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");
49 /// IVStrideUse - Keep track of one use of a strided induction variable, where
50 /// the stride is stored externally. The Offset member keeps track of the
51 /// offset from the IV, User is the actual user of the operand, and 'Operand'
52 /// is the operand # of the User that is the use.
53 struct VISIBILITY_HIDDEN IVStrideUse {
56 Value *OperandValToReplace;
58 // isUseOfPostIncrementedValue - True if this should use the
59 // post-incremented version of this IV, not the preincremented version.
60 // This can only be set in special cases, such as the terminating setcc
61 // instruction for a loop or uses dominated by the loop.
62 bool isUseOfPostIncrementedValue;
64 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
65 : Offset(Offs), User(U), OperandValToReplace(O),
66 isUseOfPostIncrementedValue(false) {}
69 /// IVUsersOfOneStride - This structure keeps track of all instructions that
70 /// have an operand that is based on the trip count multiplied by some stride.
71 /// The stride for all of these users is common and kept external to this
73 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
74 /// Users - Keep track of all of the users of this stride as well as the
75 /// initial value and the operand that uses the IV.
76 std::vector<IVStrideUse> Users;
78 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
79 Users.push_back(IVStrideUse(Offset, User, Operand));
83 /// IVInfo - This structure keeps track of one IV expression inserted during
84 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
85 /// well as the PHI node and increment value created for rewrite.
86 struct VISIBILITY_HIDDEN IVExpr {
93 : Stride(SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)),
94 Base (SCEVUnknown::getIntegerSCEV(0, Type::Int32Ty)) {}
95 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
97 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
100 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
101 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
102 struct VISIBILITY_HIDDEN IVsOfOneStride {
103 std::vector<IVExpr> IVs;
105 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
107 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
111 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
115 const TargetData *TD;
116 const Type *UIntPtrTy;
119 /// IVUsesByStride - Keep track of all uses of induction variables that we
120 /// are interested in. The key of the map is the stride of the access.
121 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
123 /// IVsByStride - Keep track of all IVs that have been inserted for a
124 /// particular stride.
125 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
127 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
128 /// We use this to iterate over the IVUsesByStride collection without being
129 /// dependent on random ordering of pointers in the process.
130 std::vector<SCEVHandle> StrideOrder;
132 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
133 /// of the casted version of each value. This is accessed by
134 /// getCastedVersionOf.
135 std::map<Value*, Value*> CastedPointers;
137 /// DeadInsts - Keep track of instructions we may have made dead, so that
138 /// we can remove them after we are done working.
139 std::set<Instruction*> DeadInsts;
141 /// TLI - Keep a pointer of a TargetLowering to consult for determining
142 /// transformation profitability.
143 const TargetLowering *TLI;
146 LoopStrengthReduce(const TargetLowering *tli = NULL) : TLI(tli) {
149 bool runOnLoop(Loop *L, LPPassManager &LPM);
151 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
152 // We split critical edges, so we change the CFG. However, we do update
153 // many analyses if they are around.
154 AU.addPreservedID(LoopSimplifyID);
155 AU.addPreserved<LoopInfo>();
156 AU.addPreserved<ETForest>();
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);
176 bool FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
177 const SCEVHandle *&CondStride);
179 unsigned CheckForIVReuse(const SCEVHandle&, IVExpr&, const Type*,
180 const std::vector<BasedUser>& UsersToProcess);
182 bool ValidStride(int64_t, const std::vector<BasedUser>& UsersToProcess);
184 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
185 IVUsersOfOneStride &Uses,
186 Loop *L, bool isOnlyStride);
187 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
189 RegisterPass<LoopStrengthReduce> X("loop-reduce", "Loop Strength Reduction");
192 LoopPass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
193 return new LoopStrengthReduce(TLI);
196 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
197 /// assumes that the Value* V is of integer or pointer type only.
199 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
201 if (V->getType() == UIntPtrTy) return V;
202 if (Constant *CB = dyn_cast<Constant>(V))
203 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
205 Value *&New = CastedPointers[V];
208 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
209 DeadInsts.insert(cast<Instruction>(New));
214 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
215 /// specified set are trivially dead, delete them and see if this makes any of
216 /// their operands subsequently dead.
217 void LoopStrengthReduce::
218 DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
219 while (!Insts.empty()) {
220 Instruction *I = *Insts.begin();
221 Insts.erase(Insts.begin());
222 if (isInstructionTriviallyDead(I)) {
223 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
224 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
226 SE->deleteInstructionFromRecords(I);
227 I->eraseFromParent();
234 /// GetExpressionSCEV - Compute and return the SCEV for the specified
236 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp, Loop *L) {
237 // Pointer to pointer bitcast instructions return the same value as their
239 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
240 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
241 return SE->getSCEV(BCI);
242 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)), L);
247 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
248 // If this is a GEP that SE doesn't know about, compute it now and insert it.
249 // If this is not a GEP, or if we have already done this computation, just let
251 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
252 if (!GEP || SE->hasSCEV(GEP))
253 return SE->getSCEV(Exp);
255 // Analyze all of the subscripts of this getelementptr instruction, looking
256 // for uses that are determined by the trip count of L. First, skip all
257 // operands the are not dependent on the IV.
