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 std::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 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) {
602 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator();
603 InsertLoop = InsertLoop->getParentLoop();
606 // If there is no immediate value, skip the next part.
608 return Rewriter.expandCodeFor(NewBase, BaseInsertPt);
610 Value *Base = Rewriter.expandCodeFor(NewBase, BaseInsertPt);
612 // If we are inserting the base and imm values in the same block, make sure to
613 // adjust the IP position if insertion reused a result.
614 if (IP == BaseInsertPt)
615 IP = Rewriter.getInsertionPoint();
617 // Always emit the immediate (if non-zero) into the same block as the user.
618 SCEVHandle NewValSCEV = SE->getAddExpr(SE->getUnknown(Base), Imm);
619 return Rewriter.expandCodeFor(NewValSCEV, IP);
624 // Once we rewrite the code to insert the new IVs we want, update the
625 // operands of Inst to use the new expression 'NewBase', with 'Imm' added
626 // to it. NewBasePt is the last instruction which contributes to the
627 // value of NewBase in the case that it's a diffferent instruction from
628 // the PHI that NewBase is computed from, or null otherwise.
630 void BasedUser::RewriteInstructionToUseNewBase(const SCEVHandle &NewBase,
631 Instruction *NewBasePt,
632 SCEVExpander &Rewriter, Loop *L, Pass *P,
633 SmallVectorImpl<Instruction*> &DeadInsts){
634 if (!isa<PHINode>(Inst)) {
635 // By default, insert code at the user instruction.
636 BasicBlock::iterator InsertPt = Inst;
638 // However, if the Operand is itself an instruction, the (potentially
639 // complex) inserted code may be shared by many users. Because of this, we
640 // want to emit code for the computation of the operand right before its old
641 // computation. This is usually safe, because we obviously used to use the
642 // computation when it was computed in its current block. However, in some
643 // cases (e.g. use of a post-incremented induction variable) the NewBase
644 // value will be pinned to live somewhere after the original computation.
645 // In this case, we have to back off.
646 if (!isUseOfPostIncrementedValue) {
647 if (NewBasePt && isa<PHINode>(OperandValToReplace)) {
648 InsertPt = NewBasePt;
650 } else if (Instruction *OpInst
651 = dyn_cast<Instruction>(OperandValToReplace)) {
653 while (isa<PHINode>(InsertPt)) ++InsertPt;
656 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
657 // Adjust the type back to match the Inst. Note that we can't use InsertPt
658 // here because the SCEVExpander may have inserted the instructions after
659 // that point, in its efforts to avoid inserting redundant expressions.
660 if (isa<PointerType>(OperandValToReplace->getType())) {
661 NewVal = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
663 OperandValToReplace->getType());
665 // Replace the use of the operand Value with the new Phi we just created.
666 Inst->replaceUsesOfWith(OperandValToReplace, NewVal);
667 DOUT << " CHANGED: IMM =" << *Imm;
668 DOUT << " \tNEWBASE =" << *NewBase;
669 DOUT << " \tInst = " << *Inst;
673 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm
674 // expression into each operand block that uses it. Note that PHI nodes can
675 // have multiple entries for the same predecessor. We use a map to make sure
676 // that a PHI node only has a single Value* for each predecessor (which also
677 // prevents us from inserting duplicate code in some blocks).
678 DenseMap<BasicBlock*, Value*> InsertedCode;
679 PHINode *PN = cast<PHINode>(Inst);
680 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
681 if (PN->getIncomingValue(i) == OperandValToReplace) {
682 // If this is a critical edge, split the edge so that we do not insert the
683 // code on all predecessor/successor paths. We do this unless this is the
684 // canonical backedge for this loop, as this can make some inserted code
685 // be in an illegal position.
686 BasicBlock *PHIPred = PN->getIncomingBlock(i);
687 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 &&
688 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) {
690 // First step, split the critical edge.
691 SplitCriticalEdge(PHIPred, PN->getParent(), P, false);
693 // Next step: move the basic block. In particular, if the PHI node
694 // is outside of the loop, and PredTI is in the loop, we want to
695 // move the block to be immediately before the PHI block, not
696 // immediately after PredTI.
697 if (L->contains(PHIPred) && !L->contains(PN->getParent())) {
698 BasicBlock *NewBB = PN->getIncomingBlock(i);
699 NewBB->moveBefore(PN->getParent());
702 // Splitting the edge can reduce the number of PHI entries we have.
703 e = PN->getNumIncomingValues();
706 Value *&Code = InsertedCode[PN->getIncomingBlock(i)];
708 // Insert the code into the end of the predecessor block.
709 Instruction *InsertPt = PN->getIncomingBlock(i)->getTerminator();
710 Code = InsertCodeForBaseAtPosition(NewBase, Rewriter, InsertPt, L);
712 // Adjust the type back to match the PHI. Note that we can't use
713 // InsertPt here because the SCEVExpander may have inserted its
714 // instructions after that point, in its efforts to avoid inserting
715 // redundant expressions.
716 if (isa<PointerType>(PN->getType())) {
717 Code = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr,
723 // Replace the use of the operand Value with the new Phi we just created.
724 PN->setIncomingValue(i, Code);
729 // PHI node might have become a constant value after SplitCriticalEdge.
730 DeadInsts.push_back(Inst);
732 DOUT << " CHANGED: IMM =" << *Imm << " Inst = " << *Inst;
736 /// isTargetConstant - Return true if the following can be referenced by the
737 /// immediate field of a target instruction.
738 static bool isTargetConstant(const SCEVHandle &V, const Type *UseTy,
739 const TargetLowering *TLI) {
740 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) {
741 int64_t VC = SC->getValue()->getSExtValue();
743 TargetLowering::AddrMode AM;
745 return TLI->isLegalAddressingMode(AM, UseTy);
747 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field.
748 return (VC > -(1 << 16) && VC < (1 << 16)-1);
752 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V))
753 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue()))
754 if (TLI && CE->getOpcode() == Instruction::PtrToInt) {
755 Constant *Op0 = CE->getOperand(0);
756 if (GlobalValue *GV = dyn_cast<GlobalValue>(Op0)) {
757 TargetLowering::AddrMode AM;
759 return TLI->isLegalAddressingMode(AM, UseTy);
765 /// MoveLoopVariantsToImediateField - Move any subexpressions from Val that are
766 /// loop varying to the Imm operand.
767 static void MoveLoopVariantsToImediateField(SCEVHandle &Val, SCEVHandle &Imm,
768 Loop *L, ScalarEvolution *SE) {
769 if (Val->isLoopInvariant(L)) return; // Nothing to do.
771 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
772 std::vector<SCEVHandle> NewOps;
773 NewOps.reserve(SAE->getNumOperands());
775 for (unsigned i = 0; i != SAE->getNumOperands(); ++i)
776 if (!SAE->getOperand(i)->isLoopInvariant(L)) {
777 // If this is a loop-variant expression, it must stay in the immediate
778 // field of the expression.
