1 //===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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/IntrinsicInst.h"
23 #include "llvm/Type.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/ScalarEvolutionExpander.h"
29 #include "llvm/Support/CFG.h"
30 #include "llvm/Support/GetElementPtrTypeIterator.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/Target/TargetLowering.h"
43 STATISTIC(NumReduced , "Number of GEPs strength reduced");
44 STATISTIC(NumInserted, "Number of PHIs inserted");
45 STATISTIC(NumVariable, "Number of PHIs with variable strides");
46 STATISTIC(NumEliminated, "Number of strides eliminated");
47 STATISTIC(NumShadow, "Number of Shadow IVs optimized");
53 /// IVStrideUse - Keep track of one use of a strided induction variable, where
54 /// the stride is stored externally. The Offset member keeps track of the
55 /// offset from the IV, User is the actual user of the operand, and
56 /// 'OperandValToReplace' is the operand of the User that is the use.
57 struct VISIBILITY_HIDDEN IVStrideUse {
60 Value *OperandValToReplace;
62 // isUseOfPostIncrementedValue - True if this should use the
63 // post-incremented version of this IV, not the preincremented version.
64 // This can only be set in special cases, such as the terminating setcc
65 // instruction for a loop or uses dominated by the loop.
66 bool isUseOfPostIncrementedValue;
68 IVStrideUse(const SCEVHandle &Offs, Instruction *U, Value *O)
69 : Offset(Offs), User(U), OperandValToReplace(O),
70 isUseOfPostIncrementedValue(false) {}
73 /// IVUsersOfOneStride - This structure keeps track of all instructions that
74 /// have an operand that is based on the trip count multiplied by some stride.
75 /// The stride for all of these users is common and kept external to this
77 struct VISIBILITY_HIDDEN IVUsersOfOneStride {
78 /// Users - Keep track of all of the users of this stride as well as the
79 /// initial value and the operand that uses the IV.
80 std::vector<IVStrideUse> Users;
82 void addUser(const SCEVHandle &Offset,Instruction *User, Value *Operand) {
83 Users.push_back(IVStrideUse(Offset, User, Operand));
87 /// IVInfo - This structure keeps track of one IV expression inserted during
88 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as
89 /// well as the PHI node and increment value created for rewrite.
90 struct VISIBILITY_HIDDEN IVExpr {
96 IVExpr(const SCEVHandle &stride, const SCEVHandle &base, PHINode *phi,
98 : Stride(stride), Base(base), PHI(phi), IncV(incv) {}
101 /// IVsOfOneStride - This structure keeps track of all IV expression inserted
102 /// during StrengthReduceStridedIVUsers for a particular stride of the IV.
103 struct VISIBILITY_HIDDEN IVsOfOneStride {
104 std::vector<IVExpr> IVs;
106 void addIV(const SCEVHandle &Stride, const SCEVHandle &Base, PHINode *PHI,
108 IVs.push_back(IVExpr(Stride, Base, PHI, IncV));
112 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass {
116 const TargetData *TD;
117 const Type *UIntPtrTy;
120 /// IVUsesByStride - Keep track of all uses of induction variables that we
121 /// are interested in. The key of the map is the stride of the access.
122 std::map<SCEVHandle, IVUsersOfOneStride> IVUsesByStride;
124 /// IVsByStride - Keep track of all IVs that have been inserted for a
125 /// particular stride.
126 std::map<SCEVHandle, IVsOfOneStride> IVsByStride;
128 /// StrideOrder - An ordering of the keys in IVUsesByStride that is stable:
129 /// We use this to iterate over the IVUsesByStride collection without being
130 /// dependent on random ordering of pointers in the process.
131 SmallVector<SCEVHandle, 16> StrideOrder;
133 /// CastedValues - As we need to cast values to uintptr_t, this keeps track
134 /// of the casted version of each value. This is accessed by
135 /// getCastedVersionOf.
136 DenseMap<Value*, Value*> CastedPointers;
138 /// DeadInsts - Keep track of instructions we may have made dead, so that
139 /// we can remove them after we are done working.
140 SmallVector<Instruction*, 16> DeadInsts;
142 /// TLI - Keep a pointer of a TargetLowering to consult for determining
143 /// transformation profitability.
144 const TargetLowering *TLI;
147 static char ID; // Pass ID, replacement for typeid
148 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) :
149 LoopPass(&ID), TLI(tli) {
152 bool runOnLoop(Loop *L, LPPassManager &LPM);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 // We split critical edges, so we change the CFG. However, we do update
156 // many analyses if they are around.
157 AU.addPreservedID(LoopSimplifyID);
158 AU.addPreserved<LoopInfo>();
159 AU.addPreserved<DominanceFrontier>();
160 AU.addPreserved<DominatorTree>();
162 AU.addRequiredID(LoopSimplifyID);
163 AU.addRequired<LoopInfo>();
164 AU.addRequired<DominatorTree>();
165 AU.addRequired<TargetData>();
166 AU.addRequired<ScalarEvolution>();
167 AU.addPreserved<ScalarEvolution>();
170 /// getCastedVersionOf - Return the specified value casted to uintptr_t.
172 Value *getCastedVersionOf(Instruction::CastOps opcode, Value *V);
174 bool AddUsersIfInteresting(Instruction *I, Loop *L,
175 SmallPtrSet<Instruction*,16> &Processed);
176 SCEVHandle GetExpressionSCEV(Instruction *E);
177 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond,
178 IVStrideUse* &CondUse,
179 const SCEVHandle* &CondStride);
180 void OptimizeIndvars(Loop *L);
182 /// OptimizeShadowIV - If IV is used in a int-to-float cast
183 /// inside the loop then try to eliminate the cast opeation.
184 void OptimizeShadowIV(Loop *L);
186 /// OptimizeSMax - Rewrite the loop's terminating condition
187 /// if it uses an smax computation.
188 ICmpInst *OptimizeSMax(Loop *L, ICmpInst *Cond,
189 IVStrideUse* &CondUse);
191 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
192 const SCEVHandle *&CondStride);
193 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy);
194 unsigned CheckForIVReuse(bool, bool, const SCEVHandle&,
195 IVExpr&, const Type*,
196 const std::vector<BasedUser>& UsersToProcess);
197 bool ValidStride(bool, int64_t,
198 const std::vector<BasedUser>& UsersToProcess);
199 SCEVHandle CollectIVUsers(const SCEVHandle &Stride,
200 IVUsersOfOneStride &Uses,
202 bool &AllUsesAreAddresses,
203 std::vector<BasedUser> &UsersToProcess);
204 void StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
205 IVUsersOfOneStride &Uses,
206 Loop *L, bool isOnlyStride);
207 void DeleteTriviallyDeadInstructions();
211 char LoopStrengthReduce::ID = 0;
212 static RegisterPass<LoopStrengthReduce>
213 X("loop-reduce", "Loop Strength Reduction");
215 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
216 return new LoopStrengthReduce(TLI);
219 /// getCastedVersionOf - Return the specified value casted to uintptr_t. This
220 /// assumes that the Value* V is of integer or pointer type only.
222 Value *LoopStrengthReduce::getCastedVersionOf(Instruction::CastOps opcode,
224 if (V->getType() == UIntPtrTy) return V;
225 if (Constant *CB = dyn_cast<Constant>(V))
226 return ConstantExpr::getCast(opcode, CB, UIntPtrTy);
228 Value *&New = CastedPointers[V];
231 New = SCEVExpander::InsertCastOfTo(opcode, V, UIntPtrTy);
232 DeadInsts.push_back(cast<Instruction>(New));
237 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
238 /// specified set are trivially dead, delete them and see if this makes any of
239 /// their operands subsequently dead.
240 void LoopStrengthReduce::DeleteTriviallyDeadInstructions() {
241 if (DeadInsts.empty()) return;
243 // Sort the deadinsts list so that we can trivially eliminate duplicates as we
244 // go. The code below never adds a non-dead instruction to the worklist, but
245 // callers may not be so careful.
246 array_pod_sort(DeadInsts.begin(), DeadInsts.end());
248 // Drop duplicate instructions and those with uses.