259 // Build up the base expression. Insert an LLVM cast of the pointer to
261 SCEVHandle GEPVal = SCEVUnknown::get(
262 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
264 gep_type_iterator GTI = gep_type_begin(GEP);
266 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
267 // If this is a use of a recurrence that we can analyze, and it comes before
268 // Op does in the GEP operand list, we will handle this when we process this
270 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
271 const StructLayout *SL = TD->getStructLayout(STy);
272 unsigned Idx = cast<ConstantInt>(GEP->getOperand(i))->getZExtValue();
273 uint64_t Offset = SL->getElementOffset(Idx);
274 GEPVal = SCEVAddExpr::get(GEPVal,
275 SCEVUnknown::getIntegerSCEV(Offset, UIntPtrTy));
277 unsigned GEPOpiBits =
278 GEP->getOperand(i)->getType()->getPrimitiveSizeInBits();
279 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
280 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
281 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
282 Instruction::BitCast));
283 Value *OpVal = getCastedVersionOf(opcode, GEP->getOperand(i));
284 SCEVHandle Idx = SE->getSCEV(OpVal);
286 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType());
288 Idx = SCEVMulExpr::get(Idx,
289 SCEVConstant::get(ConstantInt::get(UIntPtrTy,
291 GEPVal = SCEVAddExpr::get(GEPVal, Idx);
295 SE->setSCEV(GEP, GEPVal);
299 /// getSCEVStartAndStride - Compute the start and stride of this expression,
300 /// returning false if the expression is not a start/stride pair, or true if it
301 /// is. The stride must be a loop invariant expression, but the start may be
302 /// a mix of loop invariant and loop variant expressions.
303 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
304 SCEVHandle &Start, SCEVHandle &Stride) {
305 SCEVHandle TheAddRec = Start; // Initialize to zero.
307 // If the outer level is an AddExpr, the operands are all start values except
308 // for a nested AddRecExpr.
309 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
310 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
311 if (SCEVAddRecExpr *AddRec =
312 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
313 if (AddRec->getLoop() == L)
314 TheAddRec = SCEVAddExpr::get(AddRec, TheAddRec);
316 return false; // Nested IV of some sort?
318 Start = SCEVAddExpr::get(Start, AE->getOperand(i));
321 } else if (isa<SCEVAddRecExpr>(SH)) {
324 return false; // not analyzable.
327 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
328 if (!AddRec || AddRec->getLoop() != L) return false;
330 // FIXME: Generalize to non-affine IV's.
331 if (!AddRec->isAffine()) return false;
333 Start = SCEVAddExpr::get(Start, AddRec->getOperand(0));
335 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
336 DOUT << "[" << L->getHeader()->getName()
337 << "] Variable stride: " << *AddRec << "\n";
339 Stride = AddRec->getOperand(1);
343 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
344 /// and now we need to decide whether the user should use the preinc or post-inc
345 /// value. If this user should use the post-inc version of the IV, return true.
347 /// Choosing wrong here can break dominance properties (if we choose to use the
348 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
349 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
350 /// should use the post-inc value).
351 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
352 Loop *L, ETForest *EF, Pass *P) {
353 // If the user is in the loop, use the preinc value.
354 if (L->contains(User->getParent())) return false;
356 BasicBlock *LatchBlock = L->getLoopLatch();
358 // Ok, the user is outside of the loop. If it is dominated by the latch
359 // block, use the post-inc value.
360 if (EF->dominates(LatchBlock, User->getParent()))
363 // There is one case we have to be careful of: PHI nodes. These little guys
364 // can live in blocks that do not dominate the latch block, but (since their
365 // uses occur in the predecessor block, not the block the PHI lives in) should
366 // still use the post-inc value. Check for this case now.
367 PHINode *PN = dyn_cast<PHINode>(User);
368 if (!PN) return false; // not a phi, not dominated by latch block.
370 // Look at all of the uses of IV by the PHI node. If any use corresponds to
371 // a block that is not dominated by the latch block, give up and use the
372 // preincremented value.
373 unsigned NumUses = 0;
374 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
375 if (PN->getIncomingValue(i) == IV) {
377 if (!EF->dominates(LatchBlock, PN->getIncomingBlock(i)))
381 // Okay, all uses of IV by PN are in predecessor blocks that really are
382 // dominated by the latch block. Split the critical edges and use the
383 // post-incremented value.
384 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
385 if (PN->getIncomingValue(i) == IV) {
386 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P,
388 // Splitting the critical edge can reduce the number of entries in this
390 e = PN->getNumIncomingValues();
391 if (--NumUses == 0) break;
399 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
400 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
401 /// return true. Otherwise, return false.
402 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
403 std::set<Instruction*> &Processed) {
404 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
405 return false; // Void and FP expressions cannot be reduced.
406 if (!Processed.insert(I).second)
407 return true; // Instruction already handled.
409 // Get the symbolic expression for this instruction.
410 SCEVHandle ISE = GetExpressionSCEV(I, L);
411 if (isa<SCEVCouldNotCompute>(ISE)) return false;
413 // Get the start and stride for this expression.
414 SCEVHandle Start = SCEVUnknown::getIntegerSCEV(0, ISE->getType());
415 SCEVHandle Stride = Start;
416 if (!getSCEVStartAndStride(ISE, L, Start, Stride))
417 return false; // Non-reducible symbolic expression, bail out.
419 std::vector<Instruction *> IUsers;
420 // Collect all I uses now because IVUseShouldUsePostIncValue may
421 // invalidate use_iterator.
422 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
423 IUsers.push_back(cast<Instruction>(*UI));
425 for (unsigned iused_index = 0, iused_size = IUsers.size();
426 iused_index != iused_size; ++iused_index) {
428 Instruction *User = IUsers[iused_index];
430 // Do not infinitely recurse on PHI nodes.