779 Imm = SE->getAddExpr(Imm, SAE->getOperand(i));
781 NewOps.push_back(SAE->getOperand(i));
785 Val = SE->getIntegerSCEV(0, Val->getType());
787 Val = SE->getAddExpr(NewOps);
788 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
789 // Try to pull immediates out of the start value of nested addrec's.
790 SCEVHandle Start = SARE->getStart();
791 MoveLoopVariantsToImediateField(Start, Imm, L, SE);
793 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
795 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
797 // Otherwise, all of Val is variant, move the whole thing over.
798 Imm = SE->getAddExpr(Imm, Val);
799 Val = SE->getIntegerSCEV(0, Val->getType());
804 /// MoveImmediateValues - Look at Val, and pull out any additions of constants
805 /// that can fit into the immediate field of instructions in the target.
806 /// Accumulate these immediate values into the Imm value.
807 static void MoveImmediateValues(const TargetLowering *TLI,
809 SCEVHandle &Val, SCEVHandle &Imm,
810 bool isAddress, Loop *L,
811 ScalarEvolution *SE) {
812 const Type *UseTy = User->getType();
813 if (StoreInst *SI = dyn_cast<StoreInst>(User))
814 UseTy = SI->getOperand(0)->getType();
816 if (SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) {
817 std::vector<SCEVHandle> NewOps;
818 NewOps.reserve(SAE->getNumOperands());
820 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) {
821 SCEVHandle NewOp = SAE->getOperand(i);
822 MoveImmediateValues(TLI, User, NewOp, Imm, isAddress, L, SE);
824 if (!NewOp->isLoopInvariant(L)) {
825 // If this is a loop-variant expression, it must stay in the immediate
826 // field of the expression.
827 Imm = SE->getAddExpr(Imm, NewOp);
829 NewOps.push_back(NewOp);
834 Val = SE->getIntegerSCEV(0, Val->getType());
836 Val = SE->getAddExpr(NewOps);
838 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) {
839 // Try to pull immediates out of the start value of nested addrec's.
840 SCEVHandle Start = SARE->getStart();
841 MoveImmediateValues(TLI, User, Start, Imm, isAddress, L, SE);
843 if (Start != SARE->getStart()) {
844 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
846 Val = SE->getAddRecExpr(Ops, SARE->getLoop());
849 } else if (SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) {
850 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field.
851 if (isAddress && isTargetConstant(SME->getOperand(0), UseTy, TLI) &&
852 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) {
854 SCEVHandle SubImm = SE->getIntegerSCEV(0, Val->getType());
855 SCEVHandle NewOp = SME->getOperand(1);
856 MoveImmediateValues(TLI, User, NewOp, SubImm, isAddress, L, SE);
858 // If we extracted something out of the subexpressions, see if we can
860 if (NewOp != SME->getOperand(1)) {
861 // Scale SubImm up by "8". If the result is a target constant, we are
863 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0));
864 if (isTargetConstant(SubImm, UseTy, TLI)) {
865 // Accumulate the immediate.
866 Imm = SE->getAddExpr(Imm, SubImm);
868 // Update what is left of 'Val'.
869 Val = SE->getMulExpr(SME->getOperand(0), NewOp);
876 // Loop-variant expressions must stay in the immediate field of the
878 if ((isAddress && isTargetConstant(Val, UseTy, TLI)) ||
879 !Val->isLoopInvariant(L)) {
880 Imm = SE->getAddExpr(Imm, Val);
881 Val = SE->getIntegerSCEV(0, Val->getType());
885 // Otherwise, no immediates to move.
889 /// SeparateSubExprs - Decompose Expr into all of the subexpressions that are
890 /// added together. This is used to reassociate common addition subexprs
891 /// together for maximal sharing when rewriting bases.
892 static void SeparateSubExprs(std::vector<SCEVHandle> &SubExprs,
894 ScalarEvolution *SE) {
895 if (SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) {
896 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j)
897 SeparateSubExprs(SubExprs, AE->getOperand(j), SE);
898 } else if (SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) {
899 SCEVHandle Zero = SE->getIntegerSCEV(0, Expr->getType());
900 if (SARE->getOperand(0) == Zero) {
901 SubExprs.push_back(Expr);
903 // Compute the addrec with zero as its base.
904 std::vector<SCEVHandle> Ops(SARE->op_begin(), SARE->op_end());
905 Ops[0] = Zero; // Start with zero base.
906 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop()));
909 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE);
911 } else if (!Expr->isZero()) {
913 SubExprs.push_back(Expr);
918 /// RemoveCommonExpressionsFromUseBases - Look through all of the uses in Bases,
919 /// removing any common subexpressions from it. Anything truly common is
920 /// removed, accumulated, and returned. This looks for things like (a+b+c) and
921 /// (a+c+d) -> (a+c). The common expression is *removed* from the Bases.
923 RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses,
924 ScalarEvolution *SE) {
925 unsigned NumUses = Uses.size();
927 // Only one use? Use its base, regardless of what it is!
928 SCEVHandle Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType());
929 SCEVHandle Result = Zero;
931 std::swap(Result, Uses[0].Base);
935 // To find common subexpressions, count how many of Uses use each expression.
936 // If any subexpressions are used Uses.size() times, they are common.
937 std::map<SCEVHandle, unsigned> SubExpressionUseCounts;
939 // UniqueSubExprs - Keep track of all of the subexpressions we see in the
940 // order we see them.
941 std::vector<SCEVHandle> UniqueSubExprs;
943 std::vector<SCEVHandle> SubExprs;
944 for (unsigned i = 0; i != NumUses; ++i) {
945 // If the base is zero (which is common), return zero now, there are no
947 if (Uses[i].Base == Zero) return Zero;
949 // Split the expression into subexprs.
950 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
951 // Add one to SubExpressionUseCounts for each subexpr present.
952 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
953 if (++SubExpressionUseCounts[SubExprs[j]] == 1)
954 UniqueSubExprs.push_back(SubExprs[j]);
958 // Now that we know how many times each is used, build Result. Iterate over
959 // UniqueSubexprs so that we have a stable ordering.
960 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) {
961 std::map<SCEVHandle, unsigned>::iterator I =
962 SubExpressionUseCounts.find(UniqueSubExprs[i]);
963 assert(I != SubExpressionUseCounts.end() && "Entry not found?");
964 if (I->second == NumUses) { // Found CSE!
965 Result = SE->getAddExpr(Result, I->first);
967 // Remove non-cse's from SubExpressionUseCounts.
968 SubExpressionUseCounts.erase(I);
972 // If we found no CSE's, return now.
973 if (Result == Zero) return Result;
975 // Otherwise, remove all of the CSE's we found from each of the base values.