249 for (unsigned i = 0, e = DeadInsts.size()-1; i < e; ++i) {
250 Instruction *I = DeadInsts[i];
251 if (!I->use_empty()) DeadInsts[i] = 0;
252 while (DeadInsts[i+1] == I && i != e)
256 while (!DeadInsts.empty()) {
257 Instruction *I = DeadInsts.back();
258 DeadInsts.pop_back();
260 if (I == 0 || !isInstructionTriviallyDead(I))
263 SE->deleteValueFromRecords(I);
265 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
266 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
269 DeadInsts.push_back(U);
273 I->eraseFromParent();
279 /// GetExpressionSCEV - Compute and return the SCEV for the specified
281 SCEVHandle LoopStrengthReduce::GetExpressionSCEV(Instruction *Exp) {
282 // Pointer to pointer bitcast instructions return the same value as their
284 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Exp)) {
285 if (SE->hasSCEV(BCI) || !isa<Instruction>(BCI->getOperand(0)))
286 return SE->getSCEV(BCI);
287 SCEVHandle R = GetExpressionSCEV(cast<Instruction>(BCI->getOperand(0)));
292 // Scalar Evolutions doesn't know how to compute SCEV's for GEP instructions.
293 // If this is a GEP that SE doesn't know about, compute it now and insert it.
294 // If this is not a GEP, or if we have already done this computation, just let
296 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Exp);
297 if (!GEP || SE->hasSCEV(GEP))
298 return SE->getSCEV(Exp);
300 // Analyze all of the subscripts of this getelementptr instruction, looking
301 // for uses that are determined by the trip count of the loop. First, skip
302 // all operands the are not dependent on the IV.
304 // Build up the base expression. Insert an LLVM cast of the pointer to
306 SCEVHandle GEPVal = SE->getUnknown(
307 getCastedVersionOf(Instruction::PtrToInt, GEP->getOperand(0)));
309 gep_type_iterator GTI = gep_type_begin(GEP);
311 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end();
312 i != e; ++i, ++GTI) {
313 // If this is a use of a recurrence that we can analyze, and it comes before
314 // Op does in the GEP operand list, we will handle this when we process this
316 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
317 const StructLayout *SL = TD->getStructLayout(STy);
318 unsigned Idx = cast<ConstantInt>(*i)->getZExtValue();
319 uint64_t Offset = SL->getElementOffset(Idx);
320 GEPVal = SE->getAddExpr(GEPVal,
321 SE->getIntegerSCEV(Offset, UIntPtrTy));
323 unsigned GEPOpiBits =
324 (*i)->getType()->getPrimitiveSizeInBits();
325 unsigned IntPtrBits = UIntPtrTy->getPrimitiveSizeInBits();
326 Instruction::CastOps opcode = (GEPOpiBits < IntPtrBits ?
327 Instruction::SExt : (GEPOpiBits > IntPtrBits ? Instruction::Trunc :
328 Instruction::BitCast));
329 Value *OpVal = getCastedVersionOf(opcode, *i);
330 SCEVHandle Idx = SE->getSCEV(OpVal);
332 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
334 Idx = SE->getMulExpr(Idx,
335 SE->getConstant(ConstantInt::get(UIntPtrTy,
337 GEPVal = SE->getAddExpr(GEPVal, Idx);
341 SE->setSCEV(GEP, GEPVal);
345 /// getSCEVStartAndStride - Compute the start and stride of this expression,
346 /// returning false if the expression is not a start/stride pair, or true if it
347 /// is. The stride must be a loop invariant expression, but the start may be
348 /// a mix of loop invariant and loop variant expressions.
349 static bool getSCEVStartAndStride(const SCEVHandle &SH, Loop *L,
350 SCEVHandle &Start, SCEVHandle &Stride,
351 ScalarEvolution *SE) {
352 SCEVHandle TheAddRec = Start; // Initialize to zero.
354 // If the outer level is an AddExpr, the operands are all start values except
355 // for a nested AddRecExpr.
356 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
357 for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
358 if (SCEVAddRecExpr *AddRec =
359 dyn_cast<SCEVAddRecExpr>(AE->getOperand(i))) {
360 if (AddRec->getLoop() == L)
361 TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
363 return false; // Nested IV of some sort?
365 Start = SE->getAddExpr(Start, AE->getOperand(i));
368 } else if (isa<SCEVAddRecExpr>(SH)) {
371 return false; // not analyzable.
374 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(TheAddRec);
375 if (!AddRec || AddRec->getLoop() != L) return false;
377 // FIXME: Generalize to non-affine IV's.
378 if (!AddRec->isAffine()) return false;
380 Start = SE->getAddExpr(Start, AddRec->getOperand(0));
382 if (!isa<SCEVConstant>(AddRec->getOperand(1)))
383 DOUT << "[" << L->getHeader()->getName()
384 << "] Variable stride: " << *AddRec << "\n";
386 Stride = AddRec->getOperand(1);
390 /// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
391 /// and now we need to decide whether the user should use the preinc or post-inc
392 /// value. If this user should use the post-inc version of the IV, return true.
394 /// Choosing wrong here can break dominance properties (if we choose to use the
395 /// post-inc value when we cannot) or it can end up adding extra live-ranges to
396 /// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
397 /// should use the post-inc value).
398 static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
399 Loop *L, DominatorTree *DT, Pass *P,
400 SmallVectorImpl<Instruction*> &DeadInsts){
401 // If the user is in the loop, use the preinc value.
402 if (L->contains(User->getParent())) return false;
404 BasicBlock *LatchBlock = L->getLoopLatch();
406 // Ok, the user is outside of the loop. If it is dominated by the latch
407 // block, use the post-inc value.
408 if (DT->dominates(LatchBlock, User->getParent()))
411 // There is one case we have to be careful of: PHI nodes. These little guys
412 // can live in blocks that do not dominate the latch block, but (since their
413 // uses occur in the predecessor block, not the block the PHI lives in) should
414 // still use the post-inc value. Check for this case now.
415 PHINode *PN = dyn_cast<PHINode>(User);
416 if (!PN) return false; // not a phi, not dominated by latch block.
418 // Look at all of the uses of IV by the PHI node. If any use corresponds to
419 // a block that is not dominated by the latch block, give up and use the
420 // preincremented value.
421 unsigned NumUses = 0;
422 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
423 if (PN->getIncomingValue(i) == IV) {
425 if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
429 // Okay, all uses of IV by PN are in predecessor blocks that really are
430 // dominated by the latch block. Split the critical edges and use the
431 // post-incremented value.
432 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
433 if (PN->getIncomingValue(i) == IV) {
434 SplitCriticalEdge(PN->getIncomingBlock(i), PN->getParent(), P, false);
435 // Splitting the critical edge can reduce the number of entries in this
437 e = PN->getNumIncomingValues();
438 if (--NumUses == 0) break;
441 // PHI node might have become a constant value after SplitCriticalEdge.
442 DeadInsts.push_back(User);
449 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
450 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
451 /// return true. Otherwise, return false.
452 bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L,
453 SmallPtrSet<Instruction*,16> &Processed) {
454 if (!I->getType()->isInteger() && !isa<PointerType>(I->getType()))
455 return false; // Void and FP expressions cannot be reduced.
456 if (!Processed.insert(I))
457 return true; // Instruction already handled.
459 // Get the symbolic expression for this instruction.
460 SCEVHandle ISE = GetExpressionSCEV(I);
461 if (isa<SCEVCouldNotCompute>(ISE)) return false;
463 // Get the start and stride for this expression.
464 SCEVHandle Start = SE->getIntegerSCEV(0, ISE->getType());
465 SCEVHandle Stride = Start;
466 if (!getSCEVStartAndStride(ISE, L, Start, Stride, SE))
467 return false; // Non-reducible symbolic expression, bail out.
469 std::vector<Instruction *> IUsers;
470 // Collect all I uses now because IVUseShouldUsePostIncValue may
471 // invalidate use_iterator.