431 if (isa<PHINode>(User) && Processed.count(User))
434 // If this is an instruction defined in a nested loop, or outside this loop,
435 // don't recurse into it.
436 bool AddUserToIVUsers = false;
437 if (LI->getLoopFor(User->getParent()) != L) {
438 DOUT << "FOUND USER in other loop: " << *User
439 << " OF SCEV: " << *ISE << "\n";
440 AddUserToIVUsers = true;
441 } else if (!AddUsersIfInteresting(User, L, Processed)) {
442 DOUT << "FOUND USER: " << *User
443 << " OF SCEV: " << *ISE << "\n";
444 AddUserToIVUsers = true;
447 if (AddUserToIVUsers) {
448 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
449 if (StrideUses.Users.empty()) // First occurance of this stride?
450 StrideOrder.push_back(Stride);
452 // Okay, we found a user that we cannot reduce. Analyze the instruction
453 // and decide what to do with it. If we are a use inside of the loop, use
454 // the value before incrementation, otherwise use it after incrementation.
455 if (IVUseShouldUsePostIncValue(User, I, L, EF, this)) {
456 // The value used will be incremented by the stride more than we are
457 // expecting, so subtract this off.
458 SCEVHandle NewStart = SCEV::getMinusSCEV(Start, Stride);
459 StrideUses.addUser(NewStart, User, I);
460 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
461 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
463 StrideUses.addUser(Start, User, I);
471 /// BasedUser - For a particular base value, keep information about how we've
472 /// partitioned the expression so far.
474 /// Base - The Base value for the PHI node that needs to be inserted for
475 /// this use. As the use is processed, information gets moved from this
476 /// field to the Imm field (below). BasedUser values are sorted by this
480 /// Inst - The instruction using the induction variable.
483 /// OperandValToReplace - The operand value of Inst to replace with the
485 Value *OperandValToReplace;
487 /// Imm - The immediate value that should be added to the base immediately
488 /// before Inst, because it will be folded into the imm field of the
492 /// EmittedBase - The actual value* to use for the base value of this
493 /// operation. This is null if we should just use zero so far.
496 // isUseOfPostIncrementedValue - True if this should use the
497 // post-incremented version of this IV, not the preincremented version.
498 // This can only be set in special cases, such as the terminating setcc
499 // instruction for a loop and uses outside the loop that are dominated by
501 bool isUseOfPostIncrementedValue;
503 BasedUser(IVStrideUse &IVSU)
504 : Base(IVSU.Offset), Inst(IVSU.User),
505 OperandValToReplace(IVSU.OperandValToReplace),
506 Imm(SCEVUnknown::getIntegerSCEV(0, Base->getType())), EmittedBase(0),
507 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
509 // Once we rewrite the code to insert the new IVs we want, update the
510 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
512 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
513 SCEVExpander &Rewriter, Loop *L,
516 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
517 SCEVExpander &Rewriter,
518 Instruction *IP, Loop *L);
523 void BasedUser::dump() const {
524 cerr << " Base=" << *Base;
525 cerr << " Imm=" << *Imm;
527 cerr << " EB=" << *EmittedBase;
529 cerr << " Inst: " << *Inst;
532 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
533 SCEVExpander &Rewriter,
534 Instruction *IP, Loop *L) {
535 // Figure out where we *really* want to insert this code. In particular, if
536 // the user is inside of a loop that is nested inside of L, we really don't
537 // want to insert this expression before the user, we'd rather pull it out as
538 // many loops as possible.
539 LoopInfo &LI = Rewriter.getLoopInfo();
540 Instruction *BaseInsertPt = IP;
542 // Figure out the most-nested loop that IP is in.
543 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
545 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
546 // the preheader of the outer-most loop where NewBase is not loop invariant.
547 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
548 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
549 InsertLoop = InsertLoop->getParentLoop();
552 // If there is no immediate value, skip the next part.
553 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm))
554 if (SC->getValue()->isZero())
555 return Rewriter.expandCodeFor(NewBase, BaseInsertPt,
556 OperandValToReplace->getType());
558 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
560 // Always emit the immediate (if non-zero) into the same block as the user.
561 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(Base), Imm);
562 return Rewriter.expandCodeFor(NewValSCEV, IP,
563 OperandValToReplace->getType());
567 // Once we rewrite the code to insert the new IVs we want, update the
568 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
570 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
571 SCEVExpander &Rewriter,
573 if (!isa<PHINode>(Inst)) {
574 // By default, insert code at the user instruction.
575 BasicBlock::iterator InsertPt = Inst;
577 // However, if the Operand is itself an instruction, the (potentially
578 // complex) inserted code may be shared by many users. Because of this, we
579 // want to emit code for the computation of the operand right before its old
580 // computation. This is usually safe, because we obviously used to use the
581 // computation when it was computed in its current block. However, in some
582 // cases (e.g. use of a post-incremented induction variable) the NewBase
583 // value will be pinned to live somewhere after the original computation.
584 // In this case, we have to back off.
585 if (!isUseOfPostIncrementedValue) {
586 if (Instruction *OpInst = dyn_cast<Instruction>(OperandValToReplace)) {
588 while (isa<PHINode>(InsertPt)) ++InsertPt;
592 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
593 // Replace the use of the operand Value with the new Phi we just created.
594 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
595 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
599 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
600 // expression into each operand block that uses it. Note that PHI nodes can
601 // have multiple entries for the same predecessor. We use a map to make sure
602 // that a PHI node only has a single Value* for each predecessor (which also
603 // prevents us from inserting duplicate code in some blocks).