976 for (unsigned i = 0; i != NumUses; ++i) {
977 // Split the expression into subexprs.
978 SeparateSubExprs(SubExprs, Uses[i].Base, SE);
980 // Remove any common subexpressions.
981 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j)
982 if (SubExpressionUseCounts.count(SubExprs[j])) {
983 SubExprs.erase(SubExprs.begin()+j);
987 // Finally, the non-shared expressions together.
988 if (SubExprs.empty())
991 Uses[i].Base = SE->getAddExpr(SubExprs);
998 /// ValidStride - Check whether the given Scale is valid for all loads and
999 /// stores in UsersToProcess.
1001 bool LoopStrengthReduce::ValidStride(bool HasBaseReg,
1003 const std::vector<BasedUser>& UsersToProcess) {
1007 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) {
1008 // If this is a load or other access, pass the type of the access in.
1009 const Type *AccessTy = Type::VoidTy;
1010 if (StoreInst *SI = dyn_cast<StoreInst>(UsersToProcess[i].Inst))
1011 AccessTy = SI->getOperand(0)->getType();
1012 else if (LoadInst *LI = dyn_cast<LoadInst>(UsersToProcess[i].Inst))
1013 AccessTy = LI->getType();
1014 else if (isa<PHINode>(UsersToProcess[i].Inst))
1017 TargetLowering::AddrMode AM;
1018 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm))
1019 AM.BaseOffs = SC->getValue()->getSExtValue();
1020 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero();
1023 // If load[imm+r*scale] is illegal, bail out.
1024 if (!TLI->isLegalAddressingMode(AM, AccessTy))
1030 /// RequiresTypeConversion - Returns true if converting Ty to NewTy is not
1032 bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1,
1036 if (TLI && TLI->isTruncateFree(Ty1, Ty2))
1038 return (!Ty1->canLosslesslyBitCastTo(Ty2) &&
1039 !(isa<PointerType>(Ty2) &&
1040 Ty1->canLosslesslyBitCastTo(UIntPtrTy)) &&
1041 !(isa<PointerType>(Ty1) &&
1042 Ty2->canLosslesslyBitCastTo(UIntPtrTy)));
1045 /// CheckForIVReuse - Returns the multiple if the stride is the multiple
1046 /// of a previous stride and it is a legal value for the target addressing
1047 /// mode scale component and optional base reg. This allows the users of
1048 /// this stride to be rewritten as prev iv * factor. It returns 0 if no
1049 /// reuse is possible.
1050 unsigned LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg,
1051 bool AllUsesAreAddresses,
1052 const SCEVHandle &Stride,
1053 IVExpr &IV, const Type *Ty,
1054 const std::vector<BasedUser>& UsersToProcess) {
1055 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) {
1056 int64_t SInt = SC->getValue()->getSExtValue();
1057 for (unsigned NewStride = 0, e = StrideOrder.size(); NewStride != e;
1059 std::map<SCEVHandle, IVsOfOneStride>::iterator SI =
1060 IVsByStride.find(StrideOrder[NewStride]);
1061 if (SI == IVsByStride.end())
1063 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1064 if (SI->first != Stride &&
1065 (unsigned(abs(SInt)) < SSInt || (SInt % SSInt) != 0))
1067 int64_t Scale = SInt / SSInt;
1068 // Check that this stride is valid for all the types used for loads and
1069 // stores; if it can be used for some and not others, we might as well use
1070 // the original stride everywhere, since we have to create the IV for it
1071 // anyway. If the scale is 1, then we don't need to worry about folding
1074 (AllUsesAreAddresses &&
1075 ValidStride(HasBaseReg, Scale, UsersToProcess)))
1076 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(),
1077 IE = SI->second.IVs.end(); II != IE; ++II)
1078 // FIXME: Only handle base == 0 for now.
1079 // Only reuse previous IV if it would not require a type conversion.
1080 if (II->Base->isZero() &&
1081 !RequiresTypeConversion(II->Base->getType(), Ty)) {
1090 /// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that
1091 /// returns true if Val's isUseOfPostIncrementedValue is true.
1092 static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) {
1093 return Val.isUseOfPostIncrementedValue;
1096 /// isNonConstantNegative - Return true if the specified scev is negated, but
1098 static bool isNonConstantNegative(const SCEVHandle &Expr) {
1099 SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr);
1100 if (!Mul) return false;
1102 // If there is a constant factor, it will be first.
1103 SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0));
1104 if (!SC) return false;
1106 // Return true if the value is negative, this matches things like (-42 * V).
1107 return SC->getValue()->getValue().isNegative();
1110 /// isAddress - Returns true if the specified instruction is using the
1111 /// specified value as an address.
1112 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
1113 bool isAddress = isa<LoadInst>(Inst);
1114 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1115 if (SI->getOperand(1) == OperandVal)
1117 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
1118 // Addressing modes can also be folded into prefetches and a variety
1120 switch (II->getIntrinsicID()) {
1122 case Intrinsic::prefetch:
1123 case Intrinsic::x86_sse2_loadu_dq:
1124 case Intrinsic::x86_sse2_loadu_pd:
1125 case Intrinsic::x86_sse_loadu_ps:
1126 case Intrinsic::x86_sse_storeu_ps:
1127 case Intrinsic::x86_sse2_storeu_pd:
1128 case Intrinsic::x86_sse2_storeu_dq:
1129 case Intrinsic::x86_sse2_storel_dq:
1130 if (II->getOperand(1) == OperandVal)
1138 // CollectIVUsers - Transform our list of users and offsets to a bit more
1139 // complex table. In this new vector, each 'BasedUser' contains 'Base', the base
1140 // of the strided accesses, as well as the old information from Uses. We
1141 // progressively move information from the Base field to the Imm field, until
1142 // we eventually have the full access expression to rewrite the use.
1143 SCEVHandle LoopStrengthReduce::CollectIVUsers(const SCEVHandle &Stride,
1144 IVUsersOfOneStride &Uses,
1146 bool &AllUsesAreAddresses,
1147 std::vector<BasedUser> &UsersToProcess) {
1148 UsersToProcess.reserve(Uses.Users.size());
1149 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) {
1150 UsersToProcess.push_back(BasedUser(Uses.Users[i], SE));
1152 // Move any loop invariant operands from the offset field to the immediate
1153 // field of the use, so that we don't try to use something before it is
1155 MoveLoopVariantsToImediateField(UsersToProcess.back().Base,
1156 UsersToProcess.back().Imm, L, SE);
1157 assert(UsersToProcess.back().Base->isLoopInvariant(L) &&
1158 "Base value is not loop invariant!");
1161 // We now have a whole bunch of uses of like-strided induction variables, but
1162 // they might all have different bases. We want to emit one PHI node for this
1163 // stride which we fold as many common expressions (between the IVs) into as
1164 // possible. Start by identifying the common expressions in the base values
1165 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find
1166 // "A+B"), emit it to the preheader, then remove the expression from the
1167 // UsersToProcess base values.