472 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
473 IUsers.push_back(cast<Instruction>(*UI));
475 for (unsigned iused_index = 0, iused_size = IUsers.size();
476 iused_index != iused_size; ++iused_index) {
478 Instruction *User = IUsers[iused_index];
480 // Do not infinitely recurse on PHI nodes.
481 if (isa<PHINode>(User) && Processed.count(User))
484 // If this is an instruction defined in a nested loop, or outside this loop,
485 // don't recurse into it.
486 bool AddUserToIVUsers = false;
487 if (LI->getLoopFor(User->getParent()) != L) {
488 DOUT << "FOUND USER in other loop: " << *User
489 << " OF SCEV: " << *ISE << "\n";
490 AddUserToIVUsers = true;
491 } else if (!AddUsersIfInteresting(User, L, Processed)) {
492 DOUT << "FOUND USER: " << *User
493 << " OF SCEV: " << *ISE << "\n";
494 AddUserToIVUsers = true;
497 if (AddUserToIVUsers) {
498 IVUsersOfOneStride &StrideUses = IVUsesByStride[Stride];
499 if (StrideUses.Users.empty()) // First occurance of this stride?
500 StrideOrder.push_back(Stride);
502 // Okay, we found a user that we cannot reduce. Analyze the instruction
503 // and decide what to do with it. If we are a use inside of the loop, use
504 // the value before incrementation, otherwise use it after incrementation.
505 if (IVUseShouldUsePostIncValue(User, I, L, DT, this, DeadInsts)) {
506 // The value used will be incremented by the stride more than we are
507 // expecting, so subtract this off.
508 SCEVHandle NewStart = SE->getMinusSCEV(Start, Stride);
509 StrideUses.addUser(NewStart, User, I);
510 StrideUses.Users.back().isUseOfPostIncrementedValue = true;
511 DOUT << " USING POSTINC SCEV, START=" << *NewStart<< "\n";
513 StrideUses.addUser(Start, User, I);
521 /// BasedUser - For a particular base value, keep information about how we've
522 /// partitioned the expression so far.
524 /// SE - The current ScalarEvolution object.
527 /// Base - The Base value for the PHI node that needs to be inserted for
528 /// this use. As the use is processed, information gets moved from this
529 /// field to the Imm field (below). BasedUser values are sorted by this
533 /// Inst - The instruction using the induction variable.
536 /// OperandValToReplace - The operand value of Inst to replace with the
538 Value *OperandValToReplace;
540 /// Imm - The immediate value that should be added to the base immediately
541 /// before Inst, because it will be folded into the imm field of the
545 /// EmittedBase - The actual value* to use for the base value of this
546 /// operation. This is null if we should just use zero so far.
549 // isUseOfPostIncrementedValue - True if this should use the
550 // post-incremented version of this IV, not the preincremented version.
551 // This can only be set in special cases, such as the terminating setcc
552 // instruction for a loop and uses outside the loop that are dominated by
554 bool isUseOfPostIncrementedValue;
556 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se)
557 : SE(se), Base(IVSU.Offset), Inst(IVSU.User),
558 OperandValToReplace(IVSU.OperandValToReplace),
559 Imm(SE->getIntegerSCEV(0, Base->getType())), EmittedBase(0),
560 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue) {}
562 // Once we rewrite the code to insert the new IVs we want, update the
563 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
565 void RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
566 Instruction *InsertPt,
567 SCEVExpander &Rewriter, Loop *L, Pass *P,
568 SmallVectorImpl<Instruction*> &DeadInsts);
570 Value *InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
571 SCEVExpander &Rewriter,
572 Instruction *IP, Loop *L);
577 void BasedUser::dump() const {
578 cerr << " Base=" << *Base;
579 cerr << " Imm=" << *Imm;
581 cerr << " EB=" << *EmittedBase;
583 cerr << " Inst: " << *Inst;
586 Value *BasedUser::InsertCodeForBaseAtPosition(const SCEVHandle &NewBase,
587 SCEVExpander &Rewriter,
588 Instruction *IP, Loop *L) {
589 // Figure out where we *really* want to insert this code. In particular, if
590 // the user is inside of a loop that is nested inside of L, we really don't
591 // want to insert this expression before the user, we'd rather pull it out as
592 // many loops as possible.
593 LoopInfo &LI = Rewriter.getLoopInfo();
594 Instruction *BaseInsertPt = IP;
596 // Figure out the most-nested loop that IP is in.
597 Loop *InsertLoop = LI.getLoopFor(IP->getParent());
599 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out
600 // the preheader of the outer-most loop where NewBase is not loop invariant.
601 if (L->contains(IP->getParent()))
602 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
603 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
604 InsertLoop = InsertLoop->getParentLoop();
607 // If there is no immediate value, skip the next part.
609 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
611 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
613 // If we are inserting the base and imm values in the same block, make sure to
614 // adjust the IP position if insertion reused a result.
615 if (IP == BaseInsertPt)
616 IP = Rewriter.getInsertionPoint();
618 // Always emit the immediate (if non-zero) into the same block as the user.
619 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
620 return Rewriter.expandCodeFor(NewValSCEV, IP);
625 // Once we rewrite the code to insert the new IVs we want, update the
626 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
627 // to it. NewBasePt is the last instruction which contributes to the
628 // value of NewBase in the case that it's a diffferent instruction from
629 // the PHI that NewBase is computed from, or null otherwise.
631 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
632 Instruction *NewBasePt,
633 SCEVExpander &Rewriter, Loop *L, Pass *P,
634 SmallVectorImpl<Instruction*> &DeadInsts){
635 if (!isa<PHINode>(Inst)) {
636 // By default, insert code at the user instruction.
637 BasicBlock::iterator InsertPt = Inst;
639 // However, if the Operand is itself an instruction, the (potentially
640 // complex) inserted code may be shared by many users. Because of this, we
641 // want to emit code for the computation of the operand right before its old
642 // computation. This is usually safe, because we obviously used to use the
643 // computation when it was computed in its current block. However, in some
644 // cases (e.g. use of a post-incremented induction variable) the NewBase
645 // value will be pinned to live somewhere after the original computation.
646 // In this case, we have to back off.
648 // If this is a use outside the loop (which means after, since it is based
649 // on a loop indvar) we use the post-incremented value, so that we don't
650 // artificially make the preinc value live out the bottom of the loop.
651 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) {
652 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
653 InsertPt = NewBasePt;
655 } else if (Instruction *OpInst
656 = dyn_cast<Instruction>(OperandValToReplace)) {
658 while (isa<PHINode>(InsertPt)) ++InsertPt;
661 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
662 // Adjust the type back to match the Inst. Note that we can't use InsertPt
663 // here because the SCEVExpander may have inserted the instructions after
664 // that point, in its efforts to avoid inserting redundant expressions.
665 if (isa<PointerType>(OperandValToReplace->getType())) {
666 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
668 OperandValToReplace->getType());
670 // Replace the use of the operand Value with the new Phi we just created.
671 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
672 DOUT << " CHANGED: IMM =" << *Imm;
673 DOUT << " \tNEWBASE =" << *NewBase;
674 DOUT << " \tInst = " << *Inst;
678 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
679 // expression into each operand block that uses it. Note that PHI nodes can
680 // have multiple entries for the same predecessor. We use a map to make sure
681 // that a PHI node only has a single Value* for each predecessor (which also
682 // prevents us from inserting duplicate code in some blocks).
683 DenseMap<BasicBlock*, Value*> InsertedCode;
684 PHINode *PN = cast<PHINode>(Inst);
685 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
686 if (PN->getIncomingValue(i) == OperandValToReplace) {
687 // If this is a critical edge, split the edge so that we do not insert the
688 // code on all predecessor/successor paths. We do this unless this is the
689 // canonical backedge for this loop, as this can make some inserted code
690 // be in an illegal position.
691 BasicBlock *PHIPred = PN->getIncomingBlock(i);
692 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
693 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
695 // First step, split the critical edge.
696 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
698 // Next step: move the basic block. In particular, if the PHI node
699 // is outside of the loop, and PredTI is in the loop, we want to
700 // move the block to be immediately before the PHI block, not
701 // immediately after PredTI.