604 std::map<BasicBlock*, Value*> InsertedCode;
605 PHINode *PN = cast<PHINode>(Inst);
606 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
607 if (PN->getIncomingValue(i) == OperandValToReplace) {
608 // If this is a critical edge, split the edge so that we do not insert the
609 // code on all predecessor/successor paths. We do this unless this is the
610 // canonical backedge for this loop, as this can make some inserted code
611 // be in an illegal position.
612 BasicBlock *PHIPred = PN->getIncomingBlock(i);
613 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
614 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
616 // First step, split the critical edge.
617 SplitCriticalEdge(PHIPred, PN->getParent(), P, true);
619 // Next step: move the basic block. In particular, if the PHI node
620 // is outside of the loop, and PredTI is in the loop, we want to
621 // move the block to be immediately before the PHI block, not
622 // immediately after PredTI.
623 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
624 BasicBlock *NewBB = PN->getIncomingBlock(i);
625 NewBB->moveBefore(PN->getParent());
628 // Splitting the edge can reduce the number of PHI entries we have.
629 e = PN->getNumIncomingValues();
632 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
634 // Insert the code into the end of the predecessor block.
635 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
636 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
639 // Replace the use of the operand Value with the new Phi we just created.
640 PN->setIncomingValue(i, Code);
644 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
648 /// isTargetConstant - Return true if the following can be referenced by the
649 /// immediate field of a target instruction.
650 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
651 const TargetLowering *TLI) {
652 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
653 int64_t VC = SC->getValue()->getSExtValue();
655 TargetLowering::AddrMode AM;
657 return TLI->isLegalAddressingMode(AM, UseTy);
659 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
660 return (VC > -(1 << 16) && VC < (1 << 16)-1);
664 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
665 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
666 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
667 Constant *Op0 = CE->getOperand(0);
668 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
669 TargetLowering::AddrMode AM;
671 return TLI->isLegalAddressingMode(AM, UseTy);
677 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
678 /// loop varying to the Imm operand.
679 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
681 if (Val->isLoopInvariant(L)) return; // Nothing to do.
683 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
684 std::vector<SCEVHandle> NewOps;
685 NewOps.reserve(SAE->getNumOperands());
687 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
688 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
689 // If this is a loop-variant expression, it must stay in the immediate
690 // field of the expression.
691 Imm = SCEVAddExpr::get(Imm, SAE->getOperand(i));
693 NewOps.push_back(SAE->getOperand(i));
697 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
699 Val = SCEVAddExpr::get(NewOps);
700 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
701 // Try to pull immediates out of the start value of nested addrec's.
702 SCEVHandle Start = SARE->getStart();
703 MoveLoopVariantsToImediateField(Start, Imm, L);
705 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
707 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
709 // Otherwise, all of Val is variant, move the whole thing over.
710 Imm = SCEVAddExpr::get(Imm, Val);
711 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
716 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
717 /// that can fit into the immediate field of instructions in the target.
718 /// Accumulate these immediate values into the Imm value.
719 static void MoveImmediateValues(const TargetLowering *TLI,
721 SCEVHandle &Val, SCEVHandle &Imm,
722 bool isAddress, Loop *L) {
723 const Type *UseTy = User->getType();
724 if (StoreInst *SI = dyn_cast<StoreInst>(User))
725 UseTy = SI->getOperand(0)->getType();
727 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
728 std::vector<SCEVHandle> NewOps;
729 NewOps.reserve(SAE->getNumOperands());
731 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
732 SCEVHandle NewOp = SAE->getOperand(i);
733 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L);
735 if (!NewOp->isLoopInvariant(L)) {
736 // If this is a loop-variant expression, it must stay in the immediate
737 // field of the expression.
738 Imm = SCEVAddExpr::get(Imm, NewOp);
740 NewOps.push_back(NewOp);
745 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
747 Val = SCEVAddExpr::get(NewOps);
749 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
750 // Try to pull immediates out of the start value of nested addrec's.
751 SCEVHandle Start = SARE->getStart();
752 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L);
754 if (Start != SARE->getStart()) {
755 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
757 Val = SCEVAddRecExpr::get(Ops, SARE->getLoop());
760 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
761 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
762 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
763 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
765 SCEVHandle SubImm = SCEVUnknown::getIntegerSCEV(0, Val->getType());
766 SCEVHandle NewOp = SME->getOperand(1);
767 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L);
769 // If we extracted something out of the subexpressions, see if we can
771 if (NewOp != SME->getOperand(1)) {
772 // Scale SubImm up by "8". If the result is a target constant, we are
774 SubImm = SCEVMulExpr::get(SubImm, SME->getOperand(0));
775 if (isTargetConstant(SubImm, UseTy, TLI)) {
776 // Accumulate the immediate.
777 Imm = SCEVAddExpr::get(Imm, SubImm);
779 // Update what is left of 'Val'.
780 Val = SCEVMulExpr::get(SME->getOperand(0), NewOp);
787 // Loop-variant expressions must stay in the immediate field of the
789 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
790 !Val->isLoopInvariant(L)) {
791 Imm = SCEVAddExpr::get(Imm, Val);
792 Val = SCEVUnknown::getIntegerSCEV(0, Val->getType());
796 // Otherwise, no immediates to move.
800 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
801 /// added together. This is used to reassociate common addition subexprs
802 /// together for maximal sharing when rewriting bases.