1168 SCEVHandle CommonExprs =
1169 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE);
1171 // Next, figure out what we can represent in the immediate fields of
1172 // instructions. If we can represent anything there, move it to the imm
1173 // fields of the BasedUsers. We do this so that it increases the commonality
1174 // of the remaining uses.
1175 unsigned NumPHI = 0;
1176 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1177 // If the user is not in the current loop, this means it is using the exit
1178 // value of the IV. Do not put anything in the base, make sure it's all in
1179 // the immediate field to allow as much factoring as possible.
1180 if (!L->contains(UsersToProcess[i].Inst->getParent())) {
1181 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm,
1182 UsersToProcess[i].Base);
1183 UsersToProcess[i].Base =
1184 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType());
1187 // Addressing modes can be folded into loads and stores. Be careful that
1188 // the store is through the expression, not of the expression though.
1190 bool isAddress = isAddressUse(UsersToProcess[i].Inst,
1191 UsersToProcess[i].OperandValToReplace);
1192 if (isa<PHINode>(UsersToProcess[i].Inst)) {
1197 // If this use isn't an address, then not all uses are addresses.
1198 if (!isAddress && !isPHI)
1199 AllUsesAreAddresses = false;
1201 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base,
1202 UsersToProcess[i].Imm, isAddress, L, SE);
1206 // If one of the use if a PHI node and all other uses are addresses, still
1207 // allow iv reuse. Essentially we are trading one constant multiplication
1208 // for one fewer iv.
1210 AllUsesAreAddresses = false;
1215 /// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single
1216 /// stride of IV. All of the users may have different starting values, and this
1217 /// may not be the only stride (we know it is if isOnlyStride is true).
1218 void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEVHandle &Stride,
1219 IVUsersOfOneStride &Uses,
1221 bool isOnlyStride) {
1222 // If all the users are moved to another stride, then there is nothing to do.
1223 if (Uses.Users.empty())
1226 // Keep track if every use in UsersToProcess is an address. If they all are,
1227 // we may be able to rewrite the entire collection of them in terms of a
1228 // smaller-stride IV.
1229 bool AllUsesAreAddresses = true;
1231 // Transform our list of users and offsets to a bit more complex table. In
1232 // this new vector, each 'BasedUser' contains 'Base' the base of the
1233 // strided accessas well as the old information from Uses. We progressively
1234 // move information from the Base field to the Imm field, until we eventually
1235 // have the full access expression to rewrite the use.
1236 std::vector<BasedUser> UsersToProcess;
1237 SCEVHandle CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses,
1240 // If we managed to find some expressions in common, we'll need to carry
1241 // their value in a register and add it in for each use. This will take up
1242 // a register operand, which potentially restricts what stride values are
1244 bool HaveCommonExprs = !CommonExprs->isZero();
1246 // If all uses are addresses, check if it is possible to reuse an IV with a
1247 // stride that is a factor of this stride. And that the multiple is a number
1248 // that can be encoded in the scale field of the target addressing mode. And
1249 // that we will have a valid instruction after this substition, including the
1250 // immediate field, if any.
1251 PHINode *NewPHI = NULL;
1253 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty),
1254 SE->getIntegerSCEV(0, Type::Int32Ty),
1256 unsigned RewriteFactor = 0;
1257 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses,
1258 Stride, ReuseIV, CommonExprs->getType(),
1260 if (RewriteFactor != 0) {
1261 DOUT << "BASED ON IV of STRIDE " << *ReuseIV.Stride
1262 << " and BASE " << *ReuseIV.Base << " :\n";
1263 NewPHI = ReuseIV.PHI;
1264 IncV = ReuseIV.IncV;
1267 const Type *ReplacedTy = CommonExprs->getType();
1269 // Now that we know what we need to do, insert the PHI node itself.
1271 DOUT << "INSERTING IV of TYPE " << *ReplacedTy << " of STRIDE "
1272 << *Stride << " and BASE " << *CommonExprs << ": ";
1274 SCEVExpander Rewriter(*SE, *LI);
1275 SCEVExpander PreheaderRewriter(*SE, *LI);
1277 BasicBlock *Preheader = L->getLoopPreheader();
1278 Instruction *PreInsertPt = Preheader->getTerminator();
1279 Instruction *PhiInsertBefore = L->getHeader()->begin();
1281 BasicBlock *LatchBlock = L->getLoopLatch();
1284 // Emit the initial base value into the loop preheader.
1286 = PreheaderRewriter.expandCodeFor(CommonExprs, PreInsertPt);
1288 if (RewriteFactor == 0) {
1289 // Create a new Phi for this base, and stick it in the loop header.
1290 NewPHI = PHINode::Create(ReplacedTy, "iv.", PhiInsertBefore);
1293 // Add common base to the new Phi node.
1294 NewPHI->addIncoming(CommonBaseV, Preheader);
1296 // If the stride is negative, insert a sub instead of an add for the
1298 bool isNegative = isNonConstantNegative(Stride);
1299 SCEVHandle IncAmount = Stride;
1301 IncAmount = SE->getNegativeSCEV(Stride);
1303 // Insert the stride into the preheader.
1304 Value *StrideV = PreheaderRewriter.expandCodeFor(IncAmount, PreInsertPt);
1305 if (!isa<ConstantInt>(StrideV)) ++NumVariable;
1307 // Emit the increment of the base value before the terminator of the loop
1308 // latch block, and add it to the Phi node.
1309 SCEVHandle IncExp = SE->getUnknown(StrideV);
1311 IncExp = SE->getNegativeSCEV(IncExp);
1312 IncExp = SE->getAddExpr(SE->getUnknown(NewPHI), IncExp);
1314 IncV = Rewriter.expandCodeFor(IncExp, LatchBlock->getTerminator());
1315 IncV->setName(NewPHI->getName()+".inc");
1316 NewPHI->addIncoming(IncV, LatchBlock);
1318 // Remember this in case a later stride is multiple of this.
1319 IVsByStride[Stride].addIV(Stride, CommonExprs, NewPHI, IncV);
1321 DOUT << " IV=%" << NewPHI->getNameStr() << " INC=%" << IncV->getNameStr();
1323 Constant *C = dyn_cast<Constant>(CommonBaseV);
1325 (!C->isNullValue() &&
1326 !isTargetConstant(SE->getUnknown(CommonBaseV), ReplacedTy, TLI)))
1327 // We want the common base emitted into the preheader! This is just
1328 // using cast as a copy so BitCast (no-op cast) is appropriate
1329 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(),
1330 "commonbase", PreInsertPt);
1334 // We want to emit code for users inside the loop first. To do this, we
1335 // rearrange BasedUser so that the entries at the end have
1336 // isUseOfPostIncrementedValue = false, because we pop off the end of the
1337 // vector (so we handle them first).