702 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
703 BasicBlock *NewBB = PN->getIncomingBlock(i);
704 NewBB->moveBefore(PN->getParent());
707 // Splitting the edge can reduce the number of PHI entries we have.
708 e = PN->getNumIncomingValues();
711 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
713 // Insert the code into the end of the predecessor block.
714 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
715 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
717 // Adjust the type back to match the PHI. Note that we can't use
718 // InsertPt here because the SCEVExpander may have inserted its
719 // instructions after that point, in its efforts to avoid inserting
720 // redundant expressions.
721 if (isa<PointerType>(PN->getType())) {
722 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
728 // Replace the use of the operand Value with the new Phi we just created.
729 PN->setIncomingValue(i, Code);
734 // PHI node might have become a constant value after SplitCriticalEdge.
735 DeadInsts.push_back(Inst);
737 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
741 /// isTargetConstant - Return true if the following can be referenced by the
742 /// immediate field of a target instruction.
743 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
744 const TargetLowering *TLI) {
745 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
746 int64_t VC = SC->getValue()->getSExtValue();
748 TargetLowering::AddrMode AM;
750 return TLI->isLegalAddressingMode(AM, UseTy);
752 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
753 return (VC > -(1 << 16) && VC < (1 << 16)-1);
757 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
758 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
759 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
760 Constant *Op0 = CE->getOperand(0);
761 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
762 TargetLowering::AddrMode AM;
764 return TLI->isLegalAddressingMode(AM, UseTy);
770 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
771 /// loop varying to the Imm operand.
772 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
773 Loop *L, ScalarEvolution *SE) {
774 if (Val->isLoopInvariant(L)) return; // Nothing to do.
776 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
777 std::vector<SCEVHandle> NewOps;
778 NewOps.reserve(SAE->getNumOperands());
780 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
781 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
782 // If this is a loop-variant expression, it must stay in the immediate
783 // field of the expression.
784 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
786 NewOps.push_back(SAE->getOperand(i));
790 Val = SE->getIntegerSCEV(0, Val->getType());
792 Val = SE->getAddExpr(NewOps);
793 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
794 // Try to pull immediates out of the start value of nested addrec's.
795 SCEVHandle Start = SARE->getStart();
796 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
798 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
800 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
802 // Otherwise, all of Val is variant, move the whole thing over.
803 Imm = SE->getAddExpr(Imm, Val);
804 Val = SE->getIntegerSCEV(0, Val->getType());
809 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
810 /// that can fit into the immediate field of instructions in the target.
811 /// Accumulate these immediate values into the Imm value.
812 static void MoveImmediateValues(const TargetLowering *TLI,
814 SCEVHandle &Val, SCEVHandle &Imm,
815 bool isAddress, Loop *L,
816 ScalarEvolution *SE) {
817 const Type *UseTy = User->getType();
818 if (StoreInst *SI = dyn_cast<StoreInst>(User))
819 UseTy = SI->getOperand(0)->getType();
821 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
822 std::vector<SCEVHandle> NewOps;
823 NewOps.reserve(SAE->getNumOperands());
825 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
826 SCEVHandle NewOp = SAE->getOperand(i);
827 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
829 if (!NewOp->isLoopInvariant(L)) {
830 // If this is a loop-variant expression, it must stay in the immediate
831 // field of the expression.
832 Imm = SE->getAddExpr(Imm, NewOp);
834 NewOps.push_back(NewOp);
839 Val = SE->getIntegerSCEV(0, Val->getType());
841 Val = SE->getAddExpr(NewOps);
843 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
844 // Try to pull immediates out of the start value of nested addrec's.
845 SCEVHandle Start = SARE->getStart();
846 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
848 if (Start != SARE->getStart()) {
849 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
851 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
854 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
855 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
856 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
857 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
859 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
860 SCEVHandle NewOp = SME->getOperand(1);
861 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
863 // If we extracted something out of the subexpressions, see if we can
865 if (NewOp != SME->getOperand(1)) {
866 // Scale SubImm up by "8". If the result is a target constant, we are
868 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
869 if (isTargetConstant(SubImm, UseTy, TLI)) {
870 // Accumulate the immediate.
871 Imm = SE->getAddExpr(Imm, SubImm);
873 // Update what is left of 'Val'.
874 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
881 // Loop-variant expressions must stay in the immediate field of the
883 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
884 !Val->isLoopInvariant(L)) {
885 Imm = SE->getAddExpr(Imm, Val);
886 Val = SE->getIntegerSCEV(0, Val->getType());
890 // Otherwise, no immediates to move.
894 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
895 /// added together. This is used to reassociate common addition subexprs
896 /// together for maximal sharing when rewriting bases.
897 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
899 ScalarEvolution *SE) {
900 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
901 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
902 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
903 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
904 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
905 if (SARE->getOperand(0) == Zero) {
906 SubExprs.push_back(Expr);
908 // Compute the addrec with zero as its base.
909 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
910 Ops[0] = Zero; // Start with zero base.
911 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
914 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
916 } else if (!Expr->isZero()) {
918 SubExprs.push_back(Expr);
923 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
924 /// removing any common subexpressions from it. Anything truly common is
925 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
926 /// (a+c+d) and computes the common (a+c) subexpression. The common expression
927 /// is *removed* from the Bases and returned.
929 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
930 ScalarEvolution *SE, Loop *L) {
931 unsigned NumUses = Uses.size();
933 // Only one use? This is a very common case, so we handle it specially and
935 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
936 SCEVHandle Result = Zero;
938 // If the use is inside the loop, use its base, regardless of what it is:
939 // it is clearly shared across all the IV's. If the use is outside the loop
940 // (which means after it) we don't want to factor anything *into* the loop,
941 // so just use 0 as the base.
942 if (L->contains(Uses[0].Inst->getParent()))
943 std::swap(Result, Uses[0].Base);
947 // To find common subexpressions, count how many of Uses use each expression.
948 // If any subexpressions are used Uses.size() times, they are common.
949 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
951 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
952 // order we see them.
953 std::vector<SCEVHandle> UniqueSubExprs;
955 std::vector<SCEVHandle> SubExprs;
956 unsigned NumUsesInsideLoop = 0;
957 for (unsigned i = 0; i != NumUses; ++i) {
958 // If the user is outside the loop, just ignore it for base computation.
959 // Since the user is outside the loop, it must be *after* the loop (if it
960 // were before, it could not be based on the loop IV). We don't want users
961 // after the loop to affect base computation of values *inside* the loop,
962 // because we can always add their offsets to the result IV after the loop
963 // is done, ensuring we get good code inside the loop.
964 if (!L->contains(Uses[i].Inst->getParent()))
968 // If the base is zero (which is common), return zero now, there are no
970 if (Uses[i].Base == Zero) return Zero;
972 // Split the expression into subexprs.
973 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
974 // Add one to SubExpressionUseCounts for each subexpr present.
975 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
976 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
977 UniqueSubExprs.push_back(SubExprs[j]);
981 // Now that we know how many times each is used, build Result. Iterate over
982 // UniqueSubexprs so that we have a stable ordering.
983 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
984 std::map<SCEVHandle, unsigned>::iterator I =
985 SubExpressionUseCounts.find(UniqueSubExprs[i]);
986 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
987 if (I->second == NumUsesInsideLoop) // Found CSE!
988 Result = SE->getAddExpr(Result, I->first);
990 // Remove non-cse's from SubExpressionUseCounts.
991 SubExpressionUseCounts.erase(I);
994 // If we found no CSE's, return now.
995 if (Result == Zero) return Result;
997 // Otherwise, remove all of the CSE's we found from each of the base values.
998 for (unsigned i = 0; i != NumUses; ++i) {
999 // Uses outside the loop don't necessarily include the common base, but
1000 // the final IV value coming into those uses does. Instead of trying to
1001 // remove the pieces of the common base, which might not be there,
1002 // subtract off the base to compensate for this.
1003 if (!L->contains(Uses[i].Inst->getParent())) {
1004 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result);
1008 // Split the expression into subexprs.
1009 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
1011 // Remove any common subexpressions.