803 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
805 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
806 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
807 SeparateSubExprs(SubExprs, AE->getOperand(j));
808 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
809 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Expr->getType());
810 if (SARE->getOperand(0) == Zero) {
811 SubExprs.push_back(Expr);
813 // Compute the addrec with zero as its base.
814 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
815 Ops[0] = Zero; // Start with zero base.
816 SubExprs.push_back(SCEVAddRecExpr::get(Ops, SARE->getLoop()));
819 SeparateSubExprs(SubExprs, SARE->getOperand(0));
821 } else if (!isa<SCEVConstant>(Expr) ||
822 !cast<SCEVConstant>(Expr)->getValue()->isZero()) {
824 SubExprs.push_back(Expr);
829 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
830 /// removing any common subexpressions from it. Anything truly common is
831 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
832 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
834 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses) {
835 unsigned NumUses = Uses.size();
837 // Only one use? Use its base, regardless of what it is!
838 SCEVHandle Zero = SCEVUnknown::getIntegerSCEV(0, Uses[0].Base->getType());
839 SCEVHandle Result = Zero;
841 std::swap(Result, Uses[0].Base);
845 // To find common subexpressions, count how many of Uses use each expression.
846 // If any subexpressions are used Uses.size() times, they are common.
847 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
849 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
850 // order we see them.
851 std::vector<SCEVHandle> UniqueSubExprs;
853 std::vector<SCEVHandle> SubExprs;
854 for (unsigned i = 0; i != NumUses; ++i) {
855 // If the base is zero (which is common), return zero now, there are no
857 if (Uses[i].Base == Zero) return Zero;
859 // Split the expression into subexprs.
860 SeparateSubExprs(SubExprs, Uses[i].Base);
861 // Add one to SubExpressionUseCounts for each subexpr present.
862 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
863 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
864 UniqueSubExprs.push_back(SubExprs[j]);
868 // Now that we know how many times each is used, build Result. Iterate over
869 // UniqueSubexprs so that we have a stable ordering.
870 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
871 std::map<SCEVHandle, unsigned>::iterator I =
872 SubExpressionUseCounts.find(UniqueSubExprs[i]);
873 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
874 if (I->second == NumUses) { // Found CSE!
875 Result = SCEVAddExpr::get(Result, I->first);
877 // Remove non-cse's from SubExpressionUseCounts.
878 SubExpressionUseCounts.erase(I);
882 // If we found no CSE's, return now.
883 if (Result == Zero) return Result;
885 // Otherwise, remove all of the CSE's we found from each of the base values.
886 for (unsigned i = 0; i != NumUses; ++i) {
887 // Split the expression into subexprs.
888 SeparateSubExprs(SubExprs, Uses[i].Base);
890 // Remove any common subexpressions.
891 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
892 if (SubExpressionUseCounts.count(SubExprs[j])) {
893 SubExprs.erase(SubExprs.begin()+j);
897 // Finally, the non-shared expressions together.
898 if (SubExprs.empty())
901 Uses[i].Base = SCEVAddExpr::get(SubExprs);
908 /// isZero - returns true if the scalar evolution expression is zero.
910 static bool isZero(SCEVHandle &V) {
911 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V))
912 return SC->getValue()->isZero();
916 /// ValidStride - Check whether the given Scale is valid for all loads and
917 /// stores in UsersToProcess.
919 bool LoopStrengthReduce::ValidStride(int64_t Scale,
920 const std::vector<BasedUser>& UsersToProcess) {
921 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
922 // If this is a load or other access, pass the type of the access in.
923 const Type *AccessTy = Type::VoidTy;
924 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
925 AccessTy = SI->getOperand(0)->getType();
926 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
927 AccessTy = LI->getType();
929 TargetLowering::AddrMode AM;
930 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
931 AM.BaseOffs = SC->getValue()->getSExtValue();
934 // If load[imm+r*scale] is illegal, bail out.
935 if (!TLI->isLegalAddressingMode(AM, AccessTy))
941 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
942 /// of a previous stride and it is a legal value for the target addressing
943 /// mode scale component. This allows the users of this stride to be rewritten
944 /// as prev iv * factor. It returns 0 if no reuse is possible.
945 unsigned LoopStrengthReduce::CheckForIVReuse(const SCEVHandle &Stride,
946 IVExpr &IV, const Type *Ty,
947 const std::vector<BasedUser>& UsersToProcess) {
950 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
951 int64_t SInt = SC->getValue()->getSExtValue();
952 if (SInt == 1) return 0;
954 for (std::map<SCEVHandle, IVsOfOneStride>::iterator SI= IVsByStride.begin(),
955 SE = IVsByStride.end(); SI != SE; ++SI) {
956 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
957 if (SInt != -SSInt &&
958 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
960 int64_t Scale = SInt / SSInt;
961 // Check that this stride is valid for all the types used for loads and
962 // stores; if it can be used for some and not others, we might as well use
963 // the original stride everywhere, since we have to create the IV for it
965 if (ValidStride(Scale, UsersToProcess))
966 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
967 IE = SI->second.IVs.end(); II != IE; ++II)
968 // FIXME: Only handle base == 0 for now.
969 // Only reuse previous IV if it would not require a type conversion.
970 if (isZero(II->Base) && II->Base->getType() == Ty) {
979 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
980 /// returns true if Val's isUseOfPostIncrementedValue is true.
981 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
982 return Val.isUseOfPostIncrementedValue;
985 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
986 /// stride of IV. All of the users may have different starting values, and this
987 /// may not be the only stride (we know it is if isOnlyStride is true).