1338 std::partition(UsersToProcess.begin(), UsersToProcess.end(),
1339 PartitionByIsUseOfPostIncrementedValue);
1341 // Sort this by base, so that things with the same base are handled
1342 // together. By partitioning first and stable-sorting later, we are
1343 // guaranteed that within each base we will pop off users from within the
1344 // loop before users outside of the loop with a particular base.
1346 // We would like to use stable_sort here, but we can't. The problem is that
1347 // SCEVHandle's don't have a deterministic ordering w.r.t to each other, so
1348 // we don't have anything to do a '<' comparison on. Because we think the
1349 // number of uses is small, do a horrible bubble sort which just relies on
1351 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) {
1352 // Get a base value.
1353 SCEVHandle Base = UsersToProcess[i].Base;
1355 // Compact everything with this base to be consequtive with this one.
1356 for (unsigned j = i+1; j != e; ++j) {
1357 if (UsersToProcess[j].Base == Base) {
1358 std::swap(UsersToProcess[i+1], UsersToProcess[j]);
1364 // Process all the users now. This outer loop handles all bases, the inner
1365 // loop handles all users of a particular base.
1366 while (!UsersToProcess.empty()) {
1367 SCEVHandle Base = UsersToProcess.back().Base;
1369 // Emit the code for Base into the preheader.
1370 Value *BaseV = PreheaderRewriter.expandCodeFor(Base, PreInsertPt);
1372 DOUT << " INSERTING code for BASE = " << *Base << ":";
1373 if (BaseV->hasName())
1374 DOUT << " Result value name = %" << BaseV->getNameStr();
1377 // If BaseV is a constant other than 0, make sure that it gets inserted into
1378 // the preheader, instead of being forward substituted into the uses. We do
1379 // this by forcing a BitCast (noop cast) to be inserted into the preheader
1381 if (Constant *C = dyn_cast<Constant>(BaseV)) {
1382 if (!C->isNullValue() && !isTargetConstant(Base, ReplacedTy, TLI)) {
1383 // We want this constant emitted into the preheader! This is just
1384 // using cast as a copy so BitCast (no-op cast) is appropriate
1385 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert",
1390 // Emit the code to add the immediate offset to the Phi value, just before
1391 // the instructions that we identified as using this stride and base.
1393 // FIXME: Use emitted users to emit other users.
1394 BasedUser &User = UsersToProcess.back();
1396 // If this instruction wants to use the post-incremented value, move it
1397 // after the post-inc and use its value instead of the PHI.
1398 Value *RewriteOp = NewPHI;
1399 if (User.isUseOfPostIncrementedValue) {
1402 // If this user is in the loop, make sure it is the last thing in the
1403 // loop to ensure it is dominated by the increment.
1404 if (L->contains(User.Inst->getParent()))
1405 User.Inst->moveBefore(LatchBlock->getTerminator());
1407 if (RewriteOp->getType() != ReplacedTy) {
1408 Instruction::CastOps opcode = Instruction::Trunc;
1409 if (ReplacedTy->getPrimitiveSizeInBits() ==
1410 RewriteOp->getType()->getPrimitiveSizeInBits())
1411 opcode = Instruction::BitCast;
1412 RewriteOp = SCEVExpander::InsertCastOfTo(opcode, RewriteOp, ReplacedTy);
1415 SCEVHandle RewriteExpr = SE->getUnknown(RewriteOp);
1417 // If we had to insert new instrutions for RewriteOp, we have to
1418 // consider that they may not have been able to end up immediately
1419 // next to RewriteOp, because non-PHI instructions may never precede
1420 // PHI instructions in a block. In this case, remember where the last
1421 // instruction was inserted so that if we're replacing a different
1422 // PHI node, we can use the later point to expand the final
1424 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp);
1425 if (RewriteOp == NewPHI) NewBasePt = 0;
1427 // Clear the SCEVExpander's expression map so that we are guaranteed
1428 // to have the code emitted where we expect it.
1431 // If we are reusing the iv, then it must be multiplied by a constant
1432 // factor take advantage of addressing mode scale component.
1433 if (RewriteFactor != 0) {
1434 RewriteExpr = SE->getMulExpr(SE->getIntegerSCEV(RewriteFactor,
1435 RewriteExpr->getType()),
1438 // The common base is emitted in the loop preheader. But since we
1439 // are reusing an IV, it has not been used to initialize the PHI node.
1440 // Add it to the expression used to rewrite the uses.
1441 if (!isa<ConstantInt>(CommonBaseV) ||
1442 !cast<ConstantInt>(CommonBaseV)->isZero())
1443 RewriteExpr = SE->getAddExpr(RewriteExpr,
1444 SE->getUnknown(CommonBaseV));
1447 // Now that we know what we need to do, insert code before User for the
1448 // immediate and any loop-variant expressions.
1449 if (!isa<ConstantInt>(BaseV) || !cast<ConstantInt>(BaseV)->isZero())
1450 // Add BaseV to the PHI value if needed.
1451 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV));
1453 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt,
1457 // Mark old value we replaced as possibly dead, so that it is eliminated
1458 // if we just replaced the last use of that value.
1459 DeadInsts.push_back(cast<Instruction>(User.OperandValToReplace));
1461 UsersToProcess.pop_back();
1464 // If there are any more users to process with the same base, process them
1465 // now. We sorted by base above, so we just have to check the last elt.
1466 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base);
1467 // TODO: Next, find out which base index is the most common, pull it out.
1470 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but
1471 // different starting values, into different PHIs.
1474 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1475 /// set the IV user and stride information and return true, otherwise return
1477 bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse,
1478 const SCEVHandle *&CondStride) {
1479 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e && !CondUse;
1481 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1482 IVUsesByStride.find(StrideOrder[Stride]);
1483 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1485 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1486 E = SI->second.Users.end(); UI != E; ++UI)
1487 if (UI->User == Cond) {
1488 // NOTE: we could handle setcc instructions with multiple uses here, but
1489 // InstCombine does it as well for simple uses, it's not clear that it
1490 // occurs enough in real life to handle.
1492 CondStride = &SI->first;
1500 // Constant strides come first which in turns are sorted by their absolute
1501 // values. If absolute values are the same, then positive strides comes first.
1503 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X
1504 struct StrideCompare {
1505 bool operator()(const SCEVHandle &LHS, const SCEVHandle &RHS) {
1506 SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS);
1507 SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS);
1509 int64_t LV = LHSC->getValue()->getSExtValue();
1510 int64_t RV = RHSC->getValue()->getSExtValue();
1511 uint64_t ALV = (LV < 0) ? -LV : LV;
1512 uint64_t ARV = (RV < 0) ? -RV : RV;
1518 return (LHSC && !RHSC);
1523 /// ChangeCompareStride - If a loop termination compare instruction is the
1524 /// only use of its stride, and the compaison is against a constant value,
1525 /// try eliminate the stride by moving the compare instruction to another
1526 /// stride and change its constant operand accordingly. e.g.