1012 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
1013 if (SubExpressionUseCounts.count(SubExprs[j])) {
1014 SubExprs.erase(SubExprs.begin()+j);
1018 // Finally, add the non-shared expressions together.
1019 if (SubExprs.empty())
1020 Uses[i].Base = Zero;
1022 Uses[i].Base = SE->getAddExpr(SubExprs);
1029 /// ValidStride - Check whether the given Scale is valid for all loads and
1030 /// stores in UsersToProcess.
1032 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1034 const std::vector<BasedUser>& UsersToProcess) {
1038 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1039 // If this is a load or other access, pass the type of the access in.
1040 const Type *AccessTy = Type::VoidTy;
1041 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1042 AccessTy = SI->getOperand(0)->getType();
1043 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1044 AccessTy = LI->getType();
1045 else if (isa<PHINode>(UsersToProcess[i].Inst))
1048 TargetLowering::AddrMode AM;
1049 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1050 AM.BaseOffs = SC->getValue()->getSExtValue();
1051 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1054 // If load[imm+r*scale] is illegal, bail out.
1055 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1061 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1063 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1067 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1069 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1070 !(isa<PointerType>(Ty2) &&
1071 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1072 !(isa<PointerType>(Ty1) &&
1073 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1076 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1077 /// of a previous stride and it is a legal value for the target addressing
1078 /// mode scale component and optional base reg. This allows the users of
1079 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1080 /// reuse is possible.
1081 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1082 bool AllUsesAreAddresses,
1083 const SCEVHandle &Stride,
1084 IVExpr &IV, const Type *Ty,
1085 const std::vector<BasedUser>& UsersToProcess) {
1086 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1087 int64_t SInt = SC->getValue()->getSExtValue();
1088 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1090 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1091 IVsByStride.find(StrideOrder[NewStride]);
1092 if (SI == IVsByStride.end())
1094 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1095 if (SI->first != Stride &&
1096 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1098 int64_t Scale = SInt / SSInt;
1099 // Check that this stride is valid for all the types used for loads and
1100 // stores; if it can be used for some and not others, we might as well use
1101 // the original stride everywhere, since we have to create the IV for it
1102 // anyway. If the scale is 1, then we don't need to worry about folding
1105 (AllUsesAreAddresses &&
1106 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1107 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1108 IE = SI->second.IVs.end(); II != IE; ++II)
1109 // FIXME: Only handle base == 0 for now.
1110 // Only reuse previous IV if it would not require a type conversion.
1111 if (II->Base->isZero() &&
1112 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1121 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1122 /// returns true if Val's isUseOfPostIncrementedValue is true.
1123 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1124 return Val.isUseOfPostIncrementedValue;
1127 /// isNonConstantNegative - Return true if the specified scev is negated, but
1129 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1130 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1131 if (!Mul) return false;
1133 // If there is a constant factor, it will be first.
1134 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1135 if (!SC) return false;
1137 // Return true if the value is negative, this matches things like (-42 * V).
1138 return SC->getValue()->getValue().isNegative();
1141 /// isAddress - Returns true if the specified instruction is using the
1142 /// specified value as an address.
1143 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
1144 bool isAddress = isa<LoadInst>(Inst);
1145 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1146 if (SI->getOperand(1) == OperandVal)
1148 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1149 // Addressing modes can also be folded into prefetches and a variety
1151 switch (II->getIntrinsicID()) {
1153 case Intrinsic::prefetch:
1154 case Intrinsic::x86_sse2_loadu_dq:
1155 case Intrinsic::x86_sse2_loadu_pd:
1156 case Intrinsic::x86_sse_loadu_ps:
1157 case Intrinsic::x86_sse_storeu_ps:
1158 case Intrinsic::x86_sse2_storeu_pd:
1159 case Intrinsic::x86_sse2_storeu_dq:
1160 case Intrinsic::x86_sse2_storel_dq:
1161 if (II->getOperand(1) == OperandVal)
1169 // CollectIVUsers - Transform our list of users and offsets to a bit more
1170 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1171 // of the strided accesses, as well as the old information from Uses. We
1172 // progressively move information from the Base field to the Imm field, until
1173 // we eventually have the full access expression to rewrite the use.
1174 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1175 IVUsersOfOneStride &Uses,
1177 bool &AllUsesAreAddresses,
1178 std::vector<BasedUser> &UsersToProcess) {
1179 UsersToProcess.reserve(Uses.Users.size());
1180 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1181 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1183 // Move any loop invariant operands from the offset field to the immediate
1184 // field of the use, so that we don't try to use something before it is
1186 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1187 UsersToProcess.back().Imm, L, SE);
1188 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1189 "Base value is not loop invariant!");
1192 // We now have a whole bunch of uses of like-strided induction variables, but
1193 // they might all have different bases. We want to emit one PHI node for this
1194 // stride which we fold as many common expressions (between the IVs) into as
1195 // possible. Start by identifying the common expressions in the base values
1196 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1197 // "A+B"), emit it to the preheader, then remove the expression from the
1198 // UsersToProcess base values.
1199 SCEVHandle CommonExprs =
1200 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L);
1202 // Next, figure out what we can represent in the immediate fields of
1203 // instructions. If we can represent anything there, move it to the imm
1204 // fields of the BasedUsers. We do this so that it increases the commonality
1205 // of the remaining uses.
1206 unsigned NumPHI = 0;
1207 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1208 // If the user is not in the current loop, this means it is using the exit
1209 // value of the IV. Do not put anything in the base, make sure it's all in
1210 // the immediate field to allow as much factoring as possible.
1211 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1212 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1213 UsersToProcess[i].Base);
1214 UsersToProcess[i].Base =
1215 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1218 // Addressing modes can be folded into loads and stores. Be careful that
1219 // the store is through the expression, not of the expression though.
1221 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1222 UsersToProcess[i].OperandValToReplace);
1223 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1228 // If this use isn't an address, then not all uses are addresses.
1229 if (!isAddress && !isPHI)
1230 AllUsesAreAddresses = false;
1232 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1233 UsersToProcess[i].Imm, isAddress, L, SE);
1237 // If one of the use if a PHI node and all other uses are addresses, still
1238 // allow iv reuse. Essentially we are trading one constant multiplication
1239 // for one fewer iv.
1241 AllUsesAreAddresses = false;
1246 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1247 /// stride of IV. All of the users may have different starting values, and this
1248 /// may not be the only stride (we know it is if isOnlyStride is true).
1249 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1250 IVUsersOfOneStride &Uses,
1252 bool isOnlyStride) {
1253 // If all the users are moved to another stride, then there is nothing to do.
1254 if (Uses.Users.empty())
1257 // Keep track if every use in UsersToProcess is an address. If they all are,
1258 // we may be able to rewrite the entire collection of them in terms of a
1259 // smaller-stride IV.
1260 bool AllUsesAreAddresses = true;
1262 // Transform our list of users and offsets to a bit more complex table. In
1263 // this new vector, each 'BasedUser' contains 'Base' the base of the
1264 // strided accessas well as the old information from Uses. We progressively
1265 // move information from the Base field to the Imm field, until we eventually
1266 // have the full access expression to rewrite the use.
1267 std::vector<BasedUser> UsersToProcess;
1268 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1271 // If we managed to find some expressions in common, we'll need to carry
1272 // their value in a register and add it in for each use. This will take up
1273 // a register operand, which potentially restricts what stride values are
1275 bool HaveCommonExprs = !CommonExprs->isZero();
1277 // If all uses are addresses, check if it is possible to reuse an IV with a
1278 // stride that is a factor of this stride. And that the multiple is a number
1279 // that can be encoded in the scale field of the target addressing mode. And
1280 // that we will have a valid instruction after this substition, including the
1281 // immediate field, if any.
1282 PHINode *NewPHI = NULL;
1284 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1285 SE->getIntegerSCEV(0, Type::Int32Ty),
1287 unsigned RewriteFactor = 0;
1288 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1289 Stride, ReuseIV, CommonExprs->getType(),
1291 if (RewriteFactor != 0) {
1292 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1293 << " and BASE " << *ReuseIV.Base << " :\n";
1294 NewPHI = ReuseIV.PHI;
1295 IncV = ReuseIV.IncV;
1298 const Type *ReplacedTy = CommonExprs->getType();
1300 // Now that we know what we need to do, insert the PHI node itself.