988 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
989 IVUsersOfOneStride &Uses,
992 // Transform our list of users and offsets to a bit more complex table. In
993 // this new vector, each 'BasedUser' contains 'Base' the base of the
994 // strided accessas well as the old information from Uses. We progressively
995 // move information from the Base field to the Imm field, until we eventually
996 // have the full access expression to rewrite the use.
997 std::vector<BasedUser> UsersToProcess;
998 UsersToProcess.reserve(Uses.Users.size());
999 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1000 UsersToProcess.push_back(Uses.Users[i]);
1002 // Move any loop invariant operands from the offset field to the immediate
1003 // field of the use, so that we don't try to use something before it is
1005 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1006 UsersToProcess.back().Imm, L);
1007 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1008 "Base value is not loop invariant!");
1011 // We now have a whole bunch of uses of like-strided induction variables, but
1012 // they might all have different bases. We want to emit one PHI node for this
1013 // stride which we fold as many common expressions (between the IVs) into as
1014 // possible. Start by identifying the common expressions in the base values
1015 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1016 // "A+B"), emit it to the preheader, then remove the expression from the
1017 // UsersToProcess base values.
1018 SCEVHandle CommonExprs =
1019 RemoveCommonExpressionsFromUseBases(UsersToProcess);
1021 // Next, figure out what we can represent in the immediate fields of
1022 // instructions. If we can represent anything there, move it to the imm
1023 // fields of the BasedUsers. We do this so that it increases the commonality
1024 // of the remaining uses.
1025 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1026 // If the user is not in the current loop, this means it is using the exit
1027 // value of the IV. Do not put anything in the base, make sure it's all in
1028 // the immediate field to allow as much factoring as possible.
1029 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1030 UsersToProcess[i].Imm = SCEVAddExpr::get(UsersToProcess[i].Imm,
1031 UsersToProcess[i].Base);
1032 UsersToProcess[i].Base =
1033 SCEVUnknown::getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1036 // Addressing modes can be folded into loads and stores. Be careful that
1037 // the store is through the expression, not of the expression though.
1038 bool isAddress = isa<LoadInst>(UsersToProcess[i].Inst);
1039 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1040 if (SI->getOperand(1) == UsersToProcess[i].OperandValToReplace)
1043 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1044 UsersToProcess[i].Imm, isAddress, L);
1048 // Check if it is possible to reuse a IV with stride that is factor of this
1049 // stride. And the multiple is a number that can be encoded in the scale
1050 // field of the target addressing mode. And we will have a valid
1051 // instruction after this substition, including the immediate field, if any.
1052 PHINode *NewPHI = NULL;
1055 unsigned RewriteFactor = CheckForIVReuse(Stride, ReuseIV,
1056 CommonExprs->getType(),
1058 if (RewriteFactor != 0) {
1059 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1060 << " and BASE " << *ReuseIV.Base << " :\n";
1061 NewPHI = ReuseIV.PHI;
1062 IncV = ReuseIV.IncV;
1065 const Type *ReplacedTy = CommonExprs->getType();
1067 // Now that we know what we need to do, insert the PHI node itself.
1069 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1070 << *Stride << " and BASE " << *CommonExprs << " :\n";
1072 SCEVExpander Rewriter(*SE, *LI);
1073 SCEVExpander PreheaderRewriter(*SE, *LI);
1075 BasicBlock *Preheader = L->getLoopPreheader();
1076 Instruction *PreInsertPt = Preheader->getTerminator();
1077 Instruction *PhiInsertBefore = L->getHeader()->begin();
1079 BasicBlock *LatchBlock = L->getLoopLatch();
1082 // Emit the initial base value into the loop preheader.
1084 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt,
1087 if (RewriteFactor == 0) {
1088 // Create a new Phi for this base, and stick it in the loop header.
1089 NewPHI = new PHINode(ReplacedTy, "iv.", PhiInsertBefore);
1092 // Add common base to the new Phi node.
1093 NewPHI->addIncoming(CommonBaseV, Preheader);
1095 // Insert the stride into the preheader.
1096 Value *StrideV = PreheaderRewriter.expandCodeFor(Stride, PreInsertPt,
1098 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1100 // Emit the increment of the base value before the terminator of the loop
1101 // latch block, and add it to the Phi node.
1102 SCEVHandle IncExp = SCEVAddExpr::get(SCEVUnknown::get(NewPHI),
1103 SCEVUnknown::get(StrideV));
1105 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator(),
1107 IncV->setName(NewPHI->getName()+".inc");
1108 NewPHI->addIncoming(IncV, LatchBlock);
1110 // Remember this in case a later stride is multiple of this.
1111 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1113 Constant *C = dyn_cast<Constant>(CommonBaseV);
1115 (!C->isNullValue() &&
1116 !isTargetConstant(SCEVUnknown::get(CommonBaseV), ReplacedTy, TLI)))
1117 // We want the common base emitted into the preheader! This is just
1118 // using cast as a copy so BitCast (no-op cast) is appropriate
1119 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1120 "commonbase", PreInsertPt);
1123 // We want to emit code for users inside the loop first. To do this, we
1124 // rearrange BasedUser so that the entries at the end have
1125 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1126 // vector (so we handle them first).
1127 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1128 PartitionByIsUseOfPostIncrementedValue);
1130 // Sort this by base, so that things with the same base are handled
1131 // together. By partitioning first and stable-sorting later, we are
1132 // guaranteed that within each base we will pop off users from within the
1133 // loop before users outside of the loop with a particular base.