1532 /// if (v2 < 10) goto loop
1537 /// if (v1 < 30) goto loop
1538 ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond,
1539 IVStrideUse* &CondUse,
1540 const SCEVHandle* &CondStride) {
1541 if (StrideOrder.size() < 2 ||
1542 IVUsesByStride[*CondStride].Users.size() != 1)
1544 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride);
1545 if (!SC) return Cond;
1546 ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1));
1547 if (!C) return Cond;
1549 ICmpInst::Predicate Predicate = Cond->getPredicate();
1550 int64_t CmpSSInt = SC->getValue()->getSExtValue();
1551 int64_t CmpVal = C->getValue().getSExtValue();
1552 unsigned BitWidth = C->getValue().getBitWidth();
1553 uint64_t SignBit = 1ULL << (BitWidth-1);
1554 const Type *CmpTy = C->getType();
1555 const Type *NewCmpTy = NULL;
1556 unsigned TyBits = CmpTy->getPrimitiveSizeInBits();
1557 unsigned NewTyBits = 0;
1558 int64_t NewCmpVal = CmpVal;
1559 SCEVHandle *NewStride = NULL;
1560 Value *NewIncV = NULL;
1563 // Check stride constant and the comparision constant signs to detect
1565 if ((CmpVal & SignBit) != (CmpSSInt & SignBit))
1568 // Look for a suitable stride / iv as replacement.
1569 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
1570 for (unsigned i = 0, e = StrideOrder.size(); i != e; ++i) {
1571 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1572 IVUsesByStride.find(StrideOrder[i]);
1573 if (!isa<SCEVConstant>(SI->first))
1575 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue();
1576 if (abs(SSInt) <= abs(CmpSSInt) || (SSInt % CmpSSInt) != 0)
1579 Scale = SSInt / CmpSSInt;
1580 NewCmpVal = CmpVal * Scale;
1581 APInt Mul = APInt(BitWidth, NewCmpVal);
1582 // Check for overflow.
1583 if (Mul.getSExtValue() != NewCmpVal) {
1588 // Watch out for overflow.
1589 if (ICmpInst::isSignedPredicate(Predicate) &&
1590 (CmpVal & SignBit) != (NewCmpVal & SignBit))
1593 if (NewCmpVal != CmpVal) {
1594 // Pick the best iv to use trying to avoid a cast.
1596 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1597 E = SI->second.Users.end(); UI != E; ++UI) {
1598 NewIncV = UI->OperandValToReplace;
1599 if (NewIncV->getType() == CmpTy)
1607 NewCmpTy = NewIncV->getType();
1608 NewTyBits = isa<PointerType>(NewCmpTy)
1609 ? UIntPtrTy->getPrimitiveSizeInBits()
1610 : NewCmpTy->getPrimitiveSizeInBits();
1611 if (RequiresTypeConversion(NewCmpTy, CmpTy)) {
1612 // Check if it is possible to rewrite it using
1613 // an iv / stride of a smaller integer type.
1614 bool TruncOk = false;
1615 if (NewCmpTy->isInteger()) {
1616 unsigned Bits = NewTyBits;
1617 if (ICmpInst::isSignedPredicate(Predicate))
1619 uint64_t Mask = (1ULL << Bits) - 1;
1620 if (((uint64_t)NewCmpVal & Mask) == (uint64_t)NewCmpVal)
1629 // Don't rewrite if use offset is non-constant and the new type is
1630 // of a different type.
1631 // FIXME: too conservative?
1632 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->Offset)) {
1637 bool AllUsesAreAddresses = true;
1638 std::vector<BasedUser> UsersToProcess;
1639 SCEVHandle CommonExprs = CollectIVUsers(SI->first, SI->second, L,
1640 AllUsesAreAddresses,
1642 // Avoid rewriting the compare instruction with an iv of new stride
1643 // if it's likely the new stride uses will be rewritten using the
1644 if (AllUsesAreAddresses &&
1645 ValidStride(!CommonExprs->isZero(), Scale, UsersToProcess)) {
1650 // If scale is negative, use swapped predicate unless it's testing
1652 if (Scale < 0 && !Cond->isEquality())
1653 Predicate = ICmpInst::getSwappedPredicate(Predicate);
1655 NewStride = &StrideOrder[i];
1660 // Forgo this transformation if it the increment happens to be
1661 // unfortunately positioned after the condition, and the condition
1662 // has multiple uses which prevent it from being moved immediately
1663 // before the branch. See
1664 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll
1665 // for an example of this situation.
1666 if (!Cond->hasOneUse()) {
1667 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end();
1673 if (NewCmpVal != CmpVal) {
1674 // Create a new compare instruction using new stride / iv.
1675 ICmpInst *OldCond = Cond;
1677 if (!isa<PointerType>(NewCmpTy))
1678 RHS = ConstantInt::get(NewCmpTy, NewCmpVal);
1680 RHS = ConstantInt::get(UIntPtrTy, NewCmpVal);
1681 RHS = SCEVExpander::InsertCastOfTo(Instruction::IntToPtr, RHS, NewCmpTy);
1683 // Insert new compare instruction.
1684 Cond = new ICmpInst(Predicate, NewIncV, RHS,
1685 L->getHeader()->getName() + ".termcond",
1688 // Remove the old compare instruction. The old indvar is probably dead too.
1689 DeadInsts.push_back(cast<Instruction>(CondUse->OperandValToReplace));
1690 SE->deleteValueFromRecords(OldCond);
1691 OldCond->replaceAllUsesWith(Cond);
1692 OldCond->eraseFromParent();
1694 IVUsesByStride[*CondStride].Users.pop_back();
1695 SCEVHandle NewOffset = TyBits == NewTyBits
1696 ? SE->getMulExpr(CondUse->Offset,
1697 SE->getConstant(ConstantInt::get(CmpTy, Scale)))
1698 : SE->getConstant(ConstantInt::get(NewCmpTy,
1699 cast<SCEVConstant>(CondUse->Offset)->getValue()->getSExtValue()*Scale));
1700 IVUsesByStride[*NewStride].addUser(NewOffset, Cond, NewIncV);
1701 CondUse = &IVUsesByStride[*NewStride].Users.back();
1702 CondStride = NewStride;
1709 /// OptimizeSMax - Rewrite the loop's terminating condition if it uses
1710 /// an smax computation.
1712 /// This is a narrow solution to a specific, but acute, problem. For loops
1718 /// } while (++i < n);
1720 /// where the comparison is signed, the trip count isn't just 'n', because
1721 /// 'n' could be negative. And unfortunately this can come up even for loops
1722 /// where the user didn't use a C do-while loop. For example, seemingly
1723 /// well-behaved top-test loops will commonly be lowered like this:
1729 /// } while (++i < n);
1732 /// and then it's possible for subsequent optimization to obscure the if
1733 /// test in such a way that indvars can't find it.