1302 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1303 << *Stride << " and BASE " << *CommonExprs << ": ";
1305 SCEVExpander Rewriter(*SE, *LI);
1306 SCEVExpander PreheaderRewriter(*SE, *LI);
1308 BasicBlock *Preheader = L->getLoopPreheader();
1309 Instruction *PreInsertPt = Preheader->getTerminator();
1310 Instruction *PhiInsertBefore = L->getHeader()->begin();
1312 BasicBlock *LatchBlock = L->getLoopLatch();
1315 // Emit the initial base value into the loop preheader.
1317 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1319 if (RewriteFactor == 0) {
1320 // Create a new Phi for this base, and stick it in the loop header.
1321 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1324 // Add common base to the new Phi node.
1325 NewPHI->addIncoming(CommonBaseV, Preheader);
1327 // If the stride is negative, insert a sub instead of an add for the
1329 bool isNegative = isNonConstantNegative(Stride);
1330 SCEVHandle IncAmount = Stride;
1332 IncAmount = SE->getNegativeSCEV(Stride);
1334 // Insert the stride into the preheader.
1335 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1336 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1338 // Emit the increment of the base value before the terminator of the loop
1339 // latch block, and add it to the Phi node.
1340 SCEVHandle IncExp = SE->getUnknown(StrideV);
1342 IncExp = SE->getNegativeSCEV(IncExp);
1343 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1345 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1346 IncV->setName(NewPHI->getName()+".inc");
1347 NewPHI->addIncoming(IncV, LatchBlock);
1349 // Remember this in case a later stride is multiple of this.
1350 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1352 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1354 Constant *C = dyn_cast<Constant>(CommonBaseV);
1356 (!C->isNullValue() &&
1357 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1358 // We want the common base emitted into the preheader! This is just
1359 // using cast as a copy so BitCast (no-op cast) is appropriate
1360 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1361 "commonbase", PreInsertPt);
1365 // We want to emit code for users inside the loop first. To do this, we
1366 // rearrange BasedUser so that the entries at the end have
1367 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1368 // vector (so we handle them first).
1369 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1370 PartitionByIsUseOfPostIncrementedValue);
1372 // Sort this by base, so that things with the same base are handled
1373 // together. By partitioning first and stable-sorting later, we are
1374 // guaranteed that within each base we will pop off users from within the
1375 // loop before users outside of the loop with a particular base.
1377 // We would like to use stable_sort here, but we can't. The problem is that
1378 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1379 // we don't have anything to do a '<' comparison on. Because we think the
1380 // number of uses is small, do a horrible bubble sort which just relies on
1382 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1383 // Get a base value.
1384 SCEVHandle Base = UsersToProcess[i].Base;
1386 // Compact everything with this base to be consequtive with this one.
1387 for (unsigned j = i+1; j != e; ++j) {
1388 if (UsersToProcess[j].Base == Base) {
1389 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1395 // Process all the users now. This outer loop handles all bases, the inner
1396 // loop handles all users of a particular base.
1397 while (!UsersToProcess.empty()) {
1398 SCEVHandle Base = UsersToProcess.back().Base;
1400 // Emit the code for Base into the preheader.
1401 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1403 DOUT << " INSERTING code for BASE = " << *Base << ":";
1404 if (BaseV->hasName())
1405 DOUT << " Result value name = %" << BaseV->getNameStr();
1408 // If BaseV is a constant other than 0, make sure that it gets inserted into
1409 // the preheader, instead of being forward substituted into the uses. We do
1410 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1412 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1413 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1414 // We want this constant emitted into the preheader! This is just
1415 // using cast as a copy so BitCast (no-op cast) is appropriate
1416 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1421 // Emit the code to add the immediate offset to the Phi value, just before
1422 // the instructions that we identified as using this stride and base.
1424 // FIXME: Use emitted users to emit other users.
1425 BasedUser &User = UsersToProcess.back();
1427 // If this instruction wants to use the post-incremented value, move it
1428 // after the post-inc and use its value instead of the PHI.
1429 Value *RewriteOp = NewPHI;
1430 if (User.isUseOfPostIncrementedValue) {
1433 // If this user is in the loop, make sure it is the last thing in the
1434 // loop to ensure it is dominated by the increment.
1435 if (L->contains(User.Inst->getParent()))
1436 User.Inst->moveBefore(LatchBlock->getTerminator());
1438 if (RewriteOp->getType() != ReplacedTy) {
1439 Instruction::CastOps opcode = Instruction::Trunc;
1440 if (ReplacedTy->getPrimitiveSizeInBits() ==
1441 RewriteOp->getType()->getPrimitiveSizeInBits())
1442 opcode = Instruction::BitCast;
1443 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1446 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1448 // If we had to insert new instrutions for RewriteOp, we have to
1449 // consider that they may not have been able to end up immediately
1450 // next to RewriteOp, because non-PHI instructions may never precede
1451 // PHI instructions in a block. In this case, remember where the last
1452 // instruction was inserted so that if we're replacing a different
1453 // PHI node, we can use the later point to expand the final
1455 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1456 if (RewriteOp == NewPHI) NewBasePt = 0;
1458 // Clear the SCEVExpander's expression map so that we are guaranteed
1459 // to have the code emitted where we expect it.
1462 // If we are reusing the iv, then it must be multiplied by a constant
1463 // factor take advantage of addressing mode scale component.
1464 if (RewriteFactor != 0) {
1465 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1466 RewriteExpr->getType()),
1469 // The common base is emitted in the loop preheader. But since we
1470 // are reusing an IV, it has not been used to initialize the PHI node.
1471 // Add it to the expression used to rewrite the uses.
1472 if (!isa<ConstantInt>(CommonBaseV) ||
1473 !cast<ConstantInt>(CommonBaseV)->isZero())
1474 RewriteExpr = SE->getAddExpr(RewriteExpr,
1475 SE->getUnknown(CommonBaseV));
1478 // Now that we know what we need to do, insert code before User for the
1479 // immediate and any loop-variant expressions.
1480 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1481 // Add BaseV to the PHI value if needed.
1482 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1484 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1488 // Mark old value we replaced as possibly dead, so that it is eliminated
1489 // if we just replaced the last use of that value.
1490 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1492 UsersToProcess.pop_back();
1495 // If there are any more users to process with the same base, process them
1496 // now. We sorted by base above, so we just have to check the last elt.
1497 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1498 // TODO: Next, find out which base index is the most common, pull it out.
1501 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1502 // different starting values, into different PHIs.
1505 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1506 /// set the IV user and stride information and return true, otherwise return
1508 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1509 const SCEVHandle *&CondStride) {
1510 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1512 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1513 IVUsesByStride.find(StrideOrder[Stride]);
1514 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1516 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1517 E = SI->second.Users.end(); UI != E; ++UI)
1518 if (UI->User == Cond) {
1519 // NOTE: we could handle setcc instructions with multiple uses here, but
1520 // InstCombine does it as well for simple uses, it's not clear that it
1521 // occurs enough in real life to handle.
1523 CondStride = &SI->first;
1531 // Constant strides come first which in turns are sorted by their absolute
1532 // values. If absolute values are the same, then positive strides comes first.
1534 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1535 struct StrideCompare {
1536 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1537 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1538 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1540 int64_t LV = LHSC->getValue()->getSExtValue();
1541 int64_t RV = RHSC->getValue()->getSExtValue();
1542 uint64_t ALV = (LV < 0) ? -LV : LV;
1543 uint64_t ARV = (RV < 0) ? -RV : RV;
1549 return (LHSC && !RHSC);
1554 /// ChangeCompareStride - If a loop termination compare instruction is the
1555 /// only use of its stride, and the compaison is against a constant value,
1556 /// try eliminate the stride by moving the compare instruction to another
1557 /// stride and change its constant operand accordingly. e.g.
