435 // the value before incrementation, otherwise use it after incrementation.
436 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
437 // The value used will be incremented by the stride more than we are
438 // expecting, so subtract this off.
439 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
440 StrideUses.addUser(NewStart, User, I);
441 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
442 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
444 StrideUses.addUser(Start, User, I);
452 /// BasedUser - For a particular base value, keep information about how we've
453 /// partitioned the expression so far.
455 /// Base - The Base value for the PHI node that needs to be inserted for
456 /// this use. As the use is processed, information gets moved from this
457 /// field to the Imm field (below). BasedUser values are sorted by this
461 /// Inst - The instruction using the induction variable.
464 /// OperandValToReplace - The operand value of Inst to replace with the
466 Value *OperandValToReplace;
468 /// Imm - The immediate value that should be added to the base immediately
469 /// before Inst, because it will be folded into the imm field of the
473 /// EmittedBase - The actual value* to use for the base value of this
474 /// operation. This is null if we should just use zero so far.
477 // isUseOfPostIncrementedValue - True if this should use the
478 // post-incremented version of this IV, not the preincremented version.
479 // This can only be set in special cases, such as the terminating setcc
480 // instruction for a loop and uses outside the loop that are dominated by
482 bool isUseOfPostIncrementedValue;
484 BasedUser(IVStrideUse &IVSU)
485 : Base(IVSU.Offset), Inst(IVSU.User),
486 OperandValToReplace(IVSU.OperandValToReplace),
487 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
488 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
490 // Once we rewrite the code to insert the new IVs we want, update the
491 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
493 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
494 SCEVExpander &Rewriter, Loop *L,
497 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
498 SCEVExpander &Rewriter,
499 Instruction *IP, Loop *L);
504 void BasedUser::dump() const {
505 cerr << " Base=" << *Base;
506 cerr << " Imm=" << *Imm;
508 cerr << " EB=" << *EmittedBase;
510 cerr << " Inst: " << *Inst;
513 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
514 SCEVExpander &Rewriter,
515 Instruction *IP, Loop *L) {
516 // Figure out where we *really* want to insert this code. In particular, if
517 // the user is inside of a loop that is nested inside of L, we really don't
518 // want to insert this expression before the user, we'd rather pull it out as
519 // many loops as possible.
520 LoopInfo &LI = Rewriter.getLoopInfo();
521 Instruction *BaseInsertPt = IP;
523 // Figure out the most-nested loop that IP is in.
524 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
526 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
527 // the preheader of the outer-most loop where NewBase is not loop invariant.
528 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
529 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
530 InsertLoop = InsertLoop->getParentLoop();
533 // If there is no immediate value, skip the next part.
534 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
535 if (SC->getValue()->isZero())
536 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
537 OperandValToReplace->getType());
539 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
541 // Always emit the immediate (if non-zero) into the same block as the user.
542 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
543 return Rewriter.expandCodeFor(NewValSCEV, IP,
544 OperandValToReplace->getType());
548 // Once we rewrite the code to insert the new IVs we want, update the
549 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
551 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
552 SCEVExpander &Rewriter,
554 if (!isa<PHINode>(Inst)) {
555 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, Inst, L);
556 // Replace the use of the operand Value with the new Phi we just created.
557 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
558 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
562 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
563 // expression into each operand block that uses it. Note that PHI nodes can
564 // have multiple entries for the same predecessor. We use a map to make sure
565 // that a PHI node only has a single Value* for each predecessor (which also
566 // prevents us from inserting duplicate code in some blocks).
567 std::map<BasicBlock*, Value*> InsertedCode;
568 PHINode *PN = cast<PHINode>(Inst);
569 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
570 if (PN->getIncomingValue(i) == OperandValToReplace) {
571 // If this is a critical edge, split the edge so that we do not insert the
572 // code on all predecessor/successor paths. We do this unless this is the
573 // canonical backedge for this loop, as this can make some inserted code
574 // be in an illegal position.
575 BasicBlock *PHIPred = PN->getIncomingBlock(i);
576 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
577 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
579 // First step, split the critical edge.
580 SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
582 // Next step: move the basic block. In particular, if the PHI node
583 // is outside of the loop, and PredTI is in the loop, we want to
584 // move the block to be immediately before the PHI block, not
585 // immediately after PredTI.
586 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
587 BasicBlock *NewBB = PN->getIncomingBlock(i);
588 NewBB->moveBefore(PN->getParent());
591 // Splitting the edge can reduce the number of PHI entries we have.
592 e = PN->getNumIncomingValues();
595 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
597 // Insert the code into the end of the predecessor block.
598 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
599 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
602 // Replace the use of the operand Value with the new Phi we just created.