1735 /// When indvars can't find the if test in loops like this, it creates a
1736 /// signed-max expression, which allows it to give the loop a canonical
1737 /// induction variable:
1740 /// smax = n < 1 ? 1 : n;
1743 /// } while (++i != smax);
1745 /// Canonical induction variables are necessary because the loop passes
1746 /// are designed around them. The most obvious example of this is the
1747 /// LoopInfo analysis, which doesn't remember trip count values. It
1748 /// expects to be able to rediscover the trip count each time it is
1749 /// needed, and it does this using a simple analyis that only succeeds if
1750 /// the loop has a canonical induction variable.
1752 /// However, when it comes time to generate code, the maximum operation
1753 /// can be quite costly, especially if it's inside of an outer loop.
1755 /// This function solves this problem by detecting this type of loop and
1756 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1757 /// the instructions for the maximum computation.
1759 ICmpInst *LoopStrengthReduce::OptimizeSMax(Loop *L, ICmpInst *Cond,
1760 IVStrideUse* &CondUse) {
1761 // Check that the loop matches the pattern we're looking for.
1762 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1763 Cond->getPredicate() != CmpInst::ICMP_NE)
1766 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1767 if (!Sel || !Sel->hasOneUse()) return Cond;
1769 SCEVHandle IterationCount = SE->getIterationCount(L);
1770 if (isa<SCEVCouldNotCompute>(IterationCount))
1772 SCEVHandle One = SE->getIntegerSCEV(1, IterationCount->getType());
1774 // Adjust for an annoying getIterationCount quirk.
1775 IterationCount = SE->getAddExpr(IterationCount, One);
1777 // Check for a max calculation that matches the pattern.
1778 SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(IterationCount);
1779 if (!SMax || SMax != SE->getSCEV(Sel)) return Cond;
1781 SCEVHandle SMaxLHS = SMax->getOperand(0);
1782 SCEVHandle SMaxRHS = SMax->getOperand(1);
1783 if (!SMaxLHS || SMaxLHS != One) return Cond;
1785 // Check the relevant induction variable for conformance to
1787 SCEVHandle IV = SE->getSCEV(Cond->getOperand(0));
1788 SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1789 if (!AR || !AR->isAffine() ||
1790 AR->getStart() != One ||
1791 AR->getStepRecurrence(*SE) != One)
1794 // Check the right operand of the select, and remember it, as it will
1795 // be used in the new comparison instruction.
1797 if (SE->getSCEV(Sel->getOperand(1)) == SMaxRHS)
1798 NewRHS = Sel->getOperand(1);
1799 else if (SE->getSCEV(Sel->getOperand(2)) == SMaxRHS)
1800 NewRHS = Sel->getOperand(2);
1801 if (!NewRHS) return Cond;
1803 // Ok, everything looks ok to change the condition into an SLT or SGE and
1804 // delete the max calculation.
1806 new ICmpInst(Cond->getPredicate() == CmpInst::ICMP_NE ?
1809 Cond->getOperand(0), NewRHS, "scmp", Cond);
1811 // Delete the max calculation instructions.
1812 SE->deleteValueFromRecords(Cond);
1813 Cond->replaceAllUsesWith(NewCond);
1814 Cond->eraseFromParent();
1815 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1816 SE->deleteValueFromRecords(Sel);
1817 Sel->eraseFromParent();
1818 if (Cmp->use_empty()) {
1819 SE->deleteValueFromRecords(Cmp);
1820 Cmp->eraseFromParent();
1822 CondUse->User = NewCond;
1826 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1827 /// inside the loop then try to eliminate the cast opeation.
1828 void LoopStrengthReduce::OptimizeShadowIV(Loop *L) {
1830 SCEVHandle IterationCount = SE->getIterationCount(L);
1831 if (isa<SCEVCouldNotCompute>(IterationCount))
1834 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e;
1836 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
1837 IVUsesByStride.find(StrideOrder[Stride]);
1838 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
1839 if (!isa<SCEVConstant>(SI->first))
1842 for (std::vector<IVStrideUse>::iterator UI = SI->second.Users.begin(),
1843 E = SI->second.Users.end(); UI != E; /* empty */) {
1844 std::vector<IVStrideUse>::iterator CandidateUI = UI;
1846 Instruction *ShadowUse = CandidateUI->User;
1847 const Type *DestTy = NULL;
1849 /* If shadow use is a int->float cast then insert a second IV
1850 to eliminate this cast.
1852 for (unsigned i = 0; i < n; ++i)
1858 for (unsigned i = 0; i < n; ++i, ++d)
1861 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->User))
1862 DestTy = UCast->getDestTy();
1863 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->User))
1864 DestTy = SCast->getDestTy();
1865 if (!DestTy) continue;
1868 /* If target does not support DestTy natively then do not apply
1869 this transformation. */
1870 MVT DVT = TLI->getValueType(DestTy);
1871 if (!TLI->isTypeLegal(DVT)) continue;
1874 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1876 if (PH->getNumIncomingValues() != 2) continue;
1878 const Type *SrcTy = PH->getType();
1879 int Mantissa = DestTy->getFPMantissaWidth();
1880 if (Mantissa == -1) continue;
1881 if ((int)TD->getTypeSizeInBits(SrcTy) > Mantissa)
1884 unsigned Entry, Latch;
1885 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1893 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1894 if (!Init) continue;
1895 ConstantFP *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1897 BinaryOperator *Incr =
1898 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1899 if (!Incr) continue;
1900 if (Incr->getOpcode() != Instruction::Add
1901 && Incr->getOpcode() != Instruction::Sub)
1904 /* Initialize new IV, double d = 0.0 in above example. */
1905 ConstantInt *C = NULL;
1906 if (Incr->getOperand(0) == PH)
1907 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1908 else if (Incr->getOperand(1) == PH)
1909 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1915 /* Add new PHINode. */
1916 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1918 /* create new increment. '++d' in above example. */
1919 ConstantFP *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1920 BinaryOperator *NewIncr =
1921 BinaryOperator::Create(Incr->getOpcode(),
1922 NewPH, CFP, "IV.S.next.", Incr);
1924 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1925 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1927 /* Remove cast operation */
1928 SE->deleteValueFromRecords(ShadowUse);
1929 ShadowUse->replaceAllUsesWith(NewPH);
1930 ShadowUse->eraseFromParent();
1931 SI->second.Users.erase(CandidateUI);
1938 // OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar
1939 // uses in the loop, look to see if we can eliminate some, in favor of using
1940 // common indvars for the different uses.
1941 void LoopStrengthReduce::OptimizeIndvars(Loop *L) {
1942 // TODO: implement optzns here.