1563 /// if (v2 < 10) goto loop
1568 /// if (v1 < 30) goto loop
1569 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1570 IVStrideUse* &CondUse,
1571 const SCEVHandle* &CondStride) {
1572 if (StrideOrder.size() < 2 ||
1573 IVUsesByStride[*CondStride].Users.size() != 1)
1575 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1576 if (!SC) return Cond;
1577 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1578 if (!C) return Cond;
1580 ICmpInst::Predicate Predicate = Cond->getPredicate();
1581 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1582 int64_t CmpVal = C->getValue().getSExtValue();
1583 unsigned BitWidth = C->getValue().getBitWidth();
1584 uint64_t SignBit = 1ULL << (BitWidth-1);
1585 const Type *CmpTy = C->getType();
1586 const Type *NewCmpTy = NULL;
1587 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1588 unsigned NewTyBits = 0;
1589 int64_t NewCmpVal = CmpVal;
1590 SCEVHandle *NewStride = NULL;
1591 Value *NewIncV = NULL;
1594 // Check stride constant and the comparision constant signs to detect
1596 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1599 // Look for a suitable stride / iv as replacement.
1600 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1601 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1602 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1603 IVUsesByStride.find(StrideOrder[i]);
1604 if (!isa<SCEVConstant>(SI->first))
1606 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1607 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1610 Scale = SSInt / CmpSSInt;
1611 NewCmpVal = CmpVal * Scale;
1612 APInt Mul = APInt(BitWidth, NewCmpVal);
1613 // Check for overflow.
1614 if (Mul.getSExtValue() != NewCmpVal) {
1619 // Watch out for overflow.
1620 if (ICmpInst::isSignedPredicate(Predicate) &&
1621 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1624 if (NewCmpVal != CmpVal) {
1625 // Pick the best iv to use trying to avoid a cast.
1627 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1628 E = SI->second.Users.end(); UI != E; ++UI) {
1629 NewIncV = UI->OperandValToReplace;
1630 if (NewIncV->getType() == CmpTy)
1638 NewCmpTy = NewIncV->getType();
1639 NewTyBits = isa<PointerType>(NewCmpTy)
1640 ? UIntPtrTy->getPrimitiveSizeInBits()
1641 : NewCmpTy->getPrimitiveSizeInBits();
1642 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1643 // Check if it is possible to rewrite it using
1644 // an iv / stride of a smaller integer type.
1645 bool TruncOk = false;
1646 if (NewCmpTy->isInteger()) {
1647 unsigned Bits = NewTyBits;
1648 if (ICmpInst::isSignedPredicate(Predicate))
1650 uint64_t Mask = (1ULL << Bits) - 1;
1651 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1660 // Don't rewrite if use offset is non-constant and the new type is
1661 // of a different type.
1662 // FIXME: too conservative?
1663 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1668 bool AllUsesAreAddresses = true;
1669 std::vector<BasedUser> UsersToProcess;
1670 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1671 AllUsesAreAddresses,
1673 // Avoid rewriting the compare instruction with an iv of new stride
1674 // if it's likely the new stride uses will be rewritten using the
1675 if (AllUsesAreAddresses &&
1676 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1681 // If scale is negative, use swapped predicate unless it's testing
1683 if (Scale < 0 && !Cond->isEquality())
1684 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1686 NewStride = &StrideOrder[i];
1691 // Forgo this transformation if it the increment happens to be
1692 // unfortunately positioned after the condition, and the condition
1693 // has multiple uses which prevent it from being moved immediately
1694 // before the branch. See
1695 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1696 // for an example of this situation.
1697 if (!Cond->hasOneUse()) {
1698 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1704 if (NewCmpVal != CmpVal) {
1705 // Create a new compare instruction using new stride / iv.
1706 ICmpInst *OldCond = Cond;
1708 if (!isa<PointerType>(NewCmpTy))
1709 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1711 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1712 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1714 // Insert new compare instruction.
1715 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1716 L->getHeader()->getName() + ".termcond",
1719 // Remove the old compare instruction. The old indvar is probably dead too.
1720 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
1721 SE->deleteValueFromRecords(OldCond);
1722 OldCond->replaceAllUsesWith(Cond);
1723 OldCond->eraseFromParent();
1725 IVUsesByStride[*CondStride].Users.pop_back();
1726 SCEVHandle NewOffset = TyBits == NewTyBits
1727 ? SE->getMulExpr(CondUse->Offset,
1728 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1729 : SE->getConstant(ConstantInt::get(NewCmpTy,
1730 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1731 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1732 CondUse = &IVUsesByStride[*NewStride].Users.back();
1733 CondStride = NewStride;
1740 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
1741 /// an smax computation.
1743 /// This is a narrow solution to a specific, but acute, problem. For loops
1749 /// } while (++i < n);
1751 /// where the comparison is signed, the trip count isn't just 'n', because
1752 /// 'n' could be negative. And unfortunately this can come up even for loops
1753 /// where the user didn't use a C do-while loop. For example, seemingly
1754 /// well-behaved top-test loops will commonly be lowered like this:
1760 /// } while (++i < n);
1763 /// and then it's possible for subsequent optimization to obscure the if
1764 /// test in such a way that indvars can't find it.
1766 /// When indvars can't find the if test in loops like this, it creates a
1767 /// signed-max expression, which allows it to give the loop a canonical
1768 /// induction variable:
1771 /// smax = n < 1 ? 1 : n;
1774 /// } while (++i != smax);
1776 /// Canonical induction variables are necessary because the loop passes
1777 /// are designed around them. The most obvious example of this is the
1778 /// LoopInfo analysis, which doesn't remember trip count values. It
1779 /// expects to be able to rediscover the trip count each time it is
1780 /// needed, and it does this using a simple analyis that only succeeds if
1781 /// the loop has a canonical induction variable.
1783 /// However, when it comes time to generate code, the maximum operation
1784 /// can be quite costly, especially if it's inside of an outer loop.
1786 /// This function solves this problem by detecting this type of loop and
1787 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1788 /// the instructions for the maximum computation.
1790 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
1791 IVStrideUse* &CondUse) {
1792 // Check that the loop matches the pattern we're looking for.
1793 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1794 Cond->getPredicate() != CmpInst::ICMP_NE)
1797 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1798 if (!Sel || !Sel->hasOneUse()) return Cond;
1800 SCEVHandle IterationCount = SE->getIterationCount(L);
1801 if (isa<SCEVCouldNotCompute>(IterationCount))
1803 SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());
1805 // Adjust for an annoying getIterationCount quirk.
1806 IterationCount = SE->getAddExpr(IterationCount, One);
1808 // Check for a max calculation that matches the pattern.
1809 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
1810 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
1812 SCEVHandle SMaxLHS = SMax->getOperand(0);
1813 SCEVHandle SMaxRHS = SMax->getOperand(1);
1814 if (!SMaxLHS || SMaxLHS != One) return Cond;
1816 // Check the relevant induction variable for conformance to
1818 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
1819 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1820 if (!AR || !AR->isAffine() ||
1821 AR->getStart() != One ||
1822 AR->getStepRecurrence(*SE) != One)
1825 // Check the right operand of the select, and remember it, as it will
1826 // be used in the new comparison instruction.
1828 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
1829 NewRHS = Sel->getOperand(1);
1830 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
1831 NewRHS = Sel->getOperand(2);
1832 if (!NewRHS) return Cond;
1834 // Ok, everything looks ok to change the condition into an SLT or SGE and
1835 // delete the max calculation.
1837 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
1840 Cond->getOperand(0), NewRHS, "scmp", Cond);
1842 // Delete the max calculation instructions.
1843 SE->deleteValueFromRecords(Cond);
1844 Cond->replaceAllUsesWith(NewCond);
1845 Cond->eraseFromParent();
1846 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1847 SE->deleteValueFromRecords(Sel);
1848 Sel->eraseFromParent();
1849 if (Cmp->use_empty()) {
1850 SE->deleteValueFromRecords(Cmp);
1851 Cmp->eraseFromParent();
1853 CondUse->User = NewCond;
1857 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1858 /// inside the loop then try to eliminate the cast opeation.