1135 // We would like to use stable_sort here, but we can't. The problem is that
1136 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1137 // we don't have anything to do a '<' comparison on. Because we think the
1138 // number of uses is small, do a horrible bubble sort which just relies on
1140 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1141 // Get a base value.
1142 SCEVHandle Base = UsersToProcess[i].Base;
1144 // Compact everything with this base to be consequetive with this one.
1145 for (unsigned j = i+1; j != e; ++j) {
1146 if (UsersToProcess[j].Base == Base) {
1147 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1153 // Process all the users now. This outer loop handles all bases, the inner
1154 // loop handles all users of a particular base.
1155 while (!UsersToProcess.empty()) {
1156 SCEVHandle Base = UsersToProcess.back().Base;
1158 DOUT << " INSERTING code for BASE = " << *Base << ":\n";
1160 // Emit the code for Base into the preheader.
1161 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt,
1164 // If BaseV is a constant other than 0, make sure that it gets inserted into
1165 // the preheader, instead of being forward substituted into the uses. We do
1166 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1168 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1169 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1170 // We want this constant emitted into the preheader! This is just
1171 // using cast as a copy so BitCast (no-op cast) is appropriate
1172 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1177 // Emit the code to add the immediate offset to the Phi value, just before
1178 // the instructions that we identified as using this stride and base.
1180 // FIXME: Use emitted users to emit other users.
1181 BasedUser &User = UsersToProcess.back();
1183 // If this instruction wants to use the post-incremented value, move it
1184 // after the post-inc and use its value instead of the PHI.
1185 Value *RewriteOp = NewPHI;
1186 if (User.isUseOfPostIncrementedValue) {
1189 // If this user is in the loop, make sure it is the last thing in the
1190 // loop to ensure it is dominated by the increment.
1191 if (L->contains(User.Inst->getParent()))
1192 User.Inst->moveBefore(LatchBlock->getTerminator());
1194 if (RewriteOp->getType() != ReplacedTy) {
1195 Instruction::CastOps opcode = Instruction::Trunc;
1196 if (ReplacedTy->getPrimitiveSizeInBits() ==
1197 RewriteOp->getType()->getPrimitiveSizeInBits())
1198 opcode = Instruction::BitCast;
1199 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1202 SCEVHandle RewriteExpr = SCEVUnknown::get(RewriteOp);
1204 // Clear the SCEVExpander's expression map so that we are guaranteed
1205 // to have the code emitted where we expect it.
1208 // If we are reusing the iv, then it must be multiplied by a constant
1209 // factor take advantage of addressing mode scale component.
1210 if (RewriteFactor != 0) {
1212 SCEVMulExpr::get(SCEVUnknown::getIntegerSCEV(RewriteFactor,
1213 RewriteExpr->getType()),
1216 // The common base is emitted in the loop preheader. But since we
1217 // are reusing an IV, it has not been used to initialize the PHI node.
1218 // Add it to the expression used to rewrite the uses.
1219 if (!isa<ConstantInt>(CommonBaseV) ||
1220 !cast<ConstantInt>(CommonBaseV)->isZero())
1221 RewriteExpr = SCEVAddExpr::get(RewriteExpr,
1222 SCEVUnknown::get(CommonBaseV));
1225 // Now that we know what we need to do, insert code before User for the
1226 // immediate and any loop-variant expressions.
1227 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1228 // Add BaseV to the PHI value if needed.
1229 RewriteExpr = SCEVAddExpr::get(RewriteExpr, SCEVUnknown::get(BaseV));
1231 User.RewriteInstructionToUseNewBase(RewriteExpr, Rewriter, L, this);
1233 // Mark old value we replaced as possibly dead, so that it is elminated
1234 // if we just replaced the last use of that value.
1235 DeadInsts.insert(cast<Instruction>(User.OperandValToReplace));
1237 UsersToProcess.pop_back();
1240 // If there are any more users to process with the same base, process them
1241 // now. We sorted by base above, so we just have to check the last elt.
1242 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1243 // TODO: Next, find out which base index is the most common, pull it out.
1246 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1247 // different starting values, into different PHIs.
1250 /// FindIVForUser - If Cond has an operand that is an expression of an IV,
1251 /// set the IV user and stride information and return true, otherwise return
1253 bool LoopStrengthReduce::FindIVForUser(ICmpInst *Cond, IVStrideUse *&CondUse,
1254 const SCEVHandle *&CondStride) {
1255 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1257 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1258 IVUsesByStride.find(StrideOrder[Stride]);
1259 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1261 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1262 E = SI->second.Users.end(); UI != E; ++UI)
1263 if (UI->User == Cond) {
1264 // NOTE: we could handle setcc instructions with multiple uses here, but
1265 // InstCombine does it as well for simple uses, it's not clear that it
1266 // occurs enough in real life to handle.
1268 CondStride = &SI->first;
1275 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1276 // uses in the loop, look to see if we can eliminate some, in favor of using
1277 // common indvars for the different uses.
1278 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1279 // TODO: implement optzns here.
1281 // Finally, get the terminating condition for the loop if possible. If we
1282 // can, we want to change it to use a post-incremented version of its
1283 // induction variable, to allow coalescing the live ranges for the IV into
1284 // one register value.
1285 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1286 BasicBlock *Preheader = L->getLoopPreheader();
1287 BasicBlock *LatchBlock =
1288 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1289 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1290 if (!TermBr || TermBr->isUnconditional() ||
1291 !isa<ICmpInst>(TermBr->getCondition()))
1293 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1295 // Search IVUsesByStride to find Cond's IVUse if there is one.