603 PN->setIncomingValue(i, Code);
607 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
611 /// isTargetConstant - Return true if the following can be referenced by the
612 /// immediate field of a target instruction.
613 static bool isTargetConstant(const SCEVHandle &V, const TargetLowering *TLI) {
614 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
615 int64_t V = SC->getValue()->getSExtValue();
617 return TLI->isLegalAddressImmediate(V);
619 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
620 return (V > -(1 << 16) && V < (1 << 16)-1);
623 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
624 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
625 if (CE->getOpcode() == Instruction::PtrToInt) {
626 Constant *Op0 = CE->getOperand(0);
627 if (isa<GlobalValue>(Op0) && TLI &&
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()->isZero()) {
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()->isZero();
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, UIntPtrTy));
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) && II->Base->getType() == Ty) {
905 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
906 /// returns true if Val's isUseOfPostIncrementedValue is true.
907 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
908 return Val.isUseOfPostIncrementedValue;
911 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
912 /// stride of IV. All of the users may have different starting values, and this
913 /// may not be the only stride (we know it is if isOnlyStride is true).
914 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
915 IVUsersOfOneStride &Uses,
918 // Transform our list of users and offsets to a bit more complex table. In
919 // this new vector, each 'BasedUser' contains 'Base' the base of the
920 // strided accessas well as the old information from Uses. We progressively
921 // move information from the Base field to the Imm field, until we eventually
922 // have the full access expression to rewrite the use.
923 std::vector<BasedUser> UsersToProcess;
924 UsersToProcess.reserve(Uses.Users.size());
925 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
926 UsersToProcess.push_back(Uses.Users[i]);
928 // Move any loop invariant operands from the offset field to the immediate
929 // field of the use, so that we don't try to use something before it is
931 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
932 UsersToProcess.back().Imm, L);
933 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
934 "Base value is not loop invariant!");
937 // We now have a whole bunch of uses of like-strided induction variables, but
938 // they might all have different bases. We want to emit one PHI node for this
939 // stride which we fold as many common expressions (between the IVs) into as
940 // possible. Start by identifying the common expressions in the base values
941 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
942 // "A+B"), emit it to the preheader, then remove the expression from the
943 // UsersToProcess base values.
944 SCEVHandle CommonExprs =
945 RemoveCommonExpressionsFromUseBases(UsersToProcess);
947 // Check if it is possible to reuse a IV with stride that is factor of this
948 // stride. And the multiple is a number that can be encoded in the scale
949 // field of the target addressing mode.
950 PHINode *NewPHI = NULL;
953 unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
954 CommonExprs->getType());
955 if (RewriteFactor != 0) {
956 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
957 << " and BASE " << *ReuseIV.Base << " :\n";
958 NewPHI = ReuseIV.PHI;
962 // Next, figure out what we can represent in the immediate fields of
963 // instructions. If we can represent anything there, move it to the imm
964 // fields of the BasedUsers. We do this so that it increases the commonality
965 // of the remaining uses.
966 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
967 // If the user is not in the current loop, this means it is using the exit
968 // value of the IV. Do not put anything in the base, make sure it's all in
969 // the immediate field to allow as much factoring as possible.
970 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
971 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
972 UsersToProcess[i].Base);
973 UsersToProcess[i].Base =
974 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
977 // Addressing modes can be folded into loads and stores. Be careful that
978 // the store is through the expression, not of the expression though.
979 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
980 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
981 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
984 MoveImmediateValues(TLI, UsersToProcess[i].Base, UsersToProcess[i].Imm,
989 // Now that we know what we need to do, insert the PHI node itself.
991 DOUT << "INSERTING IV of STRIDE " << *Stride << " and BASE "
992 << *CommonExprs << " :\n";
994 SCEVExpander Rewriter(*SE, *LI);
995 SCEVExpander PreheaderRewriter(*SE, *LI);
997 BasicBlock *Preheader = L->getLoopPreheader();
998 Instruction *PreInsertPt = Preheader->getTerminator();
999 Instruction *PhiInsertBefore = L->getHeader()->begin();
1001 BasicBlock *LatchBlock = L->getLoopLatch();
1003 const Type *ReplacedTy = CommonExprs->getType();
1005 // Emit the initial base value into the loop preheader.
1007 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
1010 if (RewriteFactor == 0) {
1011 // Create a new Phi for this base, and stick it in the loop header.
1012 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1015 // Add common base to the new Phi node.
1016 NewPHI->addIncoming(CommonBaseV, Preheader);
1018 // Insert the stride into the preheader.
1019 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
1021 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1023 // Emit the increment of the base value before the terminator of the loop
1024 // latch block, and add it to the Phi node.