1859 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
1861 SCEVHandle IterationCount = SE->getIterationCount(L);
1862 if (isa<SCEVCouldNotCompute>(IterationCount))
1865 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
1867 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1868 IVUsesByStride.find(StrideOrder[Stride]);
1869 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1870 if (!isa<SCEVConstant>(SI->first))
1873 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1874 E = SI->second.Users.end(); UI != E; /* empty */) {
1875 std::vector<IVStrideUse>::iterator CandidateUI = UI;
1877 Instruction *ShadowUse = CandidateUI->User;
1878 const Type *DestTy = NULL;
1880 /* If shadow use is a int->float cast then insert a second IV
1881 to eliminate this cast.
1883 for (unsigned i = 0; i < n; ++i)
1889 for (unsigned i = 0; i < n; ++i, ++d)
1892 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
1893 DestTy = UCast->getDestTy();
1894 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
1895 DestTy = SCast->getDestTy();
1896 if (!DestTy) continue;
1899 /* If target does not support DestTy natively then do not apply
1900 this transformation. */
1901 MVT DVT = TLI->getValueType(DestTy);
1902 if (!TLI->isTypeLegal(DVT)) continue;
1905 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1907 if (PH->getNumIncomingValues() != 2) continue;
1909 const Type *SrcTy = PH->getType();
1910 int Mantissa = DestTy->getFPMantissaWidth();
1911 if (Mantissa == -1) continue;
1912 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
1915 unsigned Entry, Latch;
1916 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1924 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1925 if (!Init) continue;
1926 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1928 BinaryOperator *Incr =
1929 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1930 if (!Incr) continue;
1931 if (Incr->getOpcode() != Instruction::Add
1932 && Incr->getOpcode() != Instruction::Sub)
1935 /* Initialize new IV, double d = 0.0 in above example. */
1936 ConstantInt *C = NULL;
1937 if (Incr->getOperand(0) == PH)
1938 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1939 else if (Incr->getOperand(1) == PH)
1940 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1946 /* Add new PHINode. */
1947 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1949 /* create new increment. '++d' in above example. */
1950 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1951 BinaryOperator *NewIncr =
1952 BinaryOperator::Create(Incr->getOpcode(),
1953 NewPH, CFP, "IV.S.next.", Incr);
1955 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1956 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1958 /* Remove cast operation */
1959 SE->deleteValueFromRecords(ShadowUse);
1960 ShadowUse->replaceAllUsesWith(NewPH);
1961 ShadowUse->eraseFromParent();
1962 SI->second.Users.erase(CandidateUI);
1969 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1970 // uses in the loop, look to see if we can eliminate some, in favor of using
1971 // common indvars for the different uses.
1972 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1973 // TODO: implement optzns here.
1975 OptimizeShadowIV(L);
1977 // Finally, get the terminating condition for the loop if possible. If we
1978 // can, we want to change it to use a post-incremented version of its
1979 // induction variable, to allow coalescing the live ranges for the IV into
1980 // one register value.
1981 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1982 BasicBlock *Preheader = L->getLoopPreheader();
1983 BasicBlock *LatchBlock =
1984 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1985 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1986 if (!TermBr || TermBr->isUnconditional() ||
1987 !isa<ICmpInst>(TermBr->getCondition()))
1989 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1991 // Search IVUsesByStride to find Cond's IVUse if there is one.
1992 IVStrideUse *CondUse = 0;
1993 const SCEVHandle *CondStride = 0;
1995 if (!FindIVUserForCond(Cond, CondUse, CondStride))
1996 return; // setcc doesn't use the IV.
1998 // If the trip count is computed in terms of an smax (due to ScalarEvolution
1999 // being unable to find a sufficient guard, for example), change the loop
2000 // comparison to use SLT instead of NE.
2001 Cond = OptimizeSMax(L, Cond, CondUse);
2003 // If possible, change stride and operands of the compare instruction to
2004 // eliminate one stride.
2005 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
2007 // It's possible for the setcc instruction to be anywhere in the loop, and
2008 // possible for it to have multiple users. If it is not immediately before
2009 // the latch block branch, move it.
2010 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
2011 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
2012 Cond->moveBefore(TermBr);
2014 // Otherwise, clone the terminating condition and insert into the loopend.
2015 Cond = cast<ICmpInst>(Cond->clone());
2016 Cond->setName(L->getHeader()->getName() + ".termcond");
2017 LatchBlock->getInstList().insert(TermBr, Cond);
2019 // Clone the IVUse, as the old use still exists!
2020 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
2021 CondUse->OperandValToReplace);
2022 CondUse = &IVUsesByStride[*CondStride].Users.back();
2026 // If we get to here, we know that we can transform the setcc instruction to
2027 // use the post-incremented version of the IV, allowing us to coalesce the
2028 // live ranges for the IV correctly.
2029 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
2030 CondUse->isUseOfPostIncrementedValue = true;
2034 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2036 LI = &getAnalysis<LoopInfo>();
2037 DT = &getAnalysis<DominatorTree>();
2038 SE = &getAnalysis<ScalarEvolution>();
2039 TD = &getAnalysis<TargetData>();
2040 UIntPtrTy = TD->getIntPtrType();
2043 // Find all uses of induction variables in this loop, and catagorize
2044 // them by stride. Start by finding all of the PHI nodes in the header for
2045 // this loop. If they are induction variables, inspect their uses.
2046 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2047 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2048 AddUsersIfInteresting(I, L, Processed);
2050 if (!IVUsesByStride.empty()) {
2051 // Optimize induction variables. Some indvar uses can be transformed to use
2052 // strides that will be needed for other purposes. A common example of this
2053 // is the exit test for the loop, which can often be rewritten to use the
2054 // computation of some other indvar to decide when to terminate the loop.
2057 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2058 // doing computation in byte values, promote to 32-bit values if safe.
2060 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2061 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2062 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2063 // Need to be careful that IV's are all the same type. Only works for
2064 // intptr_t indvars.
2066 // If we only have one stride, we can more aggressively eliminate some
2068 bool HasOneStride = IVUsesByStride.size() == 1;
2071 DOUT << "\nLSR on ";
2075 // IVsByStride keeps IVs for one particular loop.
2076 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2078 // Sort the StrideOrder so we process larger strides first.
2079 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2081 // Note: this processes each stride/type pair individually. All users
2082 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2083 // Also, note that we iterate over IVUsesByStride indirectly by using
2084 // StrideOrder. This extra layer of indirection makes the ordering of
2085 // strides deterministic - not dependent on map order.
2086 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2087 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2088 IVUsesByStride.find(StrideOrder[Stride]);
2089 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2090 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
2094 // We're done analyzing this loop; release all the state we built up for it.
2095 CastedPointers.clear();
2096 IVUsesByStride.clear();
2097 IVsByStride.clear();
2098 StrideOrder.clear();
2100 // Clean up after ourselves
2101 if (!DeadInsts.empty()) {
2102 DeleteTriviallyDeadInstructions();
2104 BasicBlock::iterator I = L->getHeader()->begin();
2105 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2106 // At this point, we know that we have killed one or more IV users.
2107 // It is worth checking to see if the cannonical indvar is also
2108 // dead, so that we can remove it as well.
2110 // We can remove a PHI if it is on a cycle in the def-use graph
2111 // where each node in the cycle has degree one, i.e. only one use,
2112 // and is an instruction with no side effects.
2114 // FIXME: this needs to eliminate an induction variable even if it's being
2115 // compared against some value to decide loop termination.
2116 if (!PN->hasOneUse())
2119 SmallPtrSet<PHINode *, 4> PHIs;
2120 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2121 J && J->hasOneUse() && !J->mayWriteToMemory();
2122 J = dyn_cast<Instruction>(*J->use_begin())) {
2123 // If we find the original PHI, we've discovered a cycle.
2125 // Break the cycle and mark the PHI for deletion.
2126 SE->deleteValueFromRecords(PN);
2127 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2128 DeadInsts.push_back(PN);
2132 // If we find a PHI more than once, we're on a cycle that
2133 // won't prove fruitful.
2134 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2138 DeleteTriviallyDeadInstructions();