1944 OptimizeShadowIV(L);
1946 // Finally, get the terminating condition for the loop if possible. If we
1947 // can, we want to change it to use a post-incremented version of its
1948 // induction variable, to allow coalescing the live ranges for the IV into
1949 // one register value.
1950 PHINode *SomePHI = cast<PHINode>(L->getHeader()->begin());
1951 BasicBlock *Preheader = L->getLoopPreheader();
1952 BasicBlock *LatchBlock =
1953 SomePHI->getIncomingBlock(SomePHI->getIncomingBlock(0) == Preheader);
1954 BranchInst *TermBr = dyn_cast<BranchInst>(LatchBlock->getTerminator());
1955 if (!TermBr || TermBr->isUnconditional() ||
1956 !isa<ICmpInst>(TermBr->getCondition()))
1958 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1960 // Search IVUsesByStride to find Cond's IVUse if there is one.
1961 IVStrideUse *CondUse = 0;
1962 const SCEVHandle *CondStride = 0;
1964 if (!FindIVUserForCond(Cond, CondUse, CondStride))
1965 return; // setcc doesn't use the IV.
1967 // If the trip count is computed in terms of an smax (due to ScalarEvolution
1968 // being unable to find a sufficient guard, for example), change the loop
1969 // comparison to use SLT instead of NE.
1970 Cond = OptimizeSMax(L, Cond, CondUse);
1972 // If possible, change stride and operands of the compare instruction to
1973 // eliminate one stride.
1974 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride);
1976 // It's possible for the setcc instruction to be anywhere in the loop, and
1977 // possible for it to have multiple users. If it is not immediately before
1978 // the latch block branch, move it.
1979 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) {
1980 if (Cond->hasOneUse()) { // Condition has a single use, just move it.
1981 Cond->moveBefore(TermBr);
1983 // Otherwise, clone the terminating condition and insert into the loopend.
1984 Cond = cast<ICmpInst>(Cond->clone());
1985 Cond->setName(L->getHeader()->getName() + ".termcond");
1986 LatchBlock->getInstList().insert(TermBr, Cond);
1988 // Clone the IVUse, as the old use still exists!
1989 IVUsesByStride[*CondStride].addUser(CondUse->Offset, Cond,
1990 CondUse->OperandValToReplace);
1991 CondUse = &IVUsesByStride[*CondStride].Users.back();
1995 // If we get to here, we know that we can transform the setcc instruction to
1996 // use the post-incremented version of the IV, allowing us to coalesce the
1997 // live ranges for the IV correctly.
1998 CondUse->Offset = SE->getMinusSCEV(CondUse->Offset, *CondStride);
1999 CondUse->isUseOfPostIncrementedValue = true;
2003 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) {
2005 LI = &getAnalysis<LoopInfo>();
2006 DT = &getAnalysis<DominatorTree>();
2007 SE = &getAnalysis<ScalarEvolution>();
2008 TD = &getAnalysis<TargetData>();
2009 UIntPtrTy = TD->getIntPtrType();
2012 // Find all uses of induction variables in this loop, and catagorize
2013 // them by stride. Start by finding all of the PHI nodes in the header for
2014 // this loop. If they are induction variables, inspect their uses.
2015 SmallPtrSet<Instruction*,16> Processed; // Don't reprocess instructions.
2016 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
2017 AddUsersIfInteresting(I, L, Processed);
2019 if (!IVUsesByStride.empty()) {
2020 // Optimize induction variables. Some indvar uses can be transformed to use
2021 // strides that will be needed for other purposes. A common example of this
2022 // is the exit test for the loop, which can often be rewritten to use the
2023 // computation of some other indvar to decide when to terminate the loop.
2026 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of
2027 // doing computation in byte values, promote to 32-bit values if safe.
2029 // FIXME: Attempt to reuse values across multiple IV's. In particular, we
2030 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should
2031 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC.
2032 // Need to be careful that IV's are all the same type. Only works for
2033 // intptr_t indvars.
2035 // If we only have one stride, we can more aggressively eliminate some
2037 bool HasOneStride = IVUsesByStride.size() == 1;
2040 DOUT << "\nLSR on ";
2044 // IVsByStride keeps IVs for one particular loop.
2045 assert(IVsByStride.empty() && "Stale entries in IVsByStride?");
2047 // Sort the StrideOrder so we process larger strides first.
2048 std::stable_sort(StrideOrder.begin(), StrideOrder.end(), StrideCompare());
2050 // Note: this processes each stride/type pair individually. All users
2051 // passed into StrengthReduceStridedIVUsers have the same type AND stride.
2052 // Also, note that we iterate over IVUsesByStride indirectly by using
2053 // StrideOrder. This extra layer of indirection makes the ordering of
2054 // strides deterministic - not dependent on map order.
2055 for (unsigned Stride = 0, e = StrideOrder.size(); Stride != e; ++Stride) {
2056 std::map<SCEVHandle, IVUsersOfOneStride>::iterator SI =
2057 IVUsesByStride.find(StrideOrder[Stride]);
2058 assert(SI != IVUsesByStride.end() && "Stride doesn't exist!");
2059 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride);
2063 // We're done analyzing this loop; release all the state we built up for it.
2064 CastedPointers.clear();
2065 IVUsesByStride.clear();
2066 IVsByStride.clear();
2067 StrideOrder.clear();
2069 // Clean up after ourselves
2070 if (!DeadInsts.empty()) {
2071 DeleteTriviallyDeadInstructions();
2073 BasicBlock::iterator I = L->getHeader()->begin();
2074 while (PHINode *PN = dyn_cast<PHINode>(I++)) {
2075 // At this point, we know that we have killed one or more IV users.
2076 // It is worth checking to see if the cannonical indvar is also
2077 // dead, so that we can remove it as well.
2079 // We can remove a PHI if it is on a cycle in the def-use graph
2080 // where each node in the cycle has degree one, i.e. only one use,
2081 // and is an instruction with no side effects.
2083 // FIXME: this needs to eliminate an induction variable even if it's being
2084 // compared against some value to decide loop termination.
2085 if (!PN->hasOneUse())
2088 SmallPtrSet<PHINode *, 4> PHIs;
2089 for (Instruction *J = dyn_cast<Instruction>(*PN->use_begin());
2090 J && J->hasOneUse() && !J->mayWriteToMemory();
2091 J = dyn_cast<Instruction>(*J->use_begin())) {
2092 // If we find the original PHI, we've discovered a cycle.
2094 // Break the cycle and mark the PHI for deletion.
2095 SE->deleteValueFromRecords(PN);
2096 PN->replaceAllUsesWith(UndefValue::get(PN->getType()));
2097 DeadInsts.push_back(PN);
2101 // If we find a PHI more than once, we're on a cycle that
2102 // won't prove fruitful.
2103 if (isa<PHINode>(J) && !PHIs.insert(cast<PHINode>(J)))
2107 DeleteTriviallyDeadInstructions();