1025 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
1026 SCEVUnknown::get(StrideV));
1028 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
1030 IncV->setName(NewPHI->getName()+".inc");
1031 NewPHI->addIncoming(IncV, LatchBlock);
1033 // Remember this in case a later stride is multiple of this.
1034 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1036 Constant *C = dyn_cast<Constant>(CommonBaseV);
1038 (!C->isNullValue() &&
1039 !isTargetConstant(SCEVUnknown::get(CommonBaseV), TLI)))
1040 // We want the common base emitted into the preheader! This is just
1041 // using cast as a copy so BitCast (no-op cast) is appropriate
1042 CommonBaseV = new BitCastInst(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 DOUT << " 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 BitCast (noop cast) to be inserted into the preheader
1091 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1092 if (!C->isNullValue() && !isTargetConstant(Base, TLI)) {
1093 // We want this constant emitted into the preheader! This is just
1094 // using cast as a copy so BitCast (no-op cast) is appropriate
1095 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1100 // Emit the code to add the immediate offset to the Phi value, just before
1101 // the instructions that we identified as using this stride and base.
1103 // FIXME: Use emitted users to emit other users.
1104 BasedUser &User = UsersToProcess.back();
1106 // If this instruction wants to use the post-incremented value, move it
1107 // after the post-inc and use its value instead of the PHI.
1108 Value *RewriteOp = NewPHI;
1109 if (User.isUseOfPostIncrementedValue) {
1112 // If this user is in the loop, make sure it is the last thing in the
1113 // loop to ensure it is dominated by the increment.
1114 if (L->contains(User.Inst->getParent()))
1115 User.Inst->moveBefore(LatchBlock->getTerminator());
1117 if (RewriteOp->getType() != ReplacedTy) {
1118 Instruction::CastOps opcode = Instruction::Trunc;
1119 if (ReplacedTy->getPrimitiveSizeInBits() ==
1120 RewriteOp->getType()->getPrimitiveSizeInBits())
1121 opcode = Instruction::BitCast;
1122 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1125 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
1127 // Clear the SCEVExpander's expression map so that we are guaranteed
1128 // to have the code emitted where we expect it.
1131 // If we are reusing the iv, then it must be multiplied by a constant
1132 // factor take advantage of addressing mode scale component.
1133 if (RewriteFactor != 0) {
1135 SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
1136 RewriteExpr->getType()),
1139 // The common base is emitted in the loop preheader. But since we
1140 // are reusing an IV, it has not been used to initialize the PHI node.
1141 // Add it to the expression used to rewrite the uses.
1142 if (!isa<ConstantInt>(CommonBaseV) ||
1143 !cast<ConstantInt>(CommonBaseV)->isZero())
1144 RewriteExpr = SCEVAddExpr::get(RewriteExpr,
1145 SCEVUnknown::get(CommonBaseV));
1148 // Now that we know what we need to do, insert code before User for the
1149 // immediate and any loop-variant expressions.
1150 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1151 // Add BaseV to the PHI value if needed.
1152 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
1154 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1156 // Mark old value we replaced as possibly dead, so that it is elminated
1157 // if we just replaced the last use of that value.
1158 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1160 UsersToProcess.pop_back();
1163 // If there are any more users to process with the same base, process them
1164 // now. We sorted by base above, so we just have to check the last elt.
1165 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1166 // TODO: Next, find out which base index is the most common, pull it out.
1169 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1170 // different starting values, into different PHIs.
1173 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1174 // uses in the loop, look to see if we can eliminate some, in favor of using
1175 // common indvars for the different uses.
1176 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1177 // TODO: implement optzns here.
1179 // Finally, get the terminating condition for the loop if possible. If we
1180 // can, we want to change it to use a post-incremented version of its
1181 // induction variable, to allow coalescing the live ranges for the IV into
1182 // one register value.
1183 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1184 BasicBlock *Preheader = L->getLoopPreheader();
1185 BasicBlock *LatchBlock =
1186 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1187 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1188 if (!TermBr || TermBr->isUnconditional() ||
1189 !isa<ICmpInst>(TermBr->getCondition()))
1191 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1193 // Search IVUsesByStride to find Cond's IVUse if there is one.
1194 IVStrideUse *CondUse = 0;
1195 const SCEVHandle *CondStride = 0;
1197 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1199 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1200 IVUsesByStride.find(StrideOrder[Stride]);
1201 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1203 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1204 E = SI->second.Users.end(); UI != E; ++UI)
1205 if (UI->User == Cond) {
1207 CondStride = &SI->first;
1208 // NOTE: we could handle setcc instructions with multiple uses here, but
1209 // InstCombine does it as well for simple uses, it's not clear that it
1210 // occurs enough in real life to handle.
1214 if (!CondUse) return; // setcc doesn't use the IV.