1296 IVStrideUse *CondUse = 0;
1297 const SCEVHandle *CondStride = 0;
1299 if (!FindIVForUser(Cond, CondUse, CondStride))
1300 return; // setcc doesn't use the IV.
1303 // It's possible for the setcc instruction to be anywhere in the loop, and
1304 // possible for it to have multiple users. If it is not immediately before
1305 // the latch block branch, move it.
1306 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1307 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1308 Cond->moveBefore(TermBr);
1310 // Otherwise, clone the terminating condition and insert into the loopend.
1311 Cond = cast<ICmpInst>(Cond->clone());
1312 Cond->setName(L->getHeader()->getName() + ".termcond");
1313 LatchBlock->getInstList().insert(TermBr, Cond);
1315 // Clone the IVUse, as the old use still exists!
1316 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1317 CondUse->OperandValToReplace);
1318 CondUse = &IVUsesByStride[*CondStride].Users.back();
1322 // If we get to here, we know that we can transform the setcc instruction to
1323 // use the post-incremented version of the IV, allowing us to coalesce the
1324 // live ranges for the IV correctly.
1325 CondUse->Offset = SCEV::getMinusSCEV(CondUse->Offset, *CondStride);
1326 CondUse->isUseOfPostIncrementedValue = true;
1330 // Constant strides come first which in turns are sorted by their absolute
1331 // values. If absolute values are the same, then positive strides comes first.
1333 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1334 struct StrideCompare {
1335 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1336 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1337 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1339 int64_t LV = LHSC->getValue()->getSExtValue();
1340 int64_t RV = RHSC->getValue()->getSExtValue();
1341 uint64_t ALV = (LV < 0) ? -LV : LV;
1342 uint64_t ARV = (RV < 0) ? -RV : RV;
1348 return (LHSC && !RHSC);
1353 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
1355 LI = &getAnalysis<LoopInfo>();
1356 EF = &getAnalysis<ETForest>();
1357 SE = &getAnalysis<ScalarEvolution>();
1358 TD = &getAnalysis<TargetData>();
1359 UIntPtrTy = TD->getIntPtrType();
1361 // Find all uses of induction variables in this loop, and catagorize
1362 // them by stride. Start by finding all of the PHI nodes in the header for
1363 // this loop. If they are induction variables, inspect their uses.
1364 std::set<Instruction*> Processed; // Don't reprocess instructions.
1365 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
1366 AddUsersIfInteresting(I, L, Processed);
1368 // If we have nothing to do, return.
1369 if (IVUsesByStride.empty()) return false;
1371 // Optimize induction variables. Some indvar uses can be transformed to use
1372 // strides that will be needed for other purposes. A common example of this
1373 // is the exit test for the loop, which can often be rewritten to use the
1374 // computation of some other indvar to decide when to terminate the loop.
1378 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
1379 // doing computation in byte values, promote to 32-bit values if safe.
1381 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
1382 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be
1383 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need
1384 // to be careful that IV's are all the same type. Only works for intptr_t
1387 // If we only have one stride, we can more aggressively eliminate some things.
1388 bool HasOneStride = IVUsesByStride.size() == 1;
1391 DOUT << "\nLSR on ";
1395 // IVsByStride keeps IVs for one particular loop.
1396 IVsByStride.clear();
1398 // Sort the StrideOrder so we process larger strides first.
1399 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1401 // Note: this processes each stride/type pair individually. All users passed
1402 // into StrengthReduceStridedIVUsers have the same type AND stride. Also,
1403 // node that we iterate over IVUsesByStride indirectly by using StrideOrder.
1404 // This extra layer of indirection makes the ordering of strides deterministic
1405 // - not dependent on map order.
1406 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
1407 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1408 IVUsesByStride.find(StrideOrder[Stride]);
1409 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1410 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
1413 // Clean up after ourselves
1414 if (!DeadInsts.empty()) {
1415 DeleteTriviallyDeadInstructions(DeadInsts);
1417 BasicBlock::iterator I = L->getHeader()->begin();
1419 while ((PN = dyn_cast<PHINode>(I))) {
1420 ++I; // Preincrement iterator to avoid invalidating it when deleting PN.
1422 // At this point, we know that we have killed one or more GEP
1423 // instructions. It is worth checking to see if the cann indvar is also
1424 // dead, so that we can remove it as well. The requirements for the cann
1425 // indvar to be considered dead are:
1426 // 1. the cann indvar has one use
1427 // 2. the use is an add instruction
1428 // 3. the add has one use
1429 // 4. the add is used by the cann indvar
1430 // If all four cases above are true, then we can remove both the add and
1432 // FIXME: this needs to eliminate an induction variable even if it's being
1433 // compared against some value to decide loop termination.
1434 if (PN->hasOneUse()) {
1435 Instruction *BO = dyn_cast<Instruction>(*PN->use_begin());
1436 if (BO && (isa<BinaryOperator>(BO) || isa<CmpInst>(BO))) {
1437 if (BO->hasOneUse() && PN == *(BO->use_begin())) {
1438 DeadInsts.insert(BO);
1439 // Break the cycle, then delete the PHI.
1440 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
1441 SE->deleteInstructionFromRecords(PN);
1442 PN->eraseFromParent();
1447 DeleteTriviallyDeadInstructions(DeadInsts);
1450 CastedPointers.clear();
1451 IVUsesByStride.clear();
1452 StrideOrder.clear();