1216 // It's possible for the setcc instruction to be anywhere in the loop, and
1217 // possible for it to have multiple users. If it is not immediately before
1218 // the latch block branch, move it.
1219 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1220 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1221 Cond->moveBefore(TermBr);
1223 // Otherwise, clone the terminating condition and insert into the loopend.
1224 Cond = cast<ICmpInst>(Cond->clone());
1225 Cond->setName(L->getHeader()->getName() + ".termcond");
1226 LatchBlock->getInstList().insert(TermBr, Cond);
1228 // Clone the IVUse, as the old use still exists!
1229 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1230 CondUse->OperandValToReplace);
1231 CondUse = &IVUsesByStride[*CondStride].Users.back();
1235 // If we get to here, we know that we can transform the setcc instruction to
1236 // use the post-incremented version of the IV, allowing us to coalesce the
1237 // live ranges for the IV correctly.
1238 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1239 CondUse->isUseOfPostIncrementedValue = true;
1243 // Constant strides come first which in turns are sorted by their absolute
1244 // values. If absolute values are the same, then positive strides comes first.
1246 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1247 struct StrideCompare {
1248 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1249 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1250 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1252 int64_t LV = LHSC->getValue()->getSExtValue();
1253 int64_t RV = RHSC->getValue()->getSExtValue();
1254 uint64_t ALV = (LV < 0) ? -LV : LV;
1255 uint64_t ARV = (RV < 0) ? -RV : RV;
1261 return (LHSC && !RHSC);
1266 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1268 LI = &getAnalysis<LoopInfo>();
1269 EF = &getAnalysis<ETForest>();
1270 SE = &getAnalysis<ScalarEvolution>();
1271 TD = &getAnalysis<TargetData>();
1272 UIntPtrTy = TD->getIntPtrType();
1274 // Find all uses of induction variables in this loop, and catagorize
1275 // them by stride. Start by finding all of the PHI nodes in the header for
1276 // this loop. If they are induction variables, inspect their uses.
1277 std::set<Instruction*> Processed; // Don't reprocess instructions.
1278 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1279 AddUsersIfInteresting(I, L, Processed);
1281 // If we have nothing to do, return.
1282 if (IVUsesByStride.empty()) return false;
1284 // Optimize induction variables. Some indvar uses can be transformed to use
1285 // strides that will be needed for other purposes. A common example of this
1286 // is the exit test for the loop, which can often be rewritten to use the
1287 // computation of some other indvar to decide when to terminate the loop.
1291 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1292 // doing computation in byte values, promote to 32-bit values if safe.
1294 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1295 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1296 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1297 // to be careful that IV's are all the same type. Only works for intptr_t
1300 // If we only have one stride, we can more aggressively eliminate some things.
1301 bool HasOneStride = IVUsesByStride.size() == 1;
1304 DOUT << "\nLSR on ";
1308 // IVsByStride keeps IVs for one particular loop.
1309 IVsByStride.clear();
1311 // Sort the StrideOrder so we process larger strides first.
1312 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1314 // Note: this processes each stride/type pair individually. All users passed
1315 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1316 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1317 // This extra layer of indirection makes the ordering of strides deterministic
1318 // - not dependent on map order.
1319 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1320 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1321 IVUsesByStride.find(StrideOrder[Stride]);
1322 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1323 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1326 // Clean up after ourselves
1327 if (!DeadInsts.empty()) {
1328 DeleteTriviallyDeadInstructions(DeadInsts);
1330 BasicBlock::iterator I = L->getHeader()->begin();
1332 while ((PN = dyn_cast<PHINode>(I))) {
1333 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1335 // At this point, we know that we have killed one or more GEP
1336 // instructions. It is worth checking to see if the cann indvar is also
1337 // dead, so that we can remove it as well. The requirements for the cann
1338 // indvar to be considered dead are:
1339 // 1. the cann indvar has one use
1340 // 2. the use is an add instruction
1341 // 3. the add has one use
1342 // 4. the add is used by the cann indvar
1343 // If all four cases above are true, then we can remove both the add and
1345 // FIXME: this needs to eliminate an induction variable even if it's being
1346 // compared against some value to decide loop termination.
1347 if (PN->hasOneUse()) {
1348 Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
1349 if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
1350 if (BO->hasOneUse() && PN == *(BO->use_begin())) {
1351 DeadInsts.insert(BO);
1352 // Break the cycle, then delete the PHI.
1353 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1354 SE->deleteInstructionFromRecords(PN);
1355 PN->eraseFromParent();
1360 DeleteTriviallyDeadInstructions(DeadInsts);
1363 CastedPointers.clear();
1364 IVUsesByStride.clear();
1365 StrideOrder.